Loading...
1/*
2 * Copyright (C) 2012 Red Hat. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include "dm-bio-prison.h"
9#include "dm-bio-record.h"
10#include "dm-cache-metadata.h"
11
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
14#include <linux/jiffies.h>
15#include <linux/init.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
18#include <linux/slab.h>
19#include <linux/vmalloc.h>
20
21#define DM_MSG_PREFIX "cache"
22
23DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
24 "A percentage of time allocated for copying to and/or from cache");
25
26/*----------------------------------------------------------------*/
27
28#define IOT_RESOLUTION 4
29
30struct io_tracker {
31 spinlock_t lock;
32
33 /*
34 * Sectors of in-flight IO.
35 */
36 sector_t in_flight;
37
38 /*
39 * The time, in jiffies, when this device became idle (if it is
40 * indeed idle).
41 */
42 unsigned long idle_time;
43 unsigned long last_update_time;
44};
45
46static void iot_init(struct io_tracker *iot)
47{
48 spin_lock_init(&iot->lock);
49 iot->in_flight = 0ul;
50 iot->idle_time = 0ul;
51 iot->last_update_time = jiffies;
52}
53
54static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
55{
56 if (iot->in_flight)
57 return false;
58
59 return time_after(jiffies, iot->idle_time + jifs);
60}
61
62static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
63{
64 bool r;
65 unsigned long flags;
66
67 spin_lock_irqsave(&iot->lock, flags);
68 r = __iot_idle_for(iot, jifs);
69 spin_unlock_irqrestore(&iot->lock, flags);
70
71 return r;
72}
73
74static void iot_io_begin(struct io_tracker *iot, sector_t len)
75{
76 unsigned long flags;
77
78 spin_lock_irqsave(&iot->lock, flags);
79 iot->in_flight += len;
80 spin_unlock_irqrestore(&iot->lock, flags);
81}
82
83static void __iot_io_end(struct io_tracker *iot, sector_t len)
84{
85 iot->in_flight -= len;
86 if (!iot->in_flight)
87 iot->idle_time = jiffies;
88}
89
90static void iot_io_end(struct io_tracker *iot, sector_t len)
91{
92 unsigned long flags;
93
94 spin_lock_irqsave(&iot->lock, flags);
95 __iot_io_end(iot, len);
96 spin_unlock_irqrestore(&iot->lock, flags);
97}
98
99/*----------------------------------------------------------------*/
100
101/*
102 * Glossary:
103 *
104 * oblock: index of an origin block
105 * cblock: index of a cache block
106 * promotion: movement of a block from origin to cache
107 * demotion: movement of a block from cache to origin
108 * migration: movement of a block between the origin and cache device,
109 * either direction
110 */
111
112/*----------------------------------------------------------------*/
113
114/*
115 * There are a couple of places where we let a bio run, but want to do some
116 * work before calling its endio function. We do this by temporarily
117 * changing the endio fn.
118 */
119struct dm_hook_info {
120 bio_end_io_t *bi_end_io;
121};
122
123static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
124 bio_end_io_t *bi_end_io, void *bi_private)
125{
126 h->bi_end_io = bio->bi_end_io;
127
128 bio->bi_end_io = bi_end_io;
129 bio->bi_private = bi_private;
130}
131
132static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
133{
134 bio->bi_end_io = h->bi_end_io;
135}
136
137/*----------------------------------------------------------------*/
138
139#define MIGRATION_POOL_SIZE 128
140#define COMMIT_PERIOD HZ
141#define MIGRATION_COUNT_WINDOW 10
142
143/*
144 * The block size of the device holding cache data must be
145 * between 32KB and 1GB.
146 */
147#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
148#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
149
150enum cache_metadata_mode {
151 CM_WRITE, /* metadata may be changed */
152 CM_READ_ONLY, /* metadata may not be changed */
153 CM_FAIL
154};
155
156enum cache_io_mode {
157 /*
158 * Data is written to cached blocks only. These blocks are marked
159 * dirty. If you lose the cache device you will lose data.
160 * Potential performance increase for both reads and writes.
161 */
162 CM_IO_WRITEBACK,
163
164 /*
165 * Data is written to both cache and origin. Blocks are never
166 * dirty. Potential performance benfit for reads only.
167 */
168 CM_IO_WRITETHROUGH,
169
170 /*
171 * A degraded mode useful for various cache coherency situations
172 * (eg, rolling back snapshots). Reads and writes always go to the
173 * origin. If a write goes to a cached oblock, then the cache
174 * block is invalidated.
175 */
176 CM_IO_PASSTHROUGH
177};
178
179struct cache_features {
180 enum cache_metadata_mode mode;
181 enum cache_io_mode io_mode;
182};
183
184struct cache_stats {
185 atomic_t read_hit;
186 atomic_t read_miss;
187 atomic_t write_hit;
188 atomic_t write_miss;
189 atomic_t demotion;
190 atomic_t promotion;
191 atomic_t copies_avoided;
192 atomic_t cache_cell_clash;
193 atomic_t commit_count;
194 atomic_t discard_count;
195};
196
197/*
198 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
199 * the one-past-the-end value.
200 */
201struct cblock_range {
202 dm_cblock_t begin;
203 dm_cblock_t end;
204};
205
206struct invalidation_request {
207 struct list_head list;
208 struct cblock_range *cblocks;
209
210 atomic_t complete;
211 int err;
212
213 wait_queue_head_t result_wait;
214};
215
216struct cache {
217 struct dm_target *ti;
218 struct dm_target_callbacks callbacks;
219
220 struct dm_cache_metadata *cmd;
221
222 /*
223 * Metadata is written to this device.
224 */
225 struct dm_dev *metadata_dev;
226
227 /*
228 * The slower of the two data devices. Typically a spindle.
229 */
230 struct dm_dev *origin_dev;
231
232 /*
233 * The faster of the two data devices. Typically an SSD.
234 */
235 struct dm_dev *cache_dev;
236
237 /*
238 * Size of the origin device in _complete_ blocks and native sectors.
239 */
240 dm_oblock_t origin_blocks;
241 sector_t origin_sectors;
242
243 /*
244 * Size of the cache device in blocks.
245 */
246 dm_cblock_t cache_size;
247
248 /*
249 * Fields for converting from sectors to blocks.
250 */
251 sector_t sectors_per_block;
252 int sectors_per_block_shift;
253
254 spinlock_t lock;
255 struct list_head deferred_cells;
256 struct bio_list deferred_bios;
257 struct bio_list deferred_flush_bios;
258 struct bio_list deferred_writethrough_bios;
259 struct list_head quiesced_migrations;
260 struct list_head completed_migrations;
261 struct list_head need_commit_migrations;
262 sector_t migration_threshold;
263 wait_queue_head_t migration_wait;
264 atomic_t nr_allocated_migrations;
265
266 /*
267 * The number of in flight migrations that are performing
268 * background io. eg, promotion, writeback.
269 */
270 atomic_t nr_io_migrations;
271
272 wait_queue_head_t quiescing_wait;
273 atomic_t quiescing;
274 atomic_t quiescing_ack;
275
276 /*
277 * cache_size entries, dirty if set
278 */
279 atomic_t nr_dirty;
280 unsigned long *dirty_bitset;
281
282 /*
283 * origin_blocks entries, discarded if set.
284 */
285 dm_dblock_t discard_nr_blocks;
286 unsigned long *discard_bitset;
287 uint32_t discard_block_size; /* a power of 2 times sectors per block */
288
289 /*
290 * Rather than reconstructing the table line for the status we just
291 * save it and regurgitate.
292 */
293 unsigned nr_ctr_args;
294 const char **ctr_args;
295
296 struct dm_kcopyd_client *copier;
297 struct workqueue_struct *wq;
298 struct work_struct worker;
299
300 struct delayed_work waker;
301 unsigned long last_commit_jiffies;
302
303 struct dm_bio_prison *prison;
304 struct dm_deferred_set *all_io_ds;
305
306 mempool_t *migration_pool;
307
308 struct dm_cache_policy *policy;
309 unsigned policy_nr_args;
310
311 bool need_tick_bio:1;
312 bool sized:1;
313 bool invalidate:1;
314 bool commit_requested:1;
315 bool loaded_mappings:1;
316 bool loaded_discards:1;
317
318 /*
319 * Cache features such as write-through.
320 */
321 struct cache_features features;
322
323 struct cache_stats stats;
324
325 /*
326 * Invalidation fields.
327 */
328 spinlock_t invalidation_lock;
329 struct list_head invalidation_requests;
330
331 struct io_tracker origin_tracker;
332};
333
334struct per_bio_data {
335 bool tick:1;
336 unsigned req_nr:2;
337 struct dm_deferred_entry *all_io_entry;
338 struct dm_hook_info hook_info;
339 sector_t len;
340
341 /*
342 * writethrough fields. These MUST remain at the end of this
343 * structure and the 'cache' member must be the first as it
344 * is used to determine the offset of the writethrough fields.
345 */
346 struct cache *cache;
347 dm_cblock_t cblock;
348 struct dm_bio_details bio_details;
349};
350
351struct dm_cache_migration {
352 struct list_head list;
353 struct cache *cache;
354
355 unsigned long start_jiffies;
356 dm_oblock_t old_oblock;
357 dm_oblock_t new_oblock;
358 dm_cblock_t cblock;
359
360 bool err:1;
361 bool discard:1;
362 bool writeback:1;
363 bool demote:1;
364 bool promote:1;
365 bool requeue_holder:1;
366 bool invalidate:1;
367
368 struct dm_bio_prison_cell *old_ocell;
369 struct dm_bio_prison_cell *new_ocell;
370};
371
372/*
373 * Processing a bio in the worker thread may require these memory
374 * allocations. We prealloc to avoid deadlocks (the same worker thread
375 * frees them back to the mempool).
376 */
377struct prealloc {
378 struct dm_cache_migration *mg;
379 struct dm_bio_prison_cell *cell1;
380 struct dm_bio_prison_cell *cell2;
381};
382
383static enum cache_metadata_mode get_cache_mode(struct cache *cache);
384
385static void wake_worker(struct cache *cache)
386{
387 queue_work(cache->wq, &cache->worker);
388}
389
390/*----------------------------------------------------------------*/
391
392static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
393{
394 /* FIXME: change to use a local slab. */
395 return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
396}
397
398static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
399{
400 dm_bio_prison_free_cell(cache->prison, cell);
401}
402
403static struct dm_cache_migration *alloc_migration(struct cache *cache)
404{
405 struct dm_cache_migration *mg;
406
407 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
408 if (mg) {
409 mg->cache = cache;
410 atomic_inc(&mg->cache->nr_allocated_migrations);
411 }
412
413 return mg;
414}
415
416static void free_migration(struct dm_cache_migration *mg)
417{
418 struct cache *cache = mg->cache;
419
420 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
421 wake_up(&cache->migration_wait);
422
423 mempool_free(mg, cache->migration_pool);
424}
425
426static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
427{
428 if (!p->mg) {
429 p->mg = alloc_migration(cache);
430 if (!p->mg)
431 return -ENOMEM;
432 }
433
434 if (!p->cell1) {
435 p->cell1 = alloc_prison_cell(cache);
436 if (!p->cell1)
437 return -ENOMEM;
438 }
439
440 if (!p->cell2) {
441 p->cell2 = alloc_prison_cell(cache);
442 if (!p->cell2)
443 return -ENOMEM;
444 }
445
446 return 0;
447}
448
449static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
450{
451 if (p->cell2)
452 free_prison_cell(cache, p->cell2);
453
454 if (p->cell1)
455 free_prison_cell(cache, p->cell1);
456
457 if (p->mg)
458 free_migration(p->mg);
459}
460
461static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
462{
463 struct dm_cache_migration *mg = p->mg;
464
465 BUG_ON(!mg);
466 p->mg = NULL;
467
468 return mg;
469}
470
471/*
472 * You must have a cell within the prealloc struct to return. If not this
473 * function will BUG() rather than returning NULL.
474 */
475static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
476{
477 struct dm_bio_prison_cell *r = NULL;
478
479 if (p->cell1) {
480 r = p->cell1;
481 p->cell1 = NULL;
482
483 } else if (p->cell2) {
484 r = p->cell2;
485 p->cell2 = NULL;
486 } else
487 BUG();
488
489 return r;
490}
491
492/*
493 * You can't have more than two cells in a prealloc struct. BUG() will be
494 * called if you try and overfill.
495 */
496static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
497{
498 if (!p->cell2)
499 p->cell2 = cell;
500
501 else if (!p->cell1)
502 p->cell1 = cell;
503
504 else
505 BUG();
506}
507
508/*----------------------------------------------------------------*/
509
510static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key *key)
511{
512 key->virtual = 0;
513 key->dev = 0;
514 key->block_begin = from_oblock(begin);
515 key->block_end = from_oblock(end);
516}
517
518/*
519 * The caller hands in a preallocated cell, and a free function for it.
520 * The cell will be freed if there's an error, or if it wasn't used because
521 * a cell with that key already exists.
522 */
523typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
524
525static int bio_detain_range(struct cache *cache, dm_oblock_t oblock_begin, dm_oblock_t oblock_end,
526 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
527 cell_free_fn free_fn, void *free_context,
528 struct dm_bio_prison_cell **cell_result)
529{
530 int r;
531 struct dm_cell_key key;
532
533 build_key(oblock_begin, oblock_end, &key);
534 r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
535 if (r)
536 free_fn(free_context, cell_prealloc);
537
538 return r;
539}
540
541static int bio_detain(struct cache *cache, dm_oblock_t oblock,
542 struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
543 cell_free_fn free_fn, void *free_context,
544 struct dm_bio_prison_cell **cell_result)
545{
546 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
547 return bio_detain_range(cache, oblock, end, bio,
548 cell_prealloc, free_fn, free_context, cell_result);
549}
550
551static int get_cell(struct cache *cache,
552 dm_oblock_t oblock,
553 struct prealloc *structs,
554 struct dm_bio_prison_cell **cell_result)
555{
556 int r;
557 struct dm_cell_key key;
558 struct dm_bio_prison_cell *cell_prealloc;
559
560 cell_prealloc = prealloc_get_cell(structs);
561
562 build_key(oblock, to_oblock(from_oblock(oblock) + 1ULL), &key);
563 r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
564 if (r)
565 prealloc_put_cell(structs, cell_prealloc);
566
567 return r;
568}
569
570/*----------------------------------------------------------------*/
571
572static bool is_dirty(struct cache *cache, dm_cblock_t b)
573{
574 return test_bit(from_cblock(b), cache->dirty_bitset);
575}
576
577static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
578{
579 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
580 atomic_inc(&cache->nr_dirty);
581 policy_set_dirty(cache->policy, oblock);
582 }
583}
584
585static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
586{
587 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
588 policy_clear_dirty(cache->policy, oblock);
589 if (atomic_dec_return(&cache->nr_dirty) == 0)
590 dm_table_event(cache->ti->table);
591 }
592}
593
594/*----------------------------------------------------------------*/
595
596static bool block_size_is_power_of_two(struct cache *cache)
597{
598 return cache->sectors_per_block_shift >= 0;
599}
600
601/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
602#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
603__always_inline
604#endif
605static dm_block_t block_div(dm_block_t b, uint32_t n)
606{
607 do_div(b, n);
608
609 return b;
610}
611
612static dm_block_t oblocks_per_dblock(struct cache *cache)
613{
614 dm_block_t oblocks = cache->discard_block_size;
615
616 if (block_size_is_power_of_two(cache))
617 oblocks >>= cache->sectors_per_block_shift;
618 else
619 oblocks = block_div(oblocks, cache->sectors_per_block);
620
621 return oblocks;
622}
623
624static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
625{
626 return to_dblock(block_div(from_oblock(oblock),
627 oblocks_per_dblock(cache)));
628}
629
630static dm_oblock_t dblock_to_oblock(struct cache *cache, dm_dblock_t dblock)
631{
632 return to_oblock(from_dblock(dblock) * oblocks_per_dblock(cache));
633}
634
635static void set_discard(struct cache *cache, dm_dblock_t b)
636{
637 unsigned long flags;
638
639 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
640 atomic_inc(&cache->stats.discard_count);
641
642 spin_lock_irqsave(&cache->lock, flags);
643 set_bit(from_dblock(b), cache->discard_bitset);
644 spin_unlock_irqrestore(&cache->lock, flags);
645}
646
647static void clear_discard(struct cache *cache, dm_dblock_t b)
648{
649 unsigned long flags;
650
651 spin_lock_irqsave(&cache->lock, flags);
652 clear_bit(from_dblock(b), cache->discard_bitset);
653 spin_unlock_irqrestore(&cache->lock, flags);
654}
655
656static bool is_discarded(struct cache *cache, dm_dblock_t b)
657{
658 int r;
659 unsigned long flags;
660
661 spin_lock_irqsave(&cache->lock, flags);
662 r = test_bit(from_dblock(b), cache->discard_bitset);
663 spin_unlock_irqrestore(&cache->lock, flags);
664
665 return r;
666}
667
668static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
669{
670 int r;
671 unsigned long flags;
672
673 spin_lock_irqsave(&cache->lock, flags);
674 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
675 cache->discard_bitset);
676 spin_unlock_irqrestore(&cache->lock, flags);
677
678 return r;
679}
680
681/*----------------------------------------------------------------*/
682
683static void load_stats(struct cache *cache)
684{
685 struct dm_cache_statistics stats;
686
687 dm_cache_metadata_get_stats(cache->cmd, &stats);
688 atomic_set(&cache->stats.read_hit, stats.read_hits);
689 atomic_set(&cache->stats.read_miss, stats.read_misses);
690 atomic_set(&cache->stats.write_hit, stats.write_hits);
691 atomic_set(&cache->stats.write_miss, stats.write_misses);
692}
693
694static void save_stats(struct cache *cache)
695{
696 struct dm_cache_statistics stats;
697
698 if (get_cache_mode(cache) >= CM_READ_ONLY)
699 return;
700
701 stats.read_hits = atomic_read(&cache->stats.read_hit);
702 stats.read_misses = atomic_read(&cache->stats.read_miss);
703 stats.write_hits = atomic_read(&cache->stats.write_hit);
704 stats.write_misses = atomic_read(&cache->stats.write_miss);
705
706 dm_cache_metadata_set_stats(cache->cmd, &stats);
707}
708
709/*----------------------------------------------------------------
710 * Per bio data
711 *--------------------------------------------------------------*/
712
713/*
714 * If using writeback, leave out struct per_bio_data's writethrough fields.
715 */
716#define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
717#define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
718
719static bool writethrough_mode(struct cache_features *f)
720{
721 return f->io_mode == CM_IO_WRITETHROUGH;
722}
723
724static bool writeback_mode(struct cache_features *f)
725{
726 return f->io_mode == CM_IO_WRITEBACK;
727}
728
729static bool passthrough_mode(struct cache_features *f)
730{
731 return f->io_mode == CM_IO_PASSTHROUGH;
732}
733
734static size_t get_per_bio_data_size(struct cache *cache)
735{
736 return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
737}
738
739static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
740{
741 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
742 BUG_ON(!pb);
743 return pb;
744}
745
746static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
747{
748 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
749
750 pb->tick = false;
751 pb->req_nr = dm_bio_get_target_bio_nr(bio);
752 pb->all_io_entry = NULL;
753 pb->len = 0;
754
755 return pb;
756}
757
758/*----------------------------------------------------------------
759 * Remapping
760 *--------------------------------------------------------------*/
761static void remap_to_origin(struct cache *cache, struct bio *bio)
762{
763 bio->bi_bdev = cache->origin_dev->bdev;
764}
765
766static void remap_to_cache(struct cache *cache, struct bio *bio,
767 dm_cblock_t cblock)
768{
769 sector_t bi_sector = bio->bi_iter.bi_sector;
770 sector_t block = from_cblock(cblock);
771
772 bio->bi_bdev = cache->cache_dev->bdev;
773 if (!block_size_is_power_of_two(cache))
774 bio->bi_iter.bi_sector =
775 (block * cache->sectors_per_block) +
776 sector_div(bi_sector, cache->sectors_per_block);
777 else
778 bio->bi_iter.bi_sector =
779 (block << cache->sectors_per_block_shift) |
780 (bi_sector & (cache->sectors_per_block - 1));
781}
782
783static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
784{
785 unsigned long flags;
786 size_t pb_data_size = get_per_bio_data_size(cache);
787 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
788
789 spin_lock_irqsave(&cache->lock, flags);
790 if (cache->need_tick_bio &&
791 !(bio->bi_opf & (REQ_FUA | REQ_PREFLUSH)) &&
792 bio_op(bio) != REQ_OP_DISCARD) {
793 pb->tick = true;
794 cache->need_tick_bio = false;
795 }
796 spin_unlock_irqrestore(&cache->lock, flags);
797}
798
799static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
800 dm_oblock_t oblock)
801{
802 check_if_tick_bio_needed(cache, bio);
803 remap_to_origin(cache, bio);
804 if (bio_data_dir(bio) == WRITE)
805 clear_discard(cache, oblock_to_dblock(cache, oblock));
806}
807
808static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
809 dm_oblock_t oblock, dm_cblock_t cblock)
810{
811 check_if_tick_bio_needed(cache, bio);
812 remap_to_cache(cache, bio, cblock);
813 if (bio_data_dir(bio) == WRITE) {
814 set_dirty(cache, oblock, cblock);
815 clear_discard(cache, oblock_to_dblock(cache, oblock));
816 }
817}
818
819static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
820{
821 sector_t block_nr = bio->bi_iter.bi_sector;
822
823 if (!block_size_is_power_of_two(cache))
824 (void) sector_div(block_nr, cache->sectors_per_block);
825 else
826 block_nr >>= cache->sectors_per_block_shift;
827
828 return to_oblock(block_nr);
829}
830
831static int bio_triggers_commit(struct cache *cache, struct bio *bio)
832{
833 return bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
834}
835
836/*
837 * You must increment the deferred set whilst the prison cell is held. To
838 * encourage this, we ask for 'cell' to be passed in.
839 */
840static void inc_ds(struct cache *cache, struct bio *bio,
841 struct dm_bio_prison_cell *cell)
842{
843 size_t pb_data_size = get_per_bio_data_size(cache);
844 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
845
846 BUG_ON(!cell);
847 BUG_ON(pb->all_io_entry);
848
849 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
850}
851
852static bool accountable_bio(struct cache *cache, struct bio *bio)
853{
854 return ((bio->bi_bdev == cache->origin_dev->bdev) &&
855 bio_op(bio) != REQ_OP_DISCARD);
856}
857
858static void accounted_begin(struct cache *cache, struct bio *bio)
859{
860 size_t pb_data_size = get_per_bio_data_size(cache);
861 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
862
863 if (accountable_bio(cache, bio)) {
864 pb->len = bio_sectors(bio);
865 iot_io_begin(&cache->origin_tracker, pb->len);
866 }
867}
868
869static void accounted_complete(struct cache *cache, struct bio *bio)
870{
871 size_t pb_data_size = get_per_bio_data_size(cache);
872 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
873
874 iot_io_end(&cache->origin_tracker, pb->len);
875}
876
877static void accounted_request(struct cache *cache, struct bio *bio)
878{
879 accounted_begin(cache, bio);
880 generic_make_request(bio);
881}
882
883static void issue(struct cache *cache, struct bio *bio)
884{
885 unsigned long flags;
886
887 if (!bio_triggers_commit(cache, bio)) {
888 accounted_request(cache, bio);
889 return;
890 }
891
892 /*
893 * Batch together any bios that trigger commits and then issue a
894 * single commit for them in do_worker().
895 */
896 spin_lock_irqsave(&cache->lock, flags);
897 cache->commit_requested = true;
898 bio_list_add(&cache->deferred_flush_bios, bio);
899 spin_unlock_irqrestore(&cache->lock, flags);
900}
901
902static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
903{
904 inc_ds(cache, bio, cell);
905 issue(cache, bio);
906}
907
908static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
909{
910 unsigned long flags;
911
912 spin_lock_irqsave(&cache->lock, flags);
913 bio_list_add(&cache->deferred_writethrough_bios, bio);
914 spin_unlock_irqrestore(&cache->lock, flags);
915
916 wake_worker(cache);
917}
918
919static void writethrough_endio(struct bio *bio)
920{
921 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
922
923 dm_unhook_bio(&pb->hook_info, bio);
924
925 if (bio->bi_error) {
926 bio_endio(bio);
927 return;
928 }
929
930 dm_bio_restore(&pb->bio_details, bio);
931 remap_to_cache(pb->cache, bio, pb->cblock);
932
933 /*
934 * We can't issue this bio directly, since we're in interrupt
935 * context. So it gets put on a bio list for processing by the
936 * worker thread.
937 */
938 defer_writethrough_bio(pb->cache, bio);
939}
940
941/*
942 * When running in writethrough mode we need to send writes to clean blocks
943 * to both the cache and origin devices. In future we'd like to clone the
944 * bio and send them in parallel, but for now we're doing them in
945 * series as this is easier.
946 */
947static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
948 dm_oblock_t oblock, dm_cblock_t cblock)
949{
950 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
951
952 pb->cache = cache;
953 pb->cblock = cblock;
954 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
955 dm_bio_record(&pb->bio_details, bio);
956
957 remap_to_origin_clear_discard(pb->cache, bio, oblock);
958}
959
960/*----------------------------------------------------------------
961 * Failure modes
962 *--------------------------------------------------------------*/
963static enum cache_metadata_mode get_cache_mode(struct cache *cache)
964{
965 return cache->features.mode;
966}
967
968static const char *cache_device_name(struct cache *cache)
969{
970 return dm_device_name(dm_table_get_md(cache->ti->table));
971}
972
973static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
974{
975 const char *descs[] = {
976 "write",
977 "read-only",
978 "fail"
979 };
980
981 dm_table_event(cache->ti->table);
982 DMINFO("%s: switching cache to %s mode",
983 cache_device_name(cache), descs[(int)mode]);
984}
985
986static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
987{
988 bool needs_check;
989 enum cache_metadata_mode old_mode = get_cache_mode(cache);
990
991 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
992 DMERR("%s: unable to read needs_check flag, setting failure mode.",
993 cache_device_name(cache));
994 new_mode = CM_FAIL;
995 }
996
997 if (new_mode == CM_WRITE && needs_check) {
998 DMERR("%s: unable to switch cache to write mode until repaired.",
999 cache_device_name(cache));
1000 if (old_mode != new_mode)
1001 new_mode = old_mode;
1002 else
1003 new_mode = CM_READ_ONLY;
1004 }
1005
1006 /* Never move out of fail mode */
1007 if (old_mode == CM_FAIL)
1008 new_mode = CM_FAIL;
1009
1010 switch (new_mode) {
1011 case CM_FAIL:
1012 case CM_READ_ONLY:
1013 dm_cache_metadata_set_read_only(cache->cmd);
1014 break;
1015
1016 case CM_WRITE:
1017 dm_cache_metadata_set_read_write(cache->cmd);
1018 break;
1019 }
1020
1021 cache->features.mode = new_mode;
1022
1023 if (new_mode != old_mode)
1024 notify_mode_switch(cache, new_mode);
1025}
1026
1027static void abort_transaction(struct cache *cache)
1028{
1029 const char *dev_name = cache_device_name(cache);
1030
1031 if (get_cache_mode(cache) >= CM_READ_ONLY)
1032 return;
1033
1034 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1035 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1036 set_cache_mode(cache, CM_FAIL);
1037 }
1038
1039 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1040 if (dm_cache_metadata_abort(cache->cmd)) {
1041 DMERR("%s: failed to abort metadata transaction", dev_name);
1042 set_cache_mode(cache, CM_FAIL);
1043 }
1044}
1045
1046static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1047{
1048 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1049 cache_device_name(cache), op, r);
1050 abort_transaction(cache);
1051 set_cache_mode(cache, CM_READ_ONLY);
1052}
1053
1054/*----------------------------------------------------------------
1055 * Migration processing
1056 *
1057 * Migration covers moving data from the origin device to the cache, or
1058 * vice versa.
1059 *--------------------------------------------------------------*/
1060static void inc_io_migrations(struct cache *cache)
1061{
1062 atomic_inc(&cache->nr_io_migrations);
1063}
1064
1065static void dec_io_migrations(struct cache *cache)
1066{
1067 atomic_dec(&cache->nr_io_migrations);
1068}
1069
1070static bool discard_or_flush(struct bio *bio)
1071{
1072 return bio_op(bio) == REQ_OP_DISCARD ||
1073 bio->bi_opf & (REQ_PREFLUSH | REQ_FUA);
1074}
1075
1076static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1077{
1078 if (discard_or_flush(cell->holder)) {
1079 /*
1080 * We have to handle these bios individually.
1081 */
1082 dm_cell_release(cache->prison, cell, &cache->deferred_bios);
1083 free_prison_cell(cache, cell);
1084 } else
1085 list_add_tail(&cell->user_list, &cache->deferred_cells);
1086}
1087
1088static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1089{
1090 unsigned long flags;
1091
1092 if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1093 /*
1094 * There was no prisoner to promote to holder, the
1095 * cell has been released.
1096 */
1097 free_prison_cell(cache, cell);
1098 return;
1099 }
1100
1101 spin_lock_irqsave(&cache->lock, flags);
1102 __cell_defer(cache, cell);
1103 spin_unlock_irqrestore(&cache->lock, flags);
1104
1105 wake_worker(cache);
1106}
1107
1108static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1109{
1110 dm_cell_error(cache->prison, cell, err);
1111 free_prison_cell(cache, cell);
1112}
1113
1114static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1115{
1116 cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1117}
1118
1119static void free_io_migration(struct dm_cache_migration *mg)
1120{
1121 struct cache *cache = mg->cache;
1122
1123 dec_io_migrations(cache);
1124 free_migration(mg);
1125 wake_worker(cache);
1126}
1127
1128static void migration_failure(struct dm_cache_migration *mg)
1129{
1130 struct cache *cache = mg->cache;
1131 const char *dev_name = cache_device_name(cache);
1132
1133 if (mg->writeback) {
1134 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1135 set_dirty(cache, mg->old_oblock, mg->cblock);
1136 cell_defer(cache, mg->old_ocell, false);
1137
1138 } else if (mg->demote) {
1139 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1140 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1141
1142 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1143 if (mg->promote)
1144 cell_defer(cache, mg->new_ocell, true);
1145 } else {
1146 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1147 policy_remove_mapping(cache->policy, mg->new_oblock);
1148 cell_defer(cache, mg->new_ocell, true);
1149 }
1150
1151 free_io_migration(mg);
1152}
1153
1154static void migration_success_pre_commit(struct dm_cache_migration *mg)
1155{
1156 int r;
1157 unsigned long flags;
1158 struct cache *cache = mg->cache;
1159
1160 if (mg->writeback) {
1161 clear_dirty(cache, mg->old_oblock, mg->cblock);
1162 cell_defer(cache, mg->old_ocell, false);
1163 free_io_migration(mg);
1164 return;
1165
1166 } else if (mg->demote) {
1167 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1168 if (r) {
1169 DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1170 cache_device_name(cache));
1171 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1172 policy_force_mapping(cache->policy, mg->new_oblock,
1173 mg->old_oblock);
1174 if (mg->promote)
1175 cell_defer(cache, mg->new_ocell, true);
1176 free_io_migration(mg);
1177 return;
1178 }
1179 } else {
1180 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1181 if (r) {
1182 DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1183 cache_device_name(cache));
1184 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1185 policy_remove_mapping(cache->policy, mg->new_oblock);
1186 free_io_migration(mg);
1187 return;
1188 }
1189 }
1190
1191 spin_lock_irqsave(&cache->lock, flags);
1192 list_add_tail(&mg->list, &cache->need_commit_migrations);
1193 cache->commit_requested = true;
1194 spin_unlock_irqrestore(&cache->lock, flags);
1195}
1196
1197static void migration_success_post_commit(struct dm_cache_migration *mg)
1198{
1199 unsigned long flags;
1200 struct cache *cache = mg->cache;
1201
1202 if (mg->writeback) {
1203 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1204 cache_device_name(cache));
1205 return;
1206
1207 } else if (mg->demote) {
1208 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1209
1210 if (mg->promote) {
1211 mg->demote = false;
1212
1213 spin_lock_irqsave(&cache->lock, flags);
1214 list_add_tail(&mg->list, &cache->quiesced_migrations);
1215 spin_unlock_irqrestore(&cache->lock, flags);
1216
1217 } else {
1218 if (mg->invalidate)
1219 policy_remove_mapping(cache->policy, mg->old_oblock);
1220 free_io_migration(mg);
1221 }
1222
1223 } else {
1224 if (mg->requeue_holder) {
1225 clear_dirty(cache, mg->new_oblock, mg->cblock);
1226 cell_defer(cache, mg->new_ocell, true);
1227 } else {
1228 /*
1229 * The block was promoted via an overwrite, so it's dirty.
1230 */
1231 set_dirty(cache, mg->new_oblock, mg->cblock);
1232 bio_endio(mg->new_ocell->holder);
1233 cell_defer(cache, mg->new_ocell, false);
1234 }
1235 free_io_migration(mg);
1236 }
1237}
1238
1239static void copy_complete(int read_err, unsigned long write_err, void *context)
1240{
1241 unsigned long flags;
1242 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1243 struct cache *cache = mg->cache;
1244
1245 if (read_err || write_err)
1246 mg->err = true;
1247
1248 spin_lock_irqsave(&cache->lock, flags);
1249 list_add_tail(&mg->list, &cache->completed_migrations);
1250 spin_unlock_irqrestore(&cache->lock, flags);
1251
1252 wake_worker(cache);
1253}
1254
1255static void issue_copy(struct dm_cache_migration *mg)
1256{
1257 int r;
1258 struct dm_io_region o_region, c_region;
1259 struct cache *cache = mg->cache;
1260 sector_t cblock = from_cblock(mg->cblock);
1261
1262 o_region.bdev = cache->origin_dev->bdev;
1263 o_region.count = cache->sectors_per_block;
1264
1265 c_region.bdev = cache->cache_dev->bdev;
1266 c_region.sector = cblock * cache->sectors_per_block;
1267 c_region.count = cache->sectors_per_block;
1268
1269 if (mg->writeback || mg->demote) {
1270 /* demote */
1271 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1272 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1273 } else {
1274 /* promote */
1275 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1276 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1277 }
1278
1279 if (r < 0) {
1280 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1281 migration_failure(mg);
1282 }
1283}
1284
1285static void overwrite_endio(struct bio *bio)
1286{
1287 struct dm_cache_migration *mg = bio->bi_private;
1288 struct cache *cache = mg->cache;
1289 size_t pb_data_size = get_per_bio_data_size(cache);
1290 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1291 unsigned long flags;
1292
1293 dm_unhook_bio(&pb->hook_info, bio);
1294
1295 if (bio->bi_error)
1296 mg->err = true;
1297
1298 mg->requeue_holder = false;
1299
1300 spin_lock_irqsave(&cache->lock, flags);
1301 list_add_tail(&mg->list, &cache->completed_migrations);
1302 spin_unlock_irqrestore(&cache->lock, flags);
1303
1304 wake_worker(cache);
1305}
1306
1307static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1308{
1309 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1310 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1311
1312 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1313 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1314
1315 /*
1316 * No need to inc_ds() here, since the cell will be held for the
1317 * duration of the io.
1318 */
1319 accounted_request(mg->cache, bio);
1320}
1321
1322static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1323{
1324 return (bio_data_dir(bio) == WRITE) &&
1325 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1326}
1327
1328static void avoid_copy(struct dm_cache_migration *mg)
1329{
1330 atomic_inc(&mg->cache->stats.copies_avoided);
1331 migration_success_pre_commit(mg);
1332}
1333
1334static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1335 dm_dblock_t *b, dm_dblock_t *e)
1336{
1337 sector_t sb = bio->bi_iter.bi_sector;
1338 sector_t se = bio_end_sector(bio);
1339
1340 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1341
1342 if (se - sb < cache->discard_block_size)
1343 *e = *b;
1344 else
1345 *e = to_dblock(block_div(se, cache->discard_block_size));
1346}
1347
1348static void issue_discard(struct dm_cache_migration *mg)
1349{
1350 dm_dblock_t b, e;
1351 struct bio *bio = mg->new_ocell->holder;
1352 struct cache *cache = mg->cache;
1353
1354 calc_discard_block_range(cache, bio, &b, &e);
1355 while (b != e) {
1356 set_discard(cache, b);
1357 b = to_dblock(from_dblock(b) + 1);
1358 }
1359
1360 bio_endio(bio);
1361 cell_defer(cache, mg->new_ocell, false);
1362 free_migration(mg);
1363 wake_worker(cache);
1364}
1365
1366static void issue_copy_or_discard(struct dm_cache_migration *mg)
1367{
1368 bool avoid;
1369 struct cache *cache = mg->cache;
1370
1371 if (mg->discard) {
1372 issue_discard(mg);
1373 return;
1374 }
1375
1376 if (mg->writeback || mg->demote)
1377 avoid = !is_dirty(cache, mg->cblock) ||
1378 is_discarded_oblock(cache, mg->old_oblock);
1379 else {
1380 struct bio *bio = mg->new_ocell->holder;
1381
1382 avoid = is_discarded_oblock(cache, mg->new_oblock);
1383
1384 if (writeback_mode(&cache->features) &&
1385 !avoid && bio_writes_complete_block(cache, bio)) {
1386 issue_overwrite(mg, bio);
1387 return;
1388 }
1389 }
1390
1391 avoid ? avoid_copy(mg) : issue_copy(mg);
1392}
1393
1394static void complete_migration(struct dm_cache_migration *mg)
1395{
1396 if (mg->err)
1397 migration_failure(mg);
1398 else
1399 migration_success_pre_commit(mg);
1400}
1401
1402static void process_migrations(struct cache *cache, struct list_head *head,
1403 void (*fn)(struct dm_cache_migration *))
1404{
1405 unsigned long flags;
1406 struct list_head list;
1407 struct dm_cache_migration *mg, *tmp;
1408
1409 INIT_LIST_HEAD(&list);
1410 spin_lock_irqsave(&cache->lock, flags);
1411 list_splice_init(head, &list);
1412 spin_unlock_irqrestore(&cache->lock, flags);
1413
1414 list_for_each_entry_safe(mg, tmp, &list, list)
1415 fn(mg);
1416}
1417
1418static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1419{
1420 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1421}
1422
1423static void queue_quiesced_migration(struct dm_cache_migration *mg)
1424{
1425 unsigned long flags;
1426 struct cache *cache = mg->cache;
1427
1428 spin_lock_irqsave(&cache->lock, flags);
1429 __queue_quiesced_migration(mg);
1430 spin_unlock_irqrestore(&cache->lock, flags);
1431
1432 wake_worker(cache);
1433}
1434
1435static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1436{
1437 unsigned long flags;
1438 struct dm_cache_migration *mg, *tmp;
1439
1440 spin_lock_irqsave(&cache->lock, flags);
1441 list_for_each_entry_safe(mg, tmp, work, list)
1442 __queue_quiesced_migration(mg);
1443 spin_unlock_irqrestore(&cache->lock, flags);
1444
1445 wake_worker(cache);
1446}
1447
1448static void check_for_quiesced_migrations(struct cache *cache,
1449 struct per_bio_data *pb)
1450{
1451 struct list_head work;
1452
1453 if (!pb->all_io_entry)
1454 return;
1455
1456 INIT_LIST_HEAD(&work);
1457 dm_deferred_entry_dec(pb->all_io_entry, &work);
1458
1459 if (!list_empty(&work))
1460 queue_quiesced_migrations(cache, &work);
1461}
1462
1463static void quiesce_migration(struct dm_cache_migration *mg)
1464{
1465 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1466 queue_quiesced_migration(mg);
1467}
1468
1469static void promote(struct cache *cache, struct prealloc *structs,
1470 dm_oblock_t oblock, dm_cblock_t cblock,
1471 struct dm_bio_prison_cell *cell)
1472{
1473 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1474
1475 mg->err = false;
1476 mg->discard = false;
1477 mg->writeback = false;
1478 mg->demote = false;
1479 mg->promote = true;
1480 mg->requeue_holder = true;
1481 mg->invalidate = false;
1482 mg->cache = cache;
1483 mg->new_oblock = oblock;
1484 mg->cblock = cblock;
1485 mg->old_ocell = NULL;
1486 mg->new_ocell = cell;
1487 mg->start_jiffies = jiffies;
1488
1489 inc_io_migrations(cache);
1490 quiesce_migration(mg);
1491}
1492
1493static void writeback(struct cache *cache, struct prealloc *structs,
1494 dm_oblock_t oblock, dm_cblock_t cblock,
1495 struct dm_bio_prison_cell *cell)
1496{
1497 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1498
1499 mg->err = false;
1500 mg->discard = false;
1501 mg->writeback = true;
1502 mg->demote = false;
1503 mg->promote = false;
1504 mg->requeue_holder = true;
1505 mg->invalidate = false;
1506 mg->cache = cache;
1507 mg->old_oblock = oblock;
1508 mg->cblock = cblock;
1509 mg->old_ocell = cell;
1510 mg->new_ocell = NULL;
1511 mg->start_jiffies = jiffies;
1512
1513 inc_io_migrations(cache);
1514 quiesce_migration(mg);
1515}
1516
1517static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1518 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1519 dm_cblock_t cblock,
1520 struct dm_bio_prison_cell *old_ocell,
1521 struct dm_bio_prison_cell *new_ocell)
1522{
1523 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1524
1525 mg->err = false;
1526 mg->discard = false;
1527 mg->writeback = false;
1528 mg->demote = true;
1529 mg->promote = true;
1530 mg->requeue_holder = true;
1531 mg->invalidate = false;
1532 mg->cache = cache;
1533 mg->old_oblock = old_oblock;
1534 mg->new_oblock = new_oblock;
1535 mg->cblock = cblock;
1536 mg->old_ocell = old_ocell;
1537 mg->new_ocell = new_ocell;
1538 mg->start_jiffies = jiffies;
1539
1540 inc_io_migrations(cache);
1541 quiesce_migration(mg);
1542}
1543
1544/*
1545 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1546 * block are thrown away.
1547 */
1548static void invalidate(struct cache *cache, struct prealloc *structs,
1549 dm_oblock_t oblock, dm_cblock_t cblock,
1550 struct dm_bio_prison_cell *cell)
1551{
1552 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1553
1554 mg->err = false;
1555 mg->discard = false;
1556 mg->writeback = false;
1557 mg->demote = true;
1558 mg->promote = false;
1559 mg->requeue_holder = true;
1560 mg->invalidate = true;
1561 mg->cache = cache;
1562 mg->old_oblock = oblock;
1563 mg->cblock = cblock;
1564 mg->old_ocell = cell;
1565 mg->new_ocell = NULL;
1566 mg->start_jiffies = jiffies;
1567
1568 inc_io_migrations(cache);
1569 quiesce_migration(mg);
1570}
1571
1572static void discard(struct cache *cache, struct prealloc *structs,
1573 struct dm_bio_prison_cell *cell)
1574{
1575 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1576
1577 mg->err = false;
1578 mg->discard = true;
1579 mg->writeback = false;
1580 mg->demote = false;
1581 mg->promote = false;
1582 mg->requeue_holder = false;
1583 mg->invalidate = false;
1584 mg->cache = cache;
1585 mg->old_ocell = NULL;
1586 mg->new_ocell = cell;
1587 mg->start_jiffies = jiffies;
1588
1589 quiesce_migration(mg);
1590}
1591
1592/*----------------------------------------------------------------
1593 * bio processing
1594 *--------------------------------------------------------------*/
1595static void defer_bio(struct cache *cache, struct bio *bio)
1596{
1597 unsigned long flags;
1598
1599 spin_lock_irqsave(&cache->lock, flags);
1600 bio_list_add(&cache->deferred_bios, bio);
1601 spin_unlock_irqrestore(&cache->lock, flags);
1602
1603 wake_worker(cache);
1604}
1605
1606static void process_flush_bio(struct cache *cache, struct bio *bio)
1607{
1608 size_t pb_data_size = get_per_bio_data_size(cache);
1609 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1610
1611 BUG_ON(bio->bi_iter.bi_size);
1612 if (!pb->req_nr)
1613 remap_to_origin(cache, bio);
1614 else
1615 remap_to_cache(cache, bio, 0);
1616
1617 /*
1618 * REQ_PREFLUSH is not directed at any particular block so we don't
1619 * need to inc_ds(). REQ_FUA's are split into a write + REQ_PREFLUSH
1620 * by dm-core.
1621 */
1622 issue(cache, bio);
1623}
1624
1625static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1626 struct bio *bio)
1627{
1628 int r;
1629 dm_dblock_t b, e;
1630 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1631
1632 calc_discard_block_range(cache, bio, &b, &e);
1633 if (b == e) {
1634 bio_endio(bio);
1635 return;
1636 }
1637
1638 cell_prealloc = prealloc_get_cell(structs);
1639 r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1640 (cell_free_fn) prealloc_put_cell,
1641 structs, &new_ocell);
1642 if (r > 0)
1643 return;
1644
1645 discard(cache, structs, new_ocell);
1646}
1647
1648static bool spare_migration_bandwidth(struct cache *cache)
1649{
1650 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1651 cache->sectors_per_block;
1652 return current_volume < cache->migration_threshold;
1653}
1654
1655static void inc_hit_counter(struct cache *cache, struct bio *bio)
1656{
1657 atomic_inc(bio_data_dir(bio) == READ ?
1658 &cache->stats.read_hit : &cache->stats.write_hit);
1659}
1660
1661static void inc_miss_counter(struct cache *cache, struct bio *bio)
1662{
1663 atomic_inc(bio_data_dir(bio) == READ ?
1664 &cache->stats.read_miss : &cache->stats.write_miss);
1665}
1666
1667/*----------------------------------------------------------------*/
1668
1669struct inc_detail {
1670 struct cache *cache;
1671 struct bio_list bios_for_issue;
1672 struct bio_list unhandled_bios;
1673 bool any_writes;
1674};
1675
1676static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1677{
1678 struct bio *bio;
1679 struct inc_detail *detail = context;
1680 struct cache *cache = detail->cache;
1681
1682 inc_ds(cache, cell->holder, cell);
1683 if (bio_data_dir(cell->holder) == WRITE)
1684 detail->any_writes = true;
1685
1686 while ((bio = bio_list_pop(&cell->bios))) {
1687 if (discard_or_flush(bio)) {
1688 bio_list_add(&detail->unhandled_bios, bio);
1689 continue;
1690 }
1691
1692 if (bio_data_dir(bio) == WRITE)
1693 detail->any_writes = true;
1694
1695 bio_list_add(&detail->bios_for_issue, bio);
1696 inc_ds(cache, bio, cell);
1697 }
1698}
1699
1700// FIXME: refactor these two
1701static void remap_cell_to_origin_clear_discard(struct cache *cache,
1702 struct dm_bio_prison_cell *cell,
1703 dm_oblock_t oblock, bool issue_holder)
1704{
1705 struct bio *bio;
1706 unsigned long flags;
1707 struct inc_detail detail;
1708
1709 detail.cache = cache;
1710 bio_list_init(&detail.bios_for_issue);
1711 bio_list_init(&detail.unhandled_bios);
1712 detail.any_writes = false;
1713
1714 spin_lock_irqsave(&cache->lock, flags);
1715 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1716 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1717 spin_unlock_irqrestore(&cache->lock, flags);
1718
1719 remap_to_origin(cache, cell->holder);
1720 if (issue_holder)
1721 issue(cache, cell->holder);
1722 else
1723 accounted_begin(cache, cell->holder);
1724
1725 if (detail.any_writes)
1726 clear_discard(cache, oblock_to_dblock(cache, oblock));
1727
1728 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1729 remap_to_origin(cache, bio);
1730 issue(cache, bio);
1731 }
1732
1733 free_prison_cell(cache, cell);
1734}
1735
1736static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1737 dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1738{
1739 struct bio *bio;
1740 unsigned long flags;
1741 struct inc_detail detail;
1742
1743 detail.cache = cache;
1744 bio_list_init(&detail.bios_for_issue);
1745 bio_list_init(&detail.unhandled_bios);
1746 detail.any_writes = false;
1747
1748 spin_lock_irqsave(&cache->lock, flags);
1749 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1750 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1751 spin_unlock_irqrestore(&cache->lock, flags);
1752
1753 remap_to_cache(cache, cell->holder, cblock);
1754 if (issue_holder)
1755 issue(cache, cell->holder);
1756 else
1757 accounted_begin(cache, cell->holder);
1758
1759 if (detail.any_writes) {
1760 set_dirty(cache, oblock, cblock);
1761 clear_discard(cache, oblock_to_dblock(cache, oblock));
1762 }
1763
1764 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1765 remap_to_cache(cache, bio, cblock);
1766 issue(cache, bio);
1767 }
1768
1769 free_prison_cell(cache, cell);
1770}
1771
1772/*----------------------------------------------------------------*/
1773
1774struct old_oblock_lock {
1775 struct policy_locker locker;
1776 struct cache *cache;
1777 struct prealloc *structs;
1778 struct dm_bio_prison_cell *cell;
1779};
1780
1781static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1782{
1783 /* This should never be called */
1784 BUG();
1785 return 0;
1786}
1787
1788static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1789{
1790 struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1791 struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1792
1793 return bio_detain(l->cache, b, NULL, cell_prealloc,
1794 (cell_free_fn) prealloc_put_cell,
1795 l->structs, &l->cell);
1796}
1797
1798static void process_cell(struct cache *cache, struct prealloc *structs,
1799 struct dm_bio_prison_cell *new_ocell)
1800{
1801 int r;
1802 bool release_cell = true;
1803 struct bio *bio = new_ocell->holder;
1804 dm_oblock_t block = get_bio_block(cache, bio);
1805 struct policy_result lookup_result;
1806 bool passthrough = passthrough_mode(&cache->features);
1807 bool fast_promotion, can_migrate;
1808 struct old_oblock_lock ool;
1809
1810 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1811 can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1812
1813 ool.locker.fn = cell_locker;
1814 ool.cache = cache;
1815 ool.structs = structs;
1816 ool.cell = NULL;
1817 r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1818 bio, &ool.locker, &lookup_result);
1819
1820 if (r == -EWOULDBLOCK)
1821 /* migration has been denied */
1822 lookup_result.op = POLICY_MISS;
1823
1824 switch (lookup_result.op) {
1825 case POLICY_HIT:
1826 if (passthrough) {
1827 inc_miss_counter(cache, bio);
1828
1829 /*
1830 * Passthrough always maps to the origin,
1831 * invalidating any cache blocks that are written
1832 * to.
1833 */
1834
1835 if (bio_data_dir(bio) == WRITE) {
1836 atomic_inc(&cache->stats.demotion);
1837 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1838 release_cell = false;
1839
1840 } else {
1841 /* FIXME: factor out issue_origin() */
1842 remap_to_origin_clear_discard(cache, bio, block);
1843 inc_and_issue(cache, bio, new_ocell);
1844 }
1845 } else {
1846 inc_hit_counter(cache, bio);
1847
1848 if (bio_data_dir(bio) == WRITE &&
1849 writethrough_mode(&cache->features) &&
1850 !is_dirty(cache, lookup_result.cblock)) {
1851 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1852 inc_and_issue(cache, bio, new_ocell);
1853
1854 } else {
1855 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1856 release_cell = false;
1857 }
1858 }
1859
1860 break;
1861
1862 case POLICY_MISS:
1863 inc_miss_counter(cache, bio);
1864 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1865 release_cell = false;
1866 break;
1867
1868 case POLICY_NEW:
1869 atomic_inc(&cache->stats.promotion);
1870 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1871 release_cell = false;
1872 break;
1873
1874 case POLICY_REPLACE:
1875 atomic_inc(&cache->stats.demotion);
1876 atomic_inc(&cache->stats.promotion);
1877 demote_then_promote(cache, structs, lookup_result.old_oblock,
1878 block, lookup_result.cblock,
1879 ool.cell, new_ocell);
1880 release_cell = false;
1881 break;
1882
1883 default:
1884 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1885 cache_device_name(cache), __func__,
1886 (unsigned) lookup_result.op);
1887 bio_io_error(bio);
1888 }
1889
1890 if (release_cell)
1891 cell_defer(cache, new_ocell, false);
1892}
1893
1894static void process_bio(struct cache *cache, struct prealloc *structs,
1895 struct bio *bio)
1896{
1897 int r;
1898 dm_oblock_t block = get_bio_block(cache, bio);
1899 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1900
1901 /*
1902 * Check to see if that block is currently migrating.
1903 */
1904 cell_prealloc = prealloc_get_cell(structs);
1905 r = bio_detain(cache, block, bio, cell_prealloc,
1906 (cell_free_fn) prealloc_put_cell,
1907 structs, &new_ocell);
1908 if (r > 0)
1909 return;
1910
1911 process_cell(cache, structs, new_ocell);
1912}
1913
1914static int need_commit_due_to_time(struct cache *cache)
1915{
1916 return jiffies < cache->last_commit_jiffies ||
1917 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1918}
1919
1920/*
1921 * A non-zero return indicates read_only or fail_io mode.
1922 */
1923static int commit(struct cache *cache, bool clean_shutdown)
1924{
1925 int r;
1926
1927 if (get_cache_mode(cache) >= CM_READ_ONLY)
1928 return -EINVAL;
1929
1930 atomic_inc(&cache->stats.commit_count);
1931 r = dm_cache_commit(cache->cmd, clean_shutdown);
1932 if (r)
1933 metadata_operation_failed(cache, "dm_cache_commit", r);
1934
1935 return r;
1936}
1937
1938static int commit_if_needed(struct cache *cache)
1939{
1940 int r = 0;
1941
1942 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1943 dm_cache_changed_this_transaction(cache->cmd)) {
1944 r = commit(cache, false);
1945 cache->commit_requested = false;
1946 cache->last_commit_jiffies = jiffies;
1947 }
1948
1949 return r;
1950}
1951
1952static void process_deferred_bios(struct cache *cache)
1953{
1954 bool prealloc_used = false;
1955 unsigned long flags;
1956 struct bio_list bios;
1957 struct bio *bio;
1958 struct prealloc structs;
1959
1960 memset(&structs, 0, sizeof(structs));
1961 bio_list_init(&bios);
1962
1963 spin_lock_irqsave(&cache->lock, flags);
1964 bio_list_merge(&bios, &cache->deferred_bios);
1965 bio_list_init(&cache->deferred_bios);
1966 spin_unlock_irqrestore(&cache->lock, flags);
1967
1968 while (!bio_list_empty(&bios)) {
1969 /*
1970 * If we've got no free migration structs, and processing
1971 * this bio might require one, we pause until there are some
1972 * prepared mappings to process.
1973 */
1974 prealloc_used = true;
1975 if (prealloc_data_structs(cache, &structs)) {
1976 spin_lock_irqsave(&cache->lock, flags);
1977 bio_list_merge(&cache->deferred_bios, &bios);
1978 spin_unlock_irqrestore(&cache->lock, flags);
1979 break;
1980 }
1981
1982 bio = bio_list_pop(&bios);
1983
1984 if (bio->bi_opf & REQ_PREFLUSH)
1985 process_flush_bio(cache, bio);
1986 else if (bio_op(bio) == REQ_OP_DISCARD)
1987 process_discard_bio(cache, &structs, bio);
1988 else
1989 process_bio(cache, &structs, bio);
1990 }
1991
1992 if (prealloc_used)
1993 prealloc_free_structs(cache, &structs);
1994}
1995
1996static void process_deferred_cells(struct cache *cache)
1997{
1998 bool prealloc_used = false;
1999 unsigned long flags;
2000 struct dm_bio_prison_cell *cell, *tmp;
2001 struct list_head cells;
2002 struct prealloc structs;
2003
2004 memset(&structs, 0, sizeof(structs));
2005
2006 INIT_LIST_HEAD(&cells);
2007
2008 spin_lock_irqsave(&cache->lock, flags);
2009 list_splice_init(&cache->deferred_cells, &cells);
2010 spin_unlock_irqrestore(&cache->lock, flags);
2011
2012 list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2013 /*
2014 * If we've got no free migration structs, and processing
2015 * this bio might require one, we pause until there are some
2016 * prepared mappings to process.
2017 */
2018 prealloc_used = true;
2019 if (prealloc_data_structs(cache, &structs)) {
2020 spin_lock_irqsave(&cache->lock, flags);
2021 list_splice(&cells, &cache->deferred_cells);
2022 spin_unlock_irqrestore(&cache->lock, flags);
2023 break;
2024 }
2025
2026 process_cell(cache, &structs, cell);
2027 }
2028
2029 if (prealloc_used)
2030 prealloc_free_structs(cache, &structs);
2031}
2032
2033static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2034{
2035 unsigned long flags;
2036 struct bio_list bios;
2037 struct bio *bio;
2038
2039 bio_list_init(&bios);
2040
2041 spin_lock_irqsave(&cache->lock, flags);
2042 bio_list_merge(&bios, &cache->deferred_flush_bios);
2043 bio_list_init(&cache->deferred_flush_bios);
2044 spin_unlock_irqrestore(&cache->lock, flags);
2045
2046 /*
2047 * These bios have already been through inc_ds()
2048 */
2049 while ((bio = bio_list_pop(&bios)))
2050 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2051}
2052
2053static void process_deferred_writethrough_bios(struct cache *cache)
2054{
2055 unsigned long flags;
2056 struct bio_list bios;
2057 struct bio *bio;
2058
2059 bio_list_init(&bios);
2060
2061 spin_lock_irqsave(&cache->lock, flags);
2062 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2063 bio_list_init(&cache->deferred_writethrough_bios);
2064 spin_unlock_irqrestore(&cache->lock, flags);
2065
2066 /*
2067 * These bios have already been through inc_ds()
2068 */
2069 while ((bio = bio_list_pop(&bios)))
2070 accounted_request(cache, bio);
2071}
2072
2073static void writeback_some_dirty_blocks(struct cache *cache)
2074{
2075 bool prealloc_used = false;
2076 dm_oblock_t oblock;
2077 dm_cblock_t cblock;
2078 struct prealloc structs;
2079 struct dm_bio_prison_cell *old_ocell;
2080 bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2081
2082 memset(&structs, 0, sizeof(structs));
2083
2084 while (spare_migration_bandwidth(cache)) {
2085 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2086 break; /* no work to do */
2087
2088 prealloc_used = true;
2089 if (prealloc_data_structs(cache, &structs) ||
2090 get_cell(cache, oblock, &structs, &old_ocell)) {
2091 policy_set_dirty(cache->policy, oblock);
2092 break;
2093 }
2094
2095 writeback(cache, &structs, oblock, cblock, old_ocell);
2096 }
2097
2098 if (prealloc_used)
2099 prealloc_free_structs(cache, &structs);
2100}
2101
2102/*----------------------------------------------------------------
2103 * Invalidations.
2104 * Dropping something from the cache *without* writing back.
2105 *--------------------------------------------------------------*/
2106
2107static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2108{
2109 int r = 0;
2110 uint64_t begin = from_cblock(req->cblocks->begin);
2111 uint64_t end = from_cblock(req->cblocks->end);
2112
2113 while (begin != end) {
2114 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2115 if (!r) {
2116 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2117 if (r) {
2118 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2119 break;
2120 }
2121
2122 } else if (r == -ENODATA) {
2123 /* harmless, already unmapped */
2124 r = 0;
2125
2126 } else {
2127 DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2128 break;
2129 }
2130
2131 begin++;
2132 }
2133
2134 cache->commit_requested = true;
2135
2136 req->err = r;
2137 atomic_set(&req->complete, 1);
2138
2139 wake_up(&req->result_wait);
2140}
2141
2142static void process_invalidation_requests(struct cache *cache)
2143{
2144 struct list_head list;
2145 struct invalidation_request *req, *tmp;
2146
2147 INIT_LIST_HEAD(&list);
2148 spin_lock(&cache->invalidation_lock);
2149 list_splice_init(&cache->invalidation_requests, &list);
2150 spin_unlock(&cache->invalidation_lock);
2151
2152 list_for_each_entry_safe (req, tmp, &list, list)
2153 process_invalidation_request(cache, req);
2154}
2155
2156/*----------------------------------------------------------------
2157 * Main worker loop
2158 *--------------------------------------------------------------*/
2159static bool is_quiescing(struct cache *cache)
2160{
2161 return atomic_read(&cache->quiescing);
2162}
2163
2164static void ack_quiescing(struct cache *cache)
2165{
2166 if (is_quiescing(cache)) {
2167 atomic_inc(&cache->quiescing_ack);
2168 wake_up(&cache->quiescing_wait);
2169 }
2170}
2171
2172static void wait_for_quiescing_ack(struct cache *cache)
2173{
2174 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2175}
2176
2177static void start_quiescing(struct cache *cache)
2178{
2179 atomic_inc(&cache->quiescing);
2180 wait_for_quiescing_ack(cache);
2181}
2182
2183static void stop_quiescing(struct cache *cache)
2184{
2185 atomic_set(&cache->quiescing, 0);
2186 atomic_set(&cache->quiescing_ack, 0);
2187}
2188
2189static void wait_for_migrations(struct cache *cache)
2190{
2191 wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2192}
2193
2194static void stop_worker(struct cache *cache)
2195{
2196 cancel_delayed_work(&cache->waker);
2197 flush_workqueue(cache->wq);
2198}
2199
2200static void requeue_deferred_cells(struct cache *cache)
2201{
2202 unsigned long flags;
2203 struct list_head cells;
2204 struct dm_bio_prison_cell *cell, *tmp;
2205
2206 INIT_LIST_HEAD(&cells);
2207 spin_lock_irqsave(&cache->lock, flags);
2208 list_splice_init(&cache->deferred_cells, &cells);
2209 spin_unlock_irqrestore(&cache->lock, flags);
2210
2211 list_for_each_entry_safe(cell, tmp, &cells, user_list)
2212 cell_requeue(cache, cell);
2213}
2214
2215static void requeue_deferred_bios(struct cache *cache)
2216{
2217 struct bio *bio;
2218 struct bio_list bios;
2219
2220 bio_list_init(&bios);
2221 bio_list_merge(&bios, &cache->deferred_bios);
2222 bio_list_init(&cache->deferred_bios);
2223
2224 while ((bio = bio_list_pop(&bios))) {
2225 bio->bi_error = DM_ENDIO_REQUEUE;
2226 bio_endio(bio);
2227 }
2228}
2229
2230static int more_work(struct cache *cache)
2231{
2232 if (is_quiescing(cache))
2233 return !list_empty(&cache->quiesced_migrations) ||
2234 !list_empty(&cache->completed_migrations) ||
2235 !list_empty(&cache->need_commit_migrations);
2236 else
2237 return !bio_list_empty(&cache->deferred_bios) ||
2238 !list_empty(&cache->deferred_cells) ||
2239 !bio_list_empty(&cache->deferred_flush_bios) ||
2240 !bio_list_empty(&cache->deferred_writethrough_bios) ||
2241 !list_empty(&cache->quiesced_migrations) ||
2242 !list_empty(&cache->completed_migrations) ||
2243 !list_empty(&cache->need_commit_migrations) ||
2244 cache->invalidate;
2245}
2246
2247static void do_worker(struct work_struct *ws)
2248{
2249 struct cache *cache = container_of(ws, struct cache, worker);
2250
2251 do {
2252 if (!is_quiescing(cache)) {
2253 writeback_some_dirty_blocks(cache);
2254 process_deferred_writethrough_bios(cache);
2255 process_deferred_bios(cache);
2256 process_deferred_cells(cache);
2257 process_invalidation_requests(cache);
2258 }
2259
2260 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2261 process_migrations(cache, &cache->completed_migrations, complete_migration);
2262
2263 if (commit_if_needed(cache)) {
2264 process_deferred_flush_bios(cache, false);
2265 process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2266 } else {
2267 process_deferred_flush_bios(cache, true);
2268 process_migrations(cache, &cache->need_commit_migrations,
2269 migration_success_post_commit);
2270 }
2271
2272 ack_quiescing(cache);
2273
2274 } while (more_work(cache));
2275}
2276
2277/*
2278 * We want to commit periodically so that not too much
2279 * unwritten metadata builds up.
2280 */
2281static void do_waker(struct work_struct *ws)
2282{
2283 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2284 policy_tick(cache->policy, true);
2285 wake_worker(cache);
2286 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2287}
2288
2289/*----------------------------------------------------------------*/
2290
2291static int is_congested(struct dm_dev *dev, int bdi_bits)
2292{
2293 struct request_queue *q = bdev_get_queue(dev->bdev);
2294 return bdi_congested(&q->backing_dev_info, bdi_bits);
2295}
2296
2297static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2298{
2299 struct cache *cache = container_of(cb, struct cache, callbacks);
2300
2301 return is_congested(cache->origin_dev, bdi_bits) ||
2302 is_congested(cache->cache_dev, bdi_bits);
2303}
2304
2305/*----------------------------------------------------------------
2306 * Target methods
2307 *--------------------------------------------------------------*/
2308
2309/*
2310 * This function gets called on the error paths of the constructor, so we
2311 * have to cope with a partially initialised struct.
2312 */
2313static void destroy(struct cache *cache)
2314{
2315 unsigned i;
2316
2317 mempool_destroy(cache->migration_pool);
2318
2319 if (cache->all_io_ds)
2320 dm_deferred_set_destroy(cache->all_io_ds);
2321
2322 if (cache->prison)
2323 dm_bio_prison_destroy(cache->prison);
2324
2325 if (cache->wq)
2326 destroy_workqueue(cache->wq);
2327
2328 if (cache->dirty_bitset)
2329 free_bitset(cache->dirty_bitset);
2330
2331 if (cache->discard_bitset)
2332 free_bitset(cache->discard_bitset);
2333
2334 if (cache->copier)
2335 dm_kcopyd_client_destroy(cache->copier);
2336
2337 if (cache->cmd)
2338 dm_cache_metadata_close(cache->cmd);
2339
2340 if (cache->metadata_dev)
2341 dm_put_device(cache->ti, cache->metadata_dev);
2342
2343 if (cache->origin_dev)
2344 dm_put_device(cache->ti, cache->origin_dev);
2345
2346 if (cache->cache_dev)
2347 dm_put_device(cache->ti, cache->cache_dev);
2348
2349 if (cache->policy)
2350 dm_cache_policy_destroy(cache->policy);
2351
2352 for (i = 0; i < cache->nr_ctr_args ; i++)
2353 kfree(cache->ctr_args[i]);
2354 kfree(cache->ctr_args);
2355
2356 kfree(cache);
2357}
2358
2359static void cache_dtr(struct dm_target *ti)
2360{
2361 struct cache *cache = ti->private;
2362
2363 destroy(cache);
2364}
2365
2366static sector_t get_dev_size(struct dm_dev *dev)
2367{
2368 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2369}
2370
2371/*----------------------------------------------------------------*/
2372
2373/*
2374 * Construct a cache device mapping.
2375 *
2376 * cache <metadata dev> <cache dev> <origin dev> <block size>
2377 * <#feature args> [<feature arg>]*
2378 * <policy> <#policy args> [<policy arg>]*
2379 *
2380 * metadata dev : fast device holding the persistent metadata
2381 * cache dev : fast device holding cached data blocks
2382 * origin dev : slow device holding original data blocks
2383 * block size : cache unit size in sectors
2384 *
2385 * #feature args : number of feature arguments passed
2386 * feature args : writethrough. (The default is writeback.)
2387 *
2388 * policy : the replacement policy to use
2389 * #policy args : an even number of policy arguments corresponding
2390 * to key/value pairs passed to the policy
2391 * policy args : key/value pairs passed to the policy
2392 * E.g. 'sequential_threshold 1024'
2393 * See cache-policies.txt for details.
2394 *
2395 * Optional feature arguments are:
2396 * writethrough : write through caching that prohibits cache block
2397 * content from being different from origin block content.
2398 * Without this argument, the default behaviour is to write
2399 * back cache block contents later for performance reasons,
2400 * so they may differ from the corresponding origin blocks.
2401 */
2402struct cache_args {
2403 struct dm_target *ti;
2404
2405 struct dm_dev *metadata_dev;
2406
2407 struct dm_dev *cache_dev;
2408 sector_t cache_sectors;
2409
2410 struct dm_dev *origin_dev;
2411 sector_t origin_sectors;
2412
2413 uint32_t block_size;
2414
2415 const char *policy_name;
2416 int policy_argc;
2417 const char **policy_argv;
2418
2419 struct cache_features features;
2420};
2421
2422static void destroy_cache_args(struct cache_args *ca)
2423{
2424 if (ca->metadata_dev)
2425 dm_put_device(ca->ti, ca->metadata_dev);
2426
2427 if (ca->cache_dev)
2428 dm_put_device(ca->ti, ca->cache_dev);
2429
2430 if (ca->origin_dev)
2431 dm_put_device(ca->ti, ca->origin_dev);
2432
2433 kfree(ca);
2434}
2435
2436static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2437{
2438 if (!as->argc) {
2439 *error = "Insufficient args";
2440 return false;
2441 }
2442
2443 return true;
2444}
2445
2446static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2447 char **error)
2448{
2449 int r;
2450 sector_t metadata_dev_size;
2451 char b[BDEVNAME_SIZE];
2452
2453 if (!at_least_one_arg(as, error))
2454 return -EINVAL;
2455
2456 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2457 &ca->metadata_dev);
2458 if (r) {
2459 *error = "Error opening metadata device";
2460 return r;
2461 }
2462
2463 metadata_dev_size = get_dev_size(ca->metadata_dev);
2464 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2465 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2466 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2467
2468 return 0;
2469}
2470
2471static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2472 char **error)
2473{
2474 int r;
2475
2476 if (!at_least_one_arg(as, error))
2477 return -EINVAL;
2478
2479 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2480 &ca->cache_dev);
2481 if (r) {
2482 *error = "Error opening cache device";
2483 return r;
2484 }
2485 ca->cache_sectors = get_dev_size(ca->cache_dev);
2486
2487 return 0;
2488}
2489
2490static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2491 char **error)
2492{
2493 int r;
2494
2495 if (!at_least_one_arg(as, error))
2496 return -EINVAL;
2497
2498 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2499 &ca->origin_dev);
2500 if (r) {
2501 *error = "Error opening origin device";
2502 return r;
2503 }
2504
2505 ca->origin_sectors = get_dev_size(ca->origin_dev);
2506 if (ca->ti->len > ca->origin_sectors) {
2507 *error = "Device size larger than cached device";
2508 return -EINVAL;
2509 }
2510
2511 return 0;
2512}
2513
2514static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2515 char **error)
2516{
2517 unsigned long block_size;
2518
2519 if (!at_least_one_arg(as, error))
2520 return -EINVAL;
2521
2522 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2523 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2524 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2525 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2526 *error = "Invalid data block size";
2527 return -EINVAL;
2528 }
2529
2530 if (block_size > ca->cache_sectors) {
2531 *error = "Data block size is larger than the cache device";
2532 return -EINVAL;
2533 }
2534
2535 ca->block_size = block_size;
2536
2537 return 0;
2538}
2539
2540static void init_features(struct cache_features *cf)
2541{
2542 cf->mode = CM_WRITE;
2543 cf->io_mode = CM_IO_WRITEBACK;
2544}
2545
2546static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2547 char **error)
2548{
2549 static struct dm_arg _args[] = {
2550 {0, 1, "Invalid number of cache feature arguments"},
2551 };
2552
2553 int r;
2554 unsigned argc;
2555 const char *arg;
2556 struct cache_features *cf = &ca->features;
2557
2558 init_features(cf);
2559
2560 r = dm_read_arg_group(_args, as, &argc, error);
2561 if (r)
2562 return -EINVAL;
2563
2564 while (argc--) {
2565 arg = dm_shift_arg(as);
2566
2567 if (!strcasecmp(arg, "writeback"))
2568 cf->io_mode = CM_IO_WRITEBACK;
2569
2570 else if (!strcasecmp(arg, "writethrough"))
2571 cf->io_mode = CM_IO_WRITETHROUGH;
2572
2573 else if (!strcasecmp(arg, "passthrough"))
2574 cf->io_mode = CM_IO_PASSTHROUGH;
2575
2576 else {
2577 *error = "Unrecognised cache feature requested";
2578 return -EINVAL;
2579 }
2580 }
2581
2582 return 0;
2583}
2584
2585static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2586 char **error)
2587{
2588 static struct dm_arg _args[] = {
2589 {0, 1024, "Invalid number of policy arguments"},
2590 };
2591
2592 int r;
2593
2594 if (!at_least_one_arg(as, error))
2595 return -EINVAL;
2596
2597 ca->policy_name = dm_shift_arg(as);
2598
2599 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2600 if (r)
2601 return -EINVAL;
2602
2603 ca->policy_argv = (const char **)as->argv;
2604 dm_consume_args(as, ca->policy_argc);
2605
2606 return 0;
2607}
2608
2609static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2610 char **error)
2611{
2612 int r;
2613 struct dm_arg_set as;
2614
2615 as.argc = argc;
2616 as.argv = argv;
2617
2618 r = parse_metadata_dev(ca, &as, error);
2619 if (r)
2620 return r;
2621
2622 r = parse_cache_dev(ca, &as, error);
2623 if (r)
2624 return r;
2625
2626 r = parse_origin_dev(ca, &as, error);
2627 if (r)
2628 return r;
2629
2630 r = parse_block_size(ca, &as, error);
2631 if (r)
2632 return r;
2633
2634 r = parse_features(ca, &as, error);
2635 if (r)
2636 return r;
2637
2638 r = parse_policy(ca, &as, error);
2639 if (r)
2640 return r;
2641
2642 return 0;
2643}
2644
2645/*----------------------------------------------------------------*/
2646
2647static struct kmem_cache *migration_cache;
2648
2649#define NOT_CORE_OPTION 1
2650
2651static int process_config_option(struct cache *cache, const char *key, const char *value)
2652{
2653 unsigned long tmp;
2654
2655 if (!strcasecmp(key, "migration_threshold")) {
2656 if (kstrtoul(value, 10, &tmp))
2657 return -EINVAL;
2658
2659 cache->migration_threshold = tmp;
2660 return 0;
2661 }
2662
2663 return NOT_CORE_OPTION;
2664}
2665
2666static int set_config_value(struct cache *cache, const char *key, const char *value)
2667{
2668 int r = process_config_option(cache, key, value);
2669
2670 if (r == NOT_CORE_OPTION)
2671 r = policy_set_config_value(cache->policy, key, value);
2672
2673 if (r)
2674 DMWARN("bad config value for %s: %s", key, value);
2675
2676 return r;
2677}
2678
2679static int set_config_values(struct cache *cache, int argc, const char **argv)
2680{
2681 int r = 0;
2682
2683 if (argc & 1) {
2684 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2685 return -EINVAL;
2686 }
2687
2688 while (argc) {
2689 r = set_config_value(cache, argv[0], argv[1]);
2690 if (r)
2691 break;
2692
2693 argc -= 2;
2694 argv += 2;
2695 }
2696
2697 return r;
2698}
2699
2700static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2701 char **error)
2702{
2703 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2704 cache->cache_size,
2705 cache->origin_sectors,
2706 cache->sectors_per_block);
2707 if (IS_ERR(p)) {
2708 *error = "Error creating cache's policy";
2709 return PTR_ERR(p);
2710 }
2711 cache->policy = p;
2712
2713 return 0;
2714}
2715
2716/*
2717 * We want the discard block size to be at least the size of the cache
2718 * block size and have no more than 2^14 discard blocks across the origin.
2719 */
2720#define MAX_DISCARD_BLOCKS (1 << 14)
2721
2722static bool too_many_discard_blocks(sector_t discard_block_size,
2723 sector_t origin_size)
2724{
2725 (void) sector_div(origin_size, discard_block_size);
2726
2727 return origin_size > MAX_DISCARD_BLOCKS;
2728}
2729
2730static sector_t calculate_discard_block_size(sector_t cache_block_size,
2731 sector_t origin_size)
2732{
2733 sector_t discard_block_size = cache_block_size;
2734
2735 if (origin_size)
2736 while (too_many_discard_blocks(discard_block_size, origin_size))
2737 discard_block_size *= 2;
2738
2739 return discard_block_size;
2740}
2741
2742static void set_cache_size(struct cache *cache, dm_cblock_t size)
2743{
2744 dm_block_t nr_blocks = from_cblock(size);
2745
2746 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2747 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2748 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2749 "Please consider increasing the cache block size to reduce the overall cache block count.",
2750 (unsigned long long) nr_blocks);
2751
2752 cache->cache_size = size;
2753}
2754
2755#define DEFAULT_MIGRATION_THRESHOLD 2048
2756
2757static int cache_create(struct cache_args *ca, struct cache **result)
2758{
2759 int r = 0;
2760 char **error = &ca->ti->error;
2761 struct cache *cache;
2762 struct dm_target *ti = ca->ti;
2763 dm_block_t origin_blocks;
2764 struct dm_cache_metadata *cmd;
2765 bool may_format = ca->features.mode == CM_WRITE;
2766
2767 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2768 if (!cache)
2769 return -ENOMEM;
2770
2771 cache->ti = ca->ti;
2772 ti->private = cache;
2773 ti->num_flush_bios = 2;
2774 ti->flush_supported = true;
2775
2776 ti->num_discard_bios = 1;
2777 ti->discards_supported = true;
2778 ti->discard_zeroes_data_unsupported = true;
2779 ti->split_discard_bios = false;
2780
2781 cache->features = ca->features;
2782 ti->per_io_data_size = get_per_bio_data_size(cache);
2783
2784 cache->callbacks.congested_fn = cache_is_congested;
2785 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2786
2787 cache->metadata_dev = ca->metadata_dev;
2788 cache->origin_dev = ca->origin_dev;
2789 cache->cache_dev = ca->cache_dev;
2790
2791 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2792
2793 /* FIXME: factor out this whole section */
2794 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2795 origin_blocks = block_div(origin_blocks, ca->block_size);
2796 cache->origin_blocks = to_oblock(origin_blocks);
2797
2798 cache->sectors_per_block = ca->block_size;
2799 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2800 r = -EINVAL;
2801 goto bad;
2802 }
2803
2804 if (ca->block_size & (ca->block_size - 1)) {
2805 dm_block_t cache_size = ca->cache_sectors;
2806
2807 cache->sectors_per_block_shift = -1;
2808 cache_size = block_div(cache_size, ca->block_size);
2809 set_cache_size(cache, to_cblock(cache_size));
2810 } else {
2811 cache->sectors_per_block_shift = __ffs(ca->block_size);
2812 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2813 }
2814
2815 r = create_cache_policy(cache, ca, error);
2816 if (r)
2817 goto bad;
2818
2819 cache->policy_nr_args = ca->policy_argc;
2820 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2821
2822 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2823 if (r) {
2824 *error = "Error setting cache policy's config values";
2825 goto bad;
2826 }
2827
2828 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2829 ca->block_size, may_format,
2830 dm_cache_policy_get_hint_size(cache->policy));
2831 if (IS_ERR(cmd)) {
2832 *error = "Error creating metadata object";
2833 r = PTR_ERR(cmd);
2834 goto bad;
2835 }
2836 cache->cmd = cmd;
2837 set_cache_mode(cache, CM_WRITE);
2838 if (get_cache_mode(cache) != CM_WRITE) {
2839 *error = "Unable to get write access to metadata, please check/repair metadata.";
2840 r = -EINVAL;
2841 goto bad;
2842 }
2843
2844 if (passthrough_mode(&cache->features)) {
2845 bool all_clean;
2846
2847 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2848 if (r) {
2849 *error = "dm_cache_metadata_all_clean() failed";
2850 goto bad;
2851 }
2852
2853 if (!all_clean) {
2854 *error = "Cannot enter passthrough mode unless all blocks are clean";
2855 r = -EINVAL;
2856 goto bad;
2857 }
2858 }
2859
2860 spin_lock_init(&cache->lock);
2861 INIT_LIST_HEAD(&cache->deferred_cells);
2862 bio_list_init(&cache->deferred_bios);
2863 bio_list_init(&cache->deferred_flush_bios);
2864 bio_list_init(&cache->deferred_writethrough_bios);
2865 INIT_LIST_HEAD(&cache->quiesced_migrations);
2866 INIT_LIST_HEAD(&cache->completed_migrations);
2867 INIT_LIST_HEAD(&cache->need_commit_migrations);
2868 atomic_set(&cache->nr_allocated_migrations, 0);
2869 atomic_set(&cache->nr_io_migrations, 0);
2870 init_waitqueue_head(&cache->migration_wait);
2871
2872 init_waitqueue_head(&cache->quiescing_wait);
2873 atomic_set(&cache->quiescing, 0);
2874 atomic_set(&cache->quiescing_ack, 0);
2875
2876 r = -ENOMEM;
2877 atomic_set(&cache->nr_dirty, 0);
2878 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2879 if (!cache->dirty_bitset) {
2880 *error = "could not allocate dirty bitset";
2881 goto bad;
2882 }
2883 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2884
2885 cache->discard_block_size =
2886 calculate_discard_block_size(cache->sectors_per_block,
2887 cache->origin_sectors);
2888 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2889 cache->discard_block_size));
2890 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2891 if (!cache->discard_bitset) {
2892 *error = "could not allocate discard bitset";
2893 goto bad;
2894 }
2895 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2896
2897 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2898 if (IS_ERR(cache->copier)) {
2899 *error = "could not create kcopyd client";
2900 r = PTR_ERR(cache->copier);
2901 goto bad;
2902 }
2903
2904 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2905 if (!cache->wq) {
2906 *error = "could not create workqueue for metadata object";
2907 goto bad;
2908 }
2909 INIT_WORK(&cache->worker, do_worker);
2910 INIT_DELAYED_WORK(&cache->waker, do_waker);
2911 cache->last_commit_jiffies = jiffies;
2912
2913 cache->prison = dm_bio_prison_create();
2914 if (!cache->prison) {
2915 *error = "could not create bio prison";
2916 goto bad;
2917 }
2918
2919 cache->all_io_ds = dm_deferred_set_create();
2920 if (!cache->all_io_ds) {
2921 *error = "could not create all_io deferred set";
2922 goto bad;
2923 }
2924
2925 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2926 migration_cache);
2927 if (!cache->migration_pool) {
2928 *error = "Error creating cache's migration mempool";
2929 goto bad;
2930 }
2931
2932 cache->need_tick_bio = true;
2933 cache->sized = false;
2934 cache->invalidate = false;
2935 cache->commit_requested = false;
2936 cache->loaded_mappings = false;
2937 cache->loaded_discards = false;
2938
2939 load_stats(cache);
2940
2941 atomic_set(&cache->stats.demotion, 0);
2942 atomic_set(&cache->stats.promotion, 0);
2943 atomic_set(&cache->stats.copies_avoided, 0);
2944 atomic_set(&cache->stats.cache_cell_clash, 0);
2945 atomic_set(&cache->stats.commit_count, 0);
2946 atomic_set(&cache->stats.discard_count, 0);
2947
2948 spin_lock_init(&cache->invalidation_lock);
2949 INIT_LIST_HEAD(&cache->invalidation_requests);
2950
2951 iot_init(&cache->origin_tracker);
2952
2953 *result = cache;
2954 return 0;
2955
2956bad:
2957 destroy(cache);
2958 return r;
2959}
2960
2961static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2962{
2963 unsigned i;
2964 const char **copy;
2965
2966 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2967 if (!copy)
2968 return -ENOMEM;
2969 for (i = 0; i < argc; i++) {
2970 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2971 if (!copy[i]) {
2972 while (i--)
2973 kfree(copy[i]);
2974 kfree(copy);
2975 return -ENOMEM;
2976 }
2977 }
2978
2979 cache->nr_ctr_args = argc;
2980 cache->ctr_args = copy;
2981
2982 return 0;
2983}
2984
2985static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2986{
2987 int r = -EINVAL;
2988 struct cache_args *ca;
2989 struct cache *cache = NULL;
2990
2991 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2992 if (!ca) {
2993 ti->error = "Error allocating memory for cache";
2994 return -ENOMEM;
2995 }
2996 ca->ti = ti;
2997
2998 r = parse_cache_args(ca, argc, argv, &ti->error);
2999 if (r)
3000 goto out;
3001
3002 r = cache_create(ca, &cache);
3003 if (r)
3004 goto out;
3005
3006 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
3007 if (r) {
3008 destroy(cache);
3009 goto out;
3010 }
3011
3012 ti->private = cache;
3013
3014out:
3015 destroy_cache_args(ca);
3016 return r;
3017}
3018
3019/*----------------------------------------------------------------*/
3020
3021static int cache_map(struct dm_target *ti, struct bio *bio)
3022{
3023 struct cache *cache = ti->private;
3024
3025 int r;
3026 struct dm_bio_prison_cell *cell = NULL;
3027 dm_oblock_t block = get_bio_block(cache, bio);
3028 size_t pb_data_size = get_per_bio_data_size(cache);
3029 bool can_migrate = false;
3030 bool fast_promotion;
3031 struct policy_result lookup_result;
3032 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3033 struct old_oblock_lock ool;
3034
3035 ool.locker.fn = null_locker;
3036
3037 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3038 /*
3039 * This can only occur if the io goes to a partial block at
3040 * the end of the origin device. We don't cache these.
3041 * Just remap to the origin and carry on.
3042 */
3043 remap_to_origin(cache, bio);
3044 accounted_begin(cache, bio);
3045 return DM_MAPIO_REMAPPED;
3046 }
3047
3048 if (discard_or_flush(bio)) {
3049 defer_bio(cache, bio);
3050 return DM_MAPIO_SUBMITTED;
3051 }
3052
3053 /*
3054 * Check to see if that block is currently migrating.
3055 */
3056 cell = alloc_prison_cell(cache);
3057 if (!cell) {
3058 defer_bio(cache, bio);
3059 return DM_MAPIO_SUBMITTED;
3060 }
3061
3062 r = bio_detain(cache, block, bio, cell,
3063 (cell_free_fn) free_prison_cell,
3064 cache, &cell);
3065 if (r) {
3066 if (r < 0)
3067 defer_bio(cache, bio);
3068
3069 return DM_MAPIO_SUBMITTED;
3070 }
3071
3072 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3073
3074 r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3075 bio, &ool.locker, &lookup_result);
3076 if (r == -EWOULDBLOCK) {
3077 cell_defer(cache, cell, true);
3078 return DM_MAPIO_SUBMITTED;
3079
3080 } else if (r) {
3081 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3082 cache_device_name(cache), r);
3083 cell_defer(cache, cell, false);
3084 bio_io_error(bio);
3085 return DM_MAPIO_SUBMITTED;
3086 }
3087
3088 r = DM_MAPIO_REMAPPED;
3089 switch (lookup_result.op) {
3090 case POLICY_HIT:
3091 if (passthrough_mode(&cache->features)) {
3092 if (bio_data_dir(bio) == WRITE) {
3093 /*
3094 * We need to invalidate this block, so
3095 * defer for the worker thread.
3096 */
3097 cell_defer(cache, cell, true);
3098 r = DM_MAPIO_SUBMITTED;
3099
3100 } else {
3101 inc_miss_counter(cache, bio);
3102 remap_to_origin_clear_discard(cache, bio, block);
3103 accounted_begin(cache, bio);
3104 inc_ds(cache, bio, cell);
3105 // FIXME: we want to remap hits or misses straight
3106 // away rather than passing over to the worker.
3107 cell_defer(cache, cell, false);
3108 }
3109
3110 } else {
3111 inc_hit_counter(cache, bio);
3112 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3113 !is_dirty(cache, lookup_result.cblock)) {
3114 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3115 accounted_begin(cache, bio);
3116 inc_ds(cache, bio, cell);
3117 cell_defer(cache, cell, false);
3118
3119 } else
3120 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3121 }
3122 break;
3123
3124 case POLICY_MISS:
3125 inc_miss_counter(cache, bio);
3126 if (pb->req_nr != 0) {
3127 /*
3128 * This is a duplicate writethrough io that is no
3129 * longer needed because the block has been demoted.
3130 */
3131 bio_endio(bio);
3132 // FIXME: remap everything as a miss
3133 cell_defer(cache, cell, false);
3134 r = DM_MAPIO_SUBMITTED;
3135
3136 } else
3137 remap_cell_to_origin_clear_discard(cache, cell, block, false);
3138 break;
3139
3140 default:
3141 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3142 cache_device_name(cache), __func__,
3143 (unsigned) lookup_result.op);
3144 cell_defer(cache, cell, false);
3145 bio_io_error(bio);
3146 r = DM_MAPIO_SUBMITTED;
3147 }
3148
3149 return r;
3150}
3151
3152static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3153{
3154 struct cache *cache = ti->private;
3155 unsigned long flags;
3156 size_t pb_data_size = get_per_bio_data_size(cache);
3157 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3158
3159 if (pb->tick) {
3160 policy_tick(cache->policy, false);
3161
3162 spin_lock_irqsave(&cache->lock, flags);
3163 cache->need_tick_bio = true;
3164 spin_unlock_irqrestore(&cache->lock, flags);
3165 }
3166
3167 check_for_quiesced_migrations(cache, pb);
3168 accounted_complete(cache, bio);
3169
3170 return 0;
3171}
3172
3173static int write_dirty_bitset(struct cache *cache)
3174{
3175 unsigned i, r;
3176
3177 if (get_cache_mode(cache) >= CM_READ_ONLY)
3178 return -EINVAL;
3179
3180 for (i = 0; i < from_cblock(cache->cache_size); i++) {
3181 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3182 is_dirty(cache, to_cblock(i)));
3183 if (r) {
3184 metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3185 return r;
3186 }
3187 }
3188
3189 return 0;
3190}
3191
3192static int write_discard_bitset(struct cache *cache)
3193{
3194 unsigned i, r;
3195
3196 if (get_cache_mode(cache) >= CM_READ_ONLY)
3197 return -EINVAL;
3198
3199 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3200 cache->discard_nr_blocks);
3201 if (r) {
3202 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3203 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3204 return r;
3205 }
3206
3207 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3208 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3209 is_discarded(cache, to_dblock(i)));
3210 if (r) {
3211 metadata_operation_failed(cache, "dm_cache_set_discard", r);
3212 return r;
3213 }
3214 }
3215
3216 return 0;
3217}
3218
3219static int write_hints(struct cache *cache)
3220{
3221 int r;
3222
3223 if (get_cache_mode(cache) >= CM_READ_ONLY)
3224 return -EINVAL;
3225
3226 r = dm_cache_write_hints(cache->cmd, cache->policy);
3227 if (r) {
3228 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3229 return r;
3230 }
3231
3232 return 0;
3233}
3234
3235/*
3236 * returns true on success
3237 */
3238static bool sync_metadata(struct cache *cache)
3239{
3240 int r1, r2, r3, r4;
3241
3242 r1 = write_dirty_bitset(cache);
3243 if (r1)
3244 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3245
3246 r2 = write_discard_bitset(cache);
3247 if (r2)
3248 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3249
3250 save_stats(cache);
3251
3252 r3 = write_hints(cache);
3253 if (r3)
3254 DMERR("%s: could not write hints", cache_device_name(cache));
3255
3256 /*
3257 * If writing the above metadata failed, we still commit, but don't
3258 * set the clean shutdown flag. This will effectively force every
3259 * dirty bit to be set on reload.
3260 */
3261 r4 = commit(cache, !r1 && !r2 && !r3);
3262 if (r4)
3263 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3264
3265 return !r1 && !r2 && !r3 && !r4;
3266}
3267
3268static void cache_postsuspend(struct dm_target *ti)
3269{
3270 struct cache *cache = ti->private;
3271
3272 start_quiescing(cache);
3273 wait_for_migrations(cache);
3274 stop_worker(cache);
3275 requeue_deferred_bios(cache);
3276 requeue_deferred_cells(cache);
3277 stop_quiescing(cache);
3278
3279 if (get_cache_mode(cache) == CM_WRITE)
3280 (void) sync_metadata(cache);
3281}
3282
3283static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3284 bool dirty, uint32_t hint, bool hint_valid)
3285{
3286 int r;
3287 struct cache *cache = context;
3288
3289 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3290 if (r)
3291 return r;
3292
3293 if (dirty)
3294 set_dirty(cache, oblock, cblock);
3295 else
3296 clear_dirty(cache, oblock, cblock);
3297
3298 return 0;
3299}
3300
3301/*
3302 * The discard block size in the on disk metadata is not
3303 * neccessarily the same as we're currently using. So we have to
3304 * be careful to only set the discarded attribute if we know it
3305 * covers a complete block of the new size.
3306 */
3307struct discard_load_info {
3308 struct cache *cache;
3309
3310 /*
3311 * These blocks are sized using the on disk dblock size, rather
3312 * than the current one.
3313 */
3314 dm_block_t block_size;
3315 dm_block_t discard_begin, discard_end;
3316};
3317
3318static void discard_load_info_init(struct cache *cache,
3319 struct discard_load_info *li)
3320{
3321 li->cache = cache;
3322 li->discard_begin = li->discard_end = 0;
3323}
3324
3325static void set_discard_range(struct discard_load_info *li)
3326{
3327 sector_t b, e;
3328
3329 if (li->discard_begin == li->discard_end)
3330 return;
3331
3332 /*
3333 * Convert to sectors.
3334 */
3335 b = li->discard_begin * li->block_size;
3336 e = li->discard_end * li->block_size;
3337
3338 /*
3339 * Then convert back to the current dblock size.
3340 */
3341 b = dm_sector_div_up(b, li->cache->discard_block_size);
3342 sector_div(e, li->cache->discard_block_size);
3343
3344 /*
3345 * The origin may have shrunk, so we need to check we're still in
3346 * bounds.
3347 */
3348 if (e > from_dblock(li->cache->discard_nr_blocks))
3349 e = from_dblock(li->cache->discard_nr_blocks);
3350
3351 for (; b < e; b++)
3352 set_discard(li->cache, to_dblock(b));
3353}
3354
3355static int load_discard(void *context, sector_t discard_block_size,
3356 dm_dblock_t dblock, bool discard)
3357{
3358 struct discard_load_info *li = context;
3359
3360 li->block_size = discard_block_size;
3361
3362 if (discard) {
3363 if (from_dblock(dblock) == li->discard_end)
3364 /*
3365 * We're already in a discard range, just extend it.
3366 */
3367 li->discard_end = li->discard_end + 1ULL;
3368
3369 else {
3370 /*
3371 * Emit the old range and start a new one.
3372 */
3373 set_discard_range(li);
3374 li->discard_begin = from_dblock(dblock);
3375 li->discard_end = li->discard_begin + 1ULL;
3376 }
3377 } else {
3378 set_discard_range(li);
3379 li->discard_begin = li->discard_end = 0;
3380 }
3381
3382 return 0;
3383}
3384
3385static dm_cblock_t get_cache_dev_size(struct cache *cache)
3386{
3387 sector_t size = get_dev_size(cache->cache_dev);
3388 (void) sector_div(size, cache->sectors_per_block);
3389 return to_cblock(size);
3390}
3391
3392static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3393{
3394 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3395 return true;
3396
3397 /*
3398 * We can't drop a dirty block when shrinking the cache.
3399 */
3400 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3401 new_size = to_cblock(from_cblock(new_size) + 1);
3402 if (is_dirty(cache, new_size)) {
3403 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3404 cache_device_name(cache),
3405 (unsigned long long) from_cblock(new_size));
3406 return false;
3407 }
3408 }
3409
3410 return true;
3411}
3412
3413static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3414{
3415 int r;
3416
3417 r = dm_cache_resize(cache->cmd, new_size);
3418 if (r) {
3419 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3420 metadata_operation_failed(cache, "dm_cache_resize", r);
3421 return r;
3422 }
3423
3424 set_cache_size(cache, new_size);
3425
3426 return 0;
3427}
3428
3429static int cache_preresume(struct dm_target *ti)
3430{
3431 int r = 0;
3432 struct cache *cache = ti->private;
3433 dm_cblock_t csize = get_cache_dev_size(cache);
3434
3435 /*
3436 * Check to see if the cache has resized.
3437 */
3438 if (!cache->sized) {
3439 r = resize_cache_dev(cache, csize);
3440 if (r)
3441 return r;
3442
3443 cache->sized = true;
3444
3445 } else if (csize != cache->cache_size) {
3446 if (!can_resize(cache, csize))
3447 return -EINVAL;
3448
3449 r = resize_cache_dev(cache, csize);
3450 if (r)
3451 return r;
3452 }
3453
3454 if (!cache->loaded_mappings) {
3455 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3456 load_mapping, cache);
3457 if (r) {
3458 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3459 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3460 return r;
3461 }
3462
3463 cache->loaded_mappings = true;
3464 }
3465
3466 if (!cache->loaded_discards) {
3467 struct discard_load_info li;
3468
3469 /*
3470 * The discard bitset could have been resized, or the
3471 * discard block size changed. To be safe we start by
3472 * setting every dblock to not discarded.
3473 */
3474 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3475
3476 discard_load_info_init(cache, &li);
3477 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3478 if (r) {
3479 DMERR("%s: could not load origin discards", cache_device_name(cache));
3480 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3481 return r;
3482 }
3483 set_discard_range(&li);
3484
3485 cache->loaded_discards = true;
3486 }
3487
3488 return r;
3489}
3490
3491static void cache_resume(struct dm_target *ti)
3492{
3493 struct cache *cache = ti->private;
3494
3495 cache->need_tick_bio = true;
3496 do_waker(&cache->waker.work);
3497}
3498
3499/*
3500 * Status format:
3501 *
3502 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3503 * <cache block size> <#used cache blocks>/<#total cache blocks>
3504 * <#read hits> <#read misses> <#write hits> <#write misses>
3505 * <#demotions> <#promotions> <#dirty>
3506 * <#features> <features>*
3507 * <#core args> <core args>
3508 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3509 */
3510static void cache_status(struct dm_target *ti, status_type_t type,
3511 unsigned status_flags, char *result, unsigned maxlen)
3512{
3513 int r = 0;
3514 unsigned i;
3515 ssize_t sz = 0;
3516 dm_block_t nr_free_blocks_metadata = 0;
3517 dm_block_t nr_blocks_metadata = 0;
3518 char buf[BDEVNAME_SIZE];
3519 struct cache *cache = ti->private;
3520 dm_cblock_t residency;
3521 bool needs_check;
3522
3523 switch (type) {
3524 case STATUSTYPE_INFO:
3525 if (get_cache_mode(cache) == CM_FAIL) {
3526 DMEMIT("Fail");
3527 break;
3528 }
3529
3530 /* Commit to ensure statistics aren't out-of-date */
3531 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3532 (void) commit(cache, false);
3533
3534 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3535 if (r) {
3536 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3537 cache_device_name(cache), r);
3538 goto err;
3539 }
3540
3541 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3542 if (r) {
3543 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3544 cache_device_name(cache), r);
3545 goto err;
3546 }
3547
3548 residency = policy_residency(cache->policy);
3549
3550 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3551 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3552 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3553 (unsigned long long)nr_blocks_metadata,
3554 (unsigned long long)cache->sectors_per_block,
3555 (unsigned long long) from_cblock(residency),
3556 (unsigned long long) from_cblock(cache->cache_size),
3557 (unsigned) atomic_read(&cache->stats.read_hit),
3558 (unsigned) atomic_read(&cache->stats.read_miss),
3559 (unsigned) atomic_read(&cache->stats.write_hit),
3560 (unsigned) atomic_read(&cache->stats.write_miss),
3561 (unsigned) atomic_read(&cache->stats.demotion),
3562 (unsigned) atomic_read(&cache->stats.promotion),
3563 (unsigned long) atomic_read(&cache->nr_dirty));
3564
3565 if (writethrough_mode(&cache->features))
3566 DMEMIT("1 writethrough ");
3567
3568 else if (passthrough_mode(&cache->features))
3569 DMEMIT("1 passthrough ");
3570
3571 else if (writeback_mode(&cache->features))
3572 DMEMIT("1 writeback ");
3573
3574 else {
3575 DMERR("%s: internal error: unknown io mode: %d",
3576 cache_device_name(cache), (int) cache->features.io_mode);
3577 goto err;
3578 }
3579
3580 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3581
3582 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3583 if (sz < maxlen) {
3584 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3585 if (r)
3586 DMERR("%s: policy_emit_config_values returned %d",
3587 cache_device_name(cache), r);
3588 }
3589
3590 if (get_cache_mode(cache) == CM_READ_ONLY)
3591 DMEMIT("ro ");
3592 else
3593 DMEMIT("rw ");
3594
3595 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3596
3597 if (r || needs_check)
3598 DMEMIT("needs_check ");
3599 else
3600 DMEMIT("- ");
3601
3602 break;
3603
3604 case STATUSTYPE_TABLE:
3605 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3606 DMEMIT("%s ", buf);
3607 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3608 DMEMIT("%s ", buf);
3609 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3610 DMEMIT("%s", buf);
3611
3612 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3613 DMEMIT(" %s", cache->ctr_args[i]);
3614 if (cache->nr_ctr_args)
3615 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3616 }
3617
3618 return;
3619
3620err:
3621 DMEMIT("Error");
3622}
3623
3624/*
3625 * A cache block range can take two forms:
3626 *
3627 * i) A single cblock, eg. '3456'
3628 * ii) A begin and end cblock with dots between, eg. 123-234
3629 */
3630static int parse_cblock_range(struct cache *cache, const char *str,
3631 struct cblock_range *result)
3632{
3633 char dummy;
3634 uint64_t b, e;
3635 int r;
3636
3637 /*
3638 * Try and parse form (ii) first.
3639 */
3640 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3641 if (r < 0)
3642 return r;
3643
3644 if (r == 2) {
3645 result->begin = to_cblock(b);
3646 result->end = to_cblock(e);
3647 return 0;
3648 }
3649
3650 /*
3651 * That didn't work, try form (i).
3652 */
3653 r = sscanf(str, "%llu%c", &b, &dummy);
3654 if (r < 0)
3655 return r;
3656
3657 if (r == 1) {
3658 result->begin = to_cblock(b);
3659 result->end = to_cblock(from_cblock(result->begin) + 1u);
3660 return 0;
3661 }
3662
3663 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3664 return -EINVAL;
3665}
3666
3667static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3668{
3669 uint64_t b = from_cblock(range->begin);
3670 uint64_t e = from_cblock(range->end);
3671 uint64_t n = from_cblock(cache->cache_size);
3672
3673 if (b >= n) {
3674 DMERR("%s: begin cblock out of range: %llu >= %llu",
3675 cache_device_name(cache), b, n);
3676 return -EINVAL;
3677 }
3678
3679 if (e > n) {
3680 DMERR("%s: end cblock out of range: %llu > %llu",
3681 cache_device_name(cache), e, n);
3682 return -EINVAL;
3683 }
3684
3685 if (b >= e) {
3686 DMERR("%s: invalid cblock range: %llu >= %llu",
3687 cache_device_name(cache), b, e);
3688 return -EINVAL;
3689 }
3690
3691 return 0;
3692}
3693
3694static int request_invalidation(struct cache *cache, struct cblock_range *range)
3695{
3696 struct invalidation_request req;
3697
3698 INIT_LIST_HEAD(&req.list);
3699 req.cblocks = range;
3700 atomic_set(&req.complete, 0);
3701 req.err = 0;
3702 init_waitqueue_head(&req.result_wait);
3703
3704 spin_lock(&cache->invalidation_lock);
3705 list_add(&req.list, &cache->invalidation_requests);
3706 spin_unlock(&cache->invalidation_lock);
3707 wake_worker(cache);
3708
3709 wait_event(req.result_wait, atomic_read(&req.complete));
3710 return req.err;
3711}
3712
3713static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3714 const char **cblock_ranges)
3715{
3716 int r = 0;
3717 unsigned i;
3718 struct cblock_range range;
3719
3720 if (!passthrough_mode(&cache->features)) {
3721 DMERR("%s: cache has to be in passthrough mode for invalidation",
3722 cache_device_name(cache));
3723 return -EPERM;
3724 }
3725
3726 for (i = 0; i < count; i++) {
3727 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3728 if (r)
3729 break;
3730
3731 r = validate_cblock_range(cache, &range);
3732 if (r)
3733 break;
3734
3735 /*
3736 * Pass begin and end origin blocks to the worker and wake it.
3737 */
3738 r = request_invalidation(cache, &range);
3739 if (r)
3740 break;
3741 }
3742
3743 return r;
3744}
3745
3746/*
3747 * Supports
3748 * "<key> <value>"
3749 * and
3750 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3751 *
3752 * The key migration_threshold is supported by the cache target core.
3753 */
3754static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3755{
3756 struct cache *cache = ti->private;
3757
3758 if (!argc)
3759 return -EINVAL;
3760
3761 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3762 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3763 cache_device_name(cache));
3764 return -EOPNOTSUPP;
3765 }
3766
3767 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3768 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3769
3770 if (argc != 2)
3771 return -EINVAL;
3772
3773 return set_config_value(cache, argv[0], argv[1]);
3774}
3775
3776static int cache_iterate_devices(struct dm_target *ti,
3777 iterate_devices_callout_fn fn, void *data)
3778{
3779 int r = 0;
3780 struct cache *cache = ti->private;
3781
3782 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3783 if (!r)
3784 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3785
3786 return r;
3787}
3788
3789static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3790{
3791 /*
3792 * FIXME: these limits may be incompatible with the cache device
3793 */
3794 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3795 cache->origin_sectors);
3796 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3797}
3798
3799static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3800{
3801 struct cache *cache = ti->private;
3802 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3803
3804 /*
3805 * If the system-determined stacked limits are compatible with the
3806 * cache's blocksize (io_opt is a factor) do not override them.
3807 */
3808 if (io_opt_sectors < cache->sectors_per_block ||
3809 do_div(io_opt_sectors, cache->sectors_per_block)) {
3810 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3811 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3812 }
3813 set_discard_limits(cache, limits);
3814}
3815
3816/*----------------------------------------------------------------*/
3817
3818static struct target_type cache_target = {
3819 .name = "cache",
3820 .version = {1, 9, 0},
3821 .module = THIS_MODULE,
3822 .ctr = cache_ctr,
3823 .dtr = cache_dtr,
3824 .map = cache_map,
3825 .end_io = cache_end_io,
3826 .postsuspend = cache_postsuspend,
3827 .preresume = cache_preresume,
3828 .resume = cache_resume,
3829 .status = cache_status,
3830 .message = cache_message,
3831 .iterate_devices = cache_iterate_devices,
3832 .io_hints = cache_io_hints,
3833};
3834
3835static int __init dm_cache_init(void)
3836{
3837 int r;
3838
3839 r = dm_register_target(&cache_target);
3840 if (r) {
3841 DMERR("cache target registration failed: %d", r);
3842 return r;
3843 }
3844
3845 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3846 if (!migration_cache) {
3847 dm_unregister_target(&cache_target);
3848 return -ENOMEM;
3849 }
3850
3851 return 0;
3852}
3853
3854static void __exit dm_cache_exit(void)
3855{
3856 dm_unregister_target(&cache_target);
3857 kmem_cache_destroy(migration_cache);
3858}
3859
3860module_init(dm_cache_init);
3861module_exit(dm_cache_exit);
3862
3863MODULE_DESCRIPTION(DM_NAME " cache target");
3864MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3865MODULE_LICENSE("GPL");
1/*
2 * Copyright (C) 2012 Red Hat. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
8#include "dm-bio-prison-v2.h"
9#include "dm-bio-record.h"
10#include "dm-cache-metadata.h"
11
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
14#include <linux/jiffies.h>
15#include <linux/init.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
18#include <linux/rwsem.h>
19#include <linux/slab.h>
20#include <linux/vmalloc.h>
21
22#define DM_MSG_PREFIX "cache"
23
24DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
25 "A percentage of time allocated for copying to and/or from cache");
26
27/*----------------------------------------------------------------*/
28
29/*
30 * Glossary:
31 *
32 * oblock: index of an origin block
33 * cblock: index of a cache block
34 * promotion: movement of a block from origin to cache
35 * demotion: movement of a block from cache to origin
36 * migration: movement of a block between the origin and cache device,
37 * either direction
38 */
39
40/*----------------------------------------------------------------*/
41
42struct io_tracker {
43 spinlock_t lock;
44
45 /*
46 * Sectors of in-flight IO.
47 */
48 sector_t in_flight;
49
50 /*
51 * The time, in jiffies, when this device became idle (if it is
52 * indeed idle).
53 */
54 unsigned long idle_time;
55 unsigned long last_update_time;
56};
57
58static void iot_init(struct io_tracker *iot)
59{
60 spin_lock_init(&iot->lock);
61 iot->in_flight = 0ul;
62 iot->idle_time = 0ul;
63 iot->last_update_time = jiffies;
64}
65
66static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
67{
68 if (iot->in_flight)
69 return false;
70
71 return time_after(jiffies, iot->idle_time + jifs);
72}
73
74static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
75{
76 bool r;
77 unsigned long flags;
78
79 spin_lock_irqsave(&iot->lock, flags);
80 r = __iot_idle_for(iot, jifs);
81 spin_unlock_irqrestore(&iot->lock, flags);
82
83 return r;
84}
85
86static void iot_io_begin(struct io_tracker *iot, sector_t len)
87{
88 unsigned long flags;
89
90 spin_lock_irqsave(&iot->lock, flags);
91 iot->in_flight += len;
92 spin_unlock_irqrestore(&iot->lock, flags);
93}
94
95static void __iot_io_end(struct io_tracker *iot, sector_t len)
96{
97 if (!len)
98 return;
99
100 iot->in_flight -= len;
101 if (!iot->in_flight)
102 iot->idle_time = jiffies;
103}
104
105static void iot_io_end(struct io_tracker *iot, sector_t len)
106{
107 unsigned long flags;
108
109 spin_lock_irqsave(&iot->lock, flags);
110 __iot_io_end(iot, len);
111 spin_unlock_irqrestore(&iot->lock, flags);
112}
113
114/*----------------------------------------------------------------*/
115
116/*
117 * Represents a chunk of future work. 'input' allows continuations to pass
118 * values between themselves, typically error values.
119 */
120struct continuation {
121 struct work_struct ws;
122 blk_status_t input;
123};
124
125static inline void init_continuation(struct continuation *k,
126 void (*fn)(struct work_struct *))
127{
128 INIT_WORK(&k->ws, fn);
129 k->input = 0;
130}
131
132static inline void queue_continuation(struct workqueue_struct *wq,
133 struct continuation *k)
134{
135 queue_work(wq, &k->ws);
136}
137
138/*----------------------------------------------------------------*/
139
140/*
141 * The batcher collects together pieces of work that need a particular
142 * operation to occur before they can proceed (typically a commit).
143 */
144struct batcher {
145 /*
146 * The operation that everyone is waiting for.
147 */
148 blk_status_t (*commit_op)(void *context);
149 void *commit_context;
150
151 /*
152 * This is how bios should be issued once the commit op is complete
153 * (accounted_request).
154 */
155 void (*issue_op)(struct bio *bio, void *context);
156 void *issue_context;
157
158 /*
159 * Queued work gets put on here after commit.
160 */
161 struct workqueue_struct *wq;
162
163 spinlock_t lock;
164 struct list_head work_items;
165 struct bio_list bios;
166 struct work_struct commit_work;
167
168 bool commit_scheduled;
169};
170
171static void __commit(struct work_struct *_ws)
172{
173 struct batcher *b = container_of(_ws, struct batcher, commit_work);
174 blk_status_t r;
175 unsigned long flags;
176 struct list_head work_items;
177 struct work_struct *ws, *tmp;
178 struct continuation *k;
179 struct bio *bio;
180 struct bio_list bios;
181
182 INIT_LIST_HEAD(&work_items);
183 bio_list_init(&bios);
184
185 /*
186 * We have to grab these before the commit_op to avoid a race
187 * condition.
188 */
189 spin_lock_irqsave(&b->lock, flags);
190 list_splice_init(&b->work_items, &work_items);
191 bio_list_merge(&bios, &b->bios);
192 bio_list_init(&b->bios);
193 b->commit_scheduled = false;
194 spin_unlock_irqrestore(&b->lock, flags);
195
196 r = b->commit_op(b->commit_context);
197
198 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
199 k = container_of(ws, struct continuation, ws);
200 k->input = r;
201 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
202 queue_work(b->wq, ws);
203 }
204
205 while ((bio = bio_list_pop(&bios))) {
206 if (r) {
207 bio->bi_status = r;
208 bio_endio(bio);
209 } else
210 b->issue_op(bio, b->issue_context);
211 }
212}
213
214static void batcher_init(struct batcher *b,
215 blk_status_t (*commit_op)(void *),
216 void *commit_context,
217 void (*issue_op)(struct bio *bio, void *),
218 void *issue_context,
219 struct workqueue_struct *wq)
220{
221 b->commit_op = commit_op;
222 b->commit_context = commit_context;
223 b->issue_op = issue_op;
224 b->issue_context = issue_context;
225 b->wq = wq;
226
227 spin_lock_init(&b->lock);
228 INIT_LIST_HEAD(&b->work_items);
229 bio_list_init(&b->bios);
230 INIT_WORK(&b->commit_work, __commit);
231 b->commit_scheduled = false;
232}
233
234static void async_commit(struct batcher *b)
235{
236 queue_work(b->wq, &b->commit_work);
237}
238
239static void continue_after_commit(struct batcher *b, struct continuation *k)
240{
241 unsigned long flags;
242 bool commit_scheduled;
243
244 spin_lock_irqsave(&b->lock, flags);
245 commit_scheduled = b->commit_scheduled;
246 list_add_tail(&k->ws.entry, &b->work_items);
247 spin_unlock_irqrestore(&b->lock, flags);
248
249 if (commit_scheduled)
250 async_commit(b);
251}
252
253/*
254 * Bios are errored if commit failed.
255 */
256static void issue_after_commit(struct batcher *b, struct bio *bio)
257{
258 unsigned long flags;
259 bool commit_scheduled;
260
261 spin_lock_irqsave(&b->lock, flags);
262 commit_scheduled = b->commit_scheduled;
263 bio_list_add(&b->bios, bio);
264 spin_unlock_irqrestore(&b->lock, flags);
265
266 if (commit_scheduled)
267 async_commit(b);
268}
269
270/*
271 * Call this if some urgent work is waiting for the commit to complete.
272 */
273static void schedule_commit(struct batcher *b)
274{
275 bool immediate;
276 unsigned long flags;
277
278 spin_lock_irqsave(&b->lock, flags);
279 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
280 b->commit_scheduled = true;
281 spin_unlock_irqrestore(&b->lock, flags);
282
283 if (immediate)
284 async_commit(b);
285}
286
287/*
288 * There are a couple of places where we let a bio run, but want to do some
289 * work before calling its endio function. We do this by temporarily
290 * changing the endio fn.
291 */
292struct dm_hook_info {
293 bio_end_io_t *bi_end_io;
294};
295
296static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
297 bio_end_io_t *bi_end_io, void *bi_private)
298{
299 h->bi_end_io = bio->bi_end_io;
300
301 bio->bi_end_io = bi_end_io;
302 bio->bi_private = bi_private;
303}
304
305static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
306{
307 bio->bi_end_io = h->bi_end_io;
308}
309
310/*----------------------------------------------------------------*/
311
312#define MIGRATION_POOL_SIZE 128
313#define COMMIT_PERIOD HZ
314#define MIGRATION_COUNT_WINDOW 10
315
316/*
317 * The block size of the device holding cache data must be
318 * between 32KB and 1GB.
319 */
320#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
321#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
322
323enum cache_metadata_mode {
324 CM_WRITE, /* metadata may be changed */
325 CM_READ_ONLY, /* metadata may not be changed */
326 CM_FAIL
327};
328
329enum cache_io_mode {
330 /*
331 * Data is written to cached blocks only. These blocks are marked
332 * dirty. If you lose the cache device you will lose data.
333 * Potential performance increase for both reads and writes.
334 */
335 CM_IO_WRITEBACK,
336
337 /*
338 * Data is written to both cache and origin. Blocks are never
339 * dirty. Potential performance benfit for reads only.
340 */
341 CM_IO_WRITETHROUGH,
342
343 /*
344 * A degraded mode useful for various cache coherency situations
345 * (eg, rolling back snapshots). Reads and writes always go to the
346 * origin. If a write goes to a cached oblock, then the cache
347 * block is invalidated.
348 */
349 CM_IO_PASSTHROUGH
350};
351
352struct cache_features {
353 enum cache_metadata_mode mode;
354 enum cache_io_mode io_mode;
355 unsigned metadata_version;
356 bool discard_passdown:1;
357};
358
359struct cache_stats {
360 atomic_t read_hit;
361 atomic_t read_miss;
362 atomic_t write_hit;
363 atomic_t write_miss;
364 atomic_t demotion;
365 atomic_t promotion;
366 atomic_t writeback;
367 atomic_t copies_avoided;
368 atomic_t cache_cell_clash;
369 atomic_t commit_count;
370 atomic_t discard_count;
371};
372
373struct cache {
374 struct dm_target *ti;
375 spinlock_t lock;
376
377 /*
378 * Fields for converting from sectors to blocks.
379 */
380 int sectors_per_block_shift;
381 sector_t sectors_per_block;
382
383 struct dm_cache_metadata *cmd;
384
385 /*
386 * Metadata is written to this device.
387 */
388 struct dm_dev *metadata_dev;
389
390 /*
391 * The slower of the two data devices. Typically a spindle.
392 */
393 struct dm_dev *origin_dev;
394
395 /*
396 * The faster of the two data devices. Typically an SSD.
397 */
398 struct dm_dev *cache_dev;
399
400 /*
401 * Size of the origin device in _complete_ blocks and native sectors.
402 */
403 dm_oblock_t origin_blocks;
404 sector_t origin_sectors;
405
406 /*
407 * Size of the cache device in blocks.
408 */
409 dm_cblock_t cache_size;
410
411 /*
412 * Invalidation fields.
413 */
414 spinlock_t invalidation_lock;
415 struct list_head invalidation_requests;
416
417 sector_t migration_threshold;
418 wait_queue_head_t migration_wait;
419 atomic_t nr_allocated_migrations;
420
421 /*
422 * The number of in flight migrations that are performing
423 * background io. eg, promotion, writeback.
424 */
425 atomic_t nr_io_migrations;
426
427 struct bio_list deferred_bios;
428
429 struct rw_semaphore quiesce_lock;
430
431 struct dm_target_callbacks callbacks;
432
433 /*
434 * origin_blocks entries, discarded if set.
435 */
436 dm_dblock_t discard_nr_blocks;
437 unsigned long *discard_bitset;
438 uint32_t discard_block_size; /* a power of 2 times sectors per block */
439
440 /*
441 * Rather than reconstructing the table line for the status we just
442 * save it and regurgitate.
443 */
444 unsigned nr_ctr_args;
445 const char **ctr_args;
446
447 struct dm_kcopyd_client *copier;
448 struct work_struct deferred_bio_worker;
449 struct work_struct migration_worker;
450 struct workqueue_struct *wq;
451 struct delayed_work waker;
452 struct dm_bio_prison_v2 *prison;
453
454 /*
455 * cache_size entries, dirty if set
456 */
457 unsigned long *dirty_bitset;
458 atomic_t nr_dirty;
459
460 unsigned policy_nr_args;
461 struct dm_cache_policy *policy;
462
463 /*
464 * Cache features such as write-through.
465 */
466 struct cache_features features;
467
468 struct cache_stats stats;
469
470 bool need_tick_bio:1;
471 bool sized:1;
472 bool invalidate:1;
473 bool commit_requested:1;
474 bool loaded_mappings:1;
475 bool loaded_discards:1;
476
477 struct rw_semaphore background_work_lock;
478
479 struct batcher committer;
480 struct work_struct commit_ws;
481
482 struct io_tracker tracker;
483
484 mempool_t migration_pool;
485
486 struct bio_set bs;
487};
488
489struct per_bio_data {
490 bool tick:1;
491 unsigned req_nr:2;
492 struct dm_bio_prison_cell_v2 *cell;
493 struct dm_hook_info hook_info;
494 sector_t len;
495};
496
497struct dm_cache_migration {
498 struct continuation k;
499 struct cache *cache;
500
501 struct policy_work *op;
502 struct bio *overwrite_bio;
503 struct dm_bio_prison_cell_v2 *cell;
504
505 dm_cblock_t invalidate_cblock;
506 dm_oblock_t invalidate_oblock;
507};
508
509/*----------------------------------------------------------------*/
510
511static bool writethrough_mode(struct cache *cache)
512{
513 return cache->features.io_mode == CM_IO_WRITETHROUGH;
514}
515
516static bool writeback_mode(struct cache *cache)
517{
518 return cache->features.io_mode == CM_IO_WRITEBACK;
519}
520
521static inline bool passthrough_mode(struct cache *cache)
522{
523 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
524}
525
526/*----------------------------------------------------------------*/
527
528static void wake_deferred_bio_worker(struct cache *cache)
529{
530 queue_work(cache->wq, &cache->deferred_bio_worker);
531}
532
533static void wake_migration_worker(struct cache *cache)
534{
535 if (passthrough_mode(cache))
536 return;
537
538 queue_work(cache->wq, &cache->migration_worker);
539}
540
541/*----------------------------------------------------------------*/
542
543static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
544{
545 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
546}
547
548static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
549{
550 dm_bio_prison_free_cell_v2(cache->prison, cell);
551}
552
553static struct dm_cache_migration *alloc_migration(struct cache *cache)
554{
555 struct dm_cache_migration *mg;
556
557 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
558
559 memset(mg, 0, sizeof(*mg));
560
561 mg->cache = cache;
562 atomic_inc(&cache->nr_allocated_migrations);
563
564 return mg;
565}
566
567static void free_migration(struct dm_cache_migration *mg)
568{
569 struct cache *cache = mg->cache;
570
571 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
572 wake_up(&cache->migration_wait);
573
574 mempool_free(mg, &cache->migration_pool);
575}
576
577/*----------------------------------------------------------------*/
578
579static inline dm_oblock_t oblock_succ(dm_oblock_t b)
580{
581 return to_oblock(from_oblock(b) + 1ull);
582}
583
584static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
585{
586 key->virtual = 0;
587 key->dev = 0;
588 key->block_begin = from_oblock(begin);
589 key->block_end = from_oblock(end);
590}
591
592/*
593 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
594 * level 1 which prevents *both* READs and WRITEs.
595 */
596#define WRITE_LOCK_LEVEL 0
597#define READ_WRITE_LOCK_LEVEL 1
598
599static unsigned lock_level(struct bio *bio)
600{
601 return bio_data_dir(bio) == WRITE ?
602 WRITE_LOCK_LEVEL :
603 READ_WRITE_LOCK_LEVEL;
604}
605
606/*----------------------------------------------------------------
607 * Per bio data
608 *--------------------------------------------------------------*/
609
610static struct per_bio_data *get_per_bio_data(struct bio *bio)
611{
612 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
613 BUG_ON(!pb);
614 return pb;
615}
616
617static struct per_bio_data *init_per_bio_data(struct bio *bio)
618{
619 struct per_bio_data *pb = get_per_bio_data(bio);
620
621 pb->tick = false;
622 pb->req_nr = dm_bio_get_target_bio_nr(bio);
623 pb->cell = NULL;
624 pb->len = 0;
625
626 return pb;
627}
628
629/*----------------------------------------------------------------*/
630
631static void defer_bio(struct cache *cache, struct bio *bio)
632{
633 unsigned long flags;
634
635 spin_lock_irqsave(&cache->lock, flags);
636 bio_list_add(&cache->deferred_bios, bio);
637 spin_unlock_irqrestore(&cache->lock, flags);
638
639 wake_deferred_bio_worker(cache);
640}
641
642static void defer_bios(struct cache *cache, struct bio_list *bios)
643{
644 unsigned long flags;
645
646 spin_lock_irqsave(&cache->lock, flags);
647 bio_list_merge(&cache->deferred_bios, bios);
648 bio_list_init(bios);
649 spin_unlock_irqrestore(&cache->lock, flags);
650
651 wake_deferred_bio_worker(cache);
652}
653
654/*----------------------------------------------------------------*/
655
656static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
657{
658 bool r;
659 struct per_bio_data *pb;
660 struct dm_cell_key_v2 key;
661 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
662 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
663
664 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
665
666 build_key(oblock, end, &key);
667 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
668 if (!r) {
669 /*
670 * Failed to get the lock.
671 */
672 free_prison_cell(cache, cell_prealloc);
673 return r;
674 }
675
676 if (cell != cell_prealloc)
677 free_prison_cell(cache, cell_prealloc);
678
679 pb = get_per_bio_data(bio);
680 pb->cell = cell;
681
682 return r;
683}
684
685/*----------------------------------------------------------------*/
686
687static bool is_dirty(struct cache *cache, dm_cblock_t b)
688{
689 return test_bit(from_cblock(b), cache->dirty_bitset);
690}
691
692static void set_dirty(struct cache *cache, dm_cblock_t cblock)
693{
694 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
695 atomic_inc(&cache->nr_dirty);
696 policy_set_dirty(cache->policy, cblock);
697 }
698}
699
700/*
701 * These two are called when setting after migrations to force the policy
702 * and dirty bitset to be in sync.
703 */
704static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
705{
706 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
707 atomic_inc(&cache->nr_dirty);
708 policy_set_dirty(cache->policy, cblock);
709}
710
711static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
712{
713 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
714 if (atomic_dec_return(&cache->nr_dirty) == 0)
715 dm_table_event(cache->ti->table);
716 }
717
718 policy_clear_dirty(cache->policy, cblock);
719}
720
721/*----------------------------------------------------------------*/
722
723static bool block_size_is_power_of_two(struct cache *cache)
724{
725 return cache->sectors_per_block_shift >= 0;
726}
727
728/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
729#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
730__always_inline
731#endif
732static dm_block_t block_div(dm_block_t b, uint32_t n)
733{
734 do_div(b, n);
735
736 return b;
737}
738
739static dm_block_t oblocks_per_dblock(struct cache *cache)
740{
741 dm_block_t oblocks = cache->discard_block_size;
742
743 if (block_size_is_power_of_two(cache))
744 oblocks >>= cache->sectors_per_block_shift;
745 else
746 oblocks = block_div(oblocks, cache->sectors_per_block);
747
748 return oblocks;
749}
750
751static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
752{
753 return to_dblock(block_div(from_oblock(oblock),
754 oblocks_per_dblock(cache)));
755}
756
757static void set_discard(struct cache *cache, dm_dblock_t b)
758{
759 unsigned long flags;
760
761 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
762 atomic_inc(&cache->stats.discard_count);
763
764 spin_lock_irqsave(&cache->lock, flags);
765 set_bit(from_dblock(b), cache->discard_bitset);
766 spin_unlock_irqrestore(&cache->lock, flags);
767}
768
769static void clear_discard(struct cache *cache, dm_dblock_t b)
770{
771 unsigned long flags;
772
773 spin_lock_irqsave(&cache->lock, flags);
774 clear_bit(from_dblock(b), cache->discard_bitset);
775 spin_unlock_irqrestore(&cache->lock, flags);
776}
777
778static bool is_discarded(struct cache *cache, dm_dblock_t b)
779{
780 int r;
781 unsigned long flags;
782
783 spin_lock_irqsave(&cache->lock, flags);
784 r = test_bit(from_dblock(b), cache->discard_bitset);
785 spin_unlock_irqrestore(&cache->lock, flags);
786
787 return r;
788}
789
790static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
791{
792 int r;
793 unsigned long flags;
794
795 spin_lock_irqsave(&cache->lock, flags);
796 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
797 cache->discard_bitset);
798 spin_unlock_irqrestore(&cache->lock, flags);
799
800 return r;
801}
802
803/*----------------------------------------------------------------
804 * Remapping
805 *--------------------------------------------------------------*/
806static void remap_to_origin(struct cache *cache, struct bio *bio)
807{
808 bio_set_dev(bio, cache->origin_dev->bdev);
809}
810
811static void remap_to_cache(struct cache *cache, struct bio *bio,
812 dm_cblock_t cblock)
813{
814 sector_t bi_sector = bio->bi_iter.bi_sector;
815 sector_t block = from_cblock(cblock);
816
817 bio_set_dev(bio, cache->cache_dev->bdev);
818 if (!block_size_is_power_of_two(cache))
819 bio->bi_iter.bi_sector =
820 (block * cache->sectors_per_block) +
821 sector_div(bi_sector, cache->sectors_per_block);
822 else
823 bio->bi_iter.bi_sector =
824 (block << cache->sectors_per_block_shift) |
825 (bi_sector & (cache->sectors_per_block - 1));
826}
827
828static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
829{
830 unsigned long flags;
831 struct per_bio_data *pb;
832
833 spin_lock_irqsave(&cache->lock, flags);
834 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
835 bio_op(bio) != REQ_OP_DISCARD) {
836 pb = get_per_bio_data(bio);
837 pb->tick = true;
838 cache->need_tick_bio = false;
839 }
840 spin_unlock_irqrestore(&cache->lock, flags);
841}
842
843static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
844 dm_oblock_t oblock, bool bio_has_pbd)
845{
846 if (bio_has_pbd)
847 check_if_tick_bio_needed(cache, bio);
848 remap_to_origin(cache, bio);
849 if (bio_data_dir(bio) == WRITE)
850 clear_discard(cache, oblock_to_dblock(cache, oblock));
851}
852
853static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
854 dm_oblock_t oblock)
855{
856 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
857 __remap_to_origin_clear_discard(cache, bio, oblock, true);
858}
859
860static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
861 dm_oblock_t oblock, dm_cblock_t cblock)
862{
863 check_if_tick_bio_needed(cache, bio);
864 remap_to_cache(cache, bio, cblock);
865 if (bio_data_dir(bio) == WRITE) {
866 set_dirty(cache, cblock);
867 clear_discard(cache, oblock_to_dblock(cache, oblock));
868 }
869}
870
871static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
872{
873 sector_t block_nr = bio->bi_iter.bi_sector;
874
875 if (!block_size_is_power_of_two(cache))
876 (void) sector_div(block_nr, cache->sectors_per_block);
877 else
878 block_nr >>= cache->sectors_per_block_shift;
879
880 return to_oblock(block_nr);
881}
882
883static bool accountable_bio(struct cache *cache, struct bio *bio)
884{
885 return bio_op(bio) != REQ_OP_DISCARD;
886}
887
888static void accounted_begin(struct cache *cache, struct bio *bio)
889{
890 struct per_bio_data *pb;
891
892 if (accountable_bio(cache, bio)) {
893 pb = get_per_bio_data(bio);
894 pb->len = bio_sectors(bio);
895 iot_io_begin(&cache->tracker, pb->len);
896 }
897}
898
899static void accounted_complete(struct cache *cache, struct bio *bio)
900{
901 struct per_bio_data *pb = get_per_bio_data(bio);
902
903 iot_io_end(&cache->tracker, pb->len);
904}
905
906static void accounted_request(struct cache *cache, struct bio *bio)
907{
908 accounted_begin(cache, bio);
909 generic_make_request(bio);
910}
911
912static void issue_op(struct bio *bio, void *context)
913{
914 struct cache *cache = context;
915 accounted_request(cache, bio);
916}
917
918/*
919 * When running in writethrough mode we need to send writes to clean blocks
920 * to both the cache and origin devices. Clone the bio and send them in parallel.
921 */
922static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
923 dm_oblock_t oblock, dm_cblock_t cblock)
924{
925 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
926
927 BUG_ON(!origin_bio);
928
929 bio_chain(origin_bio, bio);
930 /*
931 * Passing false to __remap_to_origin_clear_discard() skips
932 * all code that might use per_bio_data (since clone doesn't have it)
933 */
934 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
935 submit_bio(origin_bio);
936
937 remap_to_cache(cache, bio, cblock);
938}
939
940/*----------------------------------------------------------------
941 * Failure modes
942 *--------------------------------------------------------------*/
943static enum cache_metadata_mode get_cache_mode(struct cache *cache)
944{
945 return cache->features.mode;
946}
947
948static const char *cache_device_name(struct cache *cache)
949{
950 return dm_device_name(dm_table_get_md(cache->ti->table));
951}
952
953static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
954{
955 const char *descs[] = {
956 "write",
957 "read-only",
958 "fail"
959 };
960
961 dm_table_event(cache->ti->table);
962 DMINFO("%s: switching cache to %s mode",
963 cache_device_name(cache), descs[(int)mode]);
964}
965
966static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
967{
968 bool needs_check;
969 enum cache_metadata_mode old_mode = get_cache_mode(cache);
970
971 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
972 DMERR("%s: unable to read needs_check flag, setting failure mode.",
973 cache_device_name(cache));
974 new_mode = CM_FAIL;
975 }
976
977 if (new_mode == CM_WRITE && needs_check) {
978 DMERR("%s: unable to switch cache to write mode until repaired.",
979 cache_device_name(cache));
980 if (old_mode != new_mode)
981 new_mode = old_mode;
982 else
983 new_mode = CM_READ_ONLY;
984 }
985
986 /* Never move out of fail mode */
987 if (old_mode == CM_FAIL)
988 new_mode = CM_FAIL;
989
990 switch (new_mode) {
991 case CM_FAIL:
992 case CM_READ_ONLY:
993 dm_cache_metadata_set_read_only(cache->cmd);
994 break;
995
996 case CM_WRITE:
997 dm_cache_metadata_set_read_write(cache->cmd);
998 break;
999 }
1000
1001 cache->features.mode = new_mode;
1002
1003 if (new_mode != old_mode)
1004 notify_mode_switch(cache, new_mode);
1005}
1006
1007static void abort_transaction(struct cache *cache)
1008{
1009 const char *dev_name = cache_device_name(cache);
1010
1011 if (get_cache_mode(cache) >= CM_READ_ONLY)
1012 return;
1013
1014 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1015 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1016 set_cache_mode(cache, CM_FAIL);
1017 }
1018
1019 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1020 if (dm_cache_metadata_abort(cache->cmd)) {
1021 DMERR("%s: failed to abort metadata transaction", dev_name);
1022 set_cache_mode(cache, CM_FAIL);
1023 }
1024}
1025
1026static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1027{
1028 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1029 cache_device_name(cache), op, r);
1030 abort_transaction(cache);
1031 set_cache_mode(cache, CM_READ_ONLY);
1032}
1033
1034/*----------------------------------------------------------------*/
1035
1036static void load_stats(struct cache *cache)
1037{
1038 struct dm_cache_statistics stats;
1039
1040 dm_cache_metadata_get_stats(cache->cmd, &stats);
1041 atomic_set(&cache->stats.read_hit, stats.read_hits);
1042 atomic_set(&cache->stats.read_miss, stats.read_misses);
1043 atomic_set(&cache->stats.write_hit, stats.write_hits);
1044 atomic_set(&cache->stats.write_miss, stats.write_misses);
1045}
1046
1047static void save_stats(struct cache *cache)
1048{
1049 struct dm_cache_statistics stats;
1050
1051 if (get_cache_mode(cache) >= CM_READ_ONLY)
1052 return;
1053
1054 stats.read_hits = atomic_read(&cache->stats.read_hit);
1055 stats.read_misses = atomic_read(&cache->stats.read_miss);
1056 stats.write_hits = atomic_read(&cache->stats.write_hit);
1057 stats.write_misses = atomic_read(&cache->stats.write_miss);
1058
1059 dm_cache_metadata_set_stats(cache->cmd, &stats);
1060}
1061
1062static void update_stats(struct cache_stats *stats, enum policy_operation op)
1063{
1064 switch (op) {
1065 case POLICY_PROMOTE:
1066 atomic_inc(&stats->promotion);
1067 break;
1068
1069 case POLICY_DEMOTE:
1070 atomic_inc(&stats->demotion);
1071 break;
1072
1073 case POLICY_WRITEBACK:
1074 atomic_inc(&stats->writeback);
1075 break;
1076 }
1077}
1078
1079/*----------------------------------------------------------------
1080 * Migration processing
1081 *
1082 * Migration covers moving data from the origin device to the cache, or
1083 * vice versa.
1084 *--------------------------------------------------------------*/
1085
1086static void inc_io_migrations(struct cache *cache)
1087{
1088 atomic_inc(&cache->nr_io_migrations);
1089}
1090
1091static void dec_io_migrations(struct cache *cache)
1092{
1093 atomic_dec(&cache->nr_io_migrations);
1094}
1095
1096static bool discard_or_flush(struct bio *bio)
1097{
1098 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1099}
1100
1101static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1102 dm_dblock_t *b, dm_dblock_t *e)
1103{
1104 sector_t sb = bio->bi_iter.bi_sector;
1105 sector_t se = bio_end_sector(bio);
1106
1107 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1108
1109 if (se - sb < cache->discard_block_size)
1110 *e = *b;
1111 else
1112 *e = to_dblock(block_div(se, cache->discard_block_size));
1113}
1114
1115/*----------------------------------------------------------------*/
1116
1117static void prevent_background_work(struct cache *cache)
1118{
1119 lockdep_off();
1120 down_write(&cache->background_work_lock);
1121 lockdep_on();
1122}
1123
1124static void allow_background_work(struct cache *cache)
1125{
1126 lockdep_off();
1127 up_write(&cache->background_work_lock);
1128 lockdep_on();
1129}
1130
1131static bool background_work_begin(struct cache *cache)
1132{
1133 bool r;
1134
1135 lockdep_off();
1136 r = down_read_trylock(&cache->background_work_lock);
1137 lockdep_on();
1138
1139 return r;
1140}
1141
1142static void background_work_end(struct cache *cache)
1143{
1144 lockdep_off();
1145 up_read(&cache->background_work_lock);
1146 lockdep_on();
1147}
1148
1149/*----------------------------------------------------------------*/
1150
1151static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1152{
1153 return (bio_data_dir(bio) == WRITE) &&
1154 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1155}
1156
1157static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1158{
1159 return writeback_mode(cache) &&
1160 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1161}
1162
1163static void quiesce(struct dm_cache_migration *mg,
1164 void (*continuation)(struct work_struct *))
1165{
1166 init_continuation(&mg->k, continuation);
1167 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1168}
1169
1170static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1171{
1172 struct continuation *k = container_of(ws, struct continuation, ws);
1173 return container_of(k, struct dm_cache_migration, k);
1174}
1175
1176static void copy_complete(int read_err, unsigned long write_err, void *context)
1177{
1178 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1179
1180 if (read_err || write_err)
1181 mg->k.input = BLK_STS_IOERR;
1182
1183 queue_continuation(mg->cache->wq, &mg->k);
1184}
1185
1186static void copy(struct dm_cache_migration *mg, bool promote)
1187{
1188 struct dm_io_region o_region, c_region;
1189 struct cache *cache = mg->cache;
1190
1191 o_region.bdev = cache->origin_dev->bdev;
1192 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1193 o_region.count = cache->sectors_per_block;
1194
1195 c_region.bdev = cache->cache_dev->bdev;
1196 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1197 c_region.count = cache->sectors_per_block;
1198
1199 if (promote)
1200 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1201 else
1202 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1203}
1204
1205static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1206{
1207 struct per_bio_data *pb = get_per_bio_data(bio);
1208
1209 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1210 free_prison_cell(cache, pb->cell);
1211 pb->cell = NULL;
1212}
1213
1214static void overwrite_endio(struct bio *bio)
1215{
1216 struct dm_cache_migration *mg = bio->bi_private;
1217 struct cache *cache = mg->cache;
1218 struct per_bio_data *pb = get_per_bio_data(bio);
1219
1220 dm_unhook_bio(&pb->hook_info, bio);
1221
1222 if (bio->bi_status)
1223 mg->k.input = bio->bi_status;
1224
1225 queue_continuation(cache->wq, &mg->k);
1226}
1227
1228static void overwrite(struct dm_cache_migration *mg,
1229 void (*continuation)(struct work_struct *))
1230{
1231 struct bio *bio = mg->overwrite_bio;
1232 struct per_bio_data *pb = get_per_bio_data(bio);
1233
1234 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1235
1236 /*
1237 * The overwrite bio is part of the copy operation, as such it does
1238 * not set/clear discard or dirty flags.
1239 */
1240 if (mg->op->op == POLICY_PROMOTE)
1241 remap_to_cache(mg->cache, bio, mg->op->cblock);
1242 else
1243 remap_to_origin(mg->cache, bio);
1244
1245 init_continuation(&mg->k, continuation);
1246 accounted_request(mg->cache, bio);
1247}
1248
1249/*
1250 * Migration steps:
1251 *
1252 * 1) exclusive lock preventing WRITEs
1253 * 2) quiesce
1254 * 3) copy or issue overwrite bio
1255 * 4) upgrade to exclusive lock preventing READs and WRITEs
1256 * 5) quiesce
1257 * 6) update metadata and commit
1258 * 7) unlock
1259 */
1260static void mg_complete(struct dm_cache_migration *mg, bool success)
1261{
1262 struct bio_list bios;
1263 struct cache *cache = mg->cache;
1264 struct policy_work *op = mg->op;
1265 dm_cblock_t cblock = op->cblock;
1266
1267 if (success)
1268 update_stats(&cache->stats, op->op);
1269
1270 switch (op->op) {
1271 case POLICY_PROMOTE:
1272 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1273 policy_complete_background_work(cache->policy, op, success);
1274
1275 if (mg->overwrite_bio) {
1276 if (success)
1277 force_set_dirty(cache, cblock);
1278 else if (mg->k.input)
1279 mg->overwrite_bio->bi_status = mg->k.input;
1280 else
1281 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1282 bio_endio(mg->overwrite_bio);
1283 } else {
1284 if (success)
1285 force_clear_dirty(cache, cblock);
1286 dec_io_migrations(cache);
1287 }
1288 break;
1289
1290 case POLICY_DEMOTE:
1291 /*
1292 * We clear dirty here to update the nr_dirty counter.
1293 */
1294 if (success)
1295 force_clear_dirty(cache, cblock);
1296 policy_complete_background_work(cache->policy, op, success);
1297 dec_io_migrations(cache);
1298 break;
1299
1300 case POLICY_WRITEBACK:
1301 if (success)
1302 force_clear_dirty(cache, cblock);
1303 policy_complete_background_work(cache->policy, op, success);
1304 dec_io_migrations(cache);
1305 break;
1306 }
1307
1308 bio_list_init(&bios);
1309 if (mg->cell) {
1310 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1311 free_prison_cell(cache, mg->cell);
1312 }
1313
1314 free_migration(mg);
1315 defer_bios(cache, &bios);
1316 wake_migration_worker(cache);
1317
1318 background_work_end(cache);
1319}
1320
1321static void mg_success(struct work_struct *ws)
1322{
1323 struct dm_cache_migration *mg = ws_to_mg(ws);
1324 mg_complete(mg, mg->k.input == 0);
1325}
1326
1327static void mg_update_metadata(struct work_struct *ws)
1328{
1329 int r;
1330 struct dm_cache_migration *mg = ws_to_mg(ws);
1331 struct cache *cache = mg->cache;
1332 struct policy_work *op = mg->op;
1333
1334 switch (op->op) {
1335 case POLICY_PROMOTE:
1336 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1337 if (r) {
1338 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1339 cache_device_name(cache));
1340 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1341
1342 mg_complete(mg, false);
1343 return;
1344 }
1345 mg_complete(mg, true);
1346 break;
1347
1348 case POLICY_DEMOTE:
1349 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1350 if (r) {
1351 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1352 cache_device_name(cache));
1353 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1354
1355 mg_complete(mg, false);
1356 return;
1357 }
1358
1359 /*
1360 * It would be nice if we only had to commit when a REQ_FLUSH
1361 * comes through. But there's one scenario that we have to
1362 * look out for:
1363 *
1364 * - vblock x in a cache block
1365 * - domotion occurs
1366 * - cache block gets reallocated and over written
1367 * - crash
1368 *
1369 * When we recover, because there was no commit the cache will
1370 * rollback to having the data for vblock x in the cache block.
1371 * But the cache block has since been overwritten, so it'll end
1372 * up pointing to data that was never in 'x' during the history
1373 * of the device.
1374 *
1375 * To avoid this issue we require a commit as part of the
1376 * demotion operation.
1377 */
1378 init_continuation(&mg->k, mg_success);
1379 continue_after_commit(&cache->committer, &mg->k);
1380 schedule_commit(&cache->committer);
1381 break;
1382
1383 case POLICY_WRITEBACK:
1384 mg_complete(mg, true);
1385 break;
1386 }
1387}
1388
1389static void mg_update_metadata_after_copy(struct work_struct *ws)
1390{
1391 struct dm_cache_migration *mg = ws_to_mg(ws);
1392
1393 /*
1394 * Did the copy succeed?
1395 */
1396 if (mg->k.input)
1397 mg_complete(mg, false);
1398 else
1399 mg_update_metadata(ws);
1400}
1401
1402static void mg_upgrade_lock(struct work_struct *ws)
1403{
1404 int r;
1405 struct dm_cache_migration *mg = ws_to_mg(ws);
1406
1407 /*
1408 * Did the copy succeed?
1409 */
1410 if (mg->k.input)
1411 mg_complete(mg, false);
1412
1413 else {
1414 /*
1415 * Now we want the lock to prevent both reads and writes.
1416 */
1417 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1418 READ_WRITE_LOCK_LEVEL);
1419 if (r < 0)
1420 mg_complete(mg, false);
1421
1422 else if (r)
1423 quiesce(mg, mg_update_metadata);
1424
1425 else
1426 mg_update_metadata(ws);
1427 }
1428}
1429
1430static void mg_full_copy(struct work_struct *ws)
1431{
1432 struct dm_cache_migration *mg = ws_to_mg(ws);
1433 struct cache *cache = mg->cache;
1434 struct policy_work *op = mg->op;
1435 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1436
1437 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1438 is_discarded_oblock(cache, op->oblock)) {
1439 mg_upgrade_lock(ws);
1440 return;
1441 }
1442
1443 init_continuation(&mg->k, mg_upgrade_lock);
1444 copy(mg, is_policy_promote);
1445}
1446
1447static void mg_copy(struct work_struct *ws)
1448{
1449 struct dm_cache_migration *mg = ws_to_mg(ws);
1450
1451 if (mg->overwrite_bio) {
1452 /*
1453 * No exclusive lock was held when we last checked if the bio
1454 * was optimisable. So we have to check again in case things
1455 * have changed (eg, the block may no longer be discarded).
1456 */
1457 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1458 /*
1459 * Fallback to a real full copy after doing some tidying up.
1460 */
1461 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1462 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1463 mg->overwrite_bio = NULL;
1464 inc_io_migrations(mg->cache);
1465 mg_full_copy(ws);
1466 return;
1467 }
1468
1469 /*
1470 * It's safe to do this here, even though it's new data
1471 * because all IO has been locked out of the block.
1472 *
1473 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1474 * so _not_ using mg_upgrade_lock() as continutation.
1475 */
1476 overwrite(mg, mg_update_metadata_after_copy);
1477
1478 } else
1479 mg_full_copy(ws);
1480}
1481
1482static int mg_lock_writes(struct dm_cache_migration *mg)
1483{
1484 int r;
1485 struct dm_cell_key_v2 key;
1486 struct cache *cache = mg->cache;
1487 struct dm_bio_prison_cell_v2 *prealloc;
1488
1489 prealloc = alloc_prison_cell(cache);
1490
1491 /*
1492 * Prevent writes to the block, but allow reads to continue.
1493 * Unless we're using an overwrite bio, in which case we lock
1494 * everything.
1495 */
1496 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1497 r = dm_cell_lock_v2(cache->prison, &key,
1498 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1499 prealloc, &mg->cell);
1500 if (r < 0) {
1501 free_prison_cell(cache, prealloc);
1502 mg_complete(mg, false);
1503 return r;
1504 }
1505
1506 if (mg->cell != prealloc)
1507 free_prison_cell(cache, prealloc);
1508
1509 if (r == 0)
1510 mg_copy(&mg->k.ws);
1511 else
1512 quiesce(mg, mg_copy);
1513
1514 return 0;
1515}
1516
1517static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1518{
1519 struct dm_cache_migration *mg;
1520
1521 if (!background_work_begin(cache)) {
1522 policy_complete_background_work(cache->policy, op, false);
1523 return -EPERM;
1524 }
1525
1526 mg = alloc_migration(cache);
1527
1528 mg->op = op;
1529 mg->overwrite_bio = bio;
1530
1531 if (!bio)
1532 inc_io_migrations(cache);
1533
1534 return mg_lock_writes(mg);
1535}
1536
1537/*----------------------------------------------------------------
1538 * invalidation processing
1539 *--------------------------------------------------------------*/
1540
1541static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1542{
1543 struct bio_list bios;
1544 struct cache *cache = mg->cache;
1545
1546 bio_list_init(&bios);
1547 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1548 free_prison_cell(cache, mg->cell);
1549
1550 if (!success && mg->overwrite_bio)
1551 bio_io_error(mg->overwrite_bio);
1552
1553 free_migration(mg);
1554 defer_bios(cache, &bios);
1555
1556 background_work_end(cache);
1557}
1558
1559static void invalidate_completed(struct work_struct *ws)
1560{
1561 struct dm_cache_migration *mg = ws_to_mg(ws);
1562 invalidate_complete(mg, !mg->k.input);
1563}
1564
1565static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1566{
1567 int r = policy_invalidate_mapping(cache->policy, cblock);
1568 if (!r) {
1569 r = dm_cache_remove_mapping(cache->cmd, cblock);
1570 if (r) {
1571 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1572 cache_device_name(cache));
1573 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1574 }
1575
1576 } else if (r == -ENODATA) {
1577 /*
1578 * Harmless, already unmapped.
1579 */
1580 r = 0;
1581
1582 } else
1583 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1584
1585 return r;
1586}
1587
1588static void invalidate_remove(struct work_struct *ws)
1589{
1590 int r;
1591 struct dm_cache_migration *mg = ws_to_mg(ws);
1592 struct cache *cache = mg->cache;
1593
1594 r = invalidate_cblock(cache, mg->invalidate_cblock);
1595 if (r) {
1596 invalidate_complete(mg, false);
1597 return;
1598 }
1599
1600 init_continuation(&mg->k, invalidate_completed);
1601 continue_after_commit(&cache->committer, &mg->k);
1602 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1603 mg->overwrite_bio = NULL;
1604 schedule_commit(&cache->committer);
1605}
1606
1607static int invalidate_lock(struct dm_cache_migration *mg)
1608{
1609 int r;
1610 struct dm_cell_key_v2 key;
1611 struct cache *cache = mg->cache;
1612 struct dm_bio_prison_cell_v2 *prealloc;
1613
1614 prealloc = alloc_prison_cell(cache);
1615
1616 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1617 r = dm_cell_lock_v2(cache->prison, &key,
1618 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1619 if (r < 0) {
1620 free_prison_cell(cache, prealloc);
1621 invalidate_complete(mg, false);
1622 return r;
1623 }
1624
1625 if (mg->cell != prealloc)
1626 free_prison_cell(cache, prealloc);
1627
1628 if (r)
1629 quiesce(mg, invalidate_remove);
1630
1631 else {
1632 /*
1633 * We can't call invalidate_remove() directly here because we
1634 * might still be in request context.
1635 */
1636 init_continuation(&mg->k, invalidate_remove);
1637 queue_work(cache->wq, &mg->k.ws);
1638 }
1639
1640 return 0;
1641}
1642
1643static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1644 dm_oblock_t oblock, struct bio *bio)
1645{
1646 struct dm_cache_migration *mg;
1647
1648 if (!background_work_begin(cache))
1649 return -EPERM;
1650
1651 mg = alloc_migration(cache);
1652
1653 mg->overwrite_bio = bio;
1654 mg->invalidate_cblock = cblock;
1655 mg->invalidate_oblock = oblock;
1656
1657 return invalidate_lock(mg);
1658}
1659
1660/*----------------------------------------------------------------
1661 * bio processing
1662 *--------------------------------------------------------------*/
1663
1664enum busy {
1665 IDLE,
1666 BUSY
1667};
1668
1669static enum busy spare_migration_bandwidth(struct cache *cache)
1670{
1671 bool idle = iot_idle_for(&cache->tracker, HZ);
1672 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1673 cache->sectors_per_block;
1674
1675 if (idle && current_volume <= cache->migration_threshold)
1676 return IDLE;
1677 else
1678 return BUSY;
1679}
1680
1681static void inc_hit_counter(struct cache *cache, struct bio *bio)
1682{
1683 atomic_inc(bio_data_dir(bio) == READ ?
1684 &cache->stats.read_hit : &cache->stats.write_hit);
1685}
1686
1687static void inc_miss_counter(struct cache *cache, struct bio *bio)
1688{
1689 atomic_inc(bio_data_dir(bio) == READ ?
1690 &cache->stats.read_miss : &cache->stats.write_miss);
1691}
1692
1693/*----------------------------------------------------------------*/
1694
1695static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1696 bool *commit_needed)
1697{
1698 int r, data_dir;
1699 bool rb, background_queued;
1700 dm_cblock_t cblock;
1701
1702 *commit_needed = false;
1703
1704 rb = bio_detain_shared(cache, block, bio);
1705 if (!rb) {
1706 /*
1707 * An exclusive lock is held for this block, so we have to
1708 * wait. We set the commit_needed flag so the current
1709 * transaction will be committed asap, allowing this lock
1710 * to be dropped.
1711 */
1712 *commit_needed = true;
1713 return DM_MAPIO_SUBMITTED;
1714 }
1715
1716 data_dir = bio_data_dir(bio);
1717
1718 if (optimisable_bio(cache, bio, block)) {
1719 struct policy_work *op = NULL;
1720
1721 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1722 if (unlikely(r && r != -ENOENT)) {
1723 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1724 cache_device_name(cache), r);
1725 bio_io_error(bio);
1726 return DM_MAPIO_SUBMITTED;
1727 }
1728
1729 if (r == -ENOENT && op) {
1730 bio_drop_shared_lock(cache, bio);
1731 BUG_ON(op->op != POLICY_PROMOTE);
1732 mg_start(cache, op, bio);
1733 return DM_MAPIO_SUBMITTED;
1734 }
1735 } else {
1736 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1737 if (unlikely(r && r != -ENOENT)) {
1738 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1739 cache_device_name(cache), r);
1740 bio_io_error(bio);
1741 return DM_MAPIO_SUBMITTED;
1742 }
1743
1744 if (background_queued)
1745 wake_migration_worker(cache);
1746 }
1747
1748 if (r == -ENOENT) {
1749 struct per_bio_data *pb = get_per_bio_data(bio);
1750
1751 /*
1752 * Miss.
1753 */
1754 inc_miss_counter(cache, bio);
1755 if (pb->req_nr == 0) {
1756 accounted_begin(cache, bio);
1757 remap_to_origin_clear_discard(cache, bio, block);
1758 } else {
1759 /*
1760 * This is a duplicate writethrough io that is no
1761 * longer needed because the block has been demoted.
1762 */
1763 bio_endio(bio);
1764 return DM_MAPIO_SUBMITTED;
1765 }
1766 } else {
1767 /*
1768 * Hit.
1769 */
1770 inc_hit_counter(cache, bio);
1771
1772 /*
1773 * Passthrough always maps to the origin, invalidating any
1774 * cache blocks that are written to.
1775 */
1776 if (passthrough_mode(cache)) {
1777 if (bio_data_dir(bio) == WRITE) {
1778 bio_drop_shared_lock(cache, bio);
1779 atomic_inc(&cache->stats.demotion);
1780 invalidate_start(cache, cblock, block, bio);
1781 } else
1782 remap_to_origin_clear_discard(cache, bio, block);
1783 } else {
1784 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1785 !is_dirty(cache, cblock)) {
1786 remap_to_origin_and_cache(cache, bio, block, cblock);
1787 accounted_begin(cache, bio);
1788 } else
1789 remap_to_cache_dirty(cache, bio, block, cblock);
1790 }
1791 }
1792
1793 /*
1794 * dm core turns FUA requests into a separate payload and FLUSH req.
1795 */
1796 if (bio->bi_opf & REQ_FUA) {
1797 /*
1798 * issue_after_commit will call accounted_begin a second time. So
1799 * we call accounted_complete() to avoid double accounting.
1800 */
1801 accounted_complete(cache, bio);
1802 issue_after_commit(&cache->committer, bio);
1803 *commit_needed = true;
1804 return DM_MAPIO_SUBMITTED;
1805 }
1806
1807 return DM_MAPIO_REMAPPED;
1808}
1809
1810static bool process_bio(struct cache *cache, struct bio *bio)
1811{
1812 bool commit_needed;
1813
1814 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1815 generic_make_request(bio);
1816
1817 return commit_needed;
1818}
1819
1820/*
1821 * A non-zero return indicates read_only or fail_io mode.
1822 */
1823static int commit(struct cache *cache, bool clean_shutdown)
1824{
1825 int r;
1826
1827 if (get_cache_mode(cache) >= CM_READ_ONLY)
1828 return -EINVAL;
1829
1830 atomic_inc(&cache->stats.commit_count);
1831 r = dm_cache_commit(cache->cmd, clean_shutdown);
1832 if (r)
1833 metadata_operation_failed(cache, "dm_cache_commit", r);
1834
1835 return r;
1836}
1837
1838/*
1839 * Used by the batcher.
1840 */
1841static blk_status_t commit_op(void *context)
1842{
1843 struct cache *cache = context;
1844
1845 if (dm_cache_changed_this_transaction(cache->cmd))
1846 return errno_to_blk_status(commit(cache, false));
1847
1848 return 0;
1849}
1850
1851/*----------------------------------------------------------------*/
1852
1853static bool process_flush_bio(struct cache *cache, struct bio *bio)
1854{
1855 struct per_bio_data *pb = get_per_bio_data(bio);
1856
1857 if (!pb->req_nr)
1858 remap_to_origin(cache, bio);
1859 else
1860 remap_to_cache(cache, bio, 0);
1861
1862 issue_after_commit(&cache->committer, bio);
1863 return true;
1864}
1865
1866static bool process_discard_bio(struct cache *cache, struct bio *bio)
1867{
1868 dm_dblock_t b, e;
1869
1870 // FIXME: do we need to lock the region? Or can we just assume the
1871 // user wont be so foolish as to issue discard concurrently with
1872 // other IO?
1873 calc_discard_block_range(cache, bio, &b, &e);
1874 while (b != e) {
1875 set_discard(cache, b);
1876 b = to_dblock(from_dblock(b) + 1);
1877 }
1878
1879 if (cache->features.discard_passdown) {
1880 remap_to_origin(cache, bio);
1881 generic_make_request(bio);
1882 } else
1883 bio_endio(bio);
1884
1885 return false;
1886}
1887
1888static void process_deferred_bios(struct work_struct *ws)
1889{
1890 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1891
1892 unsigned long flags;
1893 bool commit_needed = false;
1894 struct bio_list bios;
1895 struct bio *bio;
1896
1897 bio_list_init(&bios);
1898
1899 spin_lock_irqsave(&cache->lock, flags);
1900 bio_list_merge(&bios, &cache->deferred_bios);
1901 bio_list_init(&cache->deferred_bios);
1902 spin_unlock_irqrestore(&cache->lock, flags);
1903
1904 while ((bio = bio_list_pop(&bios))) {
1905 if (bio->bi_opf & REQ_PREFLUSH)
1906 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1907
1908 else if (bio_op(bio) == REQ_OP_DISCARD)
1909 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1910
1911 else
1912 commit_needed = process_bio(cache, bio) || commit_needed;
1913 }
1914
1915 if (commit_needed)
1916 schedule_commit(&cache->committer);
1917}
1918
1919/*----------------------------------------------------------------
1920 * Main worker loop
1921 *--------------------------------------------------------------*/
1922
1923static void requeue_deferred_bios(struct cache *cache)
1924{
1925 struct bio *bio;
1926 struct bio_list bios;
1927
1928 bio_list_init(&bios);
1929 bio_list_merge(&bios, &cache->deferred_bios);
1930 bio_list_init(&cache->deferred_bios);
1931
1932 while ((bio = bio_list_pop(&bios))) {
1933 bio->bi_status = BLK_STS_DM_REQUEUE;
1934 bio_endio(bio);
1935 }
1936}
1937
1938/*
1939 * We want to commit periodically so that not too much
1940 * unwritten metadata builds up.
1941 */
1942static void do_waker(struct work_struct *ws)
1943{
1944 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1945
1946 policy_tick(cache->policy, true);
1947 wake_migration_worker(cache);
1948 schedule_commit(&cache->committer);
1949 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1950}
1951
1952static void check_migrations(struct work_struct *ws)
1953{
1954 int r;
1955 struct policy_work *op;
1956 struct cache *cache = container_of(ws, struct cache, migration_worker);
1957 enum busy b;
1958
1959 for (;;) {
1960 b = spare_migration_bandwidth(cache);
1961
1962 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1963 if (r == -ENODATA)
1964 break;
1965
1966 if (r) {
1967 DMERR_LIMIT("%s: policy_background_work failed",
1968 cache_device_name(cache));
1969 break;
1970 }
1971
1972 r = mg_start(cache, op, NULL);
1973 if (r)
1974 break;
1975 }
1976}
1977
1978/*----------------------------------------------------------------
1979 * Target methods
1980 *--------------------------------------------------------------*/
1981
1982/*
1983 * This function gets called on the error paths of the constructor, so we
1984 * have to cope with a partially initialised struct.
1985 */
1986static void destroy(struct cache *cache)
1987{
1988 unsigned i;
1989
1990 mempool_exit(&cache->migration_pool);
1991
1992 if (cache->prison)
1993 dm_bio_prison_destroy_v2(cache->prison);
1994
1995 if (cache->wq)
1996 destroy_workqueue(cache->wq);
1997
1998 if (cache->dirty_bitset)
1999 free_bitset(cache->dirty_bitset);
2000
2001 if (cache->discard_bitset)
2002 free_bitset(cache->discard_bitset);
2003
2004 if (cache->copier)
2005 dm_kcopyd_client_destroy(cache->copier);
2006
2007 if (cache->cmd)
2008 dm_cache_metadata_close(cache->cmd);
2009
2010 if (cache->metadata_dev)
2011 dm_put_device(cache->ti, cache->metadata_dev);
2012
2013 if (cache->origin_dev)
2014 dm_put_device(cache->ti, cache->origin_dev);
2015
2016 if (cache->cache_dev)
2017 dm_put_device(cache->ti, cache->cache_dev);
2018
2019 if (cache->policy)
2020 dm_cache_policy_destroy(cache->policy);
2021
2022 for (i = 0; i < cache->nr_ctr_args ; i++)
2023 kfree(cache->ctr_args[i]);
2024 kfree(cache->ctr_args);
2025
2026 bioset_exit(&cache->bs);
2027
2028 kfree(cache);
2029}
2030
2031static void cache_dtr(struct dm_target *ti)
2032{
2033 struct cache *cache = ti->private;
2034
2035 destroy(cache);
2036}
2037
2038static sector_t get_dev_size(struct dm_dev *dev)
2039{
2040 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2041}
2042
2043/*----------------------------------------------------------------*/
2044
2045/*
2046 * Construct a cache device mapping.
2047 *
2048 * cache <metadata dev> <cache dev> <origin dev> <block size>
2049 * <#feature args> [<feature arg>]*
2050 * <policy> <#policy args> [<policy arg>]*
2051 *
2052 * metadata dev : fast device holding the persistent metadata
2053 * cache dev : fast device holding cached data blocks
2054 * origin dev : slow device holding original data blocks
2055 * block size : cache unit size in sectors
2056 *
2057 * #feature args : number of feature arguments passed
2058 * feature args : writethrough. (The default is writeback.)
2059 *
2060 * policy : the replacement policy to use
2061 * #policy args : an even number of policy arguments corresponding
2062 * to key/value pairs passed to the policy
2063 * policy args : key/value pairs passed to the policy
2064 * E.g. 'sequential_threshold 1024'
2065 * See cache-policies.txt for details.
2066 *
2067 * Optional feature arguments are:
2068 * writethrough : write through caching that prohibits cache block
2069 * content from being different from origin block content.
2070 * Without this argument, the default behaviour is to write
2071 * back cache block contents later for performance reasons,
2072 * so they may differ from the corresponding origin blocks.
2073 */
2074struct cache_args {
2075 struct dm_target *ti;
2076
2077 struct dm_dev *metadata_dev;
2078
2079 struct dm_dev *cache_dev;
2080 sector_t cache_sectors;
2081
2082 struct dm_dev *origin_dev;
2083 sector_t origin_sectors;
2084
2085 uint32_t block_size;
2086
2087 const char *policy_name;
2088 int policy_argc;
2089 const char **policy_argv;
2090
2091 struct cache_features features;
2092};
2093
2094static void destroy_cache_args(struct cache_args *ca)
2095{
2096 if (ca->metadata_dev)
2097 dm_put_device(ca->ti, ca->metadata_dev);
2098
2099 if (ca->cache_dev)
2100 dm_put_device(ca->ti, ca->cache_dev);
2101
2102 if (ca->origin_dev)
2103 dm_put_device(ca->ti, ca->origin_dev);
2104
2105 kfree(ca);
2106}
2107
2108static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2109{
2110 if (!as->argc) {
2111 *error = "Insufficient args";
2112 return false;
2113 }
2114
2115 return true;
2116}
2117
2118static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2119 char **error)
2120{
2121 int r;
2122 sector_t metadata_dev_size;
2123 char b[BDEVNAME_SIZE];
2124
2125 if (!at_least_one_arg(as, error))
2126 return -EINVAL;
2127
2128 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2129 &ca->metadata_dev);
2130 if (r) {
2131 *error = "Error opening metadata device";
2132 return r;
2133 }
2134
2135 metadata_dev_size = get_dev_size(ca->metadata_dev);
2136 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2137 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2138 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2139
2140 return 0;
2141}
2142
2143static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2144 char **error)
2145{
2146 int r;
2147
2148 if (!at_least_one_arg(as, error))
2149 return -EINVAL;
2150
2151 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2152 &ca->cache_dev);
2153 if (r) {
2154 *error = "Error opening cache device";
2155 return r;
2156 }
2157 ca->cache_sectors = get_dev_size(ca->cache_dev);
2158
2159 return 0;
2160}
2161
2162static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2163 char **error)
2164{
2165 int r;
2166
2167 if (!at_least_one_arg(as, error))
2168 return -EINVAL;
2169
2170 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2171 &ca->origin_dev);
2172 if (r) {
2173 *error = "Error opening origin device";
2174 return r;
2175 }
2176
2177 ca->origin_sectors = get_dev_size(ca->origin_dev);
2178 if (ca->ti->len > ca->origin_sectors) {
2179 *error = "Device size larger than cached device";
2180 return -EINVAL;
2181 }
2182
2183 return 0;
2184}
2185
2186static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2187 char **error)
2188{
2189 unsigned long block_size;
2190
2191 if (!at_least_one_arg(as, error))
2192 return -EINVAL;
2193
2194 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2195 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2196 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2197 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2198 *error = "Invalid data block size";
2199 return -EINVAL;
2200 }
2201
2202 if (block_size > ca->cache_sectors) {
2203 *error = "Data block size is larger than the cache device";
2204 return -EINVAL;
2205 }
2206
2207 ca->block_size = block_size;
2208
2209 return 0;
2210}
2211
2212static void init_features(struct cache_features *cf)
2213{
2214 cf->mode = CM_WRITE;
2215 cf->io_mode = CM_IO_WRITEBACK;
2216 cf->metadata_version = 1;
2217 cf->discard_passdown = true;
2218}
2219
2220static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2221 char **error)
2222{
2223 static const struct dm_arg _args[] = {
2224 {0, 3, "Invalid number of cache feature arguments"},
2225 };
2226
2227 int r, mode_ctr = 0;
2228 unsigned argc;
2229 const char *arg;
2230 struct cache_features *cf = &ca->features;
2231
2232 init_features(cf);
2233
2234 r = dm_read_arg_group(_args, as, &argc, error);
2235 if (r)
2236 return -EINVAL;
2237
2238 while (argc--) {
2239 arg = dm_shift_arg(as);
2240
2241 if (!strcasecmp(arg, "writeback")) {
2242 cf->io_mode = CM_IO_WRITEBACK;
2243 mode_ctr++;
2244 }
2245
2246 else if (!strcasecmp(arg, "writethrough")) {
2247 cf->io_mode = CM_IO_WRITETHROUGH;
2248 mode_ctr++;
2249 }
2250
2251 else if (!strcasecmp(arg, "passthrough")) {
2252 cf->io_mode = CM_IO_PASSTHROUGH;
2253 mode_ctr++;
2254 }
2255
2256 else if (!strcasecmp(arg, "metadata2"))
2257 cf->metadata_version = 2;
2258
2259 else if (!strcasecmp(arg, "no_discard_passdown"))
2260 cf->discard_passdown = false;
2261
2262 else {
2263 *error = "Unrecognised cache feature requested";
2264 return -EINVAL;
2265 }
2266 }
2267
2268 if (mode_ctr > 1) {
2269 *error = "Duplicate cache io_mode features requested";
2270 return -EINVAL;
2271 }
2272
2273 return 0;
2274}
2275
2276static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2277 char **error)
2278{
2279 static const struct dm_arg _args[] = {
2280 {0, 1024, "Invalid number of policy arguments"},
2281 };
2282
2283 int r;
2284
2285 if (!at_least_one_arg(as, error))
2286 return -EINVAL;
2287
2288 ca->policy_name = dm_shift_arg(as);
2289
2290 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2291 if (r)
2292 return -EINVAL;
2293
2294 ca->policy_argv = (const char **)as->argv;
2295 dm_consume_args(as, ca->policy_argc);
2296
2297 return 0;
2298}
2299
2300static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2301 char **error)
2302{
2303 int r;
2304 struct dm_arg_set as;
2305
2306 as.argc = argc;
2307 as.argv = argv;
2308
2309 r = parse_metadata_dev(ca, &as, error);
2310 if (r)
2311 return r;
2312
2313 r = parse_cache_dev(ca, &as, error);
2314 if (r)
2315 return r;
2316
2317 r = parse_origin_dev(ca, &as, error);
2318 if (r)
2319 return r;
2320
2321 r = parse_block_size(ca, &as, error);
2322 if (r)
2323 return r;
2324
2325 r = parse_features(ca, &as, error);
2326 if (r)
2327 return r;
2328
2329 r = parse_policy(ca, &as, error);
2330 if (r)
2331 return r;
2332
2333 return 0;
2334}
2335
2336/*----------------------------------------------------------------*/
2337
2338static struct kmem_cache *migration_cache;
2339
2340#define NOT_CORE_OPTION 1
2341
2342static int process_config_option(struct cache *cache, const char *key, const char *value)
2343{
2344 unsigned long tmp;
2345
2346 if (!strcasecmp(key, "migration_threshold")) {
2347 if (kstrtoul(value, 10, &tmp))
2348 return -EINVAL;
2349
2350 cache->migration_threshold = tmp;
2351 return 0;
2352 }
2353
2354 return NOT_CORE_OPTION;
2355}
2356
2357static int set_config_value(struct cache *cache, const char *key, const char *value)
2358{
2359 int r = process_config_option(cache, key, value);
2360
2361 if (r == NOT_CORE_OPTION)
2362 r = policy_set_config_value(cache->policy, key, value);
2363
2364 if (r)
2365 DMWARN("bad config value for %s: %s", key, value);
2366
2367 return r;
2368}
2369
2370static int set_config_values(struct cache *cache, int argc, const char **argv)
2371{
2372 int r = 0;
2373
2374 if (argc & 1) {
2375 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2376 return -EINVAL;
2377 }
2378
2379 while (argc) {
2380 r = set_config_value(cache, argv[0], argv[1]);
2381 if (r)
2382 break;
2383
2384 argc -= 2;
2385 argv += 2;
2386 }
2387
2388 return r;
2389}
2390
2391static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2392 char **error)
2393{
2394 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2395 cache->cache_size,
2396 cache->origin_sectors,
2397 cache->sectors_per_block);
2398 if (IS_ERR(p)) {
2399 *error = "Error creating cache's policy";
2400 return PTR_ERR(p);
2401 }
2402 cache->policy = p;
2403 BUG_ON(!cache->policy);
2404
2405 return 0;
2406}
2407
2408/*
2409 * We want the discard block size to be at least the size of the cache
2410 * block size and have no more than 2^14 discard blocks across the origin.
2411 */
2412#define MAX_DISCARD_BLOCKS (1 << 14)
2413
2414static bool too_many_discard_blocks(sector_t discard_block_size,
2415 sector_t origin_size)
2416{
2417 (void) sector_div(origin_size, discard_block_size);
2418
2419 return origin_size > MAX_DISCARD_BLOCKS;
2420}
2421
2422static sector_t calculate_discard_block_size(sector_t cache_block_size,
2423 sector_t origin_size)
2424{
2425 sector_t discard_block_size = cache_block_size;
2426
2427 if (origin_size)
2428 while (too_many_discard_blocks(discard_block_size, origin_size))
2429 discard_block_size *= 2;
2430
2431 return discard_block_size;
2432}
2433
2434static void set_cache_size(struct cache *cache, dm_cblock_t size)
2435{
2436 dm_block_t nr_blocks = from_cblock(size);
2437
2438 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2439 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2440 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2441 "Please consider increasing the cache block size to reduce the overall cache block count.",
2442 (unsigned long long) nr_blocks);
2443
2444 cache->cache_size = size;
2445}
2446
2447static int is_congested(struct dm_dev *dev, int bdi_bits)
2448{
2449 struct request_queue *q = bdev_get_queue(dev->bdev);
2450 return bdi_congested(q->backing_dev_info, bdi_bits);
2451}
2452
2453static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2454{
2455 struct cache *cache = container_of(cb, struct cache, callbacks);
2456
2457 return is_congested(cache->origin_dev, bdi_bits) ||
2458 is_congested(cache->cache_dev, bdi_bits);
2459}
2460
2461#define DEFAULT_MIGRATION_THRESHOLD 2048
2462
2463static int cache_create(struct cache_args *ca, struct cache **result)
2464{
2465 int r = 0;
2466 char **error = &ca->ti->error;
2467 struct cache *cache;
2468 struct dm_target *ti = ca->ti;
2469 dm_block_t origin_blocks;
2470 struct dm_cache_metadata *cmd;
2471 bool may_format = ca->features.mode == CM_WRITE;
2472
2473 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2474 if (!cache)
2475 return -ENOMEM;
2476
2477 cache->ti = ca->ti;
2478 ti->private = cache;
2479 ti->num_flush_bios = 2;
2480 ti->flush_supported = true;
2481
2482 ti->num_discard_bios = 1;
2483 ti->discards_supported = true;
2484
2485 ti->per_io_data_size = sizeof(struct per_bio_data);
2486
2487 cache->features = ca->features;
2488 if (writethrough_mode(cache)) {
2489 /* Create bioset for writethrough bios issued to origin */
2490 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2491 if (r)
2492 goto bad;
2493 }
2494
2495 cache->callbacks.congested_fn = cache_is_congested;
2496 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2497
2498 cache->metadata_dev = ca->metadata_dev;
2499 cache->origin_dev = ca->origin_dev;
2500 cache->cache_dev = ca->cache_dev;
2501
2502 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2503
2504 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2505 origin_blocks = block_div(origin_blocks, ca->block_size);
2506 cache->origin_blocks = to_oblock(origin_blocks);
2507
2508 cache->sectors_per_block = ca->block_size;
2509 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2510 r = -EINVAL;
2511 goto bad;
2512 }
2513
2514 if (ca->block_size & (ca->block_size - 1)) {
2515 dm_block_t cache_size = ca->cache_sectors;
2516
2517 cache->sectors_per_block_shift = -1;
2518 cache_size = block_div(cache_size, ca->block_size);
2519 set_cache_size(cache, to_cblock(cache_size));
2520 } else {
2521 cache->sectors_per_block_shift = __ffs(ca->block_size);
2522 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2523 }
2524
2525 r = create_cache_policy(cache, ca, error);
2526 if (r)
2527 goto bad;
2528
2529 cache->policy_nr_args = ca->policy_argc;
2530 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2531
2532 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2533 if (r) {
2534 *error = "Error setting cache policy's config values";
2535 goto bad;
2536 }
2537
2538 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2539 ca->block_size, may_format,
2540 dm_cache_policy_get_hint_size(cache->policy),
2541 ca->features.metadata_version);
2542 if (IS_ERR(cmd)) {
2543 *error = "Error creating metadata object";
2544 r = PTR_ERR(cmd);
2545 goto bad;
2546 }
2547 cache->cmd = cmd;
2548 set_cache_mode(cache, CM_WRITE);
2549 if (get_cache_mode(cache) != CM_WRITE) {
2550 *error = "Unable to get write access to metadata, please check/repair metadata.";
2551 r = -EINVAL;
2552 goto bad;
2553 }
2554
2555 if (passthrough_mode(cache)) {
2556 bool all_clean;
2557
2558 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2559 if (r) {
2560 *error = "dm_cache_metadata_all_clean() failed";
2561 goto bad;
2562 }
2563
2564 if (!all_clean) {
2565 *error = "Cannot enter passthrough mode unless all blocks are clean";
2566 r = -EINVAL;
2567 goto bad;
2568 }
2569
2570 policy_allow_migrations(cache->policy, false);
2571 }
2572
2573 spin_lock_init(&cache->lock);
2574 bio_list_init(&cache->deferred_bios);
2575 atomic_set(&cache->nr_allocated_migrations, 0);
2576 atomic_set(&cache->nr_io_migrations, 0);
2577 init_waitqueue_head(&cache->migration_wait);
2578
2579 r = -ENOMEM;
2580 atomic_set(&cache->nr_dirty, 0);
2581 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2582 if (!cache->dirty_bitset) {
2583 *error = "could not allocate dirty bitset";
2584 goto bad;
2585 }
2586 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2587
2588 cache->discard_block_size =
2589 calculate_discard_block_size(cache->sectors_per_block,
2590 cache->origin_sectors);
2591 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2592 cache->discard_block_size));
2593 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2594 if (!cache->discard_bitset) {
2595 *error = "could not allocate discard bitset";
2596 goto bad;
2597 }
2598 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2599
2600 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2601 if (IS_ERR(cache->copier)) {
2602 *error = "could not create kcopyd client";
2603 r = PTR_ERR(cache->copier);
2604 goto bad;
2605 }
2606
2607 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2608 if (!cache->wq) {
2609 *error = "could not create workqueue for metadata object";
2610 goto bad;
2611 }
2612 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2613 INIT_WORK(&cache->migration_worker, check_migrations);
2614 INIT_DELAYED_WORK(&cache->waker, do_waker);
2615
2616 cache->prison = dm_bio_prison_create_v2(cache->wq);
2617 if (!cache->prison) {
2618 *error = "could not create bio prison";
2619 goto bad;
2620 }
2621
2622 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2623 migration_cache);
2624 if (r) {
2625 *error = "Error creating cache's migration mempool";
2626 goto bad;
2627 }
2628
2629 cache->need_tick_bio = true;
2630 cache->sized = false;
2631 cache->invalidate = false;
2632 cache->commit_requested = false;
2633 cache->loaded_mappings = false;
2634 cache->loaded_discards = false;
2635
2636 load_stats(cache);
2637
2638 atomic_set(&cache->stats.demotion, 0);
2639 atomic_set(&cache->stats.promotion, 0);
2640 atomic_set(&cache->stats.copies_avoided, 0);
2641 atomic_set(&cache->stats.cache_cell_clash, 0);
2642 atomic_set(&cache->stats.commit_count, 0);
2643 atomic_set(&cache->stats.discard_count, 0);
2644
2645 spin_lock_init(&cache->invalidation_lock);
2646 INIT_LIST_HEAD(&cache->invalidation_requests);
2647
2648 batcher_init(&cache->committer, commit_op, cache,
2649 issue_op, cache, cache->wq);
2650 iot_init(&cache->tracker);
2651
2652 init_rwsem(&cache->background_work_lock);
2653 prevent_background_work(cache);
2654
2655 *result = cache;
2656 return 0;
2657bad:
2658 destroy(cache);
2659 return r;
2660}
2661
2662static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2663{
2664 unsigned i;
2665 const char **copy;
2666
2667 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2668 if (!copy)
2669 return -ENOMEM;
2670 for (i = 0; i < argc; i++) {
2671 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2672 if (!copy[i]) {
2673 while (i--)
2674 kfree(copy[i]);
2675 kfree(copy);
2676 return -ENOMEM;
2677 }
2678 }
2679
2680 cache->nr_ctr_args = argc;
2681 cache->ctr_args = copy;
2682
2683 return 0;
2684}
2685
2686static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2687{
2688 int r = -EINVAL;
2689 struct cache_args *ca;
2690 struct cache *cache = NULL;
2691
2692 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2693 if (!ca) {
2694 ti->error = "Error allocating memory for cache";
2695 return -ENOMEM;
2696 }
2697 ca->ti = ti;
2698
2699 r = parse_cache_args(ca, argc, argv, &ti->error);
2700 if (r)
2701 goto out;
2702
2703 r = cache_create(ca, &cache);
2704 if (r)
2705 goto out;
2706
2707 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2708 if (r) {
2709 destroy(cache);
2710 goto out;
2711 }
2712
2713 ti->private = cache;
2714out:
2715 destroy_cache_args(ca);
2716 return r;
2717}
2718
2719/*----------------------------------------------------------------*/
2720
2721static int cache_map(struct dm_target *ti, struct bio *bio)
2722{
2723 struct cache *cache = ti->private;
2724
2725 int r;
2726 bool commit_needed;
2727 dm_oblock_t block = get_bio_block(cache, bio);
2728
2729 init_per_bio_data(bio);
2730 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2731 /*
2732 * This can only occur if the io goes to a partial block at
2733 * the end of the origin device. We don't cache these.
2734 * Just remap to the origin and carry on.
2735 */
2736 remap_to_origin(cache, bio);
2737 accounted_begin(cache, bio);
2738 return DM_MAPIO_REMAPPED;
2739 }
2740
2741 if (discard_or_flush(bio)) {
2742 defer_bio(cache, bio);
2743 return DM_MAPIO_SUBMITTED;
2744 }
2745
2746 r = map_bio(cache, bio, block, &commit_needed);
2747 if (commit_needed)
2748 schedule_commit(&cache->committer);
2749
2750 return r;
2751}
2752
2753static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2754{
2755 struct cache *cache = ti->private;
2756 unsigned long flags;
2757 struct per_bio_data *pb = get_per_bio_data(bio);
2758
2759 if (pb->tick) {
2760 policy_tick(cache->policy, false);
2761
2762 spin_lock_irqsave(&cache->lock, flags);
2763 cache->need_tick_bio = true;
2764 spin_unlock_irqrestore(&cache->lock, flags);
2765 }
2766
2767 bio_drop_shared_lock(cache, bio);
2768 accounted_complete(cache, bio);
2769
2770 return DM_ENDIO_DONE;
2771}
2772
2773static int write_dirty_bitset(struct cache *cache)
2774{
2775 int r;
2776
2777 if (get_cache_mode(cache) >= CM_READ_ONLY)
2778 return -EINVAL;
2779
2780 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2781 if (r)
2782 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2783
2784 return r;
2785}
2786
2787static int write_discard_bitset(struct cache *cache)
2788{
2789 unsigned i, r;
2790
2791 if (get_cache_mode(cache) >= CM_READ_ONLY)
2792 return -EINVAL;
2793
2794 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2795 cache->discard_nr_blocks);
2796 if (r) {
2797 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2798 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2799 return r;
2800 }
2801
2802 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2803 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2804 is_discarded(cache, to_dblock(i)));
2805 if (r) {
2806 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2807 return r;
2808 }
2809 }
2810
2811 return 0;
2812}
2813
2814static int write_hints(struct cache *cache)
2815{
2816 int r;
2817
2818 if (get_cache_mode(cache) >= CM_READ_ONLY)
2819 return -EINVAL;
2820
2821 r = dm_cache_write_hints(cache->cmd, cache->policy);
2822 if (r) {
2823 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2824 return r;
2825 }
2826
2827 return 0;
2828}
2829
2830/*
2831 * returns true on success
2832 */
2833static bool sync_metadata(struct cache *cache)
2834{
2835 int r1, r2, r3, r4;
2836
2837 r1 = write_dirty_bitset(cache);
2838 if (r1)
2839 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2840
2841 r2 = write_discard_bitset(cache);
2842 if (r2)
2843 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2844
2845 save_stats(cache);
2846
2847 r3 = write_hints(cache);
2848 if (r3)
2849 DMERR("%s: could not write hints", cache_device_name(cache));
2850
2851 /*
2852 * If writing the above metadata failed, we still commit, but don't
2853 * set the clean shutdown flag. This will effectively force every
2854 * dirty bit to be set on reload.
2855 */
2856 r4 = commit(cache, !r1 && !r2 && !r3);
2857 if (r4)
2858 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2859
2860 return !r1 && !r2 && !r3 && !r4;
2861}
2862
2863static void cache_postsuspend(struct dm_target *ti)
2864{
2865 struct cache *cache = ti->private;
2866
2867 prevent_background_work(cache);
2868 BUG_ON(atomic_read(&cache->nr_io_migrations));
2869
2870 cancel_delayed_work(&cache->waker);
2871 flush_workqueue(cache->wq);
2872 WARN_ON(cache->tracker.in_flight);
2873
2874 /*
2875 * If it's a flush suspend there won't be any deferred bios, so this
2876 * call is harmless.
2877 */
2878 requeue_deferred_bios(cache);
2879
2880 if (get_cache_mode(cache) == CM_WRITE)
2881 (void) sync_metadata(cache);
2882}
2883
2884static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2885 bool dirty, uint32_t hint, bool hint_valid)
2886{
2887 int r;
2888 struct cache *cache = context;
2889
2890 if (dirty) {
2891 set_bit(from_cblock(cblock), cache->dirty_bitset);
2892 atomic_inc(&cache->nr_dirty);
2893 } else
2894 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2895
2896 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2897 if (r)
2898 return r;
2899
2900 return 0;
2901}
2902
2903/*
2904 * The discard block size in the on disk metadata is not
2905 * neccessarily the same as we're currently using. So we have to
2906 * be careful to only set the discarded attribute if we know it
2907 * covers a complete block of the new size.
2908 */
2909struct discard_load_info {
2910 struct cache *cache;
2911
2912 /*
2913 * These blocks are sized using the on disk dblock size, rather
2914 * than the current one.
2915 */
2916 dm_block_t block_size;
2917 dm_block_t discard_begin, discard_end;
2918};
2919
2920static void discard_load_info_init(struct cache *cache,
2921 struct discard_load_info *li)
2922{
2923 li->cache = cache;
2924 li->discard_begin = li->discard_end = 0;
2925}
2926
2927static void set_discard_range(struct discard_load_info *li)
2928{
2929 sector_t b, e;
2930
2931 if (li->discard_begin == li->discard_end)
2932 return;
2933
2934 /*
2935 * Convert to sectors.
2936 */
2937 b = li->discard_begin * li->block_size;
2938 e = li->discard_end * li->block_size;
2939
2940 /*
2941 * Then convert back to the current dblock size.
2942 */
2943 b = dm_sector_div_up(b, li->cache->discard_block_size);
2944 sector_div(e, li->cache->discard_block_size);
2945
2946 /*
2947 * The origin may have shrunk, so we need to check we're still in
2948 * bounds.
2949 */
2950 if (e > from_dblock(li->cache->discard_nr_blocks))
2951 e = from_dblock(li->cache->discard_nr_blocks);
2952
2953 for (; b < e; b++)
2954 set_discard(li->cache, to_dblock(b));
2955}
2956
2957static int load_discard(void *context, sector_t discard_block_size,
2958 dm_dblock_t dblock, bool discard)
2959{
2960 struct discard_load_info *li = context;
2961
2962 li->block_size = discard_block_size;
2963
2964 if (discard) {
2965 if (from_dblock(dblock) == li->discard_end)
2966 /*
2967 * We're already in a discard range, just extend it.
2968 */
2969 li->discard_end = li->discard_end + 1ULL;
2970
2971 else {
2972 /*
2973 * Emit the old range and start a new one.
2974 */
2975 set_discard_range(li);
2976 li->discard_begin = from_dblock(dblock);
2977 li->discard_end = li->discard_begin + 1ULL;
2978 }
2979 } else {
2980 set_discard_range(li);
2981 li->discard_begin = li->discard_end = 0;
2982 }
2983
2984 return 0;
2985}
2986
2987static dm_cblock_t get_cache_dev_size(struct cache *cache)
2988{
2989 sector_t size = get_dev_size(cache->cache_dev);
2990 (void) sector_div(size, cache->sectors_per_block);
2991 return to_cblock(size);
2992}
2993
2994static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2995{
2996 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2997 if (cache->sized) {
2998 DMERR("%s: unable to extend cache due to missing cache table reload",
2999 cache_device_name(cache));
3000 return false;
3001 }
3002 }
3003
3004 /*
3005 * We can't drop a dirty block when shrinking the cache.
3006 */
3007 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3008 new_size = to_cblock(from_cblock(new_size) + 1);
3009 if (is_dirty(cache, new_size)) {
3010 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3011 cache_device_name(cache),
3012 (unsigned long long) from_cblock(new_size));
3013 return false;
3014 }
3015 }
3016
3017 return true;
3018}
3019
3020static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3021{
3022 int r;
3023
3024 r = dm_cache_resize(cache->cmd, new_size);
3025 if (r) {
3026 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3027 metadata_operation_failed(cache, "dm_cache_resize", r);
3028 return r;
3029 }
3030
3031 set_cache_size(cache, new_size);
3032
3033 return 0;
3034}
3035
3036static int cache_preresume(struct dm_target *ti)
3037{
3038 int r = 0;
3039 struct cache *cache = ti->private;
3040 dm_cblock_t csize = get_cache_dev_size(cache);
3041
3042 /*
3043 * Check to see if the cache has resized.
3044 */
3045 if (!cache->sized) {
3046 r = resize_cache_dev(cache, csize);
3047 if (r)
3048 return r;
3049
3050 cache->sized = true;
3051
3052 } else if (csize != cache->cache_size) {
3053 if (!can_resize(cache, csize))
3054 return -EINVAL;
3055
3056 r = resize_cache_dev(cache, csize);
3057 if (r)
3058 return r;
3059 }
3060
3061 if (!cache->loaded_mappings) {
3062 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3063 load_mapping, cache);
3064 if (r) {
3065 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3066 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3067 return r;
3068 }
3069
3070 cache->loaded_mappings = true;
3071 }
3072
3073 if (!cache->loaded_discards) {
3074 struct discard_load_info li;
3075
3076 /*
3077 * The discard bitset could have been resized, or the
3078 * discard block size changed. To be safe we start by
3079 * setting every dblock to not discarded.
3080 */
3081 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3082
3083 discard_load_info_init(cache, &li);
3084 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3085 if (r) {
3086 DMERR("%s: could not load origin discards", cache_device_name(cache));
3087 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3088 return r;
3089 }
3090 set_discard_range(&li);
3091
3092 cache->loaded_discards = true;
3093 }
3094
3095 return r;
3096}
3097
3098static void cache_resume(struct dm_target *ti)
3099{
3100 struct cache *cache = ti->private;
3101
3102 cache->need_tick_bio = true;
3103 allow_background_work(cache);
3104 do_waker(&cache->waker.work);
3105}
3106
3107static void emit_flags(struct cache *cache, char *result,
3108 unsigned maxlen, ssize_t *sz_ptr)
3109{
3110 ssize_t sz = *sz_ptr;
3111 struct cache_features *cf = &cache->features;
3112 unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3113
3114 DMEMIT("%u ", count);
3115
3116 if (cf->metadata_version == 2)
3117 DMEMIT("metadata2 ");
3118
3119 if (writethrough_mode(cache))
3120 DMEMIT("writethrough ");
3121
3122 else if (passthrough_mode(cache))
3123 DMEMIT("passthrough ");
3124
3125 else if (writeback_mode(cache))
3126 DMEMIT("writeback ");
3127
3128 else {
3129 DMEMIT("unknown ");
3130 DMERR("%s: internal error: unknown io mode: %d",
3131 cache_device_name(cache), (int) cf->io_mode);
3132 }
3133
3134 if (!cf->discard_passdown)
3135 DMEMIT("no_discard_passdown ");
3136
3137 *sz_ptr = sz;
3138}
3139
3140/*
3141 * Status format:
3142 *
3143 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3144 * <cache block size> <#used cache blocks>/<#total cache blocks>
3145 * <#read hits> <#read misses> <#write hits> <#write misses>
3146 * <#demotions> <#promotions> <#dirty>
3147 * <#features> <features>*
3148 * <#core args> <core args>
3149 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3150 */
3151static void cache_status(struct dm_target *ti, status_type_t type,
3152 unsigned status_flags, char *result, unsigned maxlen)
3153{
3154 int r = 0;
3155 unsigned i;
3156 ssize_t sz = 0;
3157 dm_block_t nr_free_blocks_metadata = 0;
3158 dm_block_t nr_blocks_metadata = 0;
3159 char buf[BDEVNAME_SIZE];
3160 struct cache *cache = ti->private;
3161 dm_cblock_t residency;
3162 bool needs_check;
3163
3164 switch (type) {
3165 case STATUSTYPE_INFO:
3166 if (get_cache_mode(cache) == CM_FAIL) {
3167 DMEMIT("Fail");
3168 break;
3169 }
3170
3171 /* Commit to ensure statistics aren't out-of-date */
3172 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3173 (void) commit(cache, false);
3174
3175 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3176 if (r) {
3177 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3178 cache_device_name(cache), r);
3179 goto err;
3180 }
3181
3182 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3183 if (r) {
3184 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3185 cache_device_name(cache), r);
3186 goto err;
3187 }
3188
3189 residency = policy_residency(cache->policy);
3190
3191 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3192 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3193 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3194 (unsigned long long)nr_blocks_metadata,
3195 (unsigned long long)cache->sectors_per_block,
3196 (unsigned long long) from_cblock(residency),
3197 (unsigned long long) from_cblock(cache->cache_size),
3198 (unsigned) atomic_read(&cache->stats.read_hit),
3199 (unsigned) atomic_read(&cache->stats.read_miss),
3200 (unsigned) atomic_read(&cache->stats.write_hit),
3201 (unsigned) atomic_read(&cache->stats.write_miss),
3202 (unsigned) atomic_read(&cache->stats.demotion),
3203 (unsigned) atomic_read(&cache->stats.promotion),
3204 (unsigned long) atomic_read(&cache->nr_dirty));
3205
3206 emit_flags(cache, result, maxlen, &sz);
3207
3208 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3209
3210 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3211 if (sz < maxlen) {
3212 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3213 if (r)
3214 DMERR("%s: policy_emit_config_values returned %d",
3215 cache_device_name(cache), r);
3216 }
3217
3218 if (get_cache_mode(cache) == CM_READ_ONLY)
3219 DMEMIT("ro ");
3220 else
3221 DMEMIT("rw ");
3222
3223 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3224
3225 if (r || needs_check)
3226 DMEMIT("needs_check ");
3227 else
3228 DMEMIT("- ");
3229
3230 break;
3231
3232 case STATUSTYPE_TABLE:
3233 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3234 DMEMIT("%s ", buf);
3235 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3236 DMEMIT("%s ", buf);
3237 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3238 DMEMIT("%s", buf);
3239
3240 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3241 DMEMIT(" %s", cache->ctr_args[i]);
3242 if (cache->nr_ctr_args)
3243 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3244 }
3245
3246 return;
3247
3248err:
3249 DMEMIT("Error");
3250}
3251
3252/*
3253 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3254 * the one-past-the-end value.
3255 */
3256struct cblock_range {
3257 dm_cblock_t begin;
3258 dm_cblock_t end;
3259};
3260
3261/*
3262 * A cache block range can take two forms:
3263 *
3264 * i) A single cblock, eg. '3456'
3265 * ii) A begin and end cblock with a dash between, eg. 123-234
3266 */
3267static int parse_cblock_range(struct cache *cache, const char *str,
3268 struct cblock_range *result)
3269{
3270 char dummy;
3271 uint64_t b, e;
3272 int r;
3273
3274 /*
3275 * Try and parse form (ii) first.
3276 */
3277 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3278 if (r < 0)
3279 return r;
3280
3281 if (r == 2) {
3282 result->begin = to_cblock(b);
3283 result->end = to_cblock(e);
3284 return 0;
3285 }
3286
3287 /*
3288 * That didn't work, try form (i).
3289 */
3290 r = sscanf(str, "%llu%c", &b, &dummy);
3291 if (r < 0)
3292 return r;
3293
3294 if (r == 1) {
3295 result->begin = to_cblock(b);
3296 result->end = to_cblock(from_cblock(result->begin) + 1u);
3297 return 0;
3298 }
3299
3300 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3301 return -EINVAL;
3302}
3303
3304static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3305{
3306 uint64_t b = from_cblock(range->begin);
3307 uint64_t e = from_cblock(range->end);
3308 uint64_t n = from_cblock(cache->cache_size);
3309
3310 if (b >= n) {
3311 DMERR("%s: begin cblock out of range: %llu >= %llu",
3312 cache_device_name(cache), b, n);
3313 return -EINVAL;
3314 }
3315
3316 if (e > n) {
3317 DMERR("%s: end cblock out of range: %llu > %llu",
3318 cache_device_name(cache), e, n);
3319 return -EINVAL;
3320 }
3321
3322 if (b >= e) {
3323 DMERR("%s: invalid cblock range: %llu >= %llu",
3324 cache_device_name(cache), b, e);
3325 return -EINVAL;
3326 }
3327
3328 return 0;
3329}
3330
3331static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3332{
3333 return to_cblock(from_cblock(b) + 1);
3334}
3335
3336static int request_invalidation(struct cache *cache, struct cblock_range *range)
3337{
3338 int r = 0;
3339
3340 /*
3341 * We don't need to do any locking here because we know we're in
3342 * passthrough mode. There's is potential for a race between an
3343 * invalidation triggered by an io and an invalidation message. This
3344 * is harmless, we must not worry if the policy call fails.
3345 */
3346 while (range->begin != range->end) {
3347 r = invalidate_cblock(cache, range->begin);
3348 if (r)
3349 return r;
3350
3351 range->begin = cblock_succ(range->begin);
3352 }
3353
3354 cache->commit_requested = true;
3355 return r;
3356}
3357
3358static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3359 const char **cblock_ranges)
3360{
3361 int r = 0;
3362 unsigned i;
3363 struct cblock_range range;
3364
3365 if (!passthrough_mode(cache)) {
3366 DMERR("%s: cache has to be in passthrough mode for invalidation",
3367 cache_device_name(cache));
3368 return -EPERM;
3369 }
3370
3371 for (i = 0; i < count; i++) {
3372 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3373 if (r)
3374 break;
3375
3376 r = validate_cblock_range(cache, &range);
3377 if (r)
3378 break;
3379
3380 /*
3381 * Pass begin and end origin blocks to the worker and wake it.
3382 */
3383 r = request_invalidation(cache, &range);
3384 if (r)
3385 break;
3386 }
3387
3388 return r;
3389}
3390
3391/*
3392 * Supports
3393 * "<key> <value>"
3394 * and
3395 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3396 *
3397 * The key migration_threshold is supported by the cache target core.
3398 */
3399static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3400 char *result, unsigned maxlen)
3401{
3402 struct cache *cache = ti->private;
3403
3404 if (!argc)
3405 return -EINVAL;
3406
3407 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3408 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3409 cache_device_name(cache));
3410 return -EOPNOTSUPP;
3411 }
3412
3413 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3414 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3415
3416 if (argc != 2)
3417 return -EINVAL;
3418
3419 return set_config_value(cache, argv[0], argv[1]);
3420}
3421
3422static int cache_iterate_devices(struct dm_target *ti,
3423 iterate_devices_callout_fn fn, void *data)
3424{
3425 int r = 0;
3426 struct cache *cache = ti->private;
3427
3428 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3429 if (!r)
3430 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3431
3432 return r;
3433}
3434
3435static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3436{
3437 struct request_queue *q = bdev_get_queue(origin_bdev);
3438
3439 return q && blk_queue_discard(q);
3440}
3441
3442/*
3443 * If discard_passdown was enabled verify that the origin device
3444 * supports discards. Disable discard_passdown if not.
3445 */
3446static void disable_passdown_if_not_supported(struct cache *cache)
3447{
3448 struct block_device *origin_bdev = cache->origin_dev->bdev;
3449 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3450 const char *reason = NULL;
3451 char buf[BDEVNAME_SIZE];
3452
3453 if (!cache->features.discard_passdown)
3454 return;
3455
3456 if (!origin_dev_supports_discard(origin_bdev))
3457 reason = "discard unsupported";
3458
3459 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3460 reason = "max discard sectors smaller than a block";
3461
3462 if (reason) {
3463 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3464 bdevname(origin_bdev, buf), reason);
3465 cache->features.discard_passdown = false;
3466 }
3467}
3468
3469static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3470{
3471 struct block_device *origin_bdev = cache->origin_dev->bdev;
3472 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3473
3474 if (!cache->features.discard_passdown) {
3475 /* No passdown is done so setting own virtual limits */
3476 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3477 cache->origin_sectors);
3478 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3479 return;
3480 }
3481
3482 /*
3483 * cache_iterate_devices() is stacking both origin and fast device limits
3484 * but discards aren't passed to fast device, so inherit origin's limits.
3485 */
3486 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3487 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3488 limits->discard_granularity = origin_limits->discard_granularity;
3489 limits->discard_alignment = origin_limits->discard_alignment;
3490 limits->discard_misaligned = origin_limits->discard_misaligned;
3491}
3492
3493static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3494{
3495 struct cache *cache = ti->private;
3496 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3497
3498 /*
3499 * If the system-determined stacked limits are compatible with the
3500 * cache's blocksize (io_opt is a factor) do not override them.
3501 */
3502 if (io_opt_sectors < cache->sectors_per_block ||
3503 do_div(io_opt_sectors, cache->sectors_per_block)) {
3504 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3505 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3506 }
3507
3508 disable_passdown_if_not_supported(cache);
3509 set_discard_limits(cache, limits);
3510}
3511
3512/*----------------------------------------------------------------*/
3513
3514static struct target_type cache_target = {
3515 .name = "cache",
3516 .version = {2, 1, 0},
3517 .module = THIS_MODULE,
3518 .ctr = cache_ctr,
3519 .dtr = cache_dtr,
3520 .map = cache_map,
3521 .end_io = cache_end_io,
3522 .postsuspend = cache_postsuspend,
3523 .preresume = cache_preresume,
3524 .resume = cache_resume,
3525 .status = cache_status,
3526 .message = cache_message,
3527 .iterate_devices = cache_iterate_devices,
3528 .io_hints = cache_io_hints,
3529};
3530
3531static int __init dm_cache_init(void)
3532{
3533 int r;
3534
3535 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3536 if (!migration_cache)
3537 return -ENOMEM;
3538
3539 r = dm_register_target(&cache_target);
3540 if (r) {
3541 DMERR("cache target registration failed: %d", r);
3542 kmem_cache_destroy(migration_cache);
3543 return r;
3544 }
3545
3546 return 0;
3547}
3548
3549static void __exit dm_cache_exit(void)
3550{
3551 dm_unregister_target(&cache_target);
3552 kmem_cache_destroy(migration_cache);
3553}
3554
3555module_init(dm_cache_init);
3556module_exit(dm_cache_exit);
3557
3558MODULE_DESCRIPTION(DM_NAME " cache target");
3559MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3560MODULE_LICENSE("GPL");