xref: /linux/drivers/md/dm-cache-target.c (revision b68fc09be48edbc47de1a0f3d42ef8adf6c0ac55)
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 
24 DECLARE_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 
42 struct 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 
58 static 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 
66 static 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 
74 static 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 
86 static 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 
95 static 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 
105 static 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  */
120 struct continuation {
121 	struct work_struct ws;
122 	blk_status_t input;
123 };
124 
125 static 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 
132 static 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  */
144 struct 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 
171 static 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 
214 static 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 
234 static void async_commit(struct batcher *b)
235 {
236 	queue_work(b->wq, &b->commit_work);
237 }
238 
239 static 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  */
256 static 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  */
273 static 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  */
292 struct dm_hook_info {
293 	bio_end_io_t *bi_end_io;
294 };
295 
296 static 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 
305 static 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 
323 enum cache_metadata_mode {
324 	CM_WRITE,		/* metadata may be changed */
325 	CM_READ_ONLY,		/* metadata may not be changed */
326 	CM_FAIL
327 };
328 
329 enum 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 
352 struct cache_features {
353 	enum cache_metadata_mode mode;
354 	enum cache_io_mode io_mode;
355 	unsigned metadata_version;
356 };
357 
358 struct cache_stats {
359 	atomic_t read_hit;
360 	atomic_t read_miss;
361 	atomic_t write_hit;
362 	atomic_t write_miss;
363 	atomic_t demotion;
364 	atomic_t promotion;
365 	atomic_t writeback;
366 	atomic_t copies_avoided;
367 	atomic_t cache_cell_clash;
368 	atomic_t commit_count;
369 	atomic_t discard_count;
370 };
371 
372 struct cache {
373 	struct dm_target *ti;
374 	spinlock_t lock;
375 
376 	/*
377 	 * Fields for converting from sectors to blocks.
378 	 */
379 	int sectors_per_block_shift;
380 	sector_t sectors_per_block;
381 
382 	struct dm_cache_metadata *cmd;
383 
384 	/*
385 	 * Metadata is written to this device.
386 	 */
387 	struct dm_dev *metadata_dev;
388 
389 	/*
390 	 * The slower of the two data devices.  Typically a spindle.
391 	 */
392 	struct dm_dev *origin_dev;
393 
394 	/*
395 	 * The faster of the two data devices.  Typically an SSD.
396 	 */
397 	struct dm_dev *cache_dev;
398 
399 	/*
400 	 * Size of the origin device in _complete_ blocks and native sectors.
401 	 */
402 	dm_oblock_t origin_blocks;
403 	sector_t origin_sectors;
404 
405 	/*
406 	 * Size of the cache device in blocks.
407 	 */
408 	dm_cblock_t cache_size;
409 
410 	/*
411 	 * Invalidation fields.
412 	 */
413 	spinlock_t invalidation_lock;
414 	struct list_head invalidation_requests;
415 
416 	sector_t migration_threshold;
417 	wait_queue_head_t migration_wait;
418 	atomic_t nr_allocated_migrations;
419 
420 	/*
421 	 * The number of in flight migrations that are performing
422 	 * background io. eg, promotion, writeback.
423 	 */
424 	atomic_t nr_io_migrations;
425 
426 	struct bio_list deferred_bios;
427 
428 	struct rw_semaphore quiesce_lock;
429 
430 	struct dm_target_callbacks callbacks;
431 
432 	/*
433 	 * origin_blocks entries, discarded if set.
434 	 */
435 	dm_dblock_t discard_nr_blocks;
436 	unsigned long *discard_bitset;
437 	uint32_t discard_block_size; /* a power of 2 times sectors per block */
438 
439 	/*
440 	 * Rather than reconstructing the table line for the status we just
441 	 * save it and regurgitate.
442 	 */
443 	unsigned nr_ctr_args;
444 	const char **ctr_args;
445 
446 	struct dm_kcopyd_client *copier;
447 	struct work_struct deferred_bio_worker;
448 	struct work_struct migration_worker;
449 	struct workqueue_struct *wq;
450 	struct delayed_work waker;
451 	struct dm_bio_prison_v2 *prison;
452 
453 	/*
454 	 * cache_size entries, dirty if set
455 	 */
456 	unsigned long *dirty_bitset;
457 	atomic_t nr_dirty;
458 
459 	unsigned policy_nr_args;
460 	struct dm_cache_policy *policy;
461 
462 	/*
463 	 * Cache features such as write-through.
464 	 */
465 	struct cache_features features;
466 
467 	struct cache_stats stats;
468 
469 	bool need_tick_bio:1;
470 	bool sized:1;
471 	bool invalidate:1;
472 	bool commit_requested:1;
473 	bool loaded_mappings:1;
474 	bool loaded_discards:1;
475 
476 	struct rw_semaphore background_work_lock;
477 
478 	struct batcher committer;
479 	struct work_struct commit_ws;
480 
481 	struct io_tracker tracker;
482 
483 	mempool_t migration_pool;
484 
485 	struct bio_set bs;
486 };
487 
488 struct per_bio_data {
489 	bool tick:1;
490 	unsigned req_nr:2;
491 	struct dm_bio_prison_cell_v2 *cell;
492 	struct dm_hook_info hook_info;
493 	sector_t len;
494 };
495 
496 struct dm_cache_migration {
497 	struct continuation k;
498 	struct cache *cache;
499 
500 	struct policy_work *op;
501 	struct bio *overwrite_bio;
502 	struct dm_bio_prison_cell_v2 *cell;
503 
504 	dm_cblock_t invalidate_cblock;
505 	dm_oblock_t invalidate_oblock;
506 };
507 
508 /*----------------------------------------------------------------*/
509 
510 static bool writethrough_mode(struct cache *cache)
511 {
512 	return cache->features.io_mode == CM_IO_WRITETHROUGH;
513 }
514 
515 static bool writeback_mode(struct cache *cache)
516 {
517 	return cache->features.io_mode == CM_IO_WRITEBACK;
518 }
519 
520 static inline bool passthrough_mode(struct cache *cache)
521 {
522 	return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
523 }
524 
525 /*----------------------------------------------------------------*/
526 
527 static void wake_deferred_bio_worker(struct cache *cache)
528 {
529 	queue_work(cache->wq, &cache->deferred_bio_worker);
530 }
531 
532 static void wake_migration_worker(struct cache *cache)
533 {
534 	if (passthrough_mode(cache))
535 		return;
536 
537 	queue_work(cache->wq, &cache->migration_worker);
538 }
539 
540 /*----------------------------------------------------------------*/
541 
542 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
543 {
544 	return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
545 }
546 
547 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
548 {
549 	dm_bio_prison_free_cell_v2(cache->prison, cell);
550 }
551 
552 static struct dm_cache_migration *alloc_migration(struct cache *cache)
553 {
554 	struct dm_cache_migration *mg;
555 
556 	mg = mempool_alloc(&cache->migration_pool, GFP_NOWAIT);
557 	if (!mg)
558 		return NULL;
559 
560 	memset(mg, 0, sizeof(*mg));
561 
562 	mg->cache = cache;
563 	atomic_inc(&cache->nr_allocated_migrations);
564 
565 	return mg;
566 }
567 
568 static void free_migration(struct dm_cache_migration *mg)
569 {
570 	struct cache *cache = mg->cache;
571 
572 	if (atomic_dec_and_test(&cache->nr_allocated_migrations))
573 		wake_up(&cache->migration_wait);
574 
575 	mempool_free(mg, &cache->migration_pool);
576 }
577 
578 /*----------------------------------------------------------------*/
579 
580 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
581 {
582 	return to_oblock(from_oblock(b) + 1ull);
583 }
584 
585 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
586 {
587 	key->virtual = 0;
588 	key->dev = 0;
589 	key->block_begin = from_oblock(begin);
590 	key->block_end = from_oblock(end);
591 }
592 
593 /*
594  * We have two lock levels.  Level 0, which is used to prevent WRITEs, and
595  * level 1 which prevents *both* READs and WRITEs.
596  */
597 #define WRITE_LOCK_LEVEL 0
598 #define READ_WRITE_LOCK_LEVEL 1
599 
600 static unsigned lock_level(struct bio *bio)
601 {
602 	return bio_data_dir(bio) == WRITE ?
603 		WRITE_LOCK_LEVEL :
604 		READ_WRITE_LOCK_LEVEL;
605 }
606 
607 /*----------------------------------------------------------------
608  * Per bio data
609  *--------------------------------------------------------------*/
610 
611 static struct per_bio_data *get_per_bio_data(struct bio *bio)
612 {
613 	struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
614 	BUG_ON(!pb);
615 	return pb;
616 }
617 
618 static struct per_bio_data *init_per_bio_data(struct bio *bio)
619 {
620 	struct per_bio_data *pb = get_per_bio_data(bio);
621 
622 	pb->tick = false;
623 	pb->req_nr = dm_bio_get_target_bio_nr(bio);
624 	pb->cell = NULL;
625 	pb->len = 0;
626 
627 	return pb;
628 }
629 
630 /*----------------------------------------------------------------*/
631 
632 static void defer_bio(struct cache *cache, struct bio *bio)
633 {
634 	unsigned long flags;
635 
636 	spin_lock_irqsave(&cache->lock, flags);
637 	bio_list_add(&cache->deferred_bios, bio);
638 	spin_unlock_irqrestore(&cache->lock, flags);
639 
640 	wake_deferred_bio_worker(cache);
641 }
642 
643 static void defer_bios(struct cache *cache, struct bio_list *bios)
644 {
645 	unsigned long flags;
646 
647 	spin_lock_irqsave(&cache->lock, flags);
648 	bio_list_merge(&cache->deferred_bios, bios);
649 	bio_list_init(bios);
650 	spin_unlock_irqrestore(&cache->lock, flags);
651 
652 	wake_deferred_bio_worker(cache);
653 }
654 
655 /*----------------------------------------------------------------*/
656 
657 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
658 {
659 	bool r;
660 	struct per_bio_data *pb;
661 	struct dm_cell_key_v2 key;
662 	dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
663 	struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
664 
665 	cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
666 	if (!cell_prealloc) {
667 		defer_bio(cache, bio);
668 		return false;
669 	}
670 
671 	build_key(oblock, end, &key);
672 	r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
673 	if (!r) {
674 		/*
675 		 * Failed to get the lock.
676 		 */
677 		free_prison_cell(cache, cell_prealloc);
678 		return r;
679 	}
680 
681 	if (cell != cell_prealloc)
682 		free_prison_cell(cache, cell_prealloc);
683 
684 	pb = get_per_bio_data(bio);
685 	pb->cell = cell;
686 
687 	return r;
688 }
689 
690 /*----------------------------------------------------------------*/
691 
692 static bool is_dirty(struct cache *cache, dm_cblock_t b)
693 {
694 	return test_bit(from_cblock(b), cache->dirty_bitset);
695 }
696 
697 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
698 {
699 	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
700 		atomic_inc(&cache->nr_dirty);
701 		policy_set_dirty(cache->policy, cblock);
702 	}
703 }
704 
705 /*
706  * These two are called when setting after migrations to force the policy
707  * and dirty bitset to be in sync.
708  */
709 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
710 {
711 	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
712 		atomic_inc(&cache->nr_dirty);
713 	policy_set_dirty(cache->policy, cblock);
714 }
715 
716 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
717 {
718 	if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
719 		if (atomic_dec_return(&cache->nr_dirty) == 0)
720 			dm_table_event(cache->ti->table);
721 	}
722 
723 	policy_clear_dirty(cache->policy, cblock);
724 }
725 
726 /*----------------------------------------------------------------*/
727 
728 static bool block_size_is_power_of_two(struct cache *cache)
729 {
730 	return cache->sectors_per_block_shift >= 0;
731 }
732 
733 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
734 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
735 __always_inline
736 #endif
737 static dm_block_t block_div(dm_block_t b, uint32_t n)
738 {
739 	do_div(b, n);
740 
741 	return b;
742 }
743 
744 static dm_block_t oblocks_per_dblock(struct cache *cache)
745 {
746 	dm_block_t oblocks = cache->discard_block_size;
747 
748 	if (block_size_is_power_of_two(cache))
749 		oblocks >>= cache->sectors_per_block_shift;
750 	else
751 		oblocks = block_div(oblocks, cache->sectors_per_block);
752 
753 	return oblocks;
754 }
755 
756 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
757 {
758 	return to_dblock(block_div(from_oblock(oblock),
759 				   oblocks_per_dblock(cache)));
760 }
761 
762 static void set_discard(struct cache *cache, dm_dblock_t b)
763 {
764 	unsigned long flags;
765 
766 	BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
767 	atomic_inc(&cache->stats.discard_count);
768 
769 	spin_lock_irqsave(&cache->lock, flags);
770 	set_bit(from_dblock(b), cache->discard_bitset);
771 	spin_unlock_irqrestore(&cache->lock, flags);
772 }
773 
774 static void clear_discard(struct cache *cache, dm_dblock_t b)
775 {
776 	unsigned long flags;
777 
778 	spin_lock_irqsave(&cache->lock, flags);
779 	clear_bit(from_dblock(b), cache->discard_bitset);
780 	spin_unlock_irqrestore(&cache->lock, flags);
781 }
782 
783 static bool is_discarded(struct cache *cache, dm_dblock_t b)
784 {
785 	int r;
786 	unsigned long flags;
787 
788 	spin_lock_irqsave(&cache->lock, flags);
789 	r = test_bit(from_dblock(b), cache->discard_bitset);
790 	spin_unlock_irqrestore(&cache->lock, flags);
791 
792 	return r;
793 }
794 
795 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
796 {
797 	int r;
798 	unsigned long flags;
799 
800 	spin_lock_irqsave(&cache->lock, flags);
801 	r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
802 		     cache->discard_bitset);
803 	spin_unlock_irqrestore(&cache->lock, flags);
804 
805 	return r;
806 }
807 
808 /*----------------------------------------------------------------
809  * Remapping
810  *--------------------------------------------------------------*/
811 static void remap_to_origin(struct cache *cache, struct bio *bio)
812 {
813 	bio_set_dev(bio, cache->origin_dev->bdev);
814 }
815 
816 static void remap_to_cache(struct cache *cache, struct bio *bio,
817 			   dm_cblock_t cblock)
818 {
819 	sector_t bi_sector = bio->bi_iter.bi_sector;
820 	sector_t block = from_cblock(cblock);
821 
822 	bio_set_dev(bio, cache->cache_dev->bdev);
823 	if (!block_size_is_power_of_two(cache))
824 		bio->bi_iter.bi_sector =
825 			(block * cache->sectors_per_block) +
826 			sector_div(bi_sector, cache->sectors_per_block);
827 	else
828 		bio->bi_iter.bi_sector =
829 			(block << cache->sectors_per_block_shift) |
830 			(bi_sector & (cache->sectors_per_block - 1));
831 }
832 
833 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
834 {
835 	unsigned long flags;
836 	struct per_bio_data *pb;
837 
838 	spin_lock_irqsave(&cache->lock, flags);
839 	if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
840 	    bio_op(bio) != REQ_OP_DISCARD) {
841 		pb = get_per_bio_data(bio);
842 		pb->tick = true;
843 		cache->need_tick_bio = false;
844 	}
845 	spin_unlock_irqrestore(&cache->lock, flags);
846 }
847 
848 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
849 					    dm_oblock_t oblock, bool bio_has_pbd)
850 {
851 	if (bio_has_pbd)
852 		check_if_tick_bio_needed(cache, bio);
853 	remap_to_origin(cache, bio);
854 	if (bio_data_dir(bio) == WRITE)
855 		clear_discard(cache, oblock_to_dblock(cache, oblock));
856 }
857 
858 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
859 					  dm_oblock_t oblock)
860 {
861 	// FIXME: check_if_tick_bio_needed() is called way too much through this interface
862 	__remap_to_origin_clear_discard(cache, bio, oblock, true);
863 }
864 
865 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
866 				 dm_oblock_t oblock, dm_cblock_t cblock)
867 {
868 	check_if_tick_bio_needed(cache, bio);
869 	remap_to_cache(cache, bio, cblock);
870 	if (bio_data_dir(bio) == WRITE) {
871 		set_dirty(cache, cblock);
872 		clear_discard(cache, oblock_to_dblock(cache, oblock));
873 	}
874 }
875 
876 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
877 {
878 	sector_t block_nr = bio->bi_iter.bi_sector;
879 
880 	if (!block_size_is_power_of_two(cache))
881 		(void) sector_div(block_nr, cache->sectors_per_block);
882 	else
883 		block_nr >>= cache->sectors_per_block_shift;
884 
885 	return to_oblock(block_nr);
886 }
887 
888 static bool accountable_bio(struct cache *cache, struct bio *bio)
889 {
890 	return bio_op(bio) != REQ_OP_DISCARD;
891 }
892 
893 static void accounted_begin(struct cache *cache, struct bio *bio)
894 {
895 	struct per_bio_data *pb;
896 
897 	if (accountable_bio(cache, bio)) {
898 		pb = get_per_bio_data(bio);
899 		pb->len = bio_sectors(bio);
900 		iot_io_begin(&cache->tracker, pb->len);
901 	}
902 }
903 
904 static void accounted_complete(struct cache *cache, struct bio *bio)
905 {
906 	struct per_bio_data *pb = get_per_bio_data(bio);
907 
908 	iot_io_end(&cache->tracker, pb->len);
909 }
910 
911 static void accounted_request(struct cache *cache, struct bio *bio)
912 {
913 	accounted_begin(cache, bio);
914 	generic_make_request(bio);
915 }
916 
917 static void issue_op(struct bio *bio, void *context)
918 {
919 	struct cache *cache = context;
920 	accounted_request(cache, bio);
921 }
922 
923 /*
924  * When running in writethrough mode we need to send writes to clean blocks
925  * to both the cache and origin devices.  Clone the bio and send them in parallel.
926  */
927 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
928 				      dm_oblock_t oblock, dm_cblock_t cblock)
929 {
930 	struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
931 
932 	BUG_ON(!origin_bio);
933 
934 	bio_chain(origin_bio, bio);
935 	/*
936 	 * Passing false to __remap_to_origin_clear_discard() skips
937 	 * all code that might use per_bio_data (since clone doesn't have it)
938 	 */
939 	__remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
940 	submit_bio(origin_bio);
941 
942 	remap_to_cache(cache, bio, cblock);
943 }
944 
945 /*----------------------------------------------------------------
946  * Failure modes
947  *--------------------------------------------------------------*/
948 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
949 {
950 	return cache->features.mode;
951 }
952 
953 static const char *cache_device_name(struct cache *cache)
954 {
955 	return dm_device_name(dm_table_get_md(cache->ti->table));
956 }
957 
958 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
959 {
960 	const char *descs[] = {
961 		"write",
962 		"read-only",
963 		"fail"
964 	};
965 
966 	dm_table_event(cache->ti->table);
967 	DMINFO("%s: switching cache to %s mode",
968 	       cache_device_name(cache), descs[(int)mode]);
969 }
970 
971 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
972 {
973 	bool needs_check;
974 	enum cache_metadata_mode old_mode = get_cache_mode(cache);
975 
976 	if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
977 		DMERR("%s: unable to read needs_check flag, setting failure mode.",
978 		      cache_device_name(cache));
979 		new_mode = CM_FAIL;
980 	}
981 
982 	if (new_mode == CM_WRITE && needs_check) {
983 		DMERR("%s: unable to switch cache to write mode until repaired.",
984 		      cache_device_name(cache));
985 		if (old_mode != new_mode)
986 			new_mode = old_mode;
987 		else
988 			new_mode = CM_READ_ONLY;
989 	}
990 
991 	/* Never move out of fail mode */
992 	if (old_mode == CM_FAIL)
993 		new_mode = CM_FAIL;
994 
995 	switch (new_mode) {
996 	case CM_FAIL:
997 	case CM_READ_ONLY:
998 		dm_cache_metadata_set_read_only(cache->cmd);
999 		break;
1000 
1001 	case CM_WRITE:
1002 		dm_cache_metadata_set_read_write(cache->cmd);
1003 		break;
1004 	}
1005 
1006 	cache->features.mode = new_mode;
1007 
1008 	if (new_mode != old_mode)
1009 		notify_mode_switch(cache, new_mode);
1010 }
1011 
1012 static void abort_transaction(struct cache *cache)
1013 {
1014 	const char *dev_name = cache_device_name(cache);
1015 
1016 	if (get_cache_mode(cache) >= CM_READ_ONLY)
1017 		return;
1018 
1019 	if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1020 		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1021 		set_cache_mode(cache, CM_FAIL);
1022 	}
1023 
1024 	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1025 	if (dm_cache_metadata_abort(cache->cmd)) {
1026 		DMERR("%s: failed to abort metadata transaction", dev_name);
1027 		set_cache_mode(cache, CM_FAIL);
1028 	}
1029 }
1030 
1031 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1032 {
1033 	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1034 		    cache_device_name(cache), op, r);
1035 	abort_transaction(cache);
1036 	set_cache_mode(cache, CM_READ_ONLY);
1037 }
1038 
1039 /*----------------------------------------------------------------*/
1040 
1041 static void load_stats(struct cache *cache)
1042 {
1043 	struct dm_cache_statistics stats;
1044 
1045 	dm_cache_metadata_get_stats(cache->cmd, &stats);
1046 	atomic_set(&cache->stats.read_hit, stats.read_hits);
1047 	atomic_set(&cache->stats.read_miss, stats.read_misses);
1048 	atomic_set(&cache->stats.write_hit, stats.write_hits);
1049 	atomic_set(&cache->stats.write_miss, stats.write_misses);
1050 }
1051 
1052 static void save_stats(struct cache *cache)
1053 {
1054 	struct dm_cache_statistics stats;
1055 
1056 	if (get_cache_mode(cache) >= CM_READ_ONLY)
1057 		return;
1058 
1059 	stats.read_hits = atomic_read(&cache->stats.read_hit);
1060 	stats.read_misses = atomic_read(&cache->stats.read_miss);
1061 	stats.write_hits = atomic_read(&cache->stats.write_hit);
1062 	stats.write_misses = atomic_read(&cache->stats.write_miss);
1063 
1064 	dm_cache_metadata_set_stats(cache->cmd, &stats);
1065 }
1066 
1067 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1068 {
1069 	switch (op) {
1070 	case POLICY_PROMOTE:
1071 		atomic_inc(&stats->promotion);
1072 		break;
1073 
1074 	case POLICY_DEMOTE:
1075 		atomic_inc(&stats->demotion);
1076 		break;
1077 
1078 	case POLICY_WRITEBACK:
1079 		atomic_inc(&stats->writeback);
1080 		break;
1081 	}
1082 }
1083 
1084 /*----------------------------------------------------------------
1085  * Migration processing
1086  *
1087  * Migration covers moving data from the origin device to the cache, or
1088  * vice versa.
1089  *--------------------------------------------------------------*/
1090 
1091 static void inc_io_migrations(struct cache *cache)
1092 {
1093 	atomic_inc(&cache->nr_io_migrations);
1094 }
1095 
1096 static void dec_io_migrations(struct cache *cache)
1097 {
1098 	atomic_dec(&cache->nr_io_migrations);
1099 }
1100 
1101 static bool discard_or_flush(struct bio *bio)
1102 {
1103 	return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1104 }
1105 
1106 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1107 				     dm_dblock_t *b, dm_dblock_t *e)
1108 {
1109 	sector_t sb = bio->bi_iter.bi_sector;
1110 	sector_t se = bio_end_sector(bio);
1111 
1112 	*b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1113 
1114 	if (se - sb < cache->discard_block_size)
1115 		*e = *b;
1116 	else
1117 		*e = to_dblock(block_div(se, cache->discard_block_size));
1118 }
1119 
1120 /*----------------------------------------------------------------*/
1121 
1122 static void prevent_background_work(struct cache *cache)
1123 {
1124 	lockdep_off();
1125 	down_write(&cache->background_work_lock);
1126 	lockdep_on();
1127 }
1128 
1129 static void allow_background_work(struct cache *cache)
1130 {
1131 	lockdep_off();
1132 	up_write(&cache->background_work_lock);
1133 	lockdep_on();
1134 }
1135 
1136 static bool background_work_begin(struct cache *cache)
1137 {
1138 	bool r;
1139 
1140 	lockdep_off();
1141 	r = down_read_trylock(&cache->background_work_lock);
1142 	lockdep_on();
1143 
1144 	return r;
1145 }
1146 
1147 static void background_work_end(struct cache *cache)
1148 {
1149 	lockdep_off();
1150 	up_read(&cache->background_work_lock);
1151 	lockdep_on();
1152 }
1153 
1154 /*----------------------------------------------------------------*/
1155 
1156 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1157 {
1158 	return (bio_data_dir(bio) == WRITE) &&
1159 		(bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1160 }
1161 
1162 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1163 {
1164 	return writeback_mode(cache) &&
1165 		(is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1166 }
1167 
1168 static void quiesce(struct dm_cache_migration *mg,
1169 		    void (*continuation)(struct work_struct *))
1170 {
1171 	init_continuation(&mg->k, continuation);
1172 	dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1173 }
1174 
1175 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1176 {
1177 	struct continuation *k = container_of(ws, struct continuation, ws);
1178 	return container_of(k, struct dm_cache_migration, k);
1179 }
1180 
1181 static void copy_complete(int read_err, unsigned long write_err, void *context)
1182 {
1183 	struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1184 
1185 	if (read_err || write_err)
1186 		mg->k.input = BLK_STS_IOERR;
1187 
1188 	queue_continuation(mg->cache->wq, &mg->k);
1189 }
1190 
1191 static void copy(struct dm_cache_migration *mg, bool promote)
1192 {
1193 	struct dm_io_region o_region, c_region;
1194 	struct cache *cache = mg->cache;
1195 
1196 	o_region.bdev = cache->origin_dev->bdev;
1197 	o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1198 	o_region.count = cache->sectors_per_block;
1199 
1200 	c_region.bdev = cache->cache_dev->bdev;
1201 	c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1202 	c_region.count = cache->sectors_per_block;
1203 
1204 	if (promote)
1205 		dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1206 	else
1207 		dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1208 }
1209 
1210 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1211 {
1212 	struct per_bio_data *pb = get_per_bio_data(bio);
1213 
1214 	if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1215 		free_prison_cell(cache, pb->cell);
1216 	pb->cell = NULL;
1217 }
1218 
1219 static void overwrite_endio(struct bio *bio)
1220 {
1221 	struct dm_cache_migration *mg = bio->bi_private;
1222 	struct cache *cache = mg->cache;
1223 	struct per_bio_data *pb = get_per_bio_data(bio);
1224 
1225 	dm_unhook_bio(&pb->hook_info, bio);
1226 
1227 	if (bio->bi_status)
1228 		mg->k.input = bio->bi_status;
1229 
1230 	queue_continuation(cache->wq, &mg->k);
1231 }
1232 
1233 static void overwrite(struct dm_cache_migration *mg,
1234 		      void (*continuation)(struct work_struct *))
1235 {
1236 	struct bio *bio = mg->overwrite_bio;
1237 	struct per_bio_data *pb = get_per_bio_data(bio);
1238 
1239 	dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1240 
1241 	/*
1242 	 * The overwrite bio is part of the copy operation, as such it does
1243 	 * not set/clear discard or dirty flags.
1244 	 */
1245 	if (mg->op->op == POLICY_PROMOTE)
1246 		remap_to_cache(mg->cache, bio, mg->op->cblock);
1247 	else
1248 		remap_to_origin(mg->cache, bio);
1249 
1250 	init_continuation(&mg->k, continuation);
1251 	accounted_request(mg->cache, bio);
1252 }
1253 
1254 /*
1255  * Migration steps:
1256  *
1257  * 1) exclusive lock preventing WRITEs
1258  * 2) quiesce
1259  * 3) copy or issue overwrite bio
1260  * 4) upgrade to exclusive lock preventing READs and WRITEs
1261  * 5) quiesce
1262  * 6) update metadata and commit
1263  * 7) unlock
1264  */
1265 static void mg_complete(struct dm_cache_migration *mg, bool success)
1266 {
1267 	struct bio_list bios;
1268 	struct cache *cache = mg->cache;
1269 	struct policy_work *op = mg->op;
1270 	dm_cblock_t cblock = op->cblock;
1271 
1272 	if (success)
1273 		update_stats(&cache->stats, op->op);
1274 
1275 	switch (op->op) {
1276 	case POLICY_PROMOTE:
1277 		clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1278 		policy_complete_background_work(cache->policy, op, success);
1279 
1280 		if (mg->overwrite_bio) {
1281 			if (success)
1282 				force_set_dirty(cache, cblock);
1283 			else if (mg->k.input)
1284 				mg->overwrite_bio->bi_status = mg->k.input;
1285 			else
1286 				mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1287 			bio_endio(mg->overwrite_bio);
1288 		} else {
1289 			if (success)
1290 				force_clear_dirty(cache, cblock);
1291 			dec_io_migrations(cache);
1292 		}
1293 		break;
1294 
1295 	case POLICY_DEMOTE:
1296 		/*
1297 		 * We clear dirty here to update the nr_dirty counter.
1298 		 */
1299 		if (success)
1300 			force_clear_dirty(cache, cblock);
1301 		policy_complete_background_work(cache->policy, op, success);
1302 		dec_io_migrations(cache);
1303 		break;
1304 
1305 	case POLICY_WRITEBACK:
1306 		if (success)
1307 			force_clear_dirty(cache, cblock);
1308 		policy_complete_background_work(cache->policy, op, success);
1309 		dec_io_migrations(cache);
1310 		break;
1311 	}
1312 
1313 	bio_list_init(&bios);
1314 	if (mg->cell) {
1315 		if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1316 			free_prison_cell(cache, mg->cell);
1317 	}
1318 
1319 	free_migration(mg);
1320 	defer_bios(cache, &bios);
1321 	wake_migration_worker(cache);
1322 
1323 	background_work_end(cache);
1324 }
1325 
1326 static void mg_success(struct work_struct *ws)
1327 {
1328 	struct dm_cache_migration *mg = ws_to_mg(ws);
1329 	mg_complete(mg, mg->k.input == 0);
1330 }
1331 
1332 static void mg_update_metadata(struct work_struct *ws)
1333 {
1334 	int r;
1335 	struct dm_cache_migration *mg = ws_to_mg(ws);
1336 	struct cache *cache = mg->cache;
1337 	struct policy_work *op = mg->op;
1338 
1339 	switch (op->op) {
1340 	case POLICY_PROMOTE:
1341 		r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1342 		if (r) {
1343 			DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1344 				    cache_device_name(cache));
1345 			metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1346 
1347 			mg_complete(mg, false);
1348 			return;
1349 		}
1350 		mg_complete(mg, true);
1351 		break;
1352 
1353 	case POLICY_DEMOTE:
1354 		r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1355 		if (r) {
1356 			DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1357 				    cache_device_name(cache));
1358 			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1359 
1360 			mg_complete(mg, false);
1361 			return;
1362 		}
1363 
1364 		/*
1365 		 * It would be nice if we only had to commit when a REQ_FLUSH
1366 		 * comes through.  But there's one scenario that we have to
1367 		 * look out for:
1368 		 *
1369 		 * - vblock x in a cache block
1370 		 * - domotion occurs
1371 		 * - cache block gets reallocated and over written
1372 		 * - crash
1373 		 *
1374 		 * When we recover, because there was no commit the cache will
1375 		 * rollback to having the data for vblock x in the cache block.
1376 		 * But the cache block has since been overwritten, so it'll end
1377 		 * up pointing to data that was never in 'x' during the history
1378 		 * of the device.
1379 		 *
1380 		 * To avoid this issue we require a commit as part of the
1381 		 * demotion operation.
1382 		 */
1383 		init_continuation(&mg->k, mg_success);
1384 		continue_after_commit(&cache->committer, &mg->k);
1385 		schedule_commit(&cache->committer);
1386 		break;
1387 
1388 	case POLICY_WRITEBACK:
1389 		mg_complete(mg, true);
1390 		break;
1391 	}
1392 }
1393 
1394 static void mg_update_metadata_after_copy(struct work_struct *ws)
1395 {
1396 	struct dm_cache_migration *mg = ws_to_mg(ws);
1397 
1398 	/*
1399 	 * Did the copy succeed?
1400 	 */
1401 	if (mg->k.input)
1402 		mg_complete(mg, false);
1403 	else
1404 		mg_update_metadata(ws);
1405 }
1406 
1407 static void mg_upgrade_lock(struct work_struct *ws)
1408 {
1409 	int r;
1410 	struct dm_cache_migration *mg = ws_to_mg(ws);
1411 
1412 	/*
1413 	 * Did the copy succeed?
1414 	 */
1415 	if (mg->k.input)
1416 		mg_complete(mg, false);
1417 
1418 	else {
1419 		/*
1420 		 * Now we want the lock to prevent both reads and writes.
1421 		 */
1422 		r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1423 					    READ_WRITE_LOCK_LEVEL);
1424 		if (r < 0)
1425 			mg_complete(mg, false);
1426 
1427 		else if (r)
1428 			quiesce(mg, mg_update_metadata);
1429 
1430 		else
1431 			mg_update_metadata(ws);
1432 	}
1433 }
1434 
1435 static void mg_full_copy(struct work_struct *ws)
1436 {
1437 	struct dm_cache_migration *mg = ws_to_mg(ws);
1438 	struct cache *cache = mg->cache;
1439 	struct policy_work *op = mg->op;
1440 	bool is_policy_promote = (op->op == POLICY_PROMOTE);
1441 
1442 	if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1443 	    is_discarded_oblock(cache, op->oblock)) {
1444 		mg_upgrade_lock(ws);
1445 		return;
1446 	}
1447 
1448 	init_continuation(&mg->k, mg_upgrade_lock);
1449 	copy(mg, is_policy_promote);
1450 }
1451 
1452 static void mg_copy(struct work_struct *ws)
1453 {
1454 	struct dm_cache_migration *mg = ws_to_mg(ws);
1455 
1456 	if (mg->overwrite_bio) {
1457 		/*
1458 		 * No exclusive lock was held when we last checked if the bio
1459 		 * was optimisable.  So we have to check again in case things
1460 		 * have changed (eg, the block may no longer be discarded).
1461 		 */
1462 		if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1463 			/*
1464 			 * Fallback to a real full copy after doing some tidying up.
1465 			 */
1466 			bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1467 			BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1468 			mg->overwrite_bio = NULL;
1469 			inc_io_migrations(mg->cache);
1470 			mg_full_copy(ws);
1471 			return;
1472 		}
1473 
1474 		/*
1475 		 * It's safe to do this here, even though it's new data
1476 		 * because all IO has been locked out of the block.
1477 		 *
1478 		 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1479 		 * so _not_ using mg_upgrade_lock() as continutation.
1480 		 */
1481 		overwrite(mg, mg_update_metadata_after_copy);
1482 
1483 	} else
1484 		mg_full_copy(ws);
1485 }
1486 
1487 static int mg_lock_writes(struct dm_cache_migration *mg)
1488 {
1489 	int r;
1490 	struct dm_cell_key_v2 key;
1491 	struct cache *cache = mg->cache;
1492 	struct dm_bio_prison_cell_v2 *prealloc;
1493 
1494 	prealloc = alloc_prison_cell(cache);
1495 	if (!prealloc) {
1496 		DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1497 		mg_complete(mg, false);
1498 		return -ENOMEM;
1499 	}
1500 
1501 	/*
1502 	 * Prevent writes to the block, but allow reads to continue.
1503 	 * Unless we're using an overwrite bio, in which case we lock
1504 	 * everything.
1505 	 */
1506 	build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1507 	r = dm_cell_lock_v2(cache->prison, &key,
1508 			    mg->overwrite_bio ?  READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1509 			    prealloc, &mg->cell);
1510 	if (r < 0) {
1511 		free_prison_cell(cache, prealloc);
1512 		mg_complete(mg, false);
1513 		return r;
1514 	}
1515 
1516 	if (mg->cell != prealloc)
1517 		free_prison_cell(cache, prealloc);
1518 
1519 	if (r == 0)
1520 		mg_copy(&mg->k.ws);
1521 	else
1522 		quiesce(mg, mg_copy);
1523 
1524 	return 0;
1525 }
1526 
1527 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1528 {
1529 	struct dm_cache_migration *mg;
1530 
1531 	if (!background_work_begin(cache)) {
1532 		policy_complete_background_work(cache->policy, op, false);
1533 		return -EPERM;
1534 	}
1535 
1536 	mg = alloc_migration(cache);
1537 	if (!mg) {
1538 		policy_complete_background_work(cache->policy, op, false);
1539 		background_work_end(cache);
1540 		return -ENOMEM;
1541 	}
1542 
1543 	mg->op = op;
1544 	mg->overwrite_bio = bio;
1545 
1546 	if (!bio)
1547 		inc_io_migrations(cache);
1548 
1549 	return mg_lock_writes(mg);
1550 }
1551 
1552 /*----------------------------------------------------------------
1553  * invalidation processing
1554  *--------------------------------------------------------------*/
1555 
1556 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1557 {
1558 	struct bio_list bios;
1559 	struct cache *cache = mg->cache;
1560 
1561 	bio_list_init(&bios);
1562 	if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1563 		free_prison_cell(cache, mg->cell);
1564 
1565 	if (!success && mg->overwrite_bio)
1566 		bio_io_error(mg->overwrite_bio);
1567 
1568 	free_migration(mg);
1569 	defer_bios(cache, &bios);
1570 
1571 	background_work_end(cache);
1572 }
1573 
1574 static void invalidate_completed(struct work_struct *ws)
1575 {
1576 	struct dm_cache_migration *mg = ws_to_mg(ws);
1577 	invalidate_complete(mg, !mg->k.input);
1578 }
1579 
1580 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1581 {
1582 	int r = policy_invalidate_mapping(cache->policy, cblock);
1583 	if (!r) {
1584 		r = dm_cache_remove_mapping(cache->cmd, cblock);
1585 		if (r) {
1586 			DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1587 				    cache_device_name(cache));
1588 			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1589 		}
1590 
1591 	} else if (r == -ENODATA) {
1592 		/*
1593 		 * Harmless, already unmapped.
1594 		 */
1595 		r = 0;
1596 
1597 	} else
1598 		DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1599 
1600 	return r;
1601 }
1602 
1603 static void invalidate_remove(struct work_struct *ws)
1604 {
1605 	int r;
1606 	struct dm_cache_migration *mg = ws_to_mg(ws);
1607 	struct cache *cache = mg->cache;
1608 
1609 	r = invalidate_cblock(cache, mg->invalidate_cblock);
1610 	if (r) {
1611 		invalidate_complete(mg, false);
1612 		return;
1613 	}
1614 
1615 	init_continuation(&mg->k, invalidate_completed);
1616 	continue_after_commit(&cache->committer, &mg->k);
1617 	remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1618 	mg->overwrite_bio = NULL;
1619 	schedule_commit(&cache->committer);
1620 }
1621 
1622 static int invalidate_lock(struct dm_cache_migration *mg)
1623 {
1624 	int r;
1625 	struct dm_cell_key_v2 key;
1626 	struct cache *cache = mg->cache;
1627 	struct dm_bio_prison_cell_v2 *prealloc;
1628 
1629 	prealloc = alloc_prison_cell(cache);
1630 	if (!prealloc) {
1631 		invalidate_complete(mg, false);
1632 		return -ENOMEM;
1633 	}
1634 
1635 	build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1636 	r = dm_cell_lock_v2(cache->prison, &key,
1637 			    READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1638 	if (r < 0) {
1639 		free_prison_cell(cache, prealloc);
1640 		invalidate_complete(mg, false);
1641 		return r;
1642 	}
1643 
1644 	if (mg->cell != prealloc)
1645 		free_prison_cell(cache, prealloc);
1646 
1647 	if (r)
1648 		quiesce(mg, invalidate_remove);
1649 
1650 	else {
1651 		/*
1652 		 * We can't call invalidate_remove() directly here because we
1653 		 * might still be in request context.
1654 		 */
1655 		init_continuation(&mg->k, invalidate_remove);
1656 		queue_work(cache->wq, &mg->k.ws);
1657 	}
1658 
1659 	return 0;
1660 }
1661 
1662 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1663 			    dm_oblock_t oblock, struct bio *bio)
1664 {
1665 	struct dm_cache_migration *mg;
1666 
1667 	if (!background_work_begin(cache))
1668 		return -EPERM;
1669 
1670 	mg = alloc_migration(cache);
1671 	if (!mg) {
1672 		background_work_end(cache);
1673 		return -ENOMEM;
1674 	}
1675 
1676 	mg->overwrite_bio = bio;
1677 	mg->invalidate_cblock = cblock;
1678 	mg->invalidate_oblock = oblock;
1679 
1680 	return invalidate_lock(mg);
1681 }
1682 
1683 /*----------------------------------------------------------------
1684  * bio processing
1685  *--------------------------------------------------------------*/
1686 
1687 enum busy {
1688 	IDLE,
1689 	BUSY
1690 };
1691 
1692 static enum busy spare_migration_bandwidth(struct cache *cache)
1693 {
1694 	bool idle = iot_idle_for(&cache->tracker, HZ);
1695 	sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1696 		cache->sectors_per_block;
1697 
1698 	if (idle && current_volume <= cache->migration_threshold)
1699 		return IDLE;
1700 	else
1701 		return BUSY;
1702 }
1703 
1704 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1705 {
1706 	atomic_inc(bio_data_dir(bio) == READ ?
1707 		   &cache->stats.read_hit : &cache->stats.write_hit);
1708 }
1709 
1710 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1711 {
1712 	atomic_inc(bio_data_dir(bio) == READ ?
1713 		   &cache->stats.read_miss : &cache->stats.write_miss);
1714 }
1715 
1716 /*----------------------------------------------------------------*/
1717 
1718 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1719 		   bool *commit_needed)
1720 {
1721 	int r, data_dir;
1722 	bool rb, background_queued;
1723 	dm_cblock_t cblock;
1724 
1725 	*commit_needed = false;
1726 
1727 	rb = bio_detain_shared(cache, block, bio);
1728 	if (!rb) {
1729 		/*
1730 		 * An exclusive lock is held for this block, so we have to
1731 		 * wait.  We set the commit_needed flag so the current
1732 		 * transaction will be committed asap, allowing this lock
1733 		 * to be dropped.
1734 		 */
1735 		*commit_needed = true;
1736 		return DM_MAPIO_SUBMITTED;
1737 	}
1738 
1739 	data_dir = bio_data_dir(bio);
1740 
1741 	if (optimisable_bio(cache, bio, block)) {
1742 		struct policy_work *op = NULL;
1743 
1744 		r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1745 		if (unlikely(r && r != -ENOENT)) {
1746 			DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1747 				    cache_device_name(cache), r);
1748 			bio_io_error(bio);
1749 			return DM_MAPIO_SUBMITTED;
1750 		}
1751 
1752 		if (r == -ENOENT && op) {
1753 			bio_drop_shared_lock(cache, bio);
1754 			BUG_ON(op->op != POLICY_PROMOTE);
1755 			mg_start(cache, op, bio);
1756 			return DM_MAPIO_SUBMITTED;
1757 		}
1758 	} else {
1759 		r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1760 		if (unlikely(r && r != -ENOENT)) {
1761 			DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1762 				    cache_device_name(cache), r);
1763 			bio_io_error(bio);
1764 			return DM_MAPIO_SUBMITTED;
1765 		}
1766 
1767 		if (background_queued)
1768 			wake_migration_worker(cache);
1769 	}
1770 
1771 	if (r == -ENOENT) {
1772 		struct per_bio_data *pb = get_per_bio_data(bio);
1773 
1774 		/*
1775 		 * Miss.
1776 		 */
1777 		inc_miss_counter(cache, bio);
1778 		if (pb->req_nr == 0) {
1779 			accounted_begin(cache, bio);
1780 			remap_to_origin_clear_discard(cache, bio, block);
1781 		} else {
1782 			/*
1783 			 * This is a duplicate writethrough io that is no
1784 			 * longer needed because the block has been demoted.
1785 			 */
1786 			bio_endio(bio);
1787 			return DM_MAPIO_SUBMITTED;
1788 		}
1789 	} else {
1790 		/*
1791 		 * Hit.
1792 		 */
1793 		inc_hit_counter(cache, bio);
1794 
1795 		/*
1796 		 * Passthrough always maps to the origin, invalidating any
1797 		 * cache blocks that are written to.
1798 		 */
1799 		if (passthrough_mode(cache)) {
1800 			if (bio_data_dir(bio) == WRITE) {
1801 				bio_drop_shared_lock(cache, bio);
1802 				atomic_inc(&cache->stats.demotion);
1803 				invalidate_start(cache, cblock, block, bio);
1804 			} else
1805 				remap_to_origin_clear_discard(cache, bio, block);
1806 		} else {
1807 			if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1808 			    !is_dirty(cache, cblock)) {
1809 				remap_to_origin_and_cache(cache, bio, block, cblock);
1810 				accounted_begin(cache, bio);
1811 			} else
1812 				remap_to_cache_dirty(cache, bio, block, cblock);
1813 		}
1814 	}
1815 
1816 	/*
1817 	 * dm core turns FUA requests into a separate payload and FLUSH req.
1818 	 */
1819 	if (bio->bi_opf & REQ_FUA) {
1820 		/*
1821 		 * issue_after_commit will call accounted_begin a second time.  So
1822 		 * we call accounted_complete() to avoid double accounting.
1823 		 */
1824 		accounted_complete(cache, bio);
1825 		issue_after_commit(&cache->committer, bio);
1826 		*commit_needed = true;
1827 		return DM_MAPIO_SUBMITTED;
1828 	}
1829 
1830 	return DM_MAPIO_REMAPPED;
1831 }
1832 
1833 static bool process_bio(struct cache *cache, struct bio *bio)
1834 {
1835 	bool commit_needed;
1836 
1837 	if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1838 		generic_make_request(bio);
1839 
1840 	return commit_needed;
1841 }
1842 
1843 /*
1844  * A non-zero return indicates read_only or fail_io mode.
1845  */
1846 static int commit(struct cache *cache, bool clean_shutdown)
1847 {
1848 	int r;
1849 
1850 	if (get_cache_mode(cache) >= CM_READ_ONLY)
1851 		return -EINVAL;
1852 
1853 	atomic_inc(&cache->stats.commit_count);
1854 	r = dm_cache_commit(cache->cmd, clean_shutdown);
1855 	if (r)
1856 		metadata_operation_failed(cache, "dm_cache_commit", r);
1857 
1858 	return r;
1859 }
1860 
1861 /*
1862  * Used by the batcher.
1863  */
1864 static blk_status_t commit_op(void *context)
1865 {
1866 	struct cache *cache = context;
1867 
1868 	if (dm_cache_changed_this_transaction(cache->cmd))
1869 		return errno_to_blk_status(commit(cache, false));
1870 
1871 	return 0;
1872 }
1873 
1874 /*----------------------------------------------------------------*/
1875 
1876 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1877 {
1878 	struct per_bio_data *pb = get_per_bio_data(bio);
1879 
1880 	if (!pb->req_nr)
1881 		remap_to_origin(cache, bio);
1882 	else
1883 		remap_to_cache(cache, bio, 0);
1884 
1885 	issue_after_commit(&cache->committer, bio);
1886 	return true;
1887 }
1888 
1889 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1890 {
1891 	dm_dblock_t b, e;
1892 
1893 	// FIXME: do we need to lock the region?  Or can we just assume the
1894 	// user wont be so foolish as to issue discard concurrently with
1895 	// other IO?
1896 	calc_discard_block_range(cache, bio, &b, &e);
1897 	while (b != e) {
1898 		set_discard(cache, b);
1899 		b = to_dblock(from_dblock(b) + 1);
1900 	}
1901 
1902 	bio_endio(bio);
1903 
1904 	return false;
1905 }
1906 
1907 static void process_deferred_bios(struct work_struct *ws)
1908 {
1909 	struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1910 
1911 	unsigned long flags;
1912 	bool commit_needed = false;
1913 	struct bio_list bios;
1914 	struct bio *bio;
1915 
1916 	bio_list_init(&bios);
1917 
1918 	spin_lock_irqsave(&cache->lock, flags);
1919 	bio_list_merge(&bios, &cache->deferred_bios);
1920 	bio_list_init(&cache->deferred_bios);
1921 	spin_unlock_irqrestore(&cache->lock, flags);
1922 
1923 	while ((bio = bio_list_pop(&bios))) {
1924 		if (bio->bi_opf & REQ_PREFLUSH)
1925 			commit_needed = process_flush_bio(cache, bio) || commit_needed;
1926 
1927 		else if (bio_op(bio) == REQ_OP_DISCARD)
1928 			commit_needed = process_discard_bio(cache, bio) || commit_needed;
1929 
1930 		else
1931 			commit_needed = process_bio(cache, bio) || commit_needed;
1932 	}
1933 
1934 	if (commit_needed)
1935 		schedule_commit(&cache->committer);
1936 }
1937 
1938 /*----------------------------------------------------------------
1939  * Main worker loop
1940  *--------------------------------------------------------------*/
1941 
1942 static void requeue_deferred_bios(struct cache *cache)
1943 {
1944 	struct bio *bio;
1945 	struct bio_list bios;
1946 
1947 	bio_list_init(&bios);
1948 	bio_list_merge(&bios, &cache->deferred_bios);
1949 	bio_list_init(&cache->deferred_bios);
1950 
1951 	while ((bio = bio_list_pop(&bios))) {
1952 		bio->bi_status = BLK_STS_DM_REQUEUE;
1953 		bio_endio(bio);
1954 	}
1955 }
1956 
1957 /*
1958  * We want to commit periodically so that not too much
1959  * unwritten metadata builds up.
1960  */
1961 static void do_waker(struct work_struct *ws)
1962 {
1963 	struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1964 
1965 	policy_tick(cache->policy, true);
1966 	wake_migration_worker(cache);
1967 	schedule_commit(&cache->committer);
1968 	queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1969 }
1970 
1971 static void check_migrations(struct work_struct *ws)
1972 {
1973 	int r;
1974 	struct policy_work *op;
1975 	struct cache *cache = container_of(ws, struct cache, migration_worker);
1976 	enum busy b;
1977 
1978 	for (;;) {
1979 		b = spare_migration_bandwidth(cache);
1980 
1981 		r = policy_get_background_work(cache->policy, b == IDLE, &op);
1982 		if (r == -ENODATA)
1983 			break;
1984 
1985 		if (r) {
1986 			DMERR_LIMIT("%s: policy_background_work failed",
1987 				    cache_device_name(cache));
1988 			break;
1989 		}
1990 
1991 		r = mg_start(cache, op, NULL);
1992 		if (r)
1993 			break;
1994 	}
1995 }
1996 
1997 /*----------------------------------------------------------------
1998  * Target methods
1999  *--------------------------------------------------------------*/
2000 
2001 /*
2002  * This function gets called on the error paths of the constructor, so we
2003  * have to cope with a partially initialised struct.
2004  */
2005 static void destroy(struct cache *cache)
2006 {
2007 	unsigned i;
2008 
2009 	mempool_exit(&cache->migration_pool);
2010 
2011 	if (cache->prison)
2012 		dm_bio_prison_destroy_v2(cache->prison);
2013 
2014 	if (cache->wq)
2015 		destroy_workqueue(cache->wq);
2016 
2017 	if (cache->dirty_bitset)
2018 		free_bitset(cache->dirty_bitset);
2019 
2020 	if (cache->discard_bitset)
2021 		free_bitset(cache->discard_bitset);
2022 
2023 	if (cache->copier)
2024 		dm_kcopyd_client_destroy(cache->copier);
2025 
2026 	if (cache->cmd)
2027 		dm_cache_metadata_close(cache->cmd);
2028 
2029 	if (cache->metadata_dev)
2030 		dm_put_device(cache->ti, cache->metadata_dev);
2031 
2032 	if (cache->origin_dev)
2033 		dm_put_device(cache->ti, cache->origin_dev);
2034 
2035 	if (cache->cache_dev)
2036 		dm_put_device(cache->ti, cache->cache_dev);
2037 
2038 	if (cache->policy)
2039 		dm_cache_policy_destroy(cache->policy);
2040 
2041 	for (i = 0; i < cache->nr_ctr_args ; i++)
2042 		kfree(cache->ctr_args[i]);
2043 	kfree(cache->ctr_args);
2044 
2045 	bioset_exit(&cache->bs);
2046 
2047 	kfree(cache);
2048 }
2049 
2050 static void cache_dtr(struct dm_target *ti)
2051 {
2052 	struct cache *cache = ti->private;
2053 
2054 	destroy(cache);
2055 }
2056 
2057 static sector_t get_dev_size(struct dm_dev *dev)
2058 {
2059 	return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2060 }
2061 
2062 /*----------------------------------------------------------------*/
2063 
2064 /*
2065  * Construct a cache device mapping.
2066  *
2067  * cache <metadata dev> <cache dev> <origin dev> <block size>
2068  *       <#feature args> [<feature arg>]*
2069  *       <policy> <#policy args> [<policy arg>]*
2070  *
2071  * metadata dev    : fast device holding the persistent metadata
2072  * cache dev	   : fast device holding cached data blocks
2073  * origin dev	   : slow device holding original data blocks
2074  * block size	   : cache unit size in sectors
2075  *
2076  * #feature args   : number of feature arguments passed
2077  * feature args    : writethrough.  (The default is writeback.)
2078  *
2079  * policy	   : the replacement policy to use
2080  * #policy args    : an even number of policy arguments corresponding
2081  *		     to key/value pairs passed to the policy
2082  * policy args	   : key/value pairs passed to the policy
2083  *		     E.g. 'sequential_threshold 1024'
2084  *		     See cache-policies.txt for details.
2085  *
2086  * Optional feature arguments are:
2087  *   writethrough  : write through caching that prohibits cache block
2088  *		     content from being different from origin block content.
2089  *		     Without this argument, the default behaviour is to write
2090  *		     back cache block contents later for performance reasons,
2091  *		     so they may differ from the corresponding origin blocks.
2092  */
2093 struct cache_args {
2094 	struct dm_target *ti;
2095 
2096 	struct dm_dev *metadata_dev;
2097 
2098 	struct dm_dev *cache_dev;
2099 	sector_t cache_sectors;
2100 
2101 	struct dm_dev *origin_dev;
2102 	sector_t origin_sectors;
2103 
2104 	uint32_t block_size;
2105 
2106 	const char *policy_name;
2107 	int policy_argc;
2108 	const char **policy_argv;
2109 
2110 	struct cache_features features;
2111 };
2112 
2113 static void destroy_cache_args(struct cache_args *ca)
2114 {
2115 	if (ca->metadata_dev)
2116 		dm_put_device(ca->ti, ca->metadata_dev);
2117 
2118 	if (ca->cache_dev)
2119 		dm_put_device(ca->ti, ca->cache_dev);
2120 
2121 	if (ca->origin_dev)
2122 		dm_put_device(ca->ti, ca->origin_dev);
2123 
2124 	kfree(ca);
2125 }
2126 
2127 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2128 {
2129 	if (!as->argc) {
2130 		*error = "Insufficient args";
2131 		return false;
2132 	}
2133 
2134 	return true;
2135 }
2136 
2137 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2138 			      char **error)
2139 {
2140 	int r;
2141 	sector_t metadata_dev_size;
2142 	char b[BDEVNAME_SIZE];
2143 
2144 	if (!at_least_one_arg(as, error))
2145 		return -EINVAL;
2146 
2147 	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2148 			  &ca->metadata_dev);
2149 	if (r) {
2150 		*error = "Error opening metadata device";
2151 		return r;
2152 	}
2153 
2154 	metadata_dev_size = get_dev_size(ca->metadata_dev);
2155 	if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2156 		DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2157 		       bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2158 
2159 	return 0;
2160 }
2161 
2162 static int parse_cache_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->cache_dev);
2172 	if (r) {
2173 		*error = "Error opening cache device";
2174 		return r;
2175 	}
2176 	ca->cache_sectors = get_dev_size(ca->cache_dev);
2177 
2178 	return 0;
2179 }
2180 
2181 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2182 			    char **error)
2183 {
2184 	int r;
2185 
2186 	if (!at_least_one_arg(as, error))
2187 		return -EINVAL;
2188 
2189 	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2190 			  &ca->origin_dev);
2191 	if (r) {
2192 		*error = "Error opening origin device";
2193 		return r;
2194 	}
2195 
2196 	ca->origin_sectors = get_dev_size(ca->origin_dev);
2197 	if (ca->ti->len > ca->origin_sectors) {
2198 		*error = "Device size larger than cached device";
2199 		return -EINVAL;
2200 	}
2201 
2202 	return 0;
2203 }
2204 
2205 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2206 			    char **error)
2207 {
2208 	unsigned long block_size;
2209 
2210 	if (!at_least_one_arg(as, error))
2211 		return -EINVAL;
2212 
2213 	if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2214 	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2215 	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2216 	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2217 		*error = "Invalid data block size";
2218 		return -EINVAL;
2219 	}
2220 
2221 	if (block_size > ca->cache_sectors) {
2222 		*error = "Data block size is larger than the cache device";
2223 		return -EINVAL;
2224 	}
2225 
2226 	ca->block_size = block_size;
2227 
2228 	return 0;
2229 }
2230 
2231 static void init_features(struct cache_features *cf)
2232 {
2233 	cf->mode = CM_WRITE;
2234 	cf->io_mode = CM_IO_WRITEBACK;
2235 	cf->metadata_version = 1;
2236 }
2237 
2238 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2239 			  char **error)
2240 {
2241 	static const struct dm_arg _args[] = {
2242 		{0, 2, "Invalid number of cache feature arguments"},
2243 	};
2244 
2245 	int r, mode_ctr = 0;
2246 	unsigned argc;
2247 	const char *arg;
2248 	struct cache_features *cf = &ca->features;
2249 
2250 	init_features(cf);
2251 
2252 	r = dm_read_arg_group(_args, as, &argc, error);
2253 	if (r)
2254 		return -EINVAL;
2255 
2256 	while (argc--) {
2257 		arg = dm_shift_arg(as);
2258 
2259 		if (!strcasecmp(arg, "writeback")) {
2260 			cf->io_mode = CM_IO_WRITEBACK;
2261 			mode_ctr++;
2262 		}
2263 
2264 		else if (!strcasecmp(arg, "writethrough")) {
2265 			cf->io_mode = CM_IO_WRITETHROUGH;
2266 			mode_ctr++;
2267 		}
2268 
2269 		else if (!strcasecmp(arg, "passthrough")) {
2270 			cf->io_mode = CM_IO_PASSTHROUGH;
2271 			mode_ctr++;
2272 		}
2273 
2274 		else if (!strcasecmp(arg, "metadata2"))
2275 			cf->metadata_version = 2;
2276 
2277 		else {
2278 			*error = "Unrecognised cache feature requested";
2279 			return -EINVAL;
2280 		}
2281 	}
2282 
2283 	if (mode_ctr > 1) {
2284 		*error = "Duplicate cache io_mode features requested";
2285 		return -EINVAL;
2286 	}
2287 
2288 	return 0;
2289 }
2290 
2291 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2292 			char **error)
2293 {
2294 	static const struct dm_arg _args[] = {
2295 		{0, 1024, "Invalid number of policy arguments"},
2296 	};
2297 
2298 	int r;
2299 
2300 	if (!at_least_one_arg(as, error))
2301 		return -EINVAL;
2302 
2303 	ca->policy_name = dm_shift_arg(as);
2304 
2305 	r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2306 	if (r)
2307 		return -EINVAL;
2308 
2309 	ca->policy_argv = (const char **)as->argv;
2310 	dm_consume_args(as, ca->policy_argc);
2311 
2312 	return 0;
2313 }
2314 
2315 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2316 			    char **error)
2317 {
2318 	int r;
2319 	struct dm_arg_set as;
2320 
2321 	as.argc = argc;
2322 	as.argv = argv;
2323 
2324 	r = parse_metadata_dev(ca, &as, error);
2325 	if (r)
2326 		return r;
2327 
2328 	r = parse_cache_dev(ca, &as, error);
2329 	if (r)
2330 		return r;
2331 
2332 	r = parse_origin_dev(ca, &as, error);
2333 	if (r)
2334 		return r;
2335 
2336 	r = parse_block_size(ca, &as, error);
2337 	if (r)
2338 		return r;
2339 
2340 	r = parse_features(ca, &as, error);
2341 	if (r)
2342 		return r;
2343 
2344 	r = parse_policy(ca, &as, error);
2345 	if (r)
2346 		return r;
2347 
2348 	return 0;
2349 }
2350 
2351 /*----------------------------------------------------------------*/
2352 
2353 static struct kmem_cache *migration_cache;
2354 
2355 #define NOT_CORE_OPTION 1
2356 
2357 static int process_config_option(struct cache *cache, const char *key, const char *value)
2358 {
2359 	unsigned long tmp;
2360 
2361 	if (!strcasecmp(key, "migration_threshold")) {
2362 		if (kstrtoul(value, 10, &tmp))
2363 			return -EINVAL;
2364 
2365 		cache->migration_threshold = tmp;
2366 		return 0;
2367 	}
2368 
2369 	return NOT_CORE_OPTION;
2370 }
2371 
2372 static int set_config_value(struct cache *cache, const char *key, const char *value)
2373 {
2374 	int r = process_config_option(cache, key, value);
2375 
2376 	if (r == NOT_CORE_OPTION)
2377 		r = policy_set_config_value(cache->policy, key, value);
2378 
2379 	if (r)
2380 		DMWARN("bad config value for %s: %s", key, value);
2381 
2382 	return r;
2383 }
2384 
2385 static int set_config_values(struct cache *cache, int argc, const char **argv)
2386 {
2387 	int r = 0;
2388 
2389 	if (argc & 1) {
2390 		DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2391 		return -EINVAL;
2392 	}
2393 
2394 	while (argc) {
2395 		r = set_config_value(cache, argv[0], argv[1]);
2396 		if (r)
2397 			break;
2398 
2399 		argc -= 2;
2400 		argv += 2;
2401 	}
2402 
2403 	return r;
2404 }
2405 
2406 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2407 			       char **error)
2408 {
2409 	struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2410 							   cache->cache_size,
2411 							   cache->origin_sectors,
2412 							   cache->sectors_per_block);
2413 	if (IS_ERR(p)) {
2414 		*error = "Error creating cache's policy";
2415 		return PTR_ERR(p);
2416 	}
2417 	cache->policy = p;
2418 	BUG_ON(!cache->policy);
2419 
2420 	return 0;
2421 }
2422 
2423 /*
2424  * We want the discard block size to be at least the size of the cache
2425  * block size and have no more than 2^14 discard blocks across the origin.
2426  */
2427 #define MAX_DISCARD_BLOCKS (1 << 14)
2428 
2429 static bool too_many_discard_blocks(sector_t discard_block_size,
2430 				    sector_t origin_size)
2431 {
2432 	(void) sector_div(origin_size, discard_block_size);
2433 
2434 	return origin_size > MAX_DISCARD_BLOCKS;
2435 }
2436 
2437 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2438 					     sector_t origin_size)
2439 {
2440 	sector_t discard_block_size = cache_block_size;
2441 
2442 	if (origin_size)
2443 		while (too_many_discard_blocks(discard_block_size, origin_size))
2444 			discard_block_size *= 2;
2445 
2446 	return discard_block_size;
2447 }
2448 
2449 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2450 {
2451 	dm_block_t nr_blocks = from_cblock(size);
2452 
2453 	if (nr_blocks > (1 << 20) && cache->cache_size != size)
2454 		DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2455 			     "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2456 			     "Please consider increasing the cache block size to reduce the overall cache block count.",
2457 			     (unsigned long long) nr_blocks);
2458 
2459 	cache->cache_size = size;
2460 }
2461 
2462 static int is_congested(struct dm_dev *dev, int bdi_bits)
2463 {
2464 	struct request_queue *q = bdev_get_queue(dev->bdev);
2465 	return bdi_congested(q->backing_dev_info, bdi_bits);
2466 }
2467 
2468 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2469 {
2470 	struct cache *cache = container_of(cb, struct cache, callbacks);
2471 
2472 	return is_congested(cache->origin_dev, bdi_bits) ||
2473 		is_congested(cache->cache_dev, bdi_bits);
2474 }
2475 
2476 #define DEFAULT_MIGRATION_THRESHOLD 2048
2477 
2478 static int cache_create(struct cache_args *ca, struct cache **result)
2479 {
2480 	int r = 0;
2481 	char **error = &ca->ti->error;
2482 	struct cache *cache;
2483 	struct dm_target *ti = ca->ti;
2484 	dm_block_t origin_blocks;
2485 	struct dm_cache_metadata *cmd;
2486 	bool may_format = ca->features.mode == CM_WRITE;
2487 
2488 	cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2489 	if (!cache)
2490 		return -ENOMEM;
2491 
2492 	cache->ti = ca->ti;
2493 	ti->private = cache;
2494 	ti->num_flush_bios = 2;
2495 	ti->flush_supported = true;
2496 
2497 	ti->num_discard_bios = 1;
2498 	ti->discards_supported = true;
2499 	ti->split_discard_bios = false;
2500 
2501 	ti->per_io_data_size = sizeof(struct per_bio_data);
2502 
2503 	cache->features = ca->features;
2504 	if (writethrough_mode(cache)) {
2505 		/* Create bioset for writethrough bios issued to origin */
2506 		r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2507 		if (r)
2508 			goto bad;
2509 	}
2510 
2511 	cache->callbacks.congested_fn = cache_is_congested;
2512 	dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2513 
2514 	cache->metadata_dev = ca->metadata_dev;
2515 	cache->origin_dev = ca->origin_dev;
2516 	cache->cache_dev = ca->cache_dev;
2517 
2518 	ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2519 
2520 	origin_blocks = cache->origin_sectors = ca->origin_sectors;
2521 	origin_blocks = block_div(origin_blocks, ca->block_size);
2522 	cache->origin_blocks = to_oblock(origin_blocks);
2523 
2524 	cache->sectors_per_block = ca->block_size;
2525 	if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2526 		r = -EINVAL;
2527 		goto bad;
2528 	}
2529 
2530 	if (ca->block_size & (ca->block_size - 1)) {
2531 		dm_block_t cache_size = ca->cache_sectors;
2532 
2533 		cache->sectors_per_block_shift = -1;
2534 		cache_size = block_div(cache_size, ca->block_size);
2535 		set_cache_size(cache, to_cblock(cache_size));
2536 	} else {
2537 		cache->sectors_per_block_shift = __ffs(ca->block_size);
2538 		set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2539 	}
2540 
2541 	r = create_cache_policy(cache, ca, error);
2542 	if (r)
2543 		goto bad;
2544 
2545 	cache->policy_nr_args = ca->policy_argc;
2546 	cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2547 
2548 	r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2549 	if (r) {
2550 		*error = "Error setting cache policy's config values";
2551 		goto bad;
2552 	}
2553 
2554 	cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2555 				     ca->block_size, may_format,
2556 				     dm_cache_policy_get_hint_size(cache->policy),
2557 				     ca->features.metadata_version);
2558 	if (IS_ERR(cmd)) {
2559 		*error = "Error creating metadata object";
2560 		r = PTR_ERR(cmd);
2561 		goto bad;
2562 	}
2563 	cache->cmd = cmd;
2564 	set_cache_mode(cache, CM_WRITE);
2565 	if (get_cache_mode(cache) != CM_WRITE) {
2566 		*error = "Unable to get write access to metadata, please check/repair metadata.";
2567 		r = -EINVAL;
2568 		goto bad;
2569 	}
2570 
2571 	if (passthrough_mode(cache)) {
2572 		bool all_clean;
2573 
2574 		r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2575 		if (r) {
2576 			*error = "dm_cache_metadata_all_clean() failed";
2577 			goto bad;
2578 		}
2579 
2580 		if (!all_clean) {
2581 			*error = "Cannot enter passthrough mode unless all blocks are clean";
2582 			r = -EINVAL;
2583 			goto bad;
2584 		}
2585 
2586 		policy_allow_migrations(cache->policy, false);
2587 	}
2588 
2589 	spin_lock_init(&cache->lock);
2590 	bio_list_init(&cache->deferred_bios);
2591 	atomic_set(&cache->nr_allocated_migrations, 0);
2592 	atomic_set(&cache->nr_io_migrations, 0);
2593 	init_waitqueue_head(&cache->migration_wait);
2594 
2595 	r = -ENOMEM;
2596 	atomic_set(&cache->nr_dirty, 0);
2597 	cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2598 	if (!cache->dirty_bitset) {
2599 		*error = "could not allocate dirty bitset";
2600 		goto bad;
2601 	}
2602 	clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2603 
2604 	cache->discard_block_size =
2605 		calculate_discard_block_size(cache->sectors_per_block,
2606 					     cache->origin_sectors);
2607 	cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2608 							      cache->discard_block_size));
2609 	cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2610 	if (!cache->discard_bitset) {
2611 		*error = "could not allocate discard bitset";
2612 		goto bad;
2613 	}
2614 	clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2615 
2616 	cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2617 	if (IS_ERR(cache->copier)) {
2618 		*error = "could not create kcopyd client";
2619 		r = PTR_ERR(cache->copier);
2620 		goto bad;
2621 	}
2622 
2623 	cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2624 	if (!cache->wq) {
2625 		*error = "could not create workqueue for metadata object";
2626 		goto bad;
2627 	}
2628 	INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2629 	INIT_WORK(&cache->migration_worker, check_migrations);
2630 	INIT_DELAYED_WORK(&cache->waker, do_waker);
2631 
2632 	cache->prison = dm_bio_prison_create_v2(cache->wq);
2633 	if (!cache->prison) {
2634 		*error = "could not create bio prison";
2635 		goto bad;
2636 	}
2637 
2638 	r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2639 				   migration_cache);
2640 	if (r) {
2641 		*error = "Error creating cache's migration mempool";
2642 		goto bad;
2643 	}
2644 
2645 	cache->need_tick_bio = true;
2646 	cache->sized = false;
2647 	cache->invalidate = false;
2648 	cache->commit_requested = false;
2649 	cache->loaded_mappings = false;
2650 	cache->loaded_discards = false;
2651 
2652 	load_stats(cache);
2653 
2654 	atomic_set(&cache->stats.demotion, 0);
2655 	atomic_set(&cache->stats.promotion, 0);
2656 	atomic_set(&cache->stats.copies_avoided, 0);
2657 	atomic_set(&cache->stats.cache_cell_clash, 0);
2658 	atomic_set(&cache->stats.commit_count, 0);
2659 	atomic_set(&cache->stats.discard_count, 0);
2660 
2661 	spin_lock_init(&cache->invalidation_lock);
2662 	INIT_LIST_HEAD(&cache->invalidation_requests);
2663 
2664 	batcher_init(&cache->committer, commit_op, cache,
2665 		     issue_op, cache, cache->wq);
2666 	iot_init(&cache->tracker);
2667 
2668 	init_rwsem(&cache->background_work_lock);
2669 	prevent_background_work(cache);
2670 
2671 	*result = cache;
2672 	return 0;
2673 bad:
2674 	destroy(cache);
2675 	return r;
2676 }
2677 
2678 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2679 {
2680 	unsigned i;
2681 	const char **copy;
2682 
2683 	copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2684 	if (!copy)
2685 		return -ENOMEM;
2686 	for (i = 0; i < argc; i++) {
2687 		copy[i] = kstrdup(argv[i], GFP_KERNEL);
2688 		if (!copy[i]) {
2689 			while (i--)
2690 				kfree(copy[i]);
2691 			kfree(copy);
2692 			return -ENOMEM;
2693 		}
2694 	}
2695 
2696 	cache->nr_ctr_args = argc;
2697 	cache->ctr_args = copy;
2698 
2699 	return 0;
2700 }
2701 
2702 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2703 {
2704 	int r = -EINVAL;
2705 	struct cache_args *ca;
2706 	struct cache *cache = NULL;
2707 
2708 	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2709 	if (!ca) {
2710 		ti->error = "Error allocating memory for cache";
2711 		return -ENOMEM;
2712 	}
2713 	ca->ti = ti;
2714 
2715 	r = parse_cache_args(ca, argc, argv, &ti->error);
2716 	if (r)
2717 		goto out;
2718 
2719 	r = cache_create(ca, &cache);
2720 	if (r)
2721 		goto out;
2722 
2723 	r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2724 	if (r) {
2725 		destroy(cache);
2726 		goto out;
2727 	}
2728 
2729 	ti->private = cache;
2730 out:
2731 	destroy_cache_args(ca);
2732 	return r;
2733 }
2734 
2735 /*----------------------------------------------------------------*/
2736 
2737 static int cache_map(struct dm_target *ti, struct bio *bio)
2738 {
2739 	struct cache *cache = ti->private;
2740 
2741 	int r;
2742 	bool commit_needed;
2743 	dm_oblock_t block = get_bio_block(cache, bio);
2744 
2745 	init_per_bio_data(bio);
2746 	if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2747 		/*
2748 		 * This can only occur if the io goes to a partial block at
2749 		 * the end of the origin device.  We don't cache these.
2750 		 * Just remap to the origin and carry on.
2751 		 */
2752 		remap_to_origin(cache, bio);
2753 		accounted_begin(cache, bio);
2754 		return DM_MAPIO_REMAPPED;
2755 	}
2756 
2757 	if (discard_or_flush(bio)) {
2758 		defer_bio(cache, bio);
2759 		return DM_MAPIO_SUBMITTED;
2760 	}
2761 
2762 	r = map_bio(cache, bio, block, &commit_needed);
2763 	if (commit_needed)
2764 		schedule_commit(&cache->committer);
2765 
2766 	return r;
2767 }
2768 
2769 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2770 {
2771 	struct cache *cache = ti->private;
2772 	unsigned long flags;
2773 	struct per_bio_data *pb = get_per_bio_data(bio);
2774 
2775 	if (pb->tick) {
2776 		policy_tick(cache->policy, false);
2777 
2778 		spin_lock_irqsave(&cache->lock, flags);
2779 		cache->need_tick_bio = true;
2780 		spin_unlock_irqrestore(&cache->lock, flags);
2781 	}
2782 
2783 	bio_drop_shared_lock(cache, bio);
2784 	accounted_complete(cache, bio);
2785 
2786 	return DM_ENDIO_DONE;
2787 }
2788 
2789 static int write_dirty_bitset(struct cache *cache)
2790 {
2791 	int r;
2792 
2793 	if (get_cache_mode(cache) >= CM_READ_ONLY)
2794 		return -EINVAL;
2795 
2796 	r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2797 	if (r)
2798 		metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2799 
2800 	return r;
2801 }
2802 
2803 static int write_discard_bitset(struct cache *cache)
2804 {
2805 	unsigned i, r;
2806 
2807 	if (get_cache_mode(cache) >= CM_READ_ONLY)
2808 		return -EINVAL;
2809 
2810 	r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2811 					   cache->discard_nr_blocks);
2812 	if (r) {
2813 		DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2814 		metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2815 		return r;
2816 	}
2817 
2818 	for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2819 		r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2820 					 is_discarded(cache, to_dblock(i)));
2821 		if (r) {
2822 			metadata_operation_failed(cache, "dm_cache_set_discard", r);
2823 			return r;
2824 		}
2825 	}
2826 
2827 	return 0;
2828 }
2829 
2830 static int write_hints(struct cache *cache)
2831 {
2832 	int r;
2833 
2834 	if (get_cache_mode(cache) >= CM_READ_ONLY)
2835 		return -EINVAL;
2836 
2837 	r = dm_cache_write_hints(cache->cmd, cache->policy);
2838 	if (r) {
2839 		metadata_operation_failed(cache, "dm_cache_write_hints", r);
2840 		return r;
2841 	}
2842 
2843 	return 0;
2844 }
2845 
2846 /*
2847  * returns true on success
2848  */
2849 static bool sync_metadata(struct cache *cache)
2850 {
2851 	int r1, r2, r3, r4;
2852 
2853 	r1 = write_dirty_bitset(cache);
2854 	if (r1)
2855 		DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2856 
2857 	r2 = write_discard_bitset(cache);
2858 	if (r2)
2859 		DMERR("%s: could not write discard bitset", cache_device_name(cache));
2860 
2861 	save_stats(cache);
2862 
2863 	r3 = write_hints(cache);
2864 	if (r3)
2865 		DMERR("%s: could not write hints", cache_device_name(cache));
2866 
2867 	/*
2868 	 * If writing the above metadata failed, we still commit, but don't
2869 	 * set the clean shutdown flag.  This will effectively force every
2870 	 * dirty bit to be set on reload.
2871 	 */
2872 	r4 = commit(cache, !r1 && !r2 && !r3);
2873 	if (r4)
2874 		DMERR("%s: could not write cache metadata", cache_device_name(cache));
2875 
2876 	return !r1 && !r2 && !r3 && !r4;
2877 }
2878 
2879 static void cache_postsuspend(struct dm_target *ti)
2880 {
2881 	struct cache *cache = ti->private;
2882 
2883 	prevent_background_work(cache);
2884 	BUG_ON(atomic_read(&cache->nr_io_migrations));
2885 
2886 	cancel_delayed_work(&cache->waker);
2887 	flush_workqueue(cache->wq);
2888 	WARN_ON(cache->tracker.in_flight);
2889 
2890 	/*
2891 	 * If it's a flush suspend there won't be any deferred bios, so this
2892 	 * call is harmless.
2893 	 */
2894 	requeue_deferred_bios(cache);
2895 
2896 	if (get_cache_mode(cache) == CM_WRITE)
2897 		(void) sync_metadata(cache);
2898 }
2899 
2900 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2901 			bool dirty, uint32_t hint, bool hint_valid)
2902 {
2903 	int r;
2904 	struct cache *cache = context;
2905 
2906 	if (dirty) {
2907 		set_bit(from_cblock(cblock), cache->dirty_bitset);
2908 		atomic_inc(&cache->nr_dirty);
2909 	} else
2910 		clear_bit(from_cblock(cblock), cache->dirty_bitset);
2911 
2912 	r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2913 	if (r)
2914 		return r;
2915 
2916 	return 0;
2917 }
2918 
2919 /*
2920  * The discard block size in the on disk metadata is not
2921  * neccessarily the same as we're currently using.  So we have to
2922  * be careful to only set the discarded attribute if we know it
2923  * covers a complete block of the new size.
2924  */
2925 struct discard_load_info {
2926 	struct cache *cache;
2927 
2928 	/*
2929 	 * These blocks are sized using the on disk dblock size, rather
2930 	 * than the current one.
2931 	 */
2932 	dm_block_t block_size;
2933 	dm_block_t discard_begin, discard_end;
2934 };
2935 
2936 static void discard_load_info_init(struct cache *cache,
2937 				   struct discard_load_info *li)
2938 {
2939 	li->cache = cache;
2940 	li->discard_begin = li->discard_end = 0;
2941 }
2942 
2943 static void set_discard_range(struct discard_load_info *li)
2944 {
2945 	sector_t b, e;
2946 
2947 	if (li->discard_begin == li->discard_end)
2948 		return;
2949 
2950 	/*
2951 	 * Convert to sectors.
2952 	 */
2953 	b = li->discard_begin * li->block_size;
2954 	e = li->discard_end * li->block_size;
2955 
2956 	/*
2957 	 * Then convert back to the current dblock size.
2958 	 */
2959 	b = dm_sector_div_up(b, li->cache->discard_block_size);
2960 	sector_div(e, li->cache->discard_block_size);
2961 
2962 	/*
2963 	 * The origin may have shrunk, so we need to check we're still in
2964 	 * bounds.
2965 	 */
2966 	if (e > from_dblock(li->cache->discard_nr_blocks))
2967 		e = from_dblock(li->cache->discard_nr_blocks);
2968 
2969 	for (; b < e; b++)
2970 		set_discard(li->cache, to_dblock(b));
2971 }
2972 
2973 static int load_discard(void *context, sector_t discard_block_size,
2974 			dm_dblock_t dblock, bool discard)
2975 {
2976 	struct discard_load_info *li = context;
2977 
2978 	li->block_size = discard_block_size;
2979 
2980 	if (discard) {
2981 		if (from_dblock(dblock) == li->discard_end)
2982 			/*
2983 			 * We're already in a discard range, just extend it.
2984 			 */
2985 			li->discard_end = li->discard_end + 1ULL;
2986 
2987 		else {
2988 			/*
2989 			 * Emit the old range and start a new one.
2990 			 */
2991 			set_discard_range(li);
2992 			li->discard_begin = from_dblock(dblock);
2993 			li->discard_end = li->discard_begin + 1ULL;
2994 		}
2995 	} else {
2996 		set_discard_range(li);
2997 		li->discard_begin = li->discard_end = 0;
2998 	}
2999 
3000 	return 0;
3001 }
3002 
3003 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3004 {
3005 	sector_t size = get_dev_size(cache->cache_dev);
3006 	(void) sector_div(size, cache->sectors_per_block);
3007 	return to_cblock(size);
3008 }
3009 
3010 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3011 {
3012 	if (from_cblock(new_size) > from_cblock(cache->cache_size))
3013 		return true;
3014 
3015 	/*
3016 	 * We can't drop a dirty block when shrinking the cache.
3017 	 */
3018 	while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3019 		new_size = to_cblock(from_cblock(new_size) + 1);
3020 		if (is_dirty(cache, new_size)) {
3021 			DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3022 			      cache_device_name(cache),
3023 			      (unsigned long long) from_cblock(new_size));
3024 			return false;
3025 		}
3026 	}
3027 
3028 	return true;
3029 }
3030 
3031 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3032 {
3033 	int r;
3034 
3035 	r = dm_cache_resize(cache->cmd, new_size);
3036 	if (r) {
3037 		DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3038 		metadata_operation_failed(cache, "dm_cache_resize", r);
3039 		return r;
3040 	}
3041 
3042 	set_cache_size(cache, new_size);
3043 
3044 	return 0;
3045 }
3046 
3047 static int cache_preresume(struct dm_target *ti)
3048 {
3049 	int r = 0;
3050 	struct cache *cache = ti->private;
3051 	dm_cblock_t csize = get_cache_dev_size(cache);
3052 
3053 	/*
3054 	 * Check to see if the cache has resized.
3055 	 */
3056 	if (!cache->sized) {
3057 		r = resize_cache_dev(cache, csize);
3058 		if (r)
3059 			return r;
3060 
3061 		cache->sized = true;
3062 
3063 	} else if (csize != cache->cache_size) {
3064 		if (!can_resize(cache, csize))
3065 			return -EINVAL;
3066 
3067 		r = resize_cache_dev(cache, csize);
3068 		if (r)
3069 			return r;
3070 	}
3071 
3072 	if (!cache->loaded_mappings) {
3073 		r = dm_cache_load_mappings(cache->cmd, cache->policy,
3074 					   load_mapping, cache);
3075 		if (r) {
3076 			DMERR("%s: could not load cache mappings", cache_device_name(cache));
3077 			metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3078 			return r;
3079 		}
3080 
3081 		cache->loaded_mappings = true;
3082 	}
3083 
3084 	if (!cache->loaded_discards) {
3085 		struct discard_load_info li;
3086 
3087 		/*
3088 		 * The discard bitset could have been resized, or the
3089 		 * discard block size changed.  To be safe we start by
3090 		 * setting every dblock to not discarded.
3091 		 */
3092 		clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3093 
3094 		discard_load_info_init(cache, &li);
3095 		r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3096 		if (r) {
3097 			DMERR("%s: could not load origin discards", cache_device_name(cache));
3098 			metadata_operation_failed(cache, "dm_cache_load_discards", r);
3099 			return r;
3100 		}
3101 		set_discard_range(&li);
3102 
3103 		cache->loaded_discards = true;
3104 	}
3105 
3106 	return r;
3107 }
3108 
3109 static void cache_resume(struct dm_target *ti)
3110 {
3111 	struct cache *cache = ti->private;
3112 
3113 	cache->need_tick_bio = true;
3114 	allow_background_work(cache);
3115 	do_waker(&cache->waker.work);
3116 }
3117 
3118 /*
3119  * Status format:
3120  *
3121  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3122  * <cache block size> <#used cache blocks>/<#total cache blocks>
3123  * <#read hits> <#read misses> <#write hits> <#write misses>
3124  * <#demotions> <#promotions> <#dirty>
3125  * <#features> <features>*
3126  * <#core args> <core args>
3127  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3128  */
3129 static void cache_status(struct dm_target *ti, status_type_t type,
3130 			 unsigned status_flags, char *result, unsigned maxlen)
3131 {
3132 	int r = 0;
3133 	unsigned i;
3134 	ssize_t sz = 0;
3135 	dm_block_t nr_free_blocks_metadata = 0;
3136 	dm_block_t nr_blocks_metadata = 0;
3137 	char buf[BDEVNAME_SIZE];
3138 	struct cache *cache = ti->private;
3139 	dm_cblock_t residency;
3140 	bool needs_check;
3141 
3142 	switch (type) {
3143 	case STATUSTYPE_INFO:
3144 		if (get_cache_mode(cache) == CM_FAIL) {
3145 			DMEMIT("Fail");
3146 			break;
3147 		}
3148 
3149 		/* Commit to ensure statistics aren't out-of-date */
3150 		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3151 			(void) commit(cache, false);
3152 
3153 		r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3154 		if (r) {
3155 			DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3156 			      cache_device_name(cache), r);
3157 			goto err;
3158 		}
3159 
3160 		r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3161 		if (r) {
3162 			DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3163 			      cache_device_name(cache), r);
3164 			goto err;
3165 		}
3166 
3167 		residency = policy_residency(cache->policy);
3168 
3169 		DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3170 		       (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3171 		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3172 		       (unsigned long long)nr_blocks_metadata,
3173 		       (unsigned long long)cache->sectors_per_block,
3174 		       (unsigned long long) from_cblock(residency),
3175 		       (unsigned long long) from_cblock(cache->cache_size),
3176 		       (unsigned) atomic_read(&cache->stats.read_hit),
3177 		       (unsigned) atomic_read(&cache->stats.read_miss),
3178 		       (unsigned) atomic_read(&cache->stats.write_hit),
3179 		       (unsigned) atomic_read(&cache->stats.write_miss),
3180 		       (unsigned) atomic_read(&cache->stats.demotion),
3181 		       (unsigned) atomic_read(&cache->stats.promotion),
3182 		       (unsigned long) atomic_read(&cache->nr_dirty));
3183 
3184 		if (cache->features.metadata_version == 2)
3185 			DMEMIT("2 metadata2 ");
3186 		else
3187 			DMEMIT("1 ");
3188 
3189 		if (writethrough_mode(cache))
3190 			DMEMIT("writethrough ");
3191 
3192 		else if (passthrough_mode(cache))
3193 			DMEMIT("passthrough ");
3194 
3195 		else if (writeback_mode(cache))
3196 			DMEMIT("writeback ");
3197 
3198 		else {
3199 			DMERR("%s: internal error: unknown io mode: %d",
3200 			      cache_device_name(cache), (int) cache->features.io_mode);
3201 			goto err;
3202 		}
3203 
3204 		DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3205 
3206 		DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3207 		if (sz < maxlen) {
3208 			r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3209 			if (r)
3210 				DMERR("%s: policy_emit_config_values returned %d",
3211 				      cache_device_name(cache), r);
3212 		}
3213 
3214 		if (get_cache_mode(cache) == CM_READ_ONLY)
3215 			DMEMIT("ro ");
3216 		else
3217 			DMEMIT("rw ");
3218 
3219 		r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3220 
3221 		if (r || needs_check)
3222 			DMEMIT("needs_check ");
3223 		else
3224 			DMEMIT("- ");
3225 
3226 		break;
3227 
3228 	case STATUSTYPE_TABLE:
3229 		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3230 		DMEMIT("%s ", buf);
3231 		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3232 		DMEMIT("%s ", buf);
3233 		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3234 		DMEMIT("%s", buf);
3235 
3236 		for (i = 0; i < cache->nr_ctr_args - 1; i++)
3237 			DMEMIT(" %s", cache->ctr_args[i]);
3238 		if (cache->nr_ctr_args)
3239 			DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3240 	}
3241 
3242 	return;
3243 
3244 err:
3245 	DMEMIT("Error");
3246 }
3247 
3248 /*
3249  * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
3250  * the one-past-the-end value.
3251  */
3252 struct cblock_range {
3253 	dm_cblock_t begin;
3254 	dm_cblock_t end;
3255 };
3256 
3257 /*
3258  * A cache block range can take two forms:
3259  *
3260  * i) A single cblock, eg. '3456'
3261  * ii) A begin and end cblock with a dash between, eg. 123-234
3262  */
3263 static int parse_cblock_range(struct cache *cache, const char *str,
3264 			      struct cblock_range *result)
3265 {
3266 	char dummy;
3267 	uint64_t b, e;
3268 	int r;
3269 
3270 	/*
3271 	 * Try and parse form (ii) first.
3272 	 */
3273 	r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3274 	if (r < 0)
3275 		return r;
3276 
3277 	if (r == 2) {
3278 		result->begin = to_cblock(b);
3279 		result->end = to_cblock(e);
3280 		return 0;
3281 	}
3282 
3283 	/*
3284 	 * That didn't work, try form (i).
3285 	 */
3286 	r = sscanf(str, "%llu%c", &b, &dummy);
3287 	if (r < 0)
3288 		return r;
3289 
3290 	if (r == 1) {
3291 		result->begin = to_cblock(b);
3292 		result->end = to_cblock(from_cblock(result->begin) + 1u);
3293 		return 0;
3294 	}
3295 
3296 	DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3297 	return -EINVAL;
3298 }
3299 
3300 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3301 {
3302 	uint64_t b = from_cblock(range->begin);
3303 	uint64_t e = from_cblock(range->end);
3304 	uint64_t n = from_cblock(cache->cache_size);
3305 
3306 	if (b >= n) {
3307 		DMERR("%s: begin cblock out of range: %llu >= %llu",
3308 		      cache_device_name(cache), b, n);
3309 		return -EINVAL;
3310 	}
3311 
3312 	if (e > n) {
3313 		DMERR("%s: end cblock out of range: %llu > %llu",
3314 		      cache_device_name(cache), e, n);
3315 		return -EINVAL;
3316 	}
3317 
3318 	if (b >= e) {
3319 		DMERR("%s: invalid cblock range: %llu >= %llu",
3320 		      cache_device_name(cache), b, e);
3321 		return -EINVAL;
3322 	}
3323 
3324 	return 0;
3325 }
3326 
3327 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3328 {
3329 	return to_cblock(from_cblock(b) + 1);
3330 }
3331 
3332 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3333 {
3334 	int r = 0;
3335 
3336 	/*
3337 	 * We don't need to do any locking here because we know we're in
3338 	 * passthrough mode.  There's is potential for a race between an
3339 	 * invalidation triggered by an io and an invalidation message.  This
3340 	 * is harmless, we must not worry if the policy call fails.
3341 	 */
3342 	while (range->begin != range->end) {
3343 		r = invalidate_cblock(cache, range->begin);
3344 		if (r)
3345 			return r;
3346 
3347 		range->begin = cblock_succ(range->begin);
3348 	}
3349 
3350 	cache->commit_requested = true;
3351 	return r;
3352 }
3353 
3354 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3355 					      const char **cblock_ranges)
3356 {
3357 	int r = 0;
3358 	unsigned i;
3359 	struct cblock_range range;
3360 
3361 	if (!passthrough_mode(cache)) {
3362 		DMERR("%s: cache has to be in passthrough mode for invalidation",
3363 		      cache_device_name(cache));
3364 		return -EPERM;
3365 	}
3366 
3367 	for (i = 0; i < count; i++) {
3368 		r = parse_cblock_range(cache, cblock_ranges[i], &range);
3369 		if (r)
3370 			break;
3371 
3372 		r = validate_cblock_range(cache, &range);
3373 		if (r)
3374 			break;
3375 
3376 		/*
3377 		 * Pass begin and end origin blocks to the worker and wake it.
3378 		 */
3379 		r = request_invalidation(cache, &range);
3380 		if (r)
3381 			break;
3382 	}
3383 
3384 	return r;
3385 }
3386 
3387 /*
3388  * Supports
3389  *	"<key> <value>"
3390  * and
3391  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3392  *
3393  * The key migration_threshold is supported by the cache target core.
3394  */
3395 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3396 			 char *result, unsigned maxlen)
3397 {
3398 	struct cache *cache = ti->private;
3399 
3400 	if (!argc)
3401 		return -EINVAL;
3402 
3403 	if (get_cache_mode(cache) >= CM_READ_ONLY) {
3404 		DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3405 		      cache_device_name(cache));
3406 		return -EOPNOTSUPP;
3407 	}
3408 
3409 	if (!strcasecmp(argv[0], "invalidate_cblocks"))
3410 		return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3411 
3412 	if (argc != 2)
3413 		return -EINVAL;
3414 
3415 	return set_config_value(cache, argv[0], argv[1]);
3416 }
3417 
3418 static int cache_iterate_devices(struct dm_target *ti,
3419 				 iterate_devices_callout_fn fn, void *data)
3420 {
3421 	int r = 0;
3422 	struct cache *cache = ti->private;
3423 
3424 	r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3425 	if (!r)
3426 		r = fn(ti, cache->origin_dev, 0, ti->len, data);
3427 
3428 	return r;
3429 }
3430 
3431 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3432 {
3433 	/*
3434 	 * FIXME: these limits may be incompatible with the cache device
3435 	 */
3436 	limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3437 					    cache->origin_sectors);
3438 	limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3439 }
3440 
3441 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3442 {
3443 	struct cache *cache = ti->private;
3444 	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3445 
3446 	/*
3447 	 * If the system-determined stacked limits are compatible with the
3448 	 * cache's blocksize (io_opt is a factor) do not override them.
3449 	 */
3450 	if (io_opt_sectors < cache->sectors_per_block ||
3451 	    do_div(io_opt_sectors, cache->sectors_per_block)) {
3452 		blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3453 		blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3454 	}
3455 	set_discard_limits(cache, limits);
3456 }
3457 
3458 /*----------------------------------------------------------------*/
3459 
3460 static struct target_type cache_target = {
3461 	.name = "cache",
3462 	.version = {2, 0, 0},
3463 	.module = THIS_MODULE,
3464 	.ctr = cache_ctr,
3465 	.dtr = cache_dtr,
3466 	.map = cache_map,
3467 	.end_io = cache_end_io,
3468 	.postsuspend = cache_postsuspend,
3469 	.preresume = cache_preresume,
3470 	.resume = cache_resume,
3471 	.status = cache_status,
3472 	.message = cache_message,
3473 	.iterate_devices = cache_iterate_devices,
3474 	.io_hints = cache_io_hints,
3475 };
3476 
3477 static int __init dm_cache_init(void)
3478 {
3479 	int r;
3480 
3481 	migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3482 	if (!migration_cache) {
3483 		dm_unregister_target(&cache_target);
3484 		return -ENOMEM;
3485 	}
3486 
3487 	r = dm_register_target(&cache_target);
3488 	if (r) {
3489 		DMERR("cache target registration failed: %d", r);
3490 		return r;
3491 	}
3492 
3493 	return 0;
3494 }
3495 
3496 static void __exit dm_cache_exit(void)
3497 {
3498 	dm_unregister_target(&cache_target);
3499 	kmem_cache_destroy(migration_cache);
3500 }
3501 
3502 module_init(dm_cache_init);
3503 module_exit(dm_cache_exit);
3504 
3505 MODULE_DESCRIPTION(DM_NAME " cache target");
3506 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3507 MODULE_LICENSE("GPL");
3508