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