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 uint32_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_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
792 pb->tick = true;
793 cache->need_tick_bio = false;
794 }
795 spin_unlock_irqrestore(&cache->lock, flags);
796}
797
798static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
799 dm_oblock_t oblock)
800{
801 check_if_tick_bio_needed(cache, bio);
802 remap_to_origin(cache, bio);
803 if (bio_data_dir(bio) == WRITE)
804 clear_discard(cache, oblock_to_dblock(cache, oblock));
805}
806
807static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
808 dm_oblock_t oblock, dm_cblock_t cblock)
809{
810 check_if_tick_bio_needed(cache, bio);
811 remap_to_cache(cache, bio, cblock);
812 if (bio_data_dir(bio) == WRITE) {
813 set_dirty(cache, oblock, cblock);
814 clear_discard(cache, oblock_to_dblock(cache, oblock));
815 }
816}
817
818static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
819{
820 sector_t block_nr = bio->bi_iter.bi_sector;
821
822 if (!block_size_is_power_of_two(cache))
823 (void) sector_div(block_nr, cache->sectors_per_block);
824 else
825 block_nr >>= cache->sectors_per_block_shift;
826
827 return to_oblock(block_nr);
828}
829
830static int bio_triggers_commit(struct cache *cache, struct bio *bio)
831{
832 return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
833}
834
835/*
836 * You must increment the deferred set whilst the prison cell is held. To
837 * encourage this, we ask for 'cell' to be passed in.
838 */
839static void inc_ds(struct cache *cache, struct bio *bio,
840 struct dm_bio_prison_cell *cell)
841{
842 size_t pb_data_size = get_per_bio_data_size(cache);
843 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
844
845 BUG_ON(!cell);
846 BUG_ON(pb->all_io_entry);
847
848 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
849}
850
851static bool accountable_bio(struct cache *cache, struct bio *bio)
852{
853 return ((bio->bi_bdev == cache->origin_dev->bdev) &&
854 !(bio->bi_rw & REQ_DISCARD));
855}
856
857static void accounted_begin(struct cache *cache, struct bio *bio)
858{
859 size_t pb_data_size = get_per_bio_data_size(cache);
860 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
861
862 if (accountable_bio(cache, bio)) {
863 pb->len = bio_sectors(bio);
864 iot_io_begin(&cache->origin_tracker, pb->len);
865 }
866}
867
868static void accounted_complete(struct cache *cache, struct bio *bio)
869{
870 size_t pb_data_size = get_per_bio_data_size(cache);
871 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
872
873 iot_io_end(&cache->origin_tracker, pb->len);
874}
875
876static void accounted_request(struct cache *cache, struct bio *bio)
877{
878 accounted_begin(cache, bio);
879 generic_make_request(bio);
880}
881
882static void issue(struct cache *cache, struct bio *bio)
883{
884 unsigned long flags;
885
886 if (!bio_triggers_commit(cache, bio)) {
887 accounted_request(cache, bio);
888 return;
889 }
890
891 /*
892 * Batch together any bios that trigger commits and then issue a
893 * single commit for them in do_worker().
894 */
895 spin_lock_irqsave(&cache->lock, flags);
896 cache->commit_requested = true;
897 bio_list_add(&cache->deferred_flush_bios, bio);
898 spin_unlock_irqrestore(&cache->lock, flags);
899}
900
901static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
902{
903 inc_ds(cache, bio, cell);
904 issue(cache, bio);
905}
906
907static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
908{
909 unsigned long flags;
910
911 spin_lock_irqsave(&cache->lock, flags);
912 bio_list_add(&cache->deferred_writethrough_bios, bio);
913 spin_unlock_irqrestore(&cache->lock, flags);
914
915 wake_worker(cache);
916}
917
918static void writethrough_endio(struct bio *bio)
919{
920 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
921
922 dm_unhook_bio(&pb->hook_info, bio);
923
924 if (bio->bi_error) {
925 bio_endio(bio);
926 return;
927 }
928
929 dm_bio_restore(&pb->bio_details, bio);
930 remap_to_cache(pb->cache, bio, pb->cblock);
931
932 /*
933 * We can't issue this bio directly, since we're in interrupt
934 * context. So it gets put on a bio list for processing by the
935 * worker thread.
936 */
937 defer_writethrough_bio(pb->cache, bio);
938}
939
940/*
941 * When running in writethrough mode we need to send writes to clean blocks
942 * to both the cache and origin devices. In future we'd like to clone the
943 * bio and send them in parallel, but for now we're doing them in
944 * series as this is easier.
945 */
946static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
947 dm_oblock_t oblock, dm_cblock_t cblock)
948{
949 struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
950
951 pb->cache = cache;
952 pb->cblock = cblock;
953 dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
954 dm_bio_record(&pb->bio_details, bio);
955
956 remap_to_origin_clear_discard(pb->cache, bio, oblock);
957}
958
959/*----------------------------------------------------------------
960 * Failure modes
961 *--------------------------------------------------------------*/
962static enum cache_metadata_mode get_cache_mode(struct cache *cache)
963{
964 return cache->features.mode;
965}
966
967static const char *cache_device_name(struct cache *cache)
968{
969 return dm_device_name(dm_table_get_md(cache->ti->table));
970}
971
972static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
973{
974 const char *descs[] = {
975 "write",
976 "read-only",
977 "fail"
978 };
979
980 dm_table_event(cache->ti->table);
981 DMINFO("%s: switching cache to %s mode",
982 cache_device_name(cache), descs[(int)mode]);
983}
984
985static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
986{
987 bool needs_check;
988 enum cache_metadata_mode old_mode = get_cache_mode(cache);
989
990 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
991 DMERR("unable to read needs_check flag, setting failure mode");
992 new_mode = CM_FAIL;
993 }
994
995 if (new_mode == CM_WRITE && needs_check) {
996 DMERR("%s: unable to switch cache to write mode until repaired.",
997 cache_device_name(cache));
998 if (old_mode != new_mode)
999 new_mode = old_mode;
1000 else
1001 new_mode = CM_READ_ONLY;
1002 }
1003
1004 /* Never move out of fail mode */
1005 if (old_mode == CM_FAIL)
1006 new_mode = CM_FAIL;
1007
1008 switch (new_mode) {
1009 case CM_FAIL:
1010 case CM_READ_ONLY:
1011 dm_cache_metadata_set_read_only(cache->cmd);
1012 break;
1013
1014 case CM_WRITE:
1015 dm_cache_metadata_set_read_write(cache->cmd);
1016 break;
1017 }
1018
1019 cache->features.mode = new_mode;
1020
1021 if (new_mode != old_mode)
1022 notify_mode_switch(cache, new_mode);
1023}
1024
1025static void abort_transaction(struct cache *cache)
1026{
1027 const char *dev_name = cache_device_name(cache);
1028
1029 if (get_cache_mode(cache) >= CM_READ_ONLY)
1030 return;
1031
1032 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1033 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1034 set_cache_mode(cache, CM_FAIL);
1035 }
1036
1037 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1038 if (dm_cache_metadata_abort(cache->cmd)) {
1039 DMERR("%s: failed to abort metadata transaction", dev_name);
1040 set_cache_mode(cache, CM_FAIL);
1041 }
1042}
1043
1044static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1045{
1046 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1047 cache_device_name(cache), op, r);
1048 abort_transaction(cache);
1049 set_cache_mode(cache, CM_READ_ONLY);
1050}
1051
1052/*----------------------------------------------------------------
1053 * Migration processing
1054 *
1055 * Migration covers moving data from the origin device to the cache, or
1056 * vice versa.
1057 *--------------------------------------------------------------*/
1058static void inc_io_migrations(struct cache *cache)
1059{
1060 atomic_inc(&cache->nr_io_migrations);
1061}
1062
1063static void dec_io_migrations(struct cache *cache)
1064{
1065 atomic_dec(&cache->nr_io_migrations);
1066}
1067
1068static bool discard_or_flush(struct bio *bio)
1069{
1070 return bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD);
1071}
1072
1073static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1074{
1075 if (discard_or_flush(cell->holder)) {
1076 /*
1077 * We have to handle these bios individually.
1078 */
1079 dm_cell_release(cache->prison, cell, &cache->deferred_bios);
1080 free_prison_cell(cache, cell);
1081 } else
1082 list_add_tail(&cell->user_list, &cache->deferred_cells);
1083}
1084
1085static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1086{
1087 unsigned long flags;
1088
1089 if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1090 /*
1091 * There was no prisoner to promote to holder, the
1092 * cell has been released.
1093 */
1094 free_prison_cell(cache, cell);
1095 return;
1096 }
1097
1098 spin_lock_irqsave(&cache->lock, flags);
1099 __cell_defer(cache, cell);
1100 spin_unlock_irqrestore(&cache->lock, flags);
1101
1102 wake_worker(cache);
1103}
1104
1105static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1106{
1107 dm_cell_error(cache->prison, cell, err);
1108 free_prison_cell(cache, cell);
1109}
1110
1111static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1112{
1113 cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1114}
1115
1116static void free_io_migration(struct dm_cache_migration *mg)
1117{
1118 struct cache *cache = mg->cache;
1119
1120 dec_io_migrations(cache);
1121 free_migration(mg);
1122 wake_worker(cache);
1123}
1124
1125static void migration_failure(struct dm_cache_migration *mg)
1126{
1127 struct cache *cache = mg->cache;
1128 const char *dev_name = cache_device_name(cache);
1129
1130 if (mg->writeback) {
1131 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1132 set_dirty(cache, mg->old_oblock, mg->cblock);
1133 cell_defer(cache, mg->old_ocell, false);
1134
1135 } else if (mg->demote) {
1136 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1137 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1138
1139 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1140 if (mg->promote)
1141 cell_defer(cache, mg->new_ocell, true);
1142 } else {
1143 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1144 policy_remove_mapping(cache->policy, mg->new_oblock);
1145 cell_defer(cache, mg->new_ocell, true);
1146 }
1147
1148 free_io_migration(mg);
1149}
1150
1151static void migration_success_pre_commit(struct dm_cache_migration *mg)
1152{
1153 int r;
1154 unsigned long flags;
1155 struct cache *cache = mg->cache;
1156
1157 if (mg->writeback) {
1158 clear_dirty(cache, mg->old_oblock, mg->cblock);
1159 cell_defer(cache, mg->old_ocell, false);
1160 free_io_migration(mg);
1161 return;
1162
1163 } else if (mg->demote) {
1164 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1165 if (r) {
1166 DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1167 cache_device_name(cache));
1168 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1169 policy_force_mapping(cache->policy, mg->new_oblock,
1170 mg->old_oblock);
1171 if (mg->promote)
1172 cell_defer(cache, mg->new_ocell, true);
1173 free_io_migration(mg);
1174 return;
1175 }
1176 } else {
1177 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1178 if (r) {
1179 DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1180 cache_device_name(cache));
1181 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1182 policy_remove_mapping(cache->policy, mg->new_oblock);
1183 free_io_migration(mg);
1184 return;
1185 }
1186 }
1187
1188 spin_lock_irqsave(&cache->lock, flags);
1189 list_add_tail(&mg->list, &cache->need_commit_migrations);
1190 cache->commit_requested = true;
1191 spin_unlock_irqrestore(&cache->lock, flags);
1192}
1193
1194static void migration_success_post_commit(struct dm_cache_migration *mg)
1195{
1196 unsigned long flags;
1197 struct cache *cache = mg->cache;
1198
1199 if (mg->writeback) {
1200 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1201 cache_device_name(cache));
1202 return;
1203
1204 } else if (mg->demote) {
1205 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1206
1207 if (mg->promote) {
1208 mg->demote = false;
1209
1210 spin_lock_irqsave(&cache->lock, flags);
1211 list_add_tail(&mg->list, &cache->quiesced_migrations);
1212 spin_unlock_irqrestore(&cache->lock, flags);
1213
1214 } else {
1215 if (mg->invalidate)
1216 policy_remove_mapping(cache->policy, mg->old_oblock);
1217 free_io_migration(mg);
1218 }
1219
1220 } else {
1221 if (mg->requeue_holder) {
1222 clear_dirty(cache, mg->new_oblock, mg->cblock);
1223 cell_defer(cache, mg->new_ocell, true);
1224 } else {
1225 /*
1226 * The block was promoted via an overwrite, so it's dirty.
1227 */
1228 set_dirty(cache, mg->new_oblock, mg->cblock);
1229 bio_endio(mg->new_ocell->holder);
1230 cell_defer(cache, mg->new_ocell, false);
1231 }
1232 free_io_migration(mg);
1233 }
1234}
1235
1236static void copy_complete(int read_err, unsigned long write_err, void *context)
1237{
1238 unsigned long flags;
1239 struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1240 struct cache *cache = mg->cache;
1241
1242 if (read_err || write_err)
1243 mg->err = true;
1244
1245 spin_lock_irqsave(&cache->lock, flags);
1246 list_add_tail(&mg->list, &cache->completed_migrations);
1247 spin_unlock_irqrestore(&cache->lock, flags);
1248
1249 wake_worker(cache);
1250}
1251
1252static void issue_copy(struct dm_cache_migration *mg)
1253{
1254 int r;
1255 struct dm_io_region o_region, c_region;
1256 struct cache *cache = mg->cache;
1257 sector_t cblock = from_cblock(mg->cblock);
1258
1259 o_region.bdev = cache->origin_dev->bdev;
1260 o_region.count = cache->sectors_per_block;
1261
1262 c_region.bdev = cache->cache_dev->bdev;
1263 c_region.sector = cblock * cache->sectors_per_block;
1264 c_region.count = cache->sectors_per_block;
1265
1266 if (mg->writeback || mg->demote) {
1267 /* demote */
1268 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1269 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1270 } else {
1271 /* promote */
1272 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1273 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1274 }
1275
1276 if (r < 0) {
1277 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1278 migration_failure(mg);
1279 }
1280}
1281
1282static void overwrite_endio(struct bio *bio)
1283{
1284 struct dm_cache_migration *mg = bio->bi_private;
1285 struct cache *cache = mg->cache;
1286 size_t pb_data_size = get_per_bio_data_size(cache);
1287 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1288 unsigned long flags;
1289
1290 dm_unhook_bio(&pb->hook_info, bio);
1291
1292 if (bio->bi_error)
1293 mg->err = true;
1294
1295 mg->requeue_holder = false;
1296
1297 spin_lock_irqsave(&cache->lock, flags);
1298 list_add_tail(&mg->list, &cache->completed_migrations);
1299 spin_unlock_irqrestore(&cache->lock, flags);
1300
1301 wake_worker(cache);
1302}
1303
1304static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1305{
1306 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1307 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1308
1309 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1310 remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1311
1312 /*
1313 * No need to inc_ds() here, since the cell will be held for the
1314 * duration of the io.
1315 */
1316 accounted_request(mg->cache, bio);
1317}
1318
1319static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1320{
1321 return (bio_data_dir(bio) == WRITE) &&
1322 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1323}
1324
1325static void avoid_copy(struct dm_cache_migration *mg)
1326{
1327 atomic_inc(&mg->cache->stats.copies_avoided);
1328 migration_success_pre_commit(mg);
1329}
1330
1331static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1332 dm_dblock_t *b, dm_dblock_t *e)
1333{
1334 sector_t sb = bio->bi_iter.bi_sector;
1335 sector_t se = bio_end_sector(bio);
1336
1337 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1338
1339 if (se - sb < cache->discard_block_size)
1340 *e = *b;
1341 else
1342 *e = to_dblock(block_div(se, cache->discard_block_size));
1343}
1344
1345static void issue_discard(struct dm_cache_migration *mg)
1346{
1347 dm_dblock_t b, e;
1348 struct bio *bio = mg->new_ocell->holder;
1349 struct cache *cache = mg->cache;
1350
1351 calc_discard_block_range(cache, bio, &b, &e);
1352 while (b != e) {
1353 set_discard(cache, b);
1354 b = to_dblock(from_dblock(b) + 1);
1355 }
1356
1357 bio_endio(bio);
1358 cell_defer(cache, mg->new_ocell, false);
1359 free_migration(mg);
1360 wake_worker(cache);
1361}
1362
1363static void issue_copy_or_discard(struct dm_cache_migration *mg)
1364{
1365 bool avoid;
1366 struct cache *cache = mg->cache;
1367
1368 if (mg->discard) {
1369 issue_discard(mg);
1370 return;
1371 }
1372
1373 if (mg->writeback || mg->demote)
1374 avoid = !is_dirty(cache, mg->cblock) ||
1375 is_discarded_oblock(cache, mg->old_oblock);
1376 else {
1377 struct bio *bio = mg->new_ocell->holder;
1378
1379 avoid = is_discarded_oblock(cache, mg->new_oblock);
1380
1381 if (writeback_mode(&cache->features) &&
1382 !avoid && bio_writes_complete_block(cache, bio)) {
1383 issue_overwrite(mg, bio);
1384 return;
1385 }
1386 }
1387
1388 avoid ? avoid_copy(mg) : issue_copy(mg);
1389}
1390
1391static void complete_migration(struct dm_cache_migration *mg)
1392{
1393 if (mg->err)
1394 migration_failure(mg);
1395 else
1396 migration_success_pre_commit(mg);
1397}
1398
1399static void process_migrations(struct cache *cache, struct list_head *head,
1400 void (*fn)(struct dm_cache_migration *))
1401{
1402 unsigned long flags;
1403 struct list_head list;
1404 struct dm_cache_migration *mg, *tmp;
1405
1406 INIT_LIST_HEAD(&list);
1407 spin_lock_irqsave(&cache->lock, flags);
1408 list_splice_init(head, &list);
1409 spin_unlock_irqrestore(&cache->lock, flags);
1410
1411 list_for_each_entry_safe(mg, tmp, &list, list)
1412 fn(mg);
1413}
1414
1415static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1416{
1417 list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1418}
1419
1420static void queue_quiesced_migration(struct dm_cache_migration *mg)
1421{
1422 unsigned long flags;
1423 struct cache *cache = mg->cache;
1424
1425 spin_lock_irqsave(&cache->lock, flags);
1426 __queue_quiesced_migration(mg);
1427 spin_unlock_irqrestore(&cache->lock, flags);
1428
1429 wake_worker(cache);
1430}
1431
1432static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1433{
1434 unsigned long flags;
1435 struct dm_cache_migration *mg, *tmp;
1436
1437 spin_lock_irqsave(&cache->lock, flags);
1438 list_for_each_entry_safe(mg, tmp, work, list)
1439 __queue_quiesced_migration(mg);
1440 spin_unlock_irqrestore(&cache->lock, flags);
1441
1442 wake_worker(cache);
1443}
1444
1445static void check_for_quiesced_migrations(struct cache *cache,
1446 struct per_bio_data *pb)
1447{
1448 struct list_head work;
1449
1450 if (!pb->all_io_entry)
1451 return;
1452
1453 INIT_LIST_HEAD(&work);
1454 dm_deferred_entry_dec(pb->all_io_entry, &work);
1455
1456 if (!list_empty(&work))
1457 queue_quiesced_migrations(cache, &work);
1458}
1459
1460static void quiesce_migration(struct dm_cache_migration *mg)
1461{
1462 if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1463 queue_quiesced_migration(mg);
1464}
1465
1466static void promote(struct cache *cache, struct prealloc *structs,
1467 dm_oblock_t oblock, dm_cblock_t cblock,
1468 struct dm_bio_prison_cell *cell)
1469{
1470 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1471
1472 mg->err = false;
1473 mg->discard = false;
1474 mg->writeback = false;
1475 mg->demote = false;
1476 mg->promote = true;
1477 mg->requeue_holder = true;
1478 mg->invalidate = false;
1479 mg->cache = cache;
1480 mg->new_oblock = oblock;
1481 mg->cblock = cblock;
1482 mg->old_ocell = NULL;
1483 mg->new_ocell = cell;
1484 mg->start_jiffies = jiffies;
1485
1486 inc_io_migrations(cache);
1487 quiesce_migration(mg);
1488}
1489
1490static void writeback(struct cache *cache, struct prealloc *structs,
1491 dm_oblock_t oblock, dm_cblock_t cblock,
1492 struct dm_bio_prison_cell *cell)
1493{
1494 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1495
1496 mg->err = false;
1497 mg->discard = false;
1498 mg->writeback = true;
1499 mg->demote = false;
1500 mg->promote = false;
1501 mg->requeue_holder = true;
1502 mg->invalidate = false;
1503 mg->cache = cache;
1504 mg->old_oblock = oblock;
1505 mg->cblock = cblock;
1506 mg->old_ocell = cell;
1507 mg->new_ocell = NULL;
1508 mg->start_jiffies = jiffies;
1509
1510 inc_io_migrations(cache);
1511 quiesce_migration(mg);
1512}
1513
1514static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1515 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1516 dm_cblock_t cblock,
1517 struct dm_bio_prison_cell *old_ocell,
1518 struct dm_bio_prison_cell *new_ocell)
1519{
1520 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1521
1522 mg->err = false;
1523 mg->discard = false;
1524 mg->writeback = false;
1525 mg->demote = true;
1526 mg->promote = true;
1527 mg->requeue_holder = true;
1528 mg->invalidate = false;
1529 mg->cache = cache;
1530 mg->old_oblock = old_oblock;
1531 mg->new_oblock = new_oblock;
1532 mg->cblock = cblock;
1533 mg->old_ocell = old_ocell;
1534 mg->new_ocell = new_ocell;
1535 mg->start_jiffies = jiffies;
1536
1537 inc_io_migrations(cache);
1538 quiesce_migration(mg);
1539}
1540
1541/*
1542 * Invalidate a cache entry. No writeback occurs; any changes in the cache
1543 * block are thrown away.
1544 */
1545static void invalidate(struct cache *cache, struct prealloc *structs,
1546 dm_oblock_t oblock, dm_cblock_t cblock,
1547 struct dm_bio_prison_cell *cell)
1548{
1549 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1550
1551 mg->err = false;
1552 mg->discard = false;
1553 mg->writeback = false;
1554 mg->demote = true;
1555 mg->promote = false;
1556 mg->requeue_holder = true;
1557 mg->invalidate = true;
1558 mg->cache = cache;
1559 mg->old_oblock = oblock;
1560 mg->cblock = cblock;
1561 mg->old_ocell = cell;
1562 mg->new_ocell = NULL;
1563 mg->start_jiffies = jiffies;
1564
1565 inc_io_migrations(cache);
1566 quiesce_migration(mg);
1567}
1568
1569static void discard(struct cache *cache, struct prealloc *structs,
1570 struct dm_bio_prison_cell *cell)
1571{
1572 struct dm_cache_migration *mg = prealloc_get_migration(structs);
1573
1574 mg->err = false;
1575 mg->discard = true;
1576 mg->writeback = false;
1577 mg->demote = false;
1578 mg->promote = false;
1579 mg->requeue_holder = false;
1580 mg->invalidate = false;
1581 mg->cache = cache;
1582 mg->old_ocell = NULL;
1583 mg->new_ocell = cell;
1584 mg->start_jiffies = jiffies;
1585
1586 quiesce_migration(mg);
1587}
1588
1589/*----------------------------------------------------------------
1590 * bio processing
1591 *--------------------------------------------------------------*/
1592static void defer_bio(struct cache *cache, struct bio *bio)
1593{
1594 unsigned long flags;
1595
1596 spin_lock_irqsave(&cache->lock, flags);
1597 bio_list_add(&cache->deferred_bios, bio);
1598 spin_unlock_irqrestore(&cache->lock, flags);
1599
1600 wake_worker(cache);
1601}
1602
1603static void process_flush_bio(struct cache *cache, struct bio *bio)
1604{
1605 size_t pb_data_size = get_per_bio_data_size(cache);
1606 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1607
1608 BUG_ON(bio->bi_iter.bi_size);
1609 if (!pb->req_nr)
1610 remap_to_origin(cache, bio);
1611 else
1612 remap_to_cache(cache, bio, 0);
1613
1614 /*
1615 * REQ_FLUSH is not directed at any particular block so we don't
1616 * need to inc_ds(). REQ_FUA's are split into a write + REQ_FLUSH
1617 * by dm-core.
1618 */
1619 issue(cache, bio);
1620}
1621
1622static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1623 struct bio *bio)
1624{
1625 int r;
1626 dm_dblock_t b, e;
1627 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1628
1629 calc_discard_block_range(cache, bio, &b, &e);
1630 if (b == e) {
1631 bio_endio(bio);
1632 return;
1633 }
1634
1635 cell_prealloc = prealloc_get_cell(structs);
1636 r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1637 (cell_free_fn) prealloc_put_cell,
1638 structs, &new_ocell);
1639 if (r > 0)
1640 return;
1641
1642 discard(cache, structs, new_ocell);
1643}
1644
1645static bool spare_migration_bandwidth(struct cache *cache)
1646{
1647 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1648 cache->sectors_per_block;
1649 return current_volume < cache->migration_threshold;
1650}
1651
1652static void inc_hit_counter(struct cache *cache, struct bio *bio)
1653{
1654 atomic_inc(bio_data_dir(bio) == READ ?
1655 &cache->stats.read_hit : &cache->stats.write_hit);
1656}
1657
1658static void inc_miss_counter(struct cache *cache, struct bio *bio)
1659{
1660 atomic_inc(bio_data_dir(bio) == READ ?
1661 &cache->stats.read_miss : &cache->stats.write_miss);
1662}
1663
1664/*----------------------------------------------------------------*/
1665
1666struct inc_detail {
1667 struct cache *cache;
1668 struct bio_list bios_for_issue;
1669 struct bio_list unhandled_bios;
1670 bool any_writes;
1671};
1672
1673static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1674{
1675 struct bio *bio;
1676 struct inc_detail *detail = context;
1677 struct cache *cache = detail->cache;
1678
1679 inc_ds(cache, cell->holder, cell);
1680 if (bio_data_dir(cell->holder) == WRITE)
1681 detail->any_writes = true;
1682
1683 while ((bio = bio_list_pop(&cell->bios))) {
1684 if (discard_or_flush(bio)) {
1685 bio_list_add(&detail->unhandled_bios, bio);
1686 continue;
1687 }
1688
1689 if (bio_data_dir(bio) == WRITE)
1690 detail->any_writes = true;
1691
1692 bio_list_add(&detail->bios_for_issue, bio);
1693 inc_ds(cache, bio, cell);
1694 }
1695}
1696
1697// FIXME: refactor these two
1698static void remap_cell_to_origin_clear_discard(struct cache *cache,
1699 struct dm_bio_prison_cell *cell,
1700 dm_oblock_t oblock, bool issue_holder)
1701{
1702 struct bio *bio;
1703 unsigned long flags;
1704 struct inc_detail detail;
1705
1706 detail.cache = cache;
1707 bio_list_init(&detail.bios_for_issue);
1708 bio_list_init(&detail.unhandled_bios);
1709 detail.any_writes = false;
1710
1711 spin_lock_irqsave(&cache->lock, flags);
1712 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1713 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1714 spin_unlock_irqrestore(&cache->lock, flags);
1715
1716 remap_to_origin(cache, cell->holder);
1717 if (issue_holder)
1718 issue(cache, cell->holder);
1719 else
1720 accounted_begin(cache, cell->holder);
1721
1722 if (detail.any_writes)
1723 clear_discard(cache, oblock_to_dblock(cache, oblock));
1724
1725 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1726 remap_to_origin(cache, bio);
1727 issue(cache, bio);
1728 }
1729
1730 free_prison_cell(cache, cell);
1731}
1732
1733static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1734 dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1735{
1736 struct bio *bio;
1737 unsigned long flags;
1738 struct inc_detail detail;
1739
1740 detail.cache = cache;
1741 bio_list_init(&detail.bios_for_issue);
1742 bio_list_init(&detail.unhandled_bios);
1743 detail.any_writes = false;
1744
1745 spin_lock_irqsave(&cache->lock, flags);
1746 dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1747 bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1748 spin_unlock_irqrestore(&cache->lock, flags);
1749
1750 remap_to_cache(cache, cell->holder, cblock);
1751 if (issue_holder)
1752 issue(cache, cell->holder);
1753 else
1754 accounted_begin(cache, cell->holder);
1755
1756 if (detail.any_writes) {
1757 set_dirty(cache, oblock, cblock);
1758 clear_discard(cache, oblock_to_dblock(cache, oblock));
1759 }
1760
1761 while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1762 remap_to_cache(cache, bio, cblock);
1763 issue(cache, bio);
1764 }
1765
1766 free_prison_cell(cache, cell);
1767}
1768
1769/*----------------------------------------------------------------*/
1770
1771struct old_oblock_lock {
1772 struct policy_locker locker;
1773 struct cache *cache;
1774 struct prealloc *structs;
1775 struct dm_bio_prison_cell *cell;
1776};
1777
1778static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1779{
1780 /* This should never be called */
1781 BUG();
1782 return 0;
1783}
1784
1785static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1786{
1787 struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1788 struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1789
1790 return bio_detain(l->cache, b, NULL, cell_prealloc,
1791 (cell_free_fn) prealloc_put_cell,
1792 l->structs, &l->cell);
1793}
1794
1795static void process_cell(struct cache *cache, struct prealloc *structs,
1796 struct dm_bio_prison_cell *new_ocell)
1797{
1798 int r;
1799 bool release_cell = true;
1800 struct bio *bio = new_ocell->holder;
1801 dm_oblock_t block = get_bio_block(cache, bio);
1802 struct policy_result lookup_result;
1803 bool passthrough = passthrough_mode(&cache->features);
1804 bool fast_promotion, can_migrate;
1805 struct old_oblock_lock ool;
1806
1807 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1808 can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1809
1810 ool.locker.fn = cell_locker;
1811 ool.cache = cache;
1812 ool.structs = structs;
1813 ool.cell = NULL;
1814 r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1815 bio, &ool.locker, &lookup_result);
1816
1817 if (r == -EWOULDBLOCK)
1818 /* migration has been denied */
1819 lookup_result.op = POLICY_MISS;
1820
1821 switch (lookup_result.op) {
1822 case POLICY_HIT:
1823 if (passthrough) {
1824 inc_miss_counter(cache, bio);
1825
1826 /*
1827 * Passthrough always maps to the origin,
1828 * invalidating any cache blocks that are written
1829 * to.
1830 */
1831
1832 if (bio_data_dir(bio) == WRITE) {
1833 atomic_inc(&cache->stats.demotion);
1834 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1835 release_cell = false;
1836
1837 } else {
1838 /* FIXME: factor out issue_origin() */
1839 remap_to_origin_clear_discard(cache, bio, block);
1840 inc_and_issue(cache, bio, new_ocell);
1841 }
1842 } else {
1843 inc_hit_counter(cache, bio);
1844
1845 if (bio_data_dir(bio) == WRITE &&
1846 writethrough_mode(&cache->features) &&
1847 !is_dirty(cache, lookup_result.cblock)) {
1848 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1849 inc_and_issue(cache, bio, new_ocell);
1850
1851 } else {
1852 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1853 release_cell = false;
1854 }
1855 }
1856
1857 break;
1858
1859 case POLICY_MISS:
1860 inc_miss_counter(cache, bio);
1861 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1862 release_cell = false;
1863 break;
1864
1865 case POLICY_NEW:
1866 atomic_inc(&cache->stats.promotion);
1867 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1868 release_cell = false;
1869 break;
1870
1871 case POLICY_REPLACE:
1872 atomic_inc(&cache->stats.demotion);
1873 atomic_inc(&cache->stats.promotion);
1874 demote_then_promote(cache, structs, lookup_result.old_oblock,
1875 block, lookup_result.cblock,
1876 ool.cell, new_ocell);
1877 release_cell = false;
1878 break;
1879
1880 default:
1881 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1882 cache_device_name(cache), __func__,
1883 (unsigned) lookup_result.op);
1884 bio_io_error(bio);
1885 }
1886
1887 if (release_cell)
1888 cell_defer(cache, new_ocell, false);
1889}
1890
1891static void process_bio(struct cache *cache, struct prealloc *structs,
1892 struct bio *bio)
1893{
1894 int r;
1895 dm_oblock_t block = get_bio_block(cache, bio);
1896 struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1897
1898 /*
1899 * Check to see if that block is currently migrating.
1900 */
1901 cell_prealloc = prealloc_get_cell(structs);
1902 r = bio_detain(cache, block, bio, cell_prealloc,
1903 (cell_free_fn) prealloc_put_cell,
1904 structs, &new_ocell);
1905 if (r > 0)
1906 return;
1907
1908 process_cell(cache, structs, new_ocell);
1909}
1910
1911static int need_commit_due_to_time(struct cache *cache)
1912{
1913 return jiffies < cache->last_commit_jiffies ||
1914 jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1915}
1916
1917/*
1918 * A non-zero return indicates read_only or fail_io mode.
1919 */
1920static int commit(struct cache *cache, bool clean_shutdown)
1921{
1922 int r;
1923
1924 if (get_cache_mode(cache) >= CM_READ_ONLY)
1925 return -EINVAL;
1926
1927 atomic_inc(&cache->stats.commit_count);
1928 r = dm_cache_commit(cache->cmd, clean_shutdown);
1929 if (r)
1930 metadata_operation_failed(cache, "dm_cache_commit", r);
1931
1932 return r;
1933}
1934
1935static int commit_if_needed(struct cache *cache)
1936{
1937 int r = 0;
1938
1939 if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1940 dm_cache_changed_this_transaction(cache->cmd)) {
1941 r = commit(cache, false);
1942 cache->commit_requested = false;
1943 cache->last_commit_jiffies = jiffies;
1944 }
1945
1946 return r;
1947}
1948
1949static void process_deferred_bios(struct cache *cache)
1950{
1951 bool prealloc_used = false;
1952 unsigned long flags;
1953 struct bio_list bios;
1954 struct bio *bio;
1955 struct prealloc structs;
1956
1957 memset(&structs, 0, sizeof(structs));
1958 bio_list_init(&bios);
1959
1960 spin_lock_irqsave(&cache->lock, flags);
1961 bio_list_merge(&bios, &cache->deferred_bios);
1962 bio_list_init(&cache->deferred_bios);
1963 spin_unlock_irqrestore(&cache->lock, flags);
1964
1965 while (!bio_list_empty(&bios)) {
1966 /*
1967 * If we've got no free migration structs, and processing
1968 * this bio might require one, we pause until there are some
1969 * prepared mappings to process.
1970 */
1971 prealloc_used = true;
1972 if (prealloc_data_structs(cache, &structs)) {
1973 spin_lock_irqsave(&cache->lock, flags);
1974 bio_list_merge(&cache->deferred_bios, &bios);
1975 spin_unlock_irqrestore(&cache->lock, flags);
1976 break;
1977 }
1978
1979 bio = bio_list_pop(&bios);
1980
1981 if (bio->bi_rw & REQ_FLUSH)
1982 process_flush_bio(cache, bio);
1983 else if (bio->bi_rw & REQ_DISCARD)
1984 process_discard_bio(cache, &structs, bio);
1985 else
1986 process_bio(cache, &structs, bio);
1987 }
1988
1989 if (prealloc_used)
1990 prealloc_free_structs(cache, &structs);
1991}
1992
1993static void process_deferred_cells(struct cache *cache)
1994{
1995 bool prealloc_used = false;
1996 unsigned long flags;
1997 struct dm_bio_prison_cell *cell, *tmp;
1998 struct list_head cells;
1999 struct prealloc structs;
2000
2001 memset(&structs, 0, sizeof(structs));
2002
2003 INIT_LIST_HEAD(&cells);
2004
2005 spin_lock_irqsave(&cache->lock, flags);
2006 list_splice_init(&cache->deferred_cells, &cells);
2007 spin_unlock_irqrestore(&cache->lock, flags);
2008
2009 list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2010 /*
2011 * If we've got no free migration structs, and processing
2012 * this bio might require one, we pause until there are some
2013 * prepared mappings to process.
2014 */
2015 prealloc_used = true;
2016 if (prealloc_data_structs(cache, &structs)) {
2017 spin_lock_irqsave(&cache->lock, flags);
2018 list_splice(&cells, &cache->deferred_cells);
2019 spin_unlock_irqrestore(&cache->lock, flags);
2020 break;
2021 }
2022
2023 process_cell(cache, &structs, cell);
2024 }
2025
2026 if (prealloc_used)
2027 prealloc_free_structs(cache, &structs);
2028}
2029
2030static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2031{
2032 unsigned long flags;
2033 struct bio_list bios;
2034 struct bio *bio;
2035
2036 bio_list_init(&bios);
2037
2038 spin_lock_irqsave(&cache->lock, flags);
2039 bio_list_merge(&bios, &cache->deferred_flush_bios);
2040 bio_list_init(&cache->deferred_flush_bios);
2041 spin_unlock_irqrestore(&cache->lock, flags);
2042
2043 /*
2044 * These bios have already been through inc_ds()
2045 */
2046 while ((bio = bio_list_pop(&bios)))
2047 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2048}
2049
2050static void process_deferred_writethrough_bios(struct cache *cache)
2051{
2052 unsigned long flags;
2053 struct bio_list bios;
2054 struct bio *bio;
2055
2056 bio_list_init(&bios);
2057
2058 spin_lock_irqsave(&cache->lock, flags);
2059 bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2060 bio_list_init(&cache->deferred_writethrough_bios);
2061 spin_unlock_irqrestore(&cache->lock, flags);
2062
2063 /*
2064 * These bios have already been through inc_ds()
2065 */
2066 while ((bio = bio_list_pop(&bios)))
2067 accounted_request(cache, bio);
2068}
2069
2070static void writeback_some_dirty_blocks(struct cache *cache)
2071{
2072 bool prealloc_used = false;
2073 dm_oblock_t oblock;
2074 dm_cblock_t cblock;
2075 struct prealloc structs;
2076 struct dm_bio_prison_cell *old_ocell;
2077 bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2078
2079 memset(&structs, 0, sizeof(structs));
2080
2081 while (spare_migration_bandwidth(cache)) {
2082 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2083 break; /* no work to do */
2084
2085 prealloc_used = true;
2086 if (prealloc_data_structs(cache, &structs) ||
2087 get_cell(cache, oblock, &structs, &old_ocell)) {
2088 policy_set_dirty(cache->policy, oblock);
2089 break;
2090 }
2091
2092 writeback(cache, &structs, oblock, cblock, old_ocell);
2093 }
2094
2095 if (prealloc_used)
2096 prealloc_free_structs(cache, &structs);
2097}
2098
2099/*----------------------------------------------------------------
2100 * Invalidations.
2101 * Dropping something from the cache *without* writing back.
2102 *--------------------------------------------------------------*/
2103
2104static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2105{
2106 int r = 0;
2107 uint64_t begin = from_cblock(req->cblocks->begin);
2108 uint64_t end = from_cblock(req->cblocks->end);
2109
2110 while (begin != end) {
2111 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2112 if (!r) {
2113 r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2114 if (r) {
2115 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2116 break;
2117 }
2118
2119 } else if (r == -ENODATA) {
2120 /* harmless, already unmapped */
2121 r = 0;
2122
2123 } else {
2124 DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2125 break;
2126 }
2127
2128 begin++;
2129 }
2130
2131 cache->commit_requested = true;
2132
2133 req->err = r;
2134 atomic_set(&req->complete, 1);
2135
2136 wake_up(&req->result_wait);
2137}
2138
2139static void process_invalidation_requests(struct cache *cache)
2140{
2141 struct list_head list;
2142 struct invalidation_request *req, *tmp;
2143
2144 INIT_LIST_HEAD(&list);
2145 spin_lock(&cache->invalidation_lock);
2146 list_splice_init(&cache->invalidation_requests, &list);
2147 spin_unlock(&cache->invalidation_lock);
2148
2149 list_for_each_entry_safe (req, tmp, &list, list)
2150 process_invalidation_request(cache, req);
2151}
2152
2153/*----------------------------------------------------------------
2154 * Main worker loop
2155 *--------------------------------------------------------------*/
2156static bool is_quiescing(struct cache *cache)
2157{
2158 return atomic_read(&cache->quiescing);
2159}
2160
2161static void ack_quiescing(struct cache *cache)
2162{
2163 if (is_quiescing(cache)) {
2164 atomic_inc(&cache->quiescing_ack);
2165 wake_up(&cache->quiescing_wait);
2166 }
2167}
2168
2169static void wait_for_quiescing_ack(struct cache *cache)
2170{
2171 wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2172}
2173
2174static void start_quiescing(struct cache *cache)
2175{
2176 atomic_inc(&cache->quiescing);
2177 wait_for_quiescing_ack(cache);
2178}
2179
2180static void stop_quiescing(struct cache *cache)
2181{
2182 atomic_set(&cache->quiescing, 0);
2183 atomic_set(&cache->quiescing_ack, 0);
2184}
2185
2186static void wait_for_migrations(struct cache *cache)
2187{
2188 wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2189}
2190
2191static void stop_worker(struct cache *cache)
2192{
2193 cancel_delayed_work(&cache->waker);
2194 flush_workqueue(cache->wq);
2195}
2196
2197static void requeue_deferred_cells(struct cache *cache)
2198{
2199 unsigned long flags;
2200 struct list_head cells;
2201 struct dm_bio_prison_cell *cell, *tmp;
2202
2203 INIT_LIST_HEAD(&cells);
2204 spin_lock_irqsave(&cache->lock, flags);
2205 list_splice_init(&cache->deferred_cells, &cells);
2206 spin_unlock_irqrestore(&cache->lock, flags);
2207
2208 list_for_each_entry_safe(cell, tmp, &cells, user_list)
2209 cell_requeue(cache, cell);
2210}
2211
2212static void requeue_deferred_bios(struct cache *cache)
2213{
2214 struct bio *bio;
2215 struct bio_list bios;
2216
2217 bio_list_init(&bios);
2218 bio_list_merge(&bios, &cache->deferred_bios);
2219 bio_list_init(&cache->deferred_bios);
2220
2221 while ((bio = bio_list_pop(&bios))) {
2222 bio->bi_error = DM_ENDIO_REQUEUE;
2223 bio_endio(bio);
2224 }
2225}
2226
2227static int more_work(struct cache *cache)
2228{
2229 if (is_quiescing(cache))
2230 return !list_empty(&cache->quiesced_migrations) ||
2231 !list_empty(&cache->completed_migrations) ||
2232 !list_empty(&cache->need_commit_migrations);
2233 else
2234 return !bio_list_empty(&cache->deferred_bios) ||
2235 !list_empty(&cache->deferred_cells) ||
2236 !bio_list_empty(&cache->deferred_flush_bios) ||
2237 !bio_list_empty(&cache->deferred_writethrough_bios) ||
2238 !list_empty(&cache->quiesced_migrations) ||
2239 !list_empty(&cache->completed_migrations) ||
2240 !list_empty(&cache->need_commit_migrations) ||
2241 cache->invalidate;
2242}
2243
2244static void do_worker(struct work_struct *ws)
2245{
2246 struct cache *cache = container_of(ws, struct cache, worker);
2247
2248 do {
2249 if (!is_quiescing(cache)) {
2250 writeback_some_dirty_blocks(cache);
2251 process_deferred_writethrough_bios(cache);
2252 process_deferred_bios(cache);
2253 process_deferred_cells(cache);
2254 process_invalidation_requests(cache);
2255 }
2256
2257 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2258 process_migrations(cache, &cache->completed_migrations, complete_migration);
2259
2260 if (commit_if_needed(cache)) {
2261 process_deferred_flush_bios(cache, false);
2262 process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2263 } else {
2264 process_deferred_flush_bios(cache, true);
2265 process_migrations(cache, &cache->need_commit_migrations,
2266 migration_success_post_commit);
2267 }
2268
2269 ack_quiescing(cache);
2270
2271 } while (more_work(cache));
2272}
2273
2274/*
2275 * We want to commit periodically so that not too much
2276 * unwritten metadata builds up.
2277 */
2278static void do_waker(struct work_struct *ws)
2279{
2280 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2281 policy_tick(cache->policy, true);
2282 wake_worker(cache);
2283 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2284}
2285
2286/*----------------------------------------------------------------*/
2287
2288static int is_congested(struct dm_dev *dev, int bdi_bits)
2289{
2290 struct request_queue *q = bdev_get_queue(dev->bdev);
2291 return bdi_congested(&q->backing_dev_info, bdi_bits);
2292}
2293
2294static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2295{
2296 struct cache *cache = container_of(cb, struct cache, callbacks);
2297
2298 return is_congested(cache->origin_dev, bdi_bits) ||
2299 is_congested(cache->cache_dev, bdi_bits);
2300}
2301
2302/*----------------------------------------------------------------
2303 * Target methods
2304 *--------------------------------------------------------------*/
2305
2306/*
2307 * This function gets called on the error paths of the constructor, so we
2308 * have to cope with a partially initialised struct.
2309 */
2310static void destroy(struct cache *cache)
2311{
2312 unsigned i;
2313
2314 mempool_destroy(cache->migration_pool);
2315
2316 if (cache->all_io_ds)
2317 dm_deferred_set_destroy(cache->all_io_ds);
2318
2319 if (cache->prison)
2320 dm_bio_prison_destroy(cache->prison);
2321
2322 if (cache->wq)
2323 destroy_workqueue(cache->wq);
2324
2325 if (cache->dirty_bitset)
2326 free_bitset(cache->dirty_bitset);
2327
2328 if (cache->discard_bitset)
2329 free_bitset(cache->discard_bitset);
2330
2331 if (cache->copier)
2332 dm_kcopyd_client_destroy(cache->copier);
2333
2334 if (cache->cmd)
2335 dm_cache_metadata_close(cache->cmd);
2336
2337 if (cache->metadata_dev)
2338 dm_put_device(cache->ti, cache->metadata_dev);
2339
2340 if (cache->origin_dev)
2341 dm_put_device(cache->ti, cache->origin_dev);
2342
2343 if (cache->cache_dev)
2344 dm_put_device(cache->ti, cache->cache_dev);
2345
2346 if (cache->policy)
2347 dm_cache_policy_destroy(cache->policy);
2348
2349 for (i = 0; i < cache->nr_ctr_args ; i++)
2350 kfree(cache->ctr_args[i]);
2351 kfree(cache->ctr_args);
2352
2353 kfree(cache);
2354}
2355
2356static void cache_dtr(struct dm_target *ti)
2357{
2358 struct cache *cache = ti->private;
2359
2360 destroy(cache);
2361}
2362
2363static sector_t get_dev_size(struct dm_dev *dev)
2364{
2365 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2366}
2367
2368/*----------------------------------------------------------------*/
2369
2370/*
2371 * Construct a cache device mapping.
2372 *
2373 * cache <metadata dev> <cache dev> <origin dev> <block size>
2374 * <#feature args> [<feature arg>]*
2375 * <policy> <#policy args> [<policy arg>]*
2376 *
2377 * metadata dev : fast device holding the persistent metadata
2378 * cache dev : fast device holding cached data blocks
2379 * origin dev : slow device holding original data blocks
2380 * block size : cache unit size in sectors
2381 *
2382 * #feature args : number of feature arguments passed
2383 * feature args : writethrough. (The default is writeback.)
2384 *
2385 * policy : the replacement policy to use
2386 * #policy args : an even number of policy arguments corresponding
2387 * to key/value pairs passed to the policy
2388 * policy args : key/value pairs passed to the policy
2389 * E.g. 'sequential_threshold 1024'
2390 * See cache-policies.txt for details.
2391 *
2392 * Optional feature arguments are:
2393 * writethrough : write through caching that prohibits cache block
2394 * content from being different from origin block content.
2395 * Without this argument, the default behaviour is to write
2396 * back cache block contents later for performance reasons,
2397 * so they may differ from the corresponding origin blocks.
2398 */
2399struct cache_args {
2400 struct dm_target *ti;
2401
2402 struct dm_dev *metadata_dev;
2403
2404 struct dm_dev *cache_dev;
2405 sector_t cache_sectors;
2406
2407 struct dm_dev *origin_dev;
2408 sector_t origin_sectors;
2409
2410 uint32_t block_size;
2411
2412 const char *policy_name;
2413 int policy_argc;
2414 const char **policy_argv;
2415
2416 struct cache_features features;
2417};
2418
2419static void destroy_cache_args(struct cache_args *ca)
2420{
2421 if (ca->metadata_dev)
2422 dm_put_device(ca->ti, ca->metadata_dev);
2423
2424 if (ca->cache_dev)
2425 dm_put_device(ca->ti, ca->cache_dev);
2426
2427 if (ca->origin_dev)
2428 dm_put_device(ca->ti, ca->origin_dev);
2429
2430 kfree(ca);
2431}
2432
2433static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2434{
2435 if (!as->argc) {
2436 *error = "Insufficient args";
2437 return false;
2438 }
2439
2440 return true;
2441}
2442
2443static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2444 char **error)
2445{
2446 int r;
2447 sector_t metadata_dev_size;
2448 char b[BDEVNAME_SIZE];
2449
2450 if (!at_least_one_arg(as, error))
2451 return -EINVAL;
2452
2453 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2454 &ca->metadata_dev);
2455 if (r) {
2456 *error = "Error opening metadata device";
2457 return r;
2458 }
2459
2460 metadata_dev_size = get_dev_size(ca->metadata_dev);
2461 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2462 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2463 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2464
2465 return 0;
2466}
2467
2468static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2469 char **error)
2470{
2471 int r;
2472
2473 if (!at_least_one_arg(as, error))
2474 return -EINVAL;
2475
2476 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2477 &ca->cache_dev);
2478 if (r) {
2479 *error = "Error opening cache device";
2480 return r;
2481 }
2482 ca->cache_sectors = get_dev_size(ca->cache_dev);
2483
2484 return 0;
2485}
2486
2487static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2488 char **error)
2489{
2490 int r;
2491
2492 if (!at_least_one_arg(as, error))
2493 return -EINVAL;
2494
2495 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2496 &ca->origin_dev);
2497 if (r) {
2498 *error = "Error opening origin device";
2499 return r;
2500 }
2501
2502 ca->origin_sectors = get_dev_size(ca->origin_dev);
2503 if (ca->ti->len > ca->origin_sectors) {
2504 *error = "Device size larger than cached device";
2505 return -EINVAL;
2506 }
2507
2508 return 0;
2509}
2510
2511static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2512 char **error)
2513{
2514 unsigned long block_size;
2515
2516 if (!at_least_one_arg(as, error))
2517 return -EINVAL;
2518
2519 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2520 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2521 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2522 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2523 *error = "Invalid data block size";
2524 return -EINVAL;
2525 }
2526
2527 if (block_size > ca->cache_sectors) {
2528 *error = "Data block size is larger than the cache device";
2529 return -EINVAL;
2530 }
2531
2532 ca->block_size = block_size;
2533
2534 return 0;
2535}
2536
2537static void init_features(struct cache_features *cf)
2538{
2539 cf->mode = CM_WRITE;
2540 cf->io_mode = CM_IO_WRITEBACK;
2541}
2542
2543static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2544 char **error)
2545{
2546 static struct dm_arg _args[] = {
2547 {0, 1, "Invalid number of cache feature arguments"},
2548 };
2549
2550 int r;
2551 unsigned argc;
2552 const char *arg;
2553 struct cache_features *cf = &ca->features;
2554
2555 init_features(cf);
2556
2557 r = dm_read_arg_group(_args, as, &argc, error);
2558 if (r)
2559 return -EINVAL;
2560
2561 while (argc--) {
2562 arg = dm_shift_arg(as);
2563
2564 if (!strcasecmp(arg, "writeback"))
2565 cf->io_mode = CM_IO_WRITEBACK;
2566
2567 else if (!strcasecmp(arg, "writethrough"))
2568 cf->io_mode = CM_IO_WRITETHROUGH;
2569
2570 else if (!strcasecmp(arg, "passthrough"))
2571 cf->io_mode = CM_IO_PASSTHROUGH;
2572
2573 else {
2574 *error = "Unrecognised cache feature requested";
2575 return -EINVAL;
2576 }
2577 }
2578
2579 return 0;
2580}
2581
2582static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2583 char **error)
2584{
2585 static struct dm_arg _args[] = {
2586 {0, 1024, "Invalid number of policy arguments"},
2587 };
2588
2589 int r;
2590
2591 if (!at_least_one_arg(as, error))
2592 return -EINVAL;
2593
2594 ca->policy_name = dm_shift_arg(as);
2595
2596 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2597 if (r)
2598 return -EINVAL;
2599
2600 ca->policy_argv = (const char **)as->argv;
2601 dm_consume_args(as, ca->policy_argc);
2602
2603 return 0;
2604}
2605
2606static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2607 char **error)
2608{
2609 int r;
2610 struct dm_arg_set as;
2611
2612 as.argc = argc;
2613 as.argv = argv;
2614
2615 r = parse_metadata_dev(ca, &as, error);
2616 if (r)
2617 return r;
2618
2619 r = parse_cache_dev(ca, &as, error);
2620 if (r)
2621 return r;
2622
2623 r = parse_origin_dev(ca, &as, error);
2624 if (r)
2625 return r;
2626
2627 r = parse_block_size(ca, &as, error);
2628 if (r)
2629 return r;
2630
2631 r = parse_features(ca, &as, error);
2632 if (r)
2633 return r;
2634
2635 r = parse_policy(ca, &as, error);
2636 if (r)
2637 return r;
2638
2639 return 0;
2640}
2641
2642/*----------------------------------------------------------------*/
2643
2644static struct kmem_cache *migration_cache;
2645
2646#define NOT_CORE_OPTION 1
2647
2648static int process_config_option(struct cache *cache, const char *key, const char *value)
2649{
2650 unsigned long tmp;
2651
2652 if (!strcasecmp(key, "migration_threshold")) {
2653 if (kstrtoul(value, 10, &tmp))
2654 return -EINVAL;
2655
2656 cache->migration_threshold = tmp;
2657 return 0;
2658 }
2659
2660 return NOT_CORE_OPTION;
2661}
2662
2663static int set_config_value(struct cache *cache, const char *key, const char *value)
2664{
2665 int r = process_config_option(cache, key, value);
2666
2667 if (r == NOT_CORE_OPTION)
2668 r = policy_set_config_value(cache->policy, key, value);
2669
2670 if (r)
2671 DMWARN("bad config value for %s: %s", key, value);
2672
2673 return r;
2674}
2675
2676static int set_config_values(struct cache *cache, int argc, const char **argv)
2677{
2678 int r = 0;
2679
2680 if (argc & 1) {
2681 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2682 return -EINVAL;
2683 }
2684
2685 while (argc) {
2686 r = set_config_value(cache, argv[0], argv[1]);
2687 if (r)
2688 break;
2689
2690 argc -= 2;
2691 argv += 2;
2692 }
2693
2694 return r;
2695}
2696
2697static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2698 char **error)
2699{
2700 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2701 cache->cache_size,
2702 cache->origin_sectors,
2703 cache->sectors_per_block);
2704 if (IS_ERR(p)) {
2705 *error = "Error creating cache's policy";
2706 return PTR_ERR(p);
2707 }
2708 cache->policy = p;
2709
2710 return 0;
2711}
2712
2713/*
2714 * We want the discard block size to be at least the size of the cache
2715 * block size and have no more than 2^14 discard blocks across the origin.
2716 */
2717#define MAX_DISCARD_BLOCKS (1 << 14)
2718
2719static bool too_many_discard_blocks(sector_t discard_block_size,
2720 sector_t origin_size)
2721{
2722 (void) sector_div(origin_size, discard_block_size);
2723
2724 return origin_size > MAX_DISCARD_BLOCKS;
2725}
2726
2727static sector_t calculate_discard_block_size(sector_t cache_block_size,
2728 sector_t origin_size)
2729{
2730 sector_t discard_block_size = cache_block_size;
2731
2732 if (origin_size)
2733 while (too_many_discard_blocks(discard_block_size, origin_size))
2734 discard_block_size *= 2;
2735
2736 return discard_block_size;
2737}
2738
2739static void set_cache_size(struct cache *cache, dm_cblock_t size)
2740{
2741 dm_block_t nr_blocks = from_cblock(size);
2742
2743 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2744 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2745 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2746 "Please consider increasing the cache block size to reduce the overall cache block count.",
2747 (unsigned long long) nr_blocks);
2748
2749 cache->cache_size = size;
2750}
2751
2752#define DEFAULT_MIGRATION_THRESHOLD 2048
2753
2754static int cache_create(struct cache_args *ca, struct cache **result)
2755{
2756 int r = 0;
2757 char **error = &ca->ti->error;
2758 struct cache *cache;
2759 struct dm_target *ti = ca->ti;
2760 dm_block_t origin_blocks;
2761 struct dm_cache_metadata *cmd;
2762 bool may_format = ca->features.mode == CM_WRITE;
2763
2764 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2765 if (!cache)
2766 return -ENOMEM;
2767
2768 cache->ti = ca->ti;
2769 ti->private = cache;
2770 ti->num_flush_bios = 2;
2771 ti->flush_supported = true;
2772
2773 ti->num_discard_bios = 1;
2774 ti->discards_supported = true;
2775 ti->discard_zeroes_data_unsupported = true;
2776 ti->split_discard_bios = false;
2777
2778 cache->features = ca->features;
2779 ti->per_io_data_size = get_per_bio_data_size(cache);
2780
2781 cache->callbacks.congested_fn = cache_is_congested;
2782 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2783
2784 cache->metadata_dev = ca->metadata_dev;
2785 cache->origin_dev = ca->origin_dev;
2786 cache->cache_dev = ca->cache_dev;
2787
2788 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2789
2790 /* FIXME: factor out this whole section */
2791 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2792 origin_blocks = block_div(origin_blocks, ca->block_size);
2793 cache->origin_blocks = to_oblock(origin_blocks);
2794
2795 cache->sectors_per_block = ca->block_size;
2796 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2797 r = -EINVAL;
2798 goto bad;
2799 }
2800
2801 if (ca->block_size & (ca->block_size - 1)) {
2802 dm_block_t cache_size = ca->cache_sectors;
2803
2804 cache->sectors_per_block_shift = -1;
2805 cache_size = block_div(cache_size, ca->block_size);
2806 set_cache_size(cache, to_cblock(cache_size));
2807 } else {
2808 cache->sectors_per_block_shift = __ffs(ca->block_size);
2809 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2810 }
2811
2812 r = create_cache_policy(cache, ca, error);
2813 if (r)
2814 goto bad;
2815
2816 cache->policy_nr_args = ca->policy_argc;
2817 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2818
2819 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2820 if (r) {
2821 *error = "Error setting cache policy's config values";
2822 goto bad;
2823 }
2824
2825 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2826 ca->block_size, may_format,
2827 dm_cache_policy_get_hint_size(cache->policy));
2828 if (IS_ERR(cmd)) {
2829 *error = "Error creating metadata object";
2830 r = PTR_ERR(cmd);
2831 goto bad;
2832 }
2833 cache->cmd = cmd;
2834 set_cache_mode(cache, CM_WRITE);
2835 if (get_cache_mode(cache) != CM_WRITE) {
2836 *error = "Unable to get write access to metadata, please check/repair metadata.";
2837 r = -EINVAL;
2838 goto bad;
2839 }
2840
2841 if (passthrough_mode(&cache->features)) {
2842 bool all_clean;
2843
2844 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2845 if (r) {
2846 *error = "dm_cache_metadata_all_clean() failed";
2847 goto bad;
2848 }
2849
2850 if (!all_clean) {
2851 *error = "Cannot enter passthrough mode unless all blocks are clean";
2852 r = -EINVAL;
2853 goto bad;
2854 }
2855 }
2856
2857 spin_lock_init(&cache->lock);
2858 INIT_LIST_HEAD(&cache->deferred_cells);
2859 bio_list_init(&cache->deferred_bios);
2860 bio_list_init(&cache->deferred_flush_bios);
2861 bio_list_init(&cache->deferred_writethrough_bios);
2862 INIT_LIST_HEAD(&cache->quiesced_migrations);
2863 INIT_LIST_HEAD(&cache->completed_migrations);
2864 INIT_LIST_HEAD(&cache->need_commit_migrations);
2865 atomic_set(&cache->nr_allocated_migrations, 0);
2866 atomic_set(&cache->nr_io_migrations, 0);
2867 init_waitqueue_head(&cache->migration_wait);
2868
2869 init_waitqueue_head(&cache->quiescing_wait);
2870 atomic_set(&cache->quiescing, 0);
2871 atomic_set(&cache->quiescing_ack, 0);
2872
2873 r = -ENOMEM;
2874 atomic_set(&cache->nr_dirty, 0);
2875 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2876 if (!cache->dirty_bitset) {
2877 *error = "could not allocate dirty bitset";
2878 goto bad;
2879 }
2880 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2881
2882 cache->discard_block_size =
2883 calculate_discard_block_size(cache->sectors_per_block,
2884 cache->origin_sectors);
2885 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2886 cache->discard_block_size));
2887 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2888 if (!cache->discard_bitset) {
2889 *error = "could not allocate discard bitset";
2890 goto bad;
2891 }
2892 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2893
2894 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2895 if (IS_ERR(cache->copier)) {
2896 *error = "could not create kcopyd client";
2897 r = PTR_ERR(cache->copier);
2898 goto bad;
2899 }
2900
2901 cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2902 if (!cache->wq) {
2903 *error = "could not create workqueue for metadata object";
2904 goto bad;
2905 }
2906 INIT_WORK(&cache->worker, do_worker);
2907 INIT_DELAYED_WORK(&cache->waker, do_waker);
2908 cache->last_commit_jiffies = jiffies;
2909
2910 cache->prison = dm_bio_prison_create();
2911 if (!cache->prison) {
2912 *error = "could not create bio prison";
2913 goto bad;
2914 }
2915
2916 cache->all_io_ds = dm_deferred_set_create();
2917 if (!cache->all_io_ds) {
2918 *error = "could not create all_io deferred set";
2919 goto bad;
2920 }
2921
2922 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2923 migration_cache);
2924 if (!cache->migration_pool) {
2925 *error = "Error creating cache's migration mempool";
2926 goto bad;
2927 }
2928
2929 cache->need_tick_bio = true;
2930 cache->sized = false;
2931 cache->invalidate = false;
2932 cache->commit_requested = false;
2933 cache->loaded_mappings = false;
2934 cache->loaded_discards = false;
2935
2936 load_stats(cache);
2937
2938 atomic_set(&cache->stats.demotion, 0);
2939 atomic_set(&cache->stats.promotion, 0);
2940 atomic_set(&cache->stats.copies_avoided, 0);
2941 atomic_set(&cache->stats.cache_cell_clash, 0);
2942 atomic_set(&cache->stats.commit_count, 0);
2943 atomic_set(&cache->stats.discard_count, 0);
2944
2945 spin_lock_init(&cache->invalidation_lock);
2946 INIT_LIST_HEAD(&cache->invalidation_requests);
2947
2948 iot_init(&cache->origin_tracker);
2949
2950 *result = cache;
2951 return 0;
2952
2953bad:
2954 destroy(cache);
2955 return r;
2956}
2957
2958static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2959{
2960 unsigned i;
2961 const char **copy;
2962
2963 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2964 if (!copy)
2965 return -ENOMEM;
2966 for (i = 0; i < argc; i++) {
2967 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2968 if (!copy[i]) {
2969 while (i--)
2970 kfree(copy[i]);
2971 kfree(copy);
2972 return -ENOMEM;
2973 }
2974 }
2975
2976 cache->nr_ctr_args = argc;
2977 cache->ctr_args = copy;
2978
2979 return 0;
2980}
2981
2982static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2983{
2984 int r = -EINVAL;
2985 struct cache_args *ca;
2986 struct cache *cache = NULL;
2987
2988 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2989 if (!ca) {
2990 ti->error = "Error allocating memory for cache";
2991 return -ENOMEM;
2992 }
2993 ca->ti = ti;
2994
2995 r = parse_cache_args(ca, argc, argv, &ti->error);
2996 if (r)
2997 goto out;
2998
2999 r = cache_create(ca, &cache);
3000 if (r)
3001 goto out;
3002
3003 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
3004 if (r) {
3005 destroy(cache);
3006 goto out;
3007 }
3008
3009 ti->private = cache;
3010
3011out:
3012 destroy_cache_args(ca);
3013 return r;
3014}
3015
3016/*----------------------------------------------------------------*/
3017
3018static int cache_map(struct dm_target *ti, struct bio *bio)
3019{
3020 struct cache *cache = ti->private;
3021
3022 int r;
3023 struct dm_bio_prison_cell *cell = NULL;
3024 dm_oblock_t block = get_bio_block(cache, bio);
3025 size_t pb_data_size = get_per_bio_data_size(cache);
3026 bool can_migrate = false;
3027 bool fast_promotion;
3028 struct policy_result lookup_result;
3029 struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3030 struct old_oblock_lock ool;
3031
3032 ool.locker.fn = null_locker;
3033
3034 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3035 /*
3036 * This can only occur if the io goes to a partial block at
3037 * the end of the origin device. We don't cache these.
3038 * Just remap to the origin and carry on.
3039 */
3040 remap_to_origin(cache, bio);
3041 accounted_begin(cache, bio);
3042 return DM_MAPIO_REMAPPED;
3043 }
3044
3045 if (discard_or_flush(bio)) {
3046 defer_bio(cache, bio);
3047 return DM_MAPIO_SUBMITTED;
3048 }
3049
3050 /*
3051 * Check to see if that block is currently migrating.
3052 */
3053 cell = alloc_prison_cell(cache);
3054 if (!cell) {
3055 defer_bio(cache, bio);
3056 return DM_MAPIO_SUBMITTED;
3057 }
3058
3059 r = bio_detain(cache, block, bio, cell,
3060 (cell_free_fn) free_prison_cell,
3061 cache, &cell);
3062 if (r) {
3063 if (r < 0)
3064 defer_bio(cache, bio);
3065
3066 return DM_MAPIO_SUBMITTED;
3067 }
3068
3069 fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3070
3071 r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3072 bio, &ool.locker, &lookup_result);
3073 if (r == -EWOULDBLOCK) {
3074 cell_defer(cache, cell, true);
3075 return DM_MAPIO_SUBMITTED;
3076
3077 } else if (r) {
3078 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3079 cache_device_name(cache), r);
3080 cell_defer(cache, cell, false);
3081 bio_io_error(bio);
3082 return DM_MAPIO_SUBMITTED;
3083 }
3084
3085 r = DM_MAPIO_REMAPPED;
3086 switch (lookup_result.op) {
3087 case POLICY_HIT:
3088 if (passthrough_mode(&cache->features)) {
3089 if (bio_data_dir(bio) == WRITE) {
3090 /*
3091 * We need to invalidate this block, so
3092 * defer for the worker thread.
3093 */
3094 cell_defer(cache, cell, true);
3095 r = DM_MAPIO_SUBMITTED;
3096
3097 } else {
3098 inc_miss_counter(cache, bio);
3099 remap_to_origin_clear_discard(cache, bio, block);
3100 accounted_begin(cache, bio);
3101 inc_ds(cache, bio, cell);
3102 // FIXME: we want to remap hits or misses straight
3103 // away rather than passing over to the worker.
3104 cell_defer(cache, cell, false);
3105 }
3106
3107 } else {
3108 inc_hit_counter(cache, bio);
3109 if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3110 !is_dirty(cache, lookup_result.cblock)) {
3111 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3112 accounted_begin(cache, bio);
3113 inc_ds(cache, bio, cell);
3114 cell_defer(cache, cell, false);
3115
3116 } else
3117 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3118 }
3119 break;
3120
3121 case POLICY_MISS:
3122 inc_miss_counter(cache, bio);
3123 if (pb->req_nr != 0) {
3124 /*
3125 * This is a duplicate writethrough io that is no
3126 * longer needed because the block has been demoted.
3127 */
3128 bio_endio(bio);
3129 // FIXME: remap everything as a miss
3130 cell_defer(cache, cell, false);
3131 r = DM_MAPIO_SUBMITTED;
3132
3133 } else
3134 remap_cell_to_origin_clear_discard(cache, cell, block, false);
3135 break;
3136
3137 default:
3138 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3139 cache_device_name(cache), __func__,
3140 (unsigned) lookup_result.op);
3141 cell_defer(cache, cell, false);
3142 bio_io_error(bio);
3143 r = DM_MAPIO_SUBMITTED;
3144 }
3145
3146 return r;
3147}
3148
3149static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3150{
3151 struct cache *cache = ti->private;
3152 unsigned long flags;
3153 size_t pb_data_size = get_per_bio_data_size(cache);
3154 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3155
3156 if (pb->tick) {
3157 policy_tick(cache->policy, false);
3158
3159 spin_lock_irqsave(&cache->lock, flags);
3160 cache->need_tick_bio = true;
3161 spin_unlock_irqrestore(&cache->lock, flags);
3162 }
3163
3164 check_for_quiesced_migrations(cache, pb);
3165 accounted_complete(cache, bio);
3166
3167 return 0;
3168}
3169
3170static int write_dirty_bitset(struct cache *cache)
3171{
3172 unsigned i, r;
3173
3174 if (get_cache_mode(cache) >= CM_READ_ONLY)
3175 return -EINVAL;
3176
3177 for (i = 0; i < from_cblock(cache->cache_size); i++) {
3178 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3179 is_dirty(cache, to_cblock(i)));
3180 if (r) {
3181 metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3182 return r;
3183 }
3184 }
3185
3186 return 0;
3187}
3188
3189static int write_discard_bitset(struct cache *cache)
3190{
3191 unsigned i, r;
3192
3193 if (get_cache_mode(cache) >= CM_READ_ONLY)
3194 return -EINVAL;
3195
3196 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3197 cache->discard_nr_blocks);
3198 if (r) {
3199 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3200 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3201 return r;
3202 }
3203
3204 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3205 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3206 is_discarded(cache, to_dblock(i)));
3207 if (r) {
3208 metadata_operation_failed(cache, "dm_cache_set_discard", r);
3209 return r;
3210 }
3211 }
3212
3213 return 0;
3214}
3215
3216static int write_hints(struct cache *cache)
3217{
3218 int r;
3219
3220 if (get_cache_mode(cache) >= CM_READ_ONLY)
3221 return -EINVAL;
3222
3223 r = dm_cache_write_hints(cache->cmd, cache->policy);
3224 if (r) {
3225 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3226 return r;
3227 }
3228
3229 return 0;
3230}
3231
3232/*
3233 * returns true on success
3234 */
3235static bool sync_metadata(struct cache *cache)
3236{
3237 int r1, r2, r3, r4;
3238
3239 r1 = write_dirty_bitset(cache);
3240 if (r1)
3241 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3242
3243 r2 = write_discard_bitset(cache);
3244 if (r2)
3245 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3246
3247 save_stats(cache);
3248
3249 r3 = write_hints(cache);
3250 if (r3)
3251 DMERR("%s: could not write hints", cache_device_name(cache));
3252
3253 /*
3254 * If writing the above metadata failed, we still commit, but don't
3255 * set the clean shutdown flag. This will effectively force every
3256 * dirty bit to be set on reload.
3257 */
3258 r4 = commit(cache, !r1 && !r2 && !r3);
3259 if (r4)
3260 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3261
3262 return !r1 && !r2 && !r3 && !r4;
3263}
3264
3265static void cache_postsuspend(struct dm_target *ti)
3266{
3267 struct cache *cache = ti->private;
3268
3269 start_quiescing(cache);
3270 wait_for_migrations(cache);
3271 stop_worker(cache);
3272 requeue_deferred_bios(cache);
3273 requeue_deferred_cells(cache);
3274 stop_quiescing(cache);
3275
3276 if (get_cache_mode(cache) == CM_WRITE)
3277 (void) sync_metadata(cache);
3278}
3279
3280static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3281 bool dirty, uint32_t hint, bool hint_valid)
3282{
3283 int r;
3284 struct cache *cache = context;
3285
3286 r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3287 if (r)
3288 return r;
3289
3290 if (dirty)
3291 set_dirty(cache, oblock, cblock);
3292 else
3293 clear_dirty(cache, oblock, cblock);
3294
3295 return 0;
3296}
3297
3298/*
3299 * The discard block size in the on disk metadata is not
3300 * neccessarily the same as we're currently using. So we have to
3301 * be careful to only set the discarded attribute if we know it
3302 * covers a complete block of the new size.
3303 */
3304struct discard_load_info {
3305 struct cache *cache;
3306
3307 /*
3308 * These blocks are sized using the on disk dblock size, rather
3309 * than the current one.
3310 */
3311 dm_block_t block_size;
3312 dm_block_t discard_begin, discard_end;
3313};
3314
3315static void discard_load_info_init(struct cache *cache,
3316 struct discard_load_info *li)
3317{
3318 li->cache = cache;
3319 li->discard_begin = li->discard_end = 0;
3320}
3321
3322static void set_discard_range(struct discard_load_info *li)
3323{
3324 sector_t b, e;
3325
3326 if (li->discard_begin == li->discard_end)
3327 return;
3328
3329 /*
3330 * Convert to sectors.
3331 */
3332 b = li->discard_begin * li->block_size;
3333 e = li->discard_end * li->block_size;
3334
3335 /*
3336 * Then convert back to the current dblock size.
3337 */
3338 b = dm_sector_div_up(b, li->cache->discard_block_size);
3339 sector_div(e, li->cache->discard_block_size);
3340
3341 /*
3342 * The origin may have shrunk, so we need to check we're still in
3343 * bounds.
3344 */
3345 if (e > from_dblock(li->cache->discard_nr_blocks))
3346 e = from_dblock(li->cache->discard_nr_blocks);
3347
3348 for (; b < e; b++)
3349 set_discard(li->cache, to_dblock(b));
3350}
3351
3352static int load_discard(void *context, sector_t discard_block_size,
3353 dm_dblock_t dblock, bool discard)
3354{
3355 struct discard_load_info *li = context;
3356
3357 li->block_size = discard_block_size;
3358
3359 if (discard) {
3360 if (from_dblock(dblock) == li->discard_end)
3361 /*
3362 * We're already in a discard range, just extend it.
3363 */
3364 li->discard_end = li->discard_end + 1ULL;
3365
3366 else {
3367 /*
3368 * Emit the old range and start a new one.
3369 */
3370 set_discard_range(li);
3371 li->discard_begin = from_dblock(dblock);
3372 li->discard_end = li->discard_begin + 1ULL;
3373 }
3374 } else {
3375 set_discard_range(li);
3376 li->discard_begin = li->discard_end = 0;
3377 }
3378
3379 return 0;
3380}
3381
3382static dm_cblock_t get_cache_dev_size(struct cache *cache)
3383{
3384 sector_t size = get_dev_size(cache->cache_dev);
3385 (void) sector_div(size, cache->sectors_per_block);
3386 return to_cblock(size);
3387}
3388
3389static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3390{
3391 if (from_cblock(new_size) > from_cblock(cache->cache_size))
3392 return true;
3393
3394 /*
3395 * We can't drop a dirty block when shrinking the cache.
3396 */
3397 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3398 new_size = to_cblock(from_cblock(new_size) + 1);
3399 if (is_dirty(cache, new_size)) {
3400 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3401 cache_device_name(cache),
3402 (unsigned long long) from_cblock(new_size));
3403 return false;
3404 }
3405 }
3406
3407 return true;
3408}
3409
3410static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3411{
3412 int r;
3413
3414 r = dm_cache_resize(cache->cmd, new_size);
3415 if (r) {
3416 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3417 metadata_operation_failed(cache, "dm_cache_resize", r);
3418 return r;
3419 }
3420
3421 set_cache_size(cache, new_size);
3422
3423 return 0;
3424}
3425
3426static int cache_preresume(struct dm_target *ti)
3427{
3428 int r = 0;
3429 struct cache *cache = ti->private;
3430 dm_cblock_t csize = get_cache_dev_size(cache);
3431
3432 /*
3433 * Check to see if the cache has resized.
3434 */
3435 if (!cache->sized) {
3436 r = resize_cache_dev(cache, csize);
3437 if (r)
3438 return r;
3439
3440 cache->sized = true;
3441
3442 } else if (csize != cache->cache_size) {
3443 if (!can_resize(cache, csize))
3444 return -EINVAL;
3445
3446 r = resize_cache_dev(cache, csize);
3447 if (r)
3448 return r;
3449 }
3450
3451 if (!cache->loaded_mappings) {
3452 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3453 load_mapping, cache);
3454 if (r) {
3455 DMERR("%s: could not load cache mappings", cache_device_name(cache));
3456 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3457 return r;
3458 }
3459
3460 cache->loaded_mappings = true;
3461 }
3462
3463 if (!cache->loaded_discards) {
3464 struct discard_load_info li;
3465
3466 /*
3467 * The discard bitset could have been resized, or the
3468 * discard block size changed. To be safe we start by
3469 * setting every dblock to not discarded.
3470 */
3471 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3472
3473 discard_load_info_init(cache, &li);
3474 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3475 if (r) {
3476 DMERR("%s: could not load origin discards", cache_device_name(cache));
3477 metadata_operation_failed(cache, "dm_cache_load_discards", r);
3478 return r;
3479 }
3480 set_discard_range(&li);
3481
3482 cache->loaded_discards = true;
3483 }
3484
3485 return r;
3486}
3487
3488static void cache_resume(struct dm_target *ti)
3489{
3490 struct cache *cache = ti->private;
3491
3492 cache->need_tick_bio = true;
3493 do_waker(&cache->waker.work);
3494}
3495
3496/*
3497 * Status format:
3498 *
3499 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3500 * <cache block size> <#used cache blocks>/<#total cache blocks>
3501 * <#read hits> <#read misses> <#write hits> <#write misses>
3502 * <#demotions> <#promotions> <#dirty>
3503 * <#features> <features>*
3504 * <#core args> <core args>
3505 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3506 */
3507static void cache_status(struct dm_target *ti, status_type_t type,
3508 unsigned status_flags, char *result, unsigned maxlen)
3509{
3510 int r = 0;
3511 unsigned i;
3512 ssize_t sz = 0;
3513 dm_block_t nr_free_blocks_metadata = 0;
3514 dm_block_t nr_blocks_metadata = 0;
3515 char buf[BDEVNAME_SIZE];
3516 struct cache *cache = ti->private;
3517 dm_cblock_t residency;
3518 bool needs_check;
3519
3520 switch (type) {
3521 case STATUSTYPE_INFO:
3522 if (get_cache_mode(cache) == CM_FAIL) {
3523 DMEMIT("Fail");
3524 break;
3525 }
3526
3527 /* Commit to ensure statistics aren't out-of-date */
3528 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3529 (void) commit(cache, false);
3530
3531 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3532 if (r) {
3533 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3534 cache_device_name(cache), r);
3535 goto err;
3536 }
3537
3538 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3539 if (r) {
3540 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3541 cache_device_name(cache), r);
3542 goto err;
3543 }
3544
3545 residency = policy_residency(cache->policy);
3546
3547 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3548 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3549 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3550 (unsigned long long)nr_blocks_metadata,
3551 cache->sectors_per_block,
3552 (unsigned long long) from_cblock(residency),
3553 (unsigned long long) from_cblock(cache->cache_size),
3554 (unsigned) atomic_read(&cache->stats.read_hit),
3555 (unsigned) atomic_read(&cache->stats.read_miss),
3556 (unsigned) atomic_read(&cache->stats.write_hit),
3557 (unsigned) atomic_read(&cache->stats.write_miss),
3558 (unsigned) atomic_read(&cache->stats.demotion),
3559 (unsigned) atomic_read(&cache->stats.promotion),
3560 (unsigned long) atomic_read(&cache->nr_dirty));
3561
3562 if (writethrough_mode(&cache->features))
3563 DMEMIT("1 writethrough ");
3564
3565 else if (passthrough_mode(&cache->features))
3566 DMEMIT("1 passthrough ");
3567
3568 else if (writeback_mode(&cache->features))
3569 DMEMIT("1 writeback ");
3570
3571 else {
3572 DMERR("%s: internal error: unknown io mode: %d",
3573 cache_device_name(cache), (int) cache->features.io_mode);
3574 goto err;
3575 }
3576
3577 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3578
3579 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3580 if (sz < maxlen) {
3581 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3582 if (r)
3583 DMERR("%s: policy_emit_config_values returned %d",
3584 cache_device_name(cache), r);
3585 }
3586
3587 if (get_cache_mode(cache) == CM_READ_ONLY)
3588 DMEMIT("ro ");
3589 else
3590 DMEMIT("rw ");
3591
3592 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3593
3594 if (r || needs_check)
3595 DMEMIT("needs_check ");
3596 else
3597 DMEMIT("- ");
3598
3599 break;
3600
3601 case STATUSTYPE_TABLE:
3602 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3603 DMEMIT("%s ", buf);
3604 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3605 DMEMIT("%s ", buf);
3606 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3607 DMEMIT("%s", buf);
3608
3609 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3610 DMEMIT(" %s", cache->ctr_args[i]);
3611 if (cache->nr_ctr_args)
3612 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3613 }
3614
3615 return;
3616
3617err:
3618 DMEMIT("Error");
3619}
3620
3621/*
3622 * A cache block range can take two forms:
3623 *
3624 * i) A single cblock, eg. '3456'
3625 * ii) A begin and end cblock with dots between, eg. 123-234
3626 */
3627static int parse_cblock_range(struct cache *cache, const char *str,
3628 struct cblock_range *result)
3629{
3630 char dummy;
3631 uint64_t b, e;
3632 int r;
3633
3634 /*
3635 * Try and parse form (ii) first.
3636 */
3637 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3638 if (r < 0)
3639 return r;
3640
3641 if (r == 2) {
3642 result->begin = to_cblock(b);
3643 result->end = to_cblock(e);
3644 return 0;
3645 }
3646
3647 /*
3648 * That didn't work, try form (i).
3649 */
3650 r = sscanf(str, "%llu%c", &b, &dummy);
3651 if (r < 0)
3652 return r;
3653
3654 if (r == 1) {
3655 result->begin = to_cblock(b);
3656 result->end = to_cblock(from_cblock(result->begin) + 1u);
3657 return 0;
3658 }
3659
3660 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3661 return -EINVAL;
3662}
3663
3664static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3665{
3666 uint64_t b = from_cblock(range->begin);
3667 uint64_t e = from_cblock(range->end);
3668 uint64_t n = from_cblock(cache->cache_size);
3669
3670 if (b >= n) {
3671 DMERR("%s: begin cblock out of range: %llu >= %llu",
3672 cache_device_name(cache), b, n);
3673 return -EINVAL;
3674 }
3675
3676 if (e > n) {
3677 DMERR("%s: end cblock out of range: %llu > %llu",
3678 cache_device_name(cache), e, n);
3679 return -EINVAL;
3680 }
3681
3682 if (b >= e) {
3683 DMERR("%s: invalid cblock range: %llu >= %llu",
3684 cache_device_name(cache), b, e);
3685 return -EINVAL;
3686 }
3687
3688 return 0;
3689}
3690
3691static int request_invalidation(struct cache *cache, struct cblock_range *range)
3692{
3693 struct invalidation_request req;
3694
3695 INIT_LIST_HEAD(&req.list);
3696 req.cblocks = range;
3697 atomic_set(&req.complete, 0);
3698 req.err = 0;
3699 init_waitqueue_head(&req.result_wait);
3700
3701 spin_lock(&cache->invalidation_lock);
3702 list_add(&req.list, &cache->invalidation_requests);
3703 spin_unlock(&cache->invalidation_lock);
3704 wake_worker(cache);
3705
3706 wait_event(req.result_wait, atomic_read(&req.complete));
3707 return req.err;
3708}
3709
3710static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3711 const char **cblock_ranges)
3712{
3713 int r = 0;
3714 unsigned i;
3715 struct cblock_range range;
3716
3717 if (!passthrough_mode(&cache->features)) {
3718 DMERR("%s: cache has to be in passthrough mode for invalidation",
3719 cache_device_name(cache));
3720 return -EPERM;
3721 }
3722
3723 for (i = 0; i < count; i++) {
3724 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3725 if (r)
3726 break;
3727
3728 r = validate_cblock_range(cache, &range);
3729 if (r)
3730 break;
3731
3732 /*
3733 * Pass begin and end origin blocks to the worker and wake it.
3734 */
3735 r = request_invalidation(cache, &range);
3736 if (r)
3737 break;
3738 }
3739
3740 return r;
3741}
3742
3743/*
3744 * Supports
3745 * "<key> <value>"
3746 * and
3747 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3748 *
3749 * The key migration_threshold is supported by the cache target core.
3750 */
3751static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3752{
3753 struct cache *cache = ti->private;
3754
3755 if (!argc)
3756 return -EINVAL;
3757
3758 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3759 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3760 cache_device_name(cache));
3761 return -EOPNOTSUPP;
3762 }
3763
3764 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3765 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3766
3767 if (argc != 2)
3768 return -EINVAL;
3769
3770 return set_config_value(cache, argv[0], argv[1]);
3771}
3772
3773static int cache_iterate_devices(struct dm_target *ti,
3774 iterate_devices_callout_fn fn, void *data)
3775{
3776 int r = 0;
3777 struct cache *cache = ti->private;
3778
3779 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3780 if (!r)
3781 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3782
3783 return r;
3784}
3785
3786static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3787{
3788 /*
3789 * FIXME: these limits may be incompatible with the cache device
3790 */
3791 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3792 cache->origin_sectors);
3793 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3794}
3795
3796static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3797{
3798 struct cache *cache = ti->private;
3799 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3800
3801 /*
3802 * If the system-determined stacked limits are compatible with the
3803 * cache's blocksize (io_opt is a factor) do not override them.
3804 */
3805 if (io_opt_sectors < cache->sectors_per_block ||
3806 do_div(io_opt_sectors, cache->sectors_per_block)) {
3807 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3808 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3809 }
3810 set_discard_limits(cache, limits);
3811}
3812
3813/*----------------------------------------------------------------*/
3814
3815static struct target_type cache_target = {
3816 .name = "cache",
3817 .version = {1, 9, 0},
3818 .module = THIS_MODULE,
3819 .ctr = cache_ctr,
3820 .dtr = cache_dtr,
3821 .map = cache_map,
3822 .end_io = cache_end_io,
3823 .postsuspend = cache_postsuspend,
3824 .preresume = cache_preresume,
3825 .resume = cache_resume,
3826 .status = cache_status,
3827 .message = cache_message,
3828 .iterate_devices = cache_iterate_devices,
3829 .io_hints = cache_io_hints,
3830};
3831
3832static int __init dm_cache_init(void)
3833{
3834 int r;
3835
3836 r = dm_register_target(&cache_target);
3837 if (r) {
3838 DMERR("cache target registration failed: %d", r);
3839 return r;
3840 }
3841
3842 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3843 if (!migration_cache) {
3844 dm_unregister_target(&cache_target);
3845 return -ENOMEM;
3846 }
3847
3848 return 0;
3849}
3850
3851static void __exit dm_cache_exit(void)
3852{
3853 dm_unregister_target(&cache_target);
3854 kmem_cache_destroy(migration_cache);
3855}
3856
3857module_init(dm_cache_init);
3858module_exit(dm_cache_exit);
3859
3860MODULE_DESCRIPTION(DM_NAME " cache target");
3861MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3862MODULE_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.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");