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// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2012 Red Hat. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm.h"
9#include "dm-bio-prison-v2.h"
10#include "dm-bio-record.h"
11#include "dm-cache-metadata.h"
12#include "dm-io-tracker.h"
13
14#include <linux/dm-io.h>
15#include <linux/dm-kcopyd.h>
16#include <linux/jiffies.h>
17#include <linux/init.h>
18#include <linux/mempool.h>
19#include <linux/module.h>
20#include <linux/rwsem.h>
21#include <linux/slab.h>
22#include <linux/vmalloc.h>
23
24#define DM_MSG_PREFIX "cache"
25
26DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
27 "A percentage of time allocated for copying to and/or from cache");
28
29/*----------------------------------------------------------------*/
30
31/*
32 * Glossary:
33 *
34 * oblock: index of an origin block
35 * cblock: index of a cache block
36 * promotion: movement of a block from origin to cache
37 * demotion: movement of a block from cache to origin
38 * migration: movement of a block between the origin and cache device,
39 * either direction
40 */
41
42/*----------------------------------------------------------------*/
43
44/*
45 * Represents a chunk of future work. 'input' allows continuations to pass
46 * values between themselves, typically error values.
47 */
48struct continuation {
49 struct work_struct ws;
50 blk_status_t input;
51};
52
53static inline void init_continuation(struct continuation *k,
54 void (*fn)(struct work_struct *))
55{
56 INIT_WORK(&k->ws, fn);
57 k->input = 0;
58}
59
60static inline void queue_continuation(struct workqueue_struct *wq,
61 struct continuation *k)
62{
63 queue_work(wq, &k->ws);
64}
65
66/*----------------------------------------------------------------*/
67
68/*
69 * The batcher collects together pieces of work that need a particular
70 * operation to occur before they can proceed (typically a commit).
71 */
72struct batcher {
73 /*
74 * The operation that everyone is waiting for.
75 */
76 blk_status_t (*commit_op)(void *context);
77 void *commit_context;
78
79 /*
80 * This is how bios should be issued once the commit op is complete
81 * (accounted_request).
82 */
83 void (*issue_op)(struct bio *bio, void *context);
84 void *issue_context;
85
86 /*
87 * Queued work gets put on here after commit.
88 */
89 struct workqueue_struct *wq;
90
91 spinlock_t lock;
92 struct list_head work_items;
93 struct bio_list bios;
94 struct work_struct commit_work;
95
96 bool commit_scheduled;
97};
98
99static void __commit(struct work_struct *_ws)
100{
101 struct batcher *b = container_of(_ws, struct batcher, commit_work);
102 blk_status_t r;
103 struct list_head work_items;
104 struct work_struct *ws, *tmp;
105 struct continuation *k;
106 struct bio *bio;
107 struct bio_list bios;
108
109 INIT_LIST_HEAD(&work_items);
110 bio_list_init(&bios);
111
112 /*
113 * We have to grab these before the commit_op to avoid a race
114 * condition.
115 */
116 spin_lock_irq(&b->lock);
117 list_splice_init(&b->work_items, &work_items);
118 bio_list_merge(&bios, &b->bios);
119 bio_list_init(&b->bios);
120 b->commit_scheduled = false;
121 spin_unlock_irq(&b->lock);
122
123 r = b->commit_op(b->commit_context);
124
125 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
126 k = container_of(ws, struct continuation, ws);
127 k->input = r;
128 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
129 queue_work(b->wq, ws);
130 }
131
132 while ((bio = bio_list_pop(&bios))) {
133 if (r) {
134 bio->bi_status = r;
135 bio_endio(bio);
136 } else
137 b->issue_op(bio, b->issue_context);
138 }
139}
140
141static void batcher_init(struct batcher *b,
142 blk_status_t (*commit_op)(void *),
143 void *commit_context,
144 void (*issue_op)(struct bio *bio, void *),
145 void *issue_context,
146 struct workqueue_struct *wq)
147{
148 b->commit_op = commit_op;
149 b->commit_context = commit_context;
150 b->issue_op = issue_op;
151 b->issue_context = issue_context;
152 b->wq = wq;
153
154 spin_lock_init(&b->lock);
155 INIT_LIST_HEAD(&b->work_items);
156 bio_list_init(&b->bios);
157 INIT_WORK(&b->commit_work, __commit);
158 b->commit_scheduled = false;
159}
160
161static void async_commit(struct batcher *b)
162{
163 queue_work(b->wq, &b->commit_work);
164}
165
166static void continue_after_commit(struct batcher *b, struct continuation *k)
167{
168 bool commit_scheduled;
169
170 spin_lock_irq(&b->lock);
171 commit_scheduled = b->commit_scheduled;
172 list_add_tail(&k->ws.entry, &b->work_items);
173 spin_unlock_irq(&b->lock);
174
175 if (commit_scheduled)
176 async_commit(b);
177}
178
179/*
180 * Bios are errored if commit failed.
181 */
182static void issue_after_commit(struct batcher *b, struct bio *bio)
183{
184 bool commit_scheduled;
185
186 spin_lock_irq(&b->lock);
187 commit_scheduled = b->commit_scheduled;
188 bio_list_add(&b->bios, bio);
189 spin_unlock_irq(&b->lock);
190
191 if (commit_scheduled)
192 async_commit(b);
193}
194
195/*
196 * Call this if some urgent work is waiting for the commit to complete.
197 */
198static void schedule_commit(struct batcher *b)
199{
200 bool immediate;
201
202 spin_lock_irq(&b->lock);
203 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
204 b->commit_scheduled = true;
205 spin_unlock_irq(&b->lock);
206
207 if (immediate)
208 async_commit(b);
209}
210
211/*
212 * There are a couple of places where we let a bio run, but want to do some
213 * work before calling its endio function. We do this by temporarily
214 * changing the endio fn.
215 */
216struct dm_hook_info {
217 bio_end_io_t *bi_end_io;
218};
219
220static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
221 bio_end_io_t *bi_end_io, void *bi_private)
222{
223 h->bi_end_io = bio->bi_end_io;
224
225 bio->bi_end_io = bi_end_io;
226 bio->bi_private = bi_private;
227}
228
229static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
230{
231 bio->bi_end_io = h->bi_end_io;
232}
233
234/*----------------------------------------------------------------*/
235
236#define MIGRATION_POOL_SIZE 128
237#define COMMIT_PERIOD HZ
238#define MIGRATION_COUNT_WINDOW 10
239
240/*
241 * The block size of the device holding cache data must be
242 * between 32KB and 1GB.
243 */
244#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
245#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
246
247enum cache_metadata_mode {
248 CM_WRITE, /* metadata may be changed */
249 CM_READ_ONLY, /* metadata may not be changed */
250 CM_FAIL
251};
252
253enum cache_io_mode {
254 /*
255 * Data is written to cached blocks only. These blocks are marked
256 * dirty. If you lose the cache device you will lose data.
257 * Potential performance increase for both reads and writes.
258 */
259 CM_IO_WRITEBACK,
260
261 /*
262 * Data is written to both cache and origin. Blocks are never
263 * dirty. Potential performance benfit for reads only.
264 */
265 CM_IO_WRITETHROUGH,
266
267 /*
268 * A degraded mode useful for various cache coherency situations
269 * (eg, rolling back snapshots). Reads and writes always go to the
270 * origin. If a write goes to a cached oblock, then the cache
271 * block is invalidated.
272 */
273 CM_IO_PASSTHROUGH
274};
275
276struct cache_features {
277 enum cache_metadata_mode mode;
278 enum cache_io_mode io_mode;
279 unsigned int metadata_version;
280 bool discard_passdown:1;
281};
282
283struct cache_stats {
284 atomic_t read_hit;
285 atomic_t read_miss;
286 atomic_t write_hit;
287 atomic_t write_miss;
288 atomic_t demotion;
289 atomic_t promotion;
290 atomic_t writeback;
291 atomic_t copies_avoided;
292 atomic_t cache_cell_clash;
293 atomic_t commit_count;
294 atomic_t discard_count;
295};
296
297struct cache {
298 struct dm_target *ti;
299 spinlock_t lock;
300
301 /*
302 * Fields for converting from sectors to blocks.
303 */
304 int sectors_per_block_shift;
305 sector_t sectors_per_block;
306
307 struct dm_cache_metadata *cmd;
308
309 /*
310 * Metadata is written to this device.
311 */
312 struct dm_dev *metadata_dev;
313
314 /*
315 * The slower of the two data devices. Typically a spindle.
316 */
317 struct dm_dev *origin_dev;
318
319 /*
320 * The faster of the two data devices. Typically an SSD.
321 */
322 struct dm_dev *cache_dev;
323
324 /*
325 * Size of the origin device in _complete_ blocks and native sectors.
326 */
327 dm_oblock_t origin_blocks;
328 sector_t origin_sectors;
329
330 /*
331 * Size of the cache device in blocks.
332 */
333 dm_cblock_t cache_size;
334
335 /*
336 * Invalidation fields.
337 */
338 spinlock_t invalidation_lock;
339 struct list_head invalidation_requests;
340
341 sector_t migration_threshold;
342 wait_queue_head_t migration_wait;
343 atomic_t nr_allocated_migrations;
344
345 /*
346 * The number of in flight migrations that are performing
347 * background io. eg, promotion, writeback.
348 */
349 atomic_t nr_io_migrations;
350
351 struct bio_list deferred_bios;
352
353 struct rw_semaphore quiesce_lock;
354
355 /*
356 * origin_blocks entries, discarded if set.
357 */
358 dm_dblock_t discard_nr_blocks;
359 unsigned long *discard_bitset;
360 uint32_t discard_block_size; /* a power of 2 times sectors per block */
361
362 /*
363 * Rather than reconstructing the table line for the status we just
364 * save it and regurgitate.
365 */
366 unsigned int nr_ctr_args;
367 const char **ctr_args;
368
369 struct dm_kcopyd_client *copier;
370 struct work_struct deferred_bio_worker;
371 struct work_struct migration_worker;
372 struct workqueue_struct *wq;
373 struct delayed_work waker;
374 struct dm_bio_prison_v2 *prison;
375
376 /*
377 * cache_size entries, dirty if set
378 */
379 unsigned long *dirty_bitset;
380 atomic_t nr_dirty;
381
382 unsigned int policy_nr_args;
383 struct dm_cache_policy *policy;
384
385 /*
386 * Cache features such as write-through.
387 */
388 struct cache_features features;
389
390 struct cache_stats stats;
391
392 bool need_tick_bio:1;
393 bool sized:1;
394 bool invalidate:1;
395 bool commit_requested:1;
396 bool loaded_mappings:1;
397 bool loaded_discards:1;
398
399 struct rw_semaphore background_work_lock;
400
401 struct batcher committer;
402 struct work_struct commit_ws;
403
404 struct dm_io_tracker tracker;
405
406 mempool_t migration_pool;
407
408 struct bio_set bs;
409};
410
411struct per_bio_data {
412 bool tick:1;
413 unsigned int req_nr:2;
414 struct dm_bio_prison_cell_v2 *cell;
415 struct dm_hook_info hook_info;
416 sector_t len;
417};
418
419struct dm_cache_migration {
420 struct continuation k;
421 struct cache *cache;
422
423 struct policy_work *op;
424 struct bio *overwrite_bio;
425 struct dm_bio_prison_cell_v2 *cell;
426
427 dm_cblock_t invalidate_cblock;
428 dm_oblock_t invalidate_oblock;
429};
430
431/*----------------------------------------------------------------*/
432
433static bool writethrough_mode(struct cache *cache)
434{
435 return cache->features.io_mode == CM_IO_WRITETHROUGH;
436}
437
438static bool writeback_mode(struct cache *cache)
439{
440 return cache->features.io_mode == CM_IO_WRITEBACK;
441}
442
443static inline bool passthrough_mode(struct cache *cache)
444{
445 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
446}
447
448/*----------------------------------------------------------------*/
449
450static void wake_deferred_bio_worker(struct cache *cache)
451{
452 queue_work(cache->wq, &cache->deferred_bio_worker);
453}
454
455static void wake_migration_worker(struct cache *cache)
456{
457 if (passthrough_mode(cache))
458 return;
459
460 queue_work(cache->wq, &cache->migration_worker);
461}
462
463/*----------------------------------------------------------------*/
464
465static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
466{
467 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
468}
469
470static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
471{
472 dm_bio_prison_free_cell_v2(cache->prison, cell);
473}
474
475static struct dm_cache_migration *alloc_migration(struct cache *cache)
476{
477 struct dm_cache_migration *mg;
478
479 mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
480
481 memset(mg, 0, sizeof(*mg));
482
483 mg->cache = cache;
484 atomic_inc(&cache->nr_allocated_migrations);
485
486 return mg;
487}
488
489static void free_migration(struct dm_cache_migration *mg)
490{
491 struct cache *cache = mg->cache;
492
493 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
494 wake_up(&cache->migration_wait);
495
496 mempool_free(mg, &cache->migration_pool);
497}
498
499/*----------------------------------------------------------------*/
500
501static inline dm_oblock_t oblock_succ(dm_oblock_t b)
502{
503 return to_oblock(from_oblock(b) + 1ull);
504}
505
506static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
507{
508 key->virtual = 0;
509 key->dev = 0;
510 key->block_begin = from_oblock(begin);
511 key->block_end = from_oblock(end);
512}
513
514/*
515 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
516 * level 1 which prevents *both* READs and WRITEs.
517 */
518#define WRITE_LOCK_LEVEL 0
519#define READ_WRITE_LOCK_LEVEL 1
520
521static unsigned int lock_level(struct bio *bio)
522{
523 return bio_data_dir(bio) == WRITE ?
524 WRITE_LOCK_LEVEL :
525 READ_WRITE_LOCK_LEVEL;
526}
527
528/*
529 *--------------------------------------------------------------
530 * Per bio data
531 *--------------------------------------------------------------
532 */
533
534static struct per_bio_data *get_per_bio_data(struct bio *bio)
535{
536 struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
537
538 BUG_ON(!pb);
539 return pb;
540}
541
542static struct per_bio_data *init_per_bio_data(struct bio *bio)
543{
544 struct per_bio_data *pb = get_per_bio_data(bio);
545
546 pb->tick = false;
547 pb->req_nr = dm_bio_get_target_bio_nr(bio);
548 pb->cell = NULL;
549 pb->len = 0;
550
551 return pb;
552}
553
554/*----------------------------------------------------------------*/
555
556static void defer_bio(struct cache *cache, struct bio *bio)
557{
558 spin_lock_irq(&cache->lock);
559 bio_list_add(&cache->deferred_bios, bio);
560 spin_unlock_irq(&cache->lock);
561
562 wake_deferred_bio_worker(cache);
563}
564
565static void defer_bios(struct cache *cache, struct bio_list *bios)
566{
567 spin_lock_irq(&cache->lock);
568 bio_list_merge(&cache->deferred_bios, bios);
569 bio_list_init(bios);
570 spin_unlock_irq(&cache->lock);
571
572 wake_deferred_bio_worker(cache);
573}
574
575/*----------------------------------------------------------------*/
576
577static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
578{
579 bool r;
580 struct per_bio_data *pb;
581 struct dm_cell_key_v2 key;
582 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
583 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
584
585 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
586
587 build_key(oblock, end, &key);
588 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
589 if (!r) {
590 /*
591 * Failed to get the lock.
592 */
593 free_prison_cell(cache, cell_prealloc);
594 return r;
595 }
596
597 if (cell != cell_prealloc)
598 free_prison_cell(cache, cell_prealloc);
599
600 pb = get_per_bio_data(bio);
601 pb->cell = cell;
602
603 return r;
604}
605
606/*----------------------------------------------------------------*/
607
608static bool is_dirty(struct cache *cache, dm_cblock_t b)
609{
610 return test_bit(from_cblock(b), cache->dirty_bitset);
611}
612
613static void set_dirty(struct cache *cache, dm_cblock_t cblock)
614{
615 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
616 atomic_inc(&cache->nr_dirty);
617 policy_set_dirty(cache->policy, cblock);
618 }
619}
620
621/*
622 * These two are called when setting after migrations to force the policy
623 * and dirty bitset to be in sync.
624 */
625static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
626{
627 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
628 atomic_inc(&cache->nr_dirty);
629 policy_set_dirty(cache->policy, cblock);
630}
631
632static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
633{
634 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
635 if (atomic_dec_return(&cache->nr_dirty) == 0)
636 dm_table_event(cache->ti->table);
637 }
638
639 policy_clear_dirty(cache->policy, cblock);
640}
641
642/*----------------------------------------------------------------*/
643
644static bool block_size_is_power_of_two(struct cache *cache)
645{
646 return cache->sectors_per_block_shift >= 0;
647}
648
649static dm_block_t block_div(dm_block_t b, uint32_t n)
650{
651 do_div(b, n);
652
653 return b;
654}
655
656static dm_block_t oblocks_per_dblock(struct cache *cache)
657{
658 dm_block_t oblocks = cache->discard_block_size;
659
660 if (block_size_is_power_of_two(cache))
661 oblocks >>= cache->sectors_per_block_shift;
662 else
663 oblocks = block_div(oblocks, cache->sectors_per_block);
664
665 return oblocks;
666}
667
668static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
669{
670 return to_dblock(block_div(from_oblock(oblock),
671 oblocks_per_dblock(cache)));
672}
673
674static void set_discard(struct cache *cache, dm_dblock_t b)
675{
676 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
677 atomic_inc(&cache->stats.discard_count);
678
679 spin_lock_irq(&cache->lock);
680 set_bit(from_dblock(b), cache->discard_bitset);
681 spin_unlock_irq(&cache->lock);
682}
683
684static void clear_discard(struct cache *cache, dm_dblock_t b)
685{
686 spin_lock_irq(&cache->lock);
687 clear_bit(from_dblock(b), cache->discard_bitset);
688 spin_unlock_irq(&cache->lock);
689}
690
691static bool is_discarded(struct cache *cache, dm_dblock_t b)
692{
693 int r;
694
695 spin_lock_irq(&cache->lock);
696 r = test_bit(from_dblock(b), cache->discard_bitset);
697 spin_unlock_irq(&cache->lock);
698
699 return r;
700}
701
702static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
703{
704 int r;
705
706 spin_lock_irq(&cache->lock);
707 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
708 cache->discard_bitset);
709 spin_unlock_irq(&cache->lock);
710
711 return r;
712}
713
714/*
715 * -------------------------------------------------------------
716 * Remapping
717 *--------------------------------------------------------------
718 */
719static void remap_to_origin(struct cache *cache, struct bio *bio)
720{
721 bio_set_dev(bio, cache->origin_dev->bdev);
722}
723
724static void remap_to_cache(struct cache *cache, struct bio *bio,
725 dm_cblock_t cblock)
726{
727 sector_t bi_sector = bio->bi_iter.bi_sector;
728 sector_t block = from_cblock(cblock);
729
730 bio_set_dev(bio, cache->cache_dev->bdev);
731 if (!block_size_is_power_of_two(cache))
732 bio->bi_iter.bi_sector =
733 (block * cache->sectors_per_block) +
734 sector_div(bi_sector, cache->sectors_per_block);
735 else
736 bio->bi_iter.bi_sector =
737 (block << cache->sectors_per_block_shift) |
738 (bi_sector & (cache->sectors_per_block - 1));
739}
740
741static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
742{
743 struct per_bio_data *pb;
744
745 spin_lock_irq(&cache->lock);
746 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
747 bio_op(bio) != REQ_OP_DISCARD) {
748 pb = get_per_bio_data(bio);
749 pb->tick = true;
750 cache->need_tick_bio = false;
751 }
752 spin_unlock_irq(&cache->lock);
753}
754
755static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
756 dm_oblock_t oblock)
757{
758 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
759 check_if_tick_bio_needed(cache, bio);
760 remap_to_origin(cache, bio);
761 if (bio_data_dir(bio) == WRITE)
762 clear_discard(cache, oblock_to_dblock(cache, oblock));
763}
764
765static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
766 dm_oblock_t oblock, dm_cblock_t cblock)
767{
768 check_if_tick_bio_needed(cache, bio);
769 remap_to_cache(cache, bio, cblock);
770 if (bio_data_dir(bio) == WRITE) {
771 set_dirty(cache, cblock);
772 clear_discard(cache, oblock_to_dblock(cache, oblock));
773 }
774}
775
776static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
777{
778 sector_t block_nr = bio->bi_iter.bi_sector;
779
780 if (!block_size_is_power_of_two(cache))
781 (void) sector_div(block_nr, cache->sectors_per_block);
782 else
783 block_nr >>= cache->sectors_per_block_shift;
784
785 return to_oblock(block_nr);
786}
787
788static bool accountable_bio(struct cache *cache, struct bio *bio)
789{
790 return bio_op(bio) != REQ_OP_DISCARD;
791}
792
793static void accounted_begin(struct cache *cache, struct bio *bio)
794{
795 struct per_bio_data *pb;
796
797 if (accountable_bio(cache, bio)) {
798 pb = get_per_bio_data(bio);
799 pb->len = bio_sectors(bio);
800 dm_iot_io_begin(&cache->tracker, pb->len);
801 }
802}
803
804static void accounted_complete(struct cache *cache, struct bio *bio)
805{
806 struct per_bio_data *pb = get_per_bio_data(bio);
807
808 dm_iot_io_end(&cache->tracker, pb->len);
809}
810
811static void accounted_request(struct cache *cache, struct bio *bio)
812{
813 accounted_begin(cache, bio);
814 dm_submit_bio_remap(bio, NULL);
815}
816
817static void issue_op(struct bio *bio, void *context)
818{
819 struct cache *cache = context;
820
821 accounted_request(cache, bio);
822}
823
824/*
825 * When running in writethrough mode we need to send writes to clean blocks
826 * to both the cache and origin devices. Clone the bio and send them in parallel.
827 */
828static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
829 dm_oblock_t oblock, dm_cblock_t cblock)
830{
831 struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
832 GFP_NOIO, &cache->bs);
833
834 BUG_ON(!origin_bio);
835
836 bio_chain(origin_bio, bio);
837
838 if (bio_data_dir(origin_bio) == WRITE)
839 clear_discard(cache, oblock_to_dblock(cache, oblock));
840 submit_bio(origin_bio);
841
842 remap_to_cache(cache, bio, cblock);
843}
844
845/*
846 *--------------------------------------------------------------
847 * Failure modes
848 *--------------------------------------------------------------
849 */
850static enum cache_metadata_mode get_cache_mode(struct cache *cache)
851{
852 return cache->features.mode;
853}
854
855static const char *cache_device_name(struct cache *cache)
856{
857 return dm_table_device_name(cache->ti->table);
858}
859
860static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
861{
862 static const char *descs[] = {
863 "write",
864 "read-only",
865 "fail"
866 };
867
868 dm_table_event(cache->ti->table);
869 DMINFO("%s: switching cache to %s mode",
870 cache_device_name(cache), descs[(int)mode]);
871}
872
873static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
874{
875 bool needs_check;
876 enum cache_metadata_mode old_mode = get_cache_mode(cache);
877
878 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
879 DMERR("%s: unable to read needs_check flag, setting failure mode.",
880 cache_device_name(cache));
881 new_mode = CM_FAIL;
882 }
883
884 if (new_mode == CM_WRITE && needs_check) {
885 DMERR("%s: unable to switch cache to write mode until repaired.",
886 cache_device_name(cache));
887 if (old_mode != new_mode)
888 new_mode = old_mode;
889 else
890 new_mode = CM_READ_ONLY;
891 }
892
893 /* Never move out of fail mode */
894 if (old_mode == CM_FAIL)
895 new_mode = CM_FAIL;
896
897 switch (new_mode) {
898 case CM_FAIL:
899 case CM_READ_ONLY:
900 dm_cache_metadata_set_read_only(cache->cmd);
901 break;
902
903 case CM_WRITE:
904 dm_cache_metadata_set_read_write(cache->cmd);
905 break;
906 }
907
908 cache->features.mode = new_mode;
909
910 if (new_mode != old_mode)
911 notify_mode_switch(cache, new_mode);
912}
913
914static void abort_transaction(struct cache *cache)
915{
916 const char *dev_name = cache_device_name(cache);
917
918 if (get_cache_mode(cache) >= CM_READ_ONLY)
919 return;
920
921 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
922 if (dm_cache_metadata_abort(cache->cmd)) {
923 DMERR("%s: failed to abort metadata transaction", dev_name);
924 set_cache_mode(cache, CM_FAIL);
925 }
926
927 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
928 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
929 set_cache_mode(cache, CM_FAIL);
930 }
931}
932
933static void metadata_operation_failed(struct cache *cache, const char *op, int r)
934{
935 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
936 cache_device_name(cache), op, r);
937 abort_transaction(cache);
938 set_cache_mode(cache, CM_READ_ONLY);
939}
940
941/*----------------------------------------------------------------*/
942
943static void load_stats(struct cache *cache)
944{
945 struct dm_cache_statistics stats;
946
947 dm_cache_metadata_get_stats(cache->cmd, &stats);
948 atomic_set(&cache->stats.read_hit, stats.read_hits);
949 atomic_set(&cache->stats.read_miss, stats.read_misses);
950 atomic_set(&cache->stats.write_hit, stats.write_hits);
951 atomic_set(&cache->stats.write_miss, stats.write_misses);
952}
953
954static void save_stats(struct cache *cache)
955{
956 struct dm_cache_statistics stats;
957
958 if (get_cache_mode(cache) >= CM_READ_ONLY)
959 return;
960
961 stats.read_hits = atomic_read(&cache->stats.read_hit);
962 stats.read_misses = atomic_read(&cache->stats.read_miss);
963 stats.write_hits = atomic_read(&cache->stats.write_hit);
964 stats.write_misses = atomic_read(&cache->stats.write_miss);
965
966 dm_cache_metadata_set_stats(cache->cmd, &stats);
967}
968
969static void update_stats(struct cache_stats *stats, enum policy_operation op)
970{
971 switch (op) {
972 case POLICY_PROMOTE:
973 atomic_inc(&stats->promotion);
974 break;
975
976 case POLICY_DEMOTE:
977 atomic_inc(&stats->demotion);
978 break;
979
980 case POLICY_WRITEBACK:
981 atomic_inc(&stats->writeback);
982 break;
983 }
984}
985
986/*
987 *---------------------------------------------------------------------
988 * Migration processing
989 *
990 * Migration covers moving data from the origin device to the cache, or
991 * vice versa.
992 *---------------------------------------------------------------------
993 */
994static void inc_io_migrations(struct cache *cache)
995{
996 atomic_inc(&cache->nr_io_migrations);
997}
998
999static void dec_io_migrations(struct cache *cache)
1000{
1001 atomic_dec(&cache->nr_io_migrations);
1002}
1003
1004static bool discard_or_flush(struct bio *bio)
1005{
1006 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1007}
1008
1009static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1010 dm_dblock_t *b, dm_dblock_t *e)
1011{
1012 sector_t sb = bio->bi_iter.bi_sector;
1013 sector_t se = bio_end_sector(bio);
1014
1015 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1016
1017 if (se - sb < cache->discard_block_size)
1018 *e = *b;
1019 else
1020 *e = to_dblock(block_div(se, cache->discard_block_size));
1021}
1022
1023/*----------------------------------------------------------------*/
1024
1025static void prevent_background_work(struct cache *cache)
1026{
1027 lockdep_off();
1028 down_write(&cache->background_work_lock);
1029 lockdep_on();
1030}
1031
1032static void allow_background_work(struct cache *cache)
1033{
1034 lockdep_off();
1035 up_write(&cache->background_work_lock);
1036 lockdep_on();
1037}
1038
1039static bool background_work_begin(struct cache *cache)
1040{
1041 bool r;
1042
1043 lockdep_off();
1044 r = down_read_trylock(&cache->background_work_lock);
1045 lockdep_on();
1046
1047 return r;
1048}
1049
1050static void background_work_end(struct cache *cache)
1051{
1052 lockdep_off();
1053 up_read(&cache->background_work_lock);
1054 lockdep_on();
1055}
1056
1057/*----------------------------------------------------------------*/
1058
1059static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1060{
1061 return (bio_data_dir(bio) == WRITE) &&
1062 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1063}
1064
1065static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1066{
1067 return writeback_mode(cache) &&
1068 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1069}
1070
1071static void quiesce(struct dm_cache_migration *mg,
1072 void (*continuation)(struct work_struct *))
1073{
1074 init_continuation(&mg->k, continuation);
1075 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1076}
1077
1078static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1079{
1080 struct continuation *k = container_of(ws, struct continuation, ws);
1081
1082 return container_of(k, struct dm_cache_migration, k);
1083}
1084
1085static void copy_complete(int read_err, unsigned long write_err, void *context)
1086{
1087 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1088
1089 if (read_err || write_err)
1090 mg->k.input = BLK_STS_IOERR;
1091
1092 queue_continuation(mg->cache->wq, &mg->k);
1093}
1094
1095static void copy(struct dm_cache_migration *mg, bool promote)
1096{
1097 struct dm_io_region o_region, c_region;
1098 struct cache *cache = mg->cache;
1099
1100 o_region.bdev = cache->origin_dev->bdev;
1101 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1102 o_region.count = cache->sectors_per_block;
1103
1104 c_region.bdev = cache->cache_dev->bdev;
1105 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1106 c_region.count = cache->sectors_per_block;
1107
1108 if (promote)
1109 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1110 else
1111 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1112}
1113
1114static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1115{
1116 struct per_bio_data *pb = get_per_bio_data(bio);
1117
1118 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1119 free_prison_cell(cache, pb->cell);
1120 pb->cell = NULL;
1121}
1122
1123static void overwrite_endio(struct bio *bio)
1124{
1125 struct dm_cache_migration *mg = bio->bi_private;
1126 struct cache *cache = mg->cache;
1127 struct per_bio_data *pb = get_per_bio_data(bio);
1128
1129 dm_unhook_bio(&pb->hook_info, bio);
1130
1131 if (bio->bi_status)
1132 mg->k.input = bio->bi_status;
1133
1134 queue_continuation(cache->wq, &mg->k);
1135}
1136
1137static void overwrite(struct dm_cache_migration *mg,
1138 void (*continuation)(struct work_struct *))
1139{
1140 struct bio *bio = mg->overwrite_bio;
1141 struct per_bio_data *pb = get_per_bio_data(bio);
1142
1143 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1144
1145 /*
1146 * The overwrite bio is part of the copy operation, as such it does
1147 * not set/clear discard or dirty flags.
1148 */
1149 if (mg->op->op == POLICY_PROMOTE)
1150 remap_to_cache(mg->cache, bio, mg->op->cblock);
1151 else
1152 remap_to_origin(mg->cache, bio);
1153
1154 init_continuation(&mg->k, continuation);
1155 accounted_request(mg->cache, bio);
1156}
1157
1158/*
1159 * Migration steps:
1160 *
1161 * 1) exclusive lock preventing WRITEs
1162 * 2) quiesce
1163 * 3) copy or issue overwrite bio
1164 * 4) upgrade to exclusive lock preventing READs and WRITEs
1165 * 5) quiesce
1166 * 6) update metadata and commit
1167 * 7) unlock
1168 */
1169static void mg_complete(struct dm_cache_migration *mg, bool success)
1170{
1171 struct bio_list bios;
1172 struct cache *cache = mg->cache;
1173 struct policy_work *op = mg->op;
1174 dm_cblock_t cblock = op->cblock;
1175
1176 if (success)
1177 update_stats(&cache->stats, op->op);
1178
1179 switch (op->op) {
1180 case POLICY_PROMOTE:
1181 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1182 policy_complete_background_work(cache->policy, op, success);
1183
1184 if (mg->overwrite_bio) {
1185 if (success)
1186 force_set_dirty(cache, cblock);
1187 else if (mg->k.input)
1188 mg->overwrite_bio->bi_status = mg->k.input;
1189 else
1190 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1191 bio_endio(mg->overwrite_bio);
1192 } else {
1193 if (success)
1194 force_clear_dirty(cache, cblock);
1195 dec_io_migrations(cache);
1196 }
1197 break;
1198
1199 case POLICY_DEMOTE:
1200 /*
1201 * We clear dirty here to update the nr_dirty counter.
1202 */
1203 if (success)
1204 force_clear_dirty(cache, cblock);
1205 policy_complete_background_work(cache->policy, op, success);
1206 dec_io_migrations(cache);
1207 break;
1208
1209 case POLICY_WRITEBACK:
1210 if (success)
1211 force_clear_dirty(cache, cblock);
1212 policy_complete_background_work(cache->policy, op, success);
1213 dec_io_migrations(cache);
1214 break;
1215 }
1216
1217 bio_list_init(&bios);
1218 if (mg->cell) {
1219 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1220 free_prison_cell(cache, mg->cell);
1221 }
1222
1223 free_migration(mg);
1224 defer_bios(cache, &bios);
1225 wake_migration_worker(cache);
1226
1227 background_work_end(cache);
1228}
1229
1230static void mg_success(struct work_struct *ws)
1231{
1232 struct dm_cache_migration *mg = ws_to_mg(ws);
1233
1234 mg_complete(mg, mg->k.input == 0);
1235}
1236
1237static void mg_update_metadata(struct work_struct *ws)
1238{
1239 int r;
1240 struct dm_cache_migration *mg = ws_to_mg(ws);
1241 struct cache *cache = mg->cache;
1242 struct policy_work *op = mg->op;
1243
1244 switch (op->op) {
1245 case POLICY_PROMOTE:
1246 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1247 if (r) {
1248 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1249 cache_device_name(cache));
1250 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1251
1252 mg_complete(mg, false);
1253 return;
1254 }
1255 mg_complete(mg, true);
1256 break;
1257
1258 case POLICY_DEMOTE:
1259 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1260 if (r) {
1261 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1262 cache_device_name(cache));
1263 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1264
1265 mg_complete(mg, false);
1266 return;
1267 }
1268
1269 /*
1270 * It would be nice if we only had to commit when a REQ_FLUSH
1271 * comes through. But there's one scenario that we have to
1272 * look out for:
1273 *
1274 * - vblock x in a cache block
1275 * - domotion occurs
1276 * - cache block gets reallocated and over written
1277 * - crash
1278 *
1279 * When we recover, because there was no commit the cache will
1280 * rollback to having the data for vblock x in the cache block.
1281 * But the cache block has since been overwritten, so it'll end
1282 * up pointing to data that was never in 'x' during the history
1283 * of the device.
1284 *
1285 * To avoid this issue we require a commit as part of the
1286 * demotion operation.
1287 */
1288 init_continuation(&mg->k, mg_success);
1289 continue_after_commit(&cache->committer, &mg->k);
1290 schedule_commit(&cache->committer);
1291 break;
1292
1293 case POLICY_WRITEBACK:
1294 mg_complete(mg, true);
1295 break;
1296 }
1297}
1298
1299static void mg_update_metadata_after_copy(struct work_struct *ws)
1300{
1301 struct dm_cache_migration *mg = ws_to_mg(ws);
1302
1303 /*
1304 * Did the copy succeed?
1305 */
1306 if (mg->k.input)
1307 mg_complete(mg, false);
1308 else
1309 mg_update_metadata(ws);
1310}
1311
1312static void mg_upgrade_lock(struct work_struct *ws)
1313{
1314 int r;
1315 struct dm_cache_migration *mg = ws_to_mg(ws);
1316
1317 /*
1318 * Did the copy succeed?
1319 */
1320 if (mg->k.input)
1321 mg_complete(mg, false);
1322
1323 else {
1324 /*
1325 * Now we want the lock to prevent both reads and writes.
1326 */
1327 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1328 READ_WRITE_LOCK_LEVEL);
1329 if (r < 0)
1330 mg_complete(mg, false);
1331
1332 else if (r)
1333 quiesce(mg, mg_update_metadata);
1334
1335 else
1336 mg_update_metadata(ws);
1337 }
1338}
1339
1340static void mg_full_copy(struct work_struct *ws)
1341{
1342 struct dm_cache_migration *mg = ws_to_mg(ws);
1343 struct cache *cache = mg->cache;
1344 struct policy_work *op = mg->op;
1345 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1346
1347 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1348 is_discarded_oblock(cache, op->oblock)) {
1349 mg_upgrade_lock(ws);
1350 return;
1351 }
1352
1353 init_continuation(&mg->k, mg_upgrade_lock);
1354 copy(mg, is_policy_promote);
1355}
1356
1357static void mg_copy(struct work_struct *ws)
1358{
1359 struct dm_cache_migration *mg = ws_to_mg(ws);
1360
1361 if (mg->overwrite_bio) {
1362 /*
1363 * No exclusive lock was held when we last checked if the bio
1364 * was optimisable. So we have to check again in case things
1365 * have changed (eg, the block may no longer be discarded).
1366 */
1367 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1368 /*
1369 * Fallback to a real full copy after doing some tidying up.
1370 */
1371 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1372
1373 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1374 mg->overwrite_bio = NULL;
1375 inc_io_migrations(mg->cache);
1376 mg_full_copy(ws);
1377 return;
1378 }
1379
1380 /*
1381 * It's safe to do this here, even though it's new data
1382 * because all IO has been locked out of the block.
1383 *
1384 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1385 * so _not_ using mg_upgrade_lock() as continutation.
1386 */
1387 overwrite(mg, mg_update_metadata_after_copy);
1388
1389 } else
1390 mg_full_copy(ws);
1391}
1392
1393static int mg_lock_writes(struct dm_cache_migration *mg)
1394{
1395 int r;
1396 struct dm_cell_key_v2 key;
1397 struct cache *cache = mg->cache;
1398 struct dm_bio_prison_cell_v2 *prealloc;
1399
1400 prealloc = alloc_prison_cell(cache);
1401
1402 /*
1403 * Prevent writes to the block, but allow reads to continue.
1404 * Unless we're using an overwrite bio, in which case we lock
1405 * everything.
1406 */
1407 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1408 r = dm_cell_lock_v2(cache->prison, &key,
1409 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1410 prealloc, &mg->cell);
1411 if (r < 0) {
1412 free_prison_cell(cache, prealloc);
1413 mg_complete(mg, false);
1414 return r;
1415 }
1416
1417 if (mg->cell != prealloc)
1418 free_prison_cell(cache, prealloc);
1419
1420 if (r == 0)
1421 mg_copy(&mg->k.ws);
1422 else
1423 quiesce(mg, mg_copy);
1424
1425 return 0;
1426}
1427
1428static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1429{
1430 struct dm_cache_migration *mg;
1431
1432 if (!background_work_begin(cache)) {
1433 policy_complete_background_work(cache->policy, op, false);
1434 return -EPERM;
1435 }
1436
1437 mg = alloc_migration(cache);
1438
1439 mg->op = op;
1440 mg->overwrite_bio = bio;
1441
1442 if (!bio)
1443 inc_io_migrations(cache);
1444
1445 return mg_lock_writes(mg);
1446}
1447
1448/*
1449 *--------------------------------------------------------------
1450 * invalidation processing
1451 *--------------------------------------------------------------
1452 */
1453
1454static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1455{
1456 struct bio_list bios;
1457 struct cache *cache = mg->cache;
1458
1459 bio_list_init(&bios);
1460 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1461 free_prison_cell(cache, mg->cell);
1462
1463 if (!success && mg->overwrite_bio)
1464 bio_io_error(mg->overwrite_bio);
1465
1466 free_migration(mg);
1467 defer_bios(cache, &bios);
1468
1469 background_work_end(cache);
1470}
1471
1472static void invalidate_completed(struct work_struct *ws)
1473{
1474 struct dm_cache_migration *mg = ws_to_mg(ws);
1475
1476 invalidate_complete(mg, !mg->k.input);
1477}
1478
1479static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1480{
1481 int r;
1482
1483 r = policy_invalidate_mapping(cache->policy, cblock);
1484 if (!r) {
1485 r = dm_cache_remove_mapping(cache->cmd, cblock);
1486 if (r) {
1487 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1488 cache_device_name(cache));
1489 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1490 }
1491
1492 } else if (r == -ENODATA) {
1493 /*
1494 * Harmless, already unmapped.
1495 */
1496 r = 0;
1497
1498 } else
1499 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1500
1501 return r;
1502}
1503
1504static void invalidate_remove(struct work_struct *ws)
1505{
1506 int r;
1507 struct dm_cache_migration *mg = ws_to_mg(ws);
1508 struct cache *cache = mg->cache;
1509
1510 r = invalidate_cblock(cache, mg->invalidate_cblock);
1511 if (r) {
1512 invalidate_complete(mg, false);
1513 return;
1514 }
1515
1516 init_continuation(&mg->k, invalidate_completed);
1517 continue_after_commit(&cache->committer, &mg->k);
1518 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1519 mg->overwrite_bio = NULL;
1520 schedule_commit(&cache->committer);
1521}
1522
1523static int invalidate_lock(struct dm_cache_migration *mg)
1524{
1525 int r;
1526 struct dm_cell_key_v2 key;
1527 struct cache *cache = mg->cache;
1528 struct dm_bio_prison_cell_v2 *prealloc;
1529
1530 prealloc = alloc_prison_cell(cache);
1531
1532 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1533 r = dm_cell_lock_v2(cache->prison, &key,
1534 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1535 if (r < 0) {
1536 free_prison_cell(cache, prealloc);
1537 invalidate_complete(mg, false);
1538 return r;
1539 }
1540
1541 if (mg->cell != prealloc)
1542 free_prison_cell(cache, prealloc);
1543
1544 if (r)
1545 quiesce(mg, invalidate_remove);
1546
1547 else {
1548 /*
1549 * We can't call invalidate_remove() directly here because we
1550 * might still be in request context.
1551 */
1552 init_continuation(&mg->k, invalidate_remove);
1553 queue_work(cache->wq, &mg->k.ws);
1554 }
1555
1556 return 0;
1557}
1558
1559static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1560 dm_oblock_t oblock, struct bio *bio)
1561{
1562 struct dm_cache_migration *mg;
1563
1564 if (!background_work_begin(cache))
1565 return -EPERM;
1566
1567 mg = alloc_migration(cache);
1568
1569 mg->overwrite_bio = bio;
1570 mg->invalidate_cblock = cblock;
1571 mg->invalidate_oblock = oblock;
1572
1573 return invalidate_lock(mg);
1574}
1575
1576/*
1577 *--------------------------------------------------------------
1578 * bio processing
1579 *--------------------------------------------------------------
1580 */
1581
1582enum busy {
1583 IDLE,
1584 BUSY
1585};
1586
1587static enum busy spare_migration_bandwidth(struct cache *cache)
1588{
1589 bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1590 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1591 cache->sectors_per_block;
1592
1593 if (idle && current_volume <= cache->migration_threshold)
1594 return IDLE;
1595 else
1596 return BUSY;
1597}
1598
1599static void inc_hit_counter(struct cache *cache, struct bio *bio)
1600{
1601 atomic_inc(bio_data_dir(bio) == READ ?
1602 &cache->stats.read_hit : &cache->stats.write_hit);
1603}
1604
1605static void inc_miss_counter(struct cache *cache, struct bio *bio)
1606{
1607 atomic_inc(bio_data_dir(bio) == READ ?
1608 &cache->stats.read_miss : &cache->stats.write_miss);
1609}
1610
1611/*----------------------------------------------------------------*/
1612
1613static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1614 bool *commit_needed)
1615{
1616 int r, data_dir;
1617 bool rb, background_queued;
1618 dm_cblock_t cblock;
1619
1620 *commit_needed = false;
1621
1622 rb = bio_detain_shared(cache, block, bio);
1623 if (!rb) {
1624 /*
1625 * An exclusive lock is held for this block, so we have to
1626 * wait. We set the commit_needed flag so the current
1627 * transaction will be committed asap, allowing this lock
1628 * to be dropped.
1629 */
1630 *commit_needed = true;
1631 return DM_MAPIO_SUBMITTED;
1632 }
1633
1634 data_dir = bio_data_dir(bio);
1635
1636 if (optimisable_bio(cache, bio, block)) {
1637 struct policy_work *op = NULL;
1638
1639 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1640 if (unlikely(r && r != -ENOENT)) {
1641 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1642 cache_device_name(cache), r);
1643 bio_io_error(bio);
1644 return DM_MAPIO_SUBMITTED;
1645 }
1646
1647 if (r == -ENOENT && op) {
1648 bio_drop_shared_lock(cache, bio);
1649 BUG_ON(op->op != POLICY_PROMOTE);
1650 mg_start(cache, op, bio);
1651 return DM_MAPIO_SUBMITTED;
1652 }
1653 } else {
1654 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1655 if (unlikely(r && r != -ENOENT)) {
1656 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1657 cache_device_name(cache), r);
1658 bio_io_error(bio);
1659 return DM_MAPIO_SUBMITTED;
1660 }
1661
1662 if (background_queued)
1663 wake_migration_worker(cache);
1664 }
1665
1666 if (r == -ENOENT) {
1667 struct per_bio_data *pb = get_per_bio_data(bio);
1668
1669 /*
1670 * Miss.
1671 */
1672 inc_miss_counter(cache, bio);
1673 if (pb->req_nr == 0) {
1674 accounted_begin(cache, bio);
1675 remap_to_origin_clear_discard(cache, bio, block);
1676 } else {
1677 /*
1678 * This is a duplicate writethrough io that is no
1679 * longer needed because the block has been demoted.
1680 */
1681 bio_endio(bio);
1682 return DM_MAPIO_SUBMITTED;
1683 }
1684 } else {
1685 /*
1686 * Hit.
1687 */
1688 inc_hit_counter(cache, bio);
1689
1690 /*
1691 * Passthrough always maps to the origin, invalidating any
1692 * cache blocks that are written to.
1693 */
1694 if (passthrough_mode(cache)) {
1695 if (bio_data_dir(bio) == WRITE) {
1696 bio_drop_shared_lock(cache, bio);
1697 atomic_inc(&cache->stats.demotion);
1698 invalidate_start(cache, cblock, block, bio);
1699 } else
1700 remap_to_origin_clear_discard(cache, bio, block);
1701 } else {
1702 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1703 !is_dirty(cache, cblock)) {
1704 remap_to_origin_and_cache(cache, bio, block, cblock);
1705 accounted_begin(cache, bio);
1706 } else
1707 remap_to_cache_dirty(cache, bio, block, cblock);
1708 }
1709 }
1710
1711 /*
1712 * dm core turns FUA requests into a separate payload and FLUSH req.
1713 */
1714 if (bio->bi_opf & REQ_FUA) {
1715 /*
1716 * issue_after_commit will call accounted_begin a second time. So
1717 * we call accounted_complete() to avoid double accounting.
1718 */
1719 accounted_complete(cache, bio);
1720 issue_after_commit(&cache->committer, bio);
1721 *commit_needed = true;
1722 return DM_MAPIO_SUBMITTED;
1723 }
1724
1725 return DM_MAPIO_REMAPPED;
1726}
1727
1728static bool process_bio(struct cache *cache, struct bio *bio)
1729{
1730 bool commit_needed;
1731
1732 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1733 dm_submit_bio_remap(bio, NULL);
1734
1735 return commit_needed;
1736}
1737
1738/*
1739 * A non-zero return indicates read_only or fail_io mode.
1740 */
1741static int commit(struct cache *cache, bool clean_shutdown)
1742{
1743 int r;
1744
1745 if (get_cache_mode(cache) >= CM_READ_ONLY)
1746 return -EINVAL;
1747
1748 atomic_inc(&cache->stats.commit_count);
1749 r = dm_cache_commit(cache->cmd, clean_shutdown);
1750 if (r)
1751 metadata_operation_failed(cache, "dm_cache_commit", r);
1752
1753 return r;
1754}
1755
1756/*
1757 * Used by the batcher.
1758 */
1759static blk_status_t commit_op(void *context)
1760{
1761 struct cache *cache = context;
1762
1763 if (dm_cache_changed_this_transaction(cache->cmd))
1764 return errno_to_blk_status(commit(cache, false));
1765
1766 return 0;
1767}
1768
1769/*----------------------------------------------------------------*/
1770
1771static bool process_flush_bio(struct cache *cache, struct bio *bio)
1772{
1773 struct per_bio_data *pb = get_per_bio_data(bio);
1774
1775 if (!pb->req_nr)
1776 remap_to_origin(cache, bio);
1777 else
1778 remap_to_cache(cache, bio, 0);
1779
1780 issue_after_commit(&cache->committer, bio);
1781 return true;
1782}
1783
1784static bool process_discard_bio(struct cache *cache, struct bio *bio)
1785{
1786 dm_dblock_t b, e;
1787
1788 /*
1789 * FIXME: do we need to lock the region? Or can we just assume the
1790 * user wont be so foolish as to issue discard concurrently with
1791 * other IO?
1792 */
1793 calc_discard_block_range(cache, bio, &b, &e);
1794 while (b != e) {
1795 set_discard(cache, b);
1796 b = to_dblock(from_dblock(b) + 1);
1797 }
1798
1799 if (cache->features.discard_passdown) {
1800 remap_to_origin(cache, bio);
1801 dm_submit_bio_remap(bio, NULL);
1802 } else
1803 bio_endio(bio);
1804
1805 return false;
1806}
1807
1808static void process_deferred_bios(struct work_struct *ws)
1809{
1810 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1811
1812 bool commit_needed = false;
1813 struct bio_list bios;
1814 struct bio *bio;
1815
1816 bio_list_init(&bios);
1817
1818 spin_lock_irq(&cache->lock);
1819 bio_list_merge(&bios, &cache->deferred_bios);
1820 bio_list_init(&cache->deferred_bios);
1821 spin_unlock_irq(&cache->lock);
1822
1823 while ((bio = bio_list_pop(&bios))) {
1824 if (bio->bi_opf & REQ_PREFLUSH)
1825 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1826
1827 else if (bio_op(bio) == REQ_OP_DISCARD)
1828 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1829
1830 else
1831 commit_needed = process_bio(cache, bio) || commit_needed;
1832 cond_resched();
1833 }
1834
1835 if (commit_needed)
1836 schedule_commit(&cache->committer);
1837}
1838
1839/*
1840 *--------------------------------------------------------------
1841 * Main worker loop
1842 *--------------------------------------------------------------
1843 */
1844static void requeue_deferred_bios(struct cache *cache)
1845{
1846 struct bio *bio;
1847 struct bio_list bios;
1848
1849 bio_list_init(&bios);
1850 bio_list_merge(&bios, &cache->deferred_bios);
1851 bio_list_init(&cache->deferred_bios);
1852
1853 while ((bio = bio_list_pop(&bios))) {
1854 bio->bi_status = BLK_STS_DM_REQUEUE;
1855 bio_endio(bio);
1856 cond_resched();
1857 }
1858}
1859
1860/*
1861 * We want to commit periodically so that not too much
1862 * unwritten metadata builds up.
1863 */
1864static void do_waker(struct work_struct *ws)
1865{
1866 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1867
1868 policy_tick(cache->policy, true);
1869 wake_migration_worker(cache);
1870 schedule_commit(&cache->committer);
1871 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1872}
1873
1874static void check_migrations(struct work_struct *ws)
1875{
1876 int r;
1877 struct policy_work *op;
1878 struct cache *cache = container_of(ws, struct cache, migration_worker);
1879 enum busy b;
1880
1881 for (;;) {
1882 b = spare_migration_bandwidth(cache);
1883
1884 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1885 if (r == -ENODATA)
1886 break;
1887
1888 if (r) {
1889 DMERR_LIMIT("%s: policy_background_work failed",
1890 cache_device_name(cache));
1891 break;
1892 }
1893
1894 r = mg_start(cache, op, NULL);
1895 if (r)
1896 break;
1897
1898 cond_resched();
1899 }
1900}
1901
1902/*
1903 *--------------------------------------------------------------
1904 * Target methods
1905 *--------------------------------------------------------------
1906 */
1907
1908/*
1909 * This function gets called on the error paths of the constructor, so we
1910 * have to cope with a partially initialised struct.
1911 */
1912static void destroy(struct cache *cache)
1913{
1914 unsigned int i;
1915
1916 mempool_exit(&cache->migration_pool);
1917
1918 if (cache->prison)
1919 dm_bio_prison_destroy_v2(cache->prison);
1920
1921 cancel_delayed_work_sync(&cache->waker);
1922 if (cache->wq)
1923 destroy_workqueue(cache->wq);
1924
1925 if (cache->dirty_bitset)
1926 free_bitset(cache->dirty_bitset);
1927
1928 if (cache->discard_bitset)
1929 free_bitset(cache->discard_bitset);
1930
1931 if (cache->copier)
1932 dm_kcopyd_client_destroy(cache->copier);
1933
1934 if (cache->cmd)
1935 dm_cache_metadata_close(cache->cmd);
1936
1937 if (cache->metadata_dev)
1938 dm_put_device(cache->ti, cache->metadata_dev);
1939
1940 if (cache->origin_dev)
1941 dm_put_device(cache->ti, cache->origin_dev);
1942
1943 if (cache->cache_dev)
1944 dm_put_device(cache->ti, cache->cache_dev);
1945
1946 if (cache->policy)
1947 dm_cache_policy_destroy(cache->policy);
1948
1949 for (i = 0; i < cache->nr_ctr_args ; i++)
1950 kfree(cache->ctr_args[i]);
1951 kfree(cache->ctr_args);
1952
1953 bioset_exit(&cache->bs);
1954
1955 kfree(cache);
1956}
1957
1958static void cache_dtr(struct dm_target *ti)
1959{
1960 struct cache *cache = ti->private;
1961
1962 destroy(cache);
1963}
1964
1965static sector_t get_dev_size(struct dm_dev *dev)
1966{
1967 return bdev_nr_sectors(dev->bdev);
1968}
1969
1970/*----------------------------------------------------------------*/
1971
1972/*
1973 * Construct a cache device mapping.
1974 *
1975 * cache <metadata dev> <cache dev> <origin dev> <block size>
1976 * <#feature args> [<feature arg>]*
1977 * <policy> <#policy args> [<policy arg>]*
1978 *
1979 * metadata dev : fast device holding the persistent metadata
1980 * cache dev : fast device holding cached data blocks
1981 * origin dev : slow device holding original data blocks
1982 * block size : cache unit size in sectors
1983 *
1984 * #feature args : number of feature arguments passed
1985 * feature args : writethrough. (The default is writeback.)
1986 *
1987 * policy : the replacement policy to use
1988 * #policy args : an even number of policy arguments corresponding
1989 * to key/value pairs passed to the policy
1990 * policy args : key/value pairs passed to the policy
1991 * E.g. 'sequential_threshold 1024'
1992 * See cache-policies.txt for details.
1993 *
1994 * Optional feature arguments are:
1995 * writethrough : write through caching that prohibits cache block
1996 * content from being different from origin block content.
1997 * Without this argument, the default behaviour is to write
1998 * back cache block contents later for performance reasons,
1999 * so they may differ from the corresponding origin blocks.
2000 */
2001struct cache_args {
2002 struct dm_target *ti;
2003
2004 struct dm_dev *metadata_dev;
2005
2006 struct dm_dev *cache_dev;
2007 sector_t cache_sectors;
2008
2009 struct dm_dev *origin_dev;
2010 sector_t origin_sectors;
2011
2012 uint32_t block_size;
2013
2014 const char *policy_name;
2015 int policy_argc;
2016 const char **policy_argv;
2017
2018 struct cache_features features;
2019};
2020
2021static void destroy_cache_args(struct cache_args *ca)
2022{
2023 if (ca->metadata_dev)
2024 dm_put_device(ca->ti, ca->metadata_dev);
2025
2026 if (ca->cache_dev)
2027 dm_put_device(ca->ti, ca->cache_dev);
2028
2029 if (ca->origin_dev)
2030 dm_put_device(ca->ti, ca->origin_dev);
2031
2032 kfree(ca);
2033}
2034
2035static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2036{
2037 if (!as->argc) {
2038 *error = "Insufficient args";
2039 return false;
2040 }
2041
2042 return true;
2043}
2044
2045static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2046 char **error)
2047{
2048 int r;
2049 sector_t metadata_dev_size;
2050
2051 if (!at_least_one_arg(as, error))
2052 return -EINVAL;
2053
2054 r = dm_get_device(ca->ti, dm_shift_arg(as),
2055 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->metadata_dev);
2056 if (r) {
2057 *error = "Error opening metadata device";
2058 return r;
2059 }
2060
2061 metadata_dev_size = get_dev_size(ca->metadata_dev);
2062 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2063 DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2064 ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2065
2066 return 0;
2067}
2068
2069static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2070 char **error)
2071{
2072 int r;
2073
2074 if (!at_least_one_arg(as, error))
2075 return -EINVAL;
2076
2077 r = dm_get_device(ca->ti, dm_shift_arg(as),
2078 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->cache_dev);
2079 if (r) {
2080 *error = "Error opening cache device";
2081 return r;
2082 }
2083 ca->cache_sectors = get_dev_size(ca->cache_dev);
2084
2085 return 0;
2086}
2087
2088static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2089 char **error)
2090{
2091 int r;
2092
2093 if (!at_least_one_arg(as, error))
2094 return -EINVAL;
2095
2096 r = dm_get_device(ca->ti, dm_shift_arg(as),
2097 BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->origin_dev);
2098 if (r) {
2099 *error = "Error opening origin device";
2100 return r;
2101 }
2102
2103 ca->origin_sectors = get_dev_size(ca->origin_dev);
2104 if (ca->ti->len > ca->origin_sectors) {
2105 *error = "Device size larger than cached device";
2106 return -EINVAL;
2107 }
2108
2109 return 0;
2110}
2111
2112static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2113 char **error)
2114{
2115 unsigned long block_size;
2116
2117 if (!at_least_one_arg(as, error))
2118 return -EINVAL;
2119
2120 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2121 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2122 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2123 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2124 *error = "Invalid data block size";
2125 return -EINVAL;
2126 }
2127
2128 if (block_size > ca->cache_sectors) {
2129 *error = "Data block size is larger than the cache device";
2130 return -EINVAL;
2131 }
2132
2133 ca->block_size = block_size;
2134
2135 return 0;
2136}
2137
2138static void init_features(struct cache_features *cf)
2139{
2140 cf->mode = CM_WRITE;
2141 cf->io_mode = CM_IO_WRITEBACK;
2142 cf->metadata_version = 1;
2143 cf->discard_passdown = true;
2144}
2145
2146static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2147 char **error)
2148{
2149 static const struct dm_arg _args[] = {
2150 {0, 3, "Invalid number of cache feature arguments"},
2151 };
2152
2153 int r, mode_ctr = 0;
2154 unsigned int argc;
2155 const char *arg;
2156 struct cache_features *cf = &ca->features;
2157
2158 init_features(cf);
2159
2160 r = dm_read_arg_group(_args, as, &argc, error);
2161 if (r)
2162 return -EINVAL;
2163
2164 while (argc--) {
2165 arg = dm_shift_arg(as);
2166
2167 if (!strcasecmp(arg, "writeback")) {
2168 cf->io_mode = CM_IO_WRITEBACK;
2169 mode_ctr++;
2170 }
2171
2172 else if (!strcasecmp(arg, "writethrough")) {
2173 cf->io_mode = CM_IO_WRITETHROUGH;
2174 mode_ctr++;
2175 }
2176
2177 else if (!strcasecmp(arg, "passthrough")) {
2178 cf->io_mode = CM_IO_PASSTHROUGH;
2179 mode_ctr++;
2180 }
2181
2182 else if (!strcasecmp(arg, "metadata2"))
2183 cf->metadata_version = 2;
2184
2185 else if (!strcasecmp(arg, "no_discard_passdown"))
2186 cf->discard_passdown = false;
2187
2188 else {
2189 *error = "Unrecognised cache feature requested";
2190 return -EINVAL;
2191 }
2192 }
2193
2194 if (mode_ctr > 1) {
2195 *error = "Duplicate cache io_mode features requested";
2196 return -EINVAL;
2197 }
2198
2199 return 0;
2200}
2201
2202static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2203 char **error)
2204{
2205 static const struct dm_arg _args[] = {
2206 {0, 1024, "Invalid number of policy arguments"},
2207 };
2208
2209 int r;
2210
2211 if (!at_least_one_arg(as, error))
2212 return -EINVAL;
2213
2214 ca->policy_name = dm_shift_arg(as);
2215
2216 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2217 if (r)
2218 return -EINVAL;
2219
2220 ca->policy_argv = (const char **)as->argv;
2221 dm_consume_args(as, ca->policy_argc);
2222
2223 return 0;
2224}
2225
2226static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2227 char **error)
2228{
2229 int r;
2230 struct dm_arg_set as;
2231
2232 as.argc = argc;
2233 as.argv = argv;
2234
2235 r = parse_metadata_dev(ca, &as, error);
2236 if (r)
2237 return r;
2238
2239 r = parse_cache_dev(ca, &as, error);
2240 if (r)
2241 return r;
2242
2243 r = parse_origin_dev(ca, &as, error);
2244 if (r)
2245 return r;
2246
2247 r = parse_block_size(ca, &as, error);
2248 if (r)
2249 return r;
2250
2251 r = parse_features(ca, &as, error);
2252 if (r)
2253 return r;
2254
2255 r = parse_policy(ca, &as, error);
2256 if (r)
2257 return r;
2258
2259 return 0;
2260}
2261
2262/*----------------------------------------------------------------*/
2263
2264static struct kmem_cache *migration_cache;
2265
2266#define NOT_CORE_OPTION 1
2267
2268static int process_config_option(struct cache *cache, const char *key, const char *value)
2269{
2270 unsigned long tmp;
2271
2272 if (!strcasecmp(key, "migration_threshold")) {
2273 if (kstrtoul(value, 10, &tmp))
2274 return -EINVAL;
2275
2276 cache->migration_threshold = tmp;
2277 return 0;
2278 }
2279
2280 return NOT_CORE_OPTION;
2281}
2282
2283static int set_config_value(struct cache *cache, const char *key, const char *value)
2284{
2285 int r = process_config_option(cache, key, value);
2286
2287 if (r == NOT_CORE_OPTION)
2288 r = policy_set_config_value(cache->policy, key, value);
2289
2290 if (r)
2291 DMWARN("bad config value for %s: %s", key, value);
2292
2293 return r;
2294}
2295
2296static int set_config_values(struct cache *cache, int argc, const char **argv)
2297{
2298 int r = 0;
2299
2300 if (argc & 1) {
2301 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2302 return -EINVAL;
2303 }
2304
2305 while (argc) {
2306 r = set_config_value(cache, argv[0], argv[1]);
2307 if (r)
2308 break;
2309
2310 argc -= 2;
2311 argv += 2;
2312 }
2313
2314 return r;
2315}
2316
2317static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2318 char **error)
2319{
2320 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2321 cache->cache_size,
2322 cache->origin_sectors,
2323 cache->sectors_per_block);
2324 if (IS_ERR(p)) {
2325 *error = "Error creating cache's policy";
2326 return PTR_ERR(p);
2327 }
2328 cache->policy = p;
2329 BUG_ON(!cache->policy);
2330
2331 return 0;
2332}
2333
2334/*
2335 * We want the discard block size to be at least the size of the cache
2336 * block size and have no more than 2^14 discard blocks across the origin.
2337 */
2338#define MAX_DISCARD_BLOCKS (1 << 14)
2339
2340static bool too_many_discard_blocks(sector_t discard_block_size,
2341 sector_t origin_size)
2342{
2343 (void) sector_div(origin_size, discard_block_size);
2344
2345 return origin_size > MAX_DISCARD_BLOCKS;
2346}
2347
2348static sector_t calculate_discard_block_size(sector_t cache_block_size,
2349 sector_t origin_size)
2350{
2351 sector_t discard_block_size = cache_block_size;
2352
2353 if (origin_size)
2354 while (too_many_discard_blocks(discard_block_size, origin_size))
2355 discard_block_size *= 2;
2356
2357 return discard_block_size;
2358}
2359
2360static void set_cache_size(struct cache *cache, dm_cblock_t size)
2361{
2362 dm_block_t nr_blocks = from_cblock(size);
2363
2364 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2365 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2366 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2367 "Please consider increasing the cache block size to reduce the overall cache block count.",
2368 (unsigned long long) nr_blocks);
2369
2370 cache->cache_size = size;
2371}
2372
2373#define DEFAULT_MIGRATION_THRESHOLD 2048
2374
2375static int cache_create(struct cache_args *ca, struct cache **result)
2376{
2377 int r = 0;
2378 char **error = &ca->ti->error;
2379 struct cache *cache;
2380 struct dm_target *ti = ca->ti;
2381 dm_block_t origin_blocks;
2382 struct dm_cache_metadata *cmd;
2383 bool may_format = ca->features.mode == CM_WRITE;
2384
2385 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2386 if (!cache)
2387 return -ENOMEM;
2388
2389 cache->ti = ca->ti;
2390 ti->private = cache;
2391 ti->accounts_remapped_io = true;
2392 ti->num_flush_bios = 2;
2393 ti->flush_supported = true;
2394
2395 ti->num_discard_bios = 1;
2396 ti->discards_supported = true;
2397
2398 ti->per_io_data_size = sizeof(struct per_bio_data);
2399
2400 cache->features = ca->features;
2401 if (writethrough_mode(cache)) {
2402 /* Create bioset for writethrough bios issued to origin */
2403 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2404 if (r)
2405 goto bad;
2406 }
2407
2408 cache->metadata_dev = ca->metadata_dev;
2409 cache->origin_dev = ca->origin_dev;
2410 cache->cache_dev = ca->cache_dev;
2411
2412 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2413
2414 origin_blocks = cache->origin_sectors = ca->origin_sectors;
2415 origin_blocks = block_div(origin_blocks, ca->block_size);
2416 cache->origin_blocks = to_oblock(origin_blocks);
2417
2418 cache->sectors_per_block = ca->block_size;
2419 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2420 r = -EINVAL;
2421 goto bad;
2422 }
2423
2424 if (ca->block_size & (ca->block_size - 1)) {
2425 dm_block_t cache_size = ca->cache_sectors;
2426
2427 cache->sectors_per_block_shift = -1;
2428 cache_size = block_div(cache_size, ca->block_size);
2429 set_cache_size(cache, to_cblock(cache_size));
2430 } else {
2431 cache->sectors_per_block_shift = __ffs(ca->block_size);
2432 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2433 }
2434
2435 r = create_cache_policy(cache, ca, error);
2436 if (r)
2437 goto bad;
2438
2439 cache->policy_nr_args = ca->policy_argc;
2440 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2441
2442 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2443 if (r) {
2444 *error = "Error setting cache policy's config values";
2445 goto bad;
2446 }
2447
2448 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2449 ca->block_size, may_format,
2450 dm_cache_policy_get_hint_size(cache->policy),
2451 ca->features.metadata_version);
2452 if (IS_ERR(cmd)) {
2453 *error = "Error creating metadata object";
2454 r = PTR_ERR(cmd);
2455 goto bad;
2456 }
2457 cache->cmd = cmd;
2458 set_cache_mode(cache, CM_WRITE);
2459 if (get_cache_mode(cache) != CM_WRITE) {
2460 *error = "Unable to get write access to metadata, please check/repair metadata.";
2461 r = -EINVAL;
2462 goto bad;
2463 }
2464
2465 if (passthrough_mode(cache)) {
2466 bool all_clean;
2467
2468 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2469 if (r) {
2470 *error = "dm_cache_metadata_all_clean() failed";
2471 goto bad;
2472 }
2473
2474 if (!all_clean) {
2475 *error = "Cannot enter passthrough mode unless all blocks are clean";
2476 r = -EINVAL;
2477 goto bad;
2478 }
2479
2480 policy_allow_migrations(cache->policy, false);
2481 }
2482
2483 spin_lock_init(&cache->lock);
2484 bio_list_init(&cache->deferred_bios);
2485 atomic_set(&cache->nr_allocated_migrations, 0);
2486 atomic_set(&cache->nr_io_migrations, 0);
2487 init_waitqueue_head(&cache->migration_wait);
2488
2489 r = -ENOMEM;
2490 atomic_set(&cache->nr_dirty, 0);
2491 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2492 if (!cache->dirty_bitset) {
2493 *error = "could not allocate dirty bitset";
2494 goto bad;
2495 }
2496 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2497
2498 cache->discard_block_size =
2499 calculate_discard_block_size(cache->sectors_per_block,
2500 cache->origin_sectors);
2501 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2502 cache->discard_block_size));
2503 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2504 if (!cache->discard_bitset) {
2505 *error = "could not allocate discard bitset";
2506 goto bad;
2507 }
2508 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2509
2510 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2511 if (IS_ERR(cache->copier)) {
2512 *error = "could not create kcopyd client";
2513 r = PTR_ERR(cache->copier);
2514 goto bad;
2515 }
2516
2517 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2518 if (!cache->wq) {
2519 *error = "could not create workqueue for metadata object";
2520 goto bad;
2521 }
2522 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2523 INIT_WORK(&cache->migration_worker, check_migrations);
2524 INIT_DELAYED_WORK(&cache->waker, do_waker);
2525
2526 cache->prison = dm_bio_prison_create_v2(cache->wq);
2527 if (!cache->prison) {
2528 *error = "could not create bio prison";
2529 goto bad;
2530 }
2531
2532 r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2533 migration_cache);
2534 if (r) {
2535 *error = "Error creating cache's migration mempool";
2536 goto bad;
2537 }
2538
2539 cache->need_tick_bio = true;
2540 cache->sized = false;
2541 cache->invalidate = false;
2542 cache->commit_requested = false;
2543 cache->loaded_mappings = false;
2544 cache->loaded_discards = false;
2545
2546 load_stats(cache);
2547
2548 atomic_set(&cache->stats.demotion, 0);
2549 atomic_set(&cache->stats.promotion, 0);
2550 atomic_set(&cache->stats.copies_avoided, 0);
2551 atomic_set(&cache->stats.cache_cell_clash, 0);
2552 atomic_set(&cache->stats.commit_count, 0);
2553 atomic_set(&cache->stats.discard_count, 0);
2554
2555 spin_lock_init(&cache->invalidation_lock);
2556 INIT_LIST_HEAD(&cache->invalidation_requests);
2557
2558 batcher_init(&cache->committer, commit_op, cache,
2559 issue_op, cache, cache->wq);
2560 dm_iot_init(&cache->tracker);
2561
2562 init_rwsem(&cache->background_work_lock);
2563 prevent_background_work(cache);
2564
2565 *result = cache;
2566 return 0;
2567bad:
2568 destroy(cache);
2569 return r;
2570}
2571
2572static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2573{
2574 unsigned int i;
2575 const char **copy;
2576
2577 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2578 if (!copy)
2579 return -ENOMEM;
2580 for (i = 0; i < argc; i++) {
2581 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2582 if (!copy[i]) {
2583 while (i--)
2584 kfree(copy[i]);
2585 kfree(copy);
2586 return -ENOMEM;
2587 }
2588 }
2589
2590 cache->nr_ctr_args = argc;
2591 cache->ctr_args = copy;
2592
2593 return 0;
2594}
2595
2596static int cache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2597{
2598 int r = -EINVAL;
2599 struct cache_args *ca;
2600 struct cache *cache = NULL;
2601
2602 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2603 if (!ca) {
2604 ti->error = "Error allocating memory for cache";
2605 return -ENOMEM;
2606 }
2607 ca->ti = ti;
2608
2609 r = parse_cache_args(ca, argc, argv, &ti->error);
2610 if (r)
2611 goto out;
2612
2613 r = cache_create(ca, &cache);
2614 if (r)
2615 goto out;
2616
2617 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2618 if (r) {
2619 destroy(cache);
2620 goto out;
2621 }
2622
2623 ti->private = cache;
2624out:
2625 destroy_cache_args(ca);
2626 return r;
2627}
2628
2629/*----------------------------------------------------------------*/
2630
2631static int cache_map(struct dm_target *ti, struct bio *bio)
2632{
2633 struct cache *cache = ti->private;
2634
2635 int r;
2636 bool commit_needed;
2637 dm_oblock_t block = get_bio_block(cache, bio);
2638
2639 init_per_bio_data(bio);
2640 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2641 /*
2642 * This can only occur if the io goes to a partial block at
2643 * the end of the origin device. We don't cache these.
2644 * Just remap to the origin and carry on.
2645 */
2646 remap_to_origin(cache, bio);
2647 accounted_begin(cache, bio);
2648 return DM_MAPIO_REMAPPED;
2649 }
2650
2651 if (discard_or_flush(bio)) {
2652 defer_bio(cache, bio);
2653 return DM_MAPIO_SUBMITTED;
2654 }
2655
2656 r = map_bio(cache, bio, block, &commit_needed);
2657 if (commit_needed)
2658 schedule_commit(&cache->committer);
2659
2660 return r;
2661}
2662
2663static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2664{
2665 struct cache *cache = ti->private;
2666 unsigned long flags;
2667 struct per_bio_data *pb = get_per_bio_data(bio);
2668
2669 if (pb->tick) {
2670 policy_tick(cache->policy, false);
2671
2672 spin_lock_irqsave(&cache->lock, flags);
2673 cache->need_tick_bio = true;
2674 spin_unlock_irqrestore(&cache->lock, flags);
2675 }
2676
2677 bio_drop_shared_lock(cache, bio);
2678 accounted_complete(cache, bio);
2679
2680 return DM_ENDIO_DONE;
2681}
2682
2683static int write_dirty_bitset(struct cache *cache)
2684{
2685 int r;
2686
2687 if (get_cache_mode(cache) >= CM_READ_ONLY)
2688 return -EINVAL;
2689
2690 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2691 if (r)
2692 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2693
2694 return r;
2695}
2696
2697static int write_discard_bitset(struct cache *cache)
2698{
2699 unsigned int i, r;
2700
2701 if (get_cache_mode(cache) >= CM_READ_ONLY)
2702 return -EINVAL;
2703
2704 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2705 cache->discard_nr_blocks);
2706 if (r) {
2707 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2708 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2709 return r;
2710 }
2711
2712 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2713 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2714 is_discarded(cache, to_dblock(i)));
2715 if (r) {
2716 metadata_operation_failed(cache, "dm_cache_set_discard", r);
2717 return r;
2718 }
2719 }
2720
2721 return 0;
2722}
2723
2724static int write_hints(struct cache *cache)
2725{
2726 int r;
2727
2728 if (get_cache_mode(cache) >= CM_READ_ONLY)
2729 return -EINVAL;
2730
2731 r = dm_cache_write_hints(cache->cmd, cache->policy);
2732 if (r) {
2733 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2734 return r;
2735 }
2736
2737 return 0;
2738}
2739
2740/*
2741 * returns true on success
2742 */
2743static bool sync_metadata(struct cache *cache)
2744{
2745 int r1, r2, r3, r4;
2746
2747 r1 = write_dirty_bitset(cache);
2748 if (r1)
2749 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2750
2751 r2 = write_discard_bitset(cache);
2752 if (r2)
2753 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2754
2755 save_stats(cache);
2756
2757 r3 = write_hints(cache);
2758 if (r3)
2759 DMERR("%s: could not write hints", cache_device_name(cache));
2760
2761 /*
2762 * If writing the above metadata failed, we still commit, but don't
2763 * set the clean shutdown flag. This will effectively force every
2764 * dirty bit to be set on reload.
2765 */
2766 r4 = commit(cache, !r1 && !r2 && !r3);
2767 if (r4)
2768 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2769
2770 return !r1 && !r2 && !r3 && !r4;
2771}
2772
2773static void cache_postsuspend(struct dm_target *ti)
2774{
2775 struct cache *cache = ti->private;
2776
2777 prevent_background_work(cache);
2778 BUG_ON(atomic_read(&cache->nr_io_migrations));
2779
2780 cancel_delayed_work_sync(&cache->waker);
2781 drain_workqueue(cache->wq);
2782 WARN_ON(cache->tracker.in_flight);
2783
2784 /*
2785 * If it's a flush suspend there won't be any deferred bios, so this
2786 * call is harmless.
2787 */
2788 requeue_deferred_bios(cache);
2789
2790 if (get_cache_mode(cache) == CM_WRITE)
2791 (void) sync_metadata(cache);
2792}
2793
2794static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2795 bool dirty, uint32_t hint, bool hint_valid)
2796{
2797 struct cache *cache = context;
2798
2799 if (dirty) {
2800 set_bit(from_cblock(cblock), cache->dirty_bitset);
2801 atomic_inc(&cache->nr_dirty);
2802 } else
2803 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2804
2805 return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2806}
2807
2808/*
2809 * The discard block size in the on disk metadata is not
2810 * necessarily the same as we're currently using. So we have to
2811 * be careful to only set the discarded attribute if we know it
2812 * covers a complete block of the new size.
2813 */
2814struct discard_load_info {
2815 struct cache *cache;
2816
2817 /*
2818 * These blocks are sized using the on disk dblock size, rather
2819 * than the current one.
2820 */
2821 dm_block_t block_size;
2822 dm_block_t discard_begin, discard_end;
2823};
2824
2825static void discard_load_info_init(struct cache *cache,
2826 struct discard_load_info *li)
2827{
2828 li->cache = cache;
2829 li->discard_begin = li->discard_end = 0;
2830}
2831
2832static void set_discard_range(struct discard_load_info *li)
2833{
2834 sector_t b, e;
2835
2836 if (li->discard_begin == li->discard_end)
2837 return;
2838
2839 /*
2840 * Convert to sectors.
2841 */
2842 b = li->discard_begin * li->block_size;
2843 e = li->discard_end * li->block_size;
2844
2845 /*
2846 * Then convert back to the current dblock size.
2847 */
2848 b = dm_sector_div_up(b, li->cache->discard_block_size);
2849 sector_div(e, li->cache->discard_block_size);
2850
2851 /*
2852 * The origin may have shrunk, so we need to check we're still in
2853 * bounds.
2854 */
2855 if (e > from_dblock(li->cache->discard_nr_blocks))
2856 e = from_dblock(li->cache->discard_nr_blocks);
2857
2858 for (; b < e; b++)
2859 set_discard(li->cache, to_dblock(b));
2860}
2861
2862static int load_discard(void *context, sector_t discard_block_size,
2863 dm_dblock_t dblock, bool discard)
2864{
2865 struct discard_load_info *li = context;
2866
2867 li->block_size = discard_block_size;
2868
2869 if (discard) {
2870 if (from_dblock(dblock) == li->discard_end)
2871 /*
2872 * We're already in a discard range, just extend it.
2873 */
2874 li->discard_end = li->discard_end + 1ULL;
2875
2876 else {
2877 /*
2878 * Emit the old range and start a new one.
2879 */
2880 set_discard_range(li);
2881 li->discard_begin = from_dblock(dblock);
2882 li->discard_end = li->discard_begin + 1ULL;
2883 }
2884 } else {
2885 set_discard_range(li);
2886 li->discard_begin = li->discard_end = 0;
2887 }
2888
2889 return 0;
2890}
2891
2892static dm_cblock_t get_cache_dev_size(struct cache *cache)
2893{
2894 sector_t size = get_dev_size(cache->cache_dev);
2895 (void) sector_div(size, cache->sectors_per_block);
2896 return to_cblock(size);
2897}
2898
2899static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2900{
2901 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2902 if (cache->sized) {
2903 DMERR("%s: unable to extend cache due to missing cache table reload",
2904 cache_device_name(cache));
2905 return false;
2906 }
2907 }
2908
2909 /*
2910 * We can't drop a dirty block when shrinking the cache.
2911 */
2912 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2913 new_size = to_cblock(from_cblock(new_size) + 1);
2914 if (is_dirty(cache, new_size)) {
2915 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2916 cache_device_name(cache),
2917 (unsigned long long) from_cblock(new_size));
2918 return false;
2919 }
2920 }
2921
2922 return true;
2923}
2924
2925static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2926{
2927 int r;
2928
2929 r = dm_cache_resize(cache->cmd, new_size);
2930 if (r) {
2931 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2932 metadata_operation_failed(cache, "dm_cache_resize", r);
2933 return r;
2934 }
2935
2936 set_cache_size(cache, new_size);
2937
2938 return 0;
2939}
2940
2941static int cache_preresume(struct dm_target *ti)
2942{
2943 int r = 0;
2944 struct cache *cache = ti->private;
2945 dm_cblock_t csize = get_cache_dev_size(cache);
2946
2947 /*
2948 * Check to see if the cache has resized.
2949 */
2950 if (!cache->sized) {
2951 r = resize_cache_dev(cache, csize);
2952 if (r)
2953 return r;
2954
2955 cache->sized = true;
2956
2957 } else if (csize != cache->cache_size) {
2958 if (!can_resize(cache, csize))
2959 return -EINVAL;
2960
2961 r = resize_cache_dev(cache, csize);
2962 if (r)
2963 return r;
2964 }
2965
2966 if (!cache->loaded_mappings) {
2967 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2968 load_mapping, cache);
2969 if (r) {
2970 DMERR("%s: could not load cache mappings", cache_device_name(cache));
2971 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2972 return r;
2973 }
2974
2975 cache->loaded_mappings = true;
2976 }
2977
2978 if (!cache->loaded_discards) {
2979 struct discard_load_info li;
2980
2981 /*
2982 * The discard bitset could have been resized, or the
2983 * discard block size changed. To be safe we start by
2984 * setting every dblock to not discarded.
2985 */
2986 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2987
2988 discard_load_info_init(cache, &li);
2989 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2990 if (r) {
2991 DMERR("%s: could not load origin discards", cache_device_name(cache));
2992 metadata_operation_failed(cache, "dm_cache_load_discards", r);
2993 return r;
2994 }
2995 set_discard_range(&li);
2996
2997 cache->loaded_discards = true;
2998 }
2999
3000 return r;
3001}
3002
3003static void cache_resume(struct dm_target *ti)
3004{
3005 struct cache *cache = ti->private;
3006
3007 cache->need_tick_bio = true;
3008 allow_background_work(cache);
3009 do_waker(&cache->waker.work);
3010}
3011
3012static void emit_flags(struct cache *cache, char *result,
3013 unsigned int maxlen, ssize_t *sz_ptr)
3014{
3015 ssize_t sz = *sz_ptr;
3016 struct cache_features *cf = &cache->features;
3017 unsigned int count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3018
3019 DMEMIT("%u ", count);
3020
3021 if (cf->metadata_version == 2)
3022 DMEMIT("metadata2 ");
3023
3024 if (writethrough_mode(cache))
3025 DMEMIT("writethrough ");
3026
3027 else if (passthrough_mode(cache))
3028 DMEMIT("passthrough ");
3029
3030 else if (writeback_mode(cache))
3031 DMEMIT("writeback ");
3032
3033 else {
3034 DMEMIT("unknown ");
3035 DMERR("%s: internal error: unknown io mode: %d",
3036 cache_device_name(cache), (int) cf->io_mode);
3037 }
3038
3039 if (!cf->discard_passdown)
3040 DMEMIT("no_discard_passdown ");
3041
3042 *sz_ptr = sz;
3043}
3044
3045/*
3046 * Status format:
3047 *
3048 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3049 * <cache block size> <#used cache blocks>/<#total cache blocks>
3050 * <#read hits> <#read misses> <#write hits> <#write misses>
3051 * <#demotions> <#promotions> <#dirty>
3052 * <#features> <features>*
3053 * <#core args> <core args>
3054 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3055 */
3056static void cache_status(struct dm_target *ti, status_type_t type,
3057 unsigned int status_flags, char *result, unsigned int maxlen)
3058{
3059 int r = 0;
3060 unsigned int i;
3061 ssize_t sz = 0;
3062 dm_block_t nr_free_blocks_metadata = 0;
3063 dm_block_t nr_blocks_metadata = 0;
3064 char buf[BDEVNAME_SIZE];
3065 struct cache *cache = ti->private;
3066 dm_cblock_t residency;
3067 bool needs_check;
3068
3069 switch (type) {
3070 case STATUSTYPE_INFO:
3071 if (get_cache_mode(cache) == CM_FAIL) {
3072 DMEMIT("Fail");
3073 break;
3074 }
3075
3076 /* Commit to ensure statistics aren't out-of-date */
3077 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3078 (void) commit(cache, false);
3079
3080 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3081 if (r) {
3082 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3083 cache_device_name(cache), r);
3084 goto err;
3085 }
3086
3087 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3088 if (r) {
3089 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3090 cache_device_name(cache), r);
3091 goto err;
3092 }
3093
3094 residency = policy_residency(cache->policy);
3095
3096 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3097 (unsigned int)DM_CACHE_METADATA_BLOCK_SIZE,
3098 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3099 (unsigned long long)nr_blocks_metadata,
3100 (unsigned long long)cache->sectors_per_block,
3101 (unsigned long long) from_cblock(residency),
3102 (unsigned long long) from_cblock(cache->cache_size),
3103 (unsigned int) atomic_read(&cache->stats.read_hit),
3104 (unsigned int) atomic_read(&cache->stats.read_miss),
3105 (unsigned int) atomic_read(&cache->stats.write_hit),
3106 (unsigned int) atomic_read(&cache->stats.write_miss),
3107 (unsigned int) atomic_read(&cache->stats.demotion),
3108 (unsigned int) atomic_read(&cache->stats.promotion),
3109 (unsigned long) atomic_read(&cache->nr_dirty));
3110
3111 emit_flags(cache, result, maxlen, &sz);
3112
3113 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3114
3115 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3116 if (sz < maxlen) {
3117 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3118 if (r)
3119 DMERR("%s: policy_emit_config_values returned %d",
3120 cache_device_name(cache), r);
3121 }
3122
3123 if (get_cache_mode(cache) == CM_READ_ONLY)
3124 DMEMIT("ro ");
3125 else
3126 DMEMIT("rw ");
3127
3128 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3129
3130 if (r || needs_check)
3131 DMEMIT("needs_check ");
3132 else
3133 DMEMIT("- ");
3134
3135 break;
3136
3137 case STATUSTYPE_TABLE:
3138 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3139 DMEMIT("%s ", buf);
3140 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3141 DMEMIT("%s ", buf);
3142 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3143 DMEMIT("%s", buf);
3144
3145 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3146 DMEMIT(" %s", cache->ctr_args[i]);
3147 if (cache->nr_ctr_args)
3148 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3149 break;
3150
3151 case STATUSTYPE_IMA:
3152 DMEMIT_TARGET_NAME_VERSION(ti->type);
3153 if (get_cache_mode(cache) == CM_FAIL)
3154 DMEMIT(",metadata_mode=fail");
3155 else if (get_cache_mode(cache) == CM_READ_ONLY)
3156 DMEMIT(",metadata_mode=ro");
3157 else
3158 DMEMIT(",metadata_mode=rw");
3159
3160 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3161 DMEMIT(",cache_metadata_device=%s", buf);
3162 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3163 DMEMIT(",cache_device=%s", buf);
3164 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3165 DMEMIT(",cache_origin_device=%s", buf);
3166 DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3167 DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3168 DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3169 DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3170 DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3171 DMEMIT(";");
3172 break;
3173 }
3174
3175 return;
3176
3177err:
3178 DMEMIT("Error");
3179}
3180
3181/*
3182 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3183 * the one-past-the-end value.
3184 */
3185struct cblock_range {
3186 dm_cblock_t begin;
3187 dm_cblock_t end;
3188};
3189
3190/*
3191 * A cache block range can take two forms:
3192 *
3193 * i) A single cblock, eg. '3456'
3194 * ii) A begin and end cblock with a dash between, eg. 123-234
3195 */
3196static int parse_cblock_range(struct cache *cache, const char *str,
3197 struct cblock_range *result)
3198{
3199 char dummy;
3200 uint64_t b, e;
3201 int r;
3202
3203 /*
3204 * Try and parse form (ii) first.
3205 */
3206 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3207 if (r < 0)
3208 return r;
3209
3210 if (r == 2) {
3211 result->begin = to_cblock(b);
3212 result->end = to_cblock(e);
3213 return 0;
3214 }
3215
3216 /*
3217 * That didn't work, try form (i).
3218 */
3219 r = sscanf(str, "%llu%c", &b, &dummy);
3220 if (r < 0)
3221 return r;
3222
3223 if (r == 1) {
3224 result->begin = to_cblock(b);
3225 result->end = to_cblock(from_cblock(result->begin) + 1u);
3226 return 0;
3227 }
3228
3229 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3230 return -EINVAL;
3231}
3232
3233static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3234{
3235 uint64_t b = from_cblock(range->begin);
3236 uint64_t e = from_cblock(range->end);
3237 uint64_t n = from_cblock(cache->cache_size);
3238
3239 if (b >= n) {
3240 DMERR("%s: begin cblock out of range: %llu >= %llu",
3241 cache_device_name(cache), b, n);
3242 return -EINVAL;
3243 }
3244
3245 if (e > n) {
3246 DMERR("%s: end cblock out of range: %llu > %llu",
3247 cache_device_name(cache), e, n);
3248 return -EINVAL;
3249 }
3250
3251 if (b >= e) {
3252 DMERR("%s: invalid cblock range: %llu >= %llu",
3253 cache_device_name(cache), b, e);
3254 return -EINVAL;
3255 }
3256
3257 return 0;
3258}
3259
3260static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3261{
3262 return to_cblock(from_cblock(b) + 1);
3263}
3264
3265static int request_invalidation(struct cache *cache, struct cblock_range *range)
3266{
3267 int r = 0;
3268
3269 /*
3270 * We don't need to do any locking here because we know we're in
3271 * passthrough mode. There's is potential for a race between an
3272 * invalidation triggered by an io and an invalidation message. This
3273 * is harmless, we must not worry if the policy call fails.
3274 */
3275 while (range->begin != range->end) {
3276 r = invalidate_cblock(cache, range->begin);
3277 if (r)
3278 return r;
3279
3280 range->begin = cblock_succ(range->begin);
3281 }
3282
3283 cache->commit_requested = true;
3284 return r;
3285}
3286
3287static int process_invalidate_cblocks_message(struct cache *cache, unsigned int count,
3288 const char **cblock_ranges)
3289{
3290 int r = 0;
3291 unsigned int i;
3292 struct cblock_range range;
3293
3294 if (!passthrough_mode(cache)) {
3295 DMERR("%s: cache has to be in passthrough mode for invalidation",
3296 cache_device_name(cache));
3297 return -EPERM;
3298 }
3299
3300 for (i = 0; i < count; i++) {
3301 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3302 if (r)
3303 break;
3304
3305 r = validate_cblock_range(cache, &range);
3306 if (r)
3307 break;
3308
3309 /*
3310 * Pass begin and end origin blocks to the worker and wake it.
3311 */
3312 r = request_invalidation(cache, &range);
3313 if (r)
3314 break;
3315 }
3316
3317 return r;
3318}
3319
3320/*
3321 * Supports
3322 * "<key> <value>"
3323 * and
3324 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3325 *
3326 * The key migration_threshold is supported by the cache target core.
3327 */
3328static int cache_message(struct dm_target *ti, unsigned int argc, char **argv,
3329 char *result, unsigned int maxlen)
3330{
3331 struct cache *cache = ti->private;
3332
3333 if (!argc)
3334 return -EINVAL;
3335
3336 if (get_cache_mode(cache) >= CM_READ_ONLY) {
3337 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3338 cache_device_name(cache));
3339 return -EOPNOTSUPP;
3340 }
3341
3342 if (!strcasecmp(argv[0], "invalidate_cblocks"))
3343 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3344
3345 if (argc != 2)
3346 return -EINVAL;
3347
3348 return set_config_value(cache, argv[0], argv[1]);
3349}
3350
3351static int cache_iterate_devices(struct dm_target *ti,
3352 iterate_devices_callout_fn fn, void *data)
3353{
3354 int r = 0;
3355 struct cache *cache = ti->private;
3356
3357 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3358 if (!r)
3359 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3360
3361 return r;
3362}
3363
3364/*
3365 * If discard_passdown was enabled verify that the origin device
3366 * supports discards. Disable discard_passdown if not.
3367 */
3368static void disable_passdown_if_not_supported(struct cache *cache)
3369{
3370 struct block_device *origin_bdev = cache->origin_dev->bdev;
3371 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3372 const char *reason = NULL;
3373
3374 if (!cache->features.discard_passdown)
3375 return;
3376
3377 if (!bdev_max_discard_sectors(origin_bdev))
3378 reason = "discard unsupported";
3379
3380 else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3381 reason = "max discard sectors smaller than a block";
3382
3383 if (reason) {
3384 DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3385 origin_bdev, reason);
3386 cache->features.discard_passdown = false;
3387 }
3388}
3389
3390static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3391{
3392 struct block_device *origin_bdev = cache->origin_dev->bdev;
3393 struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3394
3395 if (!cache->features.discard_passdown) {
3396 /* No passdown is done so setting own virtual limits */
3397 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3398 cache->origin_sectors);
3399 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3400 return;
3401 }
3402
3403 /*
3404 * cache_iterate_devices() is stacking both origin and fast device limits
3405 * but discards aren't passed to fast device, so inherit origin's limits.
3406 */
3407 limits->max_discard_sectors = origin_limits->max_discard_sectors;
3408 limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3409 limits->discard_granularity = origin_limits->discard_granularity;
3410 limits->discard_alignment = origin_limits->discard_alignment;
3411 limits->discard_misaligned = origin_limits->discard_misaligned;
3412}
3413
3414static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3415{
3416 struct cache *cache = ti->private;
3417 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3418
3419 /*
3420 * If the system-determined stacked limits are compatible with the
3421 * cache's blocksize (io_opt is a factor) do not override them.
3422 */
3423 if (io_opt_sectors < cache->sectors_per_block ||
3424 do_div(io_opt_sectors, cache->sectors_per_block)) {
3425 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3426 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3427 }
3428
3429 disable_passdown_if_not_supported(cache);
3430 set_discard_limits(cache, limits);
3431}
3432
3433/*----------------------------------------------------------------*/
3434
3435static struct target_type cache_target = {
3436 .name = "cache",
3437 .version = {2, 2, 0},
3438 .module = THIS_MODULE,
3439 .ctr = cache_ctr,
3440 .dtr = cache_dtr,
3441 .map = cache_map,
3442 .end_io = cache_end_io,
3443 .postsuspend = cache_postsuspend,
3444 .preresume = cache_preresume,
3445 .resume = cache_resume,
3446 .status = cache_status,
3447 .message = cache_message,
3448 .iterate_devices = cache_iterate_devices,
3449 .io_hints = cache_io_hints,
3450};
3451
3452static int __init dm_cache_init(void)
3453{
3454 int r;
3455
3456 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3457 if (!migration_cache)
3458 return -ENOMEM;
3459
3460 r = dm_register_target(&cache_target);
3461 if (r) {
3462 kmem_cache_destroy(migration_cache);
3463 return r;
3464 }
3465
3466 return 0;
3467}
3468
3469static void __exit dm_cache_exit(void)
3470{
3471 dm_unregister_target(&cache_target);
3472 kmem_cache_destroy(migration_cache);
3473}
3474
3475module_init(dm_cache_init);
3476module_exit(dm_cache_exit);
3477
3478MODULE_DESCRIPTION(DM_NAME " cache target");
3479MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3480MODULE_LICENSE("GPL");