blob: 590aff275acb85199923010f48bf5f491b3fd1c2 [file] [log] [blame]
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001/*
2 * Copyright (C) 2012 Red Hat. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm.h"
Joe Thornberb29d4982016-12-15 04:57:31 -05008#include "dm-bio-prison-v2.h"
Darrick J. Wongb844fe62013-04-05 15:36:32 +01009#include "dm-bio-record.h"
Joe Thornberc6b4fcb2013-03-01 22:45:51 +000010#include "dm-cache-metadata.h"
11
12#include <linux/dm-io.h>
13#include <linux/dm-kcopyd.h>
Manuel Schölling0f30af92014-05-22 22:42:37 +020014#include <linux/jiffies.h>
Joe Thornberc6b4fcb2013-03-01 22:45:51 +000015#include <linux/init.h>
16#include <linux/mempool.h>
17#include <linux/module.h>
Joe Thornberb29d4982016-12-15 04:57:31 -050018#include <linux/rwsem.h>
Joe Thornberc6b4fcb2013-03-01 22:45:51 +000019#include <linux/slab.h>
20#include <linux/vmalloc.h>
21
22#define DM_MSG_PREFIX "cache"
23
24DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
25 "A percentage of time allocated for copying to and/or from cache");
26
27/*----------------------------------------------------------------*/
28
Joe Thornberb29d4982016-12-15 04:57:31 -050029/*
30 * Glossary:
31 *
32 * oblock: index of an origin block
33 * cblock: index of a cache block
34 * promotion: movement of a block from origin to cache
35 * demotion: movement of a block from cache to origin
36 * migration: movement of a block between the origin and cache device,
37 * either direction
38 */
39
40/*----------------------------------------------------------------*/
Joe Thornber77289d32015-05-15 13:45:30 +010041
42struct io_tracker {
43 spinlock_t lock;
44
45 /*
46 * Sectors of in-flight IO.
47 */
48 sector_t in_flight;
49
50 /*
51 * The time, in jiffies, when this device became idle (if it is
52 * indeed idle).
53 */
54 unsigned long idle_time;
55 unsigned long last_update_time;
56};
57
58static void iot_init(struct io_tracker *iot)
59{
60 spin_lock_init(&iot->lock);
61 iot->in_flight = 0ul;
62 iot->idle_time = 0ul;
63 iot->last_update_time = jiffies;
64}
65
66static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
67{
68 if (iot->in_flight)
69 return false;
70
71 return time_after(jiffies, iot->idle_time + jifs);
72}
73
74static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
75{
76 bool r;
77 unsigned long flags;
78
79 spin_lock_irqsave(&iot->lock, flags);
80 r = __iot_idle_for(iot, jifs);
81 spin_unlock_irqrestore(&iot->lock, flags);
82
83 return r;
84}
85
86static void iot_io_begin(struct io_tracker *iot, sector_t len)
87{
88 unsigned long flags;
89
90 spin_lock_irqsave(&iot->lock, flags);
91 iot->in_flight += len;
92 spin_unlock_irqrestore(&iot->lock, flags);
93}
94
95static void __iot_io_end(struct io_tracker *iot, sector_t len)
96{
Joe Thornber072792d2017-05-11 06:14:16 -040097 if (!len)
98 return;
99
Joe Thornber77289d32015-05-15 13:45:30 +0100100 iot->in_flight -= len;
101 if (!iot->in_flight)
102 iot->idle_time = jiffies;
103}
104
105static void iot_io_end(struct io_tracker *iot, sector_t len)
106{
107 unsigned long flags;
108
109 spin_lock_irqsave(&iot->lock, flags);
110 __iot_io_end(iot, len);
111 spin_unlock_irqrestore(&iot->lock, flags);
112}
113
114/*----------------------------------------------------------------*/
115
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000116/*
Joe Thornberb29d4982016-12-15 04:57:31 -0500117 * Represents a chunk of future work. 'input' allows continuations to pass
118 * values between themselves, typically error values.
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000119 */
Joe Thornberb29d4982016-12-15 04:57:31 -0500120struct continuation {
121 struct work_struct ws;
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200122 blk_status_t input;
Joe Thornberb29d4982016-12-15 04:57:31 -0500123};
124
125static inline void init_continuation(struct continuation *k,
126 void (*fn)(struct work_struct *))
127{
128 INIT_WORK(&k->ws, fn);
129 k->input = 0;
130}
131
132static inline void queue_continuation(struct workqueue_struct *wq,
133 struct continuation *k)
134{
135 queue_work(wq, &k->ws);
136}
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000137
138/*----------------------------------------------------------------*/
139
Joe Thornberc9d28d52013-10-31 13:55:48 -0400140/*
Joe Thornberb29d4982016-12-15 04:57:31 -0500141 * The batcher collects together pieces of work that need a particular
142 * operation to occur before they can proceed (typically a commit).
143 */
144struct batcher {
145 /*
146 * The operation that everyone is waiting for.
147 */
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200148 blk_status_t (*commit_op)(void *context);
Joe Thornberb29d4982016-12-15 04:57:31 -0500149 void *commit_context;
150
151 /*
152 * This is how bios should be issued once the commit op is complete
153 * (accounted_request).
154 */
155 void (*issue_op)(struct bio *bio, void *context);
156 void *issue_context;
157
158 /*
159 * Queued work gets put on here after commit.
160 */
161 struct workqueue_struct *wq;
162
163 spinlock_t lock;
164 struct list_head work_items;
165 struct bio_list bios;
166 struct work_struct commit_work;
167
168 bool commit_scheduled;
169};
170
171static void __commit(struct work_struct *_ws)
172{
173 struct batcher *b = container_of(_ws, struct batcher, commit_work);
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200174 blk_status_t r;
Joe Thornberb29d4982016-12-15 04:57:31 -0500175 unsigned long flags;
176 struct list_head work_items;
177 struct work_struct *ws, *tmp;
178 struct continuation *k;
179 struct bio *bio;
180 struct bio_list bios;
181
182 INIT_LIST_HEAD(&work_items);
183 bio_list_init(&bios);
184
185 /*
186 * We have to grab these before the commit_op to avoid a race
187 * condition.
188 */
189 spin_lock_irqsave(&b->lock, flags);
190 list_splice_init(&b->work_items, &work_items);
191 bio_list_merge(&bios, &b->bios);
192 bio_list_init(&b->bios);
193 b->commit_scheduled = false;
194 spin_unlock_irqrestore(&b->lock, flags);
195
196 r = b->commit_op(b->commit_context);
197
198 list_for_each_entry_safe(ws, tmp, &work_items, entry) {
199 k = container_of(ws, struct continuation, ws);
200 k->input = r;
201 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
202 queue_work(b->wq, ws);
203 }
204
205 while ((bio = bio_list_pop(&bios))) {
206 if (r) {
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200207 bio->bi_status = r;
Joe Thornberb29d4982016-12-15 04:57:31 -0500208 bio_endio(bio);
209 } else
210 b->issue_op(bio, b->issue_context);
211 }
212}
213
214static void batcher_init(struct batcher *b,
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +0200215 blk_status_t (*commit_op)(void *),
Joe Thornberb29d4982016-12-15 04:57:31 -0500216 void *commit_context,
217 void (*issue_op)(struct bio *bio, void *),
218 void *issue_context,
219 struct workqueue_struct *wq)
220{
221 b->commit_op = commit_op;
222 b->commit_context = commit_context;
223 b->issue_op = issue_op;
224 b->issue_context = issue_context;
225 b->wq = wq;
226
227 spin_lock_init(&b->lock);
228 INIT_LIST_HEAD(&b->work_items);
229 bio_list_init(&b->bios);
230 INIT_WORK(&b->commit_work, __commit);
231 b->commit_scheduled = false;
232}
233
234static void async_commit(struct batcher *b)
235{
236 queue_work(b->wq, &b->commit_work);
237}
238
239static void continue_after_commit(struct batcher *b, struct continuation *k)
240{
241 unsigned long flags;
242 bool commit_scheduled;
243
244 spin_lock_irqsave(&b->lock, flags);
245 commit_scheduled = b->commit_scheduled;
246 list_add_tail(&k->ws.entry, &b->work_items);
247 spin_unlock_irqrestore(&b->lock, flags);
248
249 if (commit_scheduled)
250 async_commit(b);
251}
252
253/*
254 * Bios are errored if commit failed.
255 */
256static void issue_after_commit(struct batcher *b, struct bio *bio)
257{
258 unsigned long flags;
259 bool commit_scheduled;
260
261 spin_lock_irqsave(&b->lock, flags);
262 commit_scheduled = b->commit_scheduled;
263 bio_list_add(&b->bios, bio);
264 spin_unlock_irqrestore(&b->lock, flags);
265
266 if (commit_scheduled)
267 async_commit(b);
268}
269
270/*
271 * Call this if some urgent work is waiting for the commit to complete.
272 */
273static void schedule_commit(struct batcher *b)
274{
275 bool immediate;
276 unsigned long flags;
277
278 spin_lock_irqsave(&b->lock, flags);
279 immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
280 b->commit_scheduled = true;
281 spin_unlock_irqrestore(&b->lock, flags);
282
283 if (immediate)
284 async_commit(b);
285}
286
287/*
Joe Thornberc9d28d52013-10-31 13:55:48 -0400288 * There are a couple of places where we let a bio run, but want to do some
289 * work before calling its endio function. We do this by temporarily
290 * changing the endio fn.
291 */
292struct dm_hook_info {
293 bio_end_io_t *bi_end_io;
Joe Thornberc9d28d52013-10-31 13:55:48 -0400294};
295
296static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
297 bio_end_io_t *bi_end_io, void *bi_private)
298{
299 h->bi_end_io = bio->bi_end_io;
Joe Thornberc9d28d52013-10-31 13:55:48 -0400300
301 bio->bi_end_io = bi_end_io;
302 bio->bi_private = bi_private;
303}
304
305static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
306{
307 bio->bi_end_io = h->bi_end_io;
Joe Thornberc9d28d52013-10-31 13:55:48 -0400308}
309
310/*----------------------------------------------------------------*/
311
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000312#define MIGRATION_POOL_SIZE 128
313#define COMMIT_PERIOD HZ
314#define MIGRATION_COUNT_WINDOW 10
315
316/*
Mike Snitzer05473042013-08-16 10:54:19 -0400317 * The block size of the device holding cache data must be
318 * between 32KB and 1GB.
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000319 */
320#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
Mike Snitzer05473042013-08-16 10:54:19 -0400321#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000322
Joe Thornber2ee57d52013-10-24 14:10:29 -0400323enum cache_metadata_mode {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000324 CM_WRITE, /* metadata may be changed */
325 CM_READ_ONLY, /* metadata may not be changed */
Joe Thornber028ae9f2015-04-22 16:42:35 -0400326 CM_FAIL
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000327};
328
Joe Thornber2ee57d52013-10-24 14:10:29 -0400329enum cache_io_mode {
330 /*
331 * Data is written to cached blocks only. These blocks are marked
332 * dirty. If you lose the cache device you will lose data.
333 * Potential performance increase for both reads and writes.
334 */
335 CM_IO_WRITEBACK,
336
337 /*
338 * Data is written to both cache and origin. Blocks are never
339 * dirty. Potential performance benfit for reads only.
340 */
341 CM_IO_WRITETHROUGH,
342
343 /*
344 * A degraded mode useful for various cache coherency situations
345 * (eg, rolling back snapshots). Reads and writes always go to the
346 * origin. If a write goes to a cached oblock, then the cache
347 * block is invalidated.
348 */
349 CM_IO_PASSTHROUGH
350};
351
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000352struct cache_features {
Joe Thornber2ee57d52013-10-24 14:10:29 -0400353 enum cache_metadata_mode mode;
354 enum cache_io_mode io_mode;
Joe Thornber629d0a82016-09-22 06:15:21 -0400355 unsigned metadata_version;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000356};
357
358struct cache_stats {
359 atomic_t read_hit;
360 atomic_t read_miss;
361 atomic_t write_hit;
362 atomic_t write_miss;
363 atomic_t demotion;
364 atomic_t promotion;
Joe Thornberb29d4982016-12-15 04:57:31 -0500365 atomic_t writeback;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000366 atomic_t copies_avoided;
367 atomic_t cache_cell_clash;
368 atomic_t commit_count;
369 atomic_t discard_count;
370};
371
372struct cache {
373 struct dm_target *ti;
374 struct dm_target_callbacks callbacks;
375
Mike Snitzerc9ec5d72013-08-16 10:54:21 -0400376 struct dm_cache_metadata *cmd;
377
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000378 /*
379 * Metadata is written to this device.
380 */
381 struct dm_dev *metadata_dev;
382
383 /*
384 * The slower of the two data devices. Typically a spindle.
385 */
386 struct dm_dev *origin_dev;
387
388 /*
389 * The faster of the two data devices. Typically an SSD.
390 */
391 struct dm_dev *cache_dev;
392
393 /*
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000394 * Size of the origin device in _complete_ blocks and native sectors.
395 */
396 dm_oblock_t origin_blocks;
397 sector_t origin_sectors;
398
399 /*
400 * Size of the cache device in blocks.
401 */
402 dm_cblock_t cache_size;
403
404 /*
405 * Fields for converting from sectors to blocks.
406 */
Joe Thornberca763d02017-02-09 11:46:18 -0500407 sector_t sectors_per_block;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000408 int sectors_per_block_shift;
409
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000410 spinlock_t lock;
Joe Thornber651f5fa2015-05-15 15:26:08 +0100411 struct list_head deferred_cells;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000412 struct bio_list deferred_bios;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000413 sector_t migration_threshold;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000414 wait_queue_head_t migration_wait;
Joe Thornbera59db672015-01-23 10:16:16 +0000415 atomic_t nr_allocated_migrations;
416
417 /*
418 * The number of in flight migrations that are performing
419 * background io. eg, promotion, writeback.
420 */
421 atomic_t nr_io_migrations;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000422
Joe Thornberb29d4982016-12-15 04:57:31 -0500423 struct rw_semaphore quiesce_lock;
Joe Thornber66cb1912013-10-30 17:11:58 +0000424
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000425 /*
426 * cache_size entries, dirty if set
427 */
Anssi Hannula44fa8162014-08-01 11:55:47 -0400428 atomic_t nr_dirty;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000429 unsigned long *dirty_bitset;
430
431 /*
432 * origin_blocks entries, discarded if set.
433 */
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000434 dm_dblock_t discard_nr_blocks;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000435 unsigned long *discard_bitset;
Joe Thornber08b18452014-11-06 14:38:01 +0000436 uint32_t discard_block_size; /* a power of 2 times sectors per block */
Mike Snitzerc9ec5d72013-08-16 10:54:21 -0400437
438 /*
439 * Rather than reconstructing the table line for the status we just
440 * save it and regurgitate.
441 */
442 unsigned nr_ctr_args;
443 const char **ctr_args;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000444
445 struct dm_kcopyd_client *copier;
446 struct workqueue_struct *wq;
Joe Thornberb29d4982016-12-15 04:57:31 -0500447 struct work_struct deferred_bio_worker;
Joe Thornberb29d4982016-12-15 04:57:31 -0500448 struct work_struct migration_worker;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000449 struct delayed_work waker;
Joe Thornberb29d4982016-12-15 04:57:31 -0500450 struct dm_bio_prison_v2 *prison;
Mike Snitzer80363352017-10-19 17:16:54 -0400451 struct bio_set *bs;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000452
453 mempool_t *migration_pool;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000454
455 struct dm_cache_policy *policy;
456 unsigned policy_nr_args;
457
458 bool need_tick_bio:1;
459 bool sized:1;
Joe Thornber65790ff2013-11-08 16:39:50 +0000460 bool invalidate:1;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000461 bool commit_requested:1;
462 bool loaded_mappings:1;
463 bool loaded_discards:1;
464
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000465 /*
Mike Snitzerc9ec5d72013-08-16 10:54:21 -0400466 * Cache features such as write-through.
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000467 */
Mike Snitzerc9ec5d72013-08-16 10:54:21 -0400468 struct cache_features features;
469
470 struct cache_stats stats;
Joe Thornber65790ff2013-11-08 16:39:50 +0000471
472 /*
473 * Invalidation fields.
474 */
475 spinlock_t invalidation_lock;
476 struct list_head invalidation_requests;
Joe Thornber066dbaa2015-05-15 15:18:01 +0100477
Joe Thornber701e03e42017-05-11 08:22:31 -0400478 struct io_tracker tracker;
Joe Thornberb29d4982016-12-15 04:57:31 -0500479
480 struct work_struct commit_ws;
481 struct batcher committer;
482
483 struct rw_semaphore background_work_lock;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000484};
485
486struct per_bio_data {
487 bool tick:1;
488 unsigned req_nr:2;
Joe Thornberb29d4982016-12-15 04:57:31 -0500489 struct dm_bio_prison_cell_v2 *cell;
Mike Snitzerc6eda5e2014-01-31 14:11:54 -0500490 struct dm_hook_info hook_info;
Joe Thornber066dbaa2015-05-15 15:18:01 +0100491 sector_t len;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000492};
493
494struct dm_cache_migration {
Joe Thornberb29d4982016-12-15 04:57:31 -0500495 struct continuation k;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000496 struct cache *cache;
497
Joe Thornberb29d4982016-12-15 04:57:31 -0500498 struct policy_work *op;
499 struct bio *overwrite_bio;
500 struct dm_bio_prison_cell_v2 *cell;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000501
Joe Thornberb29d4982016-12-15 04:57:31 -0500502 dm_cblock_t invalidate_cblock;
503 dm_oblock_t invalidate_oblock;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000504};
505
Joe Thornberb29d4982016-12-15 04:57:31 -0500506/*----------------------------------------------------------------*/
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000507
Mike Snitzer755b4e82017-10-19 21:01:04 -0400508static bool writethrough_mode(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000509{
Mike Snitzer755b4e82017-10-19 21:01:04 -0400510 return cache->features.io_mode == CM_IO_WRITETHROUGH;
Joe Thornberb29d4982016-12-15 04:57:31 -0500511}
512
Mike Snitzer755b4e82017-10-19 21:01:04 -0400513static bool writeback_mode(struct cache *cache)
Joe Thornberb29d4982016-12-15 04:57:31 -0500514{
Mike Snitzer755b4e82017-10-19 21:01:04 -0400515 return cache->features.io_mode == CM_IO_WRITEBACK;
Joe Thornberb29d4982016-12-15 04:57:31 -0500516}
517
Mike Snitzer755b4e82017-10-19 21:01:04 -0400518static inline bool passthrough_mode(struct cache *cache)
Joe Thornberb29d4982016-12-15 04:57:31 -0500519{
Mike Snitzer755b4e82017-10-19 21:01:04 -0400520 return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000521}
522
523/*----------------------------------------------------------------*/
524
Joe Thornberb29d4982016-12-15 04:57:31 -0500525static void wake_deferred_bio_worker(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000526{
Joe Thornberb29d4982016-12-15 04:57:31 -0500527 queue_work(cache->wq, &cache->deferred_bio_worker);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000528}
529
Joe Thornberb29d4982016-12-15 04:57:31 -0500530static void wake_migration_worker(struct cache *cache)
531{
Mike Snitzer755b4e82017-10-19 21:01:04 -0400532 if (passthrough_mode(cache))
Joe Thornberb29d4982016-12-15 04:57:31 -0500533 return;
534
535 queue_work(cache->wq, &cache->migration_worker);
536}
537
538/*----------------------------------------------------------------*/
539
540static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
541{
542 return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOWAIT);
543}
544
545static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
546{
547 dm_bio_prison_free_cell_v2(cache->prison, cell);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000548}
549
Joe Thornbera59db672015-01-23 10:16:16 +0000550static struct dm_cache_migration *alloc_migration(struct cache *cache)
551{
552 struct dm_cache_migration *mg;
553
554 mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
555 if (mg) {
556 mg->cache = cache;
557 atomic_inc(&mg->cache->nr_allocated_migrations);
558 }
559
560 return mg;
561}
562
563static void free_migration(struct dm_cache_migration *mg)
564{
Joe Thornber88bf51842015-05-27 15:39:45 +0100565 struct cache *cache = mg->cache;
Joe Thornbera59db672015-01-23 10:16:16 +0000566
Joe Thornber88bf51842015-05-27 15:39:45 +0100567 if (atomic_dec_and_test(&cache->nr_allocated_migrations))
568 wake_up(&cache->migration_wait);
569
570 mempool_free(mg, cache->migration_pool);
Joe Thornbera59db672015-01-23 10:16:16 +0000571}
572
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000573/*----------------------------------------------------------------*/
574
Joe Thornberb29d4982016-12-15 04:57:31 -0500575static inline dm_oblock_t oblock_succ(dm_oblock_t b)
576{
577 return to_oblock(from_oblock(b) + 1ull);
578}
579
580static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000581{
582 key->virtual = 0;
583 key->dev = 0;
Joe Thornber7ae34e72014-11-06 10:18:04 +0000584 key->block_begin = from_oblock(begin);
585 key->block_end = from_oblock(end);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000586}
587
588/*
Joe Thornberb29d4982016-12-15 04:57:31 -0500589 * We have two lock levels. Level 0, which is used to prevent WRITEs, and
590 * level 1 which prevents *both* READs and WRITEs.
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000591 */
Joe Thornberb29d4982016-12-15 04:57:31 -0500592#define WRITE_LOCK_LEVEL 0
593#define READ_WRITE_LOCK_LEVEL 1
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000594
Joe Thornberb29d4982016-12-15 04:57:31 -0500595static unsigned lock_level(struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000596{
Joe Thornberb29d4982016-12-15 04:57:31 -0500597 return bio_data_dir(bio) == WRITE ?
598 WRITE_LOCK_LEVEL :
599 READ_WRITE_LOCK_LEVEL;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000600}
601
Joe Thornberb29d4982016-12-15 04:57:31 -0500602/*----------------------------------------------------------------
603 * Per bio data
604 *--------------------------------------------------------------*/
605
Joe Thornberb29d4982016-12-15 04:57:31 -0500606static size_t get_per_bio_data_size(struct cache *cache)
Joe Thornber7ae34e72014-11-06 10:18:04 +0000607{
Mike Snitzerb61727d2017-10-19 17:30:20 -0400608 return sizeof(struct per_bio_data);
Joe Thornberb29d4982016-12-15 04:57:31 -0500609}
610
611static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
612{
613 struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
614 BUG_ON(!pb);
615 return pb;
616}
617
618static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
619{
620 struct per_bio_data *pb = get_per_bio_data(bio, data_size);
621
622 pb->tick = false;
623 pb->req_nr = dm_bio_get_target_bio_nr(bio);
624 pb->cell = NULL;
625 pb->len = 0;
626
627 return pb;
628}
629
630/*----------------------------------------------------------------*/
631
632static void defer_bio(struct cache *cache, struct bio *bio)
633{
634 unsigned long flags;
635
636 spin_lock_irqsave(&cache->lock, flags);
637 bio_list_add(&cache->deferred_bios, bio);
638 spin_unlock_irqrestore(&cache->lock, flags);
639
640 wake_deferred_bio_worker(cache);
641}
642
643static void defer_bios(struct cache *cache, struct bio_list *bios)
644{
645 unsigned long flags;
646
647 spin_lock_irqsave(&cache->lock, flags);
648 bio_list_merge(&cache->deferred_bios, bios);
649 bio_list_init(bios);
650 spin_unlock_irqrestore(&cache->lock, flags);
651
652 wake_deferred_bio_worker(cache);
653}
654
655/*----------------------------------------------------------------*/
656
657static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
658{
659 bool r;
660 size_t pb_size;
661 struct per_bio_data *pb;
662 struct dm_cell_key_v2 key;
Joe Thornber7ae34e72014-11-06 10:18:04 +0000663 dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
Joe Thornberb29d4982016-12-15 04:57:31 -0500664 struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
Joe Thornber7ae34e72014-11-06 10:18:04 +0000665
Joe Thornberb29d4982016-12-15 04:57:31 -0500666 cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
667 if (!cell_prealloc) {
668 defer_bio(cache, bio);
669 return false;
670 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000671
Joe Thornberb29d4982016-12-15 04:57:31 -0500672 build_key(oblock, end, &key);
673 r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
674 if (!r) {
675 /*
676 * Failed to get the lock.
677 */
678 free_prison_cell(cache, cell_prealloc);
679 return r;
680 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000681
Joe Thornberb29d4982016-12-15 04:57:31 -0500682 if (cell != cell_prealloc)
683 free_prison_cell(cache, cell_prealloc);
684
685 pb_size = get_per_bio_data_size(cache);
686 pb = get_per_bio_data(bio, pb_size);
687 pb->cell = cell;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000688
689 return r;
690}
691
Joe Thornberaeed1422013-05-10 14:37:18 +0100692/*----------------------------------------------------------------*/
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000693
694static bool is_dirty(struct cache *cache, dm_cblock_t b)
695{
696 return test_bit(from_cblock(b), cache->dirty_bitset);
697}
698
Joe Thornberb29d4982016-12-15 04:57:31 -0500699static void set_dirty(struct cache *cache, dm_cblock_t cblock)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000700{
701 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
Anssi Hannula44fa8162014-08-01 11:55:47 -0400702 atomic_inc(&cache->nr_dirty);
Joe Thornberb29d4982016-12-15 04:57:31 -0500703 policy_set_dirty(cache->policy, cblock);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000704 }
705}
706
Joe Thornberb29d4982016-12-15 04:57:31 -0500707/*
708 * These two are called when setting after migrations to force the policy
709 * and dirty bitset to be in sync.
710 */
711static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
712{
713 if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
714 atomic_inc(&cache->nr_dirty);
715 policy_set_dirty(cache->policy, cblock);
716}
717
718static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000719{
720 if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
Anssi Hannula44fa8162014-08-01 11:55:47 -0400721 if (atomic_dec_return(&cache->nr_dirty) == 0)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000722 dm_table_event(cache->ti->table);
723 }
Joe Thornberb29d4982016-12-15 04:57:31 -0500724
725 policy_clear_dirty(cache->policy, cblock);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000726}
727
728/*----------------------------------------------------------------*/
Joe Thornberaeed1422013-05-10 14:37:18 +0100729
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000730static bool block_size_is_power_of_two(struct cache *cache)
731{
732 return cache->sectors_per_block_shift >= 0;
733}
734
Mikulas Patocka43aeaa22013-07-10 23:41:17 +0100735/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
736#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
737__always_inline
738#endif
Joe Thornber414dd672013-03-20 17:21:25 +0000739static dm_block_t block_div(dm_block_t b, uint32_t n)
740{
741 do_div(b, n);
742
743 return b;
744}
745
Joe Thornber7ae34e72014-11-06 10:18:04 +0000746static dm_block_t oblocks_per_dblock(struct cache *cache)
747{
748 dm_block_t oblocks = cache->discard_block_size;
749
750 if (block_size_is_power_of_two(cache))
751 oblocks >>= cache->sectors_per_block_shift;
752 else
753 oblocks = block_div(oblocks, cache->sectors_per_block);
754
755 return oblocks;
756}
757
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000758static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
759{
Joe Thornber7ae34e72014-11-06 10:18:04 +0000760 return to_dblock(block_div(from_oblock(oblock),
761 oblocks_per_dblock(cache)));
762}
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000763
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000764static void set_discard(struct cache *cache, dm_dblock_t b)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000765{
766 unsigned long flags;
767
Joe Thornber7ae34e72014-11-06 10:18:04 +0000768 BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000769 atomic_inc(&cache->stats.discard_count);
770
771 spin_lock_irqsave(&cache->lock, flags);
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000772 set_bit(from_dblock(b), cache->discard_bitset);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000773 spin_unlock_irqrestore(&cache->lock, flags);
774}
775
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000776static void clear_discard(struct cache *cache, dm_dblock_t b)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000777{
778 unsigned long flags;
779
780 spin_lock_irqsave(&cache->lock, flags);
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000781 clear_bit(from_dblock(b), cache->discard_bitset);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000782 spin_unlock_irqrestore(&cache->lock, flags);
783}
784
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000785static bool is_discarded(struct cache *cache, dm_dblock_t b)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000786{
787 int r;
788 unsigned long flags;
789
790 spin_lock_irqsave(&cache->lock, flags);
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000791 r = test_bit(from_dblock(b), cache->discard_bitset);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000792 spin_unlock_irqrestore(&cache->lock, flags);
793
794 return r;
795}
796
797static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
798{
799 int r;
800 unsigned long flags;
801
802 spin_lock_irqsave(&cache->lock, flags);
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000803 r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
804 cache->discard_bitset);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000805 spin_unlock_irqrestore(&cache->lock, flags);
806
807 return r;
808}
809
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000810/*----------------------------------------------------------------
811 * Remapping
812 *--------------------------------------------------------------*/
813static void remap_to_origin(struct cache *cache, struct bio *bio)
814{
Christoph Hellwig74d46992017-08-23 19:10:32 +0200815 bio_set_dev(bio, cache->origin_dev->bdev);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000816}
817
818static void remap_to_cache(struct cache *cache, struct bio *bio,
819 dm_cblock_t cblock)
820{
Kent Overstreet4f024f32013-10-11 15:44:27 -0700821 sector_t bi_sector = bio->bi_iter.bi_sector;
Heinz Mauelshagene0d849f2014-02-27 22:46:48 +0100822 sector_t block = from_cblock(cblock);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000823
Christoph Hellwig74d46992017-08-23 19:10:32 +0200824 bio_set_dev(bio, cache->cache_dev->bdev);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000825 if (!block_size_is_power_of_two(cache))
Kent Overstreet4f024f32013-10-11 15:44:27 -0700826 bio->bi_iter.bi_sector =
Heinz Mauelshagene0d849f2014-02-27 22:46:48 +0100827 (block * cache->sectors_per_block) +
Kent Overstreet4f024f32013-10-11 15:44:27 -0700828 sector_div(bi_sector, cache->sectors_per_block);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000829 else
Kent Overstreet4f024f32013-10-11 15:44:27 -0700830 bio->bi_iter.bi_sector =
Heinz Mauelshagene0d849f2014-02-27 22:46:48 +0100831 (block << cache->sectors_per_block_shift) |
Kent Overstreet4f024f32013-10-11 15:44:27 -0700832 (bi_sector & (cache->sectors_per_block - 1));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000833}
834
835static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
836{
837 unsigned long flags;
Mike Snitzer19b00922013-04-05 15:36:34 +0100838 size_t pb_data_size = get_per_bio_data_size(cache);
839 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000840
841 spin_lock_irqsave(&cache->lock, flags);
Christoph Hellwigf73f44e2017-01-27 08:30:47 -0700842 if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
Mike Christiee6047142016-06-05 14:32:04 -0500843 bio_op(bio) != REQ_OP_DISCARD) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000844 pb->tick = true;
845 cache->need_tick_bio = false;
846 }
847 spin_unlock_irqrestore(&cache->lock, flags);
848}
849
Mike Snitzer80363352017-10-19 17:16:54 -0400850static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
851 dm_oblock_t oblock, bool bio_has_pbd)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000852{
Mike Snitzer80363352017-10-19 17:16:54 -0400853 if (bio_has_pbd)
854 check_if_tick_bio_needed(cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000855 remap_to_origin(cache, bio);
856 if (bio_data_dir(bio) == WRITE)
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000857 clear_discard(cache, oblock_to_dblock(cache, oblock));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000858}
859
Mike Snitzer80363352017-10-19 17:16:54 -0400860static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
861 dm_oblock_t oblock)
862{
863 // FIXME: check_if_tick_bio_needed() is called way too much through this interface
864 __remap_to_origin_clear_discard(cache, bio, oblock, true);
865}
866
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000867static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
868 dm_oblock_t oblock, dm_cblock_t cblock)
869{
Joe Thornberf8e5f012013-10-21 12:51:45 +0100870 check_if_tick_bio_needed(cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000871 remap_to_cache(cache, bio, cblock);
872 if (bio_data_dir(bio) == WRITE) {
Joe Thornberb29d4982016-12-15 04:57:31 -0500873 set_dirty(cache, cblock);
Joe Thornber1bad9bc2014-11-07 14:47:07 +0000874 clear_discard(cache, oblock_to_dblock(cache, oblock));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000875 }
876}
877
878static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
879{
Kent Overstreet4f024f32013-10-11 15:44:27 -0700880 sector_t block_nr = bio->bi_iter.bi_sector;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000881
882 if (!block_size_is_power_of_two(cache))
883 (void) sector_div(block_nr, cache->sectors_per_block);
884 else
885 block_nr >>= cache->sectors_per_block_shift;
886
887 return to_oblock(block_nr);
888}
889
Joe Thornber066dbaa2015-05-15 15:18:01 +0100890static bool accountable_bio(struct cache *cache, struct bio *bio)
891{
Joe Thornber701e03e42017-05-11 08:22:31 -0400892 return bio_op(bio) != REQ_OP_DISCARD;
Joe Thornber066dbaa2015-05-15 15:18:01 +0100893}
894
895static void accounted_begin(struct cache *cache, struct bio *bio)
896{
897 size_t pb_data_size = get_per_bio_data_size(cache);
898 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
899
900 if (accountable_bio(cache, bio)) {
901 pb->len = bio_sectors(bio);
Joe Thornber701e03e42017-05-11 08:22:31 -0400902 iot_io_begin(&cache->tracker, pb->len);
Joe Thornber066dbaa2015-05-15 15:18:01 +0100903 }
904}
905
906static void accounted_complete(struct cache *cache, struct bio *bio)
907{
908 size_t pb_data_size = get_per_bio_data_size(cache);
909 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
910
Joe Thornber701e03e42017-05-11 08:22:31 -0400911 iot_io_end(&cache->tracker, pb->len);
Joe Thornber066dbaa2015-05-15 15:18:01 +0100912}
913
914static void accounted_request(struct cache *cache, struct bio *bio)
915{
916 accounted_begin(cache, bio);
917 generic_make_request(bio);
918}
919
Joe Thornberb29d4982016-12-15 04:57:31 -0500920static void issue_op(struct bio *bio, void *context)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000921{
Joe Thornberb29d4982016-12-15 04:57:31 -0500922 struct cache *cache = context;
923 accounted_request(cache, bio);
Joe Thornber8c081b52014-05-13 16:18:38 +0100924}
925
Joe Thornbere2e74d62013-03-20 17:21:27 +0000926/*
927 * When running in writethrough mode we need to send writes to clean blocks
Mike Snitzer80363352017-10-19 17:16:54 -0400928 * to both the cache and origin devices. Clone the bio and send them in parallel.
Joe Thornbere2e74d62013-03-20 17:21:27 +0000929 */
Mike Snitzer80363352017-10-19 17:16:54 -0400930static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
931 dm_oblock_t oblock, dm_cblock_t cblock)
Joe Thornbere2e74d62013-03-20 17:21:27 +0000932{
Mike Snitzer80363352017-10-19 17:16:54 -0400933 struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, cache->bs);
Joe Thornbere2e74d62013-03-20 17:21:27 +0000934
Mike Snitzer80363352017-10-19 17:16:54 -0400935 BUG_ON(!origin_bio);
Joe Thornbere2e74d62013-03-20 17:21:27 +0000936
Mike Snitzer80363352017-10-19 17:16:54 -0400937 bio_chain(origin_bio, bio);
938 /*
939 * Passing false to __remap_to_origin_clear_discard() skips
940 * all code that might use per_bio_data (since clone doesn't have it)
941 */
942 __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
943 submit_bio(origin_bio);
944
945 remap_to_cache(cache, bio, cblock);
Joe Thornbere2e74d62013-03-20 17:21:27 +0000946}
947
Joe Thornberc6b4fcb2013-03-01 22:45:51 +0000948/*----------------------------------------------------------------
Joe Thornber028ae9f2015-04-22 16:42:35 -0400949 * Failure modes
950 *--------------------------------------------------------------*/
951static enum cache_metadata_mode get_cache_mode(struct cache *cache)
952{
953 return cache->features.mode;
954}
955
Mike Snitzerb61d9502015-04-22 17:25:56 -0400956static const char *cache_device_name(struct cache *cache)
957{
958 return dm_device_name(dm_table_get_md(cache->ti->table));
959}
960
Joe Thornber028ae9f2015-04-22 16:42:35 -0400961static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
962{
963 const char *descs[] = {
964 "write",
965 "read-only",
966 "fail"
967 };
968
969 dm_table_event(cache->ti->table);
Mike Snitzerb61d9502015-04-22 17:25:56 -0400970 DMINFO("%s: switching cache to %s mode",
971 cache_device_name(cache), descs[(int)mode]);
Joe Thornber028ae9f2015-04-22 16:42:35 -0400972}
973
974static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
975{
Joe Thornberd14fcf32016-03-10 16:20:58 +0000976 bool needs_check;
Joe Thornber028ae9f2015-04-22 16:42:35 -0400977 enum cache_metadata_mode old_mode = get_cache_mode(cache);
978
Joe Thornberd14fcf32016-03-10 16:20:58 +0000979 if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
Mike Snitzer23cab262016-10-04 12:04:08 -0400980 DMERR("%s: unable to read needs_check flag, setting failure mode.",
981 cache_device_name(cache));
Joe Thornberd14fcf32016-03-10 16:20:58 +0000982 new_mode = CM_FAIL;
983 }
984
Joe Thornber028ae9f2015-04-22 16:42:35 -0400985 if (new_mode == CM_WRITE && needs_check) {
Mike Snitzerb61d9502015-04-22 17:25:56 -0400986 DMERR("%s: unable to switch cache to write mode until repaired.",
987 cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -0400988 if (old_mode != new_mode)
989 new_mode = old_mode;
990 else
991 new_mode = CM_READ_ONLY;
992 }
993
994 /* Never move out of fail mode */
995 if (old_mode == CM_FAIL)
996 new_mode = CM_FAIL;
997
998 switch (new_mode) {
999 case CM_FAIL:
1000 case CM_READ_ONLY:
1001 dm_cache_metadata_set_read_only(cache->cmd);
1002 break;
1003
1004 case CM_WRITE:
1005 dm_cache_metadata_set_read_write(cache->cmd);
1006 break;
1007 }
1008
1009 cache->features.mode = new_mode;
1010
1011 if (new_mode != old_mode)
1012 notify_mode_switch(cache, new_mode);
1013}
1014
1015static void abort_transaction(struct cache *cache)
1016{
Mike Snitzerb61d9502015-04-22 17:25:56 -04001017 const char *dev_name = cache_device_name(cache);
1018
Joe Thornber028ae9f2015-04-22 16:42:35 -04001019 if (get_cache_mode(cache) >= CM_READ_ONLY)
1020 return;
1021
Mike Snitzerb61d9502015-04-22 17:25:56 -04001022 DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
Joe Thornber028ae9f2015-04-22 16:42:35 -04001023 if (dm_cache_metadata_abort(cache->cmd)) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04001024 DMERR("%s: failed to abort metadata transaction", dev_name);
Joe Thornber028ae9f2015-04-22 16:42:35 -04001025 set_cache_mode(cache, CM_FAIL);
1026 }
Mike Snitzereef9d8e2022-11-30 14:02:47 -05001027
1028 if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1029 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1030 set_cache_mode(cache, CM_FAIL);
1031 }
Joe Thornber028ae9f2015-04-22 16:42:35 -04001032}
1033
1034static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1035{
Mike Snitzerb61d9502015-04-22 17:25:56 -04001036 DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1037 cache_device_name(cache), op, r);
Joe Thornber028ae9f2015-04-22 16:42:35 -04001038 abort_transaction(cache);
1039 set_cache_mode(cache, CM_READ_ONLY);
1040}
1041
Joe Thornberb29d4982016-12-15 04:57:31 -05001042/*----------------------------------------------------------------*/
1043
1044static void load_stats(struct cache *cache)
1045{
1046 struct dm_cache_statistics stats;
1047
1048 dm_cache_metadata_get_stats(cache->cmd, &stats);
1049 atomic_set(&cache->stats.read_hit, stats.read_hits);
1050 atomic_set(&cache->stats.read_miss, stats.read_misses);
1051 atomic_set(&cache->stats.write_hit, stats.write_hits);
1052 atomic_set(&cache->stats.write_miss, stats.write_misses);
1053}
1054
1055static void save_stats(struct cache *cache)
1056{
1057 struct dm_cache_statistics stats;
1058
1059 if (get_cache_mode(cache) >= CM_READ_ONLY)
1060 return;
1061
1062 stats.read_hits = atomic_read(&cache->stats.read_hit);
1063 stats.read_misses = atomic_read(&cache->stats.read_miss);
1064 stats.write_hits = atomic_read(&cache->stats.write_hit);
1065 stats.write_misses = atomic_read(&cache->stats.write_miss);
1066
1067 dm_cache_metadata_set_stats(cache->cmd, &stats);
1068}
1069
1070static void update_stats(struct cache_stats *stats, enum policy_operation op)
1071{
1072 switch (op) {
1073 case POLICY_PROMOTE:
1074 atomic_inc(&stats->promotion);
1075 break;
1076
1077 case POLICY_DEMOTE:
1078 atomic_inc(&stats->demotion);
1079 break;
1080
1081 case POLICY_WRITEBACK:
1082 atomic_inc(&stats->writeback);
1083 break;
1084 }
1085}
1086
Joe Thornber028ae9f2015-04-22 16:42:35 -04001087/*----------------------------------------------------------------
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001088 * Migration processing
1089 *
1090 * Migration covers moving data from the origin device to the cache, or
1091 * vice versa.
1092 *--------------------------------------------------------------*/
Joe Thornberb29d4982016-12-15 04:57:31 -05001093
Joe Thornbera59db672015-01-23 10:16:16 +00001094static void inc_io_migrations(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001095{
Joe Thornbera59db672015-01-23 10:16:16 +00001096 atomic_inc(&cache->nr_io_migrations);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001097}
1098
Joe Thornbera59db672015-01-23 10:16:16 +00001099static void dec_io_migrations(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001100{
Joe Thornbera59db672015-01-23 10:16:16 +00001101 atomic_dec(&cache->nr_io_migrations);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001102}
1103
Joe Thornber651f5fa2015-05-15 15:26:08 +01001104static bool discard_or_flush(struct bio *bio)
1105{
Christoph Hellwigf73f44e2017-01-27 08:30:47 -07001106 return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
Joe Thornber651f5fa2015-05-15 15:26:08 +01001107}
1108
Joe Thornber7ae34e72014-11-06 10:18:04 +00001109static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1110 dm_dblock_t *b, dm_dblock_t *e)
1111{
1112 sector_t sb = bio->bi_iter.bi_sector;
1113 sector_t se = bio_end_sector(bio);
1114
1115 *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1116
1117 if (se - sb < cache->discard_block_size)
1118 *e = *b;
1119 else
1120 *e = to_dblock(block_div(se, cache->discard_block_size));
1121}
1122
Joe Thornberb29d4982016-12-15 04:57:31 -05001123/*----------------------------------------------------------------*/
1124
1125static void prevent_background_work(struct cache *cache)
Joe Thornber7ae34e72014-11-06 10:18:04 +00001126{
Joe Thornberb29d4982016-12-15 04:57:31 -05001127 lockdep_off();
1128 down_write(&cache->background_work_lock);
1129 lockdep_on();
Joe Thornber7ae34e72014-11-06 10:18:04 +00001130}
1131
Joe Thornberb29d4982016-12-15 04:57:31 -05001132static void allow_background_work(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001133{
Joe Thornberb29d4982016-12-15 04:57:31 -05001134 lockdep_off();
1135 up_write(&cache->background_work_lock);
1136 lockdep_on();
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001137}
1138
Joe Thornberb29d4982016-12-15 04:57:31 -05001139static bool background_work_begin(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001140{
Joe Thornberb29d4982016-12-15 04:57:31 -05001141 bool r;
1142
1143 lockdep_off();
1144 r = down_read_trylock(&cache->background_work_lock);
1145 lockdep_on();
1146
1147 return r;
1148}
1149
1150static void background_work_end(struct cache *cache)
1151{
1152 lockdep_off();
1153 up_read(&cache->background_work_lock);
1154 lockdep_on();
1155}
1156
1157/*----------------------------------------------------------------*/
1158
Joe Thornberec544ec2017-11-10 07:53:31 -05001159static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1160{
1161 return (bio_data_dir(bio) == WRITE) &&
1162 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1163}
1164
1165static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1166{
Mike Snitzer755b4e82017-10-19 21:01:04 -04001167 return writeback_mode(cache) &&
Joe Thornberec544ec2017-11-10 07:53:31 -05001168 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1169}
1170
Joe Thornberb29d4982016-12-15 04:57:31 -05001171static void quiesce(struct dm_cache_migration *mg,
1172 void (*continuation)(struct work_struct *))
1173{
1174 init_continuation(&mg->k, continuation);
1175 dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1176}
1177
1178static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1179{
1180 struct continuation *k = container_of(ws, struct continuation, ws);
1181 return container_of(k, struct dm_cache_migration, k);
1182}
1183
1184static void copy_complete(int read_err, unsigned long write_err, void *context)
1185{
1186 struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1187
1188 if (read_err || write_err)
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001189 mg->k.input = BLK_STS_IOERR;
Joe Thornberb29d4982016-12-15 04:57:31 -05001190
1191 queue_continuation(mg->cache->wq, &mg->k);
1192}
1193
1194static int copy(struct dm_cache_migration *mg, bool promote)
1195{
1196 int r;
1197 struct dm_io_region o_region, c_region;
1198 struct cache *cache = mg->cache;
1199
1200 o_region.bdev = cache->origin_dev->bdev;
1201 o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1202 o_region.count = cache->sectors_per_block;
1203
1204 c_region.bdev = cache->cache_dev->bdev;
1205 c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1206 c_region.count = cache->sectors_per_block;
1207
1208 if (promote)
1209 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001210 else
Joe Thornberb29d4982016-12-15 04:57:31 -05001211 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1212
1213 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001214}
1215
Joe Thornberb29d4982016-12-15 04:57:31 -05001216static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001217{
Joe Thornberb29d4982016-12-15 04:57:31 -05001218 size_t pb_data_size = get_per_bio_data_size(cache);
1219 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001220
Joe Thornberb29d4982016-12-15 04:57:31 -05001221 if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1222 free_prison_cell(cache, pb->cell);
1223 pb->cell = NULL;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001224}
1225
Joe Thornberb29d4982016-12-15 04:57:31 -05001226static void overwrite_endio(struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001227{
Joe Thornberb29d4982016-12-15 04:57:31 -05001228 struct dm_cache_migration *mg = bio->bi_private;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001229 struct cache *cache = mg->cache;
Joe Thornberb29d4982016-12-15 04:57:31 -05001230 size_t pb_data_size = get_per_bio_data_size(cache);
1231 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001232
Joe Thornberb29d4982016-12-15 04:57:31 -05001233 dm_unhook_bio(&pb->hook_info, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001234
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001235 if (bio->bi_status)
1236 mg->k.input = bio->bi_status;
Joe Thornberb29d4982016-12-15 04:57:31 -05001237
1238 queue_continuation(mg->cache->wq, &mg->k);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001239}
1240
Joe Thornberb29d4982016-12-15 04:57:31 -05001241static void overwrite(struct dm_cache_migration *mg,
1242 void (*continuation)(struct work_struct *))
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001243{
Joe Thornberb29d4982016-12-15 04:57:31 -05001244 struct bio *bio = mg->overwrite_bio;
1245 size_t pb_data_size = get_per_bio_data_size(mg->cache);
1246 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001247
Joe Thornberb29d4982016-12-15 04:57:31 -05001248 dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001249
Joe Thornberb29d4982016-12-15 04:57:31 -05001250 /*
1251 * The overwrite bio is part of the copy operation, as such it does
1252 * not set/clear discard or dirty flags.
1253 */
1254 if (mg->op->op == POLICY_PROMOTE)
1255 remap_to_cache(mg->cache, bio, mg->op->cblock);
1256 else
1257 remap_to_origin(mg->cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001258
Joe Thornberb29d4982016-12-15 04:57:31 -05001259 init_continuation(&mg->k, continuation);
1260 accounted_request(mg->cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001261}
1262
Joe Thornber2ee57d52013-10-24 14:10:29 -04001263/*
Joe Thornberb29d4982016-12-15 04:57:31 -05001264 * Migration steps:
1265 *
1266 * 1) exclusive lock preventing WRITEs
1267 * 2) quiesce
1268 * 3) copy or issue overwrite bio
1269 * 4) upgrade to exclusive lock preventing READs and WRITEs
1270 * 5) quiesce
1271 * 6) update metadata and commit
1272 * 7) unlock
Joe Thornber2ee57d52013-10-24 14:10:29 -04001273 */
Joe Thornberb29d4982016-12-15 04:57:31 -05001274static void mg_complete(struct dm_cache_migration *mg, bool success)
Joe Thornber2ee57d52013-10-24 14:10:29 -04001275{
Joe Thornberb29d4982016-12-15 04:57:31 -05001276 struct bio_list bios;
1277 struct cache *cache = mg->cache;
1278 struct policy_work *op = mg->op;
1279 dm_cblock_t cblock = op->cblock;
Joe Thornber2ee57d52013-10-24 14:10:29 -04001280
Joe Thornberb29d4982016-12-15 04:57:31 -05001281 if (success)
1282 update_stats(&cache->stats, op->op);
Joe Thornber2ee57d52013-10-24 14:10:29 -04001283
Joe Thornberb29d4982016-12-15 04:57:31 -05001284 switch (op->op) {
1285 case POLICY_PROMOTE:
1286 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1287 policy_complete_background_work(cache->policy, op, success);
1288
1289 if (mg->overwrite_bio) {
1290 if (success)
1291 force_set_dirty(cache, cblock);
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001292 else if (mg->k.input)
1293 mg->overwrite_bio->bi_status = mg->k.input;
Joe Thornberb29d4982016-12-15 04:57:31 -05001294 else
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001295 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
Joe Thornberb29d4982016-12-15 04:57:31 -05001296 bio_endio(mg->overwrite_bio);
1297 } else {
1298 if (success)
1299 force_clear_dirty(cache, cblock);
1300 dec_io_migrations(cache);
1301 }
1302 break;
1303
1304 case POLICY_DEMOTE:
1305 /*
1306 * We clear dirty here to update the nr_dirty counter.
1307 */
1308 if (success)
1309 force_clear_dirty(cache, cblock);
1310 policy_complete_background_work(cache->policy, op, success);
1311 dec_io_migrations(cache);
1312 break;
1313
1314 case POLICY_WRITEBACK:
1315 if (success)
1316 force_clear_dirty(cache, cblock);
1317 policy_complete_background_work(cache->policy, op, success);
1318 dec_io_migrations(cache);
1319 break;
1320 }
1321
1322 bio_list_init(&bios);
1323 if (mg->cell) {
1324 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1325 free_prison_cell(cache, mg->cell);
1326 }
1327
1328 free_migration(mg);
1329 defer_bios(cache, &bios);
1330 wake_migration_worker(cache);
1331
1332 background_work_end(cache);
Joe Thornber2ee57d52013-10-24 14:10:29 -04001333}
1334
Joe Thornberb29d4982016-12-15 04:57:31 -05001335static void mg_success(struct work_struct *ws)
Joe Thornber7ae34e72014-11-06 10:18:04 +00001336{
Joe Thornberb29d4982016-12-15 04:57:31 -05001337 struct dm_cache_migration *mg = ws_to_mg(ws);
1338 mg_complete(mg, mg->k.input == 0);
1339}
Joe Thornber7ae34e72014-11-06 10:18:04 +00001340
Joe Thornberb29d4982016-12-15 04:57:31 -05001341static void mg_update_metadata(struct work_struct *ws)
1342{
1343 int r;
1344 struct dm_cache_migration *mg = ws_to_mg(ws);
1345 struct cache *cache = mg->cache;
1346 struct policy_work *op = mg->op;
1347
1348 switch (op->op) {
1349 case POLICY_PROMOTE:
1350 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1351 if (r) {
1352 DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1353 cache_device_name(cache));
1354 metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1355
1356 mg_complete(mg, false);
1357 return;
1358 }
1359 mg_complete(mg, true);
1360 break;
1361
1362 case POLICY_DEMOTE:
1363 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1364 if (r) {
1365 DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1366 cache_device_name(cache));
1367 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1368
1369 mg_complete(mg, false);
1370 return;
1371 }
1372
1373 /*
1374 * It would be nice if we only had to commit when a REQ_FLUSH
1375 * comes through. But there's one scenario that we have to
1376 * look out for:
1377 *
1378 * - vblock x in a cache block
1379 * - domotion occurs
1380 * - cache block gets reallocated and over written
1381 * - crash
1382 *
1383 * When we recover, because there was no commit the cache will
1384 * rollback to having the data for vblock x in the cache block.
1385 * But the cache block has since been overwritten, so it'll end
1386 * up pointing to data that was never in 'x' during the history
1387 * of the device.
1388 *
1389 * To avoid this issue we require a commit as part of the
1390 * demotion operation.
1391 */
1392 init_continuation(&mg->k, mg_success);
1393 continue_after_commit(&cache->committer, &mg->k);
1394 schedule_commit(&cache->committer);
1395 break;
1396
1397 case POLICY_WRITEBACK:
1398 mg_complete(mg, true);
1399 break;
1400 }
1401}
1402
1403static void mg_update_metadata_after_copy(struct work_struct *ws)
1404{
1405 struct dm_cache_migration *mg = ws_to_mg(ws);
1406
1407 /*
1408 * Did the copy succeed?
1409 */
1410 if (mg->k.input)
1411 mg_complete(mg, false);
1412 else
1413 mg_update_metadata(ws);
1414}
1415
1416static void mg_upgrade_lock(struct work_struct *ws)
1417{
1418 int r;
1419 struct dm_cache_migration *mg = ws_to_mg(ws);
1420
1421 /*
1422 * Did the copy succeed?
1423 */
1424 if (mg->k.input)
1425 mg_complete(mg, false);
1426
1427 else {
1428 /*
1429 * Now we want the lock to prevent both reads and writes.
1430 */
1431 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1432 READ_WRITE_LOCK_LEVEL);
1433 if (r < 0)
1434 mg_complete(mg, false);
1435
1436 else if (r)
1437 quiesce(mg, mg_update_metadata);
1438
1439 else
1440 mg_update_metadata(ws);
1441 }
1442}
1443
Joe Thornberec544ec2017-11-10 07:53:31 -05001444static void mg_full_copy(struct work_struct *ws)
1445{
1446 struct dm_cache_migration *mg = ws_to_mg(ws);
1447 struct cache *cache = mg->cache;
1448 struct policy_work *op = mg->op;
1449 bool is_policy_promote = (op->op == POLICY_PROMOTE);
1450
1451 if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1452 is_discarded_oblock(cache, op->oblock)) {
1453 mg_upgrade_lock(ws);
1454 return;
1455 }
1456
1457 init_continuation(&mg->k, mg_upgrade_lock);
1458
1459 if (copy(mg, is_policy_promote)) {
1460 DMERR_LIMIT("%s: migration copy failed", cache_device_name(cache));
1461 mg->k.input = BLK_STS_IOERR;
1462 mg_complete(mg, false);
1463 }
1464}
1465
Joe Thornberb29d4982016-12-15 04:57:31 -05001466static void mg_copy(struct work_struct *ws)
1467{
Joe Thornberb29d4982016-12-15 04:57:31 -05001468 struct dm_cache_migration *mg = ws_to_mg(ws);
1469
1470 if (mg->overwrite_bio) {
1471 /*
Joe Thornberec544ec2017-11-10 07:53:31 -05001472 * No exclusive lock was held when we last checked if the bio
1473 * was optimisable. So we have to check again in case things
1474 * have changed (eg, the block may no longer be discarded).
1475 */
1476 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1477 /*
1478 * Fallback to a real full copy after doing some tidying up.
1479 */
1480 bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1481 BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1482 mg->overwrite_bio = NULL;
1483 inc_io_migrations(mg->cache);
1484 mg_full_copy(ws);
1485 return;
1486 }
1487
1488 /*
Joe Thornberb29d4982016-12-15 04:57:31 -05001489 * It's safe to do this here, even though it's new data
1490 * because all IO has been locked out of the block.
1491 *
1492 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1493 * so _not_ using mg_upgrade_lock() as continutation.
1494 */
1495 overwrite(mg, mg_update_metadata_after_copy);
1496
Joe Thornberec544ec2017-11-10 07:53:31 -05001497 } else
1498 mg_full_copy(ws);
Joe Thornberb29d4982016-12-15 04:57:31 -05001499}
1500
1501static int mg_lock_writes(struct dm_cache_migration *mg)
1502{
1503 int r;
1504 struct dm_cell_key_v2 key;
1505 struct cache *cache = mg->cache;
1506 struct dm_bio_prison_cell_v2 *prealloc;
1507
1508 prealloc = alloc_prison_cell(cache);
1509 if (!prealloc) {
1510 DMERR_LIMIT("%s: alloc_prison_cell failed", cache_device_name(cache));
1511 mg_complete(mg, false);
1512 return -ENOMEM;
1513 }
1514
1515 /*
1516 * Prevent writes to the block, but allow reads to continue.
1517 * Unless we're using an overwrite bio, in which case we lock
1518 * everything.
1519 */
1520 build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1521 r = dm_cell_lock_v2(cache->prison, &key,
1522 mg->overwrite_bio ? READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1523 prealloc, &mg->cell);
1524 if (r < 0) {
1525 free_prison_cell(cache, prealloc);
1526 mg_complete(mg, false);
1527 return r;
1528 }
1529
1530 if (mg->cell != prealloc)
1531 free_prison_cell(cache, prealloc);
1532
1533 if (r == 0)
1534 mg_copy(&mg->k.ws);
1535 else
1536 quiesce(mg, mg_copy);
1537
1538 return 0;
1539}
1540
1541static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1542{
1543 struct dm_cache_migration *mg;
1544
1545 if (!background_work_begin(cache)) {
1546 policy_complete_background_work(cache->policy, op, false);
1547 return -EPERM;
1548 }
1549
1550 mg = alloc_migration(cache);
1551 if (!mg) {
1552 policy_complete_background_work(cache->policy, op, false);
1553 background_work_end(cache);
1554 return -ENOMEM;
1555 }
1556
1557 memset(mg, 0, sizeof(*mg));
1558
Joe Thornber7ae34e72014-11-06 10:18:04 +00001559 mg->cache = cache;
Joe Thornberb29d4982016-12-15 04:57:31 -05001560 mg->op = op;
1561 mg->overwrite_bio = bio;
Joe Thornber7ae34e72014-11-06 10:18:04 +00001562
Joe Thornberb29d4982016-12-15 04:57:31 -05001563 if (!bio)
1564 inc_io_migrations(cache);
1565
1566 return mg_lock_writes(mg);
1567}
1568
1569/*----------------------------------------------------------------
1570 * invalidation processing
1571 *--------------------------------------------------------------*/
1572
1573static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1574{
1575 struct bio_list bios;
1576 struct cache *cache = mg->cache;
1577
1578 bio_list_init(&bios);
1579 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1580 free_prison_cell(cache, mg->cell);
1581
1582 if (!success && mg->overwrite_bio)
1583 bio_io_error(mg->overwrite_bio);
1584
1585 free_migration(mg);
1586 defer_bios(cache, &bios);
1587
1588 background_work_end(cache);
1589}
1590
1591static void invalidate_completed(struct work_struct *ws)
1592{
1593 struct dm_cache_migration *mg = ws_to_mg(ws);
1594 invalidate_complete(mg, !mg->k.input);
1595}
1596
1597static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1598{
1599 int r = policy_invalidate_mapping(cache->policy, cblock);
1600 if (!r) {
1601 r = dm_cache_remove_mapping(cache->cmd, cblock);
1602 if (r) {
1603 DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1604 cache_device_name(cache));
1605 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1606 }
1607
1608 } else if (r == -ENODATA) {
1609 /*
1610 * Harmless, already unmapped.
1611 */
1612 r = 0;
1613
1614 } else
1615 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1616
1617 return r;
1618}
1619
1620static void invalidate_remove(struct work_struct *ws)
1621{
1622 int r;
1623 struct dm_cache_migration *mg = ws_to_mg(ws);
1624 struct cache *cache = mg->cache;
1625
1626 r = invalidate_cblock(cache, mg->invalidate_cblock);
1627 if (r) {
1628 invalidate_complete(mg, false);
1629 return;
1630 }
1631
1632 init_continuation(&mg->k, invalidate_completed);
1633 continue_after_commit(&cache->committer, &mg->k);
1634 remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1635 mg->overwrite_bio = NULL;
1636 schedule_commit(&cache->committer);
1637}
1638
1639static int invalidate_lock(struct dm_cache_migration *mg)
1640{
1641 int r;
1642 struct dm_cell_key_v2 key;
1643 struct cache *cache = mg->cache;
1644 struct dm_bio_prison_cell_v2 *prealloc;
1645
1646 prealloc = alloc_prison_cell(cache);
1647 if (!prealloc) {
1648 invalidate_complete(mg, false);
1649 return -ENOMEM;
1650 }
1651
1652 build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1653 r = dm_cell_lock_v2(cache->prison, &key,
1654 READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1655 if (r < 0) {
1656 free_prison_cell(cache, prealloc);
1657 invalidate_complete(mg, false);
1658 return r;
1659 }
1660
1661 if (mg->cell != prealloc)
1662 free_prison_cell(cache, prealloc);
1663
1664 if (r)
1665 quiesce(mg, invalidate_remove);
1666
1667 else {
1668 /*
1669 * We can't call invalidate_remove() directly here because we
1670 * might still be in request context.
1671 */
1672 init_continuation(&mg->k, invalidate_remove);
1673 queue_work(cache->wq, &mg->k.ws);
1674 }
1675
1676 return 0;
1677}
1678
1679static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1680 dm_oblock_t oblock, struct bio *bio)
1681{
1682 struct dm_cache_migration *mg;
1683
1684 if (!background_work_begin(cache))
1685 return -EPERM;
1686
1687 mg = alloc_migration(cache);
1688 if (!mg) {
1689 background_work_end(cache);
1690 return -ENOMEM;
1691 }
1692
1693 memset(mg, 0, sizeof(*mg));
1694
1695 mg->cache = cache;
1696 mg->overwrite_bio = bio;
1697 mg->invalidate_cblock = cblock;
1698 mg->invalidate_oblock = oblock;
1699
1700 return invalidate_lock(mg);
Joe Thornber7ae34e72014-11-06 10:18:04 +00001701}
1702
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001703/*----------------------------------------------------------------
1704 * bio processing
1705 *--------------------------------------------------------------*/
Joe Thornberb29d4982016-12-15 04:57:31 -05001706
1707enum busy {
1708 IDLE,
Joe Thornberb29d4982016-12-15 04:57:31 -05001709 BUSY
1710};
1711
1712static enum busy spare_migration_bandwidth(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001713{
Joe Thornber701e03e42017-05-11 08:22:31 -04001714 bool idle = iot_idle_for(&cache->tracker, HZ);
Joe Thornbera59db672015-01-23 10:16:16 +00001715 sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001716 cache->sectors_per_block;
Joe Thornberb29d4982016-12-15 04:57:31 -05001717
Joe Thornber49b7f762017-05-11 09:07:16 -04001718 if (idle && current_volume <= cache->migration_threshold)
1719 return IDLE;
Joe Thornberb29d4982016-12-15 04:57:31 -05001720 else
Joe Thornber49b7f762017-05-11 09:07:16 -04001721 return BUSY;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001722}
1723
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001724static void inc_hit_counter(struct cache *cache, struct bio *bio)
1725{
1726 atomic_inc(bio_data_dir(bio) == READ ?
1727 &cache->stats.read_hit : &cache->stats.write_hit);
1728}
1729
1730static void inc_miss_counter(struct cache *cache, struct bio *bio)
1731{
1732 atomic_inc(bio_data_dir(bio) == READ ?
1733 &cache->stats.read_miss : &cache->stats.write_miss);
1734}
1735
Joe Thornberfb4100a2015-05-20 10:30:32 +01001736/*----------------------------------------------------------------*/
1737
Joe Thornberb29d4982016-12-15 04:57:31 -05001738static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1739 bool *commit_needed)
1740{
1741 int r, data_dir;
1742 bool rb, background_queued;
1743 dm_cblock_t cblock;
1744 size_t pb_data_size = get_per_bio_data_size(cache);
1745 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1746
1747 *commit_needed = false;
1748
1749 rb = bio_detain_shared(cache, block, bio);
1750 if (!rb) {
1751 /*
1752 * An exclusive lock is held for this block, so we have to
1753 * wait. We set the commit_needed flag so the current
1754 * transaction will be committed asap, allowing this lock
1755 * to be dropped.
1756 */
1757 *commit_needed = true;
1758 return DM_MAPIO_SUBMITTED;
1759 }
1760
1761 data_dir = bio_data_dir(bio);
1762
1763 if (optimisable_bio(cache, bio, block)) {
1764 struct policy_work *op = NULL;
1765
1766 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1767 if (unlikely(r && r != -ENOENT)) {
1768 DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1769 cache_device_name(cache), r);
1770 bio_io_error(bio);
1771 return DM_MAPIO_SUBMITTED;
Joe Thornber651f5fa2015-05-15 15:26:08 +01001772 }
1773
Joe Thornberb29d4982016-12-15 04:57:31 -05001774 if (r == -ENOENT && op) {
1775 bio_drop_shared_lock(cache, bio);
1776 BUG_ON(op->op != POLICY_PROMOTE);
1777 mg_start(cache, op, bio);
1778 return DM_MAPIO_SUBMITTED;
1779 }
1780 } else {
1781 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1782 if (unlikely(r && r != -ENOENT)) {
1783 DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1784 cache_device_name(cache), r);
1785 bio_io_error(bio);
1786 return DM_MAPIO_SUBMITTED;
Joe Thornber2ee57d52013-10-24 14:10:29 -04001787 }
1788
Joe Thornberb29d4982016-12-15 04:57:31 -05001789 if (background_queued)
1790 wake_migration_worker(cache);
1791 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001792
Joe Thornberb29d4982016-12-15 04:57:31 -05001793 if (r == -ENOENT) {
1794 /*
1795 * Miss.
1796 */
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001797 inc_miss_counter(cache, bio);
Joe Thornberb29d4982016-12-15 04:57:31 -05001798 if (pb->req_nr == 0) {
1799 accounted_begin(cache, bio);
1800 remap_to_origin_clear_discard(cache, bio, block);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001801
Joe Thornberb29d4982016-12-15 04:57:31 -05001802 } else {
1803 /*
1804 * This is a duplicate writethrough io that is no
1805 * longer needed because the block has been demoted.
1806 */
1807 bio_endio(bio);
1808 return DM_MAPIO_SUBMITTED;
1809 }
1810 } else {
1811 /*
1812 * Hit.
1813 */
1814 inc_hit_counter(cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001815
Joe Thornberb29d4982016-12-15 04:57:31 -05001816 /*
1817 * Passthrough always maps to the origin, invalidating any
1818 * cache blocks that are written to.
1819 */
Mike Snitzer755b4e82017-10-19 21:01:04 -04001820 if (passthrough_mode(cache)) {
Joe Thornberb29d4982016-12-15 04:57:31 -05001821 if (bio_data_dir(bio) == WRITE) {
1822 bio_drop_shared_lock(cache, bio);
1823 atomic_inc(&cache->stats.demotion);
1824 invalidate_start(cache, cblock, block, bio);
1825 } else
1826 remap_to_origin_clear_discard(cache, bio, block);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001827
Joe Thornberb29d4982016-12-15 04:57:31 -05001828 } else {
Mike Snitzer755b4e82017-10-19 21:01:04 -04001829 if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
Joe Thornberb29d4982016-12-15 04:57:31 -05001830 !is_dirty(cache, cblock)) {
Mike Snitzer80363352017-10-19 17:16:54 -04001831 remap_to_origin_and_cache(cache, bio, block, cblock);
Joe Thornberb29d4982016-12-15 04:57:31 -05001832 accounted_begin(cache, bio);
1833 } else
1834 remap_to_cache_dirty(cache, bio, block, cblock);
1835 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001836 }
1837
Joe Thornber651f5fa2015-05-15 15:26:08 +01001838 /*
Joe Thornberb29d4982016-12-15 04:57:31 -05001839 * dm core turns FUA requests into a separate payload and FLUSH req.
Joe Thornber651f5fa2015-05-15 15:26:08 +01001840 */
Joe Thornberb29d4982016-12-15 04:57:31 -05001841 if (bio->bi_opf & REQ_FUA) {
1842 /*
1843 * issue_after_commit will call accounted_begin a second time. So
1844 * we call accounted_complete() to avoid double accounting.
1845 */
1846 accounted_complete(cache, bio);
1847 issue_after_commit(&cache->committer, bio);
1848 *commit_needed = true;
1849 return DM_MAPIO_SUBMITTED;
1850 }
Joe Thornber651f5fa2015-05-15 15:26:08 +01001851
Joe Thornberb29d4982016-12-15 04:57:31 -05001852 return DM_MAPIO_REMAPPED;
Joe Thornber651f5fa2015-05-15 15:26:08 +01001853}
1854
Joe Thornberb29d4982016-12-15 04:57:31 -05001855static bool process_bio(struct cache *cache, struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001856{
Joe Thornberb29d4982016-12-15 04:57:31 -05001857 bool commit_needed;
1858
1859 if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1860 generic_make_request(bio);
1861
1862 return commit_needed;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001863}
1864
Joe Thornber028ae9f2015-04-22 16:42:35 -04001865/*
1866 * A non-zero return indicates read_only or fail_io mode.
1867 */
1868static int commit(struct cache *cache, bool clean_shutdown)
1869{
1870 int r;
1871
1872 if (get_cache_mode(cache) >= CM_READ_ONLY)
1873 return -EINVAL;
1874
1875 atomic_inc(&cache->stats.commit_count);
1876 r = dm_cache_commit(cache->cmd, clean_shutdown);
1877 if (r)
1878 metadata_operation_failed(cache, "dm_cache_commit", r);
1879
1880 return r;
1881}
1882
Joe Thornberb29d4982016-12-15 04:57:31 -05001883/*
1884 * Used by the batcher.
1885 */
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001886static blk_status_t commit_op(void *context)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001887{
Joe Thornberb29d4982016-12-15 04:57:31 -05001888 struct cache *cache = context;
Heinz Mauelshagenffcbcb62013-10-14 17:24:43 +02001889
Joe Thornberb29d4982016-12-15 04:57:31 -05001890 if (dm_cache_changed_this_transaction(cache->cmd))
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001891 return errno_to_blk_status(commit(cache, false));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001892
Joe Thornberb29d4982016-12-15 04:57:31 -05001893 return 0;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001894}
1895
Joe Thornberb29d4982016-12-15 04:57:31 -05001896/*----------------------------------------------------------------*/
1897
1898static bool process_flush_bio(struct cache *cache, struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001899{
Joe Thornberb29d4982016-12-15 04:57:31 -05001900 size_t pb_data_size = get_per_bio_data_size(cache);
1901 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1902
1903 if (!pb->req_nr)
1904 remap_to_origin(cache, bio);
1905 else
1906 remap_to_cache(cache, bio, 0);
1907
1908 issue_after_commit(&cache->committer, bio);
1909 return true;
1910}
1911
1912static bool process_discard_bio(struct cache *cache, struct bio *bio)
1913{
1914 dm_dblock_t b, e;
1915
1916 // FIXME: do we need to lock the region? Or can we just assume the
1917 // user wont be so foolish as to issue discard concurrently with
1918 // other IO?
1919 calc_discard_block_range(cache, bio, &b, &e);
1920 while (b != e) {
1921 set_discard(cache, b);
1922 b = to_dblock(from_dblock(b) + 1);
1923 }
1924
1925 bio_endio(bio);
1926
1927 return false;
1928}
1929
1930static void process_deferred_bios(struct work_struct *ws)
1931{
1932 struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1933
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001934 unsigned long flags;
Joe Thornberb29d4982016-12-15 04:57:31 -05001935 bool commit_needed = false;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001936 struct bio_list bios;
1937 struct bio *bio;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001938
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001939 bio_list_init(&bios);
1940
1941 spin_lock_irqsave(&cache->lock, flags);
1942 bio_list_merge(&bios, &cache->deferred_bios);
1943 bio_list_init(&cache->deferred_bios);
1944 spin_unlock_irqrestore(&cache->lock, flags);
1945
Joe Thornberb29d4982016-12-15 04:57:31 -05001946 while ((bio = bio_list_pop(&bios))) {
Jens Axboe1eff9d32016-08-05 15:35:16 -06001947 if (bio->bi_opf & REQ_PREFLUSH)
Joe Thornberb29d4982016-12-15 04:57:31 -05001948 commit_needed = process_flush_bio(cache, bio) || commit_needed;
1949
Mike Christiee6047142016-06-05 14:32:04 -05001950 else if (bio_op(bio) == REQ_OP_DISCARD)
Joe Thornberb29d4982016-12-15 04:57:31 -05001951 commit_needed = process_discard_bio(cache, bio) || commit_needed;
1952
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001953 else
Joe Thornberb29d4982016-12-15 04:57:31 -05001954 commit_needed = process_bio(cache, bio) || commit_needed;
Mike Snitzer95e4e122023-02-16 15:31:08 -05001955 cond_resched();
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001956 }
1957
Joe Thornberb29d4982016-12-15 04:57:31 -05001958 if (commit_needed)
1959 schedule_commit(&cache->committer);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001960}
1961
Joe Thornber65790ff2013-11-08 16:39:50 +00001962/*----------------------------------------------------------------
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001963 * Main worker loop
1964 *--------------------------------------------------------------*/
Joe Thornber651f5fa2015-05-15 15:26:08 +01001965
1966static void requeue_deferred_bios(struct cache *cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001967{
1968 struct bio *bio;
1969 struct bio_list bios;
1970
1971 bio_list_init(&bios);
1972 bio_list_merge(&bios, &cache->deferred_bios);
1973 bio_list_init(&cache->deferred_bios);
1974
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001975 while ((bio = bio_list_pop(&bios))) {
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02001976 bio->bi_status = BLK_STS_DM_REQUEUE;
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001977 bio_endio(bio);
Mike Snitzer95e4e122023-02-16 15:31:08 -05001978 cond_resched();
Christoph Hellwig4246a0b2015-07-20 15:29:37 +02001979 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001980}
1981
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001982/*
1983 * We want to commit periodically so that not too much
1984 * unwritten metadata builds up.
1985 */
1986static void do_waker(struct work_struct *ws)
1987{
1988 struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
Joe Thornberb29d4982016-12-15 04:57:31 -05001989
Joe Thornberfba10102015-05-29 10:20:56 +01001990 policy_tick(cache->policy, true);
Joe Thornberb29d4982016-12-15 04:57:31 -05001991 wake_migration_worker(cache);
1992 schedule_commit(&cache->committer);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001993 queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1994}
1995
Joe Thornberb29d4982016-12-15 04:57:31 -05001996static void check_migrations(struct work_struct *ws)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00001997{
Joe Thornberb29d4982016-12-15 04:57:31 -05001998 int r;
1999 struct policy_work *op;
2000 struct cache *cache = container_of(ws, struct cache, migration_worker);
2001 enum busy b;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002002
Joe Thornberb29d4982016-12-15 04:57:31 -05002003 for (;;) {
2004 b = spare_migration_bandwidth(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002005
Joe Thornberb29d4982016-12-15 04:57:31 -05002006 r = policy_get_background_work(cache->policy, b == IDLE, &op);
2007 if (r == -ENODATA)
2008 break;
2009
2010 if (r) {
2011 DMERR_LIMIT("%s: policy_background_work failed",
2012 cache_device_name(cache));
2013 break;
2014 }
2015
2016 r = mg_start(cache, op, NULL);
2017 if (r)
2018 break;
Mike Snitzer95e4e122023-02-16 15:31:08 -05002019
2020 cond_resched();
Joe Thornberb29d4982016-12-15 04:57:31 -05002021 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002022}
2023
2024/*----------------------------------------------------------------
2025 * Target methods
2026 *--------------------------------------------------------------*/
2027
2028/*
2029 * This function gets called on the error paths of the constructor, so we
2030 * have to cope with a partially initialised struct.
2031 */
2032static void destroy(struct cache *cache)
2033{
2034 unsigned i;
2035
Julia Lawall6f659852015-09-13 14:15:05 +02002036 mempool_destroy(cache->migration_pool);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002037
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002038 if (cache->prison)
Joe Thornberb29d4982016-12-15 04:57:31 -05002039 dm_bio_prison_destroy_v2(cache->prison);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002040
Luo Meng99340612022-11-29 10:48:49 +08002041 cancel_delayed_work_sync(&cache->waker);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002042 if (cache->wq)
2043 destroy_workqueue(cache->wq);
2044
2045 if (cache->dirty_bitset)
2046 free_bitset(cache->dirty_bitset);
2047
2048 if (cache->discard_bitset)
2049 free_bitset(cache->discard_bitset);
2050
2051 if (cache->copier)
2052 dm_kcopyd_client_destroy(cache->copier);
2053
2054 if (cache->cmd)
2055 dm_cache_metadata_close(cache->cmd);
2056
2057 if (cache->metadata_dev)
2058 dm_put_device(cache->ti, cache->metadata_dev);
2059
2060 if (cache->origin_dev)
2061 dm_put_device(cache->ti, cache->origin_dev);
2062
2063 if (cache->cache_dev)
2064 dm_put_device(cache->ti, cache->cache_dev);
2065
2066 if (cache->policy)
2067 dm_cache_policy_destroy(cache->policy);
2068
2069 for (i = 0; i < cache->nr_ctr_args ; i++)
2070 kfree(cache->ctr_args[i]);
2071 kfree(cache->ctr_args);
2072
Mike Snitzer80363352017-10-19 17:16:54 -04002073 if (cache->bs)
2074 bioset_free(cache->bs);
2075
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002076 kfree(cache);
2077}
2078
2079static void cache_dtr(struct dm_target *ti)
2080{
2081 struct cache *cache = ti->private;
2082
2083 destroy(cache);
2084}
2085
2086static sector_t get_dev_size(struct dm_dev *dev)
2087{
2088 return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2089}
2090
2091/*----------------------------------------------------------------*/
2092
2093/*
2094 * Construct a cache device mapping.
2095 *
2096 * cache <metadata dev> <cache dev> <origin dev> <block size>
2097 * <#feature args> [<feature arg>]*
2098 * <policy> <#policy args> [<policy arg>]*
2099 *
2100 * metadata dev : fast device holding the persistent metadata
2101 * cache dev : fast device holding cached data blocks
2102 * origin dev : slow device holding original data blocks
2103 * block size : cache unit size in sectors
2104 *
2105 * #feature args : number of feature arguments passed
2106 * feature args : writethrough. (The default is writeback.)
2107 *
2108 * policy : the replacement policy to use
2109 * #policy args : an even number of policy arguments corresponding
2110 * to key/value pairs passed to the policy
2111 * policy args : key/value pairs passed to the policy
2112 * E.g. 'sequential_threshold 1024'
2113 * See cache-policies.txt for details.
2114 *
2115 * Optional feature arguments are:
2116 * writethrough : write through caching that prohibits cache block
2117 * content from being different from origin block content.
2118 * Without this argument, the default behaviour is to write
2119 * back cache block contents later for performance reasons,
2120 * so they may differ from the corresponding origin blocks.
2121 */
2122struct cache_args {
2123 struct dm_target *ti;
2124
2125 struct dm_dev *metadata_dev;
2126
2127 struct dm_dev *cache_dev;
2128 sector_t cache_sectors;
2129
2130 struct dm_dev *origin_dev;
2131 sector_t origin_sectors;
2132
2133 uint32_t block_size;
2134
2135 const char *policy_name;
2136 int policy_argc;
2137 const char **policy_argv;
2138
2139 struct cache_features features;
2140};
2141
2142static void destroy_cache_args(struct cache_args *ca)
2143{
2144 if (ca->metadata_dev)
2145 dm_put_device(ca->ti, ca->metadata_dev);
2146
2147 if (ca->cache_dev)
2148 dm_put_device(ca->ti, ca->cache_dev);
2149
2150 if (ca->origin_dev)
2151 dm_put_device(ca->ti, ca->origin_dev);
2152
2153 kfree(ca);
2154}
2155
2156static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2157{
2158 if (!as->argc) {
2159 *error = "Insufficient args";
2160 return false;
2161 }
2162
2163 return true;
2164}
2165
2166static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2167 char **error)
2168{
2169 int r;
2170 sector_t metadata_dev_size;
2171 char b[BDEVNAME_SIZE];
2172
2173 if (!at_least_one_arg(as, error))
2174 return -EINVAL;
2175
2176 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2177 &ca->metadata_dev);
2178 if (r) {
2179 *error = "Error opening metadata device";
2180 return r;
2181 }
2182
2183 metadata_dev_size = get_dev_size(ca->metadata_dev);
2184 if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2185 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2186 bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2187
2188 return 0;
2189}
2190
2191static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2192 char **error)
2193{
2194 int r;
2195
2196 if (!at_least_one_arg(as, error))
2197 return -EINVAL;
2198
2199 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2200 &ca->cache_dev);
2201 if (r) {
2202 *error = "Error opening cache device";
2203 return r;
2204 }
2205 ca->cache_sectors = get_dev_size(ca->cache_dev);
2206
2207 return 0;
2208}
2209
2210static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2211 char **error)
2212{
2213 int r;
2214
2215 if (!at_least_one_arg(as, error))
2216 return -EINVAL;
2217
2218 r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2219 &ca->origin_dev);
2220 if (r) {
2221 *error = "Error opening origin device";
2222 return r;
2223 }
2224
2225 ca->origin_sectors = get_dev_size(ca->origin_dev);
2226 if (ca->ti->len > ca->origin_sectors) {
2227 *error = "Device size larger than cached device";
2228 return -EINVAL;
2229 }
2230
2231 return 0;
2232}
2233
2234static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2235 char **error)
2236{
Mike Snitzer05473042013-08-16 10:54:19 -04002237 unsigned long block_size;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002238
2239 if (!at_least_one_arg(as, error))
2240 return -EINVAL;
2241
Mike Snitzer05473042013-08-16 10:54:19 -04002242 if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2243 block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2244 block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2245 block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002246 *error = "Invalid data block size";
2247 return -EINVAL;
2248 }
2249
Mike Snitzer05473042013-08-16 10:54:19 -04002250 if (block_size > ca->cache_sectors) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002251 *error = "Data block size is larger than the cache device";
2252 return -EINVAL;
2253 }
2254
Mike Snitzer05473042013-08-16 10:54:19 -04002255 ca->block_size = block_size;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002256
2257 return 0;
2258}
2259
2260static void init_features(struct cache_features *cf)
2261{
2262 cf->mode = CM_WRITE;
Joe Thornber2ee57d52013-10-24 14:10:29 -04002263 cf->io_mode = CM_IO_WRITEBACK;
Joe Thornber629d0a82016-09-22 06:15:21 -04002264 cf->metadata_version = 1;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002265}
2266
2267static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2268 char **error)
2269{
Eric Biggers5916a222017-06-22 11:32:45 -07002270 static const struct dm_arg _args[] = {
Joe Thornber629d0a82016-09-22 06:15:21 -04002271 {0, 2, "Invalid number of cache feature arguments"},
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002272 };
2273
John Pittmane85940a2018-06-21 17:35:33 -04002274 int r, mode_ctr = 0;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002275 unsigned argc;
2276 const char *arg;
2277 struct cache_features *cf = &ca->features;
2278
2279 init_features(cf);
2280
2281 r = dm_read_arg_group(_args, as, &argc, error);
2282 if (r)
2283 return -EINVAL;
2284
2285 while (argc--) {
2286 arg = dm_shift_arg(as);
2287
John Pittmane85940a2018-06-21 17:35:33 -04002288 if (!strcasecmp(arg, "writeback")) {
Joe Thornber2ee57d52013-10-24 14:10:29 -04002289 cf->io_mode = CM_IO_WRITEBACK;
John Pittmane85940a2018-06-21 17:35:33 -04002290 mode_ctr++;
2291 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002292
John Pittmane85940a2018-06-21 17:35:33 -04002293 else if (!strcasecmp(arg, "writethrough")) {
Joe Thornber2ee57d52013-10-24 14:10:29 -04002294 cf->io_mode = CM_IO_WRITETHROUGH;
John Pittmane85940a2018-06-21 17:35:33 -04002295 mode_ctr++;
2296 }
Joe Thornber2ee57d52013-10-24 14:10:29 -04002297
John Pittmane85940a2018-06-21 17:35:33 -04002298 else if (!strcasecmp(arg, "passthrough")) {
Joe Thornber2ee57d52013-10-24 14:10:29 -04002299 cf->io_mode = CM_IO_PASSTHROUGH;
John Pittmane85940a2018-06-21 17:35:33 -04002300 mode_ctr++;
2301 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002302
Joe Thornber629d0a82016-09-22 06:15:21 -04002303 else if (!strcasecmp(arg, "metadata2"))
2304 cf->metadata_version = 2;
2305
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002306 else {
2307 *error = "Unrecognised cache feature requested";
2308 return -EINVAL;
2309 }
2310 }
2311
John Pittmane85940a2018-06-21 17:35:33 -04002312 if (mode_ctr > 1) {
2313 *error = "Duplicate cache io_mode features requested";
2314 return -EINVAL;
2315 }
2316
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002317 return 0;
2318}
2319
2320static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2321 char **error)
2322{
Eric Biggers5916a222017-06-22 11:32:45 -07002323 static const struct dm_arg _args[] = {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002324 {0, 1024, "Invalid number of policy arguments"},
2325 };
2326
2327 int r;
2328
2329 if (!at_least_one_arg(as, error))
2330 return -EINVAL;
2331
2332 ca->policy_name = dm_shift_arg(as);
2333
2334 r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2335 if (r)
2336 return -EINVAL;
2337
2338 ca->policy_argv = (const char **)as->argv;
2339 dm_consume_args(as, ca->policy_argc);
2340
2341 return 0;
2342}
2343
2344static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2345 char **error)
2346{
2347 int r;
2348 struct dm_arg_set as;
2349
2350 as.argc = argc;
2351 as.argv = argv;
2352
2353 r = parse_metadata_dev(ca, &as, error);
2354 if (r)
2355 return r;
2356
2357 r = parse_cache_dev(ca, &as, error);
2358 if (r)
2359 return r;
2360
2361 r = parse_origin_dev(ca, &as, error);
2362 if (r)
2363 return r;
2364
2365 r = parse_block_size(ca, &as, error);
2366 if (r)
2367 return r;
2368
2369 r = parse_features(ca, &as, error);
2370 if (r)
2371 return r;
2372
2373 r = parse_policy(ca, &as, error);
2374 if (r)
2375 return r;
2376
2377 return 0;
2378}
2379
2380/*----------------------------------------------------------------*/
2381
2382static struct kmem_cache *migration_cache;
2383
Alasdair G Kergon2c73c472013-05-10 14:37:21 +01002384#define NOT_CORE_OPTION 1
2385
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002386static int process_config_option(struct cache *cache, const char *key, const char *value)
Alasdair G Kergon2c73c472013-05-10 14:37:21 +01002387{
2388 unsigned long tmp;
2389
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002390 if (!strcasecmp(key, "migration_threshold")) {
2391 if (kstrtoul(value, 10, &tmp))
Alasdair G Kergon2c73c472013-05-10 14:37:21 +01002392 return -EINVAL;
2393
2394 cache->migration_threshold = tmp;
2395 return 0;
2396 }
2397
2398 return NOT_CORE_OPTION;
2399}
2400
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002401static int set_config_value(struct cache *cache, const char *key, const char *value)
2402{
2403 int r = process_config_option(cache, key, value);
2404
2405 if (r == NOT_CORE_OPTION)
2406 r = policy_set_config_value(cache->policy, key, value);
2407
2408 if (r)
2409 DMWARN("bad config value for %s: %s", key, value);
2410
2411 return r;
2412}
2413
2414static int set_config_values(struct cache *cache, int argc, const char **argv)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002415{
2416 int r = 0;
2417
2418 if (argc & 1) {
2419 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2420 return -EINVAL;
2421 }
2422
2423 while (argc) {
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002424 r = set_config_value(cache, argv[0], argv[1]);
2425 if (r)
2426 break;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002427
2428 argc -= 2;
2429 argv += 2;
2430 }
2431
2432 return r;
2433}
2434
2435static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2436 char **error)
2437{
Mikulas Patocka4cb3e1d2013-10-01 18:35:39 -04002438 struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2439 cache->cache_size,
2440 cache->origin_sectors,
2441 cache->sectors_per_block);
2442 if (IS_ERR(p)) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002443 *error = "Error creating cache's policy";
Mikulas Patocka4cb3e1d2013-10-01 18:35:39 -04002444 return PTR_ERR(p);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002445 }
Mikulas Patocka4cb3e1d2013-10-01 18:35:39 -04002446 cache->policy = p;
Joe Thornberb29d4982016-12-15 04:57:31 -05002447 BUG_ON(!cache->policy);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002448
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002449 return 0;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002450}
2451
Joe Thornber08b18452014-11-06 14:38:01 +00002452/*
Joe Thornber2bb812d2014-11-26 16:07:50 +00002453 * We want the discard block size to be at least the size of the cache
2454 * block size and have no more than 2^14 discard blocks across the origin.
Joe Thornber08b18452014-11-06 14:38:01 +00002455 */
2456#define MAX_DISCARD_BLOCKS (1 << 14)
2457
2458static bool too_many_discard_blocks(sector_t discard_block_size,
2459 sector_t origin_size)
2460{
2461 (void) sector_div(origin_size, discard_block_size);
2462
2463 return origin_size > MAX_DISCARD_BLOCKS;
2464}
2465
2466static sector_t calculate_discard_block_size(sector_t cache_block_size,
2467 sector_t origin_size)
2468{
Joe Thornber2bb812d2014-11-26 16:07:50 +00002469 sector_t discard_block_size = cache_block_size;
Joe Thornber08b18452014-11-06 14:38:01 +00002470
2471 if (origin_size)
2472 while (too_many_discard_blocks(discard_block_size, origin_size))
2473 discard_block_size *= 2;
2474
2475 return discard_block_size;
2476}
2477
Joe Thornberd1d92202014-11-11 11:58:32 +00002478static void set_cache_size(struct cache *cache, dm_cblock_t size)
2479{
2480 dm_block_t nr_blocks = from_cblock(size);
2481
2482 if (nr_blocks > (1 << 20) && cache->cache_size != size)
2483 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2484 "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2485 "Please consider increasing the cache block size to reduce the overall cache block count.",
2486 (unsigned long long) nr_blocks);
2487
2488 cache->cache_size = size;
2489}
2490
Joe Thornberb29d4982016-12-15 04:57:31 -05002491static int is_congested(struct dm_dev *dev, int bdi_bits)
2492{
2493 struct request_queue *q = bdev_get_queue(dev->bdev);
2494 return bdi_congested(q->backing_dev_info, bdi_bits);
2495}
2496
2497static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2498{
2499 struct cache *cache = container_of(cb, struct cache, callbacks);
2500
2501 return is_congested(cache->origin_dev, bdi_bits) ||
2502 is_congested(cache->cache_dev, bdi_bits);
2503}
2504
Joe Thornberf8350da2013-05-10 14:37:16 +01002505#define DEFAULT_MIGRATION_THRESHOLD 2048
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002506
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002507static int cache_create(struct cache_args *ca, struct cache **result)
2508{
2509 int r = 0;
2510 char **error = &ca->ti->error;
2511 struct cache *cache;
2512 struct dm_target *ti = ca->ti;
2513 dm_block_t origin_blocks;
2514 struct dm_cache_metadata *cmd;
2515 bool may_format = ca->features.mode == CM_WRITE;
2516
2517 cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2518 if (!cache)
2519 return -ENOMEM;
2520
2521 cache->ti = ca->ti;
2522 ti->private = cache;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002523 ti->num_flush_bios = 2;
2524 ti->flush_supported = true;
2525
2526 ti->num_discard_bios = 1;
2527 ti->discards_supported = true;
Joe Thornber25726292014-11-24 14:05:16 +00002528 ti->split_discard_bios = false;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002529
Joe Thornber8c5008f2013-05-10 14:37:18 +01002530 cache->features = ca->features;
Mike Snitzer30187e12016-01-31 13:28:26 -05002531 ti->per_io_data_size = get_per_bio_data_size(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002532
Mike Snitzer80363352017-10-19 17:16:54 -04002533 if (writethrough_mode(cache)) {
2534 /* Create bioset for writethrough bios issued to origin */
2535 cache->bs = bioset_create(BIO_POOL_SIZE, 0, 0);
2536 if (!cache->bs)
2537 goto bad;
2538 }
2539
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002540 cache->callbacks.congested_fn = cache_is_congested;
2541 dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2542
2543 cache->metadata_dev = ca->metadata_dev;
2544 cache->origin_dev = ca->origin_dev;
2545 cache->cache_dev = ca->cache_dev;
2546
2547 ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2548
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002549 origin_blocks = cache->origin_sectors = ca->origin_sectors;
Joe Thornber414dd672013-03-20 17:21:25 +00002550 origin_blocks = block_div(origin_blocks, ca->block_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002551 cache->origin_blocks = to_oblock(origin_blocks);
2552
2553 cache->sectors_per_block = ca->block_size;
2554 if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2555 r = -EINVAL;
2556 goto bad;
2557 }
2558
2559 if (ca->block_size & (ca->block_size - 1)) {
2560 dm_block_t cache_size = ca->cache_sectors;
2561
2562 cache->sectors_per_block_shift = -1;
Joe Thornber414dd672013-03-20 17:21:25 +00002563 cache_size = block_div(cache_size, ca->block_size);
Joe Thornberd1d92202014-11-11 11:58:32 +00002564 set_cache_size(cache, to_cblock(cache_size));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002565 } else {
2566 cache->sectors_per_block_shift = __ffs(ca->block_size);
Joe Thornberd1d92202014-11-11 11:58:32 +00002567 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002568 }
2569
2570 r = create_cache_policy(cache, ca, error);
2571 if (r)
2572 goto bad;
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002573
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002574 cache->policy_nr_args = ca->policy_argc;
Joe Thornber2f14f4b2013-05-10 14:37:21 +01002575 cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2576
2577 r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2578 if (r) {
2579 *error = "Error setting cache policy's config values";
2580 goto bad;
2581 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002582
2583 cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2584 ca->block_size, may_format,
Joe Thornber629d0a82016-09-22 06:15:21 -04002585 dm_cache_policy_get_hint_size(cache->policy),
2586 ca->features.metadata_version);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002587 if (IS_ERR(cmd)) {
2588 *error = "Error creating metadata object";
2589 r = PTR_ERR(cmd);
2590 goto bad;
2591 }
2592 cache->cmd = cmd;
Joe Thornber028ae9f2015-04-22 16:42:35 -04002593 set_cache_mode(cache, CM_WRITE);
2594 if (get_cache_mode(cache) != CM_WRITE) {
2595 *error = "Unable to get write access to metadata, please check/repair metadata.";
2596 r = -EINVAL;
2597 goto bad;
2598 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002599
Mike Snitzer755b4e82017-10-19 21:01:04 -04002600 if (passthrough_mode(cache)) {
Joe Thornber2ee57d52013-10-24 14:10:29 -04002601 bool all_clean;
2602
2603 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2604 if (r) {
2605 *error = "dm_cache_metadata_all_clean() failed";
2606 goto bad;
2607 }
2608
2609 if (!all_clean) {
2610 *error = "Cannot enter passthrough mode unless all blocks are clean";
2611 r = -EINVAL;
2612 goto bad;
2613 }
Joe Thornberb29d4982016-12-15 04:57:31 -05002614
2615 policy_allow_migrations(cache->policy, false);
Joe Thornber2ee57d52013-10-24 14:10:29 -04002616 }
2617
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002618 spin_lock_init(&cache->lock);
Joe Thornber651f5fa2015-05-15 15:26:08 +01002619 INIT_LIST_HEAD(&cache->deferred_cells);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002620 bio_list_init(&cache->deferred_bios);
Joe Thornbera59db672015-01-23 10:16:16 +00002621 atomic_set(&cache->nr_allocated_migrations, 0);
2622 atomic_set(&cache->nr_io_migrations, 0);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002623 init_waitqueue_head(&cache->migration_wait);
2624
Wei Yongjunfa4d6832013-05-10 14:37:14 +01002625 r = -ENOMEM;
Anssi Hannula44fa8162014-08-01 11:55:47 -04002626 atomic_set(&cache->nr_dirty, 0);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002627 cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2628 if (!cache->dirty_bitset) {
2629 *error = "could not allocate dirty bitset";
2630 goto bad;
2631 }
2632 clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2633
Joe Thornber08b18452014-11-06 14:38:01 +00002634 cache->discard_block_size =
2635 calculate_discard_block_size(cache->sectors_per_block,
2636 cache->origin_sectors);
Joe Thornber25726292014-11-24 14:05:16 +00002637 cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2638 cache->discard_block_size));
Joe Thornber1bad9bc2014-11-07 14:47:07 +00002639 cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002640 if (!cache->discard_bitset) {
2641 *error = "could not allocate discard bitset";
2642 goto bad;
2643 }
Joe Thornber1bad9bc2014-11-07 14:47:07 +00002644 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002645
2646 cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2647 if (IS_ERR(cache->copier)) {
2648 *error = "could not create kcopyd client";
2649 r = PTR_ERR(cache->copier);
2650 goto bad;
2651 }
2652
Joe Thornberb29d4982016-12-15 04:57:31 -05002653 cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002654 if (!cache->wq) {
2655 *error = "could not create workqueue for metadata object";
2656 goto bad;
2657 }
Joe Thornberb29d4982016-12-15 04:57:31 -05002658 INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
Joe Thornberb29d4982016-12-15 04:57:31 -05002659 INIT_WORK(&cache->migration_worker, check_migrations);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002660 INIT_DELAYED_WORK(&cache->waker, do_waker);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002661
Joe Thornberb29d4982016-12-15 04:57:31 -05002662 cache->prison = dm_bio_prison_create_v2(cache->wq);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002663 if (!cache->prison) {
2664 *error = "could not create bio prison";
2665 goto bad;
2666 }
2667
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002668 cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2669 migration_cache);
2670 if (!cache->migration_pool) {
2671 *error = "Error creating cache's migration mempool";
2672 goto bad;
2673 }
2674
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002675 cache->need_tick_bio = true;
2676 cache->sized = false;
Joe Thornber65790ff2013-11-08 16:39:50 +00002677 cache->invalidate = false;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002678 cache->commit_requested = false;
2679 cache->loaded_mappings = false;
2680 cache->loaded_discards = false;
2681
2682 load_stats(cache);
2683
2684 atomic_set(&cache->stats.demotion, 0);
2685 atomic_set(&cache->stats.promotion, 0);
2686 atomic_set(&cache->stats.copies_avoided, 0);
2687 atomic_set(&cache->stats.cache_cell_clash, 0);
2688 atomic_set(&cache->stats.commit_count, 0);
2689 atomic_set(&cache->stats.discard_count, 0);
2690
Joe Thornber65790ff2013-11-08 16:39:50 +00002691 spin_lock_init(&cache->invalidation_lock);
2692 INIT_LIST_HEAD(&cache->invalidation_requests);
2693
Joe Thornberb29d4982016-12-15 04:57:31 -05002694 batcher_init(&cache->committer, commit_op, cache,
2695 issue_op, cache, cache->wq);
Joe Thornber701e03e42017-05-11 08:22:31 -04002696 iot_init(&cache->tracker);
Joe Thornber066dbaa2015-05-15 15:18:01 +01002697
Joe Thornberb29d4982016-12-15 04:57:31 -05002698 init_rwsem(&cache->background_work_lock);
2699 prevent_background_work(cache);
2700
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002701 *result = cache;
2702 return 0;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002703bad:
2704 destroy(cache);
2705 return r;
2706}
2707
2708static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2709{
2710 unsigned i;
2711 const char **copy;
2712
2713 copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2714 if (!copy)
2715 return -ENOMEM;
2716 for (i = 0; i < argc; i++) {
2717 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2718 if (!copy[i]) {
2719 while (i--)
2720 kfree(copy[i]);
2721 kfree(copy);
2722 return -ENOMEM;
2723 }
2724 }
2725
2726 cache->nr_ctr_args = argc;
2727 cache->ctr_args = copy;
2728
2729 return 0;
2730}
2731
2732static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2733{
2734 int r = -EINVAL;
2735 struct cache_args *ca;
2736 struct cache *cache = NULL;
2737
2738 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2739 if (!ca) {
2740 ti->error = "Error allocating memory for cache";
2741 return -ENOMEM;
2742 }
2743 ca->ti = ti;
2744
2745 r = parse_cache_args(ca, argc, argv, &ti->error);
2746 if (r)
2747 goto out;
2748
2749 r = cache_create(ca, &cache);
Heinz Mauelshagen617a0b82013-03-20 17:21:26 +00002750 if (r)
2751 goto out;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002752
2753 r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2754 if (r) {
2755 destroy(cache);
2756 goto out;
2757 }
2758
2759 ti->private = cache;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002760out:
2761 destroy_cache_args(ca);
2762 return r;
2763}
2764
Joe Thornber651f5fa2015-05-15 15:26:08 +01002765/*----------------------------------------------------------------*/
2766
2767static int cache_map(struct dm_target *ti, struct bio *bio)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002768{
Joe Thornber651f5fa2015-05-15 15:26:08 +01002769 struct cache *cache = ti->private;
2770
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002771 int r;
Joe Thornberb29d4982016-12-15 04:57:31 -05002772 bool commit_needed;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002773 dm_oblock_t block = get_bio_block(cache, bio);
Mike Snitzer19b00922013-04-05 15:36:34 +01002774 size_t pb_data_size = get_per_bio_data_size(cache);
Joe Thornberfb4100a2015-05-20 10:30:32 +01002775
Joe Thornberb29d4982016-12-15 04:57:31 -05002776 init_per_bio_data(bio, pb_data_size);
Heinz Mauelshagene893fba2014-03-12 16:13:39 +01002777 if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002778 /*
2779 * This can only occur if the io goes to a partial block at
2780 * the end of the origin device. We don't cache these.
2781 * Just remap to the origin and carry on.
2782 */
Heinz Mauelshagene893fba2014-03-12 16:13:39 +01002783 remap_to_origin(cache, bio);
Joe Thornber651f5fa2015-05-15 15:26:08 +01002784 accounted_begin(cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002785 return DM_MAPIO_REMAPPED;
2786 }
2787
Joe Thornber651f5fa2015-05-15 15:26:08 +01002788 if (discard_or_flush(bio)) {
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002789 defer_bio(cache, bio);
2790 return DM_MAPIO_SUBMITTED;
2791 }
2792
Joe Thornberb29d4982016-12-15 04:57:31 -05002793 r = map_bio(cache, bio, block, &commit_needed);
2794 if (commit_needed)
2795 schedule_commit(&cache->committer);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002796
Joe Thornber2ee57d52013-10-24 14:10:29 -04002797 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002798}
2799
Christoph Hellwig4e4cbee2017-06-03 09:38:06 +02002800static int cache_end_io(struct dm_target *ti, struct bio *bio,
2801 blk_status_t *error)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002802{
2803 struct cache *cache = ti->private;
2804 unsigned long flags;
Mike Snitzer19b00922013-04-05 15:36:34 +01002805 size_t pb_data_size = get_per_bio_data_size(cache);
2806 struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002807
2808 if (pb->tick) {
Joe Thornberfba10102015-05-29 10:20:56 +01002809 policy_tick(cache->policy, false);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002810
2811 spin_lock_irqsave(&cache->lock, flags);
2812 cache->need_tick_bio = true;
2813 spin_unlock_irqrestore(&cache->lock, flags);
2814 }
2815
Joe Thornberb29d4982016-12-15 04:57:31 -05002816 bio_drop_shared_lock(cache, bio);
Joe Thornber066dbaa2015-05-15 15:18:01 +01002817 accounted_complete(cache, bio);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002818
Christoph Hellwig1be56902017-06-03 09:38:03 +02002819 return DM_ENDIO_DONE;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002820}
2821
2822static int write_dirty_bitset(struct cache *cache)
2823{
Joe Thornber629d0a82016-09-22 06:15:21 -04002824 int r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002825
Joe Thornber028ae9f2015-04-22 16:42:35 -04002826 if (get_cache_mode(cache) >= CM_READ_ONLY)
2827 return -EINVAL;
2828
Joe Thornber629d0a82016-09-22 06:15:21 -04002829 r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2830 if (r)
2831 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002832
Joe Thornber629d0a82016-09-22 06:15:21 -04002833 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002834}
2835
2836static int write_discard_bitset(struct cache *cache)
2837{
2838 unsigned i, r;
2839
Joe Thornber028ae9f2015-04-22 16:42:35 -04002840 if (get_cache_mode(cache) >= CM_READ_ONLY)
2841 return -EINVAL;
2842
Joe Thornber1bad9bc2014-11-07 14:47:07 +00002843 r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2844 cache->discard_nr_blocks);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002845 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04002846 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -04002847 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002848 return r;
2849 }
2850
Joe Thornber1bad9bc2014-11-07 14:47:07 +00002851 for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2852 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2853 is_discarded(cache, to_dblock(i)));
Joe Thornber028ae9f2015-04-22 16:42:35 -04002854 if (r) {
2855 metadata_operation_failed(cache, "dm_cache_set_discard", r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002856 return r;
Joe Thornber028ae9f2015-04-22 16:42:35 -04002857 }
2858 }
2859
2860 return 0;
2861}
2862
2863static int write_hints(struct cache *cache)
2864{
2865 int r;
2866
2867 if (get_cache_mode(cache) >= CM_READ_ONLY)
2868 return -EINVAL;
2869
2870 r = dm_cache_write_hints(cache->cmd, cache->policy);
2871 if (r) {
2872 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2873 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002874 }
2875
2876 return 0;
2877}
2878
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002879/*
2880 * returns true on success
2881 */
2882static bool sync_metadata(struct cache *cache)
2883{
2884 int r1, r2, r3, r4;
2885
2886 r1 = write_dirty_bitset(cache);
2887 if (r1)
Mike Snitzerb61d9502015-04-22 17:25:56 -04002888 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002889
2890 r2 = write_discard_bitset(cache);
2891 if (r2)
Mike Snitzerb61d9502015-04-22 17:25:56 -04002892 DMERR("%s: could not write discard bitset", cache_device_name(cache));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002893
2894 save_stats(cache);
2895
Joe Thornber028ae9f2015-04-22 16:42:35 -04002896 r3 = write_hints(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002897 if (r3)
Mike Snitzerb61d9502015-04-22 17:25:56 -04002898 DMERR("%s: could not write hints", cache_device_name(cache));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002899
2900 /*
2901 * If writing the above metadata failed, we still commit, but don't
2902 * set the clean shutdown flag. This will effectively force every
2903 * dirty bit to be set on reload.
2904 */
Joe Thornber028ae9f2015-04-22 16:42:35 -04002905 r4 = commit(cache, !r1 && !r2 && !r3);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002906 if (r4)
Mike Snitzerb61d9502015-04-22 17:25:56 -04002907 DMERR("%s: could not write cache metadata", cache_device_name(cache));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002908
2909 return !r1 && !r2 && !r3 && !r4;
2910}
2911
2912static void cache_postsuspend(struct dm_target *ti)
2913{
2914 struct cache *cache = ti->private;
2915
Joe Thornberb29d4982016-12-15 04:57:31 -05002916 prevent_background_work(cache);
2917 BUG_ON(atomic_read(&cache->nr_io_migrations));
2918
Mikulas Patocka165dc072020-02-19 10:25:45 -05002919 cancel_delayed_work_sync(&cache->waker);
2920 drain_workqueue(cache->wq);
Joe Thornber701e03e42017-05-11 08:22:31 -04002921 WARN_ON(cache->tracker.in_flight);
Joe Thornberb29d4982016-12-15 04:57:31 -05002922
2923 /*
2924 * If it's a flush suspend there won't be any deferred bios, so this
2925 * call is harmless.
2926 */
Joe Thornber651f5fa2015-05-15 15:26:08 +01002927 requeue_deferred_bios(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002928
Joe Thornber028ae9f2015-04-22 16:42:35 -04002929 if (get_cache_mode(cache) == CM_WRITE)
2930 (void) sync_metadata(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002931}
2932
2933static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2934 bool dirty, uint32_t hint, bool hint_valid)
2935{
2936 int r;
2937 struct cache *cache = context;
2938
Joe Thornber449b6682017-03-31 10:09:45 -04002939 if (dirty) {
2940 set_bit(from_cblock(cblock), cache->dirty_bitset);
2941 atomic_inc(&cache->nr_dirty);
2942 } else
2943 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2944
Joe Thornberb29d4982016-12-15 04:57:31 -05002945 r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002946 if (r)
2947 return r;
2948
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00002949 return 0;
2950}
2951
Joe Thornber3e2e1c32014-11-24 14:06:22 +00002952/*
2953 * The discard block size in the on disk metadata is not
2954 * neccessarily the same as we're currently using. So we have to
2955 * be careful to only set the discarded attribute if we know it
2956 * covers a complete block of the new size.
2957 */
2958struct discard_load_info {
2959 struct cache *cache;
2960
2961 /*
2962 * These blocks are sized using the on disk dblock size, rather
2963 * than the current one.
2964 */
2965 dm_block_t block_size;
2966 dm_block_t discard_begin, discard_end;
2967};
2968
2969static void discard_load_info_init(struct cache *cache,
2970 struct discard_load_info *li)
2971{
2972 li->cache = cache;
2973 li->discard_begin = li->discard_end = 0;
2974}
2975
2976static void set_discard_range(struct discard_load_info *li)
2977{
2978 sector_t b, e;
2979
2980 if (li->discard_begin == li->discard_end)
2981 return;
2982
2983 /*
2984 * Convert to sectors.
2985 */
2986 b = li->discard_begin * li->block_size;
2987 e = li->discard_end * li->block_size;
2988
2989 /*
2990 * Then convert back to the current dblock size.
2991 */
2992 b = dm_sector_div_up(b, li->cache->discard_block_size);
2993 sector_div(e, li->cache->discard_block_size);
2994
2995 /*
2996 * The origin may have shrunk, so we need to check we're still in
2997 * bounds.
2998 */
2999 if (e > from_dblock(li->cache->discard_nr_blocks))
3000 e = from_dblock(li->cache->discard_nr_blocks);
3001
3002 for (; b < e; b++)
3003 set_discard(li->cache, to_dblock(b));
3004}
3005
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003006static int load_discard(void *context, sector_t discard_block_size,
Joe Thornber1bad9bc2014-11-07 14:47:07 +00003007 dm_dblock_t dblock, bool discard)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003008{
Joe Thornber3e2e1c32014-11-24 14:06:22 +00003009 struct discard_load_info *li = context;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003010
Joe Thornber3e2e1c32014-11-24 14:06:22 +00003011 li->block_size = discard_block_size;
Joe Thornber1bad9bc2014-11-07 14:47:07 +00003012
Joe Thornber3e2e1c32014-11-24 14:06:22 +00003013 if (discard) {
3014 if (from_dblock(dblock) == li->discard_end)
3015 /*
3016 * We're already in a discard range, just extend it.
3017 */
3018 li->discard_end = li->discard_end + 1ULL;
3019
3020 else {
3021 /*
3022 * Emit the old range and start a new one.
3023 */
3024 set_discard_range(li);
3025 li->discard_begin = from_dblock(dblock);
3026 li->discard_end = li->discard_begin + 1ULL;
3027 }
3028 } else {
3029 set_discard_range(li);
3030 li->discard_begin = li->discard_end = 0;
3031 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003032
3033 return 0;
3034}
3035
Joe Thornberf494a9c2013-10-31 13:55:49 -04003036static dm_cblock_t get_cache_dev_size(struct cache *cache)
3037{
3038 sector_t size = get_dev_size(cache->cache_dev);
3039 (void) sector_div(size, cache->sectors_per_block);
3040 return to_cblock(size);
3041}
3042
3043static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3044{
Mike Snitzerec6ae632018-09-25 20:56:02 -04003045 if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
3046 if (cache->sized) {
3047 DMERR("%s: unable to extend cache due to missing cache table reload",
3048 cache_device_name(cache));
3049 return false;
3050 }
3051 }
Joe Thornberf494a9c2013-10-31 13:55:49 -04003052
3053 /*
3054 * We can't drop a dirty block when shrinking the cache.
3055 */
3056 while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3057 new_size = to_cblock(from_cblock(new_size) + 1);
3058 if (is_dirty(cache, new_size)) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003059 DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3060 cache_device_name(cache),
Joe Thornberf494a9c2013-10-31 13:55:49 -04003061 (unsigned long long) from_cblock(new_size));
3062 return false;
3063 }
3064 }
3065
3066 return true;
3067}
3068
3069static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3070{
3071 int r;
3072
Vincent Pelletier08844802013-11-30 12:58:42 +01003073 r = dm_cache_resize(cache->cmd, new_size);
Joe Thornberf494a9c2013-10-31 13:55:49 -04003074 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003075 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -04003076 metadata_operation_failed(cache, "dm_cache_resize", r);
Joe Thornberf494a9c2013-10-31 13:55:49 -04003077 return r;
3078 }
3079
Joe Thornberd1d92202014-11-11 11:58:32 +00003080 set_cache_size(cache, new_size);
Joe Thornberf494a9c2013-10-31 13:55:49 -04003081
3082 return 0;
3083}
3084
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003085static int cache_preresume(struct dm_target *ti)
3086{
3087 int r = 0;
3088 struct cache *cache = ti->private;
Joe Thornberf494a9c2013-10-31 13:55:49 -04003089 dm_cblock_t csize = get_cache_dev_size(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003090
3091 /*
3092 * Check to see if the cache has resized.
3093 */
Joe Thornberf494a9c2013-10-31 13:55:49 -04003094 if (!cache->sized) {
3095 r = resize_cache_dev(cache, csize);
3096 if (r)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003097 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003098
3099 cache->sized = true;
Joe Thornberf494a9c2013-10-31 13:55:49 -04003100
3101 } else if (csize != cache->cache_size) {
3102 if (!can_resize(cache, csize))
3103 return -EINVAL;
3104
3105 r = resize_cache_dev(cache, csize);
3106 if (r)
3107 return r;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003108 }
3109
3110 if (!cache->loaded_mappings) {
Mike Snitzerea2dd8c2013-03-20 17:21:28 +00003111 r = dm_cache_load_mappings(cache->cmd, cache->policy,
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003112 load_mapping, cache);
3113 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003114 DMERR("%s: could not load cache mappings", cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -04003115 metadata_operation_failed(cache, "dm_cache_load_mappings", r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003116 return r;
3117 }
3118
3119 cache->loaded_mappings = true;
3120 }
3121
3122 if (!cache->loaded_discards) {
Joe Thornber3e2e1c32014-11-24 14:06:22 +00003123 struct discard_load_info li;
3124
3125 /*
3126 * The discard bitset could have been resized, or the
3127 * discard block size changed. To be safe we start by
3128 * setting every dblock to not discarded.
3129 */
3130 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3131
3132 discard_load_info_init(cache, &li);
3133 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003134 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003135 DMERR("%s: could not load origin discards", cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -04003136 metadata_operation_failed(cache, "dm_cache_load_discards", r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003137 return r;
3138 }
Joe Thornber3e2e1c32014-11-24 14:06:22 +00003139 set_discard_range(&li);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003140
3141 cache->loaded_discards = true;
3142 }
3143
3144 return r;
3145}
3146
3147static void cache_resume(struct dm_target *ti)
3148{
3149 struct cache *cache = ti->private;
3150
3151 cache->need_tick_bio = true;
Joe Thornberb29d4982016-12-15 04:57:31 -05003152 allow_background_work(cache);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003153 do_waker(&cache->waker.work);
3154}
3155
3156/*
3157 * Status format:
3158 *
Mike Snitzer6a388612014-01-09 16:04:12 -05003159 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3160 * <cache block size> <#used cache blocks>/<#total cache blocks>
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003161 * <#read hits> <#read misses> <#write hits> <#write misses>
Mike Snitzer6a388612014-01-09 16:04:12 -05003162 * <#demotions> <#promotions> <#dirty>
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003163 * <#features> <features>*
3164 * <#core args> <core args>
Mike Snitzer255eac22015-07-15 11:42:59 -04003165 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003166 */
3167static void cache_status(struct dm_target *ti, status_type_t type,
3168 unsigned status_flags, char *result, unsigned maxlen)
3169{
3170 int r = 0;
3171 unsigned i;
3172 ssize_t sz = 0;
3173 dm_block_t nr_free_blocks_metadata = 0;
3174 dm_block_t nr_blocks_metadata = 0;
3175 char buf[BDEVNAME_SIZE];
3176 struct cache *cache = ti->private;
3177 dm_cblock_t residency;
Joe Thornberd14fcf32016-03-10 16:20:58 +00003178 bool needs_check;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003179
3180 switch (type) {
3181 case STATUSTYPE_INFO:
Joe Thornber028ae9f2015-04-22 16:42:35 -04003182 if (get_cache_mode(cache) == CM_FAIL) {
3183 DMEMIT("Fail");
3184 break;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003185 }
3186
Joe Thornber028ae9f2015-04-22 16:42:35 -04003187 /* Commit to ensure statistics aren't out-of-date */
3188 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3189 (void) commit(cache, false);
3190
Mike Snitzerb61d9502015-04-22 17:25:56 -04003191 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003192 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003193 DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3194 cache_device_name(cache), r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003195 goto err;
3196 }
3197
3198 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3199 if (r) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003200 DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3201 cache_device_name(cache), r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003202 goto err;
3203 }
3204
3205 residency = policy_residency(cache->policy);
3206
Joe Thornberca763d02017-02-09 11:46:18 -05003207 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
Mike Snitzer895b47d2014-07-14 15:37:18 -04003208 (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003209 (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3210 (unsigned long long)nr_blocks_metadata,
Joe Thornberca763d02017-02-09 11:46:18 -05003211 (unsigned long long)cache->sectors_per_block,
Mike Snitzer6a388612014-01-09 16:04:12 -05003212 (unsigned long long) from_cblock(residency),
3213 (unsigned long long) from_cblock(cache->cache_size),
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003214 (unsigned) atomic_read(&cache->stats.read_hit),
3215 (unsigned) atomic_read(&cache->stats.read_miss),
3216 (unsigned) atomic_read(&cache->stats.write_hit),
3217 (unsigned) atomic_read(&cache->stats.write_miss),
3218 (unsigned) atomic_read(&cache->stats.demotion),
3219 (unsigned) atomic_read(&cache->stats.promotion),
Anssi Hannula44fa8162014-08-01 11:55:47 -04003220 (unsigned long) atomic_read(&cache->nr_dirty));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003221
Joe Thornber629d0a82016-09-22 06:15:21 -04003222 if (cache->features.metadata_version == 2)
3223 DMEMIT("2 metadata2 ");
3224 else
3225 DMEMIT("1 ");
3226
Mike Snitzer755b4e82017-10-19 21:01:04 -04003227 if (writethrough_mode(cache))
Joe Thornber629d0a82016-09-22 06:15:21 -04003228 DMEMIT("writethrough ");
Joe Thornber2ee57d52013-10-24 14:10:29 -04003229
Mike Snitzer755b4e82017-10-19 21:01:04 -04003230 else if (passthrough_mode(cache))
Joe Thornber629d0a82016-09-22 06:15:21 -04003231 DMEMIT("passthrough ");
Joe Thornber2ee57d52013-10-24 14:10:29 -04003232
Mike Snitzer755b4e82017-10-19 21:01:04 -04003233 else if (writeback_mode(cache))
Joe Thornber629d0a82016-09-22 06:15:21 -04003234 DMEMIT("writeback ");
Joe Thornber2ee57d52013-10-24 14:10:29 -04003235
3236 else {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003237 DMERR("%s: internal error: unknown io mode: %d",
3238 cache_device_name(cache), (int) cache->features.io_mode);
Joe Thornber2ee57d52013-10-24 14:10:29 -04003239 goto err;
3240 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003241
3242 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
Mike Snitzer2e68c4e2014-01-15 21:06:55 -05003243
3244 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003245 if (sz < maxlen) {
Joe Thornber028ae9f2015-04-22 16:42:35 -04003246 r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003247 if (r)
Mike Snitzerb61d9502015-04-22 17:25:56 -04003248 DMERR("%s: policy_emit_config_values returned %d",
3249 cache_device_name(cache), r);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003250 }
3251
Joe Thornber028ae9f2015-04-22 16:42:35 -04003252 if (get_cache_mode(cache) == CM_READ_ONLY)
3253 DMEMIT("ro ");
3254 else
3255 DMEMIT("rw ");
3256
Joe Thornberd14fcf32016-03-10 16:20:58 +00003257 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3258
3259 if (r || needs_check)
Mike Snitzer255eac22015-07-15 11:42:59 -04003260 DMEMIT("needs_check ");
3261 else
3262 DMEMIT("- ");
3263
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003264 break;
3265
3266 case STATUSTYPE_TABLE:
3267 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3268 DMEMIT("%s ", buf);
3269 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3270 DMEMIT("%s ", buf);
3271 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3272 DMEMIT("%s", buf);
3273
3274 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3275 DMEMIT(" %s", cache->ctr_args[i]);
3276 if (cache->nr_ctr_args)
3277 DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3278 }
3279
3280 return;
3281
3282err:
3283 DMEMIT("Error");
3284}
3285
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003286/*
Joe Thornberb29d4982016-12-15 04:57:31 -05003287 * Defines a range of cblocks, begin to (end - 1) are in the range. end is
3288 * the one-past-the-end value.
3289 */
3290struct cblock_range {
3291 dm_cblock_t begin;
3292 dm_cblock_t end;
3293};
3294
3295/*
Joe Thornber65790ff2013-11-08 16:39:50 +00003296 * A cache block range can take two forms:
3297 *
3298 * i) A single cblock, eg. '3456'
Joe Thornberb29d4982016-12-15 04:57:31 -05003299 * ii) A begin and end cblock with a dash between, eg. 123-234
Joe Thornber65790ff2013-11-08 16:39:50 +00003300 */
3301static int parse_cblock_range(struct cache *cache, const char *str,
3302 struct cblock_range *result)
3303{
3304 char dummy;
3305 uint64_t b, e;
3306 int r;
3307
3308 /*
3309 * Try and parse form (ii) first.
3310 */
3311 r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3312 if (r < 0)
3313 return r;
3314
3315 if (r == 2) {
3316 result->begin = to_cblock(b);
3317 result->end = to_cblock(e);
3318 return 0;
3319 }
3320
3321 /*
3322 * That didn't work, try form (i).
3323 */
3324 r = sscanf(str, "%llu%c", &b, &dummy);
3325 if (r < 0)
3326 return r;
3327
3328 if (r == 1) {
3329 result->begin = to_cblock(b);
3330 result->end = to_cblock(from_cblock(result->begin) + 1u);
3331 return 0;
3332 }
3333
Mike Snitzerb61d9502015-04-22 17:25:56 -04003334 DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
Joe Thornber65790ff2013-11-08 16:39:50 +00003335 return -EINVAL;
3336}
3337
3338static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3339{
3340 uint64_t b = from_cblock(range->begin);
3341 uint64_t e = from_cblock(range->end);
3342 uint64_t n = from_cblock(cache->cache_size);
3343
3344 if (b >= n) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003345 DMERR("%s: begin cblock out of range: %llu >= %llu",
3346 cache_device_name(cache), b, n);
Joe Thornber65790ff2013-11-08 16:39:50 +00003347 return -EINVAL;
3348 }
3349
3350 if (e > n) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003351 DMERR("%s: end cblock out of range: %llu > %llu",
3352 cache_device_name(cache), e, n);
Joe Thornber65790ff2013-11-08 16:39:50 +00003353 return -EINVAL;
3354 }
3355
3356 if (b >= e) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003357 DMERR("%s: invalid cblock range: %llu >= %llu",
3358 cache_device_name(cache), b, e);
Joe Thornber65790ff2013-11-08 16:39:50 +00003359 return -EINVAL;
3360 }
3361
3362 return 0;
3363}
3364
Joe Thornberb29d4982016-12-15 04:57:31 -05003365static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3366{
3367 return to_cblock(from_cblock(b) + 1);
3368}
3369
Joe Thornber65790ff2013-11-08 16:39:50 +00003370static int request_invalidation(struct cache *cache, struct cblock_range *range)
3371{
Joe Thornberb29d4982016-12-15 04:57:31 -05003372 int r = 0;
Joe Thornber65790ff2013-11-08 16:39:50 +00003373
Joe Thornberb29d4982016-12-15 04:57:31 -05003374 /*
3375 * We don't need to do any locking here because we know we're in
3376 * passthrough mode. There's is potential for a race between an
3377 * invalidation triggered by an io and an invalidation message. This
3378 * is harmless, we must not worry if the policy call fails.
3379 */
3380 while (range->begin != range->end) {
3381 r = invalidate_cblock(cache, range->begin);
3382 if (r)
3383 return r;
Joe Thornber65790ff2013-11-08 16:39:50 +00003384
Joe Thornberb29d4982016-12-15 04:57:31 -05003385 range->begin = cblock_succ(range->begin);
3386 }
Joe Thornber65790ff2013-11-08 16:39:50 +00003387
Joe Thornberb29d4982016-12-15 04:57:31 -05003388 cache->commit_requested = true;
3389 return r;
Joe Thornber65790ff2013-11-08 16:39:50 +00003390}
3391
3392static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3393 const char **cblock_ranges)
3394{
3395 int r = 0;
3396 unsigned i;
3397 struct cblock_range range;
3398
Mike Snitzer755b4e82017-10-19 21:01:04 -04003399 if (!passthrough_mode(cache)) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003400 DMERR("%s: cache has to be in passthrough mode for invalidation",
3401 cache_device_name(cache));
Joe Thornber65790ff2013-11-08 16:39:50 +00003402 return -EPERM;
3403 }
3404
3405 for (i = 0; i < count; i++) {
3406 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3407 if (r)
3408 break;
3409
3410 r = validate_cblock_range(cache, &range);
3411 if (r)
3412 break;
3413
3414 /*
3415 * Pass begin and end origin blocks to the worker and wake it.
3416 */
3417 r = request_invalidation(cache, &range);
3418 if (r)
3419 break;
3420 }
3421
3422 return r;
3423}
3424
3425/*
3426 * Supports
3427 * "<key> <value>"
3428 * and
3429 * "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003430 *
3431 * The key migration_threshold is supported by the cache target core.
3432 */
3433static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3434{
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003435 struct cache *cache = ti->private;
3436
Joe Thornber65790ff2013-11-08 16:39:50 +00003437 if (!argc)
3438 return -EINVAL;
3439
Joe Thornber028ae9f2015-04-22 16:42:35 -04003440 if (get_cache_mode(cache) >= CM_READ_ONLY) {
Mike Snitzerb61d9502015-04-22 17:25:56 -04003441 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3442 cache_device_name(cache));
Joe Thornber028ae9f2015-04-22 16:42:35 -04003443 return -EOPNOTSUPP;
3444 }
3445
Mike Snitzer7b6b2bc2013-11-12 12:17:43 -05003446 if (!strcasecmp(argv[0], "invalidate_cblocks"))
Joe Thornber65790ff2013-11-08 16:39:50 +00003447 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3448
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003449 if (argc != 2)
3450 return -EINVAL;
3451
Joe Thornber2f14f4b2013-05-10 14:37:21 +01003452 return set_config_value(cache, argv[0], argv[1]);
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003453}
3454
3455static int cache_iterate_devices(struct dm_target *ti,
3456 iterate_devices_callout_fn fn, void *data)
3457{
3458 int r = 0;
3459 struct cache *cache = ti->private;
3460
3461 r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3462 if (!r)
3463 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3464
3465 return r;
3466}
3467
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003468static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3469{
3470 /*
3471 * FIXME: these limits may be incompatible with the cache device
3472 */
Joe Thornber7ae34e72014-11-06 10:18:04 +00003473 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3474 cache->origin_sectors);
Joe Thornber1bad9bc2014-11-07 14:47:07 +00003475 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003476}
3477
3478static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3479{
3480 struct cache *cache = ti->private;
Mike Snitzerf6109372013-08-20 15:02:41 -04003481 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003482
Mike Snitzerf6109372013-08-20 15:02:41 -04003483 /*
3484 * If the system-determined stacked limits are compatible with the
3485 * cache's blocksize (io_opt is a factor) do not override them.
3486 */
3487 if (io_opt_sectors < cache->sectors_per_block ||
3488 do_div(io_opt_sectors, cache->sectors_per_block)) {
Mike Snitzerb0246532014-07-19 13:25:46 -04003489 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
Mike Snitzerf6109372013-08-20 15:02:41 -04003490 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3491 }
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003492 set_discard_limits(cache, limits);
3493}
3494
3495/*----------------------------------------------------------------*/
3496
3497static struct target_type cache_target = {
3498 .name = "cache",
Joe Thornberb29d4982016-12-15 04:57:31 -05003499 .version = {2, 0, 0},
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003500 .module = THIS_MODULE,
3501 .ctr = cache_ctr,
3502 .dtr = cache_dtr,
3503 .map = cache_map,
3504 .end_io = cache_end_io,
3505 .postsuspend = cache_postsuspend,
3506 .preresume = cache_preresume,
3507 .resume = cache_resume,
3508 .status = cache_status,
3509 .message = cache_message,
3510 .iterate_devices = cache_iterate_devices,
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003511 .io_hints = cache_io_hints,
3512};
3513
3514static int __init dm_cache_init(void)
3515{
3516 int r;
3517
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003518 migration_cache = KMEM_CACHE(dm_cache_migration, 0);
Shenghui Wang6c8faa12018-10-07 14:45:41 +08003519 if (!migration_cache)
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003520 return -ENOMEM;
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003521
monty_pavel@sina.comfbce4292017-11-25 01:43:50 +08003522 r = dm_register_target(&cache_target);
3523 if (r) {
3524 DMERR("cache target registration failed: %d", r);
Shenghui Wang6c8faa12018-10-07 14:45:41 +08003525 kmem_cache_destroy(migration_cache);
monty_pavel@sina.comfbce4292017-11-25 01:43:50 +08003526 return r;
3527 }
3528
Joe Thornberc6b4fcb2013-03-01 22:45:51 +00003529 return 0;
3530}
3531
3532static void __exit dm_cache_exit(void)
3533{
3534 dm_unregister_target(&cache_target);
3535 kmem_cache_destroy(migration_cache);
3536}
3537
3538module_init(dm_cache_init);
3539module_exit(dm_cache_exit);
3540
3541MODULE_DESCRIPTION(DM_NAME " cache target");
3542MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3543MODULE_LICENSE("GPL");