Linux Audio

Check our new training course

Loading...
v6.2
 
   1/*
   2 * Copyright (C) 2012 Red Hat. All rights reserved.
   3 *
   4 * This file is released under the GPL.
   5 */
   6
   7#include "dm.h"
   8#include "dm-bio-prison-v2.h"
   9#include "dm-bio-record.h"
  10#include "dm-cache-metadata.h"
  11#include "dm-io-tracker.h"
  12
  13#include <linux/dm-io.h>
  14#include <linux/dm-kcopyd.h>
  15#include <linux/jiffies.h>
  16#include <linux/init.h>
  17#include <linux/mempool.h>
  18#include <linux/module.h>
  19#include <linux/rwsem.h>
  20#include <linux/slab.h>
  21#include <linux/vmalloc.h>
  22
  23#define DM_MSG_PREFIX "cache"
  24
  25DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  26	"A percentage of time allocated for copying to and/or from cache");
  27
  28/*----------------------------------------------------------------*/
  29
  30/*
  31 * Glossary:
  32 *
  33 * oblock: index of an origin block
  34 * cblock: index of a cache block
  35 * promotion: movement of a block from origin to cache
  36 * demotion: movement of a block from cache to origin
  37 * migration: movement of a block between the origin and cache device,
  38 *	      either direction
  39 */
  40
  41/*----------------------------------------------------------------*/
  42
  43/*
  44 * Represents a chunk of future work.  'input' allows continuations to pass
  45 * values between themselves, typically error values.
  46 */
  47struct continuation {
  48	struct work_struct ws;
  49	blk_status_t input;
  50};
  51
  52static inline void init_continuation(struct continuation *k,
  53				     void (*fn)(struct work_struct *))
  54{
  55	INIT_WORK(&k->ws, fn);
  56	k->input = 0;
  57}
  58
  59static inline void queue_continuation(struct workqueue_struct *wq,
  60				      struct continuation *k)
  61{
  62	queue_work(wq, &k->ws);
  63}
  64
  65/*----------------------------------------------------------------*/
  66
  67/*
  68 * The batcher collects together pieces of work that need a particular
  69 * operation to occur before they can proceed (typically a commit).
  70 */
  71struct batcher {
  72	/*
  73	 * The operation that everyone is waiting for.
  74	 */
  75	blk_status_t (*commit_op)(void *context);
  76	void *commit_context;
  77
  78	/*
  79	 * This is how bios should be issued once the commit op is complete
  80	 * (accounted_request).
  81	 */
  82	void (*issue_op)(struct bio *bio, void *context);
  83	void *issue_context;
  84
  85	/*
  86	 * Queued work gets put on here after commit.
  87	 */
  88	struct workqueue_struct *wq;
  89
  90	spinlock_t lock;
  91	struct list_head work_items;
  92	struct bio_list bios;
  93	struct work_struct commit_work;
  94
  95	bool commit_scheduled;
  96};
  97
  98static void __commit(struct work_struct *_ws)
  99{
 100	struct batcher *b = container_of(_ws, struct batcher, commit_work);
 101	blk_status_t r;
 102	struct list_head work_items;
 103	struct work_struct *ws, *tmp;
 104	struct continuation *k;
 105	struct bio *bio;
 106	struct bio_list bios;
 107
 108	INIT_LIST_HEAD(&work_items);
 109	bio_list_init(&bios);
 110
 111	/*
 112	 * We have to grab these before the commit_op to avoid a race
 113	 * condition.
 114	 */
 115	spin_lock_irq(&b->lock);
 116	list_splice_init(&b->work_items, &work_items);
 117	bio_list_merge(&bios, &b->bios);
 118	bio_list_init(&b->bios);
 119	b->commit_scheduled = false;
 120	spin_unlock_irq(&b->lock);
 121
 122	r = b->commit_op(b->commit_context);
 123
 124	list_for_each_entry_safe(ws, tmp, &work_items, entry) {
 125		k = container_of(ws, struct continuation, ws);
 126		k->input = r;
 127		INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
 128		queue_work(b->wq, ws);
 129	}
 130
 131	while ((bio = bio_list_pop(&bios))) {
 132		if (r) {
 133			bio->bi_status = r;
 134			bio_endio(bio);
 135		} else
 136			b->issue_op(bio, b->issue_context);
 137	}
 138}
 139
 140static void batcher_init(struct batcher *b,
 141			 blk_status_t (*commit_op)(void *),
 142			 void *commit_context,
 143			 void (*issue_op)(struct bio *bio, void *),
 144			 void *issue_context,
 145			 struct workqueue_struct *wq)
 146{
 147	b->commit_op = commit_op;
 148	b->commit_context = commit_context;
 149	b->issue_op = issue_op;
 150	b->issue_context = issue_context;
 151	b->wq = wq;
 152
 153	spin_lock_init(&b->lock);
 154	INIT_LIST_HEAD(&b->work_items);
 155	bio_list_init(&b->bios);
 156	INIT_WORK(&b->commit_work, __commit);
 157	b->commit_scheduled = false;
 158}
 159
 160static void async_commit(struct batcher *b)
 161{
 162	queue_work(b->wq, &b->commit_work);
 163}
 164
 165static void continue_after_commit(struct batcher *b, struct continuation *k)
 166{
 167	bool commit_scheduled;
 168
 169	spin_lock_irq(&b->lock);
 170	commit_scheduled = b->commit_scheduled;
 171	list_add_tail(&k->ws.entry, &b->work_items);
 172	spin_unlock_irq(&b->lock);
 173
 174	if (commit_scheduled)
 175		async_commit(b);
 176}
 177
 178/*
 179 * Bios are errored if commit failed.
 180 */
 181static void issue_after_commit(struct batcher *b, struct bio *bio)
 182{
 183       bool commit_scheduled;
 184
 185       spin_lock_irq(&b->lock);
 186       commit_scheduled = b->commit_scheduled;
 187       bio_list_add(&b->bios, bio);
 188       spin_unlock_irq(&b->lock);
 189
 190       if (commit_scheduled)
 191	       async_commit(b);
 192}
 193
 194/*
 195 * Call this if some urgent work is waiting for the commit to complete.
 196 */
 197static void schedule_commit(struct batcher *b)
 198{
 199	bool immediate;
 200
 201	spin_lock_irq(&b->lock);
 202	immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
 203	b->commit_scheduled = true;
 204	spin_unlock_irq(&b->lock);
 205
 206	if (immediate)
 207		async_commit(b);
 208}
 209
 210/*
 211 * There are a couple of places where we let a bio run, but want to do some
 212 * work before calling its endio function.  We do this by temporarily
 213 * changing the endio fn.
 214 */
 215struct dm_hook_info {
 216	bio_end_io_t *bi_end_io;
 217};
 218
 219static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 220			bio_end_io_t *bi_end_io, void *bi_private)
 221{
 222	h->bi_end_io = bio->bi_end_io;
 223
 224	bio->bi_end_io = bi_end_io;
 225	bio->bi_private = bi_private;
 226}
 227
 228static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 229{
 230	bio->bi_end_io = h->bi_end_io;
 231}
 232
 233/*----------------------------------------------------------------*/
 234
 235#define MIGRATION_POOL_SIZE 128
 236#define COMMIT_PERIOD HZ
 237#define MIGRATION_COUNT_WINDOW 10
 238
 239/*
 240 * The block size of the device holding cache data must be
 241 * between 32KB and 1GB.
 242 */
 243#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 244#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 245
 246enum cache_metadata_mode {
 247	CM_WRITE,		/* metadata may be changed */
 248	CM_READ_ONLY,		/* metadata may not be changed */
 249	CM_FAIL
 250};
 251
 252enum cache_io_mode {
 253	/*
 254	 * Data is written to cached blocks only.  These blocks are marked
 255	 * dirty.  If you lose the cache device you will lose data.
 256	 * Potential performance increase for both reads and writes.
 257	 */
 258	CM_IO_WRITEBACK,
 259
 260	/*
 261	 * Data is written to both cache and origin.  Blocks are never
 262	 * dirty.  Potential performance benfit for reads only.
 263	 */
 264	CM_IO_WRITETHROUGH,
 265
 266	/*
 267	 * A degraded mode useful for various cache coherency situations
 268	 * (eg, rolling back snapshots).  Reads and writes always go to the
 269	 * origin.  If a write goes to a cached oblock, then the cache
 270	 * block is invalidated.
 271	 */
 272	CM_IO_PASSTHROUGH
 273};
 274
 275struct cache_features {
 276	enum cache_metadata_mode mode;
 277	enum cache_io_mode io_mode;
 278	unsigned metadata_version;
 279	bool discard_passdown:1;
 280};
 281
 282struct cache_stats {
 283	atomic_t read_hit;
 284	atomic_t read_miss;
 285	atomic_t write_hit;
 286	atomic_t write_miss;
 287	atomic_t demotion;
 288	atomic_t promotion;
 289	atomic_t writeback;
 290	atomic_t copies_avoided;
 291	atomic_t cache_cell_clash;
 292	atomic_t commit_count;
 293	atomic_t discard_count;
 294};
 295
 296struct cache {
 297	struct dm_target *ti;
 298	spinlock_t lock;
 299
 300	/*
 301	 * Fields for converting from sectors to blocks.
 302	 */
 303	int sectors_per_block_shift;
 304	sector_t sectors_per_block;
 305
 306	struct dm_cache_metadata *cmd;
 307
 308	/*
 309	 * Metadata is written to this device.
 310	 */
 311	struct dm_dev *metadata_dev;
 312
 313	/*
 314	 * The slower of the two data devices.  Typically a spindle.
 315	 */
 316	struct dm_dev *origin_dev;
 317
 318	/*
 319	 * The faster of the two data devices.  Typically an SSD.
 320	 */
 321	struct dm_dev *cache_dev;
 322
 323	/*
 324	 * Size of the origin device in _complete_ blocks and native sectors.
 325	 */
 326	dm_oblock_t origin_blocks;
 327	sector_t origin_sectors;
 328
 329	/*
 330	 * Size of the cache device in blocks.
 331	 */
 332	dm_cblock_t cache_size;
 333
 334	/*
 335	 * Invalidation fields.
 336	 */
 337	spinlock_t invalidation_lock;
 338	struct list_head invalidation_requests;
 339
 340	sector_t migration_threshold;
 341	wait_queue_head_t migration_wait;
 342	atomic_t nr_allocated_migrations;
 343
 344	/*
 345	 * The number of in flight migrations that are performing
 346	 * background io. eg, promotion, writeback.
 347	 */
 348	atomic_t nr_io_migrations;
 349
 350	struct bio_list deferred_bios;
 351
 352	struct rw_semaphore quiesce_lock;
 353
 354	/*
 355	 * origin_blocks entries, discarded if set.
 356	 */
 357	dm_dblock_t discard_nr_blocks;
 358	unsigned long *discard_bitset;
 359	uint32_t discard_block_size; /* a power of 2 times sectors per block */
 360
 361	/*
 362	 * Rather than reconstructing the table line for the status we just
 363	 * save it and regurgitate.
 364	 */
 365	unsigned nr_ctr_args;
 366	const char **ctr_args;
 367
 368	struct dm_kcopyd_client *copier;
 369	struct work_struct deferred_bio_worker;
 370	struct work_struct migration_worker;
 371	struct workqueue_struct *wq;
 372	struct delayed_work waker;
 373	struct dm_bio_prison_v2 *prison;
 374
 375	/*
 376	 * cache_size entries, dirty if set
 377	 */
 378	unsigned long *dirty_bitset;
 379	atomic_t nr_dirty;
 380
 381	unsigned policy_nr_args;
 382	struct dm_cache_policy *policy;
 383
 384	/*
 385	 * Cache features such as write-through.
 386	 */
 387	struct cache_features features;
 388
 389	struct cache_stats stats;
 390
 391	bool need_tick_bio:1;
 392	bool sized:1;
 393	bool invalidate:1;
 394	bool commit_requested:1;
 395	bool loaded_mappings:1;
 396	bool loaded_discards:1;
 397
 398	struct rw_semaphore background_work_lock;
 399
 400	struct batcher committer;
 401	struct work_struct commit_ws;
 402
 403	struct dm_io_tracker tracker;
 404
 405	mempool_t migration_pool;
 406
 407	struct bio_set bs;
 408};
 409
 410struct per_bio_data {
 411	bool tick:1;
 412	unsigned req_nr:2;
 413	struct dm_bio_prison_cell_v2 *cell;
 414	struct dm_hook_info hook_info;
 415	sector_t len;
 416};
 417
 418struct dm_cache_migration {
 419	struct continuation k;
 420	struct cache *cache;
 421
 422	struct policy_work *op;
 423	struct bio *overwrite_bio;
 424	struct dm_bio_prison_cell_v2 *cell;
 425
 426	dm_cblock_t invalidate_cblock;
 427	dm_oblock_t invalidate_oblock;
 428};
 429
 430/*----------------------------------------------------------------*/
 431
 432static bool writethrough_mode(struct cache *cache)
 433{
 434	return cache->features.io_mode == CM_IO_WRITETHROUGH;
 435}
 436
 437static bool writeback_mode(struct cache *cache)
 438{
 439	return cache->features.io_mode == CM_IO_WRITEBACK;
 440}
 441
 442static inline bool passthrough_mode(struct cache *cache)
 443{
 444	return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
 445}
 446
 447/*----------------------------------------------------------------*/
 448
 449static void wake_deferred_bio_worker(struct cache *cache)
 450{
 451	queue_work(cache->wq, &cache->deferred_bio_worker);
 452}
 453
 454static void wake_migration_worker(struct cache *cache)
 455{
 456	if (passthrough_mode(cache))
 457		return;
 458
 459	queue_work(cache->wq, &cache->migration_worker);
 460}
 461
 462/*----------------------------------------------------------------*/
 463
 464static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
 465{
 466	return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
 467}
 468
 469static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
 470{
 471	dm_bio_prison_free_cell_v2(cache->prison, cell);
 472}
 473
 474static struct dm_cache_migration *alloc_migration(struct cache *cache)
 475{
 476	struct dm_cache_migration *mg;
 477
 478	mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
 479
 480	memset(mg, 0, sizeof(*mg));
 481
 482	mg->cache = cache;
 483	atomic_inc(&cache->nr_allocated_migrations);
 484
 485	return mg;
 486}
 487
 488static void free_migration(struct dm_cache_migration *mg)
 489{
 490	struct cache *cache = mg->cache;
 491
 492	if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 493		wake_up(&cache->migration_wait);
 494
 495	mempool_free(mg, &cache->migration_pool);
 496}
 497
 498/*----------------------------------------------------------------*/
 499
 500static inline dm_oblock_t oblock_succ(dm_oblock_t b)
 501{
 502	return to_oblock(from_oblock(b) + 1ull);
 503}
 504
 505static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
 506{
 507	key->virtual = 0;
 508	key->dev = 0;
 509	key->block_begin = from_oblock(begin);
 510	key->block_end = from_oblock(end);
 511}
 512
 513/*
 514 * We have two lock levels.  Level 0, which is used to prevent WRITEs, and
 515 * level 1 which prevents *both* READs and WRITEs.
 516 */
 517#define WRITE_LOCK_LEVEL 0
 518#define READ_WRITE_LOCK_LEVEL 1
 519
 520static unsigned lock_level(struct bio *bio)
 521{
 522	return bio_data_dir(bio) == WRITE ?
 523		WRITE_LOCK_LEVEL :
 524		READ_WRITE_LOCK_LEVEL;
 525}
 526
 527/*----------------------------------------------------------------
 
 528 * Per bio data
 529 *--------------------------------------------------------------*/
 
 530
 531static struct per_bio_data *get_per_bio_data(struct bio *bio)
 532{
 533	struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 
 534	BUG_ON(!pb);
 535	return pb;
 536}
 537
 538static struct per_bio_data *init_per_bio_data(struct bio *bio)
 539{
 540	struct per_bio_data *pb = get_per_bio_data(bio);
 541
 542	pb->tick = false;
 543	pb->req_nr = dm_bio_get_target_bio_nr(bio);
 544	pb->cell = NULL;
 545	pb->len = 0;
 546
 547	return pb;
 548}
 549
 550/*----------------------------------------------------------------*/
 551
 552static void defer_bio(struct cache *cache, struct bio *bio)
 553{
 554	spin_lock_irq(&cache->lock);
 555	bio_list_add(&cache->deferred_bios, bio);
 556	spin_unlock_irq(&cache->lock);
 557
 558	wake_deferred_bio_worker(cache);
 559}
 560
 561static void defer_bios(struct cache *cache, struct bio_list *bios)
 562{
 563	spin_lock_irq(&cache->lock);
 564	bio_list_merge(&cache->deferred_bios, bios);
 565	bio_list_init(bios);
 566	spin_unlock_irq(&cache->lock);
 567
 568	wake_deferred_bio_worker(cache);
 569}
 570
 571/*----------------------------------------------------------------*/
 572
 573static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
 574{
 575	bool r;
 576	struct per_bio_data *pb;
 577	struct dm_cell_key_v2 key;
 578	dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 579	struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
 580
 581	cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
 582
 583	build_key(oblock, end, &key);
 584	r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
 585	if (!r) {
 586		/*
 587		 * Failed to get the lock.
 588		 */
 589		free_prison_cell(cache, cell_prealloc);
 590		return r;
 591	}
 592
 593	if (cell != cell_prealloc)
 594		free_prison_cell(cache, cell_prealloc);
 595
 596	pb = get_per_bio_data(bio);
 597	pb->cell = cell;
 598
 599	return r;
 600}
 601
 602/*----------------------------------------------------------------*/
 603
 604static bool is_dirty(struct cache *cache, dm_cblock_t b)
 605{
 606	return test_bit(from_cblock(b), cache->dirty_bitset);
 607}
 608
 609static void set_dirty(struct cache *cache, dm_cblock_t cblock)
 610{
 611	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 612		atomic_inc(&cache->nr_dirty);
 613		policy_set_dirty(cache->policy, cblock);
 614	}
 615}
 616
 617/*
 618 * These two are called when setting after migrations to force the policy
 619 * and dirty bitset to be in sync.
 620 */
 621static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
 622{
 623	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
 624		atomic_inc(&cache->nr_dirty);
 625	policy_set_dirty(cache->policy, cblock);
 626}
 627
 628static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
 629{
 630	if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 631		if (atomic_dec_return(&cache->nr_dirty) == 0)
 632			dm_table_event(cache->ti->table);
 633	}
 634
 635	policy_clear_dirty(cache->policy, cblock);
 636}
 637
 638/*----------------------------------------------------------------*/
 639
 640static bool block_size_is_power_of_two(struct cache *cache)
 641{
 642	return cache->sectors_per_block_shift >= 0;
 643}
 644
 645static dm_block_t block_div(dm_block_t b, uint32_t n)
 646{
 647	do_div(b, n);
 648
 649	return b;
 650}
 651
 652static dm_block_t oblocks_per_dblock(struct cache *cache)
 653{
 654	dm_block_t oblocks = cache->discard_block_size;
 655
 656	if (block_size_is_power_of_two(cache))
 657		oblocks >>= cache->sectors_per_block_shift;
 658	else
 659		oblocks = block_div(oblocks, cache->sectors_per_block);
 660
 661	return oblocks;
 662}
 663
 664static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 665{
 666	return to_dblock(block_div(from_oblock(oblock),
 667				   oblocks_per_dblock(cache)));
 668}
 669
 670static void set_discard(struct cache *cache, dm_dblock_t b)
 671{
 672	BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 673	atomic_inc(&cache->stats.discard_count);
 674
 675	spin_lock_irq(&cache->lock);
 676	set_bit(from_dblock(b), cache->discard_bitset);
 677	spin_unlock_irq(&cache->lock);
 678}
 679
 680static void clear_discard(struct cache *cache, dm_dblock_t b)
 681{
 682	spin_lock_irq(&cache->lock);
 683	clear_bit(from_dblock(b), cache->discard_bitset);
 684	spin_unlock_irq(&cache->lock);
 685}
 686
 687static bool is_discarded(struct cache *cache, dm_dblock_t b)
 688{
 689	int r;
 
 690	spin_lock_irq(&cache->lock);
 691	r = test_bit(from_dblock(b), cache->discard_bitset);
 692	spin_unlock_irq(&cache->lock);
 693
 694	return r;
 695}
 696
 697static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 698{
 699	int r;
 
 700	spin_lock_irq(&cache->lock);
 701	r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 702		     cache->discard_bitset);
 703	spin_unlock_irq(&cache->lock);
 704
 705	return r;
 706}
 707
 708/*----------------------------------------------------------------
 
 709 * Remapping
 710 *--------------------------------------------------------------*/
 
 711static void remap_to_origin(struct cache *cache, struct bio *bio)
 712{
 713	bio_set_dev(bio, cache->origin_dev->bdev);
 714}
 715
 716static void remap_to_cache(struct cache *cache, struct bio *bio,
 717			   dm_cblock_t cblock)
 718{
 719	sector_t bi_sector = bio->bi_iter.bi_sector;
 720	sector_t block = from_cblock(cblock);
 721
 722	bio_set_dev(bio, cache->cache_dev->bdev);
 723	if (!block_size_is_power_of_two(cache))
 724		bio->bi_iter.bi_sector =
 725			(block * cache->sectors_per_block) +
 726			sector_div(bi_sector, cache->sectors_per_block);
 727	else
 728		bio->bi_iter.bi_sector =
 729			(block << cache->sectors_per_block_shift) |
 730			(bi_sector & (cache->sectors_per_block - 1));
 731}
 732
 733static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 734{
 735	struct per_bio_data *pb;
 736
 737	spin_lock_irq(&cache->lock);
 738	if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
 739	    bio_op(bio) != REQ_OP_DISCARD) {
 740		pb = get_per_bio_data(bio);
 741		pb->tick = true;
 742		cache->need_tick_bio = false;
 743	}
 744	spin_unlock_irq(&cache->lock);
 745}
 746
 747static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 748					  dm_oblock_t oblock)
 749{
 750	// FIXME: check_if_tick_bio_needed() is called way too much through this interface
 751	check_if_tick_bio_needed(cache, bio);
 752	remap_to_origin(cache, bio);
 753	if (bio_data_dir(bio) == WRITE)
 754		clear_discard(cache, oblock_to_dblock(cache, oblock));
 755}
 756
 757static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 758				 dm_oblock_t oblock, dm_cblock_t cblock)
 759{
 760	check_if_tick_bio_needed(cache, bio);
 761	remap_to_cache(cache, bio, cblock);
 762	if (bio_data_dir(bio) == WRITE) {
 763		set_dirty(cache, cblock);
 764		clear_discard(cache, oblock_to_dblock(cache, oblock));
 765	}
 766}
 767
 768static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 769{
 770	sector_t block_nr = bio->bi_iter.bi_sector;
 771
 772	if (!block_size_is_power_of_two(cache))
 773		(void) sector_div(block_nr, cache->sectors_per_block);
 774	else
 775		block_nr >>= cache->sectors_per_block_shift;
 776
 777	return to_oblock(block_nr);
 778}
 779
 780static bool accountable_bio(struct cache *cache, struct bio *bio)
 781{
 782	return bio_op(bio) != REQ_OP_DISCARD;
 783}
 784
 785static void accounted_begin(struct cache *cache, struct bio *bio)
 786{
 787	struct per_bio_data *pb;
 788
 789	if (accountable_bio(cache, bio)) {
 790		pb = get_per_bio_data(bio);
 791		pb->len = bio_sectors(bio);
 792		dm_iot_io_begin(&cache->tracker, pb->len);
 793	}
 794}
 795
 796static void accounted_complete(struct cache *cache, struct bio *bio)
 797{
 798	struct per_bio_data *pb = get_per_bio_data(bio);
 799
 800	dm_iot_io_end(&cache->tracker, pb->len);
 801}
 802
 803static void accounted_request(struct cache *cache, struct bio *bio)
 804{
 805	accounted_begin(cache, bio);
 806	dm_submit_bio_remap(bio, NULL);
 807}
 808
 809static void issue_op(struct bio *bio, void *context)
 810{
 811	struct cache *cache = context;
 
 812	accounted_request(cache, bio);
 813}
 814
 815/*
 816 * When running in writethrough mode we need to send writes to clean blocks
 817 * to both the cache and origin devices.  Clone the bio and send them in parallel.
 818 */
 819static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
 820				      dm_oblock_t oblock, dm_cblock_t cblock)
 821{
 822	struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
 823						 GFP_NOIO, &cache->bs);
 824
 825	BUG_ON(!origin_bio);
 826
 827	bio_chain(origin_bio, bio);
 828
 829	if (bio_data_dir(origin_bio) == WRITE)
 830		clear_discard(cache, oblock_to_dblock(cache, oblock));
 831	submit_bio(origin_bio);
 832
 833	remap_to_cache(cache, bio, cblock);
 834}
 835
 836/*----------------------------------------------------------------
 
 837 * Failure modes
 838 *--------------------------------------------------------------*/
 
 839static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 840{
 841	return cache->features.mode;
 842}
 843
 844static const char *cache_device_name(struct cache *cache)
 845{
 846	return dm_table_device_name(cache->ti->table);
 847}
 848
 849static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 850{
 851	const char *descs[] = {
 852		"write",
 853		"read-only",
 854		"fail"
 855	};
 856
 857	dm_table_event(cache->ti->table);
 858	DMINFO("%s: switching cache to %s mode",
 859	       cache_device_name(cache), descs[(int)mode]);
 860}
 861
 862static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 863{
 864	bool needs_check;
 865	enum cache_metadata_mode old_mode = get_cache_mode(cache);
 866
 867	if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
 868		DMERR("%s: unable to read needs_check flag, setting failure mode.",
 869		      cache_device_name(cache));
 870		new_mode = CM_FAIL;
 871	}
 872
 873	if (new_mode == CM_WRITE && needs_check) {
 874		DMERR("%s: unable to switch cache to write mode until repaired.",
 875		      cache_device_name(cache));
 876		if (old_mode != new_mode)
 877			new_mode = old_mode;
 878		else
 879			new_mode = CM_READ_ONLY;
 880	}
 881
 882	/* Never move out of fail mode */
 883	if (old_mode == CM_FAIL)
 884		new_mode = CM_FAIL;
 885
 886	switch (new_mode) {
 887	case CM_FAIL:
 888	case CM_READ_ONLY:
 889		dm_cache_metadata_set_read_only(cache->cmd);
 890		break;
 891
 892	case CM_WRITE:
 893		dm_cache_metadata_set_read_write(cache->cmd);
 894		break;
 895	}
 896
 897	cache->features.mode = new_mode;
 898
 899	if (new_mode != old_mode)
 900		notify_mode_switch(cache, new_mode);
 901}
 902
 903static void abort_transaction(struct cache *cache)
 904{
 905	const char *dev_name = cache_device_name(cache);
 906
 907	if (get_cache_mode(cache) >= CM_READ_ONLY)
 908		return;
 909
 910	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
 911	if (dm_cache_metadata_abort(cache->cmd)) {
 912		DMERR("%s: failed to abort metadata transaction", dev_name);
 913		set_cache_mode(cache, CM_FAIL);
 914	}
 915
 916	if (dm_cache_metadata_set_needs_check(cache->cmd)) {
 917		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
 918		set_cache_mode(cache, CM_FAIL);
 919	}
 920}
 921
 922static void metadata_operation_failed(struct cache *cache, const char *op, int r)
 923{
 924	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
 925		    cache_device_name(cache), op, r);
 926	abort_transaction(cache);
 927	set_cache_mode(cache, CM_READ_ONLY);
 928}
 929
 930/*----------------------------------------------------------------*/
 931
 932static void load_stats(struct cache *cache)
 933{
 934	struct dm_cache_statistics stats;
 935
 936	dm_cache_metadata_get_stats(cache->cmd, &stats);
 937	atomic_set(&cache->stats.read_hit, stats.read_hits);
 938	atomic_set(&cache->stats.read_miss, stats.read_misses);
 939	atomic_set(&cache->stats.write_hit, stats.write_hits);
 940	atomic_set(&cache->stats.write_miss, stats.write_misses);
 941}
 942
 943static void save_stats(struct cache *cache)
 944{
 945	struct dm_cache_statistics stats;
 946
 947	if (get_cache_mode(cache) >= CM_READ_ONLY)
 948		return;
 949
 950	stats.read_hits = atomic_read(&cache->stats.read_hit);
 951	stats.read_misses = atomic_read(&cache->stats.read_miss);
 952	stats.write_hits = atomic_read(&cache->stats.write_hit);
 953	stats.write_misses = atomic_read(&cache->stats.write_miss);
 954
 955	dm_cache_metadata_set_stats(cache->cmd, &stats);
 956}
 957
 958static void update_stats(struct cache_stats *stats, enum policy_operation op)
 959{
 960	switch (op) {
 961	case POLICY_PROMOTE:
 962		atomic_inc(&stats->promotion);
 963		break;
 964
 965	case POLICY_DEMOTE:
 966		atomic_inc(&stats->demotion);
 967		break;
 968
 969	case POLICY_WRITEBACK:
 970		atomic_inc(&stats->writeback);
 971		break;
 972	}
 973}
 974
 975/*----------------------------------------------------------------
 
 976 * Migration processing
 977 *
 978 * Migration covers moving data from the origin device to the cache, or
 979 * vice versa.
 980 *--------------------------------------------------------------*/
 981
 982static void inc_io_migrations(struct cache *cache)
 983{
 984	atomic_inc(&cache->nr_io_migrations);
 985}
 986
 987static void dec_io_migrations(struct cache *cache)
 988{
 989	atomic_dec(&cache->nr_io_migrations);
 990}
 991
 992static bool discard_or_flush(struct bio *bio)
 993{
 994	return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
 995}
 996
 997static void calc_discard_block_range(struct cache *cache, struct bio *bio,
 998				     dm_dblock_t *b, dm_dblock_t *e)
 999{
1000	sector_t sb = bio->bi_iter.bi_sector;
1001	sector_t se = bio_end_sector(bio);
1002
1003	*b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1004
1005	if (se - sb < cache->discard_block_size)
1006		*e = *b;
1007	else
1008		*e = to_dblock(block_div(se, cache->discard_block_size));
1009}
1010
1011/*----------------------------------------------------------------*/
1012
1013static void prevent_background_work(struct cache *cache)
1014{
1015	lockdep_off();
1016	down_write(&cache->background_work_lock);
1017	lockdep_on();
1018}
1019
1020static void allow_background_work(struct cache *cache)
1021{
1022	lockdep_off();
1023	up_write(&cache->background_work_lock);
1024	lockdep_on();
1025}
1026
1027static bool background_work_begin(struct cache *cache)
1028{
1029	bool r;
1030
1031	lockdep_off();
1032	r = down_read_trylock(&cache->background_work_lock);
1033	lockdep_on();
1034
1035	return r;
1036}
1037
1038static void background_work_end(struct cache *cache)
1039{
1040	lockdep_off();
1041	up_read(&cache->background_work_lock);
1042	lockdep_on();
1043}
1044
1045/*----------------------------------------------------------------*/
1046
1047static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1048{
1049	return (bio_data_dir(bio) == WRITE) &&
1050		(bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1051}
1052
1053static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1054{
1055	return writeback_mode(cache) &&
1056		(is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1057}
1058
1059static void quiesce(struct dm_cache_migration *mg,
1060		    void (*continuation)(struct work_struct *))
1061{
1062	init_continuation(&mg->k, continuation);
1063	dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1064}
1065
1066static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1067{
1068	struct continuation *k = container_of(ws, struct continuation, ws);
 
1069	return container_of(k, struct dm_cache_migration, k);
1070}
1071
1072static void copy_complete(int read_err, unsigned long write_err, void *context)
1073{
1074	struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1075
1076	if (read_err || write_err)
1077		mg->k.input = BLK_STS_IOERR;
1078
1079	queue_continuation(mg->cache->wq, &mg->k);
1080}
1081
1082static void copy(struct dm_cache_migration *mg, bool promote)
1083{
1084	struct dm_io_region o_region, c_region;
1085	struct cache *cache = mg->cache;
1086
1087	o_region.bdev = cache->origin_dev->bdev;
1088	o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1089	o_region.count = cache->sectors_per_block;
1090
1091	c_region.bdev = cache->cache_dev->bdev;
1092	c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1093	c_region.count = cache->sectors_per_block;
1094
1095	if (promote)
1096		dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1097	else
1098		dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1099}
1100
1101static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1102{
1103	struct per_bio_data *pb = get_per_bio_data(bio);
1104
1105	if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1106		free_prison_cell(cache, pb->cell);
1107	pb->cell = NULL;
1108}
1109
1110static void overwrite_endio(struct bio *bio)
1111{
1112	struct dm_cache_migration *mg = bio->bi_private;
1113	struct cache *cache = mg->cache;
1114	struct per_bio_data *pb = get_per_bio_data(bio);
1115
1116	dm_unhook_bio(&pb->hook_info, bio);
1117
1118	if (bio->bi_status)
1119		mg->k.input = bio->bi_status;
1120
1121	queue_continuation(cache->wq, &mg->k);
1122}
1123
1124static void overwrite(struct dm_cache_migration *mg,
1125		      void (*continuation)(struct work_struct *))
1126{
1127	struct bio *bio = mg->overwrite_bio;
1128	struct per_bio_data *pb = get_per_bio_data(bio);
1129
1130	dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1131
1132	/*
1133	 * The overwrite bio is part of the copy operation, as such it does
1134	 * not set/clear discard or dirty flags.
1135	 */
1136	if (mg->op->op == POLICY_PROMOTE)
1137		remap_to_cache(mg->cache, bio, mg->op->cblock);
1138	else
1139		remap_to_origin(mg->cache, bio);
1140
1141	init_continuation(&mg->k, continuation);
1142	accounted_request(mg->cache, bio);
1143}
1144
1145/*
1146 * Migration steps:
1147 *
1148 * 1) exclusive lock preventing WRITEs
1149 * 2) quiesce
1150 * 3) copy or issue overwrite bio
1151 * 4) upgrade to exclusive lock preventing READs and WRITEs
1152 * 5) quiesce
1153 * 6) update metadata and commit
1154 * 7) unlock
1155 */
1156static void mg_complete(struct dm_cache_migration *mg, bool success)
1157{
1158	struct bio_list bios;
1159	struct cache *cache = mg->cache;
1160	struct policy_work *op = mg->op;
1161	dm_cblock_t cblock = op->cblock;
1162
1163	if (success)
1164		update_stats(&cache->stats, op->op);
1165
1166	switch (op->op) {
1167	case POLICY_PROMOTE:
1168		clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1169		policy_complete_background_work(cache->policy, op, success);
1170
1171		if (mg->overwrite_bio) {
1172			if (success)
1173				force_set_dirty(cache, cblock);
1174			else if (mg->k.input)
1175				mg->overwrite_bio->bi_status = mg->k.input;
1176			else
1177				mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1178			bio_endio(mg->overwrite_bio);
1179		} else {
1180			if (success)
1181				force_clear_dirty(cache, cblock);
1182			dec_io_migrations(cache);
1183		}
1184		break;
1185
1186	case POLICY_DEMOTE:
1187		/*
1188		 * We clear dirty here to update the nr_dirty counter.
1189		 */
1190		if (success)
1191			force_clear_dirty(cache, cblock);
1192		policy_complete_background_work(cache->policy, op, success);
1193		dec_io_migrations(cache);
1194		break;
1195
1196	case POLICY_WRITEBACK:
1197		if (success)
1198			force_clear_dirty(cache, cblock);
1199		policy_complete_background_work(cache->policy, op, success);
1200		dec_io_migrations(cache);
1201		break;
1202	}
1203
1204	bio_list_init(&bios);
1205	if (mg->cell) {
1206		if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1207			free_prison_cell(cache, mg->cell);
1208	}
1209
1210	free_migration(mg);
1211	defer_bios(cache, &bios);
1212	wake_migration_worker(cache);
1213
1214	background_work_end(cache);
1215}
1216
1217static void mg_success(struct work_struct *ws)
1218{
1219	struct dm_cache_migration *mg = ws_to_mg(ws);
 
1220	mg_complete(mg, mg->k.input == 0);
1221}
1222
1223static void mg_update_metadata(struct work_struct *ws)
1224{
1225	int r;
1226	struct dm_cache_migration *mg = ws_to_mg(ws);
1227	struct cache *cache = mg->cache;
1228	struct policy_work *op = mg->op;
1229
1230	switch (op->op) {
1231	case POLICY_PROMOTE:
1232		r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1233		if (r) {
1234			DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1235				    cache_device_name(cache));
1236			metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1237
1238			mg_complete(mg, false);
1239			return;
1240		}
1241		mg_complete(mg, true);
1242		break;
1243
1244	case POLICY_DEMOTE:
1245		r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1246		if (r) {
1247			DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1248				    cache_device_name(cache));
1249			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1250
1251			mg_complete(mg, false);
1252			return;
1253		}
1254
1255		/*
1256		 * It would be nice if we only had to commit when a REQ_FLUSH
1257		 * comes through.  But there's one scenario that we have to
1258		 * look out for:
1259		 *
1260		 * - vblock x in a cache block
1261		 * - domotion occurs
1262		 * - cache block gets reallocated and over written
1263		 * - crash
1264		 *
1265		 * When we recover, because there was no commit the cache will
1266		 * rollback to having the data for vblock x in the cache block.
1267		 * But the cache block has since been overwritten, so it'll end
1268		 * up pointing to data that was never in 'x' during the history
1269		 * of the device.
1270		 *
1271		 * To avoid this issue we require a commit as part of the
1272		 * demotion operation.
1273		 */
1274		init_continuation(&mg->k, mg_success);
1275		continue_after_commit(&cache->committer, &mg->k);
1276		schedule_commit(&cache->committer);
1277		break;
1278
1279	case POLICY_WRITEBACK:
1280		mg_complete(mg, true);
1281		break;
1282	}
1283}
1284
1285static void mg_update_metadata_after_copy(struct work_struct *ws)
1286{
1287	struct dm_cache_migration *mg = ws_to_mg(ws);
1288
1289	/*
1290	 * Did the copy succeed?
1291	 */
1292	if (mg->k.input)
1293		mg_complete(mg, false);
1294	else
1295		mg_update_metadata(ws);
1296}
1297
1298static void mg_upgrade_lock(struct work_struct *ws)
1299{
1300	int r;
1301	struct dm_cache_migration *mg = ws_to_mg(ws);
1302
1303	/*
1304	 * Did the copy succeed?
1305	 */
1306	if (mg->k.input)
1307		mg_complete(mg, false);
1308
1309	else {
1310		/*
1311		 * Now we want the lock to prevent both reads and writes.
1312		 */
1313		r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1314					    READ_WRITE_LOCK_LEVEL);
1315		if (r < 0)
1316			mg_complete(mg, false);
1317
1318		else if (r)
1319			quiesce(mg, mg_update_metadata);
1320
1321		else
1322			mg_update_metadata(ws);
1323	}
1324}
1325
1326static void mg_full_copy(struct work_struct *ws)
1327{
1328	struct dm_cache_migration *mg = ws_to_mg(ws);
1329	struct cache *cache = mg->cache;
1330	struct policy_work *op = mg->op;
1331	bool is_policy_promote = (op->op == POLICY_PROMOTE);
1332
1333	if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1334	    is_discarded_oblock(cache, op->oblock)) {
1335		mg_upgrade_lock(ws);
1336		return;
1337	}
1338
1339	init_continuation(&mg->k, mg_upgrade_lock);
1340	copy(mg, is_policy_promote);
1341}
1342
1343static void mg_copy(struct work_struct *ws)
1344{
1345	struct dm_cache_migration *mg = ws_to_mg(ws);
1346
1347	if (mg->overwrite_bio) {
1348		/*
1349		 * No exclusive lock was held when we last checked if the bio
1350		 * was optimisable.  So we have to check again in case things
1351		 * have changed (eg, the block may no longer be discarded).
1352		 */
1353		if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1354			/*
1355			 * Fallback to a real full copy after doing some tidying up.
1356			 */
1357			bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
 
1358			BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1359			mg->overwrite_bio = NULL;
1360			inc_io_migrations(mg->cache);
1361			mg_full_copy(ws);
1362			return;
1363		}
1364
1365		/*
1366		 * It's safe to do this here, even though it's new data
1367		 * because all IO has been locked out of the block.
1368		 *
1369		 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1370		 * so _not_ using mg_upgrade_lock() as continutation.
1371		 */
1372		overwrite(mg, mg_update_metadata_after_copy);
1373
1374	} else
1375		mg_full_copy(ws);
1376}
1377
1378static int mg_lock_writes(struct dm_cache_migration *mg)
1379{
1380	int r;
1381	struct dm_cell_key_v2 key;
1382	struct cache *cache = mg->cache;
1383	struct dm_bio_prison_cell_v2 *prealloc;
1384
1385	prealloc = alloc_prison_cell(cache);
1386
1387	/*
1388	 * Prevent writes to the block, but allow reads to continue.
1389	 * Unless we're using an overwrite bio, in which case we lock
1390	 * everything.
1391	 */
1392	build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1393	r = dm_cell_lock_v2(cache->prison, &key,
1394			    mg->overwrite_bio ?  READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1395			    prealloc, &mg->cell);
1396	if (r < 0) {
1397		free_prison_cell(cache, prealloc);
1398		mg_complete(mg, false);
1399		return r;
1400	}
1401
1402	if (mg->cell != prealloc)
1403		free_prison_cell(cache, prealloc);
1404
1405	if (r == 0)
1406		mg_copy(&mg->k.ws);
1407	else
1408		quiesce(mg, mg_copy);
1409
1410	return 0;
1411}
1412
1413static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1414{
1415	struct dm_cache_migration *mg;
1416
1417	if (!background_work_begin(cache)) {
1418		policy_complete_background_work(cache->policy, op, false);
1419		return -EPERM;
1420	}
1421
1422	mg = alloc_migration(cache);
1423
1424	mg->op = op;
1425	mg->overwrite_bio = bio;
1426
1427	if (!bio)
1428		inc_io_migrations(cache);
1429
1430	return mg_lock_writes(mg);
1431}
1432
1433/*----------------------------------------------------------------
 
1434 * invalidation processing
1435 *--------------------------------------------------------------*/
 
1436
1437static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1438{
1439	struct bio_list bios;
1440	struct cache *cache = mg->cache;
1441
1442	bio_list_init(&bios);
1443	if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1444		free_prison_cell(cache, mg->cell);
1445
1446	if (!success && mg->overwrite_bio)
1447		bio_io_error(mg->overwrite_bio);
1448
1449	free_migration(mg);
1450	defer_bios(cache, &bios);
1451
1452	background_work_end(cache);
1453}
1454
1455static void invalidate_completed(struct work_struct *ws)
1456{
1457	struct dm_cache_migration *mg = ws_to_mg(ws);
 
1458	invalidate_complete(mg, !mg->k.input);
1459}
1460
1461static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1462{
1463	int r = policy_invalidate_mapping(cache->policy, cblock);
 
 
1464	if (!r) {
1465		r = dm_cache_remove_mapping(cache->cmd, cblock);
1466		if (r) {
1467			DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1468				    cache_device_name(cache));
1469			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1470		}
1471
1472	} else if (r == -ENODATA) {
1473		/*
1474		 * Harmless, already unmapped.
1475		 */
1476		r = 0;
1477
1478	} else
1479		DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1480
1481	return r;
1482}
1483
1484static void invalidate_remove(struct work_struct *ws)
1485{
1486	int r;
1487	struct dm_cache_migration *mg = ws_to_mg(ws);
1488	struct cache *cache = mg->cache;
1489
1490	r = invalidate_cblock(cache, mg->invalidate_cblock);
1491	if (r) {
1492		invalidate_complete(mg, false);
1493		return;
1494	}
1495
1496	init_continuation(&mg->k, invalidate_completed);
1497	continue_after_commit(&cache->committer, &mg->k);
1498	remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1499	mg->overwrite_bio = NULL;
1500	schedule_commit(&cache->committer);
1501}
1502
1503static int invalidate_lock(struct dm_cache_migration *mg)
1504{
1505	int r;
1506	struct dm_cell_key_v2 key;
1507	struct cache *cache = mg->cache;
1508	struct dm_bio_prison_cell_v2 *prealloc;
1509
1510	prealloc = alloc_prison_cell(cache);
1511
1512	build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1513	r = dm_cell_lock_v2(cache->prison, &key,
1514			    READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1515	if (r < 0) {
1516		free_prison_cell(cache, prealloc);
1517		invalidate_complete(mg, false);
1518		return r;
1519	}
1520
1521	if (mg->cell != prealloc)
1522		free_prison_cell(cache, prealloc);
1523
1524	if (r)
1525		quiesce(mg, invalidate_remove);
1526
1527	else {
1528		/*
1529		 * We can't call invalidate_remove() directly here because we
1530		 * might still be in request context.
1531		 */
1532		init_continuation(&mg->k, invalidate_remove);
1533		queue_work(cache->wq, &mg->k.ws);
1534	}
1535
1536	return 0;
1537}
1538
1539static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1540			    dm_oblock_t oblock, struct bio *bio)
1541{
1542	struct dm_cache_migration *mg;
1543
1544	if (!background_work_begin(cache))
1545		return -EPERM;
1546
1547	mg = alloc_migration(cache);
1548
1549	mg->overwrite_bio = bio;
1550	mg->invalidate_cblock = cblock;
1551	mg->invalidate_oblock = oblock;
1552
1553	return invalidate_lock(mg);
1554}
1555
1556/*----------------------------------------------------------------
 
1557 * bio processing
1558 *--------------------------------------------------------------*/
 
1559
1560enum busy {
1561	IDLE,
1562	BUSY
1563};
1564
1565static enum busy spare_migration_bandwidth(struct cache *cache)
1566{
1567	bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1568	sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1569		cache->sectors_per_block;
1570
1571	if (idle && current_volume <= cache->migration_threshold)
1572		return IDLE;
1573	else
1574		return BUSY;
1575}
1576
1577static void inc_hit_counter(struct cache *cache, struct bio *bio)
1578{
1579	atomic_inc(bio_data_dir(bio) == READ ?
1580		   &cache->stats.read_hit : &cache->stats.write_hit);
1581}
1582
1583static void inc_miss_counter(struct cache *cache, struct bio *bio)
1584{
1585	atomic_inc(bio_data_dir(bio) == READ ?
1586		   &cache->stats.read_miss : &cache->stats.write_miss);
1587}
1588
1589/*----------------------------------------------------------------*/
1590
1591static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1592		   bool *commit_needed)
1593{
1594	int r, data_dir;
1595	bool rb, background_queued;
1596	dm_cblock_t cblock;
1597
1598	*commit_needed = false;
1599
1600	rb = bio_detain_shared(cache, block, bio);
1601	if (!rb) {
1602		/*
1603		 * An exclusive lock is held for this block, so we have to
1604		 * wait.  We set the commit_needed flag so the current
1605		 * transaction will be committed asap, allowing this lock
1606		 * to be dropped.
1607		 */
1608		*commit_needed = true;
1609		return DM_MAPIO_SUBMITTED;
1610	}
1611
1612	data_dir = bio_data_dir(bio);
1613
1614	if (optimisable_bio(cache, bio, block)) {
1615		struct policy_work *op = NULL;
1616
1617		r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1618		if (unlikely(r && r != -ENOENT)) {
1619			DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1620				    cache_device_name(cache), r);
1621			bio_io_error(bio);
1622			return DM_MAPIO_SUBMITTED;
1623		}
1624
1625		if (r == -ENOENT && op) {
1626			bio_drop_shared_lock(cache, bio);
1627			BUG_ON(op->op != POLICY_PROMOTE);
1628			mg_start(cache, op, bio);
1629			return DM_MAPIO_SUBMITTED;
1630		}
1631	} else {
1632		r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1633		if (unlikely(r && r != -ENOENT)) {
1634			DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1635				    cache_device_name(cache), r);
1636			bio_io_error(bio);
1637			return DM_MAPIO_SUBMITTED;
1638		}
1639
1640		if (background_queued)
1641			wake_migration_worker(cache);
1642	}
1643
1644	if (r == -ENOENT) {
1645		struct per_bio_data *pb = get_per_bio_data(bio);
1646
1647		/*
1648		 * Miss.
1649		 */
1650		inc_miss_counter(cache, bio);
1651		if (pb->req_nr == 0) {
1652			accounted_begin(cache, bio);
1653			remap_to_origin_clear_discard(cache, bio, block);
1654		} else {
1655			/*
1656			 * This is a duplicate writethrough io that is no
1657			 * longer needed because the block has been demoted.
1658			 */
1659			bio_endio(bio);
1660			return DM_MAPIO_SUBMITTED;
1661		}
1662	} else {
1663		/*
1664		 * Hit.
1665		 */
1666		inc_hit_counter(cache, bio);
1667
1668		/*
1669		 * Passthrough always maps to the origin, invalidating any
1670		 * cache blocks that are written to.
1671		 */
1672		if (passthrough_mode(cache)) {
1673			if (bio_data_dir(bio) == WRITE) {
1674				bio_drop_shared_lock(cache, bio);
1675				atomic_inc(&cache->stats.demotion);
1676				invalidate_start(cache, cblock, block, bio);
1677			} else
1678				remap_to_origin_clear_discard(cache, bio, block);
1679		} else {
1680			if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1681			    !is_dirty(cache, cblock)) {
1682				remap_to_origin_and_cache(cache, bio, block, cblock);
1683				accounted_begin(cache, bio);
1684			} else
1685				remap_to_cache_dirty(cache, bio, block, cblock);
1686		}
1687	}
1688
1689	/*
1690	 * dm core turns FUA requests into a separate payload and FLUSH req.
1691	 */
1692	if (bio->bi_opf & REQ_FUA) {
1693		/*
1694		 * issue_after_commit will call accounted_begin a second time.  So
1695		 * we call accounted_complete() to avoid double accounting.
1696		 */
1697		accounted_complete(cache, bio);
1698		issue_after_commit(&cache->committer, bio);
1699		*commit_needed = true;
1700		return DM_MAPIO_SUBMITTED;
1701	}
1702
1703	return DM_MAPIO_REMAPPED;
1704}
1705
1706static bool process_bio(struct cache *cache, struct bio *bio)
1707{
1708	bool commit_needed;
1709
1710	if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1711		dm_submit_bio_remap(bio, NULL);
1712
1713	return commit_needed;
1714}
1715
1716/*
1717 * A non-zero return indicates read_only or fail_io mode.
1718 */
1719static int commit(struct cache *cache, bool clean_shutdown)
1720{
1721	int r;
1722
1723	if (get_cache_mode(cache) >= CM_READ_ONLY)
1724		return -EINVAL;
1725
1726	atomic_inc(&cache->stats.commit_count);
1727	r = dm_cache_commit(cache->cmd, clean_shutdown);
1728	if (r)
1729		metadata_operation_failed(cache, "dm_cache_commit", r);
1730
1731	return r;
1732}
1733
1734/*
1735 * Used by the batcher.
1736 */
1737static blk_status_t commit_op(void *context)
1738{
1739	struct cache *cache = context;
1740
1741	if (dm_cache_changed_this_transaction(cache->cmd))
1742		return errno_to_blk_status(commit(cache, false));
1743
1744	return 0;
1745}
1746
1747/*----------------------------------------------------------------*/
1748
1749static bool process_flush_bio(struct cache *cache, struct bio *bio)
1750{
1751	struct per_bio_data *pb = get_per_bio_data(bio);
1752
1753	if (!pb->req_nr)
1754		remap_to_origin(cache, bio);
1755	else
1756		remap_to_cache(cache, bio, 0);
1757
1758	issue_after_commit(&cache->committer, bio);
1759	return true;
1760}
1761
1762static bool process_discard_bio(struct cache *cache, struct bio *bio)
1763{
1764	dm_dblock_t b, e;
1765
1766	// FIXME: do we need to lock the region?  Or can we just assume the
1767	// user wont be so foolish as to issue discard concurrently with
1768	// other IO?
 
 
1769	calc_discard_block_range(cache, bio, &b, &e);
1770	while (b != e) {
1771		set_discard(cache, b);
1772		b = to_dblock(from_dblock(b) + 1);
1773	}
1774
1775	if (cache->features.discard_passdown) {
1776		remap_to_origin(cache, bio);
1777		dm_submit_bio_remap(bio, NULL);
1778	} else
1779		bio_endio(bio);
1780
1781	return false;
1782}
1783
1784static void process_deferred_bios(struct work_struct *ws)
1785{
1786	struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1787
1788	bool commit_needed = false;
1789	struct bio_list bios;
1790	struct bio *bio;
1791
1792	bio_list_init(&bios);
1793
1794	spin_lock_irq(&cache->lock);
1795	bio_list_merge(&bios, &cache->deferred_bios);
1796	bio_list_init(&cache->deferred_bios);
1797	spin_unlock_irq(&cache->lock);
1798
1799	while ((bio = bio_list_pop(&bios))) {
1800		if (bio->bi_opf & REQ_PREFLUSH)
1801			commit_needed = process_flush_bio(cache, bio) || commit_needed;
1802
1803		else if (bio_op(bio) == REQ_OP_DISCARD)
1804			commit_needed = process_discard_bio(cache, bio) || commit_needed;
1805
1806		else
1807			commit_needed = process_bio(cache, bio) || commit_needed;
 
1808	}
1809
1810	if (commit_needed)
1811		schedule_commit(&cache->committer);
1812}
1813
1814/*----------------------------------------------------------------
 
1815 * Main worker loop
1816 *--------------------------------------------------------------*/
1817
1818static void requeue_deferred_bios(struct cache *cache)
1819{
1820	struct bio *bio;
1821	struct bio_list bios;
1822
1823	bio_list_init(&bios);
1824	bio_list_merge(&bios, &cache->deferred_bios);
1825	bio_list_init(&cache->deferred_bios);
1826
1827	while ((bio = bio_list_pop(&bios))) {
1828		bio->bi_status = BLK_STS_DM_REQUEUE;
1829		bio_endio(bio);
 
1830	}
1831}
1832
1833/*
1834 * We want to commit periodically so that not too much
1835 * unwritten metadata builds up.
1836 */
1837static void do_waker(struct work_struct *ws)
1838{
1839	struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1840
1841	policy_tick(cache->policy, true);
1842	wake_migration_worker(cache);
1843	schedule_commit(&cache->committer);
1844	queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1845}
1846
1847static void check_migrations(struct work_struct *ws)
1848{
1849	int r;
1850	struct policy_work *op;
1851	struct cache *cache = container_of(ws, struct cache, migration_worker);
1852	enum busy b;
1853
1854	for (;;) {
1855		b = spare_migration_bandwidth(cache);
1856
1857		r = policy_get_background_work(cache->policy, b == IDLE, &op);
1858		if (r == -ENODATA)
1859			break;
1860
1861		if (r) {
1862			DMERR_LIMIT("%s: policy_background_work failed",
1863				    cache_device_name(cache));
1864			break;
1865		}
1866
1867		r = mg_start(cache, op, NULL);
1868		if (r)
1869			break;
 
 
1870	}
1871}
1872
1873/*----------------------------------------------------------------
 
1874 * Target methods
1875 *--------------------------------------------------------------*/
 
1876
1877/*
1878 * This function gets called on the error paths of the constructor, so we
1879 * have to cope with a partially initialised struct.
1880 */
1881static void destroy(struct cache *cache)
1882{
1883	unsigned i;
1884
1885	mempool_exit(&cache->migration_pool);
1886
1887	if (cache->prison)
1888		dm_bio_prison_destroy_v2(cache->prison);
1889
1890	cancel_delayed_work_sync(&cache->waker);
1891	if (cache->wq)
1892		destroy_workqueue(cache->wq);
1893
1894	if (cache->dirty_bitset)
1895		free_bitset(cache->dirty_bitset);
1896
1897	if (cache->discard_bitset)
1898		free_bitset(cache->discard_bitset);
1899
1900	if (cache->copier)
1901		dm_kcopyd_client_destroy(cache->copier);
1902
1903	if (cache->cmd)
1904		dm_cache_metadata_close(cache->cmd);
1905
1906	if (cache->metadata_dev)
1907		dm_put_device(cache->ti, cache->metadata_dev);
1908
1909	if (cache->origin_dev)
1910		dm_put_device(cache->ti, cache->origin_dev);
1911
1912	if (cache->cache_dev)
1913		dm_put_device(cache->ti, cache->cache_dev);
1914
1915	if (cache->policy)
1916		dm_cache_policy_destroy(cache->policy);
1917
1918	for (i = 0; i < cache->nr_ctr_args ; i++)
1919		kfree(cache->ctr_args[i]);
1920	kfree(cache->ctr_args);
1921
1922	bioset_exit(&cache->bs);
1923
1924	kfree(cache);
1925}
1926
1927static void cache_dtr(struct dm_target *ti)
1928{
1929	struct cache *cache = ti->private;
1930
1931	destroy(cache);
1932}
1933
1934static sector_t get_dev_size(struct dm_dev *dev)
1935{
1936	return bdev_nr_sectors(dev->bdev);
1937}
1938
1939/*----------------------------------------------------------------*/
1940
1941/*
1942 * Construct a cache device mapping.
1943 *
1944 * cache <metadata dev> <cache dev> <origin dev> <block size>
1945 *       <#feature args> [<feature arg>]*
1946 *       <policy> <#policy args> [<policy arg>]*
1947 *
1948 * metadata dev    : fast device holding the persistent metadata
1949 * cache dev	   : fast device holding cached data blocks
1950 * origin dev	   : slow device holding original data blocks
1951 * block size	   : cache unit size in sectors
1952 *
1953 * #feature args   : number of feature arguments passed
1954 * feature args    : writethrough.  (The default is writeback.)
1955 *
1956 * policy	   : the replacement policy to use
1957 * #policy args    : an even number of policy arguments corresponding
1958 *		     to key/value pairs passed to the policy
1959 * policy args	   : key/value pairs passed to the policy
1960 *		     E.g. 'sequential_threshold 1024'
1961 *		     See cache-policies.txt for details.
1962 *
1963 * Optional feature arguments are:
1964 *   writethrough  : write through caching that prohibits cache block
1965 *		     content from being different from origin block content.
1966 *		     Without this argument, the default behaviour is to write
1967 *		     back cache block contents later for performance reasons,
1968 *		     so they may differ from the corresponding origin blocks.
1969 */
1970struct cache_args {
1971	struct dm_target *ti;
1972
1973	struct dm_dev *metadata_dev;
1974
1975	struct dm_dev *cache_dev;
1976	sector_t cache_sectors;
1977
1978	struct dm_dev *origin_dev;
1979	sector_t origin_sectors;
1980
1981	uint32_t block_size;
1982
1983	const char *policy_name;
1984	int policy_argc;
1985	const char **policy_argv;
1986
1987	struct cache_features features;
1988};
1989
1990static void destroy_cache_args(struct cache_args *ca)
1991{
1992	if (ca->metadata_dev)
1993		dm_put_device(ca->ti, ca->metadata_dev);
1994
1995	if (ca->cache_dev)
1996		dm_put_device(ca->ti, ca->cache_dev);
1997
1998	if (ca->origin_dev)
1999		dm_put_device(ca->ti, ca->origin_dev);
2000
2001	kfree(ca);
2002}
2003
2004static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2005{
2006	if (!as->argc) {
2007		*error = "Insufficient args";
2008		return false;
2009	}
2010
2011	return true;
2012}
2013
2014static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2015			      char **error)
2016{
2017	int r;
2018	sector_t metadata_dev_size;
2019
2020	if (!at_least_one_arg(as, error))
2021		return -EINVAL;
2022
2023	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2024			  &ca->metadata_dev);
2025	if (r) {
2026		*error = "Error opening metadata device";
2027		return r;
2028	}
2029
2030	metadata_dev_size = get_dev_size(ca->metadata_dev);
2031	if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2032		DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2033		       ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2034
2035	return 0;
2036}
2037
2038static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2039			   char **error)
2040{
2041	int r;
2042
2043	if (!at_least_one_arg(as, error))
2044		return -EINVAL;
2045
2046	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2047			  &ca->cache_dev);
2048	if (r) {
2049		*error = "Error opening cache device";
2050		return r;
2051	}
2052	ca->cache_sectors = get_dev_size(ca->cache_dev);
2053
2054	return 0;
2055}
2056
2057static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2058			    char **error)
2059{
2060	int r;
2061
2062	if (!at_least_one_arg(as, error))
2063		return -EINVAL;
2064
2065	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2066			  &ca->origin_dev);
2067	if (r) {
2068		*error = "Error opening origin device";
2069		return r;
2070	}
2071
2072	ca->origin_sectors = get_dev_size(ca->origin_dev);
2073	if (ca->ti->len > ca->origin_sectors) {
2074		*error = "Device size larger than cached device";
2075		return -EINVAL;
2076	}
2077
2078	return 0;
2079}
2080
2081static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2082			    char **error)
2083{
2084	unsigned long block_size;
2085
2086	if (!at_least_one_arg(as, error))
2087		return -EINVAL;
2088
2089	if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2090	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2091	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2092	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2093		*error = "Invalid data block size";
2094		return -EINVAL;
2095	}
2096
2097	if (block_size > ca->cache_sectors) {
2098		*error = "Data block size is larger than the cache device";
2099		return -EINVAL;
2100	}
2101
2102	ca->block_size = block_size;
2103
2104	return 0;
2105}
2106
2107static void init_features(struct cache_features *cf)
2108{
2109	cf->mode = CM_WRITE;
2110	cf->io_mode = CM_IO_WRITEBACK;
2111	cf->metadata_version = 1;
2112	cf->discard_passdown = true;
2113}
2114
2115static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2116			  char **error)
2117{
2118	static const struct dm_arg _args[] = {
2119		{0, 3, "Invalid number of cache feature arguments"},
2120	};
2121
2122	int r, mode_ctr = 0;
2123	unsigned argc;
2124	const char *arg;
2125	struct cache_features *cf = &ca->features;
2126
2127	init_features(cf);
2128
2129	r = dm_read_arg_group(_args, as, &argc, error);
2130	if (r)
2131		return -EINVAL;
2132
2133	while (argc--) {
2134		arg = dm_shift_arg(as);
2135
2136		if (!strcasecmp(arg, "writeback")) {
2137			cf->io_mode = CM_IO_WRITEBACK;
2138			mode_ctr++;
2139		}
2140
2141		else if (!strcasecmp(arg, "writethrough")) {
2142			cf->io_mode = CM_IO_WRITETHROUGH;
2143			mode_ctr++;
2144		}
2145
2146		else if (!strcasecmp(arg, "passthrough")) {
2147			cf->io_mode = CM_IO_PASSTHROUGH;
2148			mode_ctr++;
2149		}
2150
2151		else if (!strcasecmp(arg, "metadata2"))
2152			cf->metadata_version = 2;
2153
2154		else if (!strcasecmp(arg, "no_discard_passdown"))
2155			cf->discard_passdown = false;
2156
2157		else {
2158			*error = "Unrecognised cache feature requested";
2159			return -EINVAL;
2160		}
2161	}
2162
2163	if (mode_ctr > 1) {
2164		*error = "Duplicate cache io_mode features requested";
2165		return -EINVAL;
2166	}
2167
2168	return 0;
2169}
2170
2171static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2172			char **error)
2173{
2174	static const struct dm_arg _args[] = {
2175		{0, 1024, "Invalid number of policy arguments"},
2176	};
2177
2178	int r;
2179
2180	if (!at_least_one_arg(as, error))
2181		return -EINVAL;
2182
2183	ca->policy_name = dm_shift_arg(as);
2184
2185	r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2186	if (r)
2187		return -EINVAL;
2188
2189	ca->policy_argv = (const char **)as->argv;
2190	dm_consume_args(as, ca->policy_argc);
2191
2192	return 0;
2193}
2194
2195static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2196			    char **error)
2197{
2198	int r;
2199	struct dm_arg_set as;
2200
2201	as.argc = argc;
2202	as.argv = argv;
2203
2204	r = parse_metadata_dev(ca, &as, error);
2205	if (r)
2206		return r;
2207
2208	r = parse_cache_dev(ca, &as, error);
2209	if (r)
2210		return r;
2211
2212	r = parse_origin_dev(ca, &as, error);
2213	if (r)
2214		return r;
2215
2216	r = parse_block_size(ca, &as, error);
2217	if (r)
2218		return r;
2219
2220	r = parse_features(ca, &as, error);
2221	if (r)
2222		return r;
2223
2224	r = parse_policy(ca, &as, error);
2225	if (r)
2226		return r;
2227
2228	return 0;
2229}
2230
2231/*----------------------------------------------------------------*/
2232
2233static struct kmem_cache *migration_cache;
2234
2235#define NOT_CORE_OPTION 1
2236
2237static int process_config_option(struct cache *cache, const char *key, const char *value)
2238{
2239	unsigned long tmp;
2240
2241	if (!strcasecmp(key, "migration_threshold")) {
2242		if (kstrtoul(value, 10, &tmp))
2243			return -EINVAL;
2244
2245		cache->migration_threshold = tmp;
2246		return 0;
2247	}
2248
2249	return NOT_CORE_OPTION;
2250}
2251
2252static int set_config_value(struct cache *cache, const char *key, const char *value)
2253{
2254	int r = process_config_option(cache, key, value);
2255
2256	if (r == NOT_CORE_OPTION)
2257		r = policy_set_config_value(cache->policy, key, value);
2258
2259	if (r)
2260		DMWARN("bad config value for %s: %s", key, value);
2261
2262	return r;
2263}
2264
2265static int set_config_values(struct cache *cache, int argc, const char **argv)
2266{
2267	int r = 0;
2268
2269	if (argc & 1) {
2270		DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2271		return -EINVAL;
2272	}
2273
2274	while (argc) {
2275		r = set_config_value(cache, argv[0], argv[1]);
2276		if (r)
2277			break;
2278
2279		argc -= 2;
2280		argv += 2;
2281	}
2282
2283	return r;
2284}
2285
2286static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2287			       char **error)
2288{
2289	struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2290							   cache->cache_size,
2291							   cache->origin_sectors,
2292							   cache->sectors_per_block);
2293	if (IS_ERR(p)) {
2294		*error = "Error creating cache's policy";
2295		return PTR_ERR(p);
2296	}
2297	cache->policy = p;
2298	BUG_ON(!cache->policy);
2299
2300	return 0;
2301}
2302
2303/*
2304 * We want the discard block size to be at least the size of the cache
2305 * block size and have no more than 2^14 discard blocks across the origin.
2306 */
2307#define MAX_DISCARD_BLOCKS (1 << 14)
2308
2309static bool too_many_discard_blocks(sector_t discard_block_size,
2310				    sector_t origin_size)
2311{
2312	(void) sector_div(origin_size, discard_block_size);
2313
2314	return origin_size > MAX_DISCARD_BLOCKS;
2315}
2316
2317static sector_t calculate_discard_block_size(sector_t cache_block_size,
2318					     sector_t origin_size)
2319{
2320	sector_t discard_block_size = cache_block_size;
2321
2322	if (origin_size)
2323		while (too_many_discard_blocks(discard_block_size, origin_size))
2324			discard_block_size *= 2;
2325
2326	return discard_block_size;
2327}
2328
2329static void set_cache_size(struct cache *cache, dm_cblock_t size)
2330{
2331	dm_block_t nr_blocks = from_cblock(size);
2332
2333	if (nr_blocks > (1 << 20) && cache->cache_size != size)
2334		DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2335			     "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2336			     "Please consider increasing the cache block size to reduce the overall cache block count.",
2337			     (unsigned long long) nr_blocks);
2338
2339	cache->cache_size = size;
2340}
2341
2342#define DEFAULT_MIGRATION_THRESHOLD 2048
2343
2344static int cache_create(struct cache_args *ca, struct cache **result)
2345{
2346	int r = 0;
2347	char **error = &ca->ti->error;
2348	struct cache *cache;
2349	struct dm_target *ti = ca->ti;
2350	dm_block_t origin_blocks;
2351	struct dm_cache_metadata *cmd;
2352	bool may_format = ca->features.mode == CM_WRITE;
2353
2354	cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2355	if (!cache)
2356		return -ENOMEM;
2357
2358	cache->ti = ca->ti;
2359	ti->private = cache;
2360	ti->accounts_remapped_io = true;
2361	ti->num_flush_bios = 2;
2362	ti->flush_supported = true;
2363
2364	ti->num_discard_bios = 1;
2365	ti->discards_supported = true;
2366
2367	ti->per_io_data_size = sizeof(struct per_bio_data);
2368
2369	cache->features = ca->features;
2370	if (writethrough_mode(cache)) {
2371		/* Create bioset for writethrough bios issued to origin */
2372		r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2373		if (r)
2374			goto bad;
2375	}
2376
2377	cache->metadata_dev = ca->metadata_dev;
2378	cache->origin_dev = ca->origin_dev;
2379	cache->cache_dev = ca->cache_dev;
2380
2381	ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2382
2383	origin_blocks = cache->origin_sectors = ca->origin_sectors;
2384	origin_blocks = block_div(origin_blocks, ca->block_size);
2385	cache->origin_blocks = to_oblock(origin_blocks);
2386
2387	cache->sectors_per_block = ca->block_size;
2388	if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2389		r = -EINVAL;
2390		goto bad;
2391	}
2392
2393	if (ca->block_size & (ca->block_size - 1)) {
2394		dm_block_t cache_size = ca->cache_sectors;
2395
2396		cache->sectors_per_block_shift = -1;
2397		cache_size = block_div(cache_size, ca->block_size);
2398		set_cache_size(cache, to_cblock(cache_size));
2399	} else {
2400		cache->sectors_per_block_shift = __ffs(ca->block_size);
2401		set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2402	}
2403
2404	r = create_cache_policy(cache, ca, error);
2405	if (r)
2406		goto bad;
2407
2408	cache->policy_nr_args = ca->policy_argc;
2409	cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2410
2411	r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2412	if (r) {
2413		*error = "Error setting cache policy's config values";
2414		goto bad;
2415	}
2416
2417	cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2418				     ca->block_size, may_format,
2419				     dm_cache_policy_get_hint_size(cache->policy),
2420				     ca->features.metadata_version);
2421	if (IS_ERR(cmd)) {
2422		*error = "Error creating metadata object";
2423		r = PTR_ERR(cmd);
2424		goto bad;
2425	}
2426	cache->cmd = cmd;
2427	set_cache_mode(cache, CM_WRITE);
2428	if (get_cache_mode(cache) != CM_WRITE) {
2429		*error = "Unable to get write access to metadata, please check/repair metadata.";
2430		r = -EINVAL;
2431		goto bad;
2432	}
2433
2434	if (passthrough_mode(cache)) {
2435		bool all_clean;
2436
2437		r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2438		if (r) {
2439			*error = "dm_cache_metadata_all_clean() failed";
2440			goto bad;
2441		}
2442
2443		if (!all_clean) {
2444			*error = "Cannot enter passthrough mode unless all blocks are clean";
2445			r = -EINVAL;
2446			goto bad;
2447		}
2448
2449		policy_allow_migrations(cache->policy, false);
2450	}
2451
2452	spin_lock_init(&cache->lock);
2453	bio_list_init(&cache->deferred_bios);
2454	atomic_set(&cache->nr_allocated_migrations, 0);
2455	atomic_set(&cache->nr_io_migrations, 0);
2456	init_waitqueue_head(&cache->migration_wait);
2457
2458	r = -ENOMEM;
2459	atomic_set(&cache->nr_dirty, 0);
2460	cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2461	if (!cache->dirty_bitset) {
2462		*error = "could not allocate dirty bitset";
2463		goto bad;
2464	}
2465	clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2466
2467	cache->discard_block_size =
2468		calculate_discard_block_size(cache->sectors_per_block,
2469					     cache->origin_sectors);
2470	cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2471							      cache->discard_block_size));
2472	cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2473	if (!cache->discard_bitset) {
2474		*error = "could not allocate discard bitset";
2475		goto bad;
2476	}
2477	clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2478
2479	cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2480	if (IS_ERR(cache->copier)) {
2481		*error = "could not create kcopyd client";
2482		r = PTR_ERR(cache->copier);
2483		goto bad;
2484	}
2485
2486	cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2487	if (!cache->wq) {
2488		*error = "could not create workqueue for metadata object";
2489		goto bad;
2490	}
2491	INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2492	INIT_WORK(&cache->migration_worker, check_migrations);
2493	INIT_DELAYED_WORK(&cache->waker, do_waker);
2494
2495	cache->prison = dm_bio_prison_create_v2(cache->wq);
2496	if (!cache->prison) {
2497		*error = "could not create bio prison";
2498		goto bad;
2499	}
2500
2501	r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2502				   migration_cache);
2503	if (r) {
2504		*error = "Error creating cache's migration mempool";
2505		goto bad;
2506	}
2507
2508	cache->need_tick_bio = true;
2509	cache->sized = false;
2510	cache->invalidate = false;
2511	cache->commit_requested = false;
2512	cache->loaded_mappings = false;
2513	cache->loaded_discards = false;
2514
2515	load_stats(cache);
2516
2517	atomic_set(&cache->stats.demotion, 0);
2518	atomic_set(&cache->stats.promotion, 0);
2519	atomic_set(&cache->stats.copies_avoided, 0);
2520	atomic_set(&cache->stats.cache_cell_clash, 0);
2521	atomic_set(&cache->stats.commit_count, 0);
2522	atomic_set(&cache->stats.discard_count, 0);
2523
2524	spin_lock_init(&cache->invalidation_lock);
2525	INIT_LIST_HEAD(&cache->invalidation_requests);
2526
2527	batcher_init(&cache->committer, commit_op, cache,
2528		     issue_op, cache, cache->wq);
2529	dm_iot_init(&cache->tracker);
2530
2531	init_rwsem(&cache->background_work_lock);
2532	prevent_background_work(cache);
2533
2534	*result = cache;
2535	return 0;
2536bad:
2537	destroy(cache);
2538	return r;
2539}
2540
2541static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2542{
2543	unsigned i;
2544	const char **copy;
2545
2546	copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2547	if (!copy)
2548		return -ENOMEM;
2549	for (i = 0; i < argc; i++) {
2550		copy[i] = kstrdup(argv[i], GFP_KERNEL);
2551		if (!copy[i]) {
2552			while (i--)
2553				kfree(copy[i]);
2554			kfree(copy);
2555			return -ENOMEM;
2556		}
2557	}
2558
2559	cache->nr_ctr_args = argc;
2560	cache->ctr_args = copy;
2561
2562	return 0;
2563}
2564
2565static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2566{
2567	int r = -EINVAL;
2568	struct cache_args *ca;
2569	struct cache *cache = NULL;
2570
2571	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2572	if (!ca) {
2573		ti->error = "Error allocating memory for cache";
2574		return -ENOMEM;
2575	}
2576	ca->ti = ti;
2577
2578	r = parse_cache_args(ca, argc, argv, &ti->error);
2579	if (r)
2580		goto out;
2581
2582	r = cache_create(ca, &cache);
2583	if (r)
2584		goto out;
2585
2586	r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2587	if (r) {
2588		destroy(cache);
2589		goto out;
2590	}
2591
2592	ti->private = cache;
2593out:
2594	destroy_cache_args(ca);
2595	return r;
2596}
2597
2598/*----------------------------------------------------------------*/
2599
2600static int cache_map(struct dm_target *ti, struct bio *bio)
2601{
2602	struct cache *cache = ti->private;
2603
2604	int r;
2605	bool commit_needed;
2606	dm_oblock_t block = get_bio_block(cache, bio);
2607
2608	init_per_bio_data(bio);
2609	if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2610		/*
2611		 * This can only occur if the io goes to a partial block at
2612		 * the end of the origin device.  We don't cache these.
2613		 * Just remap to the origin and carry on.
2614		 */
2615		remap_to_origin(cache, bio);
2616		accounted_begin(cache, bio);
2617		return DM_MAPIO_REMAPPED;
2618	}
2619
2620	if (discard_or_flush(bio)) {
2621		defer_bio(cache, bio);
2622		return DM_MAPIO_SUBMITTED;
2623	}
2624
2625	r = map_bio(cache, bio, block, &commit_needed);
2626	if (commit_needed)
2627		schedule_commit(&cache->committer);
2628
2629	return r;
2630}
2631
2632static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2633{
2634	struct cache *cache = ti->private;
2635	unsigned long flags;
2636	struct per_bio_data *pb = get_per_bio_data(bio);
2637
2638	if (pb->tick) {
2639		policy_tick(cache->policy, false);
2640
2641		spin_lock_irqsave(&cache->lock, flags);
2642		cache->need_tick_bio = true;
2643		spin_unlock_irqrestore(&cache->lock, flags);
2644	}
2645
2646	bio_drop_shared_lock(cache, bio);
2647	accounted_complete(cache, bio);
2648
2649	return DM_ENDIO_DONE;
2650}
2651
2652static int write_dirty_bitset(struct cache *cache)
2653{
2654	int r;
2655
2656	if (get_cache_mode(cache) >= CM_READ_ONLY)
2657		return -EINVAL;
2658
2659	r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2660	if (r)
2661		metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2662
2663	return r;
2664}
2665
2666static int write_discard_bitset(struct cache *cache)
2667{
2668	unsigned i, r;
2669
2670	if (get_cache_mode(cache) >= CM_READ_ONLY)
2671		return -EINVAL;
2672
2673	r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2674					   cache->discard_nr_blocks);
2675	if (r) {
2676		DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2677		metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2678		return r;
2679	}
2680
2681	for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2682		r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2683					 is_discarded(cache, to_dblock(i)));
2684		if (r) {
2685			metadata_operation_failed(cache, "dm_cache_set_discard", r);
2686			return r;
2687		}
2688	}
2689
2690	return 0;
2691}
2692
2693static int write_hints(struct cache *cache)
2694{
2695	int r;
2696
2697	if (get_cache_mode(cache) >= CM_READ_ONLY)
2698		return -EINVAL;
2699
2700	r = dm_cache_write_hints(cache->cmd, cache->policy);
2701	if (r) {
2702		metadata_operation_failed(cache, "dm_cache_write_hints", r);
2703		return r;
2704	}
2705
2706	return 0;
2707}
2708
2709/*
2710 * returns true on success
2711 */
2712static bool sync_metadata(struct cache *cache)
2713{
2714	int r1, r2, r3, r4;
2715
2716	r1 = write_dirty_bitset(cache);
2717	if (r1)
2718		DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2719
2720	r2 = write_discard_bitset(cache);
2721	if (r2)
2722		DMERR("%s: could not write discard bitset", cache_device_name(cache));
2723
2724	save_stats(cache);
2725
2726	r3 = write_hints(cache);
2727	if (r3)
2728		DMERR("%s: could not write hints", cache_device_name(cache));
2729
2730	/*
2731	 * If writing the above metadata failed, we still commit, but don't
2732	 * set the clean shutdown flag.  This will effectively force every
2733	 * dirty bit to be set on reload.
2734	 */
2735	r4 = commit(cache, !r1 && !r2 && !r3);
2736	if (r4)
2737		DMERR("%s: could not write cache metadata", cache_device_name(cache));
2738
2739	return !r1 && !r2 && !r3 && !r4;
2740}
2741
2742static void cache_postsuspend(struct dm_target *ti)
2743{
2744	struct cache *cache = ti->private;
2745
2746	prevent_background_work(cache);
2747	BUG_ON(atomic_read(&cache->nr_io_migrations));
2748
2749	cancel_delayed_work_sync(&cache->waker);
2750	drain_workqueue(cache->wq);
2751	WARN_ON(cache->tracker.in_flight);
2752
2753	/*
2754	 * If it's a flush suspend there won't be any deferred bios, so this
2755	 * call is harmless.
2756	 */
2757	requeue_deferred_bios(cache);
2758
2759	if (get_cache_mode(cache) == CM_WRITE)
2760		(void) sync_metadata(cache);
2761}
2762
2763static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2764			bool dirty, uint32_t hint, bool hint_valid)
2765{
2766	struct cache *cache = context;
2767
2768	if (dirty) {
2769		set_bit(from_cblock(cblock), cache->dirty_bitset);
2770		atomic_inc(&cache->nr_dirty);
2771	} else
2772		clear_bit(from_cblock(cblock), cache->dirty_bitset);
2773
2774	return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2775}
2776
2777/*
2778 * The discard block size in the on disk metadata is not
2779 * necessarily the same as we're currently using.  So we have to
2780 * be careful to only set the discarded attribute if we know it
2781 * covers a complete block of the new size.
2782 */
2783struct discard_load_info {
2784	struct cache *cache;
2785
2786	/*
2787	 * These blocks are sized using the on disk dblock size, rather
2788	 * than the current one.
2789	 */
2790	dm_block_t block_size;
2791	dm_block_t discard_begin, discard_end;
2792};
2793
2794static void discard_load_info_init(struct cache *cache,
2795				   struct discard_load_info *li)
2796{
2797	li->cache = cache;
2798	li->discard_begin = li->discard_end = 0;
2799}
2800
2801static void set_discard_range(struct discard_load_info *li)
2802{
2803	sector_t b, e;
2804
2805	if (li->discard_begin == li->discard_end)
2806		return;
2807
2808	/*
2809	 * Convert to sectors.
2810	 */
2811	b = li->discard_begin * li->block_size;
2812	e = li->discard_end * li->block_size;
2813
2814	/*
2815	 * Then convert back to the current dblock size.
2816	 */
2817	b = dm_sector_div_up(b, li->cache->discard_block_size);
2818	sector_div(e, li->cache->discard_block_size);
2819
2820	/*
2821	 * The origin may have shrunk, so we need to check we're still in
2822	 * bounds.
2823	 */
2824	if (e > from_dblock(li->cache->discard_nr_blocks))
2825		e = from_dblock(li->cache->discard_nr_blocks);
2826
2827	for (; b < e; b++)
2828		set_discard(li->cache, to_dblock(b));
2829}
2830
2831static int load_discard(void *context, sector_t discard_block_size,
2832			dm_dblock_t dblock, bool discard)
2833{
2834	struct discard_load_info *li = context;
2835
2836	li->block_size = discard_block_size;
2837
2838	if (discard) {
2839		if (from_dblock(dblock) == li->discard_end)
2840			/*
2841			 * We're already in a discard range, just extend it.
2842			 */
2843			li->discard_end = li->discard_end + 1ULL;
2844
2845		else {
2846			/*
2847			 * Emit the old range and start a new one.
2848			 */
2849			set_discard_range(li);
2850			li->discard_begin = from_dblock(dblock);
2851			li->discard_end = li->discard_begin + 1ULL;
2852		}
2853	} else {
2854		set_discard_range(li);
2855		li->discard_begin = li->discard_end = 0;
2856	}
2857
2858	return 0;
2859}
2860
2861static dm_cblock_t get_cache_dev_size(struct cache *cache)
2862{
2863	sector_t size = get_dev_size(cache->cache_dev);
2864	(void) sector_div(size, cache->sectors_per_block);
2865	return to_cblock(size);
2866}
2867
2868static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2869{
2870	if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2871		if (cache->sized) {
2872			DMERR("%s: unable to extend cache due to missing cache table reload",
2873			      cache_device_name(cache));
2874			return false;
2875		}
2876	}
2877
2878	/*
2879	 * We can't drop a dirty block when shrinking the cache.
2880	 */
2881	while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2882		new_size = to_cblock(from_cblock(new_size) + 1);
2883		if (is_dirty(cache, new_size)) {
2884			DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2885			      cache_device_name(cache),
2886			      (unsigned long long) from_cblock(new_size));
2887			return false;
2888		}
2889	}
2890
2891	return true;
2892}
2893
2894static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2895{
2896	int r;
2897
2898	r = dm_cache_resize(cache->cmd, new_size);
2899	if (r) {
2900		DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2901		metadata_operation_failed(cache, "dm_cache_resize", r);
2902		return r;
2903	}
2904
2905	set_cache_size(cache, new_size);
2906
2907	return 0;
2908}
2909
2910static int cache_preresume(struct dm_target *ti)
2911{
2912	int r = 0;
2913	struct cache *cache = ti->private;
2914	dm_cblock_t csize = get_cache_dev_size(cache);
2915
2916	/*
2917	 * Check to see if the cache has resized.
2918	 */
2919	if (!cache->sized) {
2920		r = resize_cache_dev(cache, csize);
2921		if (r)
2922			return r;
2923
2924		cache->sized = true;
2925
2926	} else if (csize != cache->cache_size) {
2927		if (!can_resize(cache, csize))
2928			return -EINVAL;
2929
2930		r = resize_cache_dev(cache, csize);
2931		if (r)
2932			return r;
2933	}
2934
2935	if (!cache->loaded_mappings) {
2936		r = dm_cache_load_mappings(cache->cmd, cache->policy,
2937					   load_mapping, cache);
2938		if (r) {
2939			DMERR("%s: could not load cache mappings", cache_device_name(cache));
2940			metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2941			return r;
2942		}
2943
2944		cache->loaded_mappings = true;
2945	}
2946
2947	if (!cache->loaded_discards) {
2948		struct discard_load_info li;
2949
2950		/*
2951		 * The discard bitset could have been resized, or the
2952		 * discard block size changed.  To be safe we start by
2953		 * setting every dblock to not discarded.
2954		 */
2955		clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2956
2957		discard_load_info_init(cache, &li);
2958		r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2959		if (r) {
2960			DMERR("%s: could not load origin discards", cache_device_name(cache));
2961			metadata_operation_failed(cache, "dm_cache_load_discards", r);
2962			return r;
2963		}
2964		set_discard_range(&li);
2965
2966		cache->loaded_discards = true;
2967	}
2968
2969	return r;
2970}
2971
2972static void cache_resume(struct dm_target *ti)
2973{
2974	struct cache *cache = ti->private;
2975
2976	cache->need_tick_bio = true;
2977	allow_background_work(cache);
2978	do_waker(&cache->waker.work);
2979}
2980
2981static void emit_flags(struct cache *cache, char *result,
2982		       unsigned maxlen, ssize_t *sz_ptr)
2983{
2984	ssize_t sz = *sz_ptr;
2985	struct cache_features *cf = &cache->features;
2986	unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
2987
2988	DMEMIT("%u ", count);
2989
2990	if (cf->metadata_version == 2)
2991		DMEMIT("metadata2 ");
2992
2993	if (writethrough_mode(cache))
2994		DMEMIT("writethrough ");
2995
2996	else if (passthrough_mode(cache))
2997		DMEMIT("passthrough ");
2998
2999	else if (writeback_mode(cache))
3000		DMEMIT("writeback ");
3001
3002	else {
3003		DMEMIT("unknown ");
3004		DMERR("%s: internal error: unknown io mode: %d",
3005		      cache_device_name(cache), (int) cf->io_mode);
3006	}
3007
3008	if (!cf->discard_passdown)
3009		DMEMIT("no_discard_passdown ");
3010
3011	*sz_ptr = sz;
3012}
3013
3014/*
3015 * Status format:
3016 *
3017 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3018 * <cache block size> <#used cache blocks>/<#total cache blocks>
3019 * <#read hits> <#read misses> <#write hits> <#write misses>
3020 * <#demotions> <#promotions> <#dirty>
3021 * <#features> <features>*
3022 * <#core args> <core args>
3023 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3024 */
3025static void cache_status(struct dm_target *ti, status_type_t type,
3026			 unsigned status_flags, char *result, unsigned maxlen)
3027{
3028	int r = 0;
3029	unsigned i;
3030	ssize_t sz = 0;
3031	dm_block_t nr_free_blocks_metadata = 0;
3032	dm_block_t nr_blocks_metadata = 0;
3033	char buf[BDEVNAME_SIZE];
3034	struct cache *cache = ti->private;
3035	dm_cblock_t residency;
3036	bool needs_check;
3037
3038	switch (type) {
3039	case STATUSTYPE_INFO:
3040		if (get_cache_mode(cache) == CM_FAIL) {
3041			DMEMIT("Fail");
3042			break;
3043		}
3044
3045		/* Commit to ensure statistics aren't out-of-date */
3046		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3047			(void) commit(cache, false);
3048
3049		r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3050		if (r) {
3051			DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3052			      cache_device_name(cache), r);
3053			goto err;
3054		}
3055
3056		r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3057		if (r) {
3058			DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3059			      cache_device_name(cache), r);
3060			goto err;
3061		}
3062
3063		residency = policy_residency(cache->policy);
3064
3065		DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3066		       (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3067		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3068		       (unsigned long long)nr_blocks_metadata,
3069		       (unsigned long long)cache->sectors_per_block,
3070		       (unsigned long long) from_cblock(residency),
3071		       (unsigned long long) from_cblock(cache->cache_size),
3072		       (unsigned) atomic_read(&cache->stats.read_hit),
3073		       (unsigned) atomic_read(&cache->stats.read_miss),
3074		       (unsigned) atomic_read(&cache->stats.write_hit),
3075		       (unsigned) atomic_read(&cache->stats.write_miss),
3076		       (unsigned) atomic_read(&cache->stats.demotion),
3077		       (unsigned) atomic_read(&cache->stats.promotion),
3078		       (unsigned long) atomic_read(&cache->nr_dirty));
3079
3080		emit_flags(cache, result, maxlen, &sz);
3081
3082		DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3083
3084		DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3085		if (sz < maxlen) {
3086			r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3087			if (r)
3088				DMERR("%s: policy_emit_config_values returned %d",
3089				      cache_device_name(cache), r);
3090		}
3091
3092		if (get_cache_mode(cache) == CM_READ_ONLY)
3093			DMEMIT("ro ");
3094		else
3095			DMEMIT("rw ");
3096
3097		r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3098
3099		if (r || needs_check)
3100			DMEMIT("needs_check ");
3101		else
3102			DMEMIT("- ");
3103
3104		break;
3105
3106	case STATUSTYPE_TABLE:
3107		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3108		DMEMIT("%s ", buf);
3109		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3110		DMEMIT("%s ", buf);
3111		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3112		DMEMIT("%s", buf);
3113
3114		for (i = 0; i < cache->nr_ctr_args - 1; i++)
3115			DMEMIT(" %s", cache->ctr_args[i]);
3116		if (cache->nr_ctr_args)
3117			DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3118		break;
3119
3120	case STATUSTYPE_IMA:
3121		DMEMIT_TARGET_NAME_VERSION(ti->type);
3122		if (get_cache_mode(cache) == CM_FAIL)
3123			DMEMIT(",metadata_mode=fail");
3124		else if (get_cache_mode(cache) == CM_READ_ONLY)
3125			DMEMIT(",metadata_mode=ro");
3126		else
3127			DMEMIT(",metadata_mode=rw");
3128
3129		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3130		DMEMIT(",cache_metadata_device=%s", buf);
3131		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3132		DMEMIT(",cache_device=%s", buf);
3133		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3134		DMEMIT(",cache_origin_device=%s", buf);
3135		DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3136		DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3137		DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3138		DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3139		DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3140		DMEMIT(";");
3141		break;
3142	}
3143
3144	return;
3145
3146err:
3147	DMEMIT("Error");
3148}
3149
3150/*
3151 * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
3152 * the one-past-the-end value.
3153 */
3154struct cblock_range {
3155	dm_cblock_t begin;
3156	dm_cblock_t end;
3157};
3158
3159/*
3160 * A cache block range can take two forms:
3161 *
3162 * i) A single cblock, eg. '3456'
3163 * ii) A begin and end cblock with a dash between, eg. 123-234
3164 */
3165static int parse_cblock_range(struct cache *cache, const char *str,
3166			      struct cblock_range *result)
3167{
3168	char dummy;
3169	uint64_t b, e;
3170	int r;
3171
3172	/*
3173	 * Try and parse form (ii) first.
3174	 */
3175	r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3176	if (r < 0)
3177		return r;
3178
3179	if (r == 2) {
3180		result->begin = to_cblock(b);
3181		result->end = to_cblock(e);
3182		return 0;
3183	}
3184
3185	/*
3186	 * That didn't work, try form (i).
3187	 */
3188	r = sscanf(str, "%llu%c", &b, &dummy);
3189	if (r < 0)
3190		return r;
3191
3192	if (r == 1) {
3193		result->begin = to_cblock(b);
3194		result->end = to_cblock(from_cblock(result->begin) + 1u);
3195		return 0;
3196	}
3197
3198	DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3199	return -EINVAL;
3200}
3201
3202static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3203{
3204	uint64_t b = from_cblock(range->begin);
3205	uint64_t e = from_cblock(range->end);
3206	uint64_t n = from_cblock(cache->cache_size);
3207
3208	if (b >= n) {
3209		DMERR("%s: begin cblock out of range: %llu >= %llu",
3210		      cache_device_name(cache), b, n);
3211		return -EINVAL;
3212	}
3213
3214	if (e > n) {
3215		DMERR("%s: end cblock out of range: %llu > %llu",
3216		      cache_device_name(cache), e, n);
3217		return -EINVAL;
3218	}
3219
3220	if (b >= e) {
3221		DMERR("%s: invalid cblock range: %llu >= %llu",
3222		      cache_device_name(cache), b, e);
3223		return -EINVAL;
3224	}
3225
3226	return 0;
3227}
3228
3229static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3230{
3231	return to_cblock(from_cblock(b) + 1);
3232}
3233
3234static int request_invalidation(struct cache *cache, struct cblock_range *range)
3235{
3236	int r = 0;
3237
3238	/*
3239	 * We don't need to do any locking here because we know we're in
3240	 * passthrough mode.  There's is potential for a race between an
3241	 * invalidation triggered by an io and an invalidation message.  This
3242	 * is harmless, we must not worry if the policy call fails.
3243	 */
3244	while (range->begin != range->end) {
3245		r = invalidate_cblock(cache, range->begin);
3246		if (r)
3247			return r;
3248
3249		range->begin = cblock_succ(range->begin);
3250	}
3251
3252	cache->commit_requested = true;
3253	return r;
3254}
3255
3256static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3257					      const char **cblock_ranges)
3258{
3259	int r = 0;
3260	unsigned i;
3261	struct cblock_range range;
3262
3263	if (!passthrough_mode(cache)) {
3264		DMERR("%s: cache has to be in passthrough mode for invalidation",
3265		      cache_device_name(cache));
3266		return -EPERM;
3267	}
3268
3269	for (i = 0; i < count; i++) {
3270		r = parse_cblock_range(cache, cblock_ranges[i], &range);
3271		if (r)
3272			break;
3273
3274		r = validate_cblock_range(cache, &range);
3275		if (r)
3276			break;
3277
3278		/*
3279		 * Pass begin and end origin blocks to the worker and wake it.
3280		 */
3281		r = request_invalidation(cache, &range);
3282		if (r)
3283			break;
3284	}
3285
3286	return r;
3287}
3288
3289/*
3290 * Supports
3291 *	"<key> <value>"
3292 * and
3293 *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3294 *
3295 * The key migration_threshold is supported by the cache target core.
3296 */
3297static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3298			 char *result, unsigned maxlen)
3299{
3300	struct cache *cache = ti->private;
3301
3302	if (!argc)
3303		return -EINVAL;
3304
3305	if (get_cache_mode(cache) >= CM_READ_ONLY) {
3306		DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3307		      cache_device_name(cache));
3308		return -EOPNOTSUPP;
3309	}
3310
3311	if (!strcasecmp(argv[0], "invalidate_cblocks"))
3312		return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3313
3314	if (argc != 2)
3315		return -EINVAL;
3316
3317	return set_config_value(cache, argv[0], argv[1]);
3318}
3319
3320static int cache_iterate_devices(struct dm_target *ti,
3321				 iterate_devices_callout_fn fn, void *data)
3322{
3323	int r = 0;
3324	struct cache *cache = ti->private;
3325
3326	r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3327	if (!r)
3328		r = fn(ti, cache->origin_dev, 0, ti->len, data);
3329
3330	return r;
3331}
3332
3333/*
3334 * If discard_passdown was enabled verify that the origin device
3335 * supports discards.  Disable discard_passdown if not.
3336 */
3337static void disable_passdown_if_not_supported(struct cache *cache)
3338{
3339	struct block_device *origin_bdev = cache->origin_dev->bdev;
3340	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3341	const char *reason = NULL;
3342
3343	if (!cache->features.discard_passdown)
3344		return;
3345
3346	if (!bdev_max_discard_sectors(origin_bdev))
3347		reason = "discard unsupported";
3348
3349	else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3350		reason = "max discard sectors smaller than a block";
3351
3352	if (reason) {
3353		DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3354		       origin_bdev, reason);
3355		cache->features.discard_passdown = false;
3356	}
3357}
3358
3359static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3360{
3361	struct block_device *origin_bdev = cache->origin_dev->bdev;
3362	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3363
3364	if (!cache->features.discard_passdown) {
3365		/* No passdown is done so setting own virtual limits */
3366		limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3367						    cache->origin_sectors);
3368		limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3369		return;
3370	}
3371
3372	/*
3373	 * cache_iterate_devices() is stacking both origin and fast device limits
3374	 * but discards aren't passed to fast device, so inherit origin's limits.
3375	 */
3376	limits->max_discard_sectors = origin_limits->max_discard_sectors;
3377	limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3378	limits->discard_granularity = origin_limits->discard_granularity;
3379	limits->discard_alignment = origin_limits->discard_alignment;
3380	limits->discard_misaligned = origin_limits->discard_misaligned;
3381}
3382
3383static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3384{
3385	struct cache *cache = ti->private;
3386	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3387
3388	/*
3389	 * If the system-determined stacked limits are compatible with the
3390	 * cache's blocksize (io_opt is a factor) do not override them.
3391	 */
3392	if (io_opt_sectors < cache->sectors_per_block ||
3393	    do_div(io_opt_sectors, cache->sectors_per_block)) {
3394		blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3395		blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3396	}
3397
3398	disable_passdown_if_not_supported(cache);
3399	set_discard_limits(cache, limits);
3400}
3401
3402/*----------------------------------------------------------------*/
3403
3404static struct target_type cache_target = {
3405	.name = "cache",
3406	.version = {2, 2, 0},
3407	.module = THIS_MODULE,
3408	.ctr = cache_ctr,
3409	.dtr = cache_dtr,
3410	.map = cache_map,
3411	.end_io = cache_end_io,
3412	.postsuspend = cache_postsuspend,
3413	.preresume = cache_preresume,
3414	.resume = cache_resume,
3415	.status = cache_status,
3416	.message = cache_message,
3417	.iterate_devices = cache_iterate_devices,
3418	.io_hints = cache_io_hints,
3419};
3420
3421static int __init dm_cache_init(void)
3422{
3423	int r;
3424
3425	migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3426	if (!migration_cache)
3427		return -ENOMEM;
3428
3429	r = dm_register_target(&cache_target);
3430	if (r) {
3431		DMERR("cache target registration failed: %d", r);
3432		kmem_cache_destroy(migration_cache);
3433		return r;
3434	}
3435
3436	return 0;
3437}
3438
3439static void __exit dm_cache_exit(void)
3440{
3441	dm_unregister_target(&cache_target);
3442	kmem_cache_destroy(migration_cache);
3443}
3444
3445module_init(dm_cache_init);
3446module_exit(dm_cache_exit);
3447
3448MODULE_DESCRIPTION(DM_NAME " cache target");
3449MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3450MODULE_LICENSE("GPL");
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2012 Red Hat. All rights reserved.
   4 *
   5 * This file is released under the GPL.
   6 */
   7
   8#include "dm.h"
   9#include "dm-bio-prison-v2.h"
  10#include "dm-bio-record.h"
  11#include "dm-cache-metadata.h"
  12#include "dm-io-tracker.h"
  13
  14#include <linux/dm-io.h>
  15#include <linux/dm-kcopyd.h>
  16#include <linux/jiffies.h>
  17#include <linux/init.h>
  18#include <linux/mempool.h>
  19#include <linux/module.h>
  20#include <linux/rwsem.h>
  21#include <linux/slab.h>
  22#include <linux/vmalloc.h>
  23
  24#define DM_MSG_PREFIX "cache"
  25
  26DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  27	"A percentage of time allocated for copying to and/or from cache");
  28
  29/*----------------------------------------------------------------*/
  30
  31/*
  32 * Glossary:
  33 *
  34 * oblock: index of an origin block
  35 * cblock: index of a cache block
  36 * promotion: movement of a block from origin to cache
  37 * demotion: movement of a block from cache to origin
  38 * migration: movement of a block between the origin and cache device,
  39 *	      either direction
  40 */
  41
  42/*----------------------------------------------------------------*/
  43
  44/*
  45 * Represents a chunk of future work.  'input' allows continuations to pass
  46 * values between themselves, typically error values.
  47 */
  48struct continuation {
  49	struct work_struct ws;
  50	blk_status_t input;
  51};
  52
  53static inline void init_continuation(struct continuation *k,
  54				     void (*fn)(struct work_struct *))
  55{
  56	INIT_WORK(&k->ws, fn);
  57	k->input = 0;
  58}
  59
  60static inline void queue_continuation(struct workqueue_struct *wq,
  61				      struct continuation *k)
  62{
  63	queue_work(wq, &k->ws);
  64}
  65
  66/*----------------------------------------------------------------*/
  67
  68/*
  69 * The batcher collects together pieces of work that need a particular
  70 * operation to occur before they can proceed (typically a commit).
  71 */
  72struct batcher {
  73	/*
  74	 * The operation that everyone is waiting for.
  75	 */
  76	blk_status_t (*commit_op)(void *context);
  77	void *commit_context;
  78
  79	/*
  80	 * This is how bios should be issued once the commit op is complete
  81	 * (accounted_request).
  82	 */
  83	void (*issue_op)(struct bio *bio, void *context);
  84	void *issue_context;
  85
  86	/*
  87	 * Queued work gets put on here after commit.
  88	 */
  89	struct workqueue_struct *wq;
  90
  91	spinlock_t lock;
  92	struct list_head work_items;
  93	struct bio_list bios;
  94	struct work_struct commit_work;
  95
  96	bool commit_scheduled;
  97};
  98
  99static void __commit(struct work_struct *_ws)
 100{
 101	struct batcher *b = container_of(_ws, struct batcher, commit_work);
 102	blk_status_t r;
 103	struct list_head work_items;
 104	struct work_struct *ws, *tmp;
 105	struct continuation *k;
 106	struct bio *bio;
 107	struct bio_list bios;
 108
 109	INIT_LIST_HEAD(&work_items);
 110	bio_list_init(&bios);
 111
 112	/*
 113	 * We have to grab these before the commit_op to avoid a race
 114	 * condition.
 115	 */
 116	spin_lock_irq(&b->lock);
 117	list_splice_init(&b->work_items, &work_items);
 118	bio_list_merge(&bios, &b->bios);
 119	bio_list_init(&b->bios);
 120	b->commit_scheduled = false;
 121	spin_unlock_irq(&b->lock);
 122
 123	r = b->commit_op(b->commit_context);
 124
 125	list_for_each_entry_safe(ws, tmp, &work_items, entry) {
 126		k = container_of(ws, struct continuation, ws);
 127		k->input = r;
 128		INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
 129		queue_work(b->wq, ws);
 130	}
 131
 132	while ((bio = bio_list_pop(&bios))) {
 133		if (r) {
 134			bio->bi_status = r;
 135			bio_endio(bio);
 136		} else
 137			b->issue_op(bio, b->issue_context);
 138	}
 139}
 140
 141static void batcher_init(struct batcher *b,
 142			 blk_status_t (*commit_op)(void *),
 143			 void *commit_context,
 144			 void (*issue_op)(struct bio *bio, void *),
 145			 void *issue_context,
 146			 struct workqueue_struct *wq)
 147{
 148	b->commit_op = commit_op;
 149	b->commit_context = commit_context;
 150	b->issue_op = issue_op;
 151	b->issue_context = issue_context;
 152	b->wq = wq;
 153
 154	spin_lock_init(&b->lock);
 155	INIT_LIST_HEAD(&b->work_items);
 156	bio_list_init(&b->bios);
 157	INIT_WORK(&b->commit_work, __commit);
 158	b->commit_scheduled = false;
 159}
 160
 161static void async_commit(struct batcher *b)
 162{
 163	queue_work(b->wq, &b->commit_work);
 164}
 165
 166static void continue_after_commit(struct batcher *b, struct continuation *k)
 167{
 168	bool commit_scheduled;
 169
 170	spin_lock_irq(&b->lock);
 171	commit_scheduled = b->commit_scheduled;
 172	list_add_tail(&k->ws.entry, &b->work_items);
 173	spin_unlock_irq(&b->lock);
 174
 175	if (commit_scheduled)
 176		async_commit(b);
 177}
 178
 179/*
 180 * Bios are errored if commit failed.
 181 */
 182static void issue_after_commit(struct batcher *b, struct bio *bio)
 183{
 184	bool commit_scheduled;
 185
 186	spin_lock_irq(&b->lock);
 187	commit_scheduled = b->commit_scheduled;
 188	bio_list_add(&b->bios, bio);
 189	spin_unlock_irq(&b->lock);
 190
 191	if (commit_scheduled)
 192		async_commit(b);
 193}
 194
 195/*
 196 * Call this if some urgent work is waiting for the commit to complete.
 197 */
 198static void schedule_commit(struct batcher *b)
 199{
 200	bool immediate;
 201
 202	spin_lock_irq(&b->lock);
 203	immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
 204	b->commit_scheduled = true;
 205	spin_unlock_irq(&b->lock);
 206
 207	if (immediate)
 208		async_commit(b);
 209}
 210
 211/*
 212 * There are a couple of places where we let a bio run, but want to do some
 213 * work before calling its endio function.  We do this by temporarily
 214 * changing the endio fn.
 215 */
 216struct dm_hook_info {
 217	bio_end_io_t *bi_end_io;
 218};
 219
 220static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 221			bio_end_io_t *bi_end_io, void *bi_private)
 222{
 223	h->bi_end_io = bio->bi_end_io;
 224
 225	bio->bi_end_io = bi_end_io;
 226	bio->bi_private = bi_private;
 227}
 228
 229static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 230{
 231	bio->bi_end_io = h->bi_end_io;
 232}
 233
 234/*----------------------------------------------------------------*/
 235
 236#define MIGRATION_POOL_SIZE 128
 237#define COMMIT_PERIOD HZ
 238#define MIGRATION_COUNT_WINDOW 10
 239
 240/*
 241 * The block size of the device holding cache data must be
 242 * between 32KB and 1GB.
 243 */
 244#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 245#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 246
 247enum cache_metadata_mode {
 248	CM_WRITE,		/* metadata may be changed */
 249	CM_READ_ONLY,		/* metadata may not be changed */
 250	CM_FAIL
 251};
 252
 253enum cache_io_mode {
 254	/*
 255	 * Data is written to cached blocks only.  These blocks are marked
 256	 * dirty.  If you lose the cache device you will lose data.
 257	 * Potential performance increase for both reads and writes.
 258	 */
 259	CM_IO_WRITEBACK,
 260
 261	/*
 262	 * Data is written to both cache and origin.  Blocks are never
 263	 * dirty.  Potential performance benfit for reads only.
 264	 */
 265	CM_IO_WRITETHROUGH,
 266
 267	/*
 268	 * A degraded mode useful for various cache coherency situations
 269	 * (eg, rolling back snapshots).  Reads and writes always go to the
 270	 * origin.  If a write goes to a cached oblock, then the cache
 271	 * block is invalidated.
 272	 */
 273	CM_IO_PASSTHROUGH
 274};
 275
 276struct cache_features {
 277	enum cache_metadata_mode mode;
 278	enum cache_io_mode io_mode;
 279	unsigned int metadata_version;
 280	bool discard_passdown:1;
 281};
 282
 283struct cache_stats {
 284	atomic_t read_hit;
 285	atomic_t read_miss;
 286	atomic_t write_hit;
 287	atomic_t write_miss;
 288	atomic_t demotion;
 289	atomic_t promotion;
 290	atomic_t writeback;
 291	atomic_t copies_avoided;
 292	atomic_t cache_cell_clash;
 293	atomic_t commit_count;
 294	atomic_t discard_count;
 295};
 296
 297struct cache {
 298	struct dm_target *ti;
 299	spinlock_t lock;
 300
 301	/*
 302	 * Fields for converting from sectors to blocks.
 303	 */
 304	int sectors_per_block_shift;
 305	sector_t sectors_per_block;
 306
 307	struct dm_cache_metadata *cmd;
 308
 309	/*
 310	 * Metadata is written to this device.
 311	 */
 312	struct dm_dev *metadata_dev;
 313
 314	/*
 315	 * The slower of the two data devices.  Typically a spindle.
 316	 */
 317	struct dm_dev *origin_dev;
 318
 319	/*
 320	 * The faster of the two data devices.  Typically an SSD.
 321	 */
 322	struct dm_dev *cache_dev;
 323
 324	/*
 325	 * Size of the origin device in _complete_ blocks and native sectors.
 326	 */
 327	dm_oblock_t origin_blocks;
 328	sector_t origin_sectors;
 329
 330	/*
 331	 * Size of the cache device in blocks.
 332	 */
 333	dm_cblock_t cache_size;
 334
 335	/*
 336	 * Invalidation fields.
 337	 */
 338	spinlock_t invalidation_lock;
 339	struct list_head invalidation_requests;
 340
 341	sector_t migration_threshold;
 342	wait_queue_head_t migration_wait;
 343	atomic_t nr_allocated_migrations;
 344
 345	/*
 346	 * The number of in flight migrations that are performing
 347	 * background io. eg, promotion, writeback.
 348	 */
 349	atomic_t nr_io_migrations;
 350
 351	struct bio_list deferred_bios;
 352
 353	struct rw_semaphore quiesce_lock;
 354
 355	/*
 356	 * origin_blocks entries, discarded if set.
 357	 */
 358	dm_dblock_t discard_nr_blocks;
 359	unsigned long *discard_bitset;
 360	uint32_t discard_block_size; /* a power of 2 times sectors per block */
 361
 362	/*
 363	 * Rather than reconstructing the table line for the status we just
 364	 * save it and regurgitate.
 365	 */
 366	unsigned int nr_ctr_args;
 367	const char **ctr_args;
 368
 369	struct dm_kcopyd_client *copier;
 370	struct work_struct deferred_bio_worker;
 371	struct work_struct migration_worker;
 372	struct workqueue_struct *wq;
 373	struct delayed_work waker;
 374	struct dm_bio_prison_v2 *prison;
 375
 376	/*
 377	 * cache_size entries, dirty if set
 378	 */
 379	unsigned long *dirty_bitset;
 380	atomic_t nr_dirty;
 381
 382	unsigned int policy_nr_args;
 383	struct dm_cache_policy *policy;
 384
 385	/*
 386	 * Cache features such as write-through.
 387	 */
 388	struct cache_features features;
 389
 390	struct cache_stats stats;
 391
 392	bool need_tick_bio:1;
 393	bool sized:1;
 394	bool invalidate:1;
 395	bool commit_requested:1;
 396	bool loaded_mappings:1;
 397	bool loaded_discards:1;
 398
 399	struct rw_semaphore background_work_lock;
 400
 401	struct batcher committer;
 402	struct work_struct commit_ws;
 403
 404	struct dm_io_tracker tracker;
 405
 406	mempool_t migration_pool;
 407
 408	struct bio_set bs;
 409};
 410
 411struct per_bio_data {
 412	bool tick:1;
 413	unsigned int req_nr:2;
 414	struct dm_bio_prison_cell_v2 *cell;
 415	struct dm_hook_info hook_info;
 416	sector_t len;
 417};
 418
 419struct dm_cache_migration {
 420	struct continuation k;
 421	struct cache *cache;
 422
 423	struct policy_work *op;
 424	struct bio *overwrite_bio;
 425	struct dm_bio_prison_cell_v2 *cell;
 426
 427	dm_cblock_t invalidate_cblock;
 428	dm_oblock_t invalidate_oblock;
 429};
 430
 431/*----------------------------------------------------------------*/
 432
 433static bool writethrough_mode(struct cache *cache)
 434{
 435	return cache->features.io_mode == CM_IO_WRITETHROUGH;
 436}
 437
 438static bool writeback_mode(struct cache *cache)
 439{
 440	return cache->features.io_mode == CM_IO_WRITEBACK;
 441}
 442
 443static inline bool passthrough_mode(struct cache *cache)
 444{
 445	return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
 446}
 447
 448/*----------------------------------------------------------------*/
 449
 450static void wake_deferred_bio_worker(struct cache *cache)
 451{
 452	queue_work(cache->wq, &cache->deferred_bio_worker);
 453}
 454
 455static void wake_migration_worker(struct cache *cache)
 456{
 457	if (passthrough_mode(cache))
 458		return;
 459
 460	queue_work(cache->wq, &cache->migration_worker);
 461}
 462
 463/*----------------------------------------------------------------*/
 464
 465static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
 466{
 467	return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
 468}
 469
 470static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
 471{
 472	dm_bio_prison_free_cell_v2(cache->prison, cell);
 473}
 474
 475static struct dm_cache_migration *alloc_migration(struct cache *cache)
 476{
 477	struct dm_cache_migration *mg;
 478
 479	mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
 480
 481	memset(mg, 0, sizeof(*mg));
 482
 483	mg->cache = cache;
 484	atomic_inc(&cache->nr_allocated_migrations);
 485
 486	return mg;
 487}
 488
 489static void free_migration(struct dm_cache_migration *mg)
 490{
 491	struct cache *cache = mg->cache;
 492
 493	if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 494		wake_up(&cache->migration_wait);
 495
 496	mempool_free(mg, &cache->migration_pool);
 497}
 498
 499/*----------------------------------------------------------------*/
 500
 501static inline dm_oblock_t oblock_succ(dm_oblock_t b)
 502{
 503	return to_oblock(from_oblock(b) + 1ull);
 504}
 505
 506static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
 507{
 508	key->virtual = 0;
 509	key->dev = 0;
 510	key->block_begin = from_oblock(begin);
 511	key->block_end = from_oblock(end);
 512}
 513
 514/*
 515 * We have two lock levels.  Level 0, which is used to prevent WRITEs, and
 516 * level 1 which prevents *both* READs and WRITEs.
 517 */
 518#define WRITE_LOCK_LEVEL 0
 519#define READ_WRITE_LOCK_LEVEL 1
 520
 521static unsigned int lock_level(struct bio *bio)
 522{
 523	return bio_data_dir(bio) == WRITE ?
 524		WRITE_LOCK_LEVEL :
 525		READ_WRITE_LOCK_LEVEL;
 526}
 527
 528/*
 529 *--------------------------------------------------------------
 530 * Per bio data
 531 *--------------------------------------------------------------
 532 */
 533
 534static struct per_bio_data *get_per_bio_data(struct bio *bio)
 535{
 536	struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 537
 538	BUG_ON(!pb);
 539	return pb;
 540}
 541
 542static struct per_bio_data *init_per_bio_data(struct bio *bio)
 543{
 544	struct per_bio_data *pb = get_per_bio_data(bio);
 545
 546	pb->tick = false;
 547	pb->req_nr = dm_bio_get_target_bio_nr(bio);
 548	pb->cell = NULL;
 549	pb->len = 0;
 550
 551	return pb;
 552}
 553
 554/*----------------------------------------------------------------*/
 555
 556static void defer_bio(struct cache *cache, struct bio *bio)
 557{
 558	spin_lock_irq(&cache->lock);
 559	bio_list_add(&cache->deferred_bios, bio);
 560	spin_unlock_irq(&cache->lock);
 561
 562	wake_deferred_bio_worker(cache);
 563}
 564
 565static void defer_bios(struct cache *cache, struct bio_list *bios)
 566{
 567	spin_lock_irq(&cache->lock);
 568	bio_list_merge(&cache->deferred_bios, bios);
 569	bio_list_init(bios);
 570	spin_unlock_irq(&cache->lock);
 571
 572	wake_deferred_bio_worker(cache);
 573}
 574
 575/*----------------------------------------------------------------*/
 576
 577static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
 578{
 579	bool r;
 580	struct per_bio_data *pb;
 581	struct dm_cell_key_v2 key;
 582	dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 583	struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
 584
 585	cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
 586
 587	build_key(oblock, end, &key);
 588	r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
 589	if (!r) {
 590		/*
 591		 * Failed to get the lock.
 592		 */
 593		free_prison_cell(cache, cell_prealloc);
 594		return r;
 595	}
 596
 597	if (cell != cell_prealloc)
 598		free_prison_cell(cache, cell_prealloc);
 599
 600	pb = get_per_bio_data(bio);
 601	pb->cell = cell;
 602
 603	return r;
 604}
 605
 606/*----------------------------------------------------------------*/
 607
 608static bool is_dirty(struct cache *cache, dm_cblock_t b)
 609{
 610	return test_bit(from_cblock(b), cache->dirty_bitset);
 611}
 612
 613static void set_dirty(struct cache *cache, dm_cblock_t cblock)
 614{
 615	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 616		atomic_inc(&cache->nr_dirty);
 617		policy_set_dirty(cache->policy, cblock);
 618	}
 619}
 620
 621/*
 622 * These two are called when setting after migrations to force the policy
 623 * and dirty bitset to be in sync.
 624 */
 625static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
 626{
 627	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
 628		atomic_inc(&cache->nr_dirty);
 629	policy_set_dirty(cache->policy, cblock);
 630}
 631
 632static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
 633{
 634	if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 635		if (atomic_dec_return(&cache->nr_dirty) == 0)
 636			dm_table_event(cache->ti->table);
 637	}
 638
 639	policy_clear_dirty(cache->policy, cblock);
 640}
 641
 642/*----------------------------------------------------------------*/
 643
 644static bool block_size_is_power_of_two(struct cache *cache)
 645{
 646	return cache->sectors_per_block_shift >= 0;
 647}
 648
 649static dm_block_t block_div(dm_block_t b, uint32_t n)
 650{
 651	do_div(b, n);
 652
 653	return b;
 654}
 655
 656static dm_block_t oblocks_per_dblock(struct cache *cache)
 657{
 658	dm_block_t oblocks = cache->discard_block_size;
 659
 660	if (block_size_is_power_of_two(cache))
 661		oblocks >>= cache->sectors_per_block_shift;
 662	else
 663		oblocks = block_div(oblocks, cache->sectors_per_block);
 664
 665	return oblocks;
 666}
 667
 668static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 669{
 670	return to_dblock(block_div(from_oblock(oblock),
 671				   oblocks_per_dblock(cache)));
 672}
 673
 674static void set_discard(struct cache *cache, dm_dblock_t b)
 675{
 676	BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 677	atomic_inc(&cache->stats.discard_count);
 678
 679	spin_lock_irq(&cache->lock);
 680	set_bit(from_dblock(b), cache->discard_bitset);
 681	spin_unlock_irq(&cache->lock);
 682}
 683
 684static void clear_discard(struct cache *cache, dm_dblock_t b)
 685{
 686	spin_lock_irq(&cache->lock);
 687	clear_bit(from_dblock(b), cache->discard_bitset);
 688	spin_unlock_irq(&cache->lock);
 689}
 690
 691static bool is_discarded(struct cache *cache, dm_dblock_t b)
 692{
 693	int r;
 694
 695	spin_lock_irq(&cache->lock);
 696	r = test_bit(from_dblock(b), cache->discard_bitset);
 697	spin_unlock_irq(&cache->lock);
 698
 699	return r;
 700}
 701
 702static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 703{
 704	int r;
 705
 706	spin_lock_irq(&cache->lock);
 707	r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 708		     cache->discard_bitset);
 709	spin_unlock_irq(&cache->lock);
 710
 711	return r;
 712}
 713
 714/*
 715 * -------------------------------------------------------------
 716 * Remapping
 717 *--------------------------------------------------------------
 718 */
 719static void remap_to_origin(struct cache *cache, struct bio *bio)
 720{
 721	bio_set_dev(bio, cache->origin_dev->bdev);
 722}
 723
 724static void remap_to_cache(struct cache *cache, struct bio *bio,
 725			   dm_cblock_t cblock)
 726{
 727	sector_t bi_sector = bio->bi_iter.bi_sector;
 728	sector_t block = from_cblock(cblock);
 729
 730	bio_set_dev(bio, cache->cache_dev->bdev);
 731	if (!block_size_is_power_of_two(cache))
 732		bio->bi_iter.bi_sector =
 733			(block * cache->sectors_per_block) +
 734			sector_div(bi_sector, cache->sectors_per_block);
 735	else
 736		bio->bi_iter.bi_sector =
 737			(block << cache->sectors_per_block_shift) |
 738			(bi_sector & (cache->sectors_per_block - 1));
 739}
 740
 741static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 742{
 743	struct per_bio_data *pb;
 744
 745	spin_lock_irq(&cache->lock);
 746	if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
 747	    bio_op(bio) != REQ_OP_DISCARD) {
 748		pb = get_per_bio_data(bio);
 749		pb->tick = true;
 750		cache->need_tick_bio = false;
 751	}
 752	spin_unlock_irq(&cache->lock);
 753}
 754
 755static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 756					  dm_oblock_t oblock)
 757{
 758	// FIXME: check_if_tick_bio_needed() is called way too much through this interface
 759	check_if_tick_bio_needed(cache, bio);
 760	remap_to_origin(cache, bio);
 761	if (bio_data_dir(bio) == WRITE)
 762		clear_discard(cache, oblock_to_dblock(cache, oblock));
 763}
 764
 765static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 766				 dm_oblock_t oblock, dm_cblock_t cblock)
 767{
 768	check_if_tick_bio_needed(cache, bio);
 769	remap_to_cache(cache, bio, cblock);
 770	if (bio_data_dir(bio) == WRITE) {
 771		set_dirty(cache, cblock);
 772		clear_discard(cache, oblock_to_dblock(cache, oblock));
 773	}
 774}
 775
 776static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 777{
 778	sector_t block_nr = bio->bi_iter.bi_sector;
 779
 780	if (!block_size_is_power_of_two(cache))
 781		(void) sector_div(block_nr, cache->sectors_per_block);
 782	else
 783		block_nr >>= cache->sectors_per_block_shift;
 784
 785	return to_oblock(block_nr);
 786}
 787
 788static bool accountable_bio(struct cache *cache, struct bio *bio)
 789{
 790	return bio_op(bio) != REQ_OP_DISCARD;
 791}
 792
 793static void accounted_begin(struct cache *cache, struct bio *bio)
 794{
 795	struct per_bio_data *pb;
 796
 797	if (accountable_bio(cache, bio)) {
 798		pb = get_per_bio_data(bio);
 799		pb->len = bio_sectors(bio);
 800		dm_iot_io_begin(&cache->tracker, pb->len);
 801	}
 802}
 803
 804static void accounted_complete(struct cache *cache, struct bio *bio)
 805{
 806	struct per_bio_data *pb = get_per_bio_data(bio);
 807
 808	dm_iot_io_end(&cache->tracker, pb->len);
 809}
 810
 811static void accounted_request(struct cache *cache, struct bio *bio)
 812{
 813	accounted_begin(cache, bio);
 814	dm_submit_bio_remap(bio, NULL);
 815}
 816
 817static void issue_op(struct bio *bio, void *context)
 818{
 819	struct cache *cache = context;
 820
 821	accounted_request(cache, bio);
 822}
 823
 824/*
 825 * When running in writethrough mode we need to send writes to clean blocks
 826 * to both the cache and origin devices.  Clone the bio and send them in parallel.
 827 */
 828static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
 829				      dm_oblock_t oblock, dm_cblock_t cblock)
 830{
 831	struct bio *origin_bio = bio_alloc_clone(cache->origin_dev->bdev, bio,
 832						 GFP_NOIO, &cache->bs);
 833
 834	BUG_ON(!origin_bio);
 835
 836	bio_chain(origin_bio, bio);
 837
 838	if (bio_data_dir(origin_bio) == WRITE)
 839		clear_discard(cache, oblock_to_dblock(cache, oblock));
 840	submit_bio(origin_bio);
 841
 842	remap_to_cache(cache, bio, cblock);
 843}
 844
 845/*
 846 *--------------------------------------------------------------
 847 * Failure modes
 848 *--------------------------------------------------------------
 849 */
 850static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 851{
 852	return cache->features.mode;
 853}
 854
 855static const char *cache_device_name(struct cache *cache)
 856{
 857	return dm_table_device_name(cache->ti->table);
 858}
 859
 860static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 861{
 862	static const char *descs[] = {
 863		"write",
 864		"read-only",
 865		"fail"
 866	};
 867
 868	dm_table_event(cache->ti->table);
 869	DMINFO("%s: switching cache to %s mode",
 870	       cache_device_name(cache), descs[(int)mode]);
 871}
 872
 873static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 874{
 875	bool needs_check;
 876	enum cache_metadata_mode old_mode = get_cache_mode(cache);
 877
 878	if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
 879		DMERR("%s: unable to read needs_check flag, setting failure mode.",
 880		      cache_device_name(cache));
 881		new_mode = CM_FAIL;
 882	}
 883
 884	if (new_mode == CM_WRITE && needs_check) {
 885		DMERR("%s: unable to switch cache to write mode until repaired.",
 886		      cache_device_name(cache));
 887		if (old_mode != new_mode)
 888			new_mode = old_mode;
 889		else
 890			new_mode = CM_READ_ONLY;
 891	}
 892
 893	/* Never move out of fail mode */
 894	if (old_mode == CM_FAIL)
 895		new_mode = CM_FAIL;
 896
 897	switch (new_mode) {
 898	case CM_FAIL:
 899	case CM_READ_ONLY:
 900		dm_cache_metadata_set_read_only(cache->cmd);
 901		break;
 902
 903	case CM_WRITE:
 904		dm_cache_metadata_set_read_write(cache->cmd);
 905		break;
 906	}
 907
 908	cache->features.mode = new_mode;
 909
 910	if (new_mode != old_mode)
 911		notify_mode_switch(cache, new_mode);
 912}
 913
 914static void abort_transaction(struct cache *cache)
 915{
 916	const char *dev_name = cache_device_name(cache);
 917
 918	if (get_cache_mode(cache) >= CM_READ_ONLY)
 919		return;
 920
 921	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
 922	if (dm_cache_metadata_abort(cache->cmd)) {
 923		DMERR("%s: failed to abort metadata transaction", dev_name);
 924		set_cache_mode(cache, CM_FAIL);
 925	}
 926
 927	if (dm_cache_metadata_set_needs_check(cache->cmd)) {
 928		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
 929		set_cache_mode(cache, CM_FAIL);
 930	}
 931}
 932
 933static void metadata_operation_failed(struct cache *cache, const char *op, int r)
 934{
 935	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
 936		    cache_device_name(cache), op, r);
 937	abort_transaction(cache);
 938	set_cache_mode(cache, CM_READ_ONLY);
 939}
 940
 941/*----------------------------------------------------------------*/
 942
 943static void load_stats(struct cache *cache)
 944{
 945	struct dm_cache_statistics stats;
 946
 947	dm_cache_metadata_get_stats(cache->cmd, &stats);
 948	atomic_set(&cache->stats.read_hit, stats.read_hits);
 949	atomic_set(&cache->stats.read_miss, stats.read_misses);
 950	atomic_set(&cache->stats.write_hit, stats.write_hits);
 951	atomic_set(&cache->stats.write_miss, stats.write_misses);
 952}
 953
 954static void save_stats(struct cache *cache)
 955{
 956	struct dm_cache_statistics stats;
 957
 958	if (get_cache_mode(cache) >= CM_READ_ONLY)
 959		return;
 960
 961	stats.read_hits = atomic_read(&cache->stats.read_hit);
 962	stats.read_misses = atomic_read(&cache->stats.read_miss);
 963	stats.write_hits = atomic_read(&cache->stats.write_hit);
 964	stats.write_misses = atomic_read(&cache->stats.write_miss);
 965
 966	dm_cache_metadata_set_stats(cache->cmd, &stats);
 967}
 968
 969static void update_stats(struct cache_stats *stats, enum policy_operation op)
 970{
 971	switch (op) {
 972	case POLICY_PROMOTE:
 973		atomic_inc(&stats->promotion);
 974		break;
 975
 976	case POLICY_DEMOTE:
 977		atomic_inc(&stats->demotion);
 978		break;
 979
 980	case POLICY_WRITEBACK:
 981		atomic_inc(&stats->writeback);
 982		break;
 983	}
 984}
 985
 986/*
 987 *---------------------------------------------------------------------
 988 * Migration processing
 989 *
 990 * Migration covers moving data from the origin device to the cache, or
 991 * vice versa.
 992 *---------------------------------------------------------------------
 993 */
 994static void inc_io_migrations(struct cache *cache)
 995{
 996	atomic_inc(&cache->nr_io_migrations);
 997}
 998
 999static void dec_io_migrations(struct cache *cache)
1000{
1001	atomic_dec(&cache->nr_io_migrations);
1002}
1003
1004static bool discard_or_flush(struct bio *bio)
1005{
1006	return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1007}
1008
1009static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1010				     dm_dblock_t *b, dm_dblock_t *e)
1011{
1012	sector_t sb = bio->bi_iter.bi_sector;
1013	sector_t se = bio_end_sector(bio);
1014
1015	*b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1016
1017	if (se - sb < cache->discard_block_size)
1018		*e = *b;
1019	else
1020		*e = to_dblock(block_div(se, cache->discard_block_size));
1021}
1022
1023/*----------------------------------------------------------------*/
1024
1025static void prevent_background_work(struct cache *cache)
1026{
1027	lockdep_off();
1028	down_write(&cache->background_work_lock);
1029	lockdep_on();
1030}
1031
1032static void allow_background_work(struct cache *cache)
1033{
1034	lockdep_off();
1035	up_write(&cache->background_work_lock);
1036	lockdep_on();
1037}
1038
1039static bool background_work_begin(struct cache *cache)
1040{
1041	bool r;
1042
1043	lockdep_off();
1044	r = down_read_trylock(&cache->background_work_lock);
1045	lockdep_on();
1046
1047	return r;
1048}
1049
1050static void background_work_end(struct cache *cache)
1051{
1052	lockdep_off();
1053	up_read(&cache->background_work_lock);
1054	lockdep_on();
1055}
1056
1057/*----------------------------------------------------------------*/
1058
1059static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1060{
1061	return (bio_data_dir(bio) == WRITE) &&
1062		(bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1063}
1064
1065static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1066{
1067	return writeback_mode(cache) &&
1068		(is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1069}
1070
1071static void quiesce(struct dm_cache_migration *mg,
1072		    void (*continuation)(struct work_struct *))
1073{
1074	init_continuation(&mg->k, continuation);
1075	dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1076}
1077
1078static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1079{
1080	struct continuation *k = container_of(ws, struct continuation, ws);
1081
1082	return container_of(k, struct dm_cache_migration, k);
1083}
1084
1085static void copy_complete(int read_err, unsigned long write_err, void *context)
1086{
1087	struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1088
1089	if (read_err || write_err)
1090		mg->k.input = BLK_STS_IOERR;
1091
1092	queue_continuation(mg->cache->wq, &mg->k);
1093}
1094
1095static void copy(struct dm_cache_migration *mg, bool promote)
1096{
1097	struct dm_io_region o_region, c_region;
1098	struct cache *cache = mg->cache;
1099
1100	o_region.bdev = cache->origin_dev->bdev;
1101	o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1102	o_region.count = cache->sectors_per_block;
1103
1104	c_region.bdev = cache->cache_dev->bdev;
1105	c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1106	c_region.count = cache->sectors_per_block;
1107
1108	if (promote)
1109		dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1110	else
1111		dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1112}
1113
1114static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1115{
1116	struct per_bio_data *pb = get_per_bio_data(bio);
1117
1118	if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1119		free_prison_cell(cache, pb->cell);
1120	pb->cell = NULL;
1121}
1122
1123static void overwrite_endio(struct bio *bio)
1124{
1125	struct dm_cache_migration *mg = bio->bi_private;
1126	struct cache *cache = mg->cache;
1127	struct per_bio_data *pb = get_per_bio_data(bio);
1128
1129	dm_unhook_bio(&pb->hook_info, bio);
1130
1131	if (bio->bi_status)
1132		mg->k.input = bio->bi_status;
1133
1134	queue_continuation(cache->wq, &mg->k);
1135}
1136
1137static void overwrite(struct dm_cache_migration *mg,
1138		      void (*continuation)(struct work_struct *))
1139{
1140	struct bio *bio = mg->overwrite_bio;
1141	struct per_bio_data *pb = get_per_bio_data(bio);
1142
1143	dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1144
1145	/*
1146	 * The overwrite bio is part of the copy operation, as such it does
1147	 * not set/clear discard or dirty flags.
1148	 */
1149	if (mg->op->op == POLICY_PROMOTE)
1150		remap_to_cache(mg->cache, bio, mg->op->cblock);
1151	else
1152		remap_to_origin(mg->cache, bio);
1153
1154	init_continuation(&mg->k, continuation);
1155	accounted_request(mg->cache, bio);
1156}
1157
1158/*
1159 * Migration steps:
1160 *
1161 * 1) exclusive lock preventing WRITEs
1162 * 2) quiesce
1163 * 3) copy or issue overwrite bio
1164 * 4) upgrade to exclusive lock preventing READs and WRITEs
1165 * 5) quiesce
1166 * 6) update metadata and commit
1167 * 7) unlock
1168 */
1169static void mg_complete(struct dm_cache_migration *mg, bool success)
1170{
1171	struct bio_list bios;
1172	struct cache *cache = mg->cache;
1173	struct policy_work *op = mg->op;
1174	dm_cblock_t cblock = op->cblock;
1175
1176	if (success)
1177		update_stats(&cache->stats, op->op);
1178
1179	switch (op->op) {
1180	case POLICY_PROMOTE:
1181		clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1182		policy_complete_background_work(cache->policy, op, success);
1183
1184		if (mg->overwrite_bio) {
1185			if (success)
1186				force_set_dirty(cache, cblock);
1187			else if (mg->k.input)
1188				mg->overwrite_bio->bi_status = mg->k.input;
1189			else
1190				mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1191			bio_endio(mg->overwrite_bio);
1192		} else {
1193			if (success)
1194				force_clear_dirty(cache, cblock);
1195			dec_io_migrations(cache);
1196		}
1197		break;
1198
1199	case POLICY_DEMOTE:
1200		/*
1201		 * We clear dirty here to update the nr_dirty counter.
1202		 */
1203		if (success)
1204			force_clear_dirty(cache, cblock);
1205		policy_complete_background_work(cache->policy, op, success);
1206		dec_io_migrations(cache);
1207		break;
1208
1209	case POLICY_WRITEBACK:
1210		if (success)
1211			force_clear_dirty(cache, cblock);
1212		policy_complete_background_work(cache->policy, op, success);
1213		dec_io_migrations(cache);
1214		break;
1215	}
1216
1217	bio_list_init(&bios);
1218	if (mg->cell) {
1219		if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1220			free_prison_cell(cache, mg->cell);
1221	}
1222
1223	free_migration(mg);
1224	defer_bios(cache, &bios);
1225	wake_migration_worker(cache);
1226
1227	background_work_end(cache);
1228}
1229
1230static void mg_success(struct work_struct *ws)
1231{
1232	struct dm_cache_migration *mg = ws_to_mg(ws);
1233
1234	mg_complete(mg, mg->k.input == 0);
1235}
1236
1237static void mg_update_metadata(struct work_struct *ws)
1238{
1239	int r;
1240	struct dm_cache_migration *mg = ws_to_mg(ws);
1241	struct cache *cache = mg->cache;
1242	struct policy_work *op = mg->op;
1243
1244	switch (op->op) {
1245	case POLICY_PROMOTE:
1246		r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1247		if (r) {
1248			DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1249				    cache_device_name(cache));
1250			metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1251
1252			mg_complete(mg, false);
1253			return;
1254		}
1255		mg_complete(mg, true);
1256		break;
1257
1258	case POLICY_DEMOTE:
1259		r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1260		if (r) {
1261			DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1262				    cache_device_name(cache));
1263			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1264
1265			mg_complete(mg, false);
1266			return;
1267		}
1268
1269		/*
1270		 * It would be nice if we only had to commit when a REQ_FLUSH
1271		 * comes through.  But there's one scenario that we have to
1272		 * look out for:
1273		 *
1274		 * - vblock x in a cache block
1275		 * - domotion occurs
1276		 * - cache block gets reallocated and over written
1277		 * - crash
1278		 *
1279		 * When we recover, because there was no commit the cache will
1280		 * rollback to having the data for vblock x in the cache block.
1281		 * But the cache block has since been overwritten, so it'll end
1282		 * up pointing to data that was never in 'x' during the history
1283		 * of the device.
1284		 *
1285		 * To avoid this issue we require a commit as part of the
1286		 * demotion operation.
1287		 */
1288		init_continuation(&mg->k, mg_success);
1289		continue_after_commit(&cache->committer, &mg->k);
1290		schedule_commit(&cache->committer);
1291		break;
1292
1293	case POLICY_WRITEBACK:
1294		mg_complete(mg, true);
1295		break;
1296	}
1297}
1298
1299static void mg_update_metadata_after_copy(struct work_struct *ws)
1300{
1301	struct dm_cache_migration *mg = ws_to_mg(ws);
1302
1303	/*
1304	 * Did the copy succeed?
1305	 */
1306	if (mg->k.input)
1307		mg_complete(mg, false);
1308	else
1309		mg_update_metadata(ws);
1310}
1311
1312static void mg_upgrade_lock(struct work_struct *ws)
1313{
1314	int r;
1315	struct dm_cache_migration *mg = ws_to_mg(ws);
1316
1317	/*
1318	 * Did the copy succeed?
1319	 */
1320	if (mg->k.input)
1321		mg_complete(mg, false);
1322
1323	else {
1324		/*
1325		 * Now we want the lock to prevent both reads and writes.
1326		 */
1327		r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1328					    READ_WRITE_LOCK_LEVEL);
1329		if (r < 0)
1330			mg_complete(mg, false);
1331
1332		else if (r)
1333			quiesce(mg, mg_update_metadata);
1334
1335		else
1336			mg_update_metadata(ws);
1337	}
1338}
1339
1340static void mg_full_copy(struct work_struct *ws)
1341{
1342	struct dm_cache_migration *mg = ws_to_mg(ws);
1343	struct cache *cache = mg->cache;
1344	struct policy_work *op = mg->op;
1345	bool is_policy_promote = (op->op == POLICY_PROMOTE);
1346
1347	if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1348	    is_discarded_oblock(cache, op->oblock)) {
1349		mg_upgrade_lock(ws);
1350		return;
1351	}
1352
1353	init_continuation(&mg->k, mg_upgrade_lock);
1354	copy(mg, is_policy_promote);
1355}
1356
1357static void mg_copy(struct work_struct *ws)
1358{
1359	struct dm_cache_migration *mg = ws_to_mg(ws);
1360
1361	if (mg->overwrite_bio) {
1362		/*
1363		 * No exclusive lock was held when we last checked if the bio
1364		 * was optimisable.  So we have to check again in case things
1365		 * have changed (eg, the block may no longer be discarded).
1366		 */
1367		if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1368			/*
1369			 * Fallback to a real full copy after doing some tidying up.
1370			 */
1371			bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1372
1373			BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1374			mg->overwrite_bio = NULL;
1375			inc_io_migrations(mg->cache);
1376			mg_full_copy(ws);
1377			return;
1378		}
1379
1380		/*
1381		 * It's safe to do this here, even though it's new data
1382		 * because all IO has been locked out of the block.
1383		 *
1384		 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1385		 * so _not_ using mg_upgrade_lock() as continutation.
1386		 */
1387		overwrite(mg, mg_update_metadata_after_copy);
1388
1389	} else
1390		mg_full_copy(ws);
1391}
1392
1393static int mg_lock_writes(struct dm_cache_migration *mg)
1394{
1395	int r;
1396	struct dm_cell_key_v2 key;
1397	struct cache *cache = mg->cache;
1398	struct dm_bio_prison_cell_v2 *prealloc;
1399
1400	prealloc = alloc_prison_cell(cache);
1401
1402	/*
1403	 * Prevent writes to the block, but allow reads to continue.
1404	 * Unless we're using an overwrite bio, in which case we lock
1405	 * everything.
1406	 */
1407	build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1408	r = dm_cell_lock_v2(cache->prison, &key,
1409			    mg->overwrite_bio ?  READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1410			    prealloc, &mg->cell);
1411	if (r < 0) {
1412		free_prison_cell(cache, prealloc);
1413		mg_complete(mg, false);
1414		return r;
1415	}
1416
1417	if (mg->cell != prealloc)
1418		free_prison_cell(cache, prealloc);
1419
1420	if (r == 0)
1421		mg_copy(&mg->k.ws);
1422	else
1423		quiesce(mg, mg_copy);
1424
1425	return 0;
1426}
1427
1428static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1429{
1430	struct dm_cache_migration *mg;
1431
1432	if (!background_work_begin(cache)) {
1433		policy_complete_background_work(cache->policy, op, false);
1434		return -EPERM;
1435	}
1436
1437	mg = alloc_migration(cache);
1438
1439	mg->op = op;
1440	mg->overwrite_bio = bio;
1441
1442	if (!bio)
1443		inc_io_migrations(cache);
1444
1445	return mg_lock_writes(mg);
1446}
1447
1448/*
1449 *--------------------------------------------------------------
1450 * invalidation processing
1451 *--------------------------------------------------------------
1452 */
1453
1454static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1455{
1456	struct bio_list bios;
1457	struct cache *cache = mg->cache;
1458
1459	bio_list_init(&bios);
1460	if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1461		free_prison_cell(cache, mg->cell);
1462
1463	if (!success && mg->overwrite_bio)
1464		bio_io_error(mg->overwrite_bio);
1465
1466	free_migration(mg);
1467	defer_bios(cache, &bios);
1468
1469	background_work_end(cache);
1470}
1471
1472static void invalidate_completed(struct work_struct *ws)
1473{
1474	struct dm_cache_migration *mg = ws_to_mg(ws);
1475
1476	invalidate_complete(mg, !mg->k.input);
1477}
1478
1479static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1480{
1481	int r;
1482
1483	r = policy_invalidate_mapping(cache->policy, cblock);
1484	if (!r) {
1485		r = dm_cache_remove_mapping(cache->cmd, cblock);
1486		if (r) {
1487			DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1488				    cache_device_name(cache));
1489			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1490		}
1491
1492	} else if (r == -ENODATA) {
1493		/*
1494		 * Harmless, already unmapped.
1495		 */
1496		r = 0;
1497
1498	} else
1499		DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1500
1501	return r;
1502}
1503
1504static void invalidate_remove(struct work_struct *ws)
1505{
1506	int r;
1507	struct dm_cache_migration *mg = ws_to_mg(ws);
1508	struct cache *cache = mg->cache;
1509
1510	r = invalidate_cblock(cache, mg->invalidate_cblock);
1511	if (r) {
1512		invalidate_complete(mg, false);
1513		return;
1514	}
1515
1516	init_continuation(&mg->k, invalidate_completed);
1517	continue_after_commit(&cache->committer, &mg->k);
1518	remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1519	mg->overwrite_bio = NULL;
1520	schedule_commit(&cache->committer);
1521}
1522
1523static int invalidate_lock(struct dm_cache_migration *mg)
1524{
1525	int r;
1526	struct dm_cell_key_v2 key;
1527	struct cache *cache = mg->cache;
1528	struct dm_bio_prison_cell_v2 *prealloc;
1529
1530	prealloc = alloc_prison_cell(cache);
1531
1532	build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1533	r = dm_cell_lock_v2(cache->prison, &key,
1534			    READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1535	if (r < 0) {
1536		free_prison_cell(cache, prealloc);
1537		invalidate_complete(mg, false);
1538		return r;
1539	}
1540
1541	if (mg->cell != prealloc)
1542		free_prison_cell(cache, prealloc);
1543
1544	if (r)
1545		quiesce(mg, invalidate_remove);
1546
1547	else {
1548		/*
1549		 * We can't call invalidate_remove() directly here because we
1550		 * might still be in request context.
1551		 */
1552		init_continuation(&mg->k, invalidate_remove);
1553		queue_work(cache->wq, &mg->k.ws);
1554	}
1555
1556	return 0;
1557}
1558
1559static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1560			    dm_oblock_t oblock, struct bio *bio)
1561{
1562	struct dm_cache_migration *mg;
1563
1564	if (!background_work_begin(cache))
1565		return -EPERM;
1566
1567	mg = alloc_migration(cache);
1568
1569	mg->overwrite_bio = bio;
1570	mg->invalidate_cblock = cblock;
1571	mg->invalidate_oblock = oblock;
1572
1573	return invalidate_lock(mg);
1574}
1575
1576/*
1577 *--------------------------------------------------------------
1578 * bio processing
1579 *--------------------------------------------------------------
1580 */
1581
1582enum busy {
1583	IDLE,
1584	BUSY
1585};
1586
1587static enum busy spare_migration_bandwidth(struct cache *cache)
1588{
1589	bool idle = dm_iot_idle_for(&cache->tracker, HZ);
1590	sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1591		cache->sectors_per_block;
1592
1593	if (idle && current_volume <= cache->migration_threshold)
1594		return IDLE;
1595	else
1596		return BUSY;
1597}
1598
1599static void inc_hit_counter(struct cache *cache, struct bio *bio)
1600{
1601	atomic_inc(bio_data_dir(bio) == READ ?
1602		   &cache->stats.read_hit : &cache->stats.write_hit);
1603}
1604
1605static void inc_miss_counter(struct cache *cache, struct bio *bio)
1606{
1607	atomic_inc(bio_data_dir(bio) == READ ?
1608		   &cache->stats.read_miss : &cache->stats.write_miss);
1609}
1610
1611/*----------------------------------------------------------------*/
1612
1613static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1614		   bool *commit_needed)
1615{
1616	int r, data_dir;
1617	bool rb, background_queued;
1618	dm_cblock_t cblock;
1619
1620	*commit_needed = false;
1621
1622	rb = bio_detain_shared(cache, block, bio);
1623	if (!rb) {
1624		/*
1625		 * An exclusive lock is held for this block, so we have to
1626		 * wait.  We set the commit_needed flag so the current
1627		 * transaction will be committed asap, allowing this lock
1628		 * to be dropped.
1629		 */
1630		*commit_needed = true;
1631		return DM_MAPIO_SUBMITTED;
1632	}
1633
1634	data_dir = bio_data_dir(bio);
1635
1636	if (optimisable_bio(cache, bio, block)) {
1637		struct policy_work *op = NULL;
1638
1639		r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1640		if (unlikely(r && r != -ENOENT)) {
1641			DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1642				    cache_device_name(cache), r);
1643			bio_io_error(bio);
1644			return DM_MAPIO_SUBMITTED;
1645		}
1646
1647		if (r == -ENOENT && op) {
1648			bio_drop_shared_lock(cache, bio);
1649			BUG_ON(op->op != POLICY_PROMOTE);
1650			mg_start(cache, op, bio);
1651			return DM_MAPIO_SUBMITTED;
1652		}
1653	} else {
1654		r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1655		if (unlikely(r && r != -ENOENT)) {
1656			DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1657				    cache_device_name(cache), r);
1658			bio_io_error(bio);
1659			return DM_MAPIO_SUBMITTED;
1660		}
1661
1662		if (background_queued)
1663			wake_migration_worker(cache);
1664	}
1665
1666	if (r == -ENOENT) {
1667		struct per_bio_data *pb = get_per_bio_data(bio);
1668
1669		/*
1670		 * Miss.
1671		 */
1672		inc_miss_counter(cache, bio);
1673		if (pb->req_nr == 0) {
1674			accounted_begin(cache, bio);
1675			remap_to_origin_clear_discard(cache, bio, block);
1676		} else {
1677			/*
1678			 * This is a duplicate writethrough io that is no
1679			 * longer needed because the block has been demoted.
1680			 */
1681			bio_endio(bio);
1682			return DM_MAPIO_SUBMITTED;
1683		}
1684	} else {
1685		/*
1686		 * Hit.
1687		 */
1688		inc_hit_counter(cache, bio);
1689
1690		/*
1691		 * Passthrough always maps to the origin, invalidating any
1692		 * cache blocks that are written to.
1693		 */
1694		if (passthrough_mode(cache)) {
1695			if (bio_data_dir(bio) == WRITE) {
1696				bio_drop_shared_lock(cache, bio);
1697				atomic_inc(&cache->stats.demotion);
1698				invalidate_start(cache, cblock, block, bio);
1699			} else
1700				remap_to_origin_clear_discard(cache, bio, block);
1701		} else {
1702			if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1703			    !is_dirty(cache, cblock)) {
1704				remap_to_origin_and_cache(cache, bio, block, cblock);
1705				accounted_begin(cache, bio);
1706			} else
1707				remap_to_cache_dirty(cache, bio, block, cblock);
1708		}
1709	}
1710
1711	/*
1712	 * dm core turns FUA requests into a separate payload and FLUSH req.
1713	 */
1714	if (bio->bi_opf & REQ_FUA) {
1715		/*
1716		 * issue_after_commit will call accounted_begin a second time.  So
1717		 * we call accounted_complete() to avoid double accounting.
1718		 */
1719		accounted_complete(cache, bio);
1720		issue_after_commit(&cache->committer, bio);
1721		*commit_needed = true;
1722		return DM_MAPIO_SUBMITTED;
1723	}
1724
1725	return DM_MAPIO_REMAPPED;
1726}
1727
1728static bool process_bio(struct cache *cache, struct bio *bio)
1729{
1730	bool commit_needed;
1731
1732	if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1733		dm_submit_bio_remap(bio, NULL);
1734
1735	return commit_needed;
1736}
1737
1738/*
1739 * A non-zero return indicates read_only or fail_io mode.
1740 */
1741static int commit(struct cache *cache, bool clean_shutdown)
1742{
1743	int r;
1744
1745	if (get_cache_mode(cache) >= CM_READ_ONLY)
1746		return -EINVAL;
1747
1748	atomic_inc(&cache->stats.commit_count);
1749	r = dm_cache_commit(cache->cmd, clean_shutdown);
1750	if (r)
1751		metadata_operation_failed(cache, "dm_cache_commit", r);
1752
1753	return r;
1754}
1755
1756/*
1757 * Used by the batcher.
1758 */
1759static blk_status_t commit_op(void *context)
1760{
1761	struct cache *cache = context;
1762
1763	if (dm_cache_changed_this_transaction(cache->cmd))
1764		return errno_to_blk_status(commit(cache, false));
1765
1766	return 0;
1767}
1768
1769/*----------------------------------------------------------------*/
1770
1771static bool process_flush_bio(struct cache *cache, struct bio *bio)
1772{
1773	struct per_bio_data *pb = get_per_bio_data(bio);
1774
1775	if (!pb->req_nr)
1776		remap_to_origin(cache, bio);
1777	else
1778		remap_to_cache(cache, bio, 0);
1779
1780	issue_after_commit(&cache->committer, bio);
1781	return true;
1782}
1783
1784static bool process_discard_bio(struct cache *cache, struct bio *bio)
1785{
1786	dm_dblock_t b, e;
1787
1788	/*
1789	 * FIXME: do we need to lock the region?  Or can we just assume the
1790	 * user wont be so foolish as to issue discard concurrently with
1791	 * other IO?
1792	 */
1793	calc_discard_block_range(cache, bio, &b, &e);
1794	while (b != e) {
1795		set_discard(cache, b);
1796		b = to_dblock(from_dblock(b) + 1);
1797	}
1798
1799	if (cache->features.discard_passdown) {
1800		remap_to_origin(cache, bio);
1801		dm_submit_bio_remap(bio, NULL);
1802	} else
1803		bio_endio(bio);
1804
1805	return false;
1806}
1807
1808static void process_deferred_bios(struct work_struct *ws)
1809{
1810	struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1811
1812	bool commit_needed = false;
1813	struct bio_list bios;
1814	struct bio *bio;
1815
1816	bio_list_init(&bios);
1817
1818	spin_lock_irq(&cache->lock);
1819	bio_list_merge(&bios, &cache->deferred_bios);
1820	bio_list_init(&cache->deferred_bios);
1821	spin_unlock_irq(&cache->lock);
1822
1823	while ((bio = bio_list_pop(&bios))) {
1824		if (bio->bi_opf & REQ_PREFLUSH)
1825			commit_needed = process_flush_bio(cache, bio) || commit_needed;
1826
1827		else if (bio_op(bio) == REQ_OP_DISCARD)
1828			commit_needed = process_discard_bio(cache, bio) || commit_needed;
1829
1830		else
1831			commit_needed = process_bio(cache, bio) || commit_needed;
1832		cond_resched();
1833	}
1834
1835	if (commit_needed)
1836		schedule_commit(&cache->committer);
1837}
1838
1839/*
1840 *--------------------------------------------------------------
1841 * Main worker loop
1842 *--------------------------------------------------------------
1843 */
1844static void requeue_deferred_bios(struct cache *cache)
1845{
1846	struct bio *bio;
1847	struct bio_list bios;
1848
1849	bio_list_init(&bios);
1850	bio_list_merge(&bios, &cache->deferred_bios);
1851	bio_list_init(&cache->deferred_bios);
1852
1853	while ((bio = bio_list_pop(&bios))) {
1854		bio->bi_status = BLK_STS_DM_REQUEUE;
1855		bio_endio(bio);
1856		cond_resched();
1857	}
1858}
1859
1860/*
1861 * We want to commit periodically so that not too much
1862 * unwritten metadata builds up.
1863 */
1864static void do_waker(struct work_struct *ws)
1865{
1866	struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1867
1868	policy_tick(cache->policy, true);
1869	wake_migration_worker(cache);
1870	schedule_commit(&cache->committer);
1871	queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1872}
1873
1874static void check_migrations(struct work_struct *ws)
1875{
1876	int r;
1877	struct policy_work *op;
1878	struct cache *cache = container_of(ws, struct cache, migration_worker);
1879	enum busy b;
1880
1881	for (;;) {
1882		b = spare_migration_bandwidth(cache);
1883
1884		r = policy_get_background_work(cache->policy, b == IDLE, &op);
1885		if (r == -ENODATA)
1886			break;
1887
1888		if (r) {
1889			DMERR_LIMIT("%s: policy_background_work failed",
1890				    cache_device_name(cache));
1891			break;
1892		}
1893
1894		r = mg_start(cache, op, NULL);
1895		if (r)
1896			break;
1897
1898		cond_resched();
1899	}
1900}
1901
1902/*
1903 *--------------------------------------------------------------
1904 * Target methods
1905 *--------------------------------------------------------------
1906 */
1907
1908/*
1909 * This function gets called on the error paths of the constructor, so we
1910 * have to cope with a partially initialised struct.
1911 */
1912static void destroy(struct cache *cache)
1913{
1914	unsigned int i;
1915
1916	mempool_exit(&cache->migration_pool);
1917
1918	if (cache->prison)
1919		dm_bio_prison_destroy_v2(cache->prison);
1920
1921	cancel_delayed_work_sync(&cache->waker);
1922	if (cache->wq)
1923		destroy_workqueue(cache->wq);
1924
1925	if (cache->dirty_bitset)
1926		free_bitset(cache->dirty_bitset);
1927
1928	if (cache->discard_bitset)
1929		free_bitset(cache->discard_bitset);
1930
1931	if (cache->copier)
1932		dm_kcopyd_client_destroy(cache->copier);
1933
1934	if (cache->cmd)
1935		dm_cache_metadata_close(cache->cmd);
1936
1937	if (cache->metadata_dev)
1938		dm_put_device(cache->ti, cache->metadata_dev);
1939
1940	if (cache->origin_dev)
1941		dm_put_device(cache->ti, cache->origin_dev);
1942
1943	if (cache->cache_dev)
1944		dm_put_device(cache->ti, cache->cache_dev);
1945
1946	if (cache->policy)
1947		dm_cache_policy_destroy(cache->policy);
1948
1949	for (i = 0; i < cache->nr_ctr_args ; i++)
1950		kfree(cache->ctr_args[i]);
1951	kfree(cache->ctr_args);
1952
1953	bioset_exit(&cache->bs);
1954
1955	kfree(cache);
1956}
1957
1958static void cache_dtr(struct dm_target *ti)
1959{
1960	struct cache *cache = ti->private;
1961
1962	destroy(cache);
1963}
1964
1965static sector_t get_dev_size(struct dm_dev *dev)
1966{
1967	return bdev_nr_sectors(dev->bdev);
1968}
1969
1970/*----------------------------------------------------------------*/
1971
1972/*
1973 * Construct a cache device mapping.
1974 *
1975 * cache <metadata dev> <cache dev> <origin dev> <block size>
1976 *       <#feature args> [<feature arg>]*
1977 *       <policy> <#policy args> [<policy arg>]*
1978 *
1979 * metadata dev    : fast device holding the persistent metadata
1980 * cache dev	   : fast device holding cached data blocks
1981 * origin dev	   : slow device holding original data blocks
1982 * block size	   : cache unit size in sectors
1983 *
1984 * #feature args   : number of feature arguments passed
1985 * feature args    : writethrough.  (The default is writeback.)
1986 *
1987 * policy	   : the replacement policy to use
1988 * #policy args    : an even number of policy arguments corresponding
1989 *		     to key/value pairs passed to the policy
1990 * policy args	   : key/value pairs passed to the policy
1991 *		     E.g. 'sequential_threshold 1024'
1992 *		     See cache-policies.txt for details.
1993 *
1994 * Optional feature arguments are:
1995 *   writethrough  : write through caching that prohibits cache block
1996 *		     content from being different from origin block content.
1997 *		     Without this argument, the default behaviour is to write
1998 *		     back cache block contents later for performance reasons,
1999 *		     so they may differ from the corresponding origin blocks.
2000 */
2001struct cache_args {
2002	struct dm_target *ti;
2003
2004	struct dm_dev *metadata_dev;
2005
2006	struct dm_dev *cache_dev;
2007	sector_t cache_sectors;
2008
2009	struct dm_dev *origin_dev;
2010	sector_t origin_sectors;
2011
2012	uint32_t block_size;
2013
2014	const char *policy_name;
2015	int policy_argc;
2016	const char **policy_argv;
2017
2018	struct cache_features features;
2019};
2020
2021static void destroy_cache_args(struct cache_args *ca)
2022{
2023	if (ca->metadata_dev)
2024		dm_put_device(ca->ti, ca->metadata_dev);
2025
2026	if (ca->cache_dev)
2027		dm_put_device(ca->ti, ca->cache_dev);
2028
2029	if (ca->origin_dev)
2030		dm_put_device(ca->ti, ca->origin_dev);
2031
2032	kfree(ca);
2033}
2034
2035static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2036{
2037	if (!as->argc) {
2038		*error = "Insufficient args";
2039		return false;
2040	}
2041
2042	return true;
2043}
2044
2045static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2046			      char **error)
2047{
2048	int r;
2049	sector_t metadata_dev_size;
2050
2051	if (!at_least_one_arg(as, error))
2052		return -EINVAL;
2053
2054	r = dm_get_device(ca->ti, dm_shift_arg(as),
2055			  BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->metadata_dev);
2056	if (r) {
2057		*error = "Error opening metadata device";
2058		return r;
2059	}
2060
2061	metadata_dev_size = get_dev_size(ca->metadata_dev);
2062	if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2063		DMWARN("Metadata device %pg is larger than %u sectors: excess space will not be used.",
2064		       ca->metadata_dev->bdev, THIN_METADATA_MAX_SECTORS);
2065
2066	return 0;
2067}
2068
2069static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2070			   char **error)
2071{
2072	int r;
2073
2074	if (!at_least_one_arg(as, error))
2075		return -EINVAL;
2076
2077	r = dm_get_device(ca->ti, dm_shift_arg(as),
2078			  BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->cache_dev);
2079	if (r) {
2080		*error = "Error opening cache device";
2081		return r;
2082	}
2083	ca->cache_sectors = get_dev_size(ca->cache_dev);
2084
2085	return 0;
2086}
2087
2088static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2089			    char **error)
2090{
2091	int r;
2092
2093	if (!at_least_one_arg(as, error))
2094		return -EINVAL;
2095
2096	r = dm_get_device(ca->ti, dm_shift_arg(as),
2097			  BLK_OPEN_READ | BLK_OPEN_WRITE, &ca->origin_dev);
2098	if (r) {
2099		*error = "Error opening origin device";
2100		return r;
2101	}
2102
2103	ca->origin_sectors = get_dev_size(ca->origin_dev);
2104	if (ca->ti->len > ca->origin_sectors) {
2105		*error = "Device size larger than cached device";
2106		return -EINVAL;
2107	}
2108
2109	return 0;
2110}
2111
2112static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2113			    char **error)
2114{
2115	unsigned long block_size;
2116
2117	if (!at_least_one_arg(as, error))
2118		return -EINVAL;
2119
2120	if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2121	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2122	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2123	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2124		*error = "Invalid data block size";
2125		return -EINVAL;
2126	}
2127
2128	if (block_size > ca->cache_sectors) {
2129		*error = "Data block size is larger than the cache device";
2130		return -EINVAL;
2131	}
2132
2133	ca->block_size = block_size;
2134
2135	return 0;
2136}
2137
2138static void init_features(struct cache_features *cf)
2139{
2140	cf->mode = CM_WRITE;
2141	cf->io_mode = CM_IO_WRITEBACK;
2142	cf->metadata_version = 1;
2143	cf->discard_passdown = true;
2144}
2145
2146static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2147			  char **error)
2148{
2149	static const struct dm_arg _args[] = {
2150		{0, 3, "Invalid number of cache feature arguments"},
2151	};
2152
2153	int r, mode_ctr = 0;
2154	unsigned int argc;
2155	const char *arg;
2156	struct cache_features *cf = &ca->features;
2157
2158	init_features(cf);
2159
2160	r = dm_read_arg_group(_args, as, &argc, error);
2161	if (r)
2162		return -EINVAL;
2163
2164	while (argc--) {
2165		arg = dm_shift_arg(as);
2166
2167		if (!strcasecmp(arg, "writeback")) {
2168			cf->io_mode = CM_IO_WRITEBACK;
2169			mode_ctr++;
2170		}
2171
2172		else if (!strcasecmp(arg, "writethrough")) {
2173			cf->io_mode = CM_IO_WRITETHROUGH;
2174			mode_ctr++;
2175		}
2176
2177		else if (!strcasecmp(arg, "passthrough")) {
2178			cf->io_mode = CM_IO_PASSTHROUGH;
2179			mode_ctr++;
2180		}
2181
2182		else if (!strcasecmp(arg, "metadata2"))
2183			cf->metadata_version = 2;
2184
2185		else if (!strcasecmp(arg, "no_discard_passdown"))
2186			cf->discard_passdown = false;
2187
2188		else {
2189			*error = "Unrecognised cache feature requested";
2190			return -EINVAL;
2191		}
2192	}
2193
2194	if (mode_ctr > 1) {
2195		*error = "Duplicate cache io_mode features requested";
2196		return -EINVAL;
2197	}
2198
2199	return 0;
2200}
2201
2202static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2203			char **error)
2204{
2205	static const struct dm_arg _args[] = {
2206		{0, 1024, "Invalid number of policy arguments"},
2207	};
2208
2209	int r;
2210
2211	if (!at_least_one_arg(as, error))
2212		return -EINVAL;
2213
2214	ca->policy_name = dm_shift_arg(as);
2215
2216	r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2217	if (r)
2218		return -EINVAL;
2219
2220	ca->policy_argv = (const char **)as->argv;
2221	dm_consume_args(as, ca->policy_argc);
2222
2223	return 0;
2224}
2225
2226static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2227			    char **error)
2228{
2229	int r;
2230	struct dm_arg_set as;
2231
2232	as.argc = argc;
2233	as.argv = argv;
2234
2235	r = parse_metadata_dev(ca, &as, error);
2236	if (r)
2237		return r;
2238
2239	r = parse_cache_dev(ca, &as, error);
2240	if (r)
2241		return r;
2242
2243	r = parse_origin_dev(ca, &as, error);
2244	if (r)
2245		return r;
2246
2247	r = parse_block_size(ca, &as, error);
2248	if (r)
2249		return r;
2250
2251	r = parse_features(ca, &as, error);
2252	if (r)
2253		return r;
2254
2255	r = parse_policy(ca, &as, error);
2256	if (r)
2257		return r;
2258
2259	return 0;
2260}
2261
2262/*----------------------------------------------------------------*/
2263
2264static struct kmem_cache *migration_cache;
2265
2266#define NOT_CORE_OPTION 1
2267
2268static int process_config_option(struct cache *cache, const char *key, const char *value)
2269{
2270	unsigned long tmp;
2271
2272	if (!strcasecmp(key, "migration_threshold")) {
2273		if (kstrtoul(value, 10, &tmp))
2274			return -EINVAL;
2275
2276		cache->migration_threshold = tmp;
2277		return 0;
2278	}
2279
2280	return NOT_CORE_OPTION;
2281}
2282
2283static int set_config_value(struct cache *cache, const char *key, const char *value)
2284{
2285	int r = process_config_option(cache, key, value);
2286
2287	if (r == NOT_CORE_OPTION)
2288		r = policy_set_config_value(cache->policy, key, value);
2289
2290	if (r)
2291		DMWARN("bad config value for %s: %s", key, value);
2292
2293	return r;
2294}
2295
2296static int set_config_values(struct cache *cache, int argc, const char **argv)
2297{
2298	int r = 0;
2299
2300	if (argc & 1) {
2301		DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2302		return -EINVAL;
2303	}
2304
2305	while (argc) {
2306		r = set_config_value(cache, argv[0], argv[1]);
2307		if (r)
2308			break;
2309
2310		argc -= 2;
2311		argv += 2;
2312	}
2313
2314	return r;
2315}
2316
2317static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2318			       char **error)
2319{
2320	struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2321							   cache->cache_size,
2322							   cache->origin_sectors,
2323							   cache->sectors_per_block);
2324	if (IS_ERR(p)) {
2325		*error = "Error creating cache's policy";
2326		return PTR_ERR(p);
2327	}
2328	cache->policy = p;
2329	BUG_ON(!cache->policy);
2330
2331	return 0;
2332}
2333
2334/*
2335 * We want the discard block size to be at least the size of the cache
2336 * block size and have no more than 2^14 discard blocks across the origin.
2337 */
2338#define MAX_DISCARD_BLOCKS (1 << 14)
2339
2340static bool too_many_discard_blocks(sector_t discard_block_size,
2341				    sector_t origin_size)
2342{
2343	(void) sector_div(origin_size, discard_block_size);
2344
2345	return origin_size > MAX_DISCARD_BLOCKS;
2346}
2347
2348static sector_t calculate_discard_block_size(sector_t cache_block_size,
2349					     sector_t origin_size)
2350{
2351	sector_t discard_block_size = cache_block_size;
2352
2353	if (origin_size)
2354		while (too_many_discard_blocks(discard_block_size, origin_size))
2355			discard_block_size *= 2;
2356
2357	return discard_block_size;
2358}
2359
2360static void set_cache_size(struct cache *cache, dm_cblock_t size)
2361{
2362	dm_block_t nr_blocks = from_cblock(size);
2363
2364	if (nr_blocks > (1 << 20) && cache->cache_size != size)
2365		DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2366			     "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2367			     "Please consider increasing the cache block size to reduce the overall cache block count.",
2368			     (unsigned long long) nr_blocks);
2369
2370	cache->cache_size = size;
2371}
2372
2373#define DEFAULT_MIGRATION_THRESHOLD 2048
2374
2375static int cache_create(struct cache_args *ca, struct cache **result)
2376{
2377	int r = 0;
2378	char **error = &ca->ti->error;
2379	struct cache *cache;
2380	struct dm_target *ti = ca->ti;
2381	dm_block_t origin_blocks;
2382	struct dm_cache_metadata *cmd;
2383	bool may_format = ca->features.mode == CM_WRITE;
2384
2385	cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2386	if (!cache)
2387		return -ENOMEM;
2388
2389	cache->ti = ca->ti;
2390	ti->private = cache;
2391	ti->accounts_remapped_io = true;
2392	ti->num_flush_bios = 2;
2393	ti->flush_supported = true;
2394
2395	ti->num_discard_bios = 1;
2396	ti->discards_supported = true;
2397
2398	ti->per_io_data_size = sizeof(struct per_bio_data);
2399
2400	cache->features = ca->features;
2401	if (writethrough_mode(cache)) {
2402		/* Create bioset for writethrough bios issued to origin */
2403		r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2404		if (r)
2405			goto bad;
2406	}
2407
2408	cache->metadata_dev = ca->metadata_dev;
2409	cache->origin_dev = ca->origin_dev;
2410	cache->cache_dev = ca->cache_dev;
2411
2412	ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2413
2414	origin_blocks = cache->origin_sectors = ca->origin_sectors;
2415	origin_blocks = block_div(origin_blocks, ca->block_size);
2416	cache->origin_blocks = to_oblock(origin_blocks);
2417
2418	cache->sectors_per_block = ca->block_size;
2419	if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2420		r = -EINVAL;
2421		goto bad;
2422	}
2423
2424	if (ca->block_size & (ca->block_size - 1)) {
2425		dm_block_t cache_size = ca->cache_sectors;
2426
2427		cache->sectors_per_block_shift = -1;
2428		cache_size = block_div(cache_size, ca->block_size);
2429		set_cache_size(cache, to_cblock(cache_size));
2430	} else {
2431		cache->sectors_per_block_shift = __ffs(ca->block_size);
2432		set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2433	}
2434
2435	r = create_cache_policy(cache, ca, error);
2436	if (r)
2437		goto bad;
2438
2439	cache->policy_nr_args = ca->policy_argc;
2440	cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2441
2442	r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2443	if (r) {
2444		*error = "Error setting cache policy's config values";
2445		goto bad;
2446	}
2447
2448	cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2449				     ca->block_size, may_format,
2450				     dm_cache_policy_get_hint_size(cache->policy),
2451				     ca->features.metadata_version);
2452	if (IS_ERR(cmd)) {
2453		*error = "Error creating metadata object";
2454		r = PTR_ERR(cmd);
2455		goto bad;
2456	}
2457	cache->cmd = cmd;
2458	set_cache_mode(cache, CM_WRITE);
2459	if (get_cache_mode(cache) != CM_WRITE) {
2460		*error = "Unable to get write access to metadata, please check/repair metadata.";
2461		r = -EINVAL;
2462		goto bad;
2463	}
2464
2465	if (passthrough_mode(cache)) {
2466		bool all_clean;
2467
2468		r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2469		if (r) {
2470			*error = "dm_cache_metadata_all_clean() failed";
2471			goto bad;
2472		}
2473
2474		if (!all_clean) {
2475			*error = "Cannot enter passthrough mode unless all blocks are clean";
2476			r = -EINVAL;
2477			goto bad;
2478		}
2479
2480		policy_allow_migrations(cache->policy, false);
2481	}
2482
2483	spin_lock_init(&cache->lock);
2484	bio_list_init(&cache->deferred_bios);
2485	atomic_set(&cache->nr_allocated_migrations, 0);
2486	atomic_set(&cache->nr_io_migrations, 0);
2487	init_waitqueue_head(&cache->migration_wait);
2488
2489	r = -ENOMEM;
2490	atomic_set(&cache->nr_dirty, 0);
2491	cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2492	if (!cache->dirty_bitset) {
2493		*error = "could not allocate dirty bitset";
2494		goto bad;
2495	}
2496	clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2497
2498	cache->discard_block_size =
2499		calculate_discard_block_size(cache->sectors_per_block,
2500					     cache->origin_sectors);
2501	cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2502							      cache->discard_block_size));
2503	cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2504	if (!cache->discard_bitset) {
2505		*error = "could not allocate discard bitset";
2506		goto bad;
2507	}
2508	clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2509
2510	cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2511	if (IS_ERR(cache->copier)) {
2512		*error = "could not create kcopyd client";
2513		r = PTR_ERR(cache->copier);
2514		goto bad;
2515	}
2516
2517	cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2518	if (!cache->wq) {
2519		*error = "could not create workqueue for metadata object";
2520		goto bad;
2521	}
2522	INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2523	INIT_WORK(&cache->migration_worker, check_migrations);
2524	INIT_DELAYED_WORK(&cache->waker, do_waker);
2525
2526	cache->prison = dm_bio_prison_create_v2(cache->wq);
2527	if (!cache->prison) {
2528		*error = "could not create bio prison";
2529		goto bad;
2530	}
2531
2532	r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2533				   migration_cache);
2534	if (r) {
2535		*error = "Error creating cache's migration mempool";
2536		goto bad;
2537	}
2538
2539	cache->need_tick_bio = true;
2540	cache->sized = false;
2541	cache->invalidate = false;
2542	cache->commit_requested = false;
2543	cache->loaded_mappings = false;
2544	cache->loaded_discards = false;
2545
2546	load_stats(cache);
2547
2548	atomic_set(&cache->stats.demotion, 0);
2549	atomic_set(&cache->stats.promotion, 0);
2550	atomic_set(&cache->stats.copies_avoided, 0);
2551	atomic_set(&cache->stats.cache_cell_clash, 0);
2552	atomic_set(&cache->stats.commit_count, 0);
2553	atomic_set(&cache->stats.discard_count, 0);
2554
2555	spin_lock_init(&cache->invalidation_lock);
2556	INIT_LIST_HEAD(&cache->invalidation_requests);
2557
2558	batcher_init(&cache->committer, commit_op, cache,
2559		     issue_op, cache, cache->wq);
2560	dm_iot_init(&cache->tracker);
2561
2562	init_rwsem(&cache->background_work_lock);
2563	prevent_background_work(cache);
2564
2565	*result = cache;
2566	return 0;
2567bad:
2568	destroy(cache);
2569	return r;
2570}
2571
2572static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2573{
2574	unsigned int i;
2575	const char **copy;
2576
2577	copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2578	if (!copy)
2579		return -ENOMEM;
2580	for (i = 0; i < argc; i++) {
2581		copy[i] = kstrdup(argv[i], GFP_KERNEL);
2582		if (!copy[i]) {
2583			while (i--)
2584				kfree(copy[i]);
2585			kfree(copy);
2586			return -ENOMEM;
2587		}
2588	}
2589
2590	cache->nr_ctr_args = argc;
2591	cache->ctr_args = copy;
2592
2593	return 0;
2594}
2595
2596static int cache_ctr(struct dm_target *ti, unsigned int argc, char **argv)
2597{
2598	int r = -EINVAL;
2599	struct cache_args *ca;
2600	struct cache *cache = NULL;
2601
2602	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2603	if (!ca) {
2604		ti->error = "Error allocating memory for cache";
2605		return -ENOMEM;
2606	}
2607	ca->ti = ti;
2608
2609	r = parse_cache_args(ca, argc, argv, &ti->error);
2610	if (r)
2611		goto out;
2612
2613	r = cache_create(ca, &cache);
2614	if (r)
2615		goto out;
2616
2617	r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2618	if (r) {
2619		destroy(cache);
2620		goto out;
2621	}
2622
2623	ti->private = cache;
2624out:
2625	destroy_cache_args(ca);
2626	return r;
2627}
2628
2629/*----------------------------------------------------------------*/
2630
2631static int cache_map(struct dm_target *ti, struct bio *bio)
2632{
2633	struct cache *cache = ti->private;
2634
2635	int r;
2636	bool commit_needed;
2637	dm_oblock_t block = get_bio_block(cache, bio);
2638
2639	init_per_bio_data(bio);
2640	if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2641		/*
2642		 * This can only occur if the io goes to a partial block at
2643		 * the end of the origin device.  We don't cache these.
2644		 * Just remap to the origin and carry on.
2645		 */
2646		remap_to_origin(cache, bio);
2647		accounted_begin(cache, bio);
2648		return DM_MAPIO_REMAPPED;
2649	}
2650
2651	if (discard_or_flush(bio)) {
2652		defer_bio(cache, bio);
2653		return DM_MAPIO_SUBMITTED;
2654	}
2655
2656	r = map_bio(cache, bio, block, &commit_needed);
2657	if (commit_needed)
2658		schedule_commit(&cache->committer);
2659
2660	return r;
2661}
2662
2663static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2664{
2665	struct cache *cache = ti->private;
2666	unsigned long flags;
2667	struct per_bio_data *pb = get_per_bio_data(bio);
2668
2669	if (pb->tick) {
2670		policy_tick(cache->policy, false);
2671
2672		spin_lock_irqsave(&cache->lock, flags);
2673		cache->need_tick_bio = true;
2674		spin_unlock_irqrestore(&cache->lock, flags);
2675	}
2676
2677	bio_drop_shared_lock(cache, bio);
2678	accounted_complete(cache, bio);
2679
2680	return DM_ENDIO_DONE;
2681}
2682
2683static int write_dirty_bitset(struct cache *cache)
2684{
2685	int r;
2686
2687	if (get_cache_mode(cache) >= CM_READ_ONLY)
2688		return -EINVAL;
2689
2690	r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2691	if (r)
2692		metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2693
2694	return r;
2695}
2696
2697static int write_discard_bitset(struct cache *cache)
2698{
2699	unsigned int i, r;
2700
2701	if (get_cache_mode(cache) >= CM_READ_ONLY)
2702		return -EINVAL;
2703
2704	r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2705					   cache->discard_nr_blocks);
2706	if (r) {
2707		DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2708		metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2709		return r;
2710	}
2711
2712	for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2713		r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2714					 is_discarded(cache, to_dblock(i)));
2715		if (r) {
2716			metadata_operation_failed(cache, "dm_cache_set_discard", r);
2717			return r;
2718		}
2719	}
2720
2721	return 0;
2722}
2723
2724static int write_hints(struct cache *cache)
2725{
2726	int r;
2727
2728	if (get_cache_mode(cache) >= CM_READ_ONLY)
2729		return -EINVAL;
2730
2731	r = dm_cache_write_hints(cache->cmd, cache->policy);
2732	if (r) {
2733		metadata_operation_failed(cache, "dm_cache_write_hints", r);
2734		return r;
2735	}
2736
2737	return 0;
2738}
2739
2740/*
2741 * returns true on success
2742 */
2743static bool sync_metadata(struct cache *cache)
2744{
2745	int r1, r2, r3, r4;
2746
2747	r1 = write_dirty_bitset(cache);
2748	if (r1)
2749		DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2750
2751	r2 = write_discard_bitset(cache);
2752	if (r2)
2753		DMERR("%s: could not write discard bitset", cache_device_name(cache));
2754
2755	save_stats(cache);
2756
2757	r3 = write_hints(cache);
2758	if (r3)
2759		DMERR("%s: could not write hints", cache_device_name(cache));
2760
2761	/*
2762	 * If writing the above metadata failed, we still commit, but don't
2763	 * set the clean shutdown flag.  This will effectively force every
2764	 * dirty bit to be set on reload.
2765	 */
2766	r4 = commit(cache, !r1 && !r2 && !r3);
2767	if (r4)
2768		DMERR("%s: could not write cache metadata", cache_device_name(cache));
2769
2770	return !r1 && !r2 && !r3 && !r4;
2771}
2772
2773static void cache_postsuspend(struct dm_target *ti)
2774{
2775	struct cache *cache = ti->private;
2776
2777	prevent_background_work(cache);
2778	BUG_ON(atomic_read(&cache->nr_io_migrations));
2779
2780	cancel_delayed_work_sync(&cache->waker);
2781	drain_workqueue(cache->wq);
2782	WARN_ON(cache->tracker.in_flight);
2783
2784	/*
2785	 * If it's a flush suspend there won't be any deferred bios, so this
2786	 * call is harmless.
2787	 */
2788	requeue_deferred_bios(cache);
2789
2790	if (get_cache_mode(cache) == CM_WRITE)
2791		(void) sync_metadata(cache);
2792}
2793
2794static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2795			bool dirty, uint32_t hint, bool hint_valid)
2796{
2797	struct cache *cache = context;
2798
2799	if (dirty) {
2800		set_bit(from_cblock(cblock), cache->dirty_bitset);
2801		atomic_inc(&cache->nr_dirty);
2802	} else
2803		clear_bit(from_cblock(cblock), cache->dirty_bitset);
2804
2805	return policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2806}
2807
2808/*
2809 * The discard block size in the on disk metadata is not
2810 * necessarily the same as we're currently using.  So we have to
2811 * be careful to only set the discarded attribute if we know it
2812 * covers a complete block of the new size.
2813 */
2814struct discard_load_info {
2815	struct cache *cache;
2816
2817	/*
2818	 * These blocks are sized using the on disk dblock size, rather
2819	 * than the current one.
2820	 */
2821	dm_block_t block_size;
2822	dm_block_t discard_begin, discard_end;
2823};
2824
2825static void discard_load_info_init(struct cache *cache,
2826				   struct discard_load_info *li)
2827{
2828	li->cache = cache;
2829	li->discard_begin = li->discard_end = 0;
2830}
2831
2832static void set_discard_range(struct discard_load_info *li)
2833{
2834	sector_t b, e;
2835
2836	if (li->discard_begin == li->discard_end)
2837		return;
2838
2839	/*
2840	 * Convert to sectors.
2841	 */
2842	b = li->discard_begin * li->block_size;
2843	e = li->discard_end * li->block_size;
2844
2845	/*
2846	 * Then convert back to the current dblock size.
2847	 */
2848	b = dm_sector_div_up(b, li->cache->discard_block_size);
2849	sector_div(e, li->cache->discard_block_size);
2850
2851	/*
2852	 * The origin may have shrunk, so we need to check we're still in
2853	 * bounds.
2854	 */
2855	if (e > from_dblock(li->cache->discard_nr_blocks))
2856		e = from_dblock(li->cache->discard_nr_blocks);
2857
2858	for (; b < e; b++)
2859		set_discard(li->cache, to_dblock(b));
2860}
2861
2862static int load_discard(void *context, sector_t discard_block_size,
2863			dm_dblock_t dblock, bool discard)
2864{
2865	struct discard_load_info *li = context;
2866
2867	li->block_size = discard_block_size;
2868
2869	if (discard) {
2870		if (from_dblock(dblock) == li->discard_end)
2871			/*
2872			 * We're already in a discard range, just extend it.
2873			 */
2874			li->discard_end = li->discard_end + 1ULL;
2875
2876		else {
2877			/*
2878			 * Emit the old range and start a new one.
2879			 */
2880			set_discard_range(li);
2881			li->discard_begin = from_dblock(dblock);
2882			li->discard_end = li->discard_begin + 1ULL;
2883		}
2884	} else {
2885		set_discard_range(li);
2886		li->discard_begin = li->discard_end = 0;
2887	}
2888
2889	return 0;
2890}
2891
2892static dm_cblock_t get_cache_dev_size(struct cache *cache)
2893{
2894	sector_t size = get_dev_size(cache->cache_dev);
2895	(void) sector_div(size, cache->sectors_per_block);
2896	return to_cblock(size);
2897}
2898
2899static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2900{
2901	if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2902		if (cache->sized) {
2903			DMERR("%s: unable to extend cache due to missing cache table reload",
2904			      cache_device_name(cache));
2905			return false;
2906		}
2907	}
2908
2909	/*
2910	 * We can't drop a dirty block when shrinking the cache.
2911	 */
2912	while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2913		new_size = to_cblock(from_cblock(new_size) + 1);
2914		if (is_dirty(cache, new_size)) {
2915			DMERR("%s: unable to shrink cache; cache block %llu is dirty",
2916			      cache_device_name(cache),
2917			      (unsigned long long) from_cblock(new_size));
2918			return false;
2919		}
2920	}
2921
2922	return true;
2923}
2924
2925static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2926{
2927	int r;
2928
2929	r = dm_cache_resize(cache->cmd, new_size);
2930	if (r) {
2931		DMERR("%s: could not resize cache metadata", cache_device_name(cache));
2932		metadata_operation_failed(cache, "dm_cache_resize", r);
2933		return r;
2934	}
2935
2936	set_cache_size(cache, new_size);
2937
2938	return 0;
2939}
2940
2941static int cache_preresume(struct dm_target *ti)
2942{
2943	int r = 0;
2944	struct cache *cache = ti->private;
2945	dm_cblock_t csize = get_cache_dev_size(cache);
2946
2947	/*
2948	 * Check to see if the cache has resized.
2949	 */
2950	if (!cache->sized) {
2951		r = resize_cache_dev(cache, csize);
2952		if (r)
2953			return r;
2954
2955		cache->sized = true;
2956
2957	} else if (csize != cache->cache_size) {
2958		if (!can_resize(cache, csize))
2959			return -EINVAL;
2960
2961		r = resize_cache_dev(cache, csize);
2962		if (r)
2963			return r;
2964	}
2965
2966	if (!cache->loaded_mappings) {
2967		r = dm_cache_load_mappings(cache->cmd, cache->policy,
2968					   load_mapping, cache);
2969		if (r) {
2970			DMERR("%s: could not load cache mappings", cache_device_name(cache));
2971			metadata_operation_failed(cache, "dm_cache_load_mappings", r);
2972			return r;
2973		}
2974
2975		cache->loaded_mappings = true;
2976	}
2977
2978	if (!cache->loaded_discards) {
2979		struct discard_load_info li;
2980
2981		/*
2982		 * The discard bitset could have been resized, or the
2983		 * discard block size changed.  To be safe we start by
2984		 * setting every dblock to not discarded.
2985		 */
2986		clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2987
2988		discard_load_info_init(cache, &li);
2989		r = dm_cache_load_discards(cache->cmd, load_discard, &li);
2990		if (r) {
2991			DMERR("%s: could not load origin discards", cache_device_name(cache));
2992			metadata_operation_failed(cache, "dm_cache_load_discards", r);
2993			return r;
2994		}
2995		set_discard_range(&li);
2996
2997		cache->loaded_discards = true;
2998	}
2999
3000	return r;
3001}
3002
3003static void cache_resume(struct dm_target *ti)
3004{
3005	struct cache *cache = ti->private;
3006
3007	cache->need_tick_bio = true;
3008	allow_background_work(cache);
3009	do_waker(&cache->waker.work);
3010}
3011
3012static void emit_flags(struct cache *cache, char *result,
3013		       unsigned int maxlen, ssize_t *sz_ptr)
3014{
3015	ssize_t sz = *sz_ptr;
3016	struct cache_features *cf = &cache->features;
3017	unsigned int count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3018
3019	DMEMIT("%u ", count);
3020
3021	if (cf->metadata_version == 2)
3022		DMEMIT("metadata2 ");
3023
3024	if (writethrough_mode(cache))
3025		DMEMIT("writethrough ");
3026
3027	else if (passthrough_mode(cache))
3028		DMEMIT("passthrough ");
3029
3030	else if (writeback_mode(cache))
3031		DMEMIT("writeback ");
3032
3033	else {
3034		DMEMIT("unknown ");
3035		DMERR("%s: internal error: unknown io mode: %d",
3036		      cache_device_name(cache), (int) cf->io_mode);
3037	}
3038
3039	if (!cf->discard_passdown)
3040		DMEMIT("no_discard_passdown ");
3041
3042	*sz_ptr = sz;
3043}
3044
3045/*
3046 * Status format:
3047 *
3048 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3049 * <cache block size> <#used cache blocks>/<#total cache blocks>
3050 * <#read hits> <#read misses> <#write hits> <#write misses>
3051 * <#demotions> <#promotions> <#dirty>
3052 * <#features> <features>*
3053 * <#core args> <core args>
3054 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3055 */
3056static void cache_status(struct dm_target *ti, status_type_t type,
3057			 unsigned int status_flags, char *result, unsigned int maxlen)
3058{
3059	int r = 0;
3060	unsigned int i;
3061	ssize_t sz = 0;
3062	dm_block_t nr_free_blocks_metadata = 0;
3063	dm_block_t nr_blocks_metadata = 0;
3064	char buf[BDEVNAME_SIZE];
3065	struct cache *cache = ti->private;
3066	dm_cblock_t residency;
3067	bool needs_check;
3068
3069	switch (type) {
3070	case STATUSTYPE_INFO:
3071		if (get_cache_mode(cache) == CM_FAIL) {
3072			DMEMIT("Fail");
3073			break;
3074		}
3075
3076		/* Commit to ensure statistics aren't out-of-date */
3077		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3078			(void) commit(cache, false);
3079
3080		r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3081		if (r) {
3082			DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3083			      cache_device_name(cache), r);
3084			goto err;
3085		}
3086
3087		r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3088		if (r) {
3089			DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3090			      cache_device_name(cache), r);
3091			goto err;
3092		}
3093
3094		residency = policy_residency(cache->policy);
3095
3096		DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3097		       (unsigned int)DM_CACHE_METADATA_BLOCK_SIZE,
3098		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3099		       (unsigned long long)nr_blocks_metadata,
3100		       (unsigned long long)cache->sectors_per_block,
3101		       (unsigned long long) from_cblock(residency),
3102		       (unsigned long long) from_cblock(cache->cache_size),
3103		       (unsigned int) atomic_read(&cache->stats.read_hit),
3104		       (unsigned int) atomic_read(&cache->stats.read_miss),
3105		       (unsigned int) atomic_read(&cache->stats.write_hit),
3106		       (unsigned int) atomic_read(&cache->stats.write_miss),
3107		       (unsigned int) atomic_read(&cache->stats.demotion),
3108		       (unsigned int) atomic_read(&cache->stats.promotion),
3109		       (unsigned long) atomic_read(&cache->nr_dirty));
3110
3111		emit_flags(cache, result, maxlen, &sz);
3112
3113		DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3114
3115		DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3116		if (sz < maxlen) {
3117			r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3118			if (r)
3119				DMERR("%s: policy_emit_config_values returned %d",
3120				      cache_device_name(cache), r);
3121		}
3122
3123		if (get_cache_mode(cache) == CM_READ_ONLY)
3124			DMEMIT("ro ");
3125		else
3126			DMEMIT("rw ");
3127
3128		r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3129
3130		if (r || needs_check)
3131			DMEMIT("needs_check ");
3132		else
3133			DMEMIT("- ");
3134
3135		break;
3136
3137	case STATUSTYPE_TABLE:
3138		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3139		DMEMIT("%s ", buf);
3140		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3141		DMEMIT("%s ", buf);
3142		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3143		DMEMIT("%s", buf);
3144
3145		for (i = 0; i < cache->nr_ctr_args - 1; i++)
3146			DMEMIT(" %s", cache->ctr_args[i]);
3147		if (cache->nr_ctr_args)
3148			DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3149		break;
3150
3151	case STATUSTYPE_IMA:
3152		DMEMIT_TARGET_NAME_VERSION(ti->type);
3153		if (get_cache_mode(cache) == CM_FAIL)
3154			DMEMIT(",metadata_mode=fail");
3155		else if (get_cache_mode(cache) == CM_READ_ONLY)
3156			DMEMIT(",metadata_mode=ro");
3157		else
3158			DMEMIT(",metadata_mode=rw");
3159
3160		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3161		DMEMIT(",cache_metadata_device=%s", buf);
3162		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3163		DMEMIT(",cache_device=%s", buf);
3164		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3165		DMEMIT(",cache_origin_device=%s", buf);
3166		DMEMIT(",writethrough=%c", writethrough_mode(cache) ? 'y' : 'n');
3167		DMEMIT(",writeback=%c", writeback_mode(cache) ? 'y' : 'n');
3168		DMEMIT(",passthrough=%c", passthrough_mode(cache) ? 'y' : 'n');
3169		DMEMIT(",metadata2=%c", cache->features.metadata_version == 2 ? 'y' : 'n');
3170		DMEMIT(",no_discard_passdown=%c", cache->features.discard_passdown ? 'n' : 'y');
3171		DMEMIT(";");
3172		break;
3173	}
3174
3175	return;
3176
3177err:
3178	DMEMIT("Error");
3179}
3180
3181/*
3182 * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
3183 * the one-past-the-end value.
3184 */
3185struct cblock_range {
3186	dm_cblock_t begin;
3187	dm_cblock_t end;
3188};
3189
3190/*
3191 * A cache block range can take two forms:
3192 *
3193 * i) A single cblock, eg. '3456'
3194 * ii) A begin and end cblock with a dash between, eg. 123-234
3195 */
3196static int parse_cblock_range(struct cache *cache, const char *str,
3197			      struct cblock_range *result)
3198{
3199	char dummy;
3200	uint64_t b, e;
3201	int r;
3202
3203	/*
3204	 * Try and parse form (ii) first.
3205	 */
3206	r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3207	if (r < 0)
3208		return r;
3209
3210	if (r == 2) {
3211		result->begin = to_cblock(b);
3212		result->end = to_cblock(e);
3213		return 0;
3214	}
3215
3216	/*
3217	 * That didn't work, try form (i).
3218	 */
3219	r = sscanf(str, "%llu%c", &b, &dummy);
3220	if (r < 0)
3221		return r;
3222
3223	if (r == 1) {
3224		result->begin = to_cblock(b);
3225		result->end = to_cblock(from_cblock(result->begin) + 1u);
3226		return 0;
3227	}
3228
3229	DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3230	return -EINVAL;
3231}
3232
3233static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3234{
3235	uint64_t b = from_cblock(range->begin);
3236	uint64_t e = from_cblock(range->end);
3237	uint64_t n = from_cblock(cache->cache_size);
3238
3239	if (b >= n) {
3240		DMERR("%s: begin cblock out of range: %llu >= %llu",
3241		      cache_device_name(cache), b, n);
3242		return -EINVAL;
3243	}
3244
3245	if (e > n) {
3246		DMERR("%s: end cblock out of range: %llu > %llu",
3247		      cache_device_name(cache), e, n);
3248		return -EINVAL;
3249	}
3250
3251	if (b >= e) {
3252		DMERR("%s: invalid cblock range: %llu >= %llu",
3253		      cache_device_name(cache), b, e);
3254		return -EINVAL;
3255	}
3256
3257	return 0;
3258}
3259
3260static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3261{
3262	return to_cblock(from_cblock(b) + 1);
3263}
3264
3265static int request_invalidation(struct cache *cache, struct cblock_range *range)
3266{
3267	int r = 0;
3268
3269	/*
3270	 * We don't need to do any locking here because we know we're in
3271	 * passthrough mode.  There's is potential for a race between an
3272	 * invalidation triggered by an io and an invalidation message.  This
3273	 * is harmless, we must not worry if the policy call fails.
3274	 */
3275	while (range->begin != range->end) {
3276		r = invalidate_cblock(cache, range->begin);
3277		if (r)
3278			return r;
3279
3280		range->begin = cblock_succ(range->begin);
3281	}
3282
3283	cache->commit_requested = true;
3284	return r;
3285}
3286
3287static int process_invalidate_cblocks_message(struct cache *cache, unsigned int count,
3288					      const char **cblock_ranges)
3289{
3290	int r = 0;
3291	unsigned int i;
3292	struct cblock_range range;
3293
3294	if (!passthrough_mode(cache)) {
3295		DMERR("%s: cache has to be in passthrough mode for invalidation",
3296		      cache_device_name(cache));
3297		return -EPERM;
3298	}
3299
3300	for (i = 0; i < count; i++) {
3301		r = parse_cblock_range(cache, cblock_ranges[i], &range);
3302		if (r)
3303			break;
3304
3305		r = validate_cblock_range(cache, &range);
3306		if (r)
3307			break;
3308
3309		/*
3310		 * Pass begin and end origin blocks to the worker and wake it.
3311		 */
3312		r = request_invalidation(cache, &range);
3313		if (r)
3314			break;
3315	}
3316
3317	return r;
3318}
3319
3320/*
3321 * Supports
3322 *	"<key> <value>"
3323 * and
3324 *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3325 *
3326 * The key migration_threshold is supported by the cache target core.
3327 */
3328static int cache_message(struct dm_target *ti, unsigned int argc, char **argv,
3329			 char *result, unsigned int maxlen)
3330{
3331	struct cache *cache = ti->private;
3332
3333	if (!argc)
3334		return -EINVAL;
3335
3336	if (get_cache_mode(cache) >= CM_READ_ONLY) {
3337		DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3338		      cache_device_name(cache));
3339		return -EOPNOTSUPP;
3340	}
3341
3342	if (!strcasecmp(argv[0], "invalidate_cblocks"))
3343		return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3344
3345	if (argc != 2)
3346		return -EINVAL;
3347
3348	return set_config_value(cache, argv[0], argv[1]);
3349}
3350
3351static int cache_iterate_devices(struct dm_target *ti,
3352				 iterate_devices_callout_fn fn, void *data)
3353{
3354	int r = 0;
3355	struct cache *cache = ti->private;
3356
3357	r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3358	if (!r)
3359		r = fn(ti, cache->origin_dev, 0, ti->len, data);
3360
3361	return r;
3362}
3363
3364/*
3365 * If discard_passdown was enabled verify that the origin device
3366 * supports discards.  Disable discard_passdown if not.
3367 */
3368static void disable_passdown_if_not_supported(struct cache *cache)
3369{
3370	struct block_device *origin_bdev = cache->origin_dev->bdev;
3371	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3372	const char *reason = NULL;
3373
3374	if (!cache->features.discard_passdown)
3375		return;
3376
3377	if (!bdev_max_discard_sectors(origin_bdev))
3378		reason = "discard unsupported";
3379
3380	else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3381		reason = "max discard sectors smaller than a block";
3382
3383	if (reason) {
3384		DMWARN("Origin device (%pg) %s: Disabling discard passdown.",
3385		       origin_bdev, reason);
3386		cache->features.discard_passdown = false;
3387	}
3388}
3389
3390static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3391{
3392	struct block_device *origin_bdev = cache->origin_dev->bdev;
3393	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3394
3395	if (!cache->features.discard_passdown) {
3396		/* No passdown is done so setting own virtual limits */
3397		limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3398						    cache->origin_sectors);
3399		limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3400		return;
3401	}
3402
3403	/*
3404	 * cache_iterate_devices() is stacking both origin and fast device limits
3405	 * but discards aren't passed to fast device, so inherit origin's limits.
3406	 */
3407	limits->max_discard_sectors = origin_limits->max_discard_sectors;
3408	limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3409	limits->discard_granularity = origin_limits->discard_granularity;
3410	limits->discard_alignment = origin_limits->discard_alignment;
3411	limits->discard_misaligned = origin_limits->discard_misaligned;
3412}
3413
3414static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3415{
3416	struct cache *cache = ti->private;
3417	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3418
3419	/*
3420	 * If the system-determined stacked limits are compatible with the
3421	 * cache's blocksize (io_opt is a factor) do not override them.
3422	 */
3423	if (io_opt_sectors < cache->sectors_per_block ||
3424	    do_div(io_opt_sectors, cache->sectors_per_block)) {
3425		blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3426		blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3427	}
3428
3429	disable_passdown_if_not_supported(cache);
3430	set_discard_limits(cache, limits);
3431}
3432
3433/*----------------------------------------------------------------*/
3434
3435static struct target_type cache_target = {
3436	.name = "cache",
3437	.version = {2, 2, 0},
3438	.module = THIS_MODULE,
3439	.ctr = cache_ctr,
3440	.dtr = cache_dtr,
3441	.map = cache_map,
3442	.end_io = cache_end_io,
3443	.postsuspend = cache_postsuspend,
3444	.preresume = cache_preresume,
3445	.resume = cache_resume,
3446	.status = cache_status,
3447	.message = cache_message,
3448	.iterate_devices = cache_iterate_devices,
3449	.io_hints = cache_io_hints,
3450};
3451
3452static int __init dm_cache_init(void)
3453{
3454	int r;
3455
3456	migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3457	if (!migration_cache)
3458		return -ENOMEM;
3459
3460	r = dm_register_target(&cache_target);
3461	if (r) {
 
3462		kmem_cache_destroy(migration_cache);
3463		return r;
3464	}
3465
3466	return 0;
3467}
3468
3469static void __exit dm_cache_exit(void)
3470{
3471	dm_unregister_target(&cache_target);
3472	kmem_cache_destroy(migration_cache);
3473}
3474
3475module_init(dm_cache_init);
3476module_exit(dm_cache_exit);
3477
3478MODULE_DESCRIPTION(DM_NAME " cache target");
3479MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3480MODULE_LICENSE("GPL");