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");
v5.4
   1/*
   2 * Copyright (C) 2012 Red Hat. All rights reserved.
   3 *
   4 * This file is released under the GPL.
   5 */
   6
   7#include "dm.h"
   8#include "dm-bio-prison-v2.h"
   9#include "dm-bio-record.h"
  10#include "dm-cache-metadata.h"
 
  11
  12#include <linux/dm-io.h>
  13#include <linux/dm-kcopyd.h>
  14#include <linux/jiffies.h>
  15#include <linux/init.h>
  16#include <linux/mempool.h>
  17#include <linux/module.h>
  18#include <linux/rwsem.h>
  19#include <linux/slab.h>
  20#include <linux/vmalloc.h>
  21
  22#define DM_MSG_PREFIX "cache"
  23
  24DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  25	"A percentage of time allocated for copying to and/or from cache");
  26
  27/*----------------------------------------------------------------*/
  28
  29/*
  30 * Glossary:
  31 *
  32 * oblock: index of an origin block
  33 * cblock: index of a cache block
  34 * promotion: movement of a block from origin to cache
  35 * demotion: movement of a block from cache to origin
  36 * migration: movement of a block between the origin and cache device,
  37 *	      either direction
  38 */
  39
  40/*----------------------------------------------------------------*/
  41
  42struct io_tracker {
  43	spinlock_t lock;
  44
  45	/*
  46	 * Sectors of in-flight IO.
  47	 */
  48	sector_t in_flight;
  49
  50	/*
  51	 * The time, in jiffies, when this device became idle (if it is
  52	 * indeed idle).
  53	 */
  54	unsigned long idle_time;
  55	unsigned long last_update_time;
  56};
  57
  58static void iot_init(struct io_tracker *iot)
  59{
  60	spin_lock_init(&iot->lock);
  61	iot->in_flight = 0ul;
  62	iot->idle_time = 0ul;
  63	iot->last_update_time = jiffies;
  64}
  65
  66static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  67{
  68	if (iot->in_flight)
  69		return false;
  70
  71	return time_after(jiffies, iot->idle_time + jifs);
  72}
  73
  74static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  75{
  76	bool r;
  77	unsigned long flags;
  78
  79	spin_lock_irqsave(&iot->lock, flags);
  80	r = __iot_idle_for(iot, jifs);
  81	spin_unlock_irqrestore(&iot->lock, flags);
  82
  83	return r;
  84}
  85
  86static void iot_io_begin(struct io_tracker *iot, sector_t len)
  87{
  88	unsigned long flags;
  89
  90	spin_lock_irqsave(&iot->lock, flags);
  91	iot->in_flight += len;
  92	spin_unlock_irqrestore(&iot->lock, flags);
  93}
  94
  95static void __iot_io_end(struct io_tracker *iot, sector_t len)
  96{
  97	if (!len)
  98		return;
  99
 100	iot->in_flight -= len;
 101	if (!iot->in_flight)
 102		iot->idle_time = jiffies;
 103}
 104
 105static void iot_io_end(struct io_tracker *iot, sector_t len)
 106{
 107	unsigned long flags;
 108
 109	spin_lock_irqsave(&iot->lock, flags);
 110	__iot_io_end(iot, len);
 111	spin_unlock_irqrestore(&iot->lock, flags);
 112}
 113
 114/*----------------------------------------------------------------*/
 115
 116/*
 117 * Represents a chunk of future work.  'input' allows continuations to pass
 118 * values between themselves, typically error values.
 119 */
 120struct continuation {
 121	struct work_struct ws;
 122	blk_status_t input;
 123};
 124
 125static inline void init_continuation(struct continuation *k,
 126				     void (*fn)(struct work_struct *))
 127{
 128	INIT_WORK(&k->ws, fn);
 129	k->input = 0;
 130}
 131
 132static inline void queue_continuation(struct workqueue_struct *wq,
 133				      struct continuation *k)
 134{
 135	queue_work(wq, &k->ws);
 136}
 137
 138/*----------------------------------------------------------------*/
 139
 140/*
 141 * The batcher collects together pieces of work that need a particular
 142 * operation to occur before they can proceed (typically a commit).
 143 */
 144struct batcher {
 145	/*
 146	 * The operation that everyone is waiting for.
 147	 */
 148	blk_status_t (*commit_op)(void *context);
 149	void *commit_context;
 150
 151	/*
 152	 * This is how bios should be issued once the commit op is complete
 153	 * (accounted_request).
 154	 */
 155	void (*issue_op)(struct bio *bio, void *context);
 156	void *issue_context;
 157
 158	/*
 159	 * Queued work gets put on here after commit.
 160	 */
 161	struct workqueue_struct *wq;
 162
 163	spinlock_t lock;
 164	struct list_head work_items;
 165	struct bio_list bios;
 166	struct work_struct commit_work;
 167
 168	bool commit_scheduled;
 169};
 170
 171static void __commit(struct work_struct *_ws)
 172{
 173	struct batcher *b = container_of(_ws, struct batcher, commit_work);
 174	blk_status_t r;
 175	unsigned long flags;
 176	struct list_head work_items;
 177	struct work_struct *ws, *tmp;
 178	struct continuation *k;
 179	struct bio *bio;
 180	struct bio_list bios;
 181
 182	INIT_LIST_HEAD(&work_items);
 183	bio_list_init(&bios);
 184
 185	/*
 186	 * We have to grab these before the commit_op to avoid a race
 187	 * condition.
 188	 */
 189	spin_lock_irqsave(&b->lock, flags);
 190	list_splice_init(&b->work_items, &work_items);
 191	bio_list_merge(&bios, &b->bios);
 192	bio_list_init(&b->bios);
 193	b->commit_scheduled = false;
 194	spin_unlock_irqrestore(&b->lock, flags);
 195
 196	r = b->commit_op(b->commit_context);
 197
 198	list_for_each_entry_safe(ws, tmp, &work_items, entry) {
 199		k = container_of(ws, struct continuation, ws);
 200		k->input = r;
 201		INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
 202		queue_work(b->wq, ws);
 203	}
 204
 205	while ((bio = bio_list_pop(&bios))) {
 206		if (r) {
 207			bio->bi_status = r;
 208			bio_endio(bio);
 209		} else
 210			b->issue_op(bio, b->issue_context);
 211	}
 212}
 213
 214static void batcher_init(struct batcher *b,
 215			 blk_status_t (*commit_op)(void *),
 216			 void *commit_context,
 217			 void (*issue_op)(struct bio *bio, void *),
 218			 void *issue_context,
 219			 struct workqueue_struct *wq)
 220{
 221	b->commit_op = commit_op;
 222	b->commit_context = commit_context;
 223	b->issue_op = issue_op;
 224	b->issue_context = issue_context;
 225	b->wq = wq;
 226
 227	spin_lock_init(&b->lock);
 228	INIT_LIST_HEAD(&b->work_items);
 229	bio_list_init(&b->bios);
 230	INIT_WORK(&b->commit_work, __commit);
 231	b->commit_scheduled = false;
 232}
 233
 234static void async_commit(struct batcher *b)
 235{
 236	queue_work(b->wq, &b->commit_work);
 237}
 238
 239static void continue_after_commit(struct batcher *b, struct continuation *k)
 240{
 241	unsigned long flags;
 242	bool commit_scheduled;
 243
 244	spin_lock_irqsave(&b->lock, flags);
 245	commit_scheduled = b->commit_scheduled;
 246	list_add_tail(&k->ws.entry, &b->work_items);
 247	spin_unlock_irqrestore(&b->lock, flags);
 248
 249	if (commit_scheduled)
 250		async_commit(b);
 251}
 252
 253/*
 254 * Bios are errored if commit failed.
 255 */
 256static void issue_after_commit(struct batcher *b, struct bio *bio)
 257{
 258       unsigned long flags;
 259       bool commit_scheduled;
 260
 261       spin_lock_irqsave(&b->lock, flags);
 262       commit_scheduled = b->commit_scheduled;
 263       bio_list_add(&b->bios, bio);
 264       spin_unlock_irqrestore(&b->lock, flags);
 265
 266       if (commit_scheduled)
 267	       async_commit(b);
 268}
 269
 270/*
 271 * Call this if some urgent work is waiting for the commit to complete.
 272 */
 273static void schedule_commit(struct batcher *b)
 274{
 275	bool immediate;
 276	unsigned long flags;
 277
 278	spin_lock_irqsave(&b->lock, flags);
 279	immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
 280	b->commit_scheduled = true;
 281	spin_unlock_irqrestore(&b->lock, flags);
 282
 283	if (immediate)
 284		async_commit(b);
 285}
 286
 287/*
 288 * There are a couple of places where we let a bio run, but want to do some
 289 * work before calling its endio function.  We do this by temporarily
 290 * changing the endio fn.
 291 */
 292struct dm_hook_info {
 293	bio_end_io_t *bi_end_io;
 294};
 295
 296static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 297			bio_end_io_t *bi_end_io, void *bi_private)
 298{
 299	h->bi_end_io = bio->bi_end_io;
 300
 301	bio->bi_end_io = bi_end_io;
 302	bio->bi_private = bi_private;
 303}
 304
 305static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 306{
 307	bio->bi_end_io = h->bi_end_io;
 308}
 309
 310/*----------------------------------------------------------------*/
 311
 312#define MIGRATION_POOL_SIZE 128
 313#define COMMIT_PERIOD HZ
 314#define MIGRATION_COUNT_WINDOW 10
 315
 316/*
 317 * The block size of the device holding cache data must be
 318 * between 32KB and 1GB.
 319 */
 320#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 321#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 322
 323enum cache_metadata_mode {
 324	CM_WRITE,		/* metadata may be changed */
 325	CM_READ_ONLY,		/* metadata may not be changed */
 326	CM_FAIL
 327};
 328
 329enum cache_io_mode {
 330	/*
 331	 * Data is written to cached blocks only.  These blocks are marked
 332	 * dirty.  If you lose the cache device you will lose data.
 333	 * Potential performance increase for both reads and writes.
 334	 */
 335	CM_IO_WRITEBACK,
 336
 337	/*
 338	 * Data is written to both cache and origin.  Blocks are never
 339	 * dirty.  Potential performance benfit for reads only.
 340	 */
 341	CM_IO_WRITETHROUGH,
 342
 343	/*
 344	 * A degraded mode useful for various cache coherency situations
 345	 * (eg, rolling back snapshots).  Reads and writes always go to the
 346	 * origin.  If a write goes to a cached oblock, then the cache
 347	 * block is invalidated.
 348	 */
 349	CM_IO_PASSTHROUGH
 350};
 351
 352struct cache_features {
 353	enum cache_metadata_mode mode;
 354	enum cache_io_mode io_mode;
 355	unsigned metadata_version;
 356	bool discard_passdown:1;
 357};
 358
 359struct cache_stats {
 360	atomic_t read_hit;
 361	atomic_t read_miss;
 362	atomic_t write_hit;
 363	atomic_t write_miss;
 364	atomic_t demotion;
 365	atomic_t promotion;
 366	atomic_t writeback;
 367	atomic_t copies_avoided;
 368	atomic_t cache_cell_clash;
 369	atomic_t commit_count;
 370	atomic_t discard_count;
 371};
 372
 373struct cache {
 374	struct dm_target *ti;
 375	spinlock_t lock;
 376
 377	/*
 378	 * Fields for converting from sectors to blocks.
 379	 */
 380	int sectors_per_block_shift;
 381	sector_t sectors_per_block;
 382
 383	struct dm_cache_metadata *cmd;
 384
 385	/*
 386	 * Metadata is written to this device.
 387	 */
 388	struct dm_dev *metadata_dev;
 389
 390	/*
 391	 * The slower of the two data devices.  Typically a spindle.
 392	 */
 393	struct dm_dev *origin_dev;
 394
 395	/*
 396	 * The faster of the two data devices.  Typically an SSD.
 397	 */
 398	struct dm_dev *cache_dev;
 399
 400	/*
 401	 * Size of the origin device in _complete_ blocks and native sectors.
 402	 */
 403	dm_oblock_t origin_blocks;
 404	sector_t origin_sectors;
 405
 406	/*
 407	 * Size of the cache device in blocks.
 408	 */
 409	dm_cblock_t cache_size;
 410
 411	/*
 412	 * Invalidation fields.
 413	 */
 414	spinlock_t invalidation_lock;
 415	struct list_head invalidation_requests;
 416
 417	sector_t migration_threshold;
 418	wait_queue_head_t migration_wait;
 419	atomic_t nr_allocated_migrations;
 420
 421	/*
 422	 * The number of in flight migrations that are performing
 423	 * background io. eg, promotion, writeback.
 424	 */
 425	atomic_t nr_io_migrations;
 426
 427	struct bio_list deferred_bios;
 428
 429	struct rw_semaphore quiesce_lock;
 430
 431	struct dm_target_callbacks callbacks;
 432
 433	/*
 434	 * origin_blocks entries, discarded if set.
 435	 */
 436	dm_dblock_t discard_nr_blocks;
 437	unsigned long *discard_bitset;
 438	uint32_t discard_block_size; /* a power of 2 times sectors per block */
 439
 440	/*
 441	 * Rather than reconstructing the table line for the status we just
 442	 * save it and regurgitate.
 443	 */
 444	unsigned nr_ctr_args;
 445	const char **ctr_args;
 446
 447	struct dm_kcopyd_client *copier;
 448	struct work_struct deferred_bio_worker;
 449	struct work_struct migration_worker;
 450	struct workqueue_struct *wq;
 451	struct delayed_work waker;
 452	struct dm_bio_prison_v2 *prison;
 453
 454	/*
 455	 * cache_size entries, dirty if set
 456	 */
 457	unsigned long *dirty_bitset;
 458	atomic_t nr_dirty;
 459
 460	unsigned policy_nr_args;
 461	struct dm_cache_policy *policy;
 462
 463	/*
 464	 * Cache features such as write-through.
 465	 */
 466	struct cache_features features;
 467
 468	struct cache_stats stats;
 469
 470	bool need_tick_bio:1;
 471	bool sized:1;
 472	bool invalidate:1;
 473	bool commit_requested:1;
 474	bool loaded_mappings:1;
 475	bool loaded_discards:1;
 476
 477	struct rw_semaphore background_work_lock;
 478
 479	struct batcher committer;
 480	struct work_struct commit_ws;
 481
 482	struct io_tracker tracker;
 483
 484	mempool_t migration_pool;
 485
 486	struct bio_set bs;
 487};
 488
 489struct per_bio_data {
 490	bool tick:1;
 491	unsigned req_nr:2;
 492	struct dm_bio_prison_cell_v2 *cell;
 493	struct dm_hook_info hook_info;
 494	sector_t len;
 495};
 496
 497struct dm_cache_migration {
 498	struct continuation k;
 499	struct cache *cache;
 500
 501	struct policy_work *op;
 502	struct bio *overwrite_bio;
 503	struct dm_bio_prison_cell_v2 *cell;
 504
 505	dm_cblock_t invalidate_cblock;
 506	dm_oblock_t invalidate_oblock;
 507};
 508
 509/*----------------------------------------------------------------*/
 510
 511static bool writethrough_mode(struct cache *cache)
 512{
 513	return cache->features.io_mode == CM_IO_WRITETHROUGH;
 514}
 515
 516static bool writeback_mode(struct cache *cache)
 517{
 518	return cache->features.io_mode == CM_IO_WRITEBACK;
 519}
 520
 521static inline bool passthrough_mode(struct cache *cache)
 522{
 523	return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
 524}
 525
 526/*----------------------------------------------------------------*/
 527
 528static void wake_deferred_bio_worker(struct cache *cache)
 529{
 530	queue_work(cache->wq, &cache->deferred_bio_worker);
 531}
 532
 533static void wake_migration_worker(struct cache *cache)
 534{
 535	if (passthrough_mode(cache))
 536		return;
 537
 538	queue_work(cache->wq, &cache->migration_worker);
 539}
 540
 541/*----------------------------------------------------------------*/
 542
 543static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
 544{
 545	return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
 546}
 547
 548static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
 549{
 550	dm_bio_prison_free_cell_v2(cache->prison, cell);
 551}
 552
 553static struct dm_cache_migration *alloc_migration(struct cache *cache)
 554{
 555	struct dm_cache_migration *mg;
 556
 557	mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
 558
 559	memset(mg, 0, sizeof(*mg));
 560
 561	mg->cache = cache;
 562	atomic_inc(&cache->nr_allocated_migrations);
 563
 564	return mg;
 565}
 566
 567static void free_migration(struct dm_cache_migration *mg)
 568{
 569	struct cache *cache = mg->cache;
 570
 571	if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 572		wake_up(&cache->migration_wait);
 573
 574	mempool_free(mg, &cache->migration_pool);
 575}
 576
 577/*----------------------------------------------------------------*/
 578
 579static inline dm_oblock_t oblock_succ(dm_oblock_t b)
 580{
 581	return to_oblock(from_oblock(b) + 1ull);
 582}
 583
 584static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
 585{
 586	key->virtual = 0;
 587	key->dev = 0;
 588	key->block_begin = from_oblock(begin);
 589	key->block_end = from_oblock(end);
 590}
 591
 592/*
 593 * We have two lock levels.  Level 0, which is used to prevent WRITEs, and
 594 * level 1 which prevents *both* READs and WRITEs.
 595 */
 596#define WRITE_LOCK_LEVEL 0
 597#define READ_WRITE_LOCK_LEVEL 1
 598
 599static unsigned lock_level(struct bio *bio)
 600{
 601	return bio_data_dir(bio) == WRITE ?
 602		WRITE_LOCK_LEVEL :
 603		READ_WRITE_LOCK_LEVEL;
 604}
 605
 606/*----------------------------------------------------------------
 607 * Per bio data
 608 *--------------------------------------------------------------*/
 609
 610static struct per_bio_data *get_per_bio_data(struct bio *bio)
 611{
 612	struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 613	BUG_ON(!pb);
 614	return pb;
 615}
 616
 617static struct per_bio_data *init_per_bio_data(struct bio *bio)
 618{
 619	struct per_bio_data *pb = get_per_bio_data(bio);
 620
 621	pb->tick = false;
 622	pb->req_nr = dm_bio_get_target_bio_nr(bio);
 623	pb->cell = NULL;
 624	pb->len = 0;
 625
 626	return pb;
 627}
 628
 629/*----------------------------------------------------------------*/
 630
 631static void defer_bio(struct cache *cache, struct bio *bio)
 632{
 633	unsigned long flags;
 634
 635	spin_lock_irqsave(&cache->lock, flags);
 636	bio_list_add(&cache->deferred_bios, bio);
 637	spin_unlock_irqrestore(&cache->lock, flags);
 638
 639	wake_deferred_bio_worker(cache);
 640}
 641
 642static void defer_bios(struct cache *cache, struct bio_list *bios)
 643{
 644	unsigned long flags;
 645
 646	spin_lock_irqsave(&cache->lock, flags);
 647	bio_list_merge(&cache->deferred_bios, bios);
 648	bio_list_init(bios);
 649	spin_unlock_irqrestore(&cache->lock, flags);
 650
 651	wake_deferred_bio_worker(cache);
 652}
 653
 654/*----------------------------------------------------------------*/
 655
 656static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
 657{
 658	bool r;
 659	struct per_bio_data *pb;
 660	struct dm_cell_key_v2 key;
 661	dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 662	struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
 663
 664	cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
 665
 666	build_key(oblock, end, &key);
 667	r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
 668	if (!r) {
 669		/*
 670		 * Failed to get the lock.
 671		 */
 672		free_prison_cell(cache, cell_prealloc);
 673		return r;
 674	}
 675
 676	if (cell != cell_prealloc)
 677		free_prison_cell(cache, cell_prealloc);
 678
 679	pb = get_per_bio_data(bio);
 680	pb->cell = cell;
 681
 682	return r;
 683}
 684
 685/*----------------------------------------------------------------*/
 686
 687static bool is_dirty(struct cache *cache, dm_cblock_t b)
 688{
 689	return test_bit(from_cblock(b), cache->dirty_bitset);
 690}
 691
 692static void set_dirty(struct cache *cache, dm_cblock_t cblock)
 693{
 694	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 695		atomic_inc(&cache->nr_dirty);
 696		policy_set_dirty(cache->policy, cblock);
 697	}
 698}
 699
 700/*
 701 * These two are called when setting after migrations to force the policy
 702 * and dirty bitset to be in sync.
 703 */
 704static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
 705{
 706	if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
 707		atomic_inc(&cache->nr_dirty);
 708	policy_set_dirty(cache->policy, cblock);
 709}
 710
 711static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
 712{
 713	if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 714		if (atomic_dec_return(&cache->nr_dirty) == 0)
 715			dm_table_event(cache->ti->table);
 716	}
 717
 718	policy_clear_dirty(cache->policy, cblock);
 719}
 720
 721/*----------------------------------------------------------------*/
 722
 723static bool block_size_is_power_of_two(struct cache *cache)
 724{
 725	return cache->sectors_per_block_shift >= 0;
 726}
 727
 728/* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
 729#if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
 730__always_inline
 731#endif
 732static dm_block_t block_div(dm_block_t b, uint32_t n)
 733{
 734	do_div(b, n);
 735
 736	return b;
 737}
 738
 739static dm_block_t oblocks_per_dblock(struct cache *cache)
 740{
 741	dm_block_t oblocks = cache->discard_block_size;
 742
 743	if (block_size_is_power_of_two(cache))
 744		oblocks >>= cache->sectors_per_block_shift;
 745	else
 746		oblocks = block_div(oblocks, cache->sectors_per_block);
 747
 748	return oblocks;
 749}
 750
 751static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 752{
 753	return to_dblock(block_div(from_oblock(oblock),
 754				   oblocks_per_dblock(cache)));
 755}
 756
 757static void set_discard(struct cache *cache, dm_dblock_t b)
 758{
 759	unsigned long flags;
 760
 761	BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 762	atomic_inc(&cache->stats.discard_count);
 763
 764	spin_lock_irqsave(&cache->lock, flags);
 765	set_bit(from_dblock(b), cache->discard_bitset);
 766	spin_unlock_irqrestore(&cache->lock, flags);
 767}
 768
 769static void clear_discard(struct cache *cache, dm_dblock_t b)
 770{
 771	unsigned long flags;
 772
 773	spin_lock_irqsave(&cache->lock, flags);
 774	clear_bit(from_dblock(b), cache->discard_bitset);
 775	spin_unlock_irqrestore(&cache->lock, flags);
 776}
 777
 778static bool is_discarded(struct cache *cache, dm_dblock_t b)
 779{
 780	int r;
 781	unsigned long flags;
 782
 783	spin_lock_irqsave(&cache->lock, flags);
 784	r = test_bit(from_dblock(b), cache->discard_bitset);
 785	spin_unlock_irqrestore(&cache->lock, flags);
 786
 787	return r;
 788}
 789
 790static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 791{
 792	int r;
 793	unsigned long flags;
 794
 795	spin_lock_irqsave(&cache->lock, flags);
 796	r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 797		     cache->discard_bitset);
 798	spin_unlock_irqrestore(&cache->lock, flags);
 799
 800	return r;
 801}
 802
 803/*----------------------------------------------------------------
 804 * Remapping
 805 *--------------------------------------------------------------*/
 806static void remap_to_origin(struct cache *cache, struct bio *bio)
 807{
 808	bio_set_dev(bio, cache->origin_dev->bdev);
 809}
 810
 811static void remap_to_cache(struct cache *cache, struct bio *bio,
 812			   dm_cblock_t cblock)
 813{
 814	sector_t bi_sector = bio->bi_iter.bi_sector;
 815	sector_t block = from_cblock(cblock);
 816
 817	bio_set_dev(bio, cache->cache_dev->bdev);
 818	if (!block_size_is_power_of_two(cache))
 819		bio->bi_iter.bi_sector =
 820			(block * cache->sectors_per_block) +
 821			sector_div(bi_sector, cache->sectors_per_block);
 822	else
 823		bio->bi_iter.bi_sector =
 824			(block << cache->sectors_per_block_shift) |
 825			(bi_sector & (cache->sectors_per_block - 1));
 826}
 827
 828static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 829{
 830	unsigned long flags;
 831	struct per_bio_data *pb;
 832
 833	spin_lock_irqsave(&cache->lock, flags);
 834	if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
 835	    bio_op(bio) != REQ_OP_DISCARD) {
 836		pb = get_per_bio_data(bio);
 837		pb->tick = true;
 838		cache->need_tick_bio = false;
 839	}
 840	spin_unlock_irqrestore(&cache->lock, flags);
 841}
 842
 843static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 844					    dm_oblock_t oblock, bool bio_has_pbd)
 845{
 846	if (bio_has_pbd)
 847		check_if_tick_bio_needed(cache, bio);
 848	remap_to_origin(cache, bio);
 849	if (bio_data_dir(bio) == WRITE)
 850		clear_discard(cache, oblock_to_dblock(cache, oblock));
 851}
 852
 853static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 854					  dm_oblock_t oblock)
 855{
 856	// FIXME: check_if_tick_bio_needed() is called way too much through this interface
 857	__remap_to_origin_clear_discard(cache, bio, oblock, true);
 858}
 859
 860static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 861				 dm_oblock_t oblock, dm_cblock_t cblock)
 862{
 863	check_if_tick_bio_needed(cache, bio);
 864	remap_to_cache(cache, bio, cblock);
 865	if (bio_data_dir(bio) == WRITE) {
 866		set_dirty(cache, cblock);
 867		clear_discard(cache, oblock_to_dblock(cache, oblock));
 868	}
 869}
 870
 871static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 872{
 873	sector_t block_nr = bio->bi_iter.bi_sector;
 874
 875	if (!block_size_is_power_of_two(cache))
 876		(void) sector_div(block_nr, cache->sectors_per_block);
 877	else
 878		block_nr >>= cache->sectors_per_block_shift;
 879
 880	return to_oblock(block_nr);
 881}
 882
 883static bool accountable_bio(struct cache *cache, struct bio *bio)
 884{
 885	return bio_op(bio) != REQ_OP_DISCARD;
 886}
 887
 888static void accounted_begin(struct cache *cache, struct bio *bio)
 889{
 890	struct per_bio_data *pb;
 891
 892	if (accountable_bio(cache, bio)) {
 893		pb = get_per_bio_data(bio);
 894		pb->len = bio_sectors(bio);
 895		iot_io_begin(&cache->tracker, pb->len);
 896	}
 897}
 898
 899static void accounted_complete(struct cache *cache, struct bio *bio)
 900{
 901	struct per_bio_data *pb = get_per_bio_data(bio);
 902
 903	iot_io_end(&cache->tracker, pb->len);
 904}
 905
 906static void accounted_request(struct cache *cache, struct bio *bio)
 907{
 908	accounted_begin(cache, bio);
 909	generic_make_request(bio);
 910}
 911
 912static void issue_op(struct bio *bio, void *context)
 913{
 914	struct cache *cache = context;
 915	accounted_request(cache, bio);
 916}
 917
 918/*
 919 * When running in writethrough mode we need to send writes to clean blocks
 920 * to both the cache and origin devices.  Clone the bio and send them in parallel.
 921 */
 922static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
 923				      dm_oblock_t oblock, dm_cblock_t cblock)
 924{
 925	struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
 
 926
 927	BUG_ON(!origin_bio);
 928
 929	bio_chain(origin_bio, bio);
 930	/*
 931	 * Passing false to __remap_to_origin_clear_discard() skips
 932	 * all code that might use per_bio_data (since clone doesn't have it)
 933	 */
 934	__remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
 935	submit_bio(origin_bio);
 936
 937	remap_to_cache(cache, bio, cblock);
 938}
 939
 940/*----------------------------------------------------------------
 941 * Failure modes
 942 *--------------------------------------------------------------*/
 943static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 944{
 945	return cache->features.mode;
 946}
 947
 948static const char *cache_device_name(struct cache *cache)
 949{
 950	return dm_device_name(dm_table_get_md(cache->ti->table));
 951}
 952
 953static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 954{
 955	const char *descs[] = {
 956		"write",
 957		"read-only",
 958		"fail"
 959	};
 960
 961	dm_table_event(cache->ti->table);
 962	DMINFO("%s: switching cache to %s mode",
 963	       cache_device_name(cache), descs[(int)mode]);
 964}
 965
 966static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 967{
 968	bool needs_check;
 969	enum cache_metadata_mode old_mode = get_cache_mode(cache);
 970
 971	if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
 972		DMERR("%s: unable to read needs_check flag, setting failure mode.",
 973		      cache_device_name(cache));
 974		new_mode = CM_FAIL;
 975	}
 976
 977	if (new_mode == CM_WRITE && needs_check) {
 978		DMERR("%s: unable to switch cache to write mode until repaired.",
 979		      cache_device_name(cache));
 980		if (old_mode != new_mode)
 981			new_mode = old_mode;
 982		else
 983			new_mode = CM_READ_ONLY;
 984	}
 985
 986	/* Never move out of fail mode */
 987	if (old_mode == CM_FAIL)
 988		new_mode = CM_FAIL;
 989
 990	switch (new_mode) {
 991	case CM_FAIL:
 992	case CM_READ_ONLY:
 993		dm_cache_metadata_set_read_only(cache->cmd);
 994		break;
 995
 996	case CM_WRITE:
 997		dm_cache_metadata_set_read_write(cache->cmd);
 998		break;
 999	}
1000
1001	cache->features.mode = new_mode;
1002
1003	if (new_mode != old_mode)
1004		notify_mode_switch(cache, new_mode);
1005}
1006
1007static void abort_transaction(struct cache *cache)
1008{
1009	const char *dev_name = cache_device_name(cache);
1010
1011	if (get_cache_mode(cache) >= CM_READ_ONLY)
1012		return;
1013
1014	if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1015		DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
 
1016		set_cache_mode(cache, CM_FAIL);
1017	}
1018
1019	DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1020	if (dm_cache_metadata_abort(cache->cmd)) {
1021		DMERR("%s: failed to abort metadata transaction", dev_name);
1022		set_cache_mode(cache, CM_FAIL);
1023	}
1024}
1025
1026static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1027{
1028	DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1029		    cache_device_name(cache), op, r);
1030	abort_transaction(cache);
1031	set_cache_mode(cache, CM_READ_ONLY);
1032}
1033
1034/*----------------------------------------------------------------*/
1035
1036static void load_stats(struct cache *cache)
1037{
1038	struct dm_cache_statistics stats;
1039
1040	dm_cache_metadata_get_stats(cache->cmd, &stats);
1041	atomic_set(&cache->stats.read_hit, stats.read_hits);
1042	atomic_set(&cache->stats.read_miss, stats.read_misses);
1043	atomic_set(&cache->stats.write_hit, stats.write_hits);
1044	atomic_set(&cache->stats.write_miss, stats.write_misses);
1045}
1046
1047static void save_stats(struct cache *cache)
1048{
1049	struct dm_cache_statistics stats;
1050
1051	if (get_cache_mode(cache) >= CM_READ_ONLY)
1052		return;
1053
1054	stats.read_hits = atomic_read(&cache->stats.read_hit);
1055	stats.read_misses = atomic_read(&cache->stats.read_miss);
1056	stats.write_hits = atomic_read(&cache->stats.write_hit);
1057	stats.write_misses = atomic_read(&cache->stats.write_miss);
1058
1059	dm_cache_metadata_set_stats(cache->cmd, &stats);
1060}
1061
1062static void update_stats(struct cache_stats *stats, enum policy_operation op)
1063{
1064	switch (op) {
1065	case POLICY_PROMOTE:
1066		atomic_inc(&stats->promotion);
1067		break;
1068
1069	case POLICY_DEMOTE:
1070		atomic_inc(&stats->demotion);
1071		break;
1072
1073	case POLICY_WRITEBACK:
1074		atomic_inc(&stats->writeback);
1075		break;
1076	}
1077}
1078
1079/*----------------------------------------------------------------
1080 * Migration processing
1081 *
1082 * Migration covers moving data from the origin device to the cache, or
1083 * vice versa.
1084 *--------------------------------------------------------------*/
1085
1086static void inc_io_migrations(struct cache *cache)
1087{
1088	atomic_inc(&cache->nr_io_migrations);
1089}
1090
1091static void dec_io_migrations(struct cache *cache)
1092{
1093	atomic_dec(&cache->nr_io_migrations);
1094}
1095
1096static bool discard_or_flush(struct bio *bio)
1097{
1098	return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1099}
1100
1101static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1102				     dm_dblock_t *b, dm_dblock_t *e)
1103{
1104	sector_t sb = bio->bi_iter.bi_sector;
1105	sector_t se = bio_end_sector(bio);
1106
1107	*b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1108
1109	if (se - sb < cache->discard_block_size)
1110		*e = *b;
1111	else
1112		*e = to_dblock(block_div(se, cache->discard_block_size));
1113}
1114
1115/*----------------------------------------------------------------*/
1116
1117static void prevent_background_work(struct cache *cache)
1118{
1119	lockdep_off();
1120	down_write(&cache->background_work_lock);
1121	lockdep_on();
1122}
1123
1124static void allow_background_work(struct cache *cache)
1125{
1126	lockdep_off();
1127	up_write(&cache->background_work_lock);
1128	lockdep_on();
1129}
1130
1131static bool background_work_begin(struct cache *cache)
1132{
1133	bool r;
1134
1135	lockdep_off();
1136	r = down_read_trylock(&cache->background_work_lock);
1137	lockdep_on();
1138
1139	return r;
1140}
1141
1142static void background_work_end(struct cache *cache)
1143{
1144	lockdep_off();
1145	up_read(&cache->background_work_lock);
1146	lockdep_on();
1147}
1148
1149/*----------------------------------------------------------------*/
1150
1151static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1152{
1153	return (bio_data_dir(bio) == WRITE) &&
1154		(bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1155}
1156
1157static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1158{
1159	return writeback_mode(cache) &&
1160		(is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1161}
1162
1163static void quiesce(struct dm_cache_migration *mg,
1164		    void (*continuation)(struct work_struct *))
1165{
1166	init_continuation(&mg->k, continuation);
1167	dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1168}
1169
1170static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1171{
1172	struct continuation *k = container_of(ws, struct continuation, ws);
1173	return container_of(k, struct dm_cache_migration, k);
1174}
1175
1176static void copy_complete(int read_err, unsigned long write_err, void *context)
1177{
1178	struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1179
1180	if (read_err || write_err)
1181		mg->k.input = BLK_STS_IOERR;
1182
1183	queue_continuation(mg->cache->wq, &mg->k);
1184}
1185
1186static void copy(struct dm_cache_migration *mg, bool promote)
1187{
1188	struct dm_io_region o_region, c_region;
1189	struct cache *cache = mg->cache;
1190
1191	o_region.bdev = cache->origin_dev->bdev;
1192	o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1193	o_region.count = cache->sectors_per_block;
1194
1195	c_region.bdev = cache->cache_dev->bdev;
1196	c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1197	c_region.count = cache->sectors_per_block;
1198
1199	if (promote)
1200		dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1201	else
1202		dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1203}
1204
1205static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1206{
1207	struct per_bio_data *pb = get_per_bio_data(bio);
1208
1209	if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1210		free_prison_cell(cache, pb->cell);
1211	pb->cell = NULL;
1212}
1213
1214static void overwrite_endio(struct bio *bio)
1215{
1216	struct dm_cache_migration *mg = bio->bi_private;
1217	struct cache *cache = mg->cache;
1218	struct per_bio_data *pb = get_per_bio_data(bio);
1219
1220	dm_unhook_bio(&pb->hook_info, bio);
1221
1222	if (bio->bi_status)
1223		mg->k.input = bio->bi_status;
1224
1225	queue_continuation(cache->wq, &mg->k);
1226}
1227
1228static void overwrite(struct dm_cache_migration *mg,
1229		      void (*continuation)(struct work_struct *))
1230{
1231	struct bio *bio = mg->overwrite_bio;
1232	struct per_bio_data *pb = get_per_bio_data(bio);
1233
1234	dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1235
1236	/*
1237	 * The overwrite bio is part of the copy operation, as such it does
1238	 * not set/clear discard or dirty flags.
1239	 */
1240	if (mg->op->op == POLICY_PROMOTE)
1241		remap_to_cache(mg->cache, bio, mg->op->cblock);
1242	else
1243		remap_to_origin(mg->cache, bio);
1244
1245	init_continuation(&mg->k, continuation);
1246	accounted_request(mg->cache, bio);
1247}
1248
1249/*
1250 * Migration steps:
1251 *
1252 * 1) exclusive lock preventing WRITEs
1253 * 2) quiesce
1254 * 3) copy or issue overwrite bio
1255 * 4) upgrade to exclusive lock preventing READs and WRITEs
1256 * 5) quiesce
1257 * 6) update metadata and commit
1258 * 7) unlock
1259 */
1260static void mg_complete(struct dm_cache_migration *mg, bool success)
1261{
1262	struct bio_list bios;
1263	struct cache *cache = mg->cache;
1264	struct policy_work *op = mg->op;
1265	dm_cblock_t cblock = op->cblock;
1266
1267	if (success)
1268		update_stats(&cache->stats, op->op);
1269
1270	switch (op->op) {
1271	case POLICY_PROMOTE:
1272		clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1273		policy_complete_background_work(cache->policy, op, success);
1274
1275		if (mg->overwrite_bio) {
1276			if (success)
1277				force_set_dirty(cache, cblock);
1278			else if (mg->k.input)
1279				mg->overwrite_bio->bi_status = mg->k.input;
1280			else
1281				mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1282			bio_endio(mg->overwrite_bio);
1283		} else {
1284			if (success)
1285				force_clear_dirty(cache, cblock);
1286			dec_io_migrations(cache);
1287		}
1288		break;
1289
1290	case POLICY_DEMOTE:
1291		/*
1292		 * We clear dirty here to update the nr_dirty counter.
1293		 */
1294		if (success)
1295			force_clear_dirty(cache, cblock);
1296		policy_complete_background_work(cache->policy, op, success);
1297		dec_io_migrations(cache);
1298		break;
1299
1300	case POLICY_WRITEBACK:
1301		if (success)
1302			force_clear_dirty(cache, cblock);
1303		policy_complete_background_work(cache->policy, op, success);
1304		dec_io_migrations(cache);
1305		break;
1306	}
1307
1308	bio_list_init(&bios);
1309	if (mg->cell) {
1310		if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1311			free_prison_cell(cache, mg->cell);
1312	}
1313
1314	free_migration(mg);
1315	defer_bios(cache, &bios);
1316	wake_migration_worker(cache);
1317
1318	background_work_end(cache);
1319}
1320
1321static void mg_success(struct work_struct *ws)
1322{
1323	struct dm_cache_migration *mg = ws_to_mg(ws);
1324	mg_complete(mg, mg->k.input == 0);
1325}
1326
1327static void mg_update_metadata(struct work_struct *ws)
1328{
1329	int r;
1330	struct dm_cache_migration *mg = ws_to_mg(ws);
1331	struct cache *cache = mg->cache;
1332	struct policy_work *op = mg->op;
1333
1334	switch (op->op) {
1335	case POLICY_PROMOTE:
1336		r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1337		if (r) {
1338			DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1339				    cache_device_name(cache));
1340			metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1341
1342			mg_complete(mg, false);
1343			return;
1344		}
1345		mg_complete(mg, true);
1346		break;
1347
1348	case POLICY_DEMOTE:
1349		r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1350		if (r) {
1351			DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1352				    cache_device_name(cache));
1353			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1354
1355			mg_complete(mg, false);
1356			return;
1357		}
1358
1359		/*
1360		 * It would be nice if we only had to commit when a REQ_FLUSH
1361		 * comes through.  But there's one scenario that we have to
1362		 * look out for:
1363		 *
1364		 * - vblock x in a cache block
1365		 * - domotion occurs
1366		 * - cache block gets reallocated and over written
1367		 * - crash
1368		 *
1369		 * When we recover, because there was no commit the cache will
1370		 * rollback to having the data for vblock x in the cache block.
1371		 * But the cache block has since been overwritten, so it'll end
1372		 * up pointing to data that was never in 'x' during the history
1373		 * of the device.
1374		 *
1375		 * To avoid this issue we require a commit as part of the
1376		 * demotion operation.
1377		 */
1378		init_continuation(&mg->k, mg_success);
1379		continue_after_commit(&cache->committer, &mg->k);
1380		schedule_commit(&cache->committer);
1381		break;
1382
1383	case POLICY_WRITEBACK:
1384		mg_complete(mg, true);
1385		break;
1386	}
1387}
1388
1389static void mg_update_metadata_after_copy(struct work_struct *ws)
1390{
1391	struct dm_cache_migration *mg = ws_to_mg(ws);
1392
1393	/*
1394	 * Did the copy succeed?
1395	 */
1396	if (mg->k.input)
1397		mg_complete(mg, false);
1398	else
1399		mg_update_metadata(ws);
1400}
1401
1402static void mg_upgrade_lock(struct work_struct *ws)
1403{
1404	int r;
1405	struct dm_cache_migration *mg = ws_to_mg(ws);
1406
1407	/*
1408	 * Did the copy succeed?
1409	 */
1410	if (mg->k.input)
1411		mg_complete(mg, false);
1412
1413	else {
1414		/*
1415		 * Now we want the lock to prevent both reads and writes.
1416		 */
1417		r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1418					    READ_WRITE_LOCK_LEVEL);
1419		if (r < 0)
1420			mg_complete(mg, false);
1421
1422		else if (r)
1423			quiesce(mg, mg_update_metadata);
1424
1425		else
1426			mg_update_metadata(ws);
1427	}
1428}
1429
1430static void mg_full_copy(struct work_struct *ws)
1431{
1432	struct dm_cache_migration *mg = ws_to_mg(ws);
1433	struct cache *cache = mg->cache;
1434	struct policy_work *op = mg->op;
1435	bool is_policy_promote = (op->op == POLICY_PROMOTE);
1436
1437	if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1438	    is_discarded_oblock(cache, op->oblock)) {
1439		mg_upgrade_lock(ws);
1440		return;
1441	}
1442
1443	init_continuation(&mg->k, mg_upgrade_lock);
1444	copy(mg, is_policy_promote);
1445}
1446
1447static void mg_copy(struct work_struct *ws)
1448{
1449	struct dm_cache_migration *mg = ws_to_mg(ws);
1450
1451	if (mg->overwrite_bio) {
1452		/*
1453		 * No exclusive lock was held when we last checked if the bio
1454		 * was optimisable.  So we have to check again in case things
1455		 * have changed (eg, the block may no longer be discarded).
1456		 */
1457		if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1458			/*
1459			 * Fallback to a real full copy after doing some tidying up.
1460			 */
1461			bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1462			BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1463			mg->overwrite_bio = NULL;
1464			inc_io_migrations(mg->cache);
1465			mg_full_copy(ws);
1466			return;
1467		}
1468
1469		/*
1470		 * It's safe to do this here, even though it's new data
1471		 * because all IO has been locked out of the block.
1472		 *
1473		 * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1474		 * so _not_ using mg_upgrade_lock() as continutation.
1475		 */
1476		overwrite(mg, mg_update_metadata_after_copy);
1477
1478	} else
1479		mg_full_copy(ws);
1480}
1481
1482static int mg_lock_writes(struct dm_cache_migration *mg)
1483{
1484	int r;
1485	struct dm_cell_key_v2 key;
1486	struct cache *cache = mg->cache;
1487	struct dm_bio_prison_cell_v2 *prealloc;
1488
1489	prealloc = alloc_prison_cell(cache);
1490
1491	/*
1492	 * Prevent writes to the block, but allow reads to continue.
1493	 * Unless we're using an overwrite bio, in which case we lock
1494	 * everything.
1495	 */
1496	build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1497	r = dm_cell_lock_v2(cache->prison, &key,
1498			    mg->overwrite_bio ?  READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1499			    prealloc, &mg->cell);
1500	if (r < 0) {
1501		free_prison_cell(cache, prealloc);
1502		mg_complete(mg, false);
1503		return r;
1504	}
1505
1506	if (mg->cell != prealloc)
1507		free_prison_cell(cache, prealloc);
1508
1509	if (r == 0)
1510		mg_copy(&mg->k.ws);
1511	else
1512		quiesce(mg, mg_copy);
1513
1514	return 0;
1515}
1516
1517static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1518{
1519	struct dm_cache_migration *mg;
1520
1521	if (!background_work_begin(cache)) {
1522		policy_complete_background_work(cache->policy, op, false);
1523		return -EPERM;
1524	}
1525
1526	mg = alloc_migration(cache);
1527
1528	mg->op = op;
1529	mg->overwrite_bio = bio;
1530
1531	if (!bio)
1532		inc_io_migrations(cache);
1533
1534	return mg_lock_writes(mg);
1535}
1536
1537/*----------------------------------------------------------------
1538 * invalidation processing
1539 *--------------------------------------------------------------*/
1540
1541static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1542{
1543	struct bio_list bios;
1544	struct cache *cache = mg->cache;
1545
1546	bio_list_init(&bios);
1547	if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1548		free_prison_cell(cache, mg->cell);
1549
1550	if (!success && mg->overwrite_bio)
1551		bio_io_error(mg->overwrite_bio);
1552
1553	free_migration(mg);
1554	defer_bios(cache, &bios);
1555
1556	background_work_end(cache);
1557}
1558
1559static void invalidate_completed(struct work_struct *ws)
1560{
1561	struct dm_cache_migration *mg = ws_to_mg(ws);
1562	invalidate_complete(mg, !mg->k.input);
1563}
1564
1565static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1566{
1567	int r = policy_invalidate_mapping(cache->policy, cblock);
1568	if (!r) {
1569		r = dm_cache_remove_mapping(cache->cmd, cblock);
1570		if (r) {
1571			DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1572				    cache_device_name(cache));
1573			metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1574		}
1575
1576	} else if (r == -ENODATA) {
1577		/*
1578		 * Harmless, already unmapped.
1579		 */
1580		r = 0;
1581
1582	} else
1583		DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1584
1585	return r;
1586}
1587
1588static void invalidate_remove(struct work_struct *ws)
1589{
1590	int r;
1591	struct dm_cache_migration *mg = ws_to_mg(ws);
1592	struct cache *cache = mg->cache;
1593
1594	r = invalidate_cblock(cache, mg->invalidate_cblock);
1595	if (r) {
1596		invalidate_complete(mg, false);
1597		return;
1598	}
1599
1600	init_continuation(&mg->k, invalidate_completed);
1601	continue_after_commit(&cache->committer, &mg->k);
1602	remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1603	mg->overwrite_bio = NULL;
1604	schedule_commit(&cache->committer);
1605}
1606
1607static int invalidate_lock(struct dm_cache_migration *mg)
1608{
1609	int r;
1610	struct dm_cell_key_v2 key;
1611	struct cache *cache = mg->cache;
1612	struct dm_bio_prison_cell_v2 *prealloc;
1613
1614	prealloc = alloc_prison_cell(cache);
1615
1616	build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1617	r = dm_cell_lock_v2(cache->prison, &key,
1618			    READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1619	if (r < 0) {
1620		free_prison_cell(cache, prealloc);
1621		invalidate_complete(mg, false);
1622		return r;
1623	}
1624
1625	if (mg->cell != prealloc)
1626		free_prison_cell(cache, prealloc);
1627
1628	if (r)
1629		quiesce(mg, invalidate_remove);
1630
1631	else {
1632		/*
1633		 * We can't call invalidate_remove() directly here because we
1634		 * might still be in request context.
1635		 */
1636		init_continuation(&mg->k, invalidate_remove);
1637		queue_work(cache->wq, &mg->k.ws);
1638	}
1639
1640	return 0;
1641}
1642
1643static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1644			    dm_oblock_t oblock, struct bio *bio)
1645{
1646	struct dm_cache_migration *mg;
1647
1648	if (!background_work_begin(cache))
1649		return -EPERM;
1650
1651	mg = alloc_migration(cache);
1652
1653	mg->overwrite_bio = bio;
1654	mg->invalidate_cblock = cblock;
1655	mg->invalidate_oblock = oblock;
1656
1657	return invalidate_lock(mg);
1658}
1659
1660/*----------------------------------------------------------------
1661 * bio processing
1662 *--------------------------------------------------------------*/
1663
1664enum busy {
1665	IDLE,
1666	BUSY
1667};
1668
1669static enum busy spare_migration_bandwidth(struct cache *cache)
1670{
1671	bool idle = iot_idle_for(&cache->tracker, HZ);
1672	sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1673		cache->sectors_per_block;
1674
1675	if (idle && current_volume <= cache->migration_threshold)
1676		return IDLE;
1677	else
1678		return BUSY;
1679}
1680
1681static void inc_hit_counter(struct cache *cache, struct bio *bio)
1682{
1683	atomic_inc(bio_data_dir(bio) == READ ?
1684		   &cache->stats.read_hit : &cache->stats.write_hit);
1685}
1686
1687static void inc_miss_counter(struct cache *cache, struct bio *bio)
1688{
1689	atomic_inc(bio_data_dir(bio) == READ ?
1690		   &cache->stats.read_miss : &cache->stats.write_miss);
1691}
1692
1693/*----------------------------------------------------------------*/
1694
1695static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1696		   bool *commit_needed)
1697{
1698	int r, data_dir;
1699	bool rb, background_queued;
1700	dm_cblock_t cblock;
1701
1702	*commit_needed = false;
1703
1704	rb = bio_detain_shared(cache, block, bio);
1705	if (!rb) {
1706		/*
1707		 * An exclusive lock is held for this block, so we have to
1708		 * wait.  We set the commit_needed flag so the current
1709		 * transaction will be committed asap, allowing this lock
1710		 * to be dropped.
1711		 */
1712		*commit_needed = true;
1713		return DM_MAPIO_SUBMITTED;
1714	}
1715
1716	data_dir = bio_data_dir(bio);
1717
1718	if (optimisable_bio(cache, bio, block)) {
1719		struct policy_work *op = NULL;
1720
1721		r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1722		if (unlikely(r && r != -ENOENT)) {
1723			DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1724				    cache_device_name(cache), r);
1725			bio_io_error(bio);
1726			return DM_MAPIO_SUBMITTED;
1727		}
1728
1729		if (r == -ENOENT && op) {
1730			bio_drop_shared_lock(cache, bio);
1731			BUG_ON(op->op != POLICY_PROMOTE);
1732			mg_start(cache, op, bio);
1733			return DM_MAPIO_SUBMITTED;
1734		}
1735	} else {
1736		r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1737		if (unlikely(r && r != -ENOENT)) {
1738			DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1739				    cache_device_name(cache), r);
1740			bio_io_error(bio);
1741			return DM_MAPIO_SUBMITTED;
1742		}
1743
1744		if (background_queued)
1745			wake_migration_worker(cache);
1746	}
1747
1748	if (r == -ENOENT) {
1749		struct per_bio_data *pb = get_per_bio_data(bio);
1750
1751		/*
1752		 * Miss.
1753		 */
1754		inc_miss_counter(cache, bio);
1755		if (pb->req_nr == 0) {
1756			accounted_begin(cache, bio);
1757			remap_to_origin_clear_discard(cache, bio, block);
1758		} else {
1759			/*
1760			 * This is a duplicate writethrough io that is no
1761			 * longer needed because the block has been demoted.
1762			 */
1763			bio_endio(bio);
1764			return DM_MAPIO_SUBMITTED;
1765		}
1766	} else {
1767		/*
1768		 * Hit.
1769		 */
1770		inc_hit_counter(cache, bio);
1771
1772		/*
1773		 * Passthrough always maps to the origin, invalidating any
1774		 * cache blocks that are written to.
1775		 */
1776		if (passthrough_mode(cache)) {
1777			if (bio_data_dir(bio) == WRITE) {
1778				bio_drop_shared_lock(cache, bio);
1779				atomic_inc(&cache->stats.demotion);
1780				invalidate_start(cache, cblock, block, bio);
1781			} else
1782				remap_to_origin_clear_discard(cache, bio, block);
1783		} else {
1784			if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1785			    !is_dirty(cache, cblock)) {
1786				remap_to_origin_and_cache(cache, bio, block, cblock);
1787				accounted_begin(cache, bio);
1788			} else
1789				remap_to_cache_dirty(cache, bio, block, cblock);
1790		}
1791	}
1792
1793	/*
1794	 * dm core turns FUA requests into a separate payload and FLUSH req.
1795	 */
1796	if (bio->bi_opf & REQ_FUA) {
1797		/*
1798		 * issue_after_commit will call accounted_begin a second time.  So
1799		 * we call accounted_complete() to avoid double accounting.
1800		 */
1801		accounted_complete(cache, bio);
1802		issue_after_commit(&cache->committer, bio);
1803		*commit_needed = true;
1804		return DM_MAPIO_SUBMITTED;
1805	}
1806
1807	return DM_MAPIO_REMAPPED;
1808}
1809
1810static bool process_bio(struct cache *cache, struct bio *bio)
1811{
1812	bool commit_needed;
1813
1814	if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1815		generic_make_request(bio);
1816
1817	return commit_needed;
1818}
1819
1820/*
1821 * A non-zero return indicates read_only or fail_io mode.
1822 */
1823static int commit(struct cache *cache, bool clean_shutdown)
1824{
1825	int r;
1826
1827	if (get_cache_mode(cache) >= CM_READ_ONLY)
1828		return -EINVAL;
1829
1830	atomic_inc(&cache->stats.commit_count);
1831	r = dm_cache_commit(cache->cmd, clean_shutdown);
1832	if (r)
1833		metadata_operation_failed(cache, "dm_cache_commit", r);
1834
1835	return r;
1836}
1837
1838/*
1839 * Used by the batcher.
1840 */
1841static blk_status_t commit_op(void *context)
1842{
1843	struct cache *cache = context;
1844
1845	if (dm_cache_changed_this_transaction(cache->cmd))
1846		return errno_to_blk_status(commit(cache, false));
1847
1848	return 0;
1849}
1850
1851/*----------------------------------------------------------------*/
1852
1853static bool process_flush_bio(struct cache *cache, struct bio *bio)
1854{
1855	struct per_bio_data *pb = get_per_bio_data(bio);
1856
1857	if (!pb->req_nr)
1858		remap_to_origin(cache, bio);
1859	else
1860		remap_to_cache(cache, bio, 0);
1861
1862	issue_after_commit(&cache->committer, bio);
1863	return true;
1864}
1865
1866static bool process_discard_bio(struct cache *cache, struct bio *bio)
1867{
1868	dm_dblock_t b, e;
1869
1870	// FIXME: do we need to lock the region?  Or can we just assume the
1871	// user wont be so foolish as to issue discard concurrently with
1872	// other IO?
1873	calc_discard_block_range(cache, bio, &b, &e);
1874	while (b != e) {
1875		set_discard(cache, b);
1876		b = to_dblock(from_dblock(b) + 1);
1877	}
1878
1879	if (cache->features.discard_passdown) {
1880		remap_to_origin(cache, bio);
1881		generic_make_request(bio);
1882	} else
1883		bio_endio(bio);
1884
1885	return false;
1886}
1887
1888static void process_deferred_bios(struct work_struct *ws)
1889{
1890	struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1891
1892	unsigned long flags;
1893	bool commit_needed = false;
1894	struct bio_list bios;
1895	struct bio *bio;
1896
1897	bio_list_init(&bios);
1898
1899	spin_lock_irqsave(&cache->lock, flags);
1900	bio_list_merge(&bios, &cache->deferred_bios);
1901	bio_list_init(&cache->deferred_bios);
1902	spin_unlock_irqrestore(&cache->lock, flags);
1903
1904	while ((bio = bio_list_pop(&bios))) {
1905		if (bio->bi_opf & REQ_PREFLUSH)
1906			commit_needed = process_flush_bio(cache, bio) || commit_needed;
1907
1908		else if (bio_op(bio) == REQ_OP_DISCARD)
1909			commit_needed = process_discard_bio(cache, bio) || commit_needed;
1910
1911		else
1912			commit_needed = process_bio(cache, bio) || commit_needed;
1913	}
1914
1915	if (commit_needed)
1916		schedule_commit(&cache->committer);
1917}
1918
1919/*----------------------------------------------------------------
1920 * Main worker loop
1921 *--------------------------------------------------------------*/
1922
1923static void requeue_deferred_bios(struct cache *cache)
1924{
1925	struct bio *bio;
1926	struct bio_list bios;
1927
1928	bio_list_init(&bios);
1929	bio_list_merge(&bios, &cache->deferred_bios);
1930	bio_list_init(&cache->deferred_bios);
1931
1932	while ((bio = bio_list_pop(&bios))) {
1933		bio->bi_status = BLK_STS_DM_REQUEUE;
1934		bio_endio(bio);
1935	}
1936}
1937
1938/*
1939 * We want to commit periodically so that not too much
1940 * unwritten metadata builds up.
1941 */
1942static void do_waker(struct work_struct *ws)
1943{
1944	struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1945
1946	policy_tick(cache->policy, true);
1947	wake_migration_worker(cache);
1948	schedule_commit(&cache->committer);
1949	queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1950}
1951
1952static void check_migrations(struct work_struct *ws)
1953{
1954	int r;
1955	struct policy_work *op;
1956	struct cache *cache = container_of(ws, struct cache, migration_worker);
1957	enum busy b;
1958
1959	for (;;) {
1960		b = spare_migration_bandwidth(cache);
1961
1962		r = policy_get_background_work(cache->policy, b == IDLE, &op);
1963		if (r == -ENODATA)
1964			break;
1965
1966		if (r) {
1967			DMERR_LIMIT("%s: policy_background_work failed",
1968				    cache_device_name(cache));
1969			break;
1970		}
1971
1972		r = mg_start(cache, op, NULL);
1973		if (r)
1974			break;
1975	}
1976}
1977
1978/*----------------------------------------------------------------
1979 * Target methods
1980 *--------------------------------------------------------------*/
1981
1982/*
1983 * This function gets called on the error paths of the constructor, so we
1984 * have to cope with a partially initialised struct.
1985 */
1986static void destroy(struct cache *cache)
1987{
1988	unsigned i;
1989
1990	mempool_exit(&cache->migration_pool);
1991
1992	if (cache->prison)
1993		dm_bio_prison_destroy_v2(cache->prison);
1994
 
1995	if (cache->wq)
1996		destroy_workqueue(cache->wq);
1997
1998	if (cache->dirty_bitset)
1999		free_bitset(cache->dirty_bitset);
2000
2001	if (cache->discard_bitset)
2002		free_bitset(cache->discard_bitset);
2003
2004	if (cache->copier)
2005		dm_kcopyd_client_destroy(cache->copier);
2006
2007	if (cache->cmd)
2008		dm_cache_metadata_close(cache->cmd);
2009
2010	if (cache->metadata_dev)
2011		dm_put_device(cache->ti, cache->metadata_dev);
2012
2013	if (cache->origin_dev)
2014		dm_put_device(cache->ti, cache->origin_dev);
2015
2016	if (cache->cache_dev)
2017		dm_put_device(cache->ti, cache->cache_dev);
2018
2019	if (cache->policy)
2020		dm_cache_policy_destroy(cache->policy);
2021
2022	for (i = 0; i < cache->nr_ctr_args ; i++)
2023		kfree(cache->ctr_args[i]);
2024	kfree(cache->ctr_args);
2025
2026	bioset_exit(&cache->bs);
2027
2028	kfree(cache);
2029}
2030
2031static void cache_dtr(struct dm_target *ti)
2032{
2033	struct cache *cache = ti->private;
2034
2035	destroy(cache);
2036}
2037
2038static sector_t get_dev_size(struct dm_dev *dev)
2039{
2040	return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2041}
2042
2043/*----------------------------------------------------------------*/
2044
2045/*
2046 * Construct a cache device mapping.
2047 *
2048 * cache <metadata dev> <cache dev> <origin dev> <block size>
2049 *       <#feature args> [<feature arg>]*
2050 *       <policy> <#policy args> [<policy arg>]*
2051 *
2052 * metadata dev    : fast device holding the persistent metadata
2053 * cache dev	   : fast device holding cached data blocks
2054 * origin dev	   : slow device holding original data blocks
2055 * block size	   : cache unit size in sectors
2056 *
2057 * #feature args   : number of feature arguments passed
2058 * feature args    : writethrough.  (The default is writeback.)
2059 *
2060 * policy	   : the replacement policy to use
2061 * #policy args    : an even number of policy arguments corresponding
2062 *		     to key/value pairs passed to the policy
2063 * policy args	   : key/value pairs passed to the policy
2064 *		     E.g. 'sequential_threshold 1024'
2065 *		     See cache-policies.txt for details.
2066 *
2067 * Optional feature arguments are:
2068 *   writethrough  : write through caching that prohibits cache block
2069 *		     content from being different from origin block content.
2070 *		     Without this argument, the default behaviour is to write
2071 *		     back cache block contents later for performance reasons,
2072 *		     so they may differ from the corresponding origin blocks.
2073 */
2074struct cache_args {
2075	struct dm_target *ti;
2076
2077	struct dm_dev *metadata_dev;
2078
2079	struct dm_dev *cache_dev;
2080	sector_t cache_sectors;
2081
2082	struct dm_dev *origin_dev;
2083	sector_t origin_sectors;
2084
2085	uint32_t block_size;
2086
2087	const char *policy_name;
2088	int policy_argc;
2089	const char **policy_argv;
2090
2091	struct cache_features features;
2092};
2093
2094static void destroy_cache_args(struct cache_args *ca)
2095{
2096	if (ca->metadata_dev)
2097		dm_put_device(ca->ti, ca->metadata_dev);
2098
2099	if (ca->cache_dev)
2100		dm_put_device(ca->ti, ca->cache_dev);
2101
2102	if (ca->origin_dev)
2103		dm_put_device(ca->ti, ca->origin_dev);
2104
2105	kfree(ca);
2106}
2107
2108static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2109{
2110	if (!as->argc) {
2111		*error = "Insufficient args";
2112		return false;
2113	}
2114
2115	return true;
2116}
2117
2118static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2119			      char **error)
2120{
2121	int r;
2122	sector_t metadata_dev_size;
2123	char b[BDEVNAME_SIZE];
2124
2125	if (!at_least_one_arg(as, error))
2126		return -EINVAL;
2127
2128	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2129			  &ca->metadata_dev);
2130	if (r) {
2131		*error = "Error opening metadata device";
2132		return r;
2133	}
2134
2135	metadata_dev_size = get_dev_size(ca->metadata_dev);
2136	if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2137		DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2138		       bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2139
2140	return 0;
2141}
2142
2143static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2144			   char **error)
2145{
2146	int r;
2147
2148	if (!at_least_one_arg(as, error))
2149		return -EINVAL;
2150
2151	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2152			  &ca->cache_dev);
2153	if (r) {
2154		*error = "Error opening cache device";
2155		return r;
2156	}
2157	ca->cache_sectors = get_dev_size(ca->cache_dev);
2158
2159	return 0;
2160}
2161
2162static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2163			    char **error)
2164{
2165	int r;
2166
2167	if (!at_least_one_arg(as, error))
2168		return -EINVAL;
2169
2170	r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2171			  &ca->origin_dev);
2172	if (r) {
2173		*error = "Error opening origin device";
2174		return r;
2175	}
2176
2177	ca->origin_sectors = get_dev_size(ca->origin_dev);
2178	if (ca->ti->len > ca->origin_sectors) {
2179		*error = "Device size larger than cached device";
2180		return -EINVAL;
2181	}
2182
2183	return 0;
2184}
2185
2186static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2187			    char **error)
2188{
2189	unsigned long block_size;
2190
2191	if (!at_least_one_arg(as, error))
2192		return -EINVAL;
2193
2194	if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2195	    block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2196	    block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2197	    block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2198		*error = "Invalid data block size";
2199		return -EINVAL;
2200	}
2201
2202	if (block_size > ca->cache_sectors) {
2203		*error = "Data block size is larger than the cache device";
2204		return -EINVAL;
2205	}
2206
2207	ca->block_size = block_size;
2208
2209	return 0;
2210}
2211
2212static void init_features(struct cache_features *cf)
2213{
2214	cf->mode = CM_WRITE;
2215	cf->io_mode = CM_IO_WRITEBACK;
2216	cf->metadata_version = 1;
2217	cf->discard_passdown = true;
2218}
2219
2220static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2221			  char **error)
2222{
2223	static const struct dm_arg _args[] = {
2224		{0, 3, "Invalid number of cache feature arguments"},
2225	};
2226
2227	int r, mode_ctr = 0;
2228	unsigned argc;
2229	const char *arg;
2230	struct cache_features *cf = &ca->features;
2231
2232	init_features(cf);
2233
2234	r = dm_read_arg_group(_args, as, &argc, error);
2235	if (r)
2236		return -EINVAL;
2237
2238	while (argc--) {
2239		arg = dm_shift_arg(as);
2240
2241		if (!strcasecmp(arg, "writeback")) {
2242			cf->io_mode = CM_IO_WRITEBACK;
2243			mode_ctr++;
2244		}
2245
2246		else if (!strcasecmp(arg, "writethrough")) {
2247			cf->io_mode = CM_IO_WRITETHROUGH;
2248			mode_ctr++;
2249		}
2250
2251		else if (!strcasecmp(arg, "passthrough")) {
2252			cf->io_mode = CM_IO_PASSTHROUGH;
2253			mode_ctr++;
2254		}
2255
2256		else if (!strcasecmp(arg, "metadata2"))
2257			cf->metadata_version = 2;
2258
2259		else if (!strcasecmp(arg, "no_discard_passdown"))
2260			cf->discard_passdown = false;
2261
2262		else {
2263			*error = "Unrecognised cache feature requested";
2264			return -EINVAL;
2265		}
2266	}
2267
2268	if (mode_ctr > 1) {
2269		*error = "Duplicate cache io_mode features requested";
2270		return -EINVAL;
2271	}
2272
2273	return 0;
2274}
2275
2276static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2277			char **error)
2278{
2279	static const struct dm_arg _args[] = {
2280		{0, 1024, "Invalid number of policy arguments"},
2281	};
2282
2283	int r;
2284
2285	if (!at_least_one_arg(as, error))
2286		return -EINVAL;
2287
2288	ca->policy_name = dm_shift_arg(as);
2289
2290	r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2291	if (r)
2292		return -EINVAL;
2293
2294	ca->policy_argv = (const char **)as->argv;
2295	dm_consume_args(as, ca->policy_argc);
2296
2297	return 0;
2298}
2299
2300static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2301			    char **error)
2302{
2303	int r;
2304	struct dm_arg_set as;
2305
2306	as.argc = argc;
2307	as.argv = argv;
2308
2309	r = parse_metadata_dev(ca, &as, error);
2310	if (r)
2311		return r;
2312
2313	r = parse_cache_dev(ca, &as, error);
2314	if (r)
2315		return r;
2316
2317	r = parse_origin_dev(ca, &as, error);
2318	if (r)
2319		return r;
2320
2321	r = parse_block_size(ca, &as, error);
2322	if (r)
2323		return r;
2324
2325	r = parse_features(ca, &as, error);
2326	if (r)
2327		return r;
2328
2329	r = parse_policy(ca, &as, error);
2330	if (r)
2331		return r;
2332
2333	return 0;
2334}
2335
2336/*----------------------------------------------------------------*/
2337
2338static struct kmem_cache *migration_cache;
2339
2340#define NOT_CORE_OPTION 1
2341
2342static int process_config_option(struct cache *cache, const char *key, const char *value)
2343{
2344	unsigned long tmp;
2345
2346	if (!strcasecmp(key, "migration_threshold")) {
2347		if (kstrtoul(value, 10, &tmp))
2348			return -EINVAL;
2349
2350		cache->migration_threshold = tmp;
2351		return 0;
2352	}
2353
2354	return NOT_CORE_OPTION;
2355}
2356
2357static int set_config_value(struct cache *cache, const char *key, const char *value)
2358{
2359	int r = process_config_option(cache, key, value);
2360
2361	if (r == NOT_CORE_OPTION)
2362		r = policy_set_config_value(cache->policy, key, value);
2363
2364	if (r)
2365		DMWARN("bad config value for %s: %s", key, value);
2366
2367	return r;
2368}
2369
2370static int set_config_values(struct cache *cache, int argc, const char **argv)
2371{
2372	int r = 0;
2373
2374	if (argc & 1) {
2375		DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2376		return -EINVAL;
2377	}
2378
2379	while (argc) {
2380		r = set_config_value(cache, argv[0], argv[1]);
2381		if (r)
2382			break;
2383
2384		argc -= 2;
2385		argv += 2;
2386	}
2387
2388	return r;
2389}
2390
2391static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2392			       char **error)
2393{
2394	struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2395							   cache->cache_size,
2396							   cache->origin_sectors,
2397							   cache->sectors_per_block);
2398	if (IS_ERR(p)) {
2399		*error = "Error creating cache's policy";
2400		return PTR_ERR(p);
2401	}
2402	cache->policy = p;
2403	BUG_ON(!cache->policy);
2404
2405	return 0;
2406}
2407
2408/*
2409 * We want the discard block size to be at least the size of the cache
2410 * block size and have no more than 2^14 discard blocks across the origin.
2411 */
2412#define MAX_DISCARD_BLOCKS (1 << 14)
2413
2414static bool too_many_discard_blocks(sector_t discard_block_size,
2415				    sector_t origin_size)
2416{
2417	(void) sector_div(origin_size, discard_block_size);
2418
2419	return origin_size > MAX_DISCARD_BLOCKS;
2420}
2421
2422static sector_t calculate_discard_block_size(sector_t cache_block_size,
2423					     sector_t origin_size)
2424{
2425	sector_t discard_block_size = cache_block_size;
2426
2427	if (origin_size)
2428		while (too_many_discard_blocks(discard_block_size, origin_size))
2429			discard_block_size *= 2;
2430
2431	return discard_block_size;
2432}
2433
2434static void set_cache_size(struct cache *cache, dm_cblock_t size)
2435{
2436	dm_block_t nr_blocks = from_cblock(size);
2437
2438	if (nr_blocks > (1 << 20) && cache->cache_size != size)
2439		DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2440			     "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2441			     "Please consider increasing the cache block size to reduce the overall cache block count.",
2442			     (unsigned long long) nr_blocks);
2443
2444	cache->cache_size = size;
2445}
2446
2447static int is_congested(struct dm_dev *dev, int bdi_bits)
2448{
2449	struct request_queue *q = bdev_get_queue(dev->bdev);
2450	return bdi_congested(q->backing_dev_info, bdi_bits);
2451}
2452
2453static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2454{
2455	struct cache *cache = container_of(cb, struct cache, callbacks);
2456
2457	return is_congested(cache->origin_dev, bdi_bits) ||
2458		is_congested(cache->cache_dev, bdi_bits);
2459}
2460
2461#define DEFAULT_MIGRATION_THRESHOLD 2048
2462
2463static int cache_create(struct cache_args *ca, struct cache **result)
2464{
2465	int r = 0;
2466	char **error = &ca->ti->error;
2467	struct cache *cache;
2468	struct dm_target *ti = ca->ti;
2469	dm_block_t origin_blocks;
2470	struct dm_cache_metadata *cmd;
2471	bool may_format = ca->features.mode == CM_WRITE;
2472
2473	cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2474	if (!cache)
2475		return -ENOMEM;
2476
2477	cache->ti = ca->ti;
2478	ti->private = cache;
 
2479	ti->num_flush_bios = 2;
2480	ti->flush_supported = true;
2481
2482	ti->num_discard_bios = 1;
2483	ti->discards_supported = true;
2484
2485	ti->per_io_data_size = sizeof(struct per_bio_data);
2486
2487	cache->features = ca->features;
2488	if (writethrough_mode(cache)) {
2489		/* Create bioset for writethrough bios issued to origin */
2490		r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2491		if (r)
2492			goto bad;
2493	}
2494
2495	cache->callbacks.congested_fn = cache_is_congested;
2496	dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2497
2498	cache->metadata_dev = ca->metadata_dev;
2499	cache->origin_dev = ca->origin_dev;
2500	cache->cache_dev = ca->cache_dev;
2501
2502	ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2503
2504	origin_blocks = cache->origin_sectors = ca->origin_sectors;
2505	origin_blocks = block_div(origin_blocks, ca->block_size);
2506	cache->origin_blocks = to_oblock(origin_blocks);
2507
2508	cache->sectors_per_block = ca->block_size;
2509	if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2510		r = -EINVAL;
2511		goto bad;
2512	}
2513
2514	if (ca->block_size & (ca->block_size - 1)) {
2515		dm_block_t cache_size = ca->cache_sectors;
2516
2517		cache->sectors_per_block_shift = -1;
2518		cache_size = block_div(cache_size, ca->block_size);
2519		set_cache_size(cache, to_cblock(cache_size));
2520	} else {
2521		cache->sectors_per_block_shift = __ffs(ca->block_size);
2522		set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2523	}
2524
2525	r = create_cache_policy(cache, ca, error);
2526	if (r)
2527		goto bad;
2528
2529	cache->policy_nr_args = ca->policy_argc;
2530	cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2531
2532	r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2533	if (r) {
2534		*error = "Error setting cache policy's config values";
2535		goto bad;
2536	}
2537
2538	cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2539				     ca->block_size, may_format,
2540				     dm_cache_policy_get_hint_size(cache->policy),
2541				     ca->features.metadata_version);
2542	if (IS_ERR(cmd)) {
2543		*error = "Error creating metadata object";
2544		r = PTR_ERR(cmd);
2545		goto bad;
2546	}
2547	cache->cmd = cmd;
2548	set_cache_mode(cache, CM_WRITE);
2549	if (get_cache_mode(cache) != CM_WRITE) {
2550		*error = "Unable to get write access to metadata, please check/repair metadata.";
2551		r = -EINVAL;
2552		goto bad;
2553	}
2554
2555	if (passthrough_mode(cache)) {
2556		bool all_clean;
2557
2558		r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2559		if (r) {
2560			*error = "dm_cache_metadata_all_clean() failed";
2561			goto bad;
2562		}
2563
2564		if (!all_clean) {
2565			*error = "Cannot enter passthrough mode unless all blocks are clean";
2566			r = -EINVAL;
2567			goto bad;
2568		}
2569
2570		policy_allow_migrations(cache->policy, false);
2571	}
2572
2573	spin_lock_init(&cache->lock);
2574	bio_list_init(&cache->deferred_bios);
2575	atomic_set(&cache->nr_allocated_migrations, 0);
2576	atomic_set(&cache->nr_io_migrations, 0);
2577	init_waitqueue_head(&cache->migration_wait);
2578
2579	r = -ENOMEM;
2580	atomic_set(&cache->nr_dirty, 0);
2581	cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2582	if (!cache->dirty_bitset) {
2583		*error = "could not allocate dirty bitset";
2584		goto bad;
2585	}
2586	clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2587
2588	cache->discard_block_size =
2589		calculate_discard_block_size(cache->sectors_per_block,
2590					     cache->origin_sectors);
2591	cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2592							      cache->discard_block_size));
2593	cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2594	if (!cache->discard_bitset) {
2595		*error = "could not allocate discard bitset";
2596		goto bad;
2597	}
2598	clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2599
2600	cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2601	if (IS_ERR(cache->copier)) {
2602		*error = "could not create kcopyd client";
2603		r = PTR_ERR(cache->copier);
2604		goto bad;
2605	}
2606
2607	cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2608	if (!cache->wq) {
2609		*error = "could not create workqueue for metadata object";
2610		goto bad;
2611	}
2612	INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2613	INIT_WORK(&cache->migration_worker, check_migrations);
2614	INIT_DELAYED_WORK(&cache->waker, do_waker);
2615
2616	cache->prison = dm_bio_prison_create_v2(cache->wq);
2617	if (!cache->prison) {
2618		*error = "could not create bio prison";
2619		goto bad;
2620	}
2621
2622	r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2623				   migration_cache);
2624	if (r) {
2625		*error = "Error creating cache's migration mempool";
2626		goto bad;
2627	}
2628
2629	cache->need_tick_bio = true;
2630	cache->sized = false;
2631	cache->invalidate = false;
2632	cache->commit_requested = false;
2633	cache->loaded_mappings = false;
2634	cache->loaded_discards = false;
2635
2636	load_stats(cache);
2637
2638	atomic_set(&cache->stats.demotion, 0);
2639	atomic_set(&cache->stats.promotion, 0);
2640	atomic_set(&cache->stats.copies_avoided, 0);
2641	atomic_set(&cache->stats.cache_cell_clash, 0);
2642	atomic_set(&cache->stats.commit_count, 0);
2643	atomic_set(&cache->stats.discard_count, 0);
2644
2645	spin_lock_init(&cache->invalidation_lock);
2646	INIT_LIST_HEAD(&cache->invalidation_requests);
2647
2648	batcher_init(&cache->committer, commit_op, cache,
2649		     issue_op, cache, cache->wq);
2650	iot_init(&cache->tracker);
2651
2652	init_rwsem(&cache->background_work_lock);
2653	prevent_background_work(cache);
2654
2655	*result = cache;
2656	return 0;
2657bad:
2658	destroy(cache);
2659	return r;
2660}
2661
2662static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2663{
2664	unsigned i;
2665	const char **copy;
2666
2667	copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2668	if (!copy)
2669		return -ENOMEM;
2670	for (i = 0; i < argc; i++) {
2671		copy[i] = kstrdup(argv[i], GFP_KERNEL);
2672		if (!copy[i]) {
2673			while (i--)
2674				kfree(copy[i]);
2675			kfree(copy);
2676			return -ENOMEM;
2677		}
2678	}
2679
2680	cache->nr_ctr_args = argc;
2681	cache->ctr_args = copy;
2682
2683	return 0;
2684}
2685
2686static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2687{
2688	int r = -EINVAL;
2689	struct cache_args *ca;
2690	struct cache *cache = NULL;
2691
2692	ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2693	if (!ca) {
2694		ti->error = "Error allocating memory for cache";
2695		return -ENOMEM;
2696	}
2697	ca->ti = ti;
2698
2699	r = parse_cache_args(ca, argc, argv, &ti->error);
2700	if (r)
2701		goto out;
2702
2703	r = cache_create(ca, &cache);
2704	if (r)
2705		goto out;
2706
2707	r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2708	if (r) {
2709		destroy(cache);
2710		goto out;
2711	}
2712
2713	ti->private = cache;
2714out:
2715	destroy_cache_args(ca);
2716	return r;
2717}
2718
2719/*----------------------------------------------------------------*/
2720
2721static int cache_map(struct dm_target *ti, struct bio *bio)
2722{
2723	struct cache *cache = ti->private;
2724
2725	int r;
2726	bool commit_needed;
2727	dm_oblock_t block = get_bio_block(cache, bio);
2728
2729	init_per_bio_data(bio);
2730	if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2731		/*
2732		 * This can only occur if the io goes to a partial block at
2733		 * the end of the origin device.  We don't cache these.
2734		 * Just remap to the origin and carry on.
2735		 */
2736		remap_to_origin(cache, bio);
2737		accounted_begin(cache, bio);
2738		return DM_MAPIO_REMAPPED;
2739	}
2740
2741	if (discard_or_flush(bio)) {
2742		defer_bio(cache, bio);
2743		return DM_MAPIO_SUBMITTED;
2744	}
2745
2746	r = map_bio(cache, bio, block, &commit_needed);
2747	if (commit_needed)
2748		schedule_commit(&cache->committer);
2749
2750	return r;
2751}
2752
2753static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2754{
2755	struct cache *cache = ti->private;
2756	unsigned long flags;
2757	struct per_bio_data *pb = get_per_bio_data(bio);
2758
2759	if (pb->tick) {
2760		policy_tick(cache->policy, false);
2761
2762		spin_lock_irqsave(&cache->lock, flags);
2763		cache->need_tick_bio = true;
2764		spin_unlock_irqrestore(&cache->lock, flags);
2765	}
2766
2767	bio_drop_shared_lock(cache, bio);
2768	accounted_complete(cache, bio);
2769
2770	return DM_ENDIO_DONE;
2771}
2772
2773static int write_dirty_bitset(struct cache *cache)
2774{
2775	int r;
2776
2777	if (get_cache_mode(cache) >= CM_READ_ONLY)
2778		return -EINVAL;
2779
2780	r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2781	if (r)
2782		metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2783
2784	return r;
2785}
2786
2787static int write_discard_bitset(struct cache *cache)
2788{
2789	unsigned i, r;
2790
2791	if (get_cache_mode(cache) >= CM_READ_ONLY)
2792		return -EINVAL;
2793
2794	r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2795					   cache->discard_nr_blocks);
2796	if (r) {
2797		DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2798		metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2799		return r;
2800	}
2801
2802	for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2803		r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2804					 is_discarded(cache, to_dblock(i)));
2805		if (r) {
2806			metadata_operation_failed(cache, "dm_cache_set_discard", r);
2807			return r;
2808		}
2809	}
2810
2811	return 0;
2812}
2813
2814static int write_hints(struct cache *cache)
2815{
2816	int r;
2817
2818	if (get_cache_mode(cache) >= CM_READ_ONLY)
2819		return -EINVAL;
2820
2821	r = dm_cache_write_hints(cache->cmd, cache->policy);
2822	if (r) {
2823		metadata_operation_failed(cache, "dm_cache_write_hints", r);
2824		return r;
2825	}
2826
2827	return 0;
2828}
2829
2830/*
2831 * returns true on success
2832 */
2833static bool sync_metadata(struct cache *cache)
2834{
2835	int r1, r2, r3, r4;
2836
2837	r1 = write_dirty_bitset(cache);
2838	if (r1)
2839		DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2840
2841	r2 = write_discard_bitset(cache);
2842	if (r2)
2843		DMERR("%s: could not write discard bitset", cache_device_name(cache));
2844
2845	save_stats(cache);
2846
2847	r3 = write_hints(cache);
2848	if (r3)
2849		DMERR("%s: could not write hints", cache_device_name(cache));
2850
2851	/*
2852	 * If writing the above metadata failed, we still commit, but don't
2853	 * set the clean shutdown flag.  This will effectively force every
2854	 * dirty bit to be set on reload.
2855	 */
2856	r4 = commit(cache, !r1 && !r2 && !r3);
2857	if (r4)
2858		DMERR("%s: could not write cache metadata", cache_device_name(cache));
2859
2860	return !r1 && !r2 && !r3 && !r4;
2861}
2862
2863static void cache_postsuspend(struct dm_target *ti)
2864{
2865	struct cache *cache = ti->private;
2866
2867	prevent_background_work(cache);
2868	BUG_ON(atomic_read(&cache->nr_io_migrations));
2869
2870	cancel_delayed_work(&cache->waker);
2871	flush_workqueue(cache->wq);
2872	WARN_ON(cache->tracker.in_flight);
2873
2874	/*
2875	 * If it's a flush suspend there won't be any deferred bios, so this
2876	 * call is harmless.
2877	 */
2878	requeue_deferred_bios(cache);
2879
2880	if (get_cache_mode(cache) == CM_WRITE)
2881		(void) sync_metadata(cache);
2882}
2883
2884static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2885			bool dirty, uint32_t hint, bool hint_valid)
2886{
2887	int r;
2888	struct cache *cache = context;
2889
2890	if (dirty) {
2891		set_bit(from_cblock(cblock), cache->dirty_bitset);
2892		atomic_inc(&cache->nr_dirty);
2893	} else
2894		clear_bit(from_cblock(cblock), cache->dirty_bitset);
2895
2896	r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2897	if (r)
2898		return r;
2899
2900	return 0;
2901}
2902
2903/*
2904 * The discard block size in the on disk metadata is not
2905 * neccessarily the same as we're currently using.  So we have to
2906 * be careful to only set the discarded attribute if we know it
2907 * covers a complete block of the new size.
2908 */
2909struct discard_load_info {
2910	struct cache *cache;
2911
2912	/*
2913	 * These blocks are sized using the on disk dblock size, rather
2914	 * than the current one.
2915	 */
2916	dm_block_t block_size;
2917	dm_block_t discard_begin, discard_end;
2918};
2919
2920static void discard_load_info_init(struct cache *cache,
2921				   struct discard_load_info *li)
2922{
2923	li->cache = cache;
2924	li->discard_begin = li->discard_end = 0;
2925}
2926
2927static void set_discard_range(struct discard_load_info *li)
2928{
2929	sector_t b, e;
2930
2931	if (li->discard_begin == li->discard_end)
2932		return;
2933
2934	/*
2935	 * Convert to sectors.
2936	 */
2937	b = li->discard_begin * li->block_size;
2938	e = li->discard_end * li->block_size;
2939
2940	/*
2941	 * Then convert back to the current dblock size.
2942	 */
2943	b = dm_sector_div_up(b, li->cache->discard_block_size);
2944	sector_div(e, li->cache->discard_block_size);
2945
2946	/*
2947	 * The origin may have shrunk, so we need to check we're still in
2948	 * bounds.
2949	 */
2950	if (e > from_dblock(li->cache->discard_nr_blocks))
2951		e = from_dblock(li->cache->discard_nr_blocks);
2952
2953	for (; b < e; b++)
2954		set_discard(li->cache, to_dblock(b));
2955}
2956
2957static int load_discard(void *context, sector_t discard_block_size,
2958			dm_dblock_t dblock, bool discard)
2959{
2960	struct discard_load_info *li = context;
2961
2962	li->block_size = discard_block_size;
2963
2964	if (discard) {
2965		if (from_dblock(dblock) == li->discard_end)
2966			/*
2967			 * We're already in a discard range, just extend it.
2968			 */
2969			li->discard_end = li->discard_end + 1ULL;
2970
2971		else {
2972			/*
2973			 * Emit the old range and start a new one.
2974			 */
2975			set_discard_range(li);
2976			li->discard_begin = from_dblock(dblock);
2977			li->discard_end = li->discard_begin + 1ULL;
2978		}
2979	} else {
2980		set_discard_range(li);
2981		li->discard_begin = li->discard_end = 0;
2982	}
2983
2984	return 0;
2985}
2986
2987static dm_cblock_t get_cache_dev_size(struct cache *cache)
2988{
2989	sector_t size = get_dev_size(cache->cache_dev);
2990	(void) sector_div(size, cache->sectors_per_block);
2991	return to_cblock(size);
2992}
2993
2994static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2995{
2996	if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2997		if (cache->sized) {
2998			DMERR("%s: unable to extend cache due to missing cache table reload",
2999			      cache_device_name(cache));
3000			return false;
3001		}
3002	}
3003
3004	/*
3005	 * We can't drop a dirty block when shrinking the cache.
3006	 */
3007	while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3008		new_size = to_cblock(from_cblock(new_size) + 1);
3009		if (is_dirty(cache, new_size)) {
3010			DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3011			      cache_device_name(cache),
3012			      (unsigned long long) from_cblock(new_size));
3013			return false;
3014		}
3015	}
3016
3017	return true;
3018}
3019
3020static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3021{
3022	int r;
3023
3024	r = dm_cache_resize(cache->cmd, new_size);
3025	if (r) {
3026		DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3027		metadata_operation_failed(cache, "dm_cache_resize", r);
3028		return r;
3029	}
3030
3031	set_cache_size(cache, new_size);
3032
3033	return 0;
3034}
3035
3036static int cache_preresume(struct dm_target *ti)
3037{
3038	int r = 0;
3039	struct cache *cache = ti->private;
3040	dm_cblock_t csize = get_cache_dev_size(cache);
3041
3042	/*
3043	 * Check to see if the cache has resized.
3044	 */
3045	if (!cache->sized) {
3046		r = resize_cache_dev(cache, csize);
3047		if (r)
3048			return r;
3049
3050		cache->sized = true;
3051
3052	} else if (csize != cache->cache_size) {
3053		if (!can_resize(cache, csize))
3054			return -EINVAL;
3055
3056		r = resize_cache_dev(cache, csize);
3057		if (r)
3058			return r;
3059	}
3060
3061	if (!cache->loaded_mappings) {
3062		r = dm_cache_load_mappings(cache->cmd, cache->policy,
3063					   load_mapping, cache);
3064		if (r) {
3065			DMERR("%s: could not load cache mappings", cache_device_name(cache));
3066			metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3067			return r;
3068		}
3069
3070		cache->loaded_mappings = true;
3071	}
3072
3073	if (!cache->loaded_discards) {
3074		struct discard_load_info li;
3075
3076		/*
3077		 * The discard bitset could have been resized, or the
3078		 * discard block size changed.  To be safe we start by
3079		 * setting every dblock to not discarded.
3080		 */
3081		clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3082
3083		discard_load_info_init(cache, &li);
3084		r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3085		if (r) {
3086			DMERR("%s: could not load origin discards", cache_device_name(cache));
3087			metadata_operation_failed(cache, "dm_cache_load_discards", r);
3088			return r;
3089		}
3090		set_discard_range(&li);
3091
3092		cache->loaded_discards = true;
3093	}
3094
3095	return r;
3096}
3097
3098static void cache_resume(struct dm_target *ti)
3099{
3100	struct cache *cache = ti->private;
3101
3102	cache->need_tick_bio = true;
3103	allow_background_work(cache);
3104	do_waker(&cache->waker.work);
3105}
3106
3107static void emit_flags(struct cache *cache, char *result,
3108		       unsigned maxlen, ssize_t *sz_ptr)
3109{
3110	ssize_t sz = *sz_ptr;
3111	struct cache_features *cf = &cache->features;
3112	unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3113
3114	DMEMIT("%u ", count);
3115
3116	if (cf->metadata_version == 2)
3117		DMEMIT("metadata2 ");
3118
3119	if (writethrough_mode(cache))
3120		DMEMIT("writethrough ");
3121
3122	else if (passthrough_mode(cache))
3123		DMEMIT("passthrough ");
3124
3125	else if (writeback_mode(cache))
3126		DMEMIT("writeback ");
3127
3128	else {
3129		DMEMIT("unknown ");
3130		DMERR("%s: internal error: unknown io mode: %d",
3131		      cache_device_name(cache), (int) cf->io_mode);
3132	}
3133
3134	if (!cf->discard_passdown)
3135		DMEMIT("no_discard_passdown ");
3136
3137	*sz_ptr = sz;
3138}
3139
3140/*
3141 * Status format:
3142 *
3143 * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3144 * <cache block size> <#used cache blocks>/<#total cache blocks>
3145 * <#read hits> <#read misses> <#write hits> <#write misses>
3146 * <#demotions> <#promotions> <#dirty>
3147 * <#features> <features>*
3148 * <#core args> <core args>
3149 * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3150 */
3151static void cache_status(struct dm_target *ti, status_type_t type,
3152			 unsigned status_flags, char *result, unsigned maxlen)
3153{
3154	int r = 0;
3155	unsigned i;
3156	ssize_t sz = 0;
3157	dm_block_t nr_free_blocks_metadata = 0;
3158	dm_block_t nr_blocks_metadata = 0;
3159	char buf[BDEVNAME_SIZE];
3160	struct cache *cache = ti->private;
3161	dm_cblock_t residency;
3162	bool needs_check;
3163
3164	switch (type) {
3165	case STATUSTYPE_INFO:
3166		if (get_cache_mode(cache) == CM_FAIL) {
3167			DMEMIT("Fail");
3168			break;
3169		}
3170
3171		/* Commit to ensure statistics aren't out-of-date */
3172		if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3173			(void) commit(cache, false);
3174
3175		r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3176		if (r) {
3177			DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3178			      cache_device_name(cache), r);
3179			goto err;
3180		}
3181
3182		r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3183		if (r) {
3184			DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3185			      cache_device_name(cache), r);
3186			goto err;
3187		}
3188
3189		residency = policy_residency(cache->policy);
3190
3191		DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3192		       (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3193		       (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3194		       (unsigned long long)nr_blocks_metadata,
3195		       (unsigned long long)cache->sectors_per_block,
3196		       (unsigned long long) from_cblock(residency),
3197		       (unsigned long long) from_cblock(cache->cache_size),
3198		       (unsigned) atomic_read(&cache->stats.read_hit),
3199		       (unsigned) atomic_read(&cache->stats.read_miss),
3200		       (unsigned) atomic_read(&cache->stats.write_hit),
3201		       (unsigned) atomic_read(&cache->stats.write_miss),
3202		       (unsigned) atomic_read(&cache->stats.demotion),
3203		       (unsigned) atomic_read(&cache->stats.promotion),
3204		       (unsigned long) atomic_read(&cache->nr_dirty));
3205
3206		emit_flags(cache, result, maxlen, &sz);
3207
3208		DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3209
3210		DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3211		if (sz < maxlen) {
3212			r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3213			if (r)
3214				DMERR("%s: policy_emit_config_values returned %d",
3215				      cache_device_name(cache), r);
3216		}
3217
3218		if (get_cache_mode(cache) == CM_READ_ONLY)
3219			DMEMIT("ro ");
3220		else
3221			DMEMIT("rw ");
3222
3223		r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3224
3225		if (r || needs_check)
3226			DMEMIT("needs_check ");
3227		else
3228			DMEMIT("- ");
3229
3230		break;
3231
3232	case STATUSTYPE_TABLE:
3233		format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3234		DMEMIT("%s ", buf);
3235		format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3236		DMEMIT("%s ", buf);
3237		format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3238		DMEMIT("%s", buf);
3239
3240		for (i = 0; i < cache->nr_ctr_args - 1; i++)
3241			DMEMIT(" %s", cache->ctr_args[i]);
3242		if (cache->nr_ctr_args)
3243			DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3244	}
3245
3246	return;
3247
3248err:
3249	DMEMIT("Error");
3250}
3251
3252/*
3253 * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
3254 * the one-past-the-end value.
3255 */
3256struct cblock_range {
3257	dm_cblock_t begin;
3258	dm_cblock_t end;
3259};
3260
3261/*
3262 * A cache block range can take two forms:
3263 *
3264 * i) A single cblock, eg. '3456'
3265 * ii) A begin and end cblock with a dash between, eg. 123-234
3266 */
3267static int parse_cblock_range(struct cache *cache, const char *str,
3268			      struct cblock_range *result)
3269{
3270	char dummy;
3271	uint64_t b, e;
3272	int r;
3273
3274	/*
3275	 * Try and parse form (ii) first.
3276	 */
3277	r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3278	if (r < 0)
3279		return r;
3280
3281	if (r == 2) {
3282		result->begin = to_cblock(b);
3283		result->end = to_cblock(e);
3284		return 0;
3285	}
3286
3287	/*
3288	 * That didn't work, try form (i).
3289	 */
3290	r = sscanf(str, "%llu%c", &b, &dummy);
3291	if (r < 0)
3292		return r;
3293
3294	if (r == 1) {
3295		result->begin = to_cblock(b);
3296		result->end = to_cblock(from_cblock(result->begin) + 1u);
3297		return 0;
3298	}
3299
3300	DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3301	return -EINVAL;
3302}
3303
3304static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3305{
3306	uint64_t b = from_cblock(range->begin);
3307	uint64_t e = from_cblock(range->end);
3308	uint64_t n = from_cblock(cache->cache_size);
3309
3310	if (b >= n) {
3311		DMERR("%s: begin cblock out of range: %llu >= %llu",
3312		      cache_device_name(cache), b, n);
3313		return -EINVAL;
3314	}
3315
3316	if (e > n) {
3317		DMERR("%s: end cblock out of range: %llu > %llu",
3318		      cache_device_name(cache), e, n);
3319		return -EINVAL;
3320	}
3321
3322	if (b >= e) {
3323		DMERR("%s: invalid cblock range: %llu >= %llu",
3324		      cache_device_name(cache), b, e);
3325		return -EINVAL;
3326	}
3327
3328	return 0;
3329}
3330
3331static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3332{
3333	return to_cblock(from_cblock(b) + 1);
3334}
3335
3336static int request_invalidation(struct cache *cache, struct cblock_range *range)
3337{
3338	int r = 0;
3339
3340	/*
3341	 * We don't need to do any locking here because we know we're in
3342	 * passthrough mode.  There's is potential for a race between an
3343	 * invalidation triggered by an io and an invalidation message.  This
3344	 * is harmless, we must not worry if the policy call fails.
3345	 */
3346	while (range->begin != range->end) {
3347		r = invalidate_cblock(cache, range->begin);
3348		if (r)
3349			return r;
3350
3351		range->begin = cblock_succ(range->begin);
3352	}
3353
3354	cache->commit_requested = true;
3355	return r;
3356}
3357
3358static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3359					      const char **cblock_ranges)
3360{
3361	int r = 0;
3362	unsigned i;
3363	struct cblock_range range;
3364
3365	if (!passthrough_mode(cache)) {
3366		DMERR("%s: cache has to be in passthrough mode for invalidation",
3367		      cache_device_name(cache));
3368		return -EPERM;
3369	}
3370
3371	for (i = 0; i < count; i++) {
3372		r = parse_cblock_range(cache, cblock_ranges[i], &range);
3373		if (r)
3374			break;
3375
3376		r = validate_cblock_range(cache, &range);
3377		if (r)
3378			break;
3379
3380		/*
3381		 * Pass begin and end origin blocks to the worker and wake it.
3382		 */
3383		r = request_invalidation(cache, &range);
3384		if (r)
3385			break;
3386	}
3387
3388	return r;
3389}
3390
3391/*
3392 * Supports
3393 *	"<key> <value>"
3394 * and
3395 *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3396 *
3397 * The key migration_threshold is supported by the cache target core.
3398 */
3399static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3400			 char *result, unsigned maxlen)
3401{
3402	struct cache *cache = ti->private;
3403
3404	if (!argc)
3405		return -EINVAL;
3406
3407	if (get_cache_mode(cache) >= CM_READ_ONLY) {
3408		DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3409		      cache_device_name(cache));
3410		return -EOPNOTSUPP;
3411	}
3412
3413	if (!strcasecmp(argv[0], "invalidate_cblocks"))
3414		return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3415
3416	if (argc != 2)
3417		return -EINVAL;
3418
3419	return set_config_value(cache, argv[0], argv[1]);
3420}
3421
3422static int cache_iterate_devices(struct dm_target *ti,
3423				 iterate_devices_callout_fn fn, void *data)
3424{
3425	int r = 0;
3426	struct cache *cache = ti->private;
3427
3428	r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3429	if (!r)
3430		r = fn(ti, cache->origin_dev, 0, ti->len, data);
3431
3432	return r;
3433}
3434
3435static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3436{
3437	struct request_queue *q = bdev_get_queue(origin_bdev);
3438
3439	return q && blk_queue_discard(q);
3440}
3441
3442/*
3443 * If discard_passdown was enabled verify that the origin device
3444 * supports discards.  Disable discard_passdown if not.
3445 */
3446static void disable_passdown_if_not_supported(struct cache *cache)
3447{
3448	struct block_device *origin_bdev = cache->origin_dev->bdev;
3449	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3450	const char *reason = NULL;
3451	char buf[BDEVNAME_SIZE];
3452
3453	if (!cache->features.discard_passdown)
3454		return;
3455
3456	if (!origin_dev_supports_discard(origin_bdev))
3457		reason = "discard unsupported";
3458
3459	else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3460		reason = "max discard sectors smaller than a block";
3461
3462	if (reason) {
3463		DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3464		       bdevname(origin_bdev, buf), reason);
3465		cache->features.discard_passdown = false;
3466	}
3467}
3468
3469static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3470{
3471	struct block_device *origin_bdev = cache->origin_dev->bdev;
3472	struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3473
3474	if (!cache->features.discard_passdown) {
3475		/* No passdown is done so setting own virtual limits */
3476		limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3477						    cache->origin_sectors);
3478		limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3479		return;
3480	}
3481
3482	/*
3483	 * cache_iterate_devices() is stacking both origin and fast device limits
3484	 * but discards aren't passed to fast device, so inherit origin's limits.
3485	 */
3486	limits->max_discard_sectors = origin_limits->max_discard_sectors;
3487	limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3488	limits->discard_granularity = origin_limits->discard_granularity;
3489	limits->discard_alignment = origin_limits->discard_alignment;
3490	limits->discard_misaligned = origin_limits->discard_misaligned;
3491}
3492
3493static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3494{
3495	struct cache *cache = ti->private;
3496	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3497
3498	/*
3499	 * If the system-determined stacked limits are compatible with the
3500	 * cache's blocksize (io_opt is a factor) do not override them.
3501	 */
3502	if (io_opt_sectors < cache->sectors_per_block ||
3503	    do_div(io_opt_sectors, cache->sectors_per_block)) {
3504		blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3505		blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3506	}
3507
3508	disable_passdown_if_not_supported(cache);
3509	set_discard_limits(cache, limits);
3510}
3511
3512/*----------------------------------------------------------------*/
3513
3514static struct target_type cache_target = {
3515	.name = "cache",
3516	.version = {2, 1, 0},
3517	.module = THIS_MODULE,
3518	.ctr = cache_ctr,
3519	.dtr = cache_dtr,
3520	.map = cache_map,
3521	.end_io = cache_end_io,
3522	.postsuspend = cache_postsuspend,
3523	.preresume = cache_preresume,
3524	.resume = cache_resume,
3525	.status = cache_status,
3526	.message = cache_message,
3527	.iterate_devices = cache_iterate_devices,
3528	.io_hints = cache_io_hints,
3529};
3530
3531static int __init dm_cache_init(void)
3532{
3533	int r;
3534
3535	migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3536	if (!migration_cache)
3537		return -ENOMEM;
3538
3539	r = dm_register_target(&cache_target);
3540	if (r) {
3541		DMERR("cache target registration failed: %d", r);
3542		kmem_cache_destroy(migration_cache);
3543		return r;
3544	}
3545
3546	return 0;
3547}
3548
3549static void __exit dm_cache_exit(void)
3550{
3551	dm_unregister_target(&cache_target);
3552	kmem_cache_destroy(migration_cache);
3553}
3554
3555module_init(dm_cache_init);
3556module_exit(dm_cache_exit);
3557
3558MODULE_DESCRIPTION(DM_NAME " cache target");
3559MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3560MODULE_LICENSE("GPL");