drivers/md/dm-cache-target.c

   1 /*
   2  * Copyright (C) 2012 Red Hat. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm.h"
   8 #include "dm-bio-prison.h"
   9 #include "dm-bio-record.h"
  10 #include "dm-cache-metadata.h"
  11
  12 #include <linux/dm-io.h>
  13 #include <linux/dm-kcopyd.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/slab.h>
  18 #include <linux/vmalloc.h>
  19
  20 #define DM_MSG_PREFIX "cache"
  21
  22 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  23         "A percentage of time allocated for copying to and/or from cache");
  24
  25 /*----------------------------------------------------------------*/
  26
  27 /*
  28  * Glossary:
  29  *
  30  * oblock: index of an origin block
  31  * cblock: index of a cache block
  32  * promotion: movement of a block from origin to cache
  33  * demotion: movement of a block from cache to origin
  34  * migration: movement of a block between the origin and cache device,
  35  *            either direction
  36  */
  37
  38 /*----------------------------------------------------------------*/
  39
  40 static size_t bitset_size_in_bytes(unsigned nr_entries)
  41 {
  42         return sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
  43 }
  44
  45 static unsigned long *alloc_bitset(unsigned nr_entries)
  46 {
  47         size_t s = bitset_size_in_bytes(nr_entries);
  48         return vzalloc(s);
  49 }
  50
  51 static void clear_bitset(void *bitset, unsigned nr_entries)
  52 {
  53         size_t s = bitset_size_in_bytes(nr_entries);
  54         memset(bitset, 0, s);
  55 }
  56
  57 static void free_bitset(unsigned long *bits)
  58 {
  59         vfree(bits);
  60 }
  61
  62 /*----------------------------------------------------------------*/
  63
  64 #define PRISON_CELLS 1024
  65 #define MIGRATION_POOL_SIZE 128
  66 #define COMMIT_PERIOD HZ
  67 #define MIGRATION_COUNT_WINDOW 10
  68
  69 /*
  70  * The block size of the device holding cache data must be >= 32KB
  71  */
  72 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
  73
  74 /*
  75  * FIXME: the cache is read/write for the time being.
  76  */
  77 enum cache_mode {
  78         CM_WRITE,               /* metadata may be changed */
  79         CM_READ_ONLY,           /* metadata may not be changed */
  80 };
  81
  82 struct cache_features {
  83         enum cache_mode mode;
  84         bool write_through:1;
  85 };
  86
  87 struct cache_stats {
  88         atomic_t read_hit;
  89         atomic_t read_miss;
  90         atomic_t write_hit;
  91         atomic_t write_miss;
  92         atomic_t demotion;
  93         atomic_t promotion;
  94         atomic_t copies_avoided;
  95         atomic_t cache_cell_clash;
  96         atomic_t commit_count;
  97         atomic_t discard_count;
  98 };
  99
 100 struct cache {
 101         struct dm_target *ti;
 102         struct dm_target_callbacks callbacks;
 103
 104         /*
 105          * Metadata is written to this device.
 106          */
 107         struct dm_dev *metadata_dev;
 108
 109         /*
 110          * The slower of the two data devices.  Typically a spindle.
 111          */
 112         struct dm_dev *origin_dev;
 113
 114         /*
 115          * The faster of the two data devices.  Typically an SSD.
 116          */
 117         struct dm_dev *cache_dev;
 118
 119         /*
 120          * Cache features such as write-through.
 121          */
 122         struct cache_features features;
 123
 124         /*
 125          * Size of the origin device in _complete_ blocks and native sectors.
 126          */
 127         dm_oblock_t origin_blocks;
 128         sector_t origin_sectors;
 129
 130         /*
 131          * Size of the cache device in blocks.
 132          */
 133         dm_cblock_t cache_size;
 134
 135         /*
 136          * Fields for converting from sectors to blocks.
 137          */
 138         uint32_t sectors_per_block;
 139         int sectors_per_block_shift;
 140
 141         struct dm_cache_metadata *cmd;
 142
 143         spinlock_t lock;
 144         struct bio_list deferred_bios;
 145         struct bio_list deferred_flush_bios;
 146         struct bio_list deferred_writethrough_bios;
 147         struct list_head quiesced_migrations;
 148         struct list_head completed_migrations;
 149         struct list_head need_commit_migrations;
 150         sector_t migration_threshold;
 151         atomic_t nr_migrations;
 152         wait_queue_head_t migration_wait;
 153
 154         /*
 155          * cache_size entries, dirty if set
 156          */
 157         dm_cblock_t nr_dirty;
 158         unsigned long *dirty_bitset;
 159
 160         /*
 161          * origin_blocks entries, discarded if set.
 162          */
 163         uint32_t discard_block_size; /* a power of 2 times sectors per block */
 164         dm_dblock_t discard_nr_blocks;
 165         unsigned long *discard_bitset;
 166
 167         struct dm_kcopyd_client *copier;
 168         struct workqueue_struct *wq;
 169         struct work_struct worker;
 170
 171         struct delayed_work waker;
 172         unsigned long last_commit_jiffies;
 173
 174         struct dm_bio_prison *prison;
 175         struct dm_deferred_set *all_io_ds;
 176
 177         mempool_t *migration_pool;
 178         struct dm_cache_migration *next_migration;
 179
 180         struct dm_cache_policy *policy;
 181         unsigned policy_nr_args;
 182
 183         bool need_tick_bio:1;
 184         bool sized:1;
 185         bool quiescing:1;
 186         bool commit_requested:1;
 187         bool loaded_mappings:1;
 188         bool loaded_discards:1;
 189
 190         struct cache_stats stats;
 191
 192         /*
 193          * Rather than reconstructing the table line for the status we just
 194          * save it and regurgitate.
 195          */
 196         unsigned nr_ctr_args;
 197         const char **ctr_args;
 198 };
 199
 200 struct per_bio_data {
 201         bool tick:1;
 202         unsigned req_nr:2;
 203         struct dm_deferred_entry *all_io_entry;
 204
 205         /*
 206          * writethrough fields.  These MUST remain at the end of this
 207          * structure and the 'cache' member must be the first as it
 208          * is used to determine the offset of the writethrough fields.
 209          */
 210         struct cache *cache;
 211         dm_cblock_t cblock;
 212         bio_end_io_t *saved_bi_end_io;
 213         struct dm_bio_details bio_details;
 214 };
 215
 216 struct dm_cache_migration {
 217         struct list_head list;
 218         struct cache *cache;
 219
 220         unsigned long start_jiffies;
 221         dm_oblock_t old_oblock;
 222         dm_oblock_t new_oblock;
 223         dm_cblock_t cblock;
 224
 225         bool err:1;
 226         bool writeback:1;
 227         bool demote:1;
 228         bool promote:1;
 229
 230         struct dm_bio_prison_cell *old_ocell;
 231         struct dm_bio_prison_cell *new_ocell;
 232 };
 233
 234 /*
 235  * Processing a bio in the worker thread may require these memory
 236  * allocations.  We prealloc to avoid deadlocks (the same worker thread
 237  * frees them back to the mempool).
 238  */
 239 struct prealloc {
 240         struct dm_cache_migration *mg;
 241         struct dm_bio_prison_cell *cell1;
 242         struct dm_bio_prison_cell *cell2;
 243 };
 244
 245 static void wake_worker(struct cache *cache)
 246 {
 247         queue_work(cache->wq, &cache->worker);
 248 }
 249
 250 /*----------------------------------------------------------------*/
 251
 252 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
 253 {
 254         /* FIXME: change to use a local slab. */
 255         return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
 256 }
 257
 258 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
 259 {
 260         dm_bio_prison_free_cell(cache->prison, cell);
 261 }
 262
 263 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
 264 {
 265         if (!p->mg) {
 266                 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
 267                 if (!p->mg)
 268                         return -ENOMEM;
 269         }
 270
 271         if (!p->cell1) {
 272                 p->cell1 = alloc_prison_cell(cache);
 273                 if (!p->cell1)
 274                         return -ENOMEM;
 275         }
 276
 277         if (!p->cell2) {
 278                 p->cell2 = alloc_prison_cell(cache);
 279                 if (!p->cell2)
 280                         return -ENOMEM;
 281         }
 282
 283         return 0;
 284 }
 285
 286 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
 287 {
 288         if (p->cell2)
 289                 free_prison_cell(cache, p->cell2);
 290
 291         if (p->cell1)
 292                 free_prison_cell(cache, p->cell1);
 293
 294         if (p->mg)
 295                 mempool_free(p->mg, cache->migration_pool);
 296 }
 297
 298 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
 299 {
 300         struct dm_cache_migration *mg = p->mg;
 301
 302         BUG_ON(!mg);
 303         p->mg = NULL;
 304
 305         return mg;
 306 }
 307
 308 /*
 309  * You must have a cell within the prealloc struct to return.  If not this
 310  * function will BUG() rather than returning NULL.
 311  */
 312 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
 313 {
 314         struct dm_bio_prison_cell *r = NULL;
 315
 316         if (p->cell1) {
 317                 r = p->cell1;
 318                 p->cell1 = NULL;
 319
 320         } else if (p->cell2) {
 321                 r = p->cell2;
 322                 p->cell2 = NULL;
 323         } else
 324                 BUG();
 325
 326         return r;
 327 }
 328
 329 /*
 330  * You can't have more than two cells in a prealloc struct.  BUG() will be
 331  * called if you try and overfill.
 332  */
 333 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
 334 {
 335         if (!p->cell2)
 336                 p->cell2 = cell;
 337
 338         else if (!p->cell1)
 339                 p->cell1 = cell;
 340
 341         else
 342                 BUG();
 343 }
 344
 345 /*----------------------------------------------------------------*/
 346
 347 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
 348 {
 349         key->virtual = 0;
 350         key->dev = 0;
 351         key->block = from_oblock(oblock);
 352 }
 353
 354 /*
 355  * The caller hands in a preallocated cell, and a free function for it.
 356  * The cell will be freed if there's an error, or if it wasn't used because
 357  * a cell with that key already exists.
 358  */
 359 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
 360
 361 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
 362                       struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 363                       cell_free_fn free_fn, void *free_context,
 364                       struct dm_bio_prison_cell **cell_result)
 365 {
 366         int r;
 367         struct dm_cell_key key;
 368
 369         build_key(oblock, &key);
 370         r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
 371         if (r)
 372                 free_fn(free_context, cell_prealloc);
 373
 374         return r;
 375 }
 376
 377 static int get_cell(struct cache *cache,
 378                     dm_oblock_t oblock,
 379                     struct prealloc *structs,
 380                     struct dm_bio_prison_cell **cell_result)
 381 {
 382         int r;
 383         struct dm_cell_key key;
 384         struct dm_bio_prison_cell *cell_prealloc;
 385
 386         cell_prealloc = prealloc_get_cell(structs);
 387
 388         build_key(oblock, &key);
 389         r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
 390         if (r)
 391                 prealloc_put_cell(structs, cell_prealloc);
 392
 393         return r;
 394 }
 395
 396 /*----------------------------------------------------------------*/
 397
 398 static bool is_dirty(struct cache *cache, dm_cblock_t b)
 399 {
 400         return test_bit(from_cblock(b), cache->dirty_bitset);
 401 }
 402
 403 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 404 {
 405         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 406                 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
 407                 policy_set_dirty(cache->policy, oblock);
 408         }
 409 }
 410
 411 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 412 {
 413         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 414                 policy_clear_dirty(cache->policy, oblock);
 415                 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
 416                 if (!from_cblock(cache->nr_dirty))
 417                         dm_table_event(cache->ti->table);
 418         }
 419 }
 420
 421 /*----------------------------------------------------------------*/
 422
 423 static bool block_size_is_power_of_two(struct cache *cache)
 424 {
 425         return cache->sectors_per_block_shift >= 0;
 426 }
 427
 428 static dm_block_t block_div(dm_block_t b, uint32_t n)
 429 {
 430         do_div(b, n);
 431
 432         return b;
 433 }
 434
 435 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 436 {
 437         uint32_t discard_blocks = cache->discard_block_size;
 438         dm_block_t b = from_oblock(oblock);
 439
 440         if (!block_size_is_power_of_two(cache))
 441                 discard_blocks = discard_blocks / cache->sectors_per_block;
 442         else
 443                 discard_blocks >>= cache->sectors_per_block_shift;
 444
 445         b = block_div(b, discard_blocks);
 446
 447         return to_dblock(b);
 448 }
 449
 450 static void set_discard(struct cache *cache, dm_dblock_t b)
 451 {
 452         unsigned long flags;
 453
 454         atomic_inc(&cache->stats.discard_count);
 455
 456         spin_lock_irqsave(&cache->lock, flags);
 457         set_bit(from_dblock(b), cache->discard_bitset);
 458         spin_unlock_irqrestore(&cache->lock, flags);
 459 }
 460
 461 static void clear_discard(struct cache *cache, dm_dblock_t b)
 462 {
 463         unsigned long flags;
 464
 465         spin_lock_irqsave(&cache->lock, flags);
 466         clear_bit(from_dblock(b), cache->discard_bitset);
 467         spin_unlock_irqrestore(&cache->lock, flags);
 468 }
 469
 470 static bool is_discarded(struct cache *cache, dm_dblock_t b)
 471 {
 472         int r;
 473         unsigned long flags;
 474
 475         spin_lock_irqsave(&cache->lock, flags);
 476         r = test_bit(from_dblock(b), cache->discard_bitset);
 477         spin_unlock_irqrestore(&cache->lock, flags);
 478
 479         return r;
 480 }
 481
 482 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 483 {
 484         int r;
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(&cache->lock, flags);
 488         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 489                      cache->discard_bitset);
 490         spin_unlock_irqrestore(&cache->lock, flags);
 491
 492         return r;
 493 }
 494
 495 /*----------------------------------------------------------------*/
 496
 497 static void load_stats(struct cache *cache)
 498 {
 499         struct dm_cache_statistics stats;
 500
 501         dm_cache_metadata_get_stats(cache->cmd, &stats);
 502         atomic_set(&cache->stats.read_hit, stats.read_hits);
 503         atomic_set(&cache->stats.read_miss, stats.read_misses);
 504         atomic_set(&cache->stats.write_hit, stats.write_hits);
 505         atomic_set(&cache->stats.write_miss, stats.write_misses);
 506 }
 507
 508 static void save_stats(struct cache *cache)
 509 {
 510         struct dm_cache_statistics stats;
 511
 512         stats.read_hits = atomic_read(&cache->stats.read_hit);
 513         stats.read_misses = atomic_read(&cache->stats.read_miss);
 514         stats.write_hits = atomic_read(&cache->stats.write_hit);
 515         stats.write_misses = atomic_read(&cache->stats.write_miss);
 516
 517         dm_cache_metadata_set_stats(cache->cmd, &stats);
 518 }
 519
 520 /*----------------------------------------------------------------
 521  * Per bio data
 522  *--------------------------------------------------------------*/
 523
 524 /*
 525  * If using writeback, leave out struct per_bio_data's writethrough fields.
 526  */
 527 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
 528 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
 529
 530 static size_t get_per_bio_data_size(struct cache *cache)
 531 {
 532         return cache->features.write_through ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
 533 }
 534
 535 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
 536 {
 537         struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
 538         BUG_ON(!pb);
 539         return pb;
 540 }
 541
 542 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
 543 {
 544         struct per_bio_data *pb = get_per_bio_data(bio, data_size);
 545
 546         pb->tick = false;
 547         pb->req_nr = dm_bio_get_target_bio_nr(bio);
 548         pb->all_io_entry = NULL;
 549
 550         return pb;
 551 }
 552
 553 /*----------------------------------------------------------------
 554  * Remapping
 555  *--------------------------------------------------------------*/
 556 static void remap_to_origin(struct cache *cache, struct bio *bio)
 557 {
 558         bio->bi_bdev = cache->origin_dev->bdev;
 559 }
 560
 561 static void remap_to_cache(struct cache *cache, struct bio *bio,
 562                            dm_cblock_t cblock)
 563 {
 564         sector_t bi_sector = bio->bi_sector;
 565
 566         bio->bi_bdev = cache->cache_dev->bdev;
 567         if (!block_size_is_power_of_two(cache))
 568                 bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
 569                                 sector_div(bi_sector, cache->sectors_per_block);
 570         else
 571                 bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
 572                                 (bi_sector & (cache->sectors_per_block - 1));
 573 }
 574
 575 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 576 {
 577         unsigned long flags;
 578         size_t pb_data_size = get_per_bio_data_size(cache);
 579         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 580
 581         spin_lock_irqsave(&cache->lock, flags);
 582         if (cache->need_tick_bio &&
 583             !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
 584                 pb->tick = true;
 585                 cache->need_tick_bio = false;
 586         }
 587         spin_unlock_irqrestore(&cache->lock, flags);
 588 }
 589
 590 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 591                                   dm_oblock_t oblock)
 592 {
 593         check_if_tick_bio_needed(cache, bio);
 594         remap_to_origin(cache, bio);
 595         if (bio_data_dir(bio) == WRITE)
 596                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 597 }
 598
 599 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 600                                  dm_oblock_t oblock, dm_cblock_t cblock)
 601 {
 602         remap_to_cache(cache, bio, cblock);
 603         if (bio_data_dir(bio) == WRITE) {
 604                 set_dirty(cache, oblock, cblock);
 605                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 606         }
 607 }
 608
 609 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 610 {
 611         sector_t block_nr = bio->bi_sector;
 612
 613         if (!block_size_is_power_of_two(cache))
 614                 (void) sector_div(block_nr, cache->sectors_per_block);
 615         else
 616                 block_nr >>= cache->sectors_per_block_shift;
 617
 618         return to_oblock(block_nr);
 619 }
 620
 621 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
 622 {
 623         return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
 624 }
 625
 626 static void issue(struct cache *cache, struct bio *bio)
 627 {
 628         unsigned long flags;
 629
 630         if (!bio_triggers_commit(cache, bio)) {
 631                 generic_make_request(bio);
 632                 return;
 633         }
 634
 635         /*
 636          * Batch together any bios that trigger commits and then issue a
 637          * single commit for them in do_worker().
 638          */
 639         spin_lock_irqsave(&cache->lock, flags);
 640         cache->commit_requested = true;
 641         bio_list_add(&cache->deferred_flush_bios, bio);
 642         spin_unlock_irqrestore(&cache->lock, flags);
 643 }
 644
 645 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
 646 {
 647         unsigned long flags;
 648
 649         spin_lock_irqsave(&cache->lock, flags);
 650         bio_list_add(&cache->deferred_writethrough_bios, bio);
 651         spin_unlock_irqrestore(&cache->lock, flags);
 652
 653         wake_worker(cache);
 654 }
 655
 656 static void writethrough_endio(struct bio *bio, int err)
 657 {
 658         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 659         bio->bi_end_io = pb->saved_bi_end_io;
 660
 661         if (err) {
 662                 bio_endio(bio, err);
 663                 return;
 664         }
 665
 666         dm_bio_restore(&pb->bio_details, bio);
 667         remap_to_cache(pb->cache, bio, pb->cblock);
 668
 669         /*
 670          * We can't issue this bio directly, since we're in interrupt
 671          * context.  So it gets put on a bio list for processing by the
 672          * worker thread.
 673          */
 674         defer_writethrough_bio(pb->cache, bio);
 675 }
 676
 677 /*
 678  * When running in writethrough mode we need to send writes to clean blocks
 679  * to both the cache and origin devices.  In future we'd like to clone the
 680  * bio and send them in parallel, but for now we're doing them in
 681  * series as this is easier.
 682  */
 683 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
 684                                        dm_oblock_t oblock, dm_cblock_t cblock)
 685 {
 686         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 687
 688         pb->cache = cache;
 689         pb->cblock = cblock;
 690         pb->saved_bi_end_io = bio->bi_end_io;
 691         dm_bio_record(&pb->bio_details, bio);
 692         bio->bi_end_io = writethrough_endio;
 693
 694         remap_to_origin_clear_discard(pb->cache, bio, oblock);
 695 }
 696
 697 /*----------------------------------------------------------------
 698  * Migration processing
 699  *
 700  * Migration covers moving data from the origin device to the cache, or
 701  * vice versa.
 702  *--------------------------------------------------------------*/
 703 static void free_migration(struct dm_cache_migration *mg)
 704 {
 705         mempool_free(mg, mg->cache->migration_pool);
 706 }
 707
 708 static void inc_nr_migrations(struct cache *cache)
 709 {
 710         atomic_inc(&cache->nr_migrations);
 711 }
 712
 713 static void dec_nr_migrations(struct cache *cache)
 714 {
 715         atomic_dec(&cache->nr_migrations);
 716
 717         /*
 718          * Wake the worker in case we're suspending the target.
 719          */
 720         wake_up(&cache->migration_wait);
 721 }
 722
 723 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
 724                          bool holder)
 725 {
 726         (holder ? dm_cell_release : dm_cell_release_no_holder)
 727                 (cache->prison, cell, &cache->deferred_bios);
 728         free_prison_cell(cache, cell);
 729 }
 730
 731 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
 732                        bool holder)
 733 {
 734         unsigned long flags;
 735
 736         spin_lock_irqsave(&cache->lock, flags);
 737         __cell_defer(cache, cell, holder);
 738         spin_unlock_irqrestore(&cache->lock, flags);
 739
 740         wake_worker(cache);
 741 }
 742
 743 static void cleanup_migration(struct dm_cache_migration *mg)
 744 {
 745         dec_nr_migrations(mg->cache);
 746         free_migration(mg);
 747 }
 748
 749 static void migration_failure(struct dm_cache_migration *mg)
 750 {
 751         struct cache *cache = mg->cache;
 752
 753         if (mg->writeback) {
 754                 DMWARN_LIMIT("writeback failed; couldn't copy block");
 755                 set_dirty(cache, mg->old_oblock, mg->cblock);
 756                 cell_defer(cache, mg->old_ocell, false);
 757
 758         } else if (mg->demote) {
 759                 DMWARN_LIMIT("demotion failed; couldn't copy block");
 760                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
 761
 762                 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
 763                 if (mg->promote)
 764                         cell_defer(cache, mg->new_ocell, 1);
 765         } else {
 766                 DMWARN_LIMIT("promotion failed; couldn't copy block");
 767                 policy_remove_mapping(cache->policy, mg->new_oblock);
 768                 cell_defer(cache, mg->new_ocell, 1);
 769         }
 770
 771         cleanup_migration(mg);
 772 }
 773
 774 static void migration_success_pre_commit(struct dm_cache_migration *mg)
 775 {
 776         unsigned long flags;
 777         struct cache *cache = mg->cache;
 778
 779         if (mg->writeback) {
 780                 cell_defer(cache, mg->old_ocell, false);
 781                 clear_dirty(cache, mg->old_oblock, mg->cblock);
 782                 cleanup_migration(mg);
 783                 return;
 784
 785         } else if (mg->demote) {
 786                 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
 787                         DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
 788                         policy_force_mapping(cache->policy, mg->new_oblock,
 789                                              mg->old_oblock);
 790                         if (mg->promote)
 791                                 cell_defer(cache, mg->new_ocell, true);
 792                         cleanup_migration(mg);
 793                         return;
 794                 }
 795         } else {
 796                 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
 797                         DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
 798                         policy_remove_mapping(cache->policy, mg->new_oblock);
 799                         cleanup_migration(mg);
 800                         return;
 801                 }
 802         }
 803
 804         spin_lock_irqsave(&cache->lock, flags);
 805         list_add_tail(&mg->list, &cache->need_commit_migrations);
 806         cache->commit_requested = true;
 807         spin_unlock_irqrestore(&cache->lock, flags);
 808 }
 809
 810 static void migration_success_post_commit(struct dm_cache_migration *mg)
 811 {
 812         unsigned long flags;
 813         struct cache *cache = mg->cache;
 814
 815         if (mg->writeback) {
 816                 DMWARN("writeback unexpectedly triggered commit");
 817                 return;
 818
 819         } else if (mg->demote) {
 820                 cell_defer(cache, mg->old_ocell, mg->promote ? 0 : 1);
 821
 822                 if (mg->promote) {
 823                         mg->demote = false;
 824
 825                         spin_lock_irqsave(&cache->lock, flags);
 826                         list_add_tail(&mg->list, &cache->quiesced_migrations);
 827                         spin_unlock_irqrestore(&cache->lock, flags);
 828
 829                 } else
 830                         cleanup_migration(mg);
 831
 832         } else {
 833                 cell_defer(cache, mg->new_ocell, true);
 834                 clear_dirty(cache, mg->new_oblock, mg->cblock);
 835                 cleanup_migration(mg);
 836         }
 837 }
 838
 839 static void copy_complete(int read_err, unsigned long write_err, void *context)
 840 {
 841         unsigned long flags;
 842         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
 843         struct cache *cache = mg->cache;
 844
 845         if (read_err || write_err)
 846                 mg->err = true;
 847
 848         spin_lock_irqsave(&cache->lock, flags);
 849         list_add_tail(&mg->list, &cache->completed_migrations);
 850         spin_unlock_irqrestore(&cache->lock, flags);
 851
 852         wake_worker(cache);
 853 }
 854
 855 static void issue_copy_real(struct dm_cache_migration *mg)
 856 {
 857         int r;
 858         struct dm_io_region o_region, c_region;
 859         struct cache *cache = mg->cache;
 860
 861         o_region.bdev = cache->origin_dev->bdev;
 862         o_region.count = cache->sectors_per_block;
 863
 864         c_region.bdev = cache->cache_dev->bdev;
 865         c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
 866         c_region.count = cache->sectors_per_block;
 867
 868         if (mg->writeback || mg->demote) {
 869                 /* demote */
 870                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
 871                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
 872         } else {
 873                 /* promote */
 874                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
 875                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
 876         }
 877
 878         if (r < 0)
 879                 migration_failure(mg);
 880 }
 881
 882 static void avoid_copy(struct dm_cache_migration *mg)
 883 {
 884         atomic_inc(&mg->cache->stats.copies_avoided);
 885         migration_success_pre_commit(mg);
 886 }
 887
 888 static void issue_copy(struct dm_cache_migration *mg)
 889 {
 890         bool avoid;
 891         struct cache *cache = mg->cache;
 892
 893         if (mg->writeback || mg->demote)
 894                 avoid = !is_dirty(cache, mg->cblock) ||
 895                         is_discarded_oblock(cache, mg->old_oblock);
 896         else
 897                 avoid = is_discarded_oblock(cache, mg->new_oblock);
 898
 899         avoid ? avoid_copy(mg) : issue_copy_real(mg);
 900 }
 901
 902 static void complete_migration(struct dm_cache_migration *mg)
 903 {
 904         if (mg->err)
 905                 migration_failure(mg);
 906         else
 907                 migration_success_pre_commit(mg);
 908 }
 909
 910 static void process_migrations(struct cache *cache, struct list_head *head,
 911                                void (*fn)(struct dm_cache_migration *))
 912 {
 913         unsigned long flags;
 914         struct list_head list;
 915         struct dm_cache_migration *mg, *tmp;
 916
 917         INIT_LIST_HEAD(&list);
 918         spin_lock_irqsave(&cache->lock, flags);
 919         list_splice_init(head, &list);
 920         spin_unlock_irqrestore(&cache->lock, flags);
 921
 922         list_for_each_entry_safe(mg, tmp, &list, list)
 923                 fn(mg);
 924 }
 925
 926 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
 927 {
 928         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
 929 }
 930
 931 static void queue_quiesced_migration(struct dm_cache_migration *mg)
 932 {
 933         unsigned long flags;
 934         struct cache *cache = mg->cache;
 935
 936         spin_lock_irqsave(&cache->lock, flags);
 937         __queue_quiesced_migration(mg);
 938         spin_unlock_irqrestore(&cache->lock, flags);
 939
 940         wake_worker(cache);
 941 }
 942
 943 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
 944 {
 945         unsigned long flags;
 946         struct dm_cache_migration *mg, *tmp;
 947
 948         spin_lock_irqsave(&cache->lock, flags);
 949         list_for_each_entry_safe(mg, tmp, work, list)
 950                 __queue_quiesced_migration(mg);
 951         spin_unlock_irqrestore(&cache->lock, flags);
 952
 953         wake_worker(cache);
 954 }
 955
 956 static void check_for_quiesced_migrations(struct cache *cache,
 957                                           struct per_bio_data *pb)
 958 {
 959         struct list_head work;
 960
 961         if (!pb->all_io_entry)
 962                 return;
 963
 964         INIT_LIST_HEAD(&work);
 965         if (pb->all_io_entry)
 966                 dm_deferred_entry_dec(pb->all_io_entry, &work);
 967
 968         if (!list_empty(&work))
 969                 queue_quiesced_migrations(cache, &work);
 970 }
 971
 972 static void quiesce_migration(struct dm_cache_migration *mg)
 973 {
 974         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
 975                 queue_quiesced_migration(mg);
 976 }
 977
 978 static void promote(struct cache *cache, struct prealloc *structs,
 979                     dm_oblock_t oblock, dm_cblock_t cblock,
 980                     struct dm_bio_prison_cell *cell)
 981 {
 982         struct dm_cache_migration *mg = prealloc_get_migration(structs);
 983
 984         mg->err = false;
 985         mg->writeback = false;
 986         mg->demote = false;
 987         mg->promote = true;
 988         mg->cache = cache;
 989         mg->new_oblock = oblock;
 990         mg->cblock = cblock;
 991         mg->old_ocell = NULL;
 992         mg->new_ocell = cell;
 993         mg->start_jiffies = jiffies;
 994
 995         inc_nr_migrations(cache);
 996         quiesce_migration(mg);
 997 }
 998
 999 static void writeback(struct cache *cache, struct prealloc *structs,
1000                       dm_oblock_t oblock, dm_cblock_t cblock,
1001                       struct dm_bio_prison_cell *cell)
1002 {
1003         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1004
1005         mg->err = false;
1006         mg->writeback = true;
1007         mg->demote = false;
1008         mg->promote = false;
1009         mg->cache = cache;
1010         mg->old_oblock = oblock;
1011         mg->cblock = cblock;
1012         mg->old_ocell = cell;
1013         mg->new_ocell = NULL;
1014         mg->start_jiffies = jiffies;
1015
1016         inc_nr_migrations(cache);
1017         quiesce_migration(mg);
1018 }
1019
1020 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1021                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1022                                 dm_cblock_t cblock,
1023                                 struct dm_bio_prison_cell *old_ocell,
1024                                 struct dm_bio_prison_cell *new_ocell)
1025 {
1026         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1027
1028         mg->err = false;
1029         mg->writeback = false;
1030         mg->demote = true;
1031         mg->promote = true;
1032         mg->cache = cache;
1033         mg->old_oblock = old_oblock;
1034         mg->new_oblock = new_oblock;
1035         mg->cblock = cblock;
1036         mg->old_ocell = old_ocell;
1037         mg->new_ocell = new_ocell;
1038         mg->start_jiffies = jiffies;
1039
1040         inc_nr_migrations(cache);
1041         quiesce_migration(mg);
1042 }
1043
1044 /*----------------------------------------------------------------
1045  * bio processing
1046  *--------------------------------------------------------------*/
1047 static void defer_bio(struct cache *cache, struct bio *bio)
1048 {
1049         unsigned long flags;
1050
1051         spin_lock_irqsave(&cache->lock, flags);
1052         bio_list_add(&cache->deferred_bios, bio);
1053         spin_unlock_irqrestore(&cache->lock, flags);
1054
1055         wake_worker(cache);
1056 }
1057
1058 static void process_flush_bio(struct cache *cache, struct bio *bio)
1059 {
1060         size_t pb_data_size = get_per_bio_data_size(cache);
1061         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1062
1063         BUG_ON(bio->bi_size);
1064         if (!pb->req_nr)
1065                 remap_to_origin(cache, bio);
1066         else
1067                 remap_to_cache(cache, bio, 0);
1068
1069         issue(cache, bio);
1070 }
1071
1072 /*
1073  * People generally discard large parts of a device, eg, the whole device
1074  * when formatting.  Splitting these large discards up into cache block
1075  * sized ios and then quiescing (always neccessary for discard) takes too
1076  * long.
1077  *
1078  * We keep it simple, and allow any size of discard to come in, and just
1079  * mark off blocks on the discard bitset.  No passdown occurs!
1080  *
1081  * To implement passdown we need to change the bio_prison such that a cell
1082  * can have a key that spans many blocks.
1083  */
1084 static void process_discard_bio(struct cache *cache, struct bio *bio)
1085 {
1086         dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
1087                                                   cache->discard_block_size);
1088         dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
1089         dm_block_t b;
1090
1091         end_block = block_div(end_block, cache->discard_block_size);
1092
1093         for (b = start_block; b < end_block; b++)
1094                 set_discard(cache, to_dblock(b));
1095
1096         bio_endio(bio, 0);
1097 }
1098
1099 static bool spare_migration_bandwidth(struct cache *cache)
1100 {
1101         sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1102                 cache->sectors_per_block;
1103         return current_volume < cache->migration_threshold;
1104 }
1105
1106 static bool is_writethrough_io(struct cache *cache, struct bio *bio,
1107                                dm_cblock_t cblock)
1108 {
1109         return bio_data_dir(bio) == WRITE &&
1110                 cache->features.write_through && !is_dirty(cache, cblock);
1111 }
1112
1113 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1114 {
1115         atomic_inc(bio_data_dir(bio) == READ ?
1116                    &cache->stats.read_hit : &cache->stats.write_hit);
1117 }
1118
1119 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1120 {
1121         atomic_inc(bio_data_dir(bio) == READ ?
1122                    &cache->stats.read_miss : &cache->stats.write_miss);
1123 }
1124
1125 static void process_bio(struct cache *cache, struct prealloc *structs,
1126                         struct bio *bio)
1127 {
1128         int r;
1129         bool release_cell = true;
1130         dm_oblock_t block = get_bio_block(cache, bio);
1131         struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1132         struct policy_result lookup_result;
1133         size_t pb_data_size = get_per_bio_data_size(cache);
1134         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1135         bool discarded_block = is_discarded_oblock(cache, block);
1136         bool can_migrate = discarded_block || spare_migration_bandwidth(cache);
1137
1138         /*
1139          * Check to see if that block is currently migrating.
1140          */
1141         cell_prealloc = prealloc_get_cell(structs);
1142         r = bio_detain(cache, block, bio, cell_prealloc,
1143                        (cell_free_fn) prealloc_put_cell,
1144                        structs, &new_ocell);
1145         if (r > 0)
1146                 return;
1147
1148         r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1149                        bio, &lookup_result);
1150
1151         if (r == -EWOULDBLOCK)
1152                 /* migration has been denied */
1153                 lookup_result.op = POLICY_MISS;
1154
1155         switch (lookup_result.op) {
1156         case POLICY_HIT:
1157                 inc_hit_counter(cache, bio);
1158                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1159
1160                 if (is_writethrough_io(cache, bio, lookup_result.cblock))
1161                         remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1162                 else
1163                         remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
1164
1165                 issue(cache, bio);
1166                 break;
1167
1168         case POLICY_MISS:
1169                 inc_miss_counter(cache, bio);
1170                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1171                 remap_to_origin_clear_discard(cache, bio, block);
1172                 issue(cache, bio);
1173                 break;
1174
1175         case POLICY_NEW:
1176                 atomic_inc(&cache->stats.promotion);
1177                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1178                 release_cell = false;
1179                 break;
1180
1181         case POLICY_REPLACE:
1182                 cell_prealloc = prealloc_get_cell(structs);
1183                 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1184                                (cell_free_fn) prealloc_put_cell,
1185                                structs, &old_ocell);
1186                 if (r > 0) {
1187                         /*
1188                          * We have to be careful to avoid lock inversion of
1189                          * the cells.  So we back off, and wait for the
1190                          * old_ocell to become free.
1191                          */
1192                         policy_force_mapping(cache->policy, block,
1193                                              lookup_result.old_oblock);
1194                         atomic_inc(&cache->stats.cache_cell_clash);
1195                         break;
1196                 }
1197                 atomic_inc(&cache->stats.demotion);
1198                 atomic_inc(&cache->stats.promotion);
1199
1200                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1201                                     block, lookup_result.cblock,
1202                                     old_ocell, new_ocell);
1203                 release_cell = false;
1204                 break;
1205
1206         default:
1207                 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1208                             (unsigned) lookup_result.op);
1209                 bio_io_error(bio);
1210         }
1211
1212         if (release_cell)
1213                 cell_defer(cache, new_ocell, false);
1214 }
1215
1216 static int need_commit_due_to_time(struct cache *cache)
1217 {
1218         return jiffies < cache->last_commit_jiffies ||
1219                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1220 }
1221
1222 static int commit_if_needed(struct cache *cache)
1223 {
1224         if (dm_cache_changed_this_transaction(cache->cmd) &&
1225             (cache->commit_requested || need_commit_due_to_time(cache))) {
1226                 atomic_inc(&cache->stats.commit_count);
1227                 cache->last_commit_jiffies = jiffies;
1228                 cache->commit_requested = false;
1229                 return dm_cache_commit(cache->cmd, false);
1230         }
1231
1232         return 0;
1233 }
1234
1235 static void process_deferred_bios(struct cache *cache)
1236 {
1237         unsigned long flags;
1238         struct bio_list bios;
1239         struct bio *bio;
1240         struct prealloc structs;
1241
1242         memset(&structs, 0, sizeof(structs));
1243         bio_list_init(&bios);
1244
1245         spin_lock_irqsave(&cache->lock, flags);
1246         bio_list_merge(&bios, &cache->deferred_bios);
1247         bio_list_init(&cache->deferred_bios);
1248         spin_unlock_irqrestore(&cache->lock, flags);
1249
1250         while (!bio_list_empty(&bios)) {
1251                 /*
1252                  * If we've got no free migration structs, and processing
1253                  * this bio might require one, we pause until there are some
1254                  * prepared mappings to process.
1255                  */
1256                 if (prealloc_data_structs(cache, &structs)) {
1257                         spin_lock_irqsave(&cache->lock, flags);
1258                         bio_list_merge(&cache->deferred_bios, &bios);
1259                         spin_unlock_irqrestore(&cache->lock, flags);
1260                         break;
1261                 }
1262
1263                 bio = bio_list_pop(&bios);
1264
1265                 if (bio->bi_rw & REQ_FLUSH)
1266                         process_flush_bio(cache, bio);
1267                 else if (bio->bi_rw & REQ_DISCARD)
1268                         process_discard_bio(cache, bio);
1269                 else
1270                         process_bio(cache, &structs, bio);
1271         }
1272
1273         prealloc_free_structs(cache, &structs);
1274 }
1275
1276 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1277 {
1278         unsigned long flags;
1279         struct bio_list bios;
1280         struct bio *bio;
1281
1282         bio_list_init(&bios);
1283
1284         spin_lock_irqsave(&cache->lock, flags);
1285         bio_list_merge(&bios, &cache->deferred_flush_bios);
1286         bio_list_init(&cache->deferred_flush_bios);
1287         spin_unlock_irqrestore(&cache->lock, flags);
1288
1289         while ((bio = bio_list_pop(&bios)))
1290                 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1291 }
1292
1293 static void process_deferred_writethrough_bios(struct cache *cache)
1294 {
1295         unsigned long flags;
1296         struct bio_list bios;
1297         struct bio *bio;
1298
1299         bio_list_init(&bios);
1300
1301         spin_lock_irqsave(&cache->lock, flags);
1302         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1303         bio_list_init(&cache->deferred_writethrough_bios);
1304         spin_unlock_irqrestore(&cache->lock, flags);
1305
1306         while ((bio = bio_list_pop(&bios)))
1307                 generic_make_request(bio);
1308 }
1309
1310 static void writeback_some_dirty_blocks(struct cache *cache)
1311 {
1312         int r = 0;
1313         dm_oblock_t oblock;
1314         dm_cblock_t cblock;
1315         struct prealloc structs;
1316         struct dm_bio_prison_cell *old_ocell;
1317
1318         memset(&structs, 0, sizeof(structs));
1319
1320         while (spare_migration_bandwidth(cache)) {
1321                 if (prealloc_data_structs(cache, &structs))
1322                         break;
1323
1324                 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1325                 if (r)
1326                         break;
1327
1328                 r = get_cell(cache, oblock, &structs, &old_ocell);
1329                 if (r) {
1330                         policy_set_dirty(cache->policy, oblock);
1331                         break;
1332                 }
1333
1334                 writeback(cache, &structs, oblock, cblock, old_ocell);
1335         }
1336
1337         prealloc_free_structs(cache, &structs);
1338 }
1339
1340 /*----------------------------------------------------------------
1341  * Main worker loop
1342  *--------------------------------------------------------------*/
1343 static void start_quiescing(struct cache *cache)
1344 {
1345         unsigned long flags;
1346
1347         spin_lock_irqsave(&cache->lock, flags);
1348         cache->quiescing = 1;
1349         spin_unlock_irqrestore(&cache->lock, flags);
1350 }
1351
1352 static void stop_quiescing(struct cache *cache)
1353 {
1354         unsigned long flags;
1355
1356         spin_lock_irqsave(&cache->lock, flags);
1357         cache->quiescing = 0;
1358         spin_unlock_irqrestore(&cache->lock, flags);
1359 }
1360
1361 static bool is_quiescing(struct cache *cache)
1362 {
1363         int r;
1364         unsigned long flags;
1365
1366         spin_lock_irqsave(&cache->lock, flags);
1367         r = cache->quiescing;
1368         spin_unlock_irqrestore(&cache->lock, flags);
1369
1370         return r;
1371 }
1372
1373 static void wait_for_migrations(struct cache *cache)
1374 {
1375         wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1376 }
1377
1378 static void stop_worker(struct cache *cache)
1379 {
1380         cancel_delayed_work(&cache->waker);
1381         flush_workqueue(cache->wq);
1382 }
1383
1384 static void requeue_deferred_io(struct cache *cache)
1385 {
1386         struct bio *bio;
1387         struct bio_list bios;
1388
1389         bio_list_init(&bios);
1390         bio_list_merge(&bios, &cache->deferred_bios);
1391         bio_list_init(&cache->deferred_bios);
1392
1393         while ((bio = bio_list_pop(&bios)))
1394                 bio_endio(bio, DM_ENDIO_REQUEUE);
1395 }
1396
1397 static int more_work(struct cache *cache)
1398 {
1399         if (is_quiescing(cache))
1400                 return !list_empty(&cache->quiesced_migrations) ||
1401                         !list_empty(&cache->completed_migrations) ||
1402                         !list_empty(&cache->need_commit_migrations);
1403         else
1404                 return !bio_list_empty(&cache->deferred_bios) ||
1405                         !bio_list_empty(&cache->deferred_flush_bios) ||
1406                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
1407                         !list_empty(&cache->quiesced_migrations) ||
1408                         !list_empty(&cache->completed_migrations) ||
1409                         !list_empty(&cache->need_commit_migrations);
1410 }
1411
1412 static void do_worker(struct work_struct *ws)
1413 {
1414         struct cache *cache = container_of(ws, struct cache, worker);
1415
1416         do {
1417                 if (!is_quiescing(cache))
1418                         process_deferred_bios(cache);
1419
1420                 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1421                 process_migrations(cache, &cache->completed_migrations, complete_migration);
1422
1423                 writeback_some_dirty_blocks(cache);
1424
1425                 process_deferred_writethrough_bios(cache);
1426
1427                 if (commit_if_needed(cache)) {
1428                         process_deferred_flush_bios(cache, false);
1429
1430                         /*
1431                          * FIXME: rollback metadata or just go into a
1432                          * failure mode and error everything
1433                          */
1434                 } else {
1435                         process_deferred_flush_bios(cache, true);
1436                         process_migrations(cache, &cache->need_commit_migrations,
1437                                            migration_success_post_commit);
1438                 }
1439         } while (more_work(cache));
1440 }
1441
1442 /*
1443  * We want to commit periodically so that not too much
1444  * unwritten metadata builds up.
1445  */
1446 static void do_waker(struct work_struct *ws)
1447 {
1448         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1449         policy_tick(cache->policy);
1450         wake_worker(cache);
1451         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1452 }
1453
1454 /*----------------------------------------------------------------*/
1455
1456 static int is_congested(struct dm_dev *dev, int bdi_bits)
1457 {
1458         struct request_queue *q = bdev_get_queue(dev->bdev);
1459         return bdi_congested(&q->backing_dev_info, bdi_bits);
1460 }
1461
1462 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1463 {
1464         struct cache *cache = container_of(cb, struct cache, callbacks);
1465
1466         return is_congested(cache->origin_dev, bdi_bits) ||
1467                 is_congested(cache->cache_dev, bdi_bits);
1468 }
1469
1470 /*----------------------------------------------------------------
1471  * Target methods
1472  *--------------------------------------------------------------*/
1473
1474 /*
1475  * This function gets called on the error paths of the constructor, so we
1476  * have to cope with a partially initialised struct.
1477  */
1478 static void destroy(struct cache *cache)
1479 {
1480         unsigned i;
1481
1482         if (cache->next_migration)
1483                 mempool_free(cache->next_migration, cache->migration_pool);
1484
1485         if (cache->migration_pool)
1486                 mempool_destroy(cache->migration_pool);
1487
1488         if (cache->all_io_ds)
1489                 dm_deferred_set_destroy(cache->all_io_ds);
1490
1491         if (cache->prison)
1492                 dm_bio_prison_destroy(cache->prison);
1493
1494         if (cache->wq)
1495                 destroy_workqueue(cache->wq);
1496
1497         if (cache->dirty_bitset)
1498                 free_bitset(cache->dirty_bitset);
1499
1500         if (cache->discard_bitset)
1501                 free_bitset(cache->discard_bitset);
1502
1503         if (cache->copier)
1504                 dm_kcopyd_client_destroy(cache->copier);
1505
1506         if (cache->cmd)
1507                 dm_cache_metadata_close(cache->cmd);
1508
1509         if (cache->metadata_dev)
1510                 dm_put_device(cache->ti, cache->metadata_dev);
1511
1512         if (cache->origin_dev)
1513                 dm_put_device(cache->ti, cache->origin_dev);
1514
1515         if (cache->cache_dev)
1516                 dm_put_device(cache->ti, cache->cache_dev);
1517
1518         if (cache->policy)
1519                 dm_cache_policy_destroy(cache->policy);
1520
1521         for (i = 0; i < cache->nr_ctr_args ; i++)
1522                 kfree(cache->ctr_args[i]);
1523         kfree(cache->ctr_args);
1524
1525         kfree(cache);
1526 }
1527
1528 static void cache_dtr(struct dm_target *ti)
1529 {
1530         struct cache *cache = ti->private;
1531
1532         destroy(cache);
1533 }
1534
1535 static sector_t get_dev_size(struct dm_dev *dev)
1536 {
1537         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1538 }
1539
1540 /*----------------------------------------------------------------*/
1541
1542 /*
1543  * Construct a cache device mapping.
1544  *
1545  * cache <metadata dev> <cache dev> <origin dev> <block size>
1546  *       <#feature args> [<feature arg>]*
1547  *       <policy> <#policy args> [<policy arg>]*
1548  *
1549  * metadata dev    : fast device holding the persistent metadata
1550  * cache dev       : fast device holding cached data blocks
1551  * origin dev      : slow device holding original data blocks
1552  * block size      : cache unit size in sectors
1553  *
1554  * #feature args   : number of feature arguments passed
1555  * feature args    : writethrough.  (The default is writeback.)
1556  *
1557  * policy          : the replacement policy to use
1558  * #policy args    : an even number of policy arguments corresponding
1559  *                   to key/value pairs passed to the policy
1560  * policy args     : key/value pairs passed to the policy
1561  *                   E.g. 'sequential_threshold 1024'
1562  *                   See cache-policies.txt for details.
1563  *
1564  * Optional feature arguments are:
1565  *   writethrough  : write through caching that prohibits cache block
1566  *                   content from being different from origin block content.
1567  *                   Without this argument, the default behaviour is to write
1568  *                   back cache block contents later for performance reasons,
1569  *                   so they may differ from the corresponding origin blocks.
1570  */
1571 struct cache_args {
1572         struct dm_target *ti;
1573
1574         struct dm_dev *metadata_dev;
1575
1576         struct dm_dev *cache_dev;
1577         sector_t cache_sectors;
1578
1579         struct dm_dev *origin_dev;
1580         sector_t origin_sectors;
1581
1582         uint32_t block_size;
1583
1584         const char *policy_name;
1585         int policy_argc;
1586         const char **policy_argv;
1587
1588         struct cache_features features;
1589 };
1590
1591 static void destroy_cache_args(struct cache_args *ca)
1592 {
1593         if (ca->metadata_dev)
1594                 dm_put_device(ca->ti, ca->metadata_dev);
1595
1596         if (ca->cache_dev)
1597                 dm_put_device(ca->ti, ca->cache_dev);
1598
1599         if (ca->origin_dev)
1600                 dm_put_device(ca->ti, ca->origin_dev);
1601
1602         kfree(ca);
1603 }
1604
1605 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1606 {
1607         if (!as->argc) {
1608                 *error = "Insufficient args";
1609                 return false;
1610         }
1611
1612         return true;
1613 }
1614
1615 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1616                               char **error)
1617 {
1618         int r;
1619         sector_t metadata_dev_size;
1620         char b[BDEVNAME_SIZE];
1621
1622         if (!at_least_one_arg(as, error))
1623                 return -EINVAL;
1624
1625         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1626                           &ca->metadata_dev);
1627         if (r) {
1628                 *error = "Error opening metadata device";
1629                 return r;
1630         }
1631
1632         metadata_dev_size = get_dev_size(ca->metadata_dev);
1633         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1634                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1635                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1636
1637         return 0;
1638 }
1639
1640 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1641                            char **error)
1642 {
1643         int r;
1644
1645         if (!at_least_one_arg(as, error))
1646                 return -EINVAL;
1647
1648         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1649                           &ca->cache_dev);
1650         if (r) {
1651                 *error = "Error opening cache device";
1652                 return r;
1653         }
1654         ca->cache_sectors = get_dev_size(ca->cache_dev);
1655
1656         return 0;
1657 }
1658
1659 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1660                             char **error)
1661 {
1662         int r;
1663
1664         if (!at_least_one_arg(as, error))
1665                 return -EINVAL;
1666
1667         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1668                           &ca->origin_dev);
1669         if (r) {
1670                 *error = "Error opening origin device";
1671                 return r;
1672         }
1673
1674         ca->origin_sectors = get_dev_size(ca->origin_dev);
1675         if (ca->ti->len > ca->origin_sectors) {
1676                 *error = "Device size larger than cached device";
1677                 return -EINVAL;
1678         }
1679
1680         return 0;
1681 }
1682
1683 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1684                             char **error)
1685 {
1686         unsigned long tmp;
1687
1688         if (!at_least_one_arg(as, error))
1689                 return -EINVAL;
1690
1691         if (kstrtoul(dm_shift_arg(as), 10, &tmp) || !tmp ||
1692             tmp < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1693             tmp & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1694                 *error = "Invalid data block size";
1695                 return -EINVAL;
1696         }
1697
1698         if (tmp > ca->cache_sectors) {
1699                 *error = "Data block size is larger than the cache device";
1700                 return -EINVAL;
1701         }
1702
1703         ca->block_size = tmp;
1704
1705         return 0;
1706 }
1707
1708 static void init_features(struct cache_features *cf)
1709 {
1710         cf->mode = CM_WRITE;
1711         cf->write_through = false;
1712 }
1713
1714 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1715                           char **error)
1716 {
1717         static struct dm_arg _args[] = {
1718                 {0, 1, "Invalid number of cache feature arguments"},
1719         };
1720
1721         int r;
1722         unsigned argc;
1723         const char *arg;
1724         struct cache_features *cf = &ca->features;
1725
1726         init_features(cf);
1727
1728         r = dm_read_arg_group(_args, as, &argc, error);
1729         if (r)
1730                 return -EINVAL;
1731
1732         while (argc--) {
1733                 arg = dm_shift_arg(as);
1734
1735                 if (!strcasecmp(arg, "writeback"))
1736                         cf->write_through = false;
1737
1738                 else if (!strcasecmp(arg, "writethrough"))
1739                         cf->write_through = true;
1740
1741                 else {
1742                         *error = "Unrecognised cache feature requested";
1743                         return -EINVAL;
1744                 }
1745         }
1746
1747         return 0;
1748 }
1749
1750 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
1751                         char **error)
1752 {
1753         static struct dm_arg _args[] = {
1754                 {0, 1024, "Invalid number of policy arguments"},
1755         };
1756
1757         int r;
1758
1759         if (!at_least_one_arg(as, error))
1760                 return -EINVAL;
1761
1762         ca->policy_name = dm_shift_arg(as);
1763
1764         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
1765         if (r)
1766                 return -EINVAL;
1767
1768         ca->policy_argv = (const char **)as->argv;
1769         dm_consume_args(as, ca->policy_argc);
1770
1771         return 0;
1772 }
1773
1774 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
1775                             char **error)
1776 {
1777         int r;
1778         struct dm_arg_set as;
1779
1780         as.argc = argc;
1781         as.argv = argv;
1782
1783         r = parse_metadata_dev(ca, &as, error);
1784         if (r)
1785                 return r;
1786
1787         r = parse_cache_dev(ca, &as, error);
1788         if (r)
1789                 return r;
1790
1791         r = parse_origin_dev(ca, &as, error);
1792         if (r)
1793                 return r;
1794
1795         r = parse_block_size(ca, &as, error);
1796         if (r)
1797                 return r;
1798
1799         r = parse_features(ca, &as, error);
1800         if (r)
1801                 return r;
1802
1803         r = parse_policy(ca, &as, error);
1804         if (r)
1805                 return r;
1806
1807         return 0;
1808 }
1809
1810 /*----------------------------------------------------------------*/
1811
1812 static struct kmem_cache *migration_cache;
1813
1814 #define NOT_CORE_OPTION 1
1815
1816 static int process_config_option(struct cache *cache, const char *key, const char *value)
1817 {
1818         unsigned long tmp;
1819
1820         if (!strcasecmp(key, "migration_threshold")) {
1821                 if (kstrtoul(value, 10, &tmp))
1822                         return -EINVAL;
1823
1824                 cache->migration_threshold = tmp;
1825                 return 0;
1826         }
1827
1828         return NOT_CORE_OPTION;
1829 }
1830
1831 static int set_config_value(struct cache *cache, const char *key, const char *value)
1832 {
1833         int r = process_config_option(cache, key, value);
1834
1835         if (r == NOT_CORE_OPTION)
1836                 r = policy_set_config_value(cache->policy, key, value);
1837
1838         if (r)
1839                 DMWARN("bad config value for %s: %s", key, value);
1840
1841         return r;
1842 }
1843
1844 static int set_config_values(struct cache *cache, int argc, const char **argv)
1845 {
1846         int r = 0;
1847
1848         if (argc & 1) {
1849                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
1850                 return -EINVAL;
1851         }
1852
1853         while (argc) {
1854                 r = set_config_value(cache, argv[0], argv[1]);
1855                 if (r)
1856                         break;
1857
1858                 argc -= 2;
1859                 argv += 2;
1860         }
1861
1862         return r;
1863 }
1864
1865 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
1866                                char **error)
1867 {
1868         cache->policy = dm_cache_policy_create(ca->policy_name,
1869                                                cache->cache_size,
1870                                                cache->origin_sectors,
1871                                                cache->sectors_per_block);
1872         if (!cache->policy) {
1873                 *error = "Error creating cache's policy";
1874                 return -ENOMEM;
1875         }
1876
1877         return 0;
1878 }
1879
1880 /*
1881  * We want the discard block size to be a power of two, at least the size
1882  * of the cache block size, and have no more than 2^14 discard blocks
1883  * across the origin.
1884  */
1885 #define MAX_DISCARD_BLOCKS (1 << 14)
1886
1887 static bool too_many_discard_blocks(sector_t discard_block_size,
1888                                     sector_t origin_size)
1889 {
1890         (void) sector_div(origin_size, discard_block_size);
1891
1892         return origin_size > MAX_DISCARD_BLOCKS;
1893 }
1894
1895 static sector_t calculate_discard_block_size(sector_t cache_block_size,
1896                                              sector_t origin_size)
1897 {
1898         sector_t discard_block_size;
1899
1900         discard_block_size = roundup_pow_of_two(cache_block_size);
1901
1902         if (origin_size)
1903                 while (too_many_discard_blocks(discard_block_size, origin_size))
1904                         discard_block_size *= 2;
1905
1906         return discard_block_size;
1907 }
1908
1909 #define DEFAULT_MIGRATION_THRESHOLD 2048
1910
1911 static int cache_create(struct cache_args *ca, struct cache **result)
1912 {
1913         int r = 0;
1914         char **error = &ca->ti->error;
1915         struct cache *cache;
1916         struct dm_target *ti = ca->ti;
1917         dm_block_t origin_blocks;
1918         struct dm_cache_metadata *cmd;
1919         bool may_format = ca->features.mode == CM_WRITE;
1920
1921         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
1922         if (!cache)
1923                 return -ENOMEM;
1924
1925         cache->ti = ca->ti;
1926         ti->private = cache;
1927         ti->num_flush_bios = 2;
1928         ti->flush_supported = true;
1929
1930         ti->num_discard_bios = 1;
1931         ti->discards_supported = true;
1932         ti->discard_zeroes_data_unsupported = true;
1933
1934         cache->features = ca->features;
1935         ti->per_bio_data_size = get_per_bio_data_size(cache);
1936
1937         cache->callbacks.congested_fn = cache_is_congested;
1938         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
1939
1940         cache->metadata_dev = ca->metadata_dev;
1941         cache->origin_dev = ca->origin_dev;
1942         cache->cache_dev = ca->cache_dev;
1943
1944         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
1945
1946         /* FIXME: factor out this whole section */
1947         origin_blocks = cache->origin_sectors = ca->origin_sectors;
1948         origin_blocks = block_div(origin_blocks, ca->block_size);
1949         cache->origin_blocks = to_oblock(origin_blocks);
1950
1951         cache->sectors_per_block = ca->block_size;
1952         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
1953                 r = -EINVAL;
1954                 goto bad;
1955         }
1956
1957         if (ca->block_size & (ca->block_size - 1)) {
1958                 dm_block_t cache_size = ca->cache_sectors;
1959
1960                 cache->sectors_per_block_shift = -1;
1961                 cache_size = block_div(cache_size, ca->block_size);
1962                 cache->cache_size = to_cblock(cache_size);
1963         } else {
1964                 cache->sectors_per_block_shift = __ffs(ca->block_size);
1965                 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
1966         }
1967
1968         r = create_cache_policy(cache, ca, error);
1969         if (r)
1970                 goto bad;
1971
1972         cache->policy_nr_args = ca->policy_argc;
1973         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
1974
1975         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
1976         if (r) {
1977                 *error = "Error setting cache policy's config values";
1978                 goto bad;
1979         }
1980
1981         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
1982                                      ca->block_size, may_format,
1983                                      dm_cache_policy_get_hint_size(cache->policy));
1984         if (IS_ERR(cmd)) {
1985                 *error = "Error creating metadata object";
1986                 r = PTR_ERR(cmd);
1987                 goto bad;
1988         }
1989         cache->cmd = cmd;
1990
1991         spin_lock_init(&cache->lock);
1992         bio_list_init(&cache->deferred_bios);
1993         bio_list_init(&cache->deferred_flush_bios);
1994         bio_list_init(&cache->deferred_writethrough_bios);
1995         INIT_LIST_HEAD(&cache->quiesced_migrations);
1996         INIT_LIST_HEAD(&cache->completed_migrations);
1997         INIT_LIST_HEAD(&cache->need_commit_migrations);
1998         atomic_set(&cache->nr_migrations, 0);
1999         init_waitqueue_head(&cache->migration_wait);
2000
2001         r = -ENOMEM;
2002         cache->nr_dirty = 0;
2003         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2004         if (!cache->dirty_bitset) {
2005                 *error = "could not allocate dirty bitset";
2006                 goto bad;
2007         }
2008         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2009
2010         cache->discard_block_size =
2011                 calculate_discard_block_size(cache->sectors_per_block,
2012                                              cache->origin_sectors);
2013         cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2014         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2015         if (!cache->discard_bitset) {
2016                 *error = "could not allocate discard bitset";
2017                 goto bad;
2018         }
2019         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2020
2021         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2022         if (IS_ERR(cache->copier)) {
2023                 *error = "could not create kcopyd client";
2024                 r = PTR_ERR(cache->copier);
2025                 goto bad;
2026         }
2027
2028         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2029         if (!cache->wq) {
2030                 *error = "could not create workqueue for metadata object";
2031                 goto bad;
2032         }
2033         INIT_WORK(&cache->worker, do_worker);
2034         INIT_DELAYED_WORK(&cache->waker, do_waker);
2035         cache->last_commit_jiffies = jiffies;
2036
2037         cache->prison = dm_bio_prison_create(PRISON_CELLS);
2038         if (!cache->prison) {
2039                 *error = "could not create bio prison";
2040                 goto bad;
2041         }
2042
2043         cache->all_io_ds = dm_deferred_set_create();
2044         if (!cache->all_io_ds) {
2045                 *error = "could not create all_io deferred set";
2046                 goto bad;
2047         }
2048
2049         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2050                                                          migration_cache);
2051         if (!cache->migration_pool) {
2052                 *error = "Error creating cache's migration mempool";
2053                 goto bad;
2054         }
2055
2056         cache->next_migration = NULL;
2057
2058         cache->need_tick_bio = true;
2059         cache->sized = false;
2060         cache->quiescing = false;
2061         cache->commit_requested = false;
2062         cache->loaded_mappings = false;
2063         cache->loaded_discards = false;
2064
2065         load_stats(cache);
2066
2067         atomic_set(&cache->stats.demotion, 0);
2068         atomic_set(&cache->stats.promotion, 0);
2069         atomic_set(&cache->stats.copies_avoided, 0);
2070         atomic_set(&cache->stats.cache_cell_clash, 0);
2071         atomic_set(&cache->stats.commit_count, 0);
2072         atomic_set(&cache->stats.discard_count, 0);
2073
2074         *result = cache;
2075         return 0;
2076
2077 bad:
2078         destroy(cache);
2079         return r;
2080 }
2081
2082 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2083 {
2084         unsigned i;
2085         const char **copy;
2086
2087         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2088         if (!copy)
2089                 return -ENOMEM;
2090         for (i = 0; i < argc; i++) {
2091                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2092                 if (!copy[i]) {
2093                         while (i--)
2094                                 kfree(copy[i]);
2095                         kfree(copy);
2096                         return -ENOMEM;
2097                 }
2098         }
2099
2100         cache->nr_ctr_args = argc;
2101         cache->ctr_args = copy;
2102
2103         return 0;
2104 }
2105
2106 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2107 {
2108         int r = -EINVAL;
2109         struct cache_args *ca;
2110         struct cache *cache = NULL;
2111
2112         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2113         if (!ca) {
2114                 ti->error = "Error allocating memory for cache";
2115                 return -ENOMEM;
2116         }
2117         ca->ti = ti;
2118
2119         r = parse_cache_args(ca, argc, argv, &ti->error);
2120         if (r)
2121                 goto out;
2122
2123         r = cache_create(ca, &cache);
2124         if (r)
2125                 goto out;
2126
2127         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2128         if (r) {
2129                 destroy(cache);
2130                 goto out;
2131         }
2132
2133         ti->private = cache;
2134
2135 out:
2136         destroy_cache_args(ca);
2137         return r;
2138 }
2139
2140 static int cache_map(struct dm_target *ti, struct bio *bio)
2141 {
2142         struct cache *cache = ti->private;
2143
2144         int r;
2145         dm_oblock_t block = get_bio_block(cache, bio);
2146         size_t pb_data_size = get_per_bio_data_size(cache);
2147         bool can_migrate = false;
2148         bool discarded_block;
2149         struct dm_bio_prison_cell *cell;
2150         struct policy_result lookup_result;
2151         struct per_bio_data *pb;
2152
2153         if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2154                 /*
2155                  * This can only occur if the io goes to a partial block at
2156                  * the end of the origin device.  We don't cache these.
2157                  * Just remap to the origin and carry on.
2158                  */
2159                 remap_to_origin_clear_discard(cache, bio, block);
2160                 return DM_MAPIO_REMAPPED;
2161         }
2162
2163         pb = init_per_bio_data(bio, pb_data_size);
2164
2165         if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2166                 defer_bio(cache, bio);
2167                 return DM_MAPIO_SUBMITTED;
2168         }
2169
2170         /*
2171          * Check to see if that block is currently migrating.
2172          */
2173         cell = alloc_prison_cell(cache);
2174         if (!cell) {
2175                 defer_bio(cache, bio);
2176                 return DM_MAPIO_SUBMITTED;
2177         }
2178
2179         r = bio_detain(cache, block, bio, cell,
2180                        (cell_free_fn) free_prison_cell,
2181                        cache, &cell);
2182         if (r) {
2183                 if (r < 0)
2184                         defer_bio(cache, bio);
2185
2186                 return DM_MAPIO_SUBMITTED;
2187         }
2188
2189         discarded_block = is_discarded_oblock(cache, block);
2190
2191         r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2192                        bio, &lookup_result);
2193         if (r == -EWOULDBLOCK) {
2194                 cell_defer(cache, cell, true);
2195                 return DM_MAPIO_SUBMITTED;
2196
2197         } else if (r) {
2198                 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2199                 bio_io_error(bio);
2200                 return DM_MAPIO_SUBMITTED;
2201         }
2202
2203         switch (lookup_result.op) {
2204         case POLICY_HIT:
2205                 inc_hit_counter(cache, bio);
2206                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2207
2208                 if (is_writethrough_io(cache, bio, lookup_result.cblock))
2209                         remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2210                 else
2211                         remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2212
2213                 cell_defer(cache, cell, false);
2214                 break;
2215
2216         case POLICY_MISS:
2217                 inc_miss_counter(cache, bio);
2218                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2219
2220                 if (pb->req_nr != 0) {
2221                         /*
2222                          * This is a duplicate writethrough io that is no
2223                          * longer needed because the block has been demoted.
2224                          */
2225                         bio_endio(bio, 0);
2226                         cell_defer(cache, cell, false);
2227                         return DM_MAPIO_SUBMITTED;
2228                 } else {
2229                         remap_to_origin_clear_discard(cache, bio, block);
2230                         cell_defer(cache, cell, false);
2231                 }
2232                 break;
2233
2234         default:
2235                 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2236                             (unsigned) lookup_result.op);
2237                 bio_io_error(bio);
2238                 return DM_MAPIO_SUBMITTED;
2239         }
2240
2241         return DM_MAPIO_REMAPPED;
2242 }
2243
2244 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2245 {
2246         struct cache *cache = ti->private;
2247         unsigned long flags;
2248         size_t pb_data_size = get_per_bio_data_size(cache);
2249         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2250
2251         if (pb->tick) {
2252                 policy_tick(cache->policy);
2253
2254                 spin_lock_irqsave(&cache->lock, flags);
2255                 cache->need_tick_bio = true;
2256                 spin_unlock_irqrestore(&cache->lock, flags);
2257         }
2258
2259         check_for_quiesced_migrations(cache, pb);
2260
2261         return 0;
2262 }
2263
2264 static int write_dirty_bitset(struct cache *cache)
2265 {
2266         unsigned i, r;
2267
2268         for (i = 0; i < from_cblock(cache->cache_size); i++) {
2269                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2270                                        is_dirty(cache, to_cblock(i)));
2271                 if (r)
2272                         return r;
2273         }
2274
2275         return 0;
2276 }
2277
2278 static int write_discard_bitset(struct cache *cache)
2279 {
2280         unsigned i, r;
2281
2282         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2283                                            cache->discard_nr_blocks);
2284         if (r) {
2285                 DMERR("could not resize on-disk discard bitset");
2286                 return r;
2287         }
2288
2289         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2290                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2291                                          is_discarded(cache, to_dblock(i)));
2292                 if (r)
2293                         return r;
2294         }
2295
2296         return 0;
2297 }
2298
2299 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2300                      uint32_t hint)
2301 {
2302         struct cache *cache = context;
2303         return dm_cache_save_hint(cache->cmd, cblock, hint);
2304 }
2305
2306 static int write_hints(struct cache *cache)
2307 {
2308         int r;
2309
2310         r = dm_cache_begin_hints(cache->cmd, cache->policy);
2311         if (r) {
2312                 DMERR("dm_cache_begin_hints failed");
2313                 return r;
2314         }
2315
2316         r = policy_walk_mappings(cache->policy, save_hint, cache);
2317         if (r)
2318                 DMERR("policy_walk_mappings failed");
2319
2320         return r;
2321 }
2322
2323 /*
2324  * returns true on success
2325  */
2326 static bool sync_metadata(struct cache *cache)
2327 {
2328         int r1, r2, r3, r4;
2329
2330         r1 = write_dirty_bitset(cache);
2331         if (r1)
2332                 DMERR("could not write dirty bitset");
2333
2334         r2 = write_discard_bitset(cache);
2335         if (r2)
2336                 DMERR("could not write discard bitset");
2337
2338         save_stats(cache);
2339
2340         r3 = write_hints(cache);
2341         if (r3)
2342                 DMERR("could not write hints");
2343
2344         /*
2345          * If writing the above metadata failed, we still commit, but don't
2346          * set the clean shutdown flag.  This will effectively force every
2347          * dirty bit to be set on reload.
2348          */
2349         r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2350         if (r4)
2351                 DMERR("could not write cache metadata.  Data loss may occur.");
2352
2353         return !r1 && !r2 && !r3 && !r4;
2354 }
2355
2356 static void cache_postsuspend(struct dm_target *ti)
2357 {
2358         struct cache *cache = ti->private;
2359
2360         start_quiescing(cache);
2361         wait_for_migrations(cache);
2362         stop_worker(cache);
2363         requeue_deferred_io(cache);
2364         stop_quiescing(cache);
2365
2366         (void) sync_metadata(cache);
2367 }
2368
2369 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2370                         bool dirty, uint32_t hint, bool hint_valid)
2371 {
2372         int r;
2373         struct cache *cache = context;
2374
2375         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2376         if (r)
2377                 return r;
2378
2379         if (dirty)
2380                 set_dirty(cache, oblock, cblock);
2381         else
2382                 clear_dirty(cache, oblock, cblock);
2383
2384         return 0;
2385 }
2386
2387 static int load_discard(void *context, sector_t discard_block_size,
2388                         dm_dblock_t dblock, bool discard)
2389 {
2390         struct cache *cache = context;
2391
2392         /* FIXME: handle mis-matched block size */
2393
2394         if (discard)
2395                 set_discard(cache, dblock);
2396         else
2397                 clear_discard(cache, dblock);
2398
2399         return 0;
2400 }
2401
2402 static int cache_preresume(struct dm_target *ti)
2403 {
2404         int r = 0;
2405         struct cache *cache = ti->private;
2406         sector_t actual_cache_size = get_dev_size(cache->cache_dev);
2407         (void) sector_div(actual_cache_size, cache->sectors_per_block);
2408
2409         /*
2410          * Check to see if the cache has resized.
2411          */
2412         if (from_cblock(cache->cache_size) != actual_cache_size || !cache->sized) {
2413                 cache->cache_size = to_cblock(actual_cache_size);
2414
2415                 r = dm_cache_resize(cache->cmd, cache->cache_size);
2416                 if (r) {
2417                         DMERR("could not resize cache metadata");
2418                         return r;
2419                 }
2420
2421                 cache->sized = true;
2422         }
2423
2424         if (!cache->loaded_mappings) {
2425                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2426                                            load_mapping, cache);
2427                 if (r) {
2428                         DMERR("could not load cache mappings");
2429                         return r;
2430                 }
2431
2432                 cache->loaded_mappings = true;
2433         }
2434
2435         if (!cache->loaded_discards) {
2436                 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2437                 if (r) {
2438                         DMERR("could not load origin discards");
2439                         return r;
2440                 }
2441
2442                 cache->loaded_discards = true;
2443         }
2444
2445         return r;
2446 }
2447
2448 static void cache_resume(struct dm_target *ti)
2449 {
2450         struct cache *cache = ti->private;
2451
2452         cache->need_tick_bio = true;
2453         do_waker(&cache->waker.work);
2454 }
2455
2456 /*
2457  * Status format:
2458  *
2459  * <#used metadata blocks>/<#total metadata blocks>
2460  * <#read hits> <#read misses> <#write hits> <#write misses>
2461  * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2462  * <#features> <features>*
2463  * <#core args> <core args>
2464  * <#policy args> <policy args>*
2465  */
2466 static void cache_status(struct dm_target *ti, status_type_t type,
2467                          unsigned status_flags, char *result, unsigned maxlen)
2468 {
2469         int r = 0;
2470         unsigned i;
2471         ssize_t sz = 0;
2472         dm_block_t nr_free_blocks_metadata = 0;
2473         dm_block_t nr_blocks_metadata = 0;
2474         char buf[BDEVNAME_SIZE];
2475         struct cache *cache = ti->private;
2476         dm_cblock_t residency;
2477
2478         switch (type) {
2479         case STATUSTYPE_INFO:
2480                 /* Commit to ensure statistics aren't out-of-date */
2481                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2482                         r = dm_cache_commit(cache->cmd, false);
2483                         if (r)
2484                                 DMERR("could not commit metadata for accurate status");
2485                 }
2486
2487                 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2488                                                            &nr_free_blocks_metadata);
2489                 if (r) {
2490                         DMERR("could not get metadata free block count");
2491                         goto err;
2492                 }
2493
2494                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2495                 if (r) {
2496                         DMERR("could not get metadata device size");
2497                         goto err;
2498                 }
2499
2500                 residency = policy_residency(cache->policy);
2501
2502                 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2503                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2504                        (unsigned long long)nr_blocks_metadata,
2505                        (unsigned) atomic_read(&cache->stats.read_hit),
2506                        (unsigned) atomic_read(&cache->stats.read_miss),
2507                        (unsigned) atomic_read(&cache->stats.write_hit),
2508                        (unsigned) atomic_read(&cache->stats.write_miss),
2509                        (unsigned) atomic_read(&cache->stats.demotion),
2510                        (unsigned) atomic_read(&cache->stats.promotion),
2511                        (unsigned long long) from_cblock(residency),
2512                        cache->nr_dirty);
2513
2514                 if (cache->features.write_through)
2515                         DMEMIT("1 writethrough ");
2516                 else
2517                         DMEMIT("0 ");
2518
2519                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2520                 if (sz < maxlen) {
2521                         r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2522                         if (r)
2523                                 DMERR("policy_emit_config_values returned %d", r);
2524                 }
2525
2526                 break;
2527
2528         case STATUSTYPE_TABLE:
2529                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2530                 DMEMIT("%s ", buf);
2531                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2532                 DMEMIT("%s ", buf);
2533                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2534                 DMEMIT("%s", buf);
2535
2536                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2537                         DMEMIT(" %s", cache->ctr_args[i]);
2538                 if (cache->nr_ctr_args)
2539                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2540         }
2541
2542         return;
2543
2544 err:
2545         DMEMIT("Error");
2546 }
2547
2548 /*
2549  * Supports <key> <value>.
2550  *
2551  * The key migration_threshold is supported by the cache target core.
2552  */
2553 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2554 {
2555         struct cache *cache = ti->private;
2556
2557         if (argc != 2)
2558                 return -EINVAL;
2559
2560         return set_config_value(cache, argv[0], argv[1]);
2561 }
2562
2563 static int cache_iterate_devices(struct dm_target *ti,
2564                                  iterate_devices_callout_fn fn, void *data)
2565 {
2566         int r = 0;
2567         struct cache *cache = ti->private;
2568
2569         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
2570         if (!r)
2571                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
2572
2573         return r;
2574 }
2575
2576 /*
2577  * We assume I/O is going to the origin (which is the volume
2578  * more likely to have restrictions e.g. by being striped).
2579  * (Looking up the exact location of the data would be expensive
2580  * and could always be out of date by the time the bio is submitted.)
2581  */
2582 static int cache_bvec_merge(struct dm_target *ti,
2583                             struct bvec_merge_data *bvm,
2584                             struct bio_vec *biovec, int max_size)
2585 {
2586         struct cache *cache = ti->private;
2587         struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
2588
2589         if (!q->merge_bvec_fn)
2590                 return max_size;
2591
2592         bvm->bi_bdev = cache->origin_dev->bdev;
2593         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2594 }
2595
2596 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
2597 {
2598         /*
2599          * FIXME: these limits may be incompatible with the cache device
2600          */
2601         limits->max_discard_sectors = cache->discard_block_size * 1024;
2602         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
2603 }
2604
2605 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
2606 {
2607         struct cache *cache = ti->private;
2608
2609         blk_limits_io_min(limits, 0);
2610         blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
2611         set_discard_limits(cache, limits);
2612 }
2613
2614 /*----------------------------------------------------------------*/
2615
2616 static struct target_type cache_target = {
2617         .name = "cache",
2618         .version = {1, 1, 1},
2619         .module = THIS_MODULE,
2620         .ctr = cache_ctr,
2621         .dtr = cache_dtr,
2622         .map = cache_map,
2623         .end_io = cache_end_io,
2624         .postsuspend = cache_postsuspend,
2625         .preresume = cache_preresume,
2626         .resume = cache_resume,
2627         .status = cache_status,
2628         .message = cache_message,
2629         .iterate_devices = cache_iterate_devices,
2630         .merge = cache_bvec_merge,
2631         .io_hints = cache_io_hints,
2632 };
2633
2634 static int __init dm_cache_init(void)
2635 {
2636         int r;
2637
2638         r = dm_register_target(&cache_target);
2639         if (r) {
2640                 DMERR("cache target registration failed: %d", r);
2641                 return r;
2642         }
2643
2644         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
2645         if (!migration_cache) {
2646                 dm_unregister_target(&cache_target);
2647                 return -ENOMEM;
2648         }
2649
2650         return 0;
2651 }
2652
2653 static void __exit dm_cache_exit(void)
2654 {
2655         dm_unregister_target(&cache_target);
2656         kmem_cache_destroy(migration_cache);
2657 }
2658
2659 module_init(dm_cache_init);
2660 module_exit(dm_cache_exit);
2661
2662 MODULE_DESCRIPTION(DM_NAME " cache target");
2663 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
2664 MODULE_LICENSE("GPL");