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