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 /*
  65  * There are a couple of places where we let a bio run, but want to do some
  66  * work before calling its endio function.  We do this by temporarily
  67  * changing the endio fn.
  68  */
  69 struct dm_hook_info {
  70         bio_end_io_t *bi_end_io;
  71         void *bi_private;
  72 };
  73
  74 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
  75                         bio_end_io_t *bi_end_io, void *bi_private)
  76 {
  77         h->bi_end_io = bio->bi_end_io;
  78         h->bi_private = bio->bi_private;
  79
  80         bio->bi_end_io = bi_end_io;
  81         bio->bi_private = bi_private;
  82 }
  83
  84 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
  85 {
  86         bio->bi_end_io = h->bi_end_io;
  87         bio->bi_private = h->bi_private;
  88 }
  89
  90 /*----------------------------------------------------------------*/
  91
  92 #define PRISON_CELLS 1024
  93 #define MIGRATION_POOL_SIZE 128
  94 #define COMMIT_PERIOD HZ
  95 #define MIGRATION_COUNT_WINDOW 10
  96
  97 /*
  98  * The block size of the device holding cache data must be
  99  * between 32KB and 1GB.
 100  */
 101 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 102 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 103
 104 /*
 105  * FIXME: the cache is read/write for the time being.
 106  */
 107 enum cache_metadata_mode {
 108         CM_WRITE,               /* metadata may be changed */
 109         CM_READ_ONLY,           /* metadata may not be changed */
 110 };
 111
 112 enum cache_io_mode {
 113         /*
 114          * Data is written to cached blocks only.  These blocks are marked
 115          * dirty.  If you lose the cache device you will lose data.
 116          * Potential performance increase for both reads and writes.
 117          */
 118         CM_IO_WRITEBACK,
 119
 120         /*
 121          * Data is written to both cache and origin.  Blocks are never
 122          * dirty.  Potential performance benfit for reads only.
 123          */
 124         CM_IO_WRITETHROUGH,
 125
 126         /*
 127          * A degraded mode useful for various cache coherency situations
 128          * (eg, rolling back snapshots).  Reads and writes always go to the
 129          * origin.  If a write goes to a cached oblock, then the cache
 130          * block is invalidated.
 131          */
 132         CM_IO_PASSTHROUGH
 133 };
 134
 135 struct cache_features {
 136         enum cache_metadata_mode mode;
 137         enum cache_io_mode io_mode;
 138 };
 139
 140 struct cache_stats {
 141         atomic_t read_hit;
 142         atomic_t read_miss;
 143         atomic_t write_hit;
 144         atomic_t write_miss;
 145         atomic_t demotion;
 146         atomic_t promotion;
 147         atomic_t copies_avoided;
 148         atomic_t cache_cell_clash;
 149         atomic_t commit_count;
 150         atomic_t discard_count;
 151 };
 152
 153 struct cache {
 154         struct dm_target *ti;
 155         struct dm_target_callbacks callbacks;
 156
 157         struct dm_cache_metadata *cmd;
 158
 159         /*
 160          * Metadata is written to this device.
 161          */
 162         struct dm_dev *metadata_dev;
 163
 164         /*
 165          * The slower of the two data devices.  Typically a spindle.
 166          */
 167         struct dm_dev *origin_dev;
 168
 169         /*
 170          * The faster of the two data devices.  Typically an SSD.
 171          */
 172         struct dm_dev *cache_dev;
 173
 174         /*
 175          * Size of the origin device in _complete_ blocks and native sectors.
 176          */
 177         dm_oblock_t origin_blocks;
 178         sector_t origin_sectors;
 179
 180         /*
 181          * Size of the cache device in blocks.
 182          */
 183         dm_cblock_t cache_size;
 184
 185         /*
 186          * Fields for converting from sectors to blocks.
 187          */
 188         uint32_t sectors_per_block;
 189         int sectors_per_block_shift;
 190
 191         spinlock_t lock;
 192         struct bio_list deferred_bios;
 193         struct bio_list deferred_flush_bios;
 194         struct bio_list deferred_writethrough_bios;
 195         struct list_head quiesced_migrations;
 196         struct list_head completed_migrations;
 197         struct list_head need_commit_migrations;
 198         sector_t migration_threshold;
 199         wait_queue_head_t migration_wait;
 200         atomic_t nr_migrations;
 201
 202         wait_queue_head_t quiescing_wait;
 203         atomic_t quiescing;
 204         atomic_t quiescing_ack;
 205
 206         /*
 207          * cache_size entries, dirty if set
 208          */
 209         dm_cblock_t nr_dirty;
 210         unsigned long *dirty_bitset;
 211
 212         /*
 213          * origin_blocks entries, discarded if set.
 214          */
 215         dm_dblock_t discard_nr_blocks;
 216         unsigned long *discard_bitset;
 217         uint32_t discard_block_size; /* a power of 2 times sectors per block */
 218
 219         /*
 220          * Rather than reconstructing the table line for the status we just
 221          * save it and regurgitate.
 222          */
 223         unsigned nr_ctr_args;
 224         const char **ctr_args;
 225
 226         struct dm_kcopyd_client *copier;
 227         struct workqueue_struct *wq;
 228         struct work_struct worker;
 229
 230         struct delayed_work waker;
 231         unsigned long last_commit_jiffies;
 232
 233         struct dm_bio_prison *prison;
 234         struct dm_deferred_set *all_io_ds;
 235
 236         mempool_t *migration_pool;
 237         struct dm_cache_migration *next_migration;
 238
 239         struct dm_cache_policy *policy;
 240         unsigned policy_nr_args;
 241
 242         bool need_tick_bio:1;
 243         bool sized:1;
 244         bool commit_requested:1;
 245         bool loaded_mappings:1;
 246         bool loaded_discards:1;
 247
 248         /*
 249          * Cache features such as write-through.
 250          */
 251         struct cache_features features;
 252
 253         struct cache_stats stats;
 254 };
 255
 256 struct per_bio_data {
 257         bool tick:1;
 258         unsigned req_nr:2;
 259         struct dm_deferred_entry *all_io_entry;
 260
 261         /*
 262          * writethrough fields.  These MUST remain at the end of this
 263          * structure and the 'cache' member must be the first as it
 264          * is used to determine the offset of the writethrough fields.
 265          */
 266         struct cache *cache;
 267         dm_cblock_t cblock;
 268         struct dm_hook_info hook_info;
 269         struct dm_bio_details bio_details;
 270 };
 271
 272 struct dm_cache_migration {
 273         struct list_head list;
 274         struct cache *cache;
 275
 276         unsigned long start_jiffies;
 277         dm_oblock_t old_oblock;
 278         dm_oblock_t new_oblock;
 279         dm_cblock_t cblock;
 280
 281         bool err:1;
 282         bool writeback:1;
 283         bool demote:1;
 284         bool promote:1;
 285         bool requeue_holder:1;
 286
 287         struct dm_bio_prison_cell *old_ocell;
 288         struct dm_bio_prison_cell *new_ocell;
 289 };
 290
 291 /*
 292  * Processing a bio in the worker thread may require these memory
 293  * allocations.  We prealloc to avoid deadlocks (the same worker thread
 294  * frees them back to the mempool).
 295  */
 296 struct prealloc {
 297         struct dm_cache_migration *mg;
 298         struct dm_bio_prison_cell *cell1;
 299         struct dm_bio_prison_cell *cell2;
 300 };
 301
 302 static void wake_worker(struct cache *cache)
 303 {
 304         queue_work(cache->wq, &cache->worker);
 305 }
 306
 307 /*----------------------------------------------------------------*/
 308
 309 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
 310 {
 311         /* FIXME: change to use a local slab. */
 312         return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
 313 }
 314
 315 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
 316 {
 317         dm_bio_prison_free_cell(cache->prison, cell);
 318 }
 319
 320 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
 321 {
 322         if (!p->mg) {
 323                 p->mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
 324                 if (!p->mg)
 325                         return -ENOMEM;
 326         }
 327
 328         if (!p->cell1) {
 329                 p->cell1 = alloc_prison_cell(cache);
 330                 if (!p->cell1)
 331                         return -ENOMEM;
 332         }
 333
 334         if (!p->cell2) {
 335                 p->cell2 = alloc_prison_cell(cache);
 336                 if (!p->cell2)
 337                         return -ENOMEM;
 338         }
 339
 340         return 0;
 341 }
 342
 343 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
 344 {
 345         if (p->cell2)
 346                 free_prison_cell(cache, p->cell2);
 347
 348         if (p->cell1)
 349                 free_prison_cell(cache, p->cell1);
 350
 351         if (p->mg)
 352                 mempool_free(p->mg, cache->migration_pool);
 353 }
 354
 355 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
 356 {
 357         struct dm_cache_migration *mg = p->mg;
 358
 359         BUG_ON(!mg);
 360         p->mg = NULL;
 361
 362         return mg;
 363 }
 364
 365 /*
 366  * You must have a cell within the prealloc struct to return.  If not this
 367  * function will BUG() rather than returning NULL.
 368  */
 369 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
 370 {
 371         struct dm_bio_prison_cell *r = NULL;
 372
 373         if (p->cell1) {
 374                 r = p->cell1;
 375                 p->cell1 = NULL;
 376
 377         } else if (p->cell2) {
 378                 r = p->cell2;
 379                 p->cell2 = NULL;
 380         } else
 381                 BUG();
 382
 383         return r;
 384 }
 385
 386 /*
 387  * You can't have more than two cells in a prealloc struct.  BUG() will be
 388  * called if you try and overfill.
 389  */
 390 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
 391 {
 392         if (!p->cell2)
 393                 p->cell2 = cell;
 394
 395         else if (!p->cell1)
 396                 p->cell1 = cell;
 397
 398         else
 399                 BUG();
 400 }
 401
 402 /*----------------------------------------------------------------*/
 403
 404 static void build_key(dm_oblock_t oblock, struct dm_cell_key *key)
 405 {
 406         key->virtual = 0;
 407         key->dev = 0;
 408         key->block = from_oblock(oblock);
 409 }
 410
 411 /*
 412  * The caller hands in a preallocated cell, and a free function for it.
 413  * The cell will be freed if there's an error, or if it wasn't used because
 414  * a cell with that key already exists.
 415  */
 416 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
 417
 418 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
 419                       struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 420                       cell_free_fn free_fn, void *free_context,
 421                       struct dm_bio_prison_cell **cell_result)
 422 {
 423         int r;
 424         struct dm_cell_key key;
 425
 426         build_key(oblock, &key);
 427         r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
 428         if (r)
 429                 free_fn(free_context, cell_prealloc);
 430
 431         return r;
 432 }
 433
 434 static int get_cell(struct cache *cache,
 435                     dm_oblock_t oblock,
 436                     struct prealloc *structs,
 437                     struct dm_bio_prison_cell **cell_result)
 438 {
 439         int r;
 440         struct dm_cell_key key;
 441         struct dm_bio_prison_cell *cell_prealloc;
 442
 443         cell_prealloc = prealloc_get_cell(structs);
 444
 445         build_key(oblock, &key);
 446         r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
 447         if (r)
 448                 prealloc_put_cell(structs, cell_prealloc);
 449
 450         return r;
 451 }
 452
 453 /*----------------------------------------------------------------*/
 454
 455 static bool is_dirty(struct cache *cache, dm_cblock_t b)
 456 {
 457         return test_bit(from_cblock(b), cache->dirty_bitset);
 458 }
 459
 460 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 461 {
 462         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 463                 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) + 1);
 464                 policy_set_dirty(cache->policy, oblock);
 465         }
 466 }
 467
 468 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 469 {
 470         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 471                 policy_clear_dirty(cache->policy, oblock);
 472                 cache->nr_dirty = to_cblock(from_cblock(cache->nr_dirty) - 1);
 473                 if (!from_cblock(cache->nr_dirty))
 474                         dm_table_event(cache->ti->table);
 475         }
 476 }
 477
 478 /*----------------------------------------------------------------*/
 479
 480 static bool block_size_is_power_of_two(struct cache *cache)
 481 {
 482         return cache->sectors_per_block_shift >= 0;
 483 }
 484
 485 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
 486 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
 487 __always_inline
 488 #endif
 489 static dm_block_t block_div(dm_block_t b, uint32_t n)
 490 {
 491         do_div(b, n);
 492
 493         return b;
 494 }
 495
 496 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 497 {
 498         uint32_t discard_blocks = cache->discard_block_size;
 499         dm_block_t b = from_oblock(oblock);
 500
 501         if (!block_size_is_power_of_two(cache))
 502                 discard_blocks = discard_blocks / cache->sectors_per_block;
 503         else
 504                 discard_blocks >>= cache->sectors_per_block_shift;
 505
 506         b = block_div(b, discard_blocks);
 507
 508         return to_dblock(b);
 509 }
 510
 511 static void set_discard(struct cache *cache, dm_dblock_t b)
 512 {
 513         unsigned long flags;
 514
 515         atomic_inc(&cache->stats.discard_count);
 516
 517         spin_lock_irqsave(&cache->lock, flags);
 518         set_bit(from_dblock(b), cache->discard_bitset);
 519         spin_unlock_irqrestore(&cache->lock, flags);
 520 }
 521
 522 static void clear_discard(struct cache *cache, dm_dblock_t b)
 523 {
 524         unsigned long flags;
 525
 526         spin_lock_irqsave(&cache->lock, flags);
 527         clear_bit(from_dblock(b), cache->discard_bitset);
 528         spin_unlock_irqrestore(&cache->lock, flags);
 529 }
 530
 531 static bool is_discarded(struct cache *cache, dm_dblock_t b)
 532 {
 533         int r;
 534         unsigned long flags;
 535
 536         spin_lock_irqsave(&cache->lock, flags);
 537         r = test_bit(from_dblock(b), cache->discard_bitset);
 538         spin_unlock_irqrestore(&cache->lock, flags);
 539
 540         return r;
 541 }
 542
 543 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 544 {
 545         int r;
 546         unsigned long flags;
 547
 548         spin_lock_irqsave(&cache->lock, flags);
 549         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 550                      cache->discard_bitset);
 551         spin_unlock_irqrestore(&cache->lock, flags);
 552
 553         return r;
 554 }
 555
 556 /*----------------------------------------------------------------*/
 557
 558 static void load_stats(struct cache *cache)
 559 {
 560         struct dm_cache_statistics stats;
 561
 562         dm_cache_metadata_get_stats(cache->cmd, &stats);
 563         atomic_set(&cache->stats.read_hit, stats.read_hits);
 564         atomic_set(&cache->stats.read_miss, stats.read_misses);
 565         atomic_set(&cache->stats.write_hit, stats.write_hits);
 566         atomic_set(&cache->stats.write_miss, stats.write_misses);
 567 }
 568
 569 static void save_stats(struct cache *cache)
 570 {
 571         struct dm_cache_statistics stats;
 572
 573         stats.read_hits = atomic_read(&cache->stats.read_hit);
 574         stats.read_misses = atomic_read(&cache->stats.read_miss);
 575         stats.write_hits = atomic_read(&cache->stats.write_hit);
 576         stats.write_misses = atomic_read(&cache->stats.write_miss);
 577
 578         dm_cache_metadata_set_stats(cache->cmd, &stats);
 579 }
 580
 581 /*----------------------------------------------------------------
 582  * Per bio data
 583  *--------------------------------------------------------------*/
 584
 585 /*
 586  * If using writeback, leave out struct per_bio_data's writethrough fields.
 587  */
 588 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
 589 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
 590
 591 static bool writethrough_mode(struct cache_features *f)
 592 {
 593         return f->io_mode == CM_IO_WRITETHROUGH;
 594 }
 595
 596 static bool writeback_mode(struct cache_features *f)
 597 {
 598         return f->io_mode == CM_IO_WRITEBACK;
 599 }
 600
 601 static bool passthrough_mode(struct cache_features *f)
 602 {
 603         return f->io_mode == CM_IO_PASSTHROUGH;
 604 }
 605
 606 static size_t get_per_bio_data_size(struct cache *cache)
 607 {
 608         return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
 609 }
 610
 611 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
 612 {
 613         struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
 614         BUG_ON(!pb);
 615         return pb;
 616 }
 617
 618 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
 619 {
 620         struct per_bio_data *pb = get_per_bio_data(bio, data_size);
 621
 622         pb->tick = false;
 623         pb->req_nr = dm_bio_get_target_bio_nr(bio);
 624         pb->all_io_entry = NULL;
 625
 626         return pb;
 627 }
 628
 629 /*----------------------------------------------------------------
 630  * Remapping
 631  *--------------------------------------------------------------*/
 632 static void remap_to_origin(struct cache *cache, struct bio *bio)
 633 {
 634         bio->bi_bdev = cache->origin_dev->bdev;
 635 }
 636
 637 static void remap_to_cache(struct cache *cache, struct bio *bio,
 638                            dm_cblock_t cblock)
 639 {
 640         sector_t bi_sector = bio->bi_sector;
 641
 642         bio->bi_bdev = cache->cache_dev->bdev;
 643         if (!block_size_is_power_of_two(cache))
 644                 bio->bi_sector = (from_cblock(cblock) * cache->sectors_per_block) +
 645                                 sector_div(bi_sector, cache->sectors_per_block);
 646         else
 647                 bio->bi_sector = (from_cblock(cblock) << cache->sectors_per_block_shift) |
 648                                 (bi_sector & (cache->sectors_per_block - 1));
 649 }
 650
 651 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 652 {
 653         unsigned long flags;
 654         size_t pb_data_size = get_per_bio_data_size(cache);
 655         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 656
 657         spin_lock_irqsave(&cache->lock, flags);
 658         if (cache->need_tick_bio &&
 659             !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
 660                 pb->tick = true;
 661                 cache->need_tick_bio = false;
 662         }
 663         spin_unlock_irqrestore(&cache->lock, flags);
 664 }
 665
 666 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 667                                   dm_oblock_t oblock)
 668 {
 669         check_if_tick_bio_needed(cache, bio);
 670         remap_to_origin(cache, bio);
 671         if (bio_data_dir(bio) == WRITE)
 672                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 673 }
 674
 675 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 676                                  dm_oblock_t oblock, dm_cblock_t cblock)
 677 {
 678         check_if_tick_bio_needed(cache, bio);
 679         remap_to_cache(cache, bio, cblock);
 680         if (bio_data_dir(bio) == WRITE) {
 681                 set_dirty(cache, oblock, cblock);
 682                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 683         }
 684 }
 685
 686 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 687 {
 688         sector_t block_nr = bio->bi_sector;
 689
 690         if (!block_size_is_power_of_two(cache))
 691                 (void) sector_div(block_nr, cache->sectors_per_block);
 692         else
 693                 block_nr >>= cache->sectors_per_block_shift;
 694
 695         return to_oblock(block_nr);
 696 }
 697
 698 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
 699 {
 700         return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
 701 }
 702
 703 static void issue(struct cache *cache, struct bio *bio)
 704 {
 705         unsigned long flags;
 706
 707         if (!bio_triggers_commit(cache, bio)) {
 708                 generic_make_request(bio);
 709                 return;
 710         }
 711
 712         /*
 713          * Batch together any bios that trigger commits and then issue a
 714          * single commit for them in do_worker().
 715          */
 716         spin_lock_irqsave(&cache->lock, flags);
 717         cache->commit_requested = true;
 718         bio_list_add(&cache->deferred_flush_bios, bio);
 719         spin_unlock_irqrestore(&cache->lock, flags);
 720 }
 721
 722 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
 723 {
 724         unsigned long flags;
 725
 726         spin_lock_irqsave(&cache->lock, flags);
 727         bio_list_add(&cache->deferred_writethrough_bios, bio);
 728         spin_unlock_irqrestore(&cache->lock, flags);
 729
 730         wake_worker(cache);
 731 }
 732
 733 static void writethrough_endio(struct bio *bio, int err)
 734 {
 735         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 736
 737         dm_unhook_bio(&pb->hook_info, bio);
 738
 739         if (err) {
 740                 bio_endio(bio, err);
 741                 return;
 742         }
 743
 744         dm_bio_restore(&pb->bio_details, bio);
 745         remap_to_cache(pb->cache, bio, pb->cblock);
 746
 747         /*
 748          * We can't issue this bio directly, since we're in interrupt
 749          * context.  So it gets put on a bio list for processing by the
 750          * worker thread.
 751          */
 752         defer_writethrough_bio(pb->cache, bio);
 753 }
 754
 755 /*
 756  * When running in writethrough mode we need to send writes to clean blocks
 757  * to both the cache and origin devices.  In future we'd like to clone the
 758  * bio and send them in parallel, but for now we're doing them in
 759  * series as this is easier.
 760  */
 761 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
 762                                        dm_oblock_t oblock, dm_cblock_t cblock)
 763 {
 764         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 765
 766         pb->cache = cache;
 767         pb->cblock = cblock;
 768         dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
 769         dm_bio_record(&pb->bio_details, bio);
 770
 771         remap_to_origin_clear_discard(pb->cache, bio, oblock);
 772 }
 773
 774 /*----------------------------------------------------------------
 775  * Migration processing
 776  *
 777  * Migration covers moving data from the origin device to the cache, or
 778  * vice versa.
 779  *--------------------------------------------------------------*/
 780 static void free_migration(struct dm_cache_migration *mg)
 781 {
 782         mempool_free(mg, mg->cache->migration_pool);
 783 }
 784
 785 static void inc_nr_migrations(struct cache *cache)
 786 {
 787         atomic_inc(&cache->nr_migrations);
 788 }
 789
 790 static void dec_nr_migrations(struct cache *cache)
 791 {
 792         atomic_dec(&cache->nr_migrations);
 793
 794         /*
 795          * Wake the worker in case we're suspending the target.
 796          */
 797         wake_up(&cache->migration_wait);
 798 }
 799
 800 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
 801                          bool holder)
 802 {
 803         (holder ? dm_cell_release : dm_cell_release_no_holder)
 804                 (cache->prison, cell, &cache->deferred_bios);
 805         free_prison_cell(cache, cell);
 806 }
 807
 808 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell,
 809                        bool holder)
 810 {
 811         unsigned long flags;
 812
 813         spin_lock_irqsave(&cache->lock, flags);
 814         __cell_defer(cache, cell, holder);
 815         spin_unlock_irqrestore(&cache->lock, flags);
 816
 817         wake_worker(cache);
 818 }
 819
 820 static void cleanup_migration(struct dm_cache_migration *mg)
 821 {
 822         struct cache *cache = mg->cache;
 823         free_migration(mg);
 824         dec_nr_migrations(cache);
 825 }
 826
 827 static void migration_failure(struct dm_cache_migration *mg)
 828 {
 829         struct cache *cache = mg->cache;
 830
 831         if (mg->writeback) {
 832                 DMWARN_LIMIT("writeback failed; couldn't copy block");
 833                 set_dirty(cache, mg->old_oblock, mg->cblock);
 834                 cell_defer(cache, mg->old_ocell, false);
 835
 836         } else if (mg->demote) {
 837                 DMWARN_LIMIT("demotion failed; couldn't copy block");
 838                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
 839
 840                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
 841                 if (mg->promote)
 842                         cell_defer(cache, mg->new_ocell, true);
 843         } else {
 844                 DMWARN_LIMIT("promotion failed; couldn't copy block");
 845                 policy_remove_mapping(cache->policy, mg->new_oblock);
 846                 cell_defer(cache, mg->new_ocell, true);
 847         }
 848
 849         cleanup_migration(mg);
 850 }
 851
 852 static void migration_success_pre_commit(struct dm_cache_migration *mg)
 853 {
 854         unsigned long flags;
 855         struct cache *cache = mg->cache;
 856
 857         if (mg->writeback) {
 858                 cell_defer(cache, mg->old_ocell, false);
 859                 clear_dirty(cache, mg->old_oblock, mg->cblock);
 860                 cleanup_migration(mg);
 861                 return;
 862
 863         } else if (mg->demote) {
 864                 if (dm_cache_remove_mapping(cache->cmd, mg->cblock)) {
 865                         DMWARN_LIMIT("demotion failed; couldn't update on disk metadata");
 866                         policy_force_mapping(cache->policy, mg->new_oblock,
 867                                              mg->old_oblock);
 868                         if (mg->promote)
 869                                 cell_defer(cache, mg->new_ocell, true);
 870                         cleanup_migration(mg);
 871                         return;
 872                 }
 873         } else {
 874                 if (dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock)) {
 875                         DMWARN_LIMIT("promotion failed; couldn't update on disk metadata");
 876                         policy_remove_mapping(cache->policy, mg->new_oblock);
 877                         cleanup_migration(mg);
 878                         return;
 879                 }
 880         }
 881
 882         spin_lock_irqsave(&cache->lock, flags);
 883         list_add_tail(&mg->list, &cache->need_commit_migrations);
 884         cache->commit_requested = true;
 885         spin_unlock_irqrestore(&cache->lock, flags);
 886 }
 887
 888 static void migration_success_post_commit(struct dm_cache_migration *mg)
 889 {
 890         unsigned long flags;
 891         struct cache *cache = mg->cache;
 892
 893         if (mg->writeback) {
 894                 DMWARN("writeback unexpectedly triggered commit");
 895                 return;
 896
 897         } else if (mg->demote) {
 898                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
 899
 900                 if (mg->promote) {
 901                         mg->demote = false;
 902
 903                         spin_lock_irqsave(&cache->lock, flags);
 904                         list_add_tail(&mg->list, &cache->quiesced_migrations);
 905                         spin_unlock_irqrestore(&cache->lock, flags);
 906
 907                 } else
 908                         cleanup_migration(mg);
 909
 910         } else {
 911                 if (mg->requeue_holder)
 912                         cell_defer(cache, mg->new_ocell, true);
 913                 else {
 914                         bio_endio(mg->new_ocell->holder, 0);
 915                         cell_defer(cache, mg->new_ocell, false);
 916                 }
 917                 clear_dirty(cache, mg->new_oblock, mg->cblock);
 918                 cleanup_migration(mg);
 919         }
 920 }
 921
 922 static void copy_complete(int read_err, unsigned long write_err, void *context)
 923 {
 924         unsigned long flags;
 925         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
 926         struct cache *cache = mg->cache;
 927
 928         if (read_err || write_err)
 929                 mg->err = true;
 930
 931         spin_lock_irqsave(&cache->lock, flags);
 932         list_add_tail(&mg->list, &cache->completed_migrations);
 933         spin_unlock_irqrestore(&cache->lock, flags);
 934
 935         wake_worker(cache);
 936 }
 937
 938 static void issue_copy_real(struct dm_cache_migration *mg)
 939 {
 940         int r;
 941         struct dm_io_region o_region, c_region;
 942         struct cache *cache = mg->cache;
 943
 944         o_region.bdev = cache->origin_dev->bdev;
 945         o_region.count = cache->sectors_per_block;
 946
 947         c_region.bdev = cache->cache_dev->bdev;
 948         c_region.sector = from_cblock(mg->cblock) * cache->sectors_per_block;
 949         c_region.count = cache->sectors_per_block;
 950
 951         if (mg->writeback || mg->demote) {
 952                 /* demote */
 953                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
 954                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
 955         } else {
 956                 /* promote */
 957                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
 958                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
 959         }
 960
 961         if (r < 0) {
 962                 DMERR_LIMIT("issuing migration failed");
 963                 migration_failure(mg);
 964         }
 965 }
 966
 967 static void overwrite_endio(struct bio *bio, int err)
 968 {
 969         struct dm_cache_migration *mg = bio->bi_private;
 970         struct cache *cache = mg->cache;
 971         size_t pb_data_size = get_per_bio_data_size(cache);
 972         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 973         unsigned long flags;
 974
 975         if (err)
 976                 mg->err = true;
 977
 978         spin_lock_irqsave(&cache->lock, flags);
 979         list_add_tail(&mg->list, &cache->completed_migrations);
 980         dm_unhook_bio(&pb->hook_info, bio);
 981         mg->requeue_holder = false;
 982         spin_unlock_irqrestore(&cache->lock, flags);
 983
 984         wake_worker(cache);
 985 }
 986
 987 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
 988 {
 989         size_t pb_data_size = get_per_bio_data_size(mg->cache);
 990         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 991
 992         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
 993         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
 994         generic_make_request(bio);
 995 }
 996
 997 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
 998 {
 999         return (bio_data_dir(bio) == WRITE) &&
1000                 (bio->bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1001 }
1002
1003 static void avoid_copy(struct dm_cache_migration *mg)
1004 {
1005         atomic_inc(&mg->cache->stats.copies_avoided);
1006         migration_success_pre_commit(mg);
1007 }
1008
1009 static void issue_copy(struct dm_cache_migration *mg)
1010 {
1011         bool avoid;
1012         struct cache *cache = mg->cache;
1013
1014         if (mg->writeback || mg->demote)
1015                 avoid = !is_dirty(cache, mg->cblock) ||
1016                         is_discarded_oblock(cache, mg->old_oblock);
1017         else {
1018                 struct bio *bio = mg->new_ocell->holder;
1019
1020                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1021
1022                 if (!avoid && bio_writes_complete_block(cache, bio)) {
1023                         issue_overwrite(mg, bio);
1024                         return;
1025                 }
1026         }
1027
1028         avoid ? avoid_copy(mg) : issue_copy_real(mg);
1029 }
1030
1031 static void complete_migration(struct dm_cache_migration *mg)
1032 {
1033         if (mg->err)
1034                 migration_failure(mg);
1035         else
1036                 migration_success_pre_commit(mg);
1037 }
1038
1039 static void process_migrations(struct cache *cache, struct list_head *head,
1040                                void (*fn)(struct dm_cache_migration *))
1041 {
1042         unsigned long flags;
1043         struct list_head list;
1044         struct dm_cache_migration *mg, *tmp;
1045
1046         INIT_LIST_HEAD(&list);
1047         spin_lock_irqsave(&cache->lock, flags);
1048         list_splice_init(head, &list);
1049         spin_unlock_irqrestore(&cache->lock, flags);
1050
1051         list_for_each_entry_safe(mg, tmp, &list, list)
1052                 fn(mg);
1053 }
1054
1055 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1056 {
1057         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1058 }
1059
1060 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1061 {
1062         unsigned long flags;
1063         struct cache *cache = mg->cache;
1064
1065         spin_lock_irqsave(&cache->lock, flags);
1066         __queue_quiesced_migration(mg);
1067         spin_unlock_irqrestore(&cache->lock, flags);
1068
1069         wake_worker(cache);
1070 }
1071
1072 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1073 {
1074         unsigned long flags;
1075         struct dm_cache_migration *mg, *tmp;
1076
1077         spin_lock_irqsave(&cache->lock, flags);
1078         list_for_each_entry_safe(mg, tmp, work, list)
1079                 __queue_quiesced_migration(mg);
1080         spin_unlock_irqrestore(&cache->lock, flags);
1081
1082         wake_worker(cache);
1083 }
1084
1085 static void check_for_quiesced_migrations(struct cache *cache,
1086                                           struct per_bio_data *pb)
1087 {
1088         struct list_head work;
1089
1090         if (!pb->all_io_entry)
1091                 return;
1092
1093         INIT_LIST_HEAD(&work);
1094         if (pb->all_io_entry)
1095                 dm_deferred_entry_dec(pb->all_io_entry, &work);
1096
1097         if (!list_empty(&work))
1098                 queue_quiesced_migrations(cache, &work);
1099 }
1100
1101 static void quiesce_migration(struct dm_cache_migration *mg)
1102 {
1103         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1104                 queue_quiesced_migration(mg);
1105 }
1106
1107 static void promote(struct cache *cache, struct prealloc *structs,
1108                     dm_oblock_t oblock, dm_cblock_t cblock,
1109                     struct dm_bio_prison_cell *cell)
1110 {
1111         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1112
1113         mg->err = false;
1114         mg->writeback = false;
1115         mg->demote = false;
1116         mg->promote = true;
1117         mg->requeue_holder = true;
1118         mg->cache = cache;
1119         mg->new_oblock = oblock;
1120         mg->cblock = cblock;
1121         mg->old_ocell = NULL;
1122         mg->new_ocell = cell;
1123         mg->start_jiffies = jiffies;
1124
1125         inc_nr_migrations(cache);
1126         quiesce_migration(mg);
1127 }
1128
1129 static void writeback(struct cache *cache, struct prealloc *structs,
1130                       dm_oblock_t oblock, dm_cblock_t cblock,
1131                       struct dm_bio_prison_cell *cell)
1132 {
1133         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1134
1135         mg->err = false;
1136         mg->writeback = true;
1137         mg->demote = false;
1138         mg->promote = false;
1139         mg->requeue_holder = true;
1140         mg->cache = cache;
1141         mg->old_oblock = oblock;
1142         mg->cblock = cblock;
1143         mg->old_ocell = cell;
1144         mg->new_ocell = NULL;
1145         mg->start_jiffies = jiffies;
1146
1147         inc_nr_migrations(cache);
1148         quiesce_migration(mg);
1149 }
1150
1151 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1152                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1153                                 dm_cblock_t cblock,
1154                                 struct dm_bio_prison_cell *old_ocell,
1155                                 struct dm_bio_prison_cell *new_ocell)
1156 {
1157         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1158
1159         mg->err = false;
1160         mg->writeback = false;
1161         mg->demote = true;
1162         mg->promote = true;
1163         mg->requeue_holder = true;
1164         mg->cache = cache;
1165         mg->old_oblock = old_oblock;
1166         mg->new_oblock = new_oblock;
1167         mg->cblock = cblock;
1168         mg->old_ocell = old_ocell;
1169         mg->new_ocell = new_ocell;
1170         mg->start_jiffies = jiffies;
1171
1172         inc_nr_migrations(cache);
1173         quiesce_migration(mg);
1174 }
1175
1176 /*
1177  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1178  * block are thrown away.
1179  */
1180 static void invalidate(struct cache *cache, struct prealloc *structs,
1181                        dm_oblock_t oblock, dm_cblock_t cblock,
1182                        struct dm_bio_prison_cell *cell)
1183 {
1184         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1185
1186         mg->err = false;
1187         mg->writeback = false;
1188         mg->demote = true;
1189         mg->promote = false;
1190         mg->requeue_holder = true;
1191         mg->cache = cache;
1192         mg->old_oblock = oblock;
1193         mg->cblock = cblock;
1194         mg->old_ocell = cell;
1195         mg->new_ocell = NULL;
1196         mg->start_jiffies = jiffies;
1197
1198         inc_nr_migrations(cache);
1199         quiesce_migration(mg);
1200 }
1201
1202 /*----------------------------------------------------------------
1203  * bio processing
1204  *--------------------------------------------------------------*/
1205 static void defer_bio(struct cache *cache, struct bio *bio)
1206 {
1207         unsigned long flags;
1208
1209         spin_lock_irqsave(&cache->lock, flags);
1210         bio_list_add(&cache->deferred_bios, bio);
1211         spin_unlock_irqrestore(&cache->lock, flags);
1212
1213         wake_worker(cache);
1214 }
1215
1216 static void process_flush_bio(struct cache *cache, struct bio *bio)
1217 {
1218         size_t pb_data_size = get_per_bio_data_size(cache);
1219         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1220
1221         BUG_ON(bio->bi_size);
1222         if (!pb->req_nr)
1223                 remap_to_origin(cache, bio);
1224         else
1225                 remap_to_cache(cache, bio, 0);
1226
1227         issue(cache, bio);
1228 }
1229
1230 /*
1231  * People generally discard large parts of a device, eg, the whole device
1232  * when formatting.  Splitting these large discards up into cache block
1233  * sized ios and then quiescing (always neccessary for discard) takes too
1234  * long.
1235  *
1236  * We keep it simple, and allow any size of discard to come in, and just
1237  * mark off blocks on the discard bitset.  No passdown occurs!
1238  *
1239  * To implement passdown we need to change the bio_prison such that a cell
1240  * can have a key that spans many blocks.
1241  */
1242 static void process_discard_bio(struct cache *cache, struct bio *bio)
1243 {
1244         dm_block_t start_block = dm_sector_div_up(bio->bi_sector,
1245                                                   cache->discard_block_size);
1246         dm_block_t end_block = bio->bi_sector + bio_sectors(bio);
1247         dm_block_t b;
1248
1249         end_block = block_div(end_block, cache->discard_block_size);
1250
1251         for (b = start_block; b < end_block; b++)
1252                 set_discard(cache, to_dblock(b));
1253
1254         bio_endio(bio, 0);
1255 }
1256
1257 static bool spare_migration_bandwidth(struct cache *cache)
1258 {
1259         sector_t current_volume = (atomic_read(&cache->nr_migrations) + 1) *
1260                 cache->sectors_per_block;
1261         return current_volume < cache->migration_threshold;
1262 }
1263
1264 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1265 {
1266         atomic_inc(bio_data_dir(bio) == READ ?
1267                    &cache->stats.read_hit : &cache->stats.write_hit);
1268 }
1269
1270 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1271 {
1272         atomic_inc(bio_data_dir(bio) == READ ?
1273                    &cache->stats.read_miss : &cache->stats.write_miss);
1274 }
1275
1276 static void issue_cache_bio(struct cache *cache, struct bio *bio,
1277                             struct per_bio_data *pb,
1278                             dm_oblock_t oblock, dm_cblock_t cblock)
1279 {
1280         pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1281         remap_to_cache_dirty(cache, bio, oblock, cblock);
1282         issue(cache, bio);
1283 }
1284
1285 static void process_bio(struct cache *cache, struct prealloc *structs,
1286                         struct bio *bio)
1287 {
1288         int r;
1289         bool release_cell = true;
1290         dm_oblock_t block = get_bio_block(cache, bio);
1291         struct dm_bio_prison_cell *cell_prealloc, *old_ocell, *new_ocell;
1292         struct policy_result lookup_result;
1293         size_t pb_data_size = get_per_bio_data_size(cache);
1294         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1295         bool discarded_block = is_discarded_oblock(cache, block);
1296         bool passthrough = passthrough_mode(&cache->features);
1297         bool can_migrate = !passthrough && (discarded_block || spare_migration_bandwidth(cache));
1298
1299         /*
1300          * Check to see if that block is currently migrating.
1301          */
1302         cell_prealloc = prealloc_get_cell(structs);
1303         r = bio_detain(cache, block, bio, cell_prealloc,
1304                        (cell_free_fn) prealloc_put_cell,
1305                        structs, &new_ocell);
1306         if (r > 0)
1307                 return;
1308
1309         r = policy_map(cache->policy, block, true, can_migrate, discarded_block,
1310                        bio, &lookup_result);
1311
1312         if (r == -EWOULDBLOCK)
1313                 /* migration has been denied */
1314                 lookup_result.op = POLICY_MISS;
1315
1316         switch (lookup_result.op) {
1317         case POLICY_HIT:
1318                 if (passthrough) {
1319                         inc_miss_counter(cache, bio);
1320
1321                         /*
1322                          * Passthrough always maps to the origin,
1323                          * invalidating any cache blocks that are written
1324                          * to.
1325                          */
1326
1327                         if (bio_data_dir(bio) == WRITE) {
1328                                 atomic_inc(&cache->stats.demotion);
1329                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1330                                 release_cell = false;
1331
1332                         } else {
1333                                 /* FIXME: factor out issue_origin() */
1334                                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1335                                 remap_to_origin_clear_discard(cache, bio, block);
1336                                 issue(cache, bio);
1337                         }
1338                 } else {
1339                         inc_hit_counter(cache, bio);
1340
1341                         if (bio_data_dir(bio) == WRITE &&
1342                             writethrough_mode(&cache->features) &&
1343                             !is_dirty(cache, lookup_result.cblock)) {
1344                                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1345                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1346                                 issue(cache, bio);
1347                         } else
1348                                 issue_cache_bio(cache, bio, pb, block, lookup_result.cblock);
1349                 }
1350
1351                 break;
1352
1353         case POLICY_MISS:
1354                 inc_miss_counter(cache, bio);
1355                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
1356                 remap_to_origin_clear_discard(cache, bio, block);
1357                 issue(cache, bio);
1358                 break;
1359
1360         case POLICY_NEW:
1361                 atomic_inc(&cache->stats.promotion);
1362                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1363                 release_cell = false;
1364                 break;
1365
1366         case POLICY_REPLACE:
1367                 cell_prealloc = prealloc_get_cell(structs);
1368                 r = bio_detain(cache, lookup_result.old_oblock, bio, cell_prealloc,
1369                                (cell_free_fn) prealloc_put_cell,
1370                                structs, &old_ocell);
1371                 if (r > 0) {
1372                         /*
1373                          * We have to be careful to avoid lock inversion of
1374                          * the cells.  So we back off, and wait for the
1375                          * old_ocell to become free.
1376                          */
1377                         policy_force_mapping(cache->policy, block,
1378                                              lookup_result.old_oblock);
1379                         atomic_inc(&cache->stats.cache_cell_clash);
1380                         break;
1381                 }
1382                 atomic_inc(&cache->stats.demotion);
1383                 atomic_inc(&cache->stats.promotion);
1384
1385                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1386                                     block, lookup_result.cblock,
1387                                     old_ocell, new_ocell);
1388                 release_cell = false;
1389                 break;
1390
1391         default:
1392                 DMERR_LIMIT("%s: erroring bio, unknown policy op: %u", __func__,
1393                             (unsigned) lookup_result.op);
1394                 bio_io_error(bio);
1395         }
1396
1397         if (release_cell)
1398                 cell_defer(cache, new_ocell, false);
1399 }
1400
1401 static int need_commit_due_to_time(struct cache *cache)
1402 {
1403         return jiffies < cache->last_commit_jiffies ||
1404                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1405 }
1406
1407 static int commit_if_needed(struct cache *cache)
1408 {
1409         int r = 0;
1410
1411         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1412             dm_cache_changed_this_transaction(cache->cmd)) {
1413                 atomic_inc(&cache->stats.commit_count);
1414                 cache->commit_requested = false;
1415                 r = dm_cache_commit(cache->cmd, false);
1416                 cache->last_commit_jiffies = jiffies;
1417         }
1418
1419         return r;
1420 }
1421
1422 static void process_deferred_bios(struct cache *cache)
1423 {
1424         unsigned long flags;
1425         struct bio_list bios;
1426         struct bio *bio;
1427         struct prealloc structs;
1428
1429         memset(&structs, 0, sizeof(structs));
1430         bio_list_init(&bios);
1431
1432         spin_lock_irqsave(&cache->lock, flags);
1433         bio_list_merge(&bios, &cache->deferred_bios);
1434         bio_list_init(&cache->deferred_bios);
1435         spin_unlock_irqrestore(&cache->lock, flags);
1436
1437         while (!bio_list_empty(&bios)) {
1438                 /*
1439                  * If we've got no free migration structs, and processing
1440                  * this bio might require one, we pause until there are some
1441                  * prepared mappings to process.
1442                  */
1443                 if (prealloc_data_structs(cache, &structs)) {
1444                         spin_lock_irqsave(&cache->lock, flags);
1445                         bio_list_merge(&cache->deferred_bios, &bios);
1446                         spin_unlock_irqrestore(&cache->lock, flags);
1447                         break;
1448                 }
1449
1450                 bio = bio_list_pop(&bios);
1451
1452                 if (bio->bi_rw & REQ_FLUSH)
1453                         process_flush_bio(cache, bio);
1454                 else if (bio->bi_rw & REQ_DISCARD)
1455                         process_discard_bio(cache, bio);
1456                 else
1457                         process_bio(cache, &structs, bio);
1458         }
1459
1460         prealloc_free_structs(cache, &structs);
1461 }
1462
1463 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
1464 {
1465         unsigned long flags;
1466         struct bio_list bios;
1467         struct bio *bio;
1468
1469         bio_list_init(&bios);
1470
1471         spin_lock_irqsave(&cache->lock, flags);
1472         bio_list_merge(&bios, &cache->deferred_flush_bios);
1473         bio_list_init(&cache->deferred_flush_bios);
1474         spin_unlock_irqrestore(&cache->lock, flags);
1475
1476         while ((bio = bio_list_pop(&bios)))
1477                 submit_bios ? generic_make_request(bio) : bio_io_error(bio);
1478 }
1479
1480 static void process_deferred_writethrough_bios(struct cache *cache)
1481 {
1482         unsigned long flags;
1483         struct bio_list bios;
1484         struct bio *bio;
1485
1486         bio_list_init(&bios);
1487
1488         spin_lock_irqsave(&cache->lock, flags);
1489         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
1490         bio_list_init(&cache->deferred_writethrough_bios);
1491         spin_unlock_irqrestore(&cache->lock, flags);
1492
1493         while ((bio = bio_list_pop(&bios)))
1494                 generic_make_request(bio);
1495 }
1496
1497 static void writeback_some_dirty_blocks(struct cache *cache)
1498 {
1499         int r = 0;
1500         dm_oblock_t oblock;
1501         dm_cblock_t cblock;
1502         struct prealloc structs;
1503         struct dm_bio_prison_cell *old_ocell;
1504
1505         memset(&structs, 0, sizeof(structs));
1506
1507         while (spare_migration_bandwidth(cache)) {
1508                 if (prealloc_data_structs(cache, &structs))
1509                         break;
1510
1511                 r = policy_writeback_work(cache->policy, &oblock, &cblock);
1512                 if (r)
1513                         break;
1514
1515                 r = get_cell(cache, oblock, &structs, &old_ocell);
1516                 if (r) {
1517                         policy_set_dirty(cache->policy, oblock);
1518                         break;
1519                 }
1520
1521                 writeback(cache, &structs, oblock, cblock, old_ocell);
1522         }
1523
1524         prealloc_free_structs(cache, &structs);
1525 }
1526
1527 /*----------------------------------------------------------------
1528  * Main worker loop
1529  *--------------------------------------------------------------*/
1530 static bool is_quiescing(struct cache *cache)
1531 {
1532         return atomic_read(&cache->quiescing);
1533 }
1534
1535 static void ack_quiescing(struct cache *cache)
1536 {
1537         if (is_quiescing(cache)) {
1538                 atomic_inc(&cache->quiescing_ack);
1539                 wake_up(&cache->quiescing_wait);
1540         }
1541 }
1542
1543 static void wait_for_quiescing_ack(struct cache *cache)
1544 {
1545         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
1546 }
1547
1548 static void start_quiescing(struct cache *cache)
1549 {
1550         atomic_inc(&cache->quiescing);
1551         wait_for_quiescing_ack(cache);
1552 }
1553
1554 static void stop_quiescing(struct cache *cache)
1555 {
1556         atomic_set(&cache->quiescing, 0);
1557         atomic_set(&cache->quiescing_ack, 0);
1558 }
1559
1560 static void wait_for_migrations(struct cache *cache)
1561 {
1562         wait_event(cache->migration_wait, !atomic_read(&cache->nr_migrations));
1563 }
1564
1565 static void stop_worker(struct cache *cache)
1566 {
1567         cancel_delayed_work(&cache->waker);
1568         flush_workqueue(cache->wq);
1569 }
1570
1571 static void requeue_deferred_io(struct cache *cache)
1572 {
1573         struct bio *bio;
1574         struct bio_list bios;
1575
1576         bio_list_init(&bios);
1577         bio_list_merge(&bios, &cache->deferred_bios);
1578         bio_list_init(&cache->deferred_bios);
1579
1580         while ((bio = bio_list_pop(&bios)))
1581                 bio_endio(bio, DM_ENDIO_REQUEUE);
1582 }
1583
1584 static int more_work(struct cache *cache)
1585 {
1586         if (is_quiescing(cache))
1587                 return !list_empty(&cache->quiesced_migrations) ||
1588                         !list_empty(&cache->completed_migrations) ||
1589                         !list_empty(&cache->need_commit_migrations);
1590         else
1591                 return !bio_list_empty(&cache->deferred_bios) ||
1592                         !bio_list_empty(&cache->deferred_flush_bios) ||
1593                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
1594                         !list_empty(&cache->quiesced_migrations) ||
1595                         !list_empty(&cache->completed_migrations) ||
1596                         !list_empty(&cache->need_commit_migrations);
1597 }
1598
1599 static void do_worker(struct work_struct *ws)
1600 {
1601         struct cache *cache = container_of(ws, struct cache, worker);
1602
1603         do {
1604                 if (!is_quiescing(cache)) {
1605                         writeback_some_dirty_blocks(cache);
1606                         process_deferred_writethrough_bios(cache);
1607                         process_deferred_bios(cache);
1608                 }
1609
1610                 process_migrations(cache, &cache->quiesced_migrations, issue_copy);
1611                 process_migrations(cache, &cache->completed_migrations, complete_migration);
1612
1613                 if (commit_if_needed(cache)) {
1614                         process_deferred_flush_bios(cache, false);
1615
1616                         /*
1617                          * FIXME: rollback metadata or just go into a
1618                          * failure mode and error everything
1619                          */
1620                 } else {
1621                         process_deferred_flush_bios(cache, true);
1622                         process_migrations(cache, &cache->need_commit_migrations,
1623                                            migration_success_post_commit);
1624                 }
1625
1626                 ack_quiescing(cache);
1627
1628         } while (more_work(cache));
1629 }
1630
1631 /*
1632  * We want to commit periodically so that not too much
1633  * unwritten metadata builds up.
1634  */
1635 static void do_waker(struct work_struct *ws)
1636 {
1637         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1638         policy_tick(cache->policy);
1639         wake_worker(cache);
1640         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1641 }
1642
1643 /*----------------------------------------------------------------*/
1644
1645 static int is_congested(struct dm_dev *dev, int bdi_bits)
1646 {
1647         struct request_queue *q = bdev_get_queue(dev->bdev);
1648         return bdi_congested(&q->backing_dev_info, bdi_bits);
1649 }
1650
1651 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1652 {
1653         struct cache *cache = container_of(cb, struct cache, callbacks);
1654
1655         return is_congested(cache->origin_dev, bdi_bits) ||
1656                 is_congested(cache->cache_dev, bdi_bits);
1657 }
1658
1659 /*----------------------------------------------------------------
1660  * Target methods
1661  *--------------------------------------------------------------*/
1662
1663 /*
1664  * This function gets called on the error paths of the constructor, so we
1665  * have to cope with a partially initialised struct.
1666  */
1667 static void destroy(struct cache *cache)
1668 {
1669         unsigned i;
1670
1671         if (cache->next_migration)
1672                 mempool_free(cache->next_migration, cache->migration_pool);
1673
1674         if (cache->migration_pool)
1675                 mempool_destroy(cache->migration_pool);
1676
1677         if (cache->all_io_ds)
1678                 dm_deferred_set_destroy(cache->all_io_ds);
1679
1680         if (cache->prison)
1681                 dm_bio_prison_destroy(cache->prison);
1682
1683         if (cache->wq)
1684                 destroy_workqueue(cache->wq);
1685
1686         if (cache->dirty_bitset)
1687                 free_bitset(cache->dirty_bitset);
1688
1689         if (cache->discard_bitset)
1690                 free_bitset(cache->discard_bitset);
1691
1692         if (cache->copier)
1693                 dm_kcopyd_client_destroy(cache->copier);
1694
1695         if (cache->cmd)
1696                 dm_cache_metadata_close(cache->cmd);
1697
1698         if (cache->metadata_dev)
1699                 dm_put_device(cache->ti, cache->metadata_dev);
1700
1701         if (cache->origin_dev)
1702                 dm_put_device(cache->ti, cache->origin_dev);
1703
1704         if (cache->cache_dev)
1705                 dm_put_device(cache->ti, cache->cache_dev);
1706
1707         if (cache->policy)
1708                 dm_cache_policy_destroy(cache->policy);
1709
1710         for (i = 0; i < cache->nr_ctr_args ; i++)
1711                 kfree(cache->ctr_args[i]);
1712         kfree(cache->ctr_args);
1713
1714         kfree(cache);
1715 }
1716
1717 static void cache_dtr(struct dm_target *ti)
1718 {
1719         struct cache *cache = ti->private;
1720
1721         destroy(cache);
1722 }
1723
1724 static sector_t get_dev_size(struct dm_dev *dev)
1725 {
1726         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1727 }
1728
1729 /*----------------------------------------------------------------*/
1730
1731 /*
1732  * Construct a cache device mapping.
1733  *
1734  * cache <metadata dev> <cache dev> <origin dev> <block size>
1735  *       <#feature args> [<feature arg>]*
1736  *       <policy> <#policy args> [<policy arg>]*
1737  *
1738  * metadata dev    : fast device holding the persistent metadata
1739  * cache dev       : fast device holding cached data blocks
1740  * origin dev      : slow device holding original data blocks
1741  * block size      : cache unit size in sectors
1742  *
1743  * #feature args   : number of feature arguments passed
1744  * feature args    : writethrough.  (The default is writeback.)
1745  *
1746  * policy          : the replacement policy to use
1747  * #policy args    : an even number of policy arguments corresponding
1748  *                   to key/value pairs passed to the policy
1749  * policy args     : key/value pairs passed to the policy
1750  *                   E.g. 'sequential_threshold 1024'
1751  *                   See cache-policies.txt for details.
1752  *
1753  * Optional feature arguments are:
1754  *   writethrough  : write through caching that prohibits cache block
1755  *                   content from being different from origin block content.
1756  *                   Without this argument, the default behaviour is to write
1757  *                   back cache block contents later for performance reasons,
1758  *                   so they may differ from the corresponding origin blocks.
1759  */
1760 struct cache_args {
1761         struct dm_target *ti;
1762
1763         struct dm_dev *metadata_dev;
1764
1765         struct dm_dev *cache_dev;
1766         sector_t cache_sectors;
1767
1768         struct dm_dev *origin_dev;
1769         sector_t origin_sectors;
1770
1771         uint32_t block_size;
1772
1773         const char *policy_name;
1774         int policy_argc;
1775         const char **policy_argv;
1776
1777         struct cache_features features;
1778 };
1779
1780 static void destroy_cache_args(struct cache_args *ca)
1781 {
1782         if (ca->metadata_dev)
1783                 dm_put_device(ca->ti, ca->metadata_dev);
1784
1785         if (ca->cache_dev)
1786                 dm_put_device(ca->ti, ca->cache_dev);
1787
1788         if (ca->origin_dev)
1789                 dm_put_device(ca->ti, ca->origin_dev);
1790
1791         kfree(ca);
1792 }
1793
1794 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
1795 {
1796         if (!as->argc) {
1797                 *error = "Insufficient args";
1798                 return false;
1799         }
1800
1801         return true;
1802 }
1803
1804 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
1805                               char **error)
1806 {
1807         int r;
1808         sector_t metadata_dev_size;
1809         char b[BDEVNAME_SIZE];
1810
1811         if (!at_least_one_arg(as, error))
1812                 return -EINVAL;
1813
1814         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1815                           &ca->metadata_dev);
1816         if (r) {
1817                 *error = "Error opening metadata device";
1818                 return r;
1819         }
1820
1821         metadata_dev_size = get_dev_size(ca->metadata_dev);
1822         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
1823                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
1824                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
1825
1826         return 0;
1827 }
1828
1829 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
1830                            char **error)
1831 {
1832         int r;
1833
1834         if (!at_least_one_arg(as, error))
1835                 return -EINVAL;
1836
1837         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1838                           &ca->cache_dev);
1839         if (r) {
1840                 *error = "Error opening cache device";
1841                 return r;
1842         }
1843         ca->cache_sectors = get_dev_size(ca->cache_dev);
1844
1845         return 0;
1846 }
1847
1848 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
1849                             char **error)
1850 {
1851         int r;
1852
1853         if (!at_least_one_arg(as, error))
1854                 return -EINVAL;
1855
1856         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
1857                           &ca->origin_dev);
1858         if (r) {
1859                 *error = "Error opening origin device";
1860                 return r;
1861         }
1862
1863         ca->origin_sectors = get_dev_size(ca->origin_dev);
1864         if (ca->ti->len > ca->origin_sectors) {
1865                 *error = "Device size larger than cached device";
1866                 return -EINVAL;
1867         }
1868
1869         return 0;
1870 }
1871
1872 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
1873                             char **error)
1874 {
1875         unsigned long block_size;
1876
1877         if (!at_least_one_arg(as, error))
1878                 return -EINVAL;
1879
1880         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
1881             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
1882             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
1883             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
1884                 *error = "Invalid data block size";
1885                 return -EINVAL;
1886         }
1887
1888         if (block_size > ca->cache_sectors) {
1889                 *error = "Data block size is larger than the cache device";
1890                 return -EINVAL;
1891         }
1892
1893         ca->block_size = block_size;
1894
1895         return 0;
1896 }
1897
1898 static void init_features(struct cache_features *cf)
1899 {
1900         cf->mode = CM_WRITE;
1901         cf->io_mode = CM_IO_WRITEBACK;
1902 }
1903
1904 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
1905                           char **error)
1906 {
1907         static struct dm_arg _args[] = {
1908                 {0, 1, "Invalid number of cache feature arguments"},
1909         };
1910
1911         int r;
1912         unsigned argc;
1913         const char *arg;
1914         struct cache_features *cf = &ca->features;
1915
1916         init_features(cf);
1917
1918         r = dm_read_arg_group(_args, as, &argc, error);
1919         if (r)
1920                 return -EINVAL;
1921
1922         while (argc--) {
1923                 arg = dm_shift_arg(as);
1924
1925                 if (!strcasecmp(arg, "writeback"))
1926                         cf->io_mode = CM_IO_WRITEBACK;
1927
1928                 else if (!strcasecmp(arg, "writethrough"))
1929                         cf->io_mode = CM_IO_WRITETHROUGH;
1930
1931                 else if (!strcasecmp(arg, "passthrough"))
1932                         cf->io_mode = CM_IO_PASSTHROUGH;
1933
1934                 else {
1935                         *error = "Unrecognised cache feature requested";
1936                         return -EINVAL;
1937                 }
1938         }
1939
1940         return 0;
1941 }
1942
1943 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
1944                         char **error)
1945 {
1946         static struct dm_arg _args[] = {
1947                 {0, 1024, "Invalid number of policy arguments"},
1948         };
1949
1950         int r;
1951
1952         if (!at_least_one_arg(as, error))
1953                 return -EINVAL;
1954
1955         ca->policy_name = dm_shift_arg(as);
1956
1957         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
1958         if (r)
1959                 return -EINVAL;
1960
1961         ca->policy_argv = (const char **)as->argv;
1962         dm_consume_args(as, ca->policy_argc);
1963
1964         return 0;
1965 }
1966
1967 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
1968                             char **error)
1969 {
1970         int r;
1971         struct dm_arg_set as;
1972
1973         as.argc = argc;
1974         as.argv = argv;
1975
1976         r = parse_metadata_dev(ca, &as, error);
1977         if (r)
1978                 return r;
1979
1980         r = parse_cache_dev(ca, &as, error);
1981         if (r)
1982                 return r;
1983
1984         r = parse_origin_dev(ca, &as, error);
1985         if (r)
1986                 return r;
1987
1988         r = parse_block_size(ca, &as, error);
1989         if (r)
1990                 return r;
1991
1992         r = parse_features(ca, &as, error);
1993         if (r)
1994                 return r;
1995
1996         r = parse_policy(ca, &as, error);
1997         if (r)
1998                 return r;
1999
2000         return 0;
2001 }
2002
2003 /*----------------------------------------------------------------*/
2004
2005 static struct kmem_cache *migration_cache;
2006
2007 #define NOT_CORE_OPTION 1
2008
2009 static int process_config_option(struct cache *cache, const char *key, const char *value)
2010 {
2011         unsigned long tmp;
2012
2013         if (!strcasecmp(key, "migration_threshold")) {
2014                 if (kstrtoul(value, 10, &tmp))
2015                         return -EINVAL;
2016
2017                 cache->migration_threshold = tmp;
2018                 return 0;
2019         }
2020
2021         return NOT_CORE_OPTION;
2022 }
2023
2024 static int set_config_value(struct cache *cache, const char *key, const char *value)
2025 {
2026         int r = process_config_option(cache, key, value);
2027
2028         if (r == NOT_CORE_OPTION)
2029                 r = policy_set_config_value(cache->policy, key, value);
2030
2031         if (r)
2032                 DMWARN("bad config value for %s: %s", key, value);
2033
2034         return r;
2035 }
2036
2037 static int set_config_values(struct cache *cache, int argc, const char **argv)
2038 {
2039         int r = 0;
2040
2041         if (argc & 1) {
2042                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2043                 return -EINVAL;
2044         }
2045
2046         while (argc) {
2047                 r = set_config_value(cache, argv[0], argv[1]);
2048                 if (r)
2049                         break;
2050
2051                 argc -= 2;
2052                 argv += 2;
2053         }
2054
2055         return r;
2056 }
2057
2058 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2059                                char **error)
2060 {
2061         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2062                                                            cache->cache_size,
2063                                                            cache->origin_sectors,
2064                                                            cache->sectors_per_block);
2065         if (IS_ERR(p)) {
2066                 *error = "Error creating cache's policy";
2067                 return PTR_ERR(p);
2068         }
2069         cache->policy = p;
2070
2071         return 0;
2072 }
2073
2074 /*
2075  * We want the discard block size to be a power of two, at least the size
2076  * of the cache block size, and have no more than 2^14 discard blocks
2077  * across the origin.
2078  */
2079 #define MAX_DISCARD_BLOCKS (1 << 14)
2080
2081 static bool too_many_discard_blocks(sector_t discard_block_size,
2082                                     sector_t origin_size)
2083 {
2084         (void) sector_div(origin_size, discard_block_size);
2085
2086         return origin_size > MAX_DISCARD_BLOCKS;
2087 }
2088
2089 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2090                                              sector_t origin_size)
2091 {
2092         sector_t discard_block_size;
2093
2094         discard_block_size = roundup_pow_of_two(cache_block_size);
2095
2096         if (origin_size)
2097                 while (too_many_discard_blocks(discard_block_size, origin_size))
2098                         discard_block_size *= 2;
2099
2100         return discard_block_size;
2101 }
2102
2103 #define DEFAULT_MIGRATION_THRESHOLD 2048
2104
2105 static int cache_create(struct cache_args *ca, struct cache **result)
2106 {
2107         int r = 0;
2108         char **error = &ca->ti->error;
2109         struct cache *cache;
2110         struct dm_target *ti = ca->ti;
2111         dm_block_t origin_blocks;
2112         struct dm_cache_metadata *cmd;
2113         bool may_format = ca->features.mode == CM_WRITE;
2114
2115         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2116         if (!cache)
2117                 return -ENOMEM;
2118
2119         cache->ti = ca->ti;
2120         ti->private = cache;
2121         ti->num_flush_bios = 2;
2122         ti->flush_supported = true;
2123
2124         ti->num_discard_bios = 1;
2125         ti->discards_supported = true;
2126         ti->discard_zeroes_data_unsupported = true;
2127
2128         cache->features = ca->features;
2129         ti->per_bio_data_size = get_per_bio_data_size(cache);
2130
2131         cache->callbacks.congested_fn = cache_is_congested;
2132         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2133
2134         cache->metadata_dev = ca->metadata_dev;
2135         cache->origin_dev = ca->origin_dev;
2136         cache->cache_dev = ca->cache_dev;
2137
2138         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2139
2140         /* FIXME: factor out this whole section */
2141         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2142         origin_blocks = block_div(origin_blocks, ca->block_size);
2143         cache->origin_blocks = to_oblock(origin_blocks);
2144
2145         cache->sectors_per_block = ca->block_size;
2146         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2147                 r = -EINVAL;
2148                 goto bad;
2149         }
2150
2151         if (ca->block_size & (ca->block_size - 1)) {
2152                 dm_block_t cache_size = ca->cache_sectors;
2153
2154                 cache->sectors_per_block_shift = -1;
2155                 cache_size = block_div(cache_size, ca->block_size);
2156                 cache->cache_size = to_cblock(cache_size);
2157         } else {
2158                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2159                 cache->cache_size = to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift);
2160         }
2161
2162         r = create_cache_policy(cache, ca, error);
2163         if (r)
2164                 goto bad;
2165
2166         cache->policy_nr_args = ca->policy_argc;
2167         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2168
2169         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2170         if (r) {
2171                 *error = "Error setting cache policy's config values";
2172                 goto bad;
2173         }
2174
2175         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2176                                      ca->block_size, may_format,
2177                                      dm_cache_policy_get_hint_size(cache->policy));
2178         if (IS_ERR(cmd)) {
2179                 *error = "Error creating metadata object";
2180                 r = PTR_ERR(cmd);
2181                 goto bad;
2182         }
2183         cache->cmd = cmd;
2184
2185         if (passthrough_mode(&cache->features)) {
2186                 bool all_clean;
2187
2188                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2189                 if (r) {
2190                         *error = "dm_cache_metadata_all_clean() failed";
2191                         goto bad;
2192                 }
2193
2194                 if (!all_clean) {
2195                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2196                         r = -EINVAL;
2197                         goto bad;
2198                 }
2199         }
2200
2201         spin_lock_init(&cache->lock);
2202         bio_list_init(&cache->deferred_bios);
2203         bio_list_init(&cache->deferred_flush_bios);
2204         bio_list_init(&cache->deferred_writethrough_bios);
2205         INIT_LIST_HEAD(&cache->quiesced_migrations);
2206         INIT_LIST_HEAD(&cache->completed_migrations);
2207         INIT_LIST_HEAD(&cache->need_commit_migrations);
2208         atomic_set(&cache->nr_migrations, 0);
2209         init_waitqueue_head(&cache->migration_wait);
2210
2211         init_waitqueue_head(&cache->quiescing_wait);
2212         atomic_set(&cache->quiescing, 0);
2213         atomic_set(&cache->quiescing_ack, 0);
2214
2215         r = -ENOMEM;
2216         cache->nr_dirty = 0;
2217         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2218         if (!cache->dirty_bitset) {
2219                 *error = "could not allocate dirty bitset";
2220                 goto bad;
2221         }
2222         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2223
2224         cache->discard_block_size =
2225                 calculate_discard_block_size(cache->sectors_per_block,
2226                                              cache->origin_sectors);
2227         cache->discard_nr_blocks = oblock_to_dblock(cache, cache->origin_blocks);
2228         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2229         if (!cache->discard_bitset) {
2230                 *error = "could not allocate discard bitset";
2231                 goto bad;
2232         }
2233         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2234
2235         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2236         if (IS_ERR(cache->copier)) {
2237                 *error = "could not create kcopyd client";
2238                 r = PTR_ERR(cache->copier);
2239                 goto bad;
2240         }
2241
2242         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2243         if (!cache->wq) {
2244                 *error = "could not create workqueue for metadata object";
2245                 goto bad;
2246         }
2247         INIT_WORK(&cache->worker, do_worker);
2248         INIT_DELAYED_WORK(&cache->waker, do_waker);
2249         cache->last_commit_jiffies = jiffies;
2250
2251         cache->prison = dm_bio_prison_create(PRISON_CELLS);
2252         if (!cache->prison) {
2253                 *error = "could not create bio prison";
2254                 goto bad;
2255         }
2256
2257         cache->all_io_ds = dm_deferred_set_create();
2258         if (!cache->all_io_ds) {
2259                 *error = "could not create all_io deferred set";
2260                 goto bad;
2261         }
2262
2263         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2264                                                          migration_cache);
2265         if (!cache->migration_pool) {
2266                 *error = "Error creating cache's migration mempool";
2267                 goto bad;
2268         }
2269
2270         cache->next_migration = NULL;
2271
2272         cache->need_tick_bio = true;
2273         cache->sized = false;
2274         cache->commit_requested = false;
2275         cache->loaded_mappings = false;
2276         cache->loaded_discards = false;
2277
2278         load_stats(cache);
2279
2280         atomic_set(&cache->stats.demotion, 0);
2281         atomic_set(&cache->stats.promotion, 0);
2282         atomic_set(&cache->stats.copies_avoided, 0);
2283         atomic_set(&cache->stats.cache_cell_clash, 0);
2284         atomic_set(&cache->stats.commit_count, 0);
2285         atomic_set(&cache->stats.discard_count, 0);
2286
2287         *result = cache;
2288         return 0;
2289
2290 bad:
2291         destroy(cache);
2292         return r;
2293 }
2294
2295 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2296 {
2297         unsigned i;
2298         const char **copy;
2299
2300         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2301         if (!copy)
2302                 return -ENOMEM;
2303         for (i = 0; i < argc; i++) {
2304                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2305                 if (!copy[i]) {
2306                         while (i--)
2307                                 kfree(copy[i]);
2308                         kfree(copy);
2309                         return -ENOMEM;
2310                 }
2311         }
2312
2313         cache->nr_ctr_args = argc;
2314         cache->ctr_args = copy;
2315
2316         return 0;
2317 }
2318
2319 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2320 {
2321         int r = -EINVAL;
2322         struct cache_args *ca;
2323         struct cache *cache = NULL;
2324
2325         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2326         if (!ca) {
2327                 ti->error = "Error allocating memory for cache";
2328                 return -ENOMEM;
2329         }
2330         ca->ti = ti;
2331
2332         r = parse_cache_args(ca, argc, argv, &ti->error);
2333         if (r)
2334                 goto out;
2335
2336         r = cache_create(ca, &cache);
2337         if (r)
2338                 goto out;
2339
2340         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2341         if (r) {
2342                 destroy(cache);
2343                 goto out;
2344         }
2345
2346         ti->private = cache;
2347
2348 out:
2349         destroy_cache_args(ca);
2350         return r;
2351 }
2352
2353 static int cache_map(struct dm_target *ti, struct bio *bio)
2354 {
2355         struct cache *cache = ti->private;
2356
2357         int r;
2358         dm_oblock_t block = get_bio_block(cache, bio);
2359         size_t pb_data_size = get_per_bio_data_size(cache);
2360         bool can_migrate = false;
2361         bool discarded_block;
2362         struct dm_bio_prison_cell *cell;
2363         struct policy_result lookup_result;
2364         struct per_bio_data *pb;
2365
2366         if (from_oblock(block) > from_oblock(cache->origin_blocks)) {
2367                 /*
2368                  * This can only occur if the io goes to a partial block at
2369                  * the end of the origin device.  We don't cache these.
2370                  * Just remap to the origin and carry on.
2371                  */
2372                 remap_to_origin_clear_discard(cache, bio, block);
2373                 return DM_MAPIO_REMAPPED;
2374         }
2375
2376         pb = init_per_bio_data(bio, pb_data_size);
2377
2378         if (bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD)) {
2379                 defer_bio(cache, bio);
2380                 return DM_MAPIO_SUBMITTED;
2381         }
2382
2383         /*
2384          * Check to see if that block is currently migrating.
2385          */
2386         cell = alloc_prison_cell(cache);
2387         if (!cell) {
2388                 defer_bio(cache, bio);
2389                 return DM_MAPIO_SUBMITTED;
2390         }
2391
2392         r = bio_detain(cache, block, bio, cell,
2393                        (cell_free_fn) free_prison_cell,
2394                        cache, &cell);
2395         if (r) {
2396                 if (r < 0)
2397                         defer_bio(cache, bio);
2398
2399                 return DM_MAPIO_SUBMITTED;
2400         }
2401
2402         discarded_block = is_discarded_oblock(cache, block);
2403
2404         r = policy_map(cache->policy, block, false, can_migrate, discarded_block,
2405                        bio, &lookup_result);
2406         if (r == -EWOULDBLOCK) {
2407                 cell_defer(cache, cell, true);
2408                 return DM_MAPIO_SUBMITTED;
2409
2410         } else if (r) {
2411                 DMERR_LIMIT("Unexpected return from cache replacement policy: %d", r);
2412                 bio_io_error(bio);
2413                 return DM_MAPIO_SUBMITTED;
2414         }
2415
2416         r = DM_MAPIO_REMAPPED;
2417         switch (lookup_result.op) {
2418         case POLICY_HIT:
2419                 if (passthrough_mode(&cache->features)) {
2420                         if (bio_data_dir(bio) == WRITE) {
2421                                 /*
2422                                  * We need to invalidate this block, so
2423                                  * defer for the worker thread.
2424                                  */
2425                                 cell_defer(cache, cell, true);
2426                                 r = DM_MAPIO_SUBMITTED;
2427
2428                         } else {
2429                                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2430                                 inc_miss_counter(cache, bio);
2431                                 remap_to_origin_clear_discard(cache, bio, block);
2432
2433                                 cell_defer(cache, cell, false);
2434                         }
2435
2436                 } else {
2437                         inc_hit_counter(cache, bio);
2438
2439                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
2440                             !is_dirty(cache, lookup_result.cblock))
2441                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
2442                         else
2443                                 remap_to_cache_dirty(cache, bio, block, lookup_result.cblock);
2444
2445                         cell_defer(cache, cell, false);
2446                 }
2447                 break;
2448
2449         case POLICY_MISS:
2450                 inc_miss_counter(cache, bio);
2451                 pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
2452
2453                 if (pb->req_nr != 0) {
2454                         /*
2455                          * This is a duplicate writethrough io that is no
2456                          * longer needed because the block has been demoted.
2457                          */
2458                         bio_endio(bio, 0);
2459                         cell_defer(cache, cell, false);
2460                         return DM_MAPIO_SUBMITTED;
2461                 } else {
2462                         remap_to_origin_clear_discard(cache, bio, block);
2463                         cell_defer(cache, cell, false);
2464                 }
2465                 break;
2466
2467         default:
2468                 DMERR_LIMIT("%s: erroring bio: unknown policy op: %u", __func__,
2469                             (unsigned) lookup_result.op);
2470                 bio_io_error(bio);
2471                 r = DM_MAPIO_SUBMITTED;
2472         }
2473
2474         return r;
2475 }
2476
2477 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
2478 {
2479         struct cache *cache = ti->private;
2480         unsigned long flags;
2481         size_t pb_data_size = get_per_bio_data_size(cache);
2482         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
2483
2484         if (pb->tick) {
2485                 policy_tick(cache->policy);
2486
2487                 spin_lock_irqsave(&cache->lock, flags);
2488                 cache->need_tick_bio = true;
2489                 spin_unlock_irqrestore(&cache->lock, flags);
2490         }
2491
2492         check_for_quiesced_migrations(cache, pb);
2493
2494         return 0;
2495 }
2496
2497 static int write_dirty_bitset(struct cache *cache)
2498 {
2499         unsigned i, r;
2500
2501         for (i = 0; i < from_cblock(cache->cache_size); i++) {
2502                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
2503                                        is_dirty(cache, to_cblock(i)));
2504                 if (r)
2505                         return r;
2506         }
2507
2508         return 0;
2509 }
2510
2511 static int write_discard_bitset(struct cache *cache)
2512 {
2513         unsigned i, r;
2514
2515         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2516                                            cache->discard_nr_blocks);
2517         if (r) {
2518                 DMERR("could not resize on-disk discard bitset");
2519                 return r;
2520         }
2521
2522         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2523                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2524                                          is_discarded(cache, to_dblock(i)));
2525                 if (r)
2526                         return r;
2527         }
2528
2529         return 0;
2530 }
2531
2532 static int save_hint(void *context, dm_cblock_t cblock, dm_oblock_t oblock,
2533                      uint32_t hint)
2534 {
2535         struct cache *cache = context;
2536         return dm_cache_save_hint(cache->cmd, cblock, hint);
2537 }
2538
2539 static int write_hints(struct cache *cache)
2540 {
2541         int r;
2542
2543         r = dm_cache_begin_hints(cache->cmd, cache->policy);
2544         if (r) {
2545                 DMERR("dm_cache_begin_hints failed");
2546                 return r;
2547         }
2548
2549         r = policy_walk_mappings(cache->policy, save_hint, cache);
2550         if (r)
2551                 DMERR("policy_walk_mappings failed");
2552
2553         return r;
2554 }
2555
2556 /*
2557  * returns true on success
2558  */
2559 static bool sync_metadata(struct cache *cache)
2560 {
2561         int r1, r2, r3, r4;
2562
2563         r1 = write_dirty_bitset(cache);
2564         if (r1)
2565                 DMERR("could not write dirty bitset");
2566
2567         r2 = write_discard_bitset(cache);
2568         if (r2)
2569                 DMERR("could not write discard bitset");
2570
2571         save_stats(cache);
2572
2573         r3 = write_hints(cache);
2574         if (r3)
2575                 DMERR("could not write hints");
2576
2577         /*
2578          * If writing the above metadata failed, we still commit, but don't
2579          * set the clean shutdown flag.  This will effectively force every
2580          * dirty bit to be set on reload.
2581          */
2582         r4 = dm_cache_commit(cache->cmd, !r1 && !r2 && !r3);
2583         if (r4)
2584                 DMERR("could not write cache metadata.  Data loss may occur.");
2585
2586         return !r1 && !r2 && !r3 && !r4;
2587 }
2588
2589 static void cache_postsuspend(struct dm_target *ti)
2590 {
2591         struct cache *cache = ti->private;
2592
2593         start_quiescing(cache);
2594         wait_for_migrations(cache);
2595         stop_worker(cache);
2596         requeue_deferred_io(cache);
2597         stop_quiescing(cache);
2598
2599         (void) sync_metadata(cache);
2600 }
2601
2602 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2603                         bool dirty, uint32_t hint, bool hint_valid)
2604 {
2605         int r;
2606         struct cache *cache = context;
2607
2608         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
2609         if (r)
2610                 return r;
2611
2612         if (dirty)
2613                 set_dirty(cache, oblock, cblock);
2614         else
2615                 clear_dirty(cache, oblock, cblock);
2616
2617         return 0;
2618 }
2619
2620 static int load_discard(void *context, sector_t discard_block_size,
2621                         dm_dblock_t dblock, bool discard)
2622 {
2623         struct cache *cache = context;
2624
2625         /* FIXME: handle mis-matched block size */
2626
2627         if (discard)
2628                 set_discard(cache, dblock);
2629         else
2630                 clear_discard(cache, dblock);
2631
2632         return 0;
2633 }
2634
2635 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2636 {
2637         sector_t size = get_dev_size(cache->cache_dev);
2638         (void) sector_div(size, cache->sectors_per_block);
2639         return to_cblock(size);
2640 }
2641
2642 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2643 {
2644         if (from_cblock(new_size) > from_cblock(cache->cache_size))
2645                 return true;
2646
2647         /*
2648          * We can't drop a dirty block when shrinking the cache.
2649          */
2650         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
2651                 new_size = to_cblock(from_cblock(new_size) + 1);
2652                 if (is_dirty(cache, new_size)) {
2653                         DMERR("unable to shrink cache; cache block %llu is dirty",
2654                               (unsigned long long) from_cblock(new_size));
2655                         return false;
2656                 }
2657         }
2658
2659         return true;
2660 }
2661
2662 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
2663 {
2664         int r;
2665
2666         r = dm_cache_resize(cache->cmd, cache->cache_size);
2667         if (r) {
2668                 DMERR("could not resize cache metadata");
2669                 return r;
2670         }
2671
2672         cache->cache_size = new_size;
2673
2674         return 0;
2675 }
2676
2677 static int cache_preresume(struct dm_target *ti)
2678 {
2679         int r = 0;
2680         struct cache *cache = ti->private;
2681         dm_cblock_t csize = get_cache_dev_size(cache);
2682
2683         /*
2684          * Check to see if the cache has resized.
2685          */
2686         if (!cache->sized) {
2687                 r = resize_cache_dev(cache, csize);
2688                 if (r)
2689                         return r;
2690
2691                 cache->sized = true;
2692
2693         } else if (csize != cache->cache_size) {
2694                 if (!can_resize(cache, csize))
2695                         return -EINVAL;
2696
2697                 r = resize_cache_dev(cache, csize);
2698                 if (r)
2699                         return r;
2700         }
2701
2702         if (!cache->loaded_mappings) {
2703                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
2704                                            load_mapping, cache);
2705                 if (r) {
2706                         DMERR("could not load cache mappings");
2707                         return r;
2708                 }
2709
2710                 cache->loaded_mappings = true;
2711         }
2712
2713         if (!cache->loaded_discards) {
2714                 r = dm_cache_load_discards(cache->cmd, load_discard, cache);
2715                 if (r) {
2716                         DMERR("could not load origin discards");
2717                         return r;
2718                 }
2719
2720                 cache->loaded_discards = true;
2721         }
2722
2723         return r;
2724 }
2725
2726 static void cache_resume(struct dm_target *ti)
2727 {
2728         struct cache *cache = ti->private;
2729
2730         cache->need_tick_bio = true;
2731         do_waker(&cache->waker.work);
2732 }
2733
2734 /*
2735  * Status format:
2736  *
2737  * <#used metadata blocks>/<#total metadata blocks>
2738  * <#read hits> <#read misses> <#write hits> <#write misses>
2739  * <#demotions> <#promotions> <#blocks in cache> <#dirty>
2740  * <#features> <features>*
2741  * <#core args> <core args>
2742  * <#policy args> <policy args>*
2743  */
2744 static void cache_status(struct dm_target *ti, status_type_t type,
2745                          unsigned status_flags, char *result, unsigned maxlen)
2746 {
2747         int r = 0;
2748         unsigned i;
2749         ssize_t sz = 0;
2750         dm_block_t nr_free_blocks_metadata = 0;
2751         dm_block_t nr_blocks_metadata = 0;
2752         char buf[BDEVNAME_SIZE];
2753         struct cache *cache = ti->private;
2754         dm_cblock_t residency;
2755
2756         switch (type) {
2757         case STATUSTYPE_INFO:
2758                 /* Commit to ensure statistics aren't out-of-date */
2759                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti)) {
2760                         r = dm_cache_commit(cache->cmd, false);
2761                         if (r)
2762                                 DMERR("could not commit metadata for accurate status");
2763                 }
2764
2765                 r = dm_cache_get_free_metadata_block_count(cache->cmd,
2766                                                            &nr_free_blocks_metadata);
2767                 if (r) {
2768                         DMERR("could not get metadata free block count");
2769                         goto err;
2770                 }
2771
2772                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
2773                 if (r) {
2774                         DMERR("could not get metadata device size");
2775                         goto err;
2776                 }
2777
2778                 residency = policy_residency(cache->policy);
2779
2780                 DMEMIT("%llu/%llu %u %u %u %u %u %u %llu %u ",
2781                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
2782                        (unsigned long long)nr_blocks_metadata,
2783                        (unsigned) atomic_read(&cache->stats.read_hit),
2784                        (unsigned) atomic_read(&cache->stats.read_miss),
2785                        (unsigned) atomic_read(&cache->stats.write_hit),
2786                        (unsigned) atomic_read(&cache->stats.write_miss),
2787                        (unsigned) atomic_read(&cache->stats.demotion),
2788                        (unsigned) atomic_read(&cache->stats.promotion),
2789                        (unsigned long long) from_cblock(residency),
2790                        cache->nr_dirty);
2791
2792                 if (writethrough_mode(&cache->features))
2793                         DMEMIT("1 writethrough ");
2794
2795                 else if (passthrough_mode(&cache->features))
2796                         DMEMIT("1 passthrough ");
2797
2798                 else if (writeback_mode(&cache->features))
2799                         DMEMIT("1 writeback ");
2800
2801                 else {
2802                         DMERR("internal error: unknown io mode: %d", (int) cache->features.io_mode);
2803                         goto err;
2804                 }
2805
2806                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
2807                 if (sz < maxlen) {
2808                         r = policy_emit_config_values(cache->policy, result + sz, maxlen - sz);
2809                         if (r)
2810                                 DMERR("policy_emit_config_values returned %d", r);
2811                 }
2812
2813                 break;
2814
2815         case STATUSTYPE_TABLE:
2816                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
2817                 DMEMIT("%s ", buf);
2818                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
2819                 DMEMIT("%s ", buf);
2820                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
2821                 DMEMIT("%s", buf);
2822
2823                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
2824                         DMEMIT(" %s", cache->ctr_args[i]);
2825                 if (cache->nr_ctr_args)
2826                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
2827         }
2828
2829         return;
2830
2831 err:
2832         DMEMIT("Error");
2833 }
2834
2835 /*
2836  * Supports <key> <value>.
2837  *
2838  * The key migration_threshold is supported by the cache target core.
2839  */
2840 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
2841 {
2842         struct cache *cache = ti->private;
2843
2844         if (argc != 2)
2845                 return -EINVAL;
2846
2847         return set_config_value(cache, argv[0], argv[1]);
2848 }
2849
2850 static int cache_iterate_devices(struct dm_target *ti,
2851                                  iterate_devices_callout_fn fn, void *data)
2852 {
2853         int r = 0;
2854         struct cache *cache = ti->private;
2855
2856         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
2857         if (!r)
2858                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
2859
2860         return r;
2861 }
2862
2863 /*
2864  * We assume I/O is going to the origin (which is the volume
2865  * more likely to have restrictions e.g. by being striped).
2866  * (Looking up the exact location of the data would be expensive
2867  * and could always be out of date by the time the bio is submitted.)
2868  */
2869 static int cache_bvec_merge(struct dm_target *ti,
2870                             struct bvec_merge_data *bvm,
2871                             struct bio_vec *biovec, int max_size)
2872 {
2873         struct cache *cache = ti->private;
2874         struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
2875
2876         if (!q->merge_bvec_fn)
2877                 return max_size;
2878
2879         bvm->bi_bdev = cache->origin_dev->bdev;
2880         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
2881 }
2882
2883 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
2884 {
2885         /*
2886          * FIXME: these limits may be incompatible with the cache device
2887          */
2888         limits->max_discard_sectors = cache->discard_block_size * 1024;
2889         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
2890 }
2891
2892 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
2893 {
2894         struct cache *cache = ti->private;
2895         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
2896
2897         /*
2898          * If the system-determined stacked limits are compatible with the
2899          * cache's blocksize (io_opt is a factor) do not override them.
2900          */
2901         if (io_opt_sectors < cache->sectors_per_block ||
2902             do_div(io_opt_sectors, cache->sectors_per_block)) {
2903                 blk_limits_io_min(limits, 0);
2904                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
2905         }
2906         set_discard_limits(cache, limits);
2907 }
2908
2909 /*----------------------------------------------------------------*/
2910
2911 static struct target_type cache_target = {
2912         .name = "cache",
2913         .version = {1, 2, 0},
2914         .module = THIS_MODULE,
2915         .ctr = cache_ctr,
2916         .dtr = cache_dtr,
2917         .map = cache_map,
2918         .end_io = cache_end_io,
2919         .postsuspend = cache_postsuspend,
2920         .preresume = cache_preresume,
2921         .resume = cache_resume,
2922         .status = cache_status,
2923         .message = cache_message,
2924         .iterate_devices = cache_iterate_devices,
2925         .merge = cache_bvec_merge,
2926         .io_hints = cache_io_hints,
2927 };
2928
2929 static int __init dm_cache_init(void)
2930 {
2931         int r;
2932
2933         r = dm_register_target(&cache_target);
2934         if (r) {
2935                 DMERR("cache target registration failed: %d", r);
2936                 return r;
2937         }
2938
2939         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
2940         if (!migration_cache) {
2941                 dm_unregister_target(&cache_target);
2942                 return -ENOMEM;
2943         }
2944
2945         return 0;
2946 }
2947
2948 static void __exit dm_cache_exit(void)
2949 {
2950         dm_unregister_target(&cache_target);
2951         kmem_cache_destroy(migration_cache);
2952 }
2953
2954 module_init(dm_cache_init);
2955 module_exit(dm_cache_exit);
2956
2957 MODULE_DESCRIPTION(DM_NAME " cache target");
2958 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
2959 MODULE_LICENSE("GPL");