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/jiffies.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/vmalloc.h>
  20
  21 #define DM_MSG_PREFIX "cache"
  22
  23 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  24         "A percentage of time allocated for copying to and/or from cache");
  25
  26 /*----------------------------------------------------------------*/
  27
  28 #define IOT_RESOLUTION 4
  29
  30 struct io_tracker {
  31         spinlock_t lock;
  32
  33         /*
  34          * Sectors of in-flight IO.
  35          */
  36         sector_t in_flight;
  37
  38         /*
  39          * The time, in jiffies, when this device became idle (if it is
  40          * indeed idle).
  41          */
  42         unsigned long idle_time;
  43         unsigned long last_update_time;
  44 };
  45
  46 static void iot_init(struct io_tracker *iot)
  47 {
  48         spin_lock_init(&iot->lock);
  49         iot->in_flight = 0ul;
  50         iot->idle_time = 0ul;
  51         iot->last_update_time = jiffies;
  52 }
  53
  54 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  55 {
  56         if (iot->in_flight)
  57                 return false;
  58
  59         return time_after(jiffies, iot->idle_time + jifs);
  60 }
  61
  62 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  63 {
  64         bool r;
  65         unsigned long flags;
  66
  67         spin_lock_irqsave(&iot->lock, flags);
  68         r = __iot_idle_for(iot, jifs);
  69         spin_unlock_irqrestore(&iot->lock, flags);
  70
  71         return r;
  72 }
  73
  74 static void iot_io_begin(struct io_tracker *iot, sector_t len)
  75 {
  76         unsigned long flags;
  77
  78         spin_lock_irqsave(&iot->lock, flags);
  79         iot->in_flight += len;
  80         spin_unlock_irqrestore(&iot->lock, flags);
  81 }
  82
  83 static void __iot_io_end(struct io_tracker *iot, sector_t len)
  84 {
  85         iot->in_flight -= len;
  86         if (!iot->in_flight)
  87                 iot->idle_time = jiffies;
  88 }
  89
  90 static void iot_io_end(struct io_tracker *iot, sector_t len)
  91 {
  92         unsigned long flags;
  93
  94         spin_lock_irqsave(&iot->lock, flags);
  95         __iot_io_end(iot, len);
  96         spin_unlock_irqrestore(&iot->lock, flags);
  97 }
  98
  99 /*----------------------------------------------------------------*/
 100
 101 /*
 102  * Glossary:
 103  *
 104  * oblock: index of an origin block
 105  * cblock: index of a cache block
 106  * promotion: movement of a block from origin to cache
 107  * demotion: movement of a block from cache to origin
 108  * migration: movement of a block between the origin and cache device,
 109  *            either direction
 110  */
 111
 112 /*----------------------------------------------------------------*/
 113
 114 /*
 115  * There are a couple of places where we let a bio run, but want to do some
 116  * work before calling its endio function.  We do this by temporarily
 117  * changing the endio fn.
 118  */
 119 struct dm_hook_info {
 120         bio_end_io_t *bi_end_io;
 121         void *bi_private;
 122 };
 123
 124 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 125                         bio_end_io_t *bi_end_io, void *bi_private)
 126 {
 127         h->bi_end_io = bio->bi_end_io;
 128         h->bi_private = bio->bi_private;
 129
 130         bio->bi_end_io = bi_end_io;
 131         bio->bi_private = bi_private;
 132 }
 133
 134 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 135 {
 136         bio->bi_end_io = h->bi_end_io;
 137         bio->bi_private = h->bi_private;
 138 }
 139
 140 /*----------------------------------------------------------------*/
 141
 142 #define MIGRATION_POOL_SIZE 128
 143 #define COMMIT_PERIOD HZ
 144 #define MIGRATION_COUNT_WINDOW 10
 145
 146 /*
 147  * The block size of the device holding cache data must be
 148  * between 32KB and 1GB.
 149  */
 150 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 151 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 152
 153 enum cache_metadata_mode {
 154         CM_WRITE,               /* metadata may be changed */
 155         CM_READ_ONLY,           /* metadata may not be changed */
 156         CM_FAIL
 157 };
 158
 159 enum cache_io_mode {
 160         /*
 161          * Data is written to cached blocks only.  These blocks are marked
 162          * dirty.  If you lose the cache device you will lose data.
 163          * Potential performance increase for both reads and writes.
 164          */
 165         CM_IO_WRITEBACK,
 166
 167         /*
 168          * Data is written to both cache and origin.  Blocks are never
 169          * dirty.  Potential performance benfit for reads only.
 170          */
 171         CM_IO_WRITETHROUGH,
 172
 173         /*
 174          * A degraded mode useful for various cache coherency situations
 175          * (eg, rolling back snapshots).  Reads and writes always go to the
 176          * origin.  If a write goes to a cached oblock, then the cache
 177          * block is invalidated.
 178          */
 179         CM_IO_PASSTHROUGH
 180 };
 181
 182 struct cache_features {
 183         enum cache_metadata_mode mode;
 184         enum cache_io_mode io_mode;
 185 };
 186
 187 struct cache_stats {
 188         atomic_t read_hit;
 189         atomic_t read_miss;
 190         atomic_t write_hit;
 191         atomic_t write_miss;
 192         atomic_t demotion;
 193         atomic_t promotion;
 194         atomic_t copies_avoided;
 195         atomic_t cache_cell_clash;
 196         atomic_t commit_count;
 197         atomic_t discard_count;
 198 };
 199
 200 /*
 201  * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
 202  * the one-past-the-end value.
 203  */
 204 struct cblock_range {
 205         dm_cblock_t begin;
 206         dm_cblock_t end;
 207 };
 208
 209 struct invalidation_request {
 210         struct list_head list;
 211         struct cblock_range *cblocks;
 212
 213         atomic_t complete;
 214         int err;
 215
 216         wait_queue_head_t result_wait;
 217 };
 218
 219 struct cache {
 220         struct dm_target *ti;
 221         struct dm_target_callbacks callbacks;
 222
 223         struct dm_cache_metadata *cmd;
 224
 225         /*
 226          * Metadata is written to this device.
 227          */
 228         struct dm_dev *metadata_dev;
 229
 230         /*
 231          * The slower of the two data devices.  Typically a spindle.
 232          */
 233         struct dm_dev *origin_dev;
 234
 235         /*
 236          * The faster of the two data devices.  Typically an SSD.
 237          */
 238         struct dm_dev *cache_dev;
 239
 240         /*
 241          * Size of the origin device in _complete_ blocks and native sectors.
 242          */
 243         dm_oblock_t origin_blocks;
 244         sector_t origin_sectors;
 245
 246         /*
 247          * Size of the cache device in blocks.
 248          */
 249         dm_cblock_t cache_size;
 250
 251         /*
 252          * Fields for converting from sectors to blocks.
 253          */
 254         uint32_t sectors_per_block;
 255         int sectors_per_block_shift;
 256
 257         spinlock_t lock;
 258         struct list_head deferred_cells;
 259         struct bio_list deferred_bios;
 260         struct bio_list deferred_flush_bios;
 261         struct bio_list deferred_writethrough_bios;
 262         struct list_head quiesced_migrations;
 263         struct list_head completed_migrations;
 264         struct list_head need_commit_migrations;
 265         sector_t migration_threshold;
 266         wait_queue_head_t migration_wait;
 267         atomic_t nr_allocated_migrations;
 268
 269         /*
 270          * The number of in flight migrations that are performing
 271          * background io. eg, promotion, writeback.
 272          */
 273         atomic_t nr_io_migrations;
 274
 275         wait_queue_head_t quiescing_wait;
 276         atomic_t quiescing;
 277         atomic_t quiescing_ack;
 278
 279         /*
 280          * cache_size entries, dirty if set
 281          */
 282         atomic_t nr_dirty;
 283         unsigned long *dirty_bitset;
 284
 285         /*
 286          * origin_blocks entries, discarded if set.
 287          */
 288         dm_dblock_t discard_nr_blocks;
 289         unsigned long *discard_bitset;
 290         uint32_t discard_block_size; /* a power of 2 times sectors per block */
 291
 292         /*
 293          * Rather than reconstructing the table line for the status we just
 294          * save it and regurgitate.
 295          */
 296         unsigned nr_ctr_args;
 297         const char **ctr_args;
 298
 299         struct dm_kcopyd_client *copier;
 300         struct workqueue_struct *wq;
 301         struct work_struct worker;
 302
 303         struct delayed_work waker;
 304         unsigned long last_commit_jiffies;
 305
 306         struct dm_bio_prison *prison;
 307         struct dm_deferred_set *all_io_ds;
 308
 309         mempool_t *migration_pool;
 310
 311         struct dm_cache_policy *policy;
 312         unsigned policy_nr_args;
 313
 314         bool need_tick_bio:1;
 315         bool sized:1;
 316         bool invalidate:1;
 317         bool commit_requested:1;
 318         bool loaded_mappings:1;
 319         bool loaded_discards:1;
 320
 321         /*
 322          * Cache features such as write-through.
 323          */
 324         struct cache_features features;
 325
 326         struct cache_stats stats;
 327
 328         /*
 329          * Invalidation fields.
 330          */
 331         spinlock_t invalidation_lock;
 332         struct list_head invalidation_requests;
 333
 334         struct io_tracker origin_tracker;
 335 };
 336
 337 struct per_bio_data {
 338         bool tick:1;
 339         unsigned req_nr:2;
 340         struct dm_deferred_entry *all_io_entry;
 341         struct dm_hook_info hook_info;
 342         sector_t len;
 343
 344         /*
 345          * writethrough fields.  These MUST remain at the end of this
 346          * structure and the 'cache' member must be the first as it
 347          * is used to determine the offset of the writethrough fields.
 348          */
 349         struct cache *cache;
 350         dm_cblock_t cblock;
 351         struct dm_bio_details bio_details;
 352 };
 353
 354 struct dm_cache_migration {
 355         struct list_head list;
 356         struct cache *cache;
 357
 358         unsigned long start_jiffies;
 359         dm_oblock_t old_oblock;
 360         dm_oblock_t new_oblock;
 361         dm_cblock_t cblock;
 362
 363         bool err:1;
 364         bool discard:1;
 365         bool writeback:1;
 366         bool demote:1;
 367         bool promote:1;
 368         bool requeue_holder:1;
 369         bool invalidate:1;
 370
 371         struct dm_bio_prison_cell *old_ocell;
 372         struct dm_bio_prison_cell *new_ocell;
 373 };
 374
 375 /*
 376  * Processing a bio in the worker thread may require these memory
 377  * allocations.  We prealloc to avoid deadlocks (the same worker thread
 378  * frees them back to the mempool).
 379  */
 380 struct prealloc {
 381         struct dm_cache_migration *mg;
 382         struct dm_bio_prison_cell *cell1;
 383         struct dm_bio_prison_cell *cell2;
 384 };
 385
 386 static enum cache_metadata_mode get_cache_mode(struct cache *cache);
 387
 388 static void wake_worker(struct cache *cache)
 389 {
 390         queue_work(cache->wq, &cache->worker);
 391 }
 392
 393 /*----------------------------------------------------------------*/
 394
 395 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
 396 {
 397         /* FIXME: change to use a local slab. */
 398         return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
 399 }
 400
 401 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
 402 {
 403         dm_bio_prison_free_cell(cache->prison, cell);
 404 }
 405
 406 static struct dm_cache_migration *alloc_migration(struct cache *cache)
 407 {
 408         struct dm_cache_migration *mg;
 409
 410         mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
 411         if (mg) {
 412                 mg->cache = cache;
 413                 atomic_inc(&mg->cache->nr_allocated_migrations);
 414         }
 415
 416         return mg;
 417 }
 418
 419 static void free_migration(struct dm_cache_migration *mg)
 420 {
 421         struct cache *cache = mg->cache;
 422
 423         if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 424                 wake_up(&cache->migration_wait);
 425
 426         mempool_free(mg, cache->migration_pool);
 427 }
 428
 429 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
 430 {
 431         if (!p->mg) {
 432                 p->mg = alloc_migration(cache);
 433                 if (!p->mg)
 434                         return -ENOMEM;
 435         }
 436
 437         if (!p->cell1) {
 438                 p->cell1 = alloc_prison_cell(cache);
 439                 if (!p->cell1)
 440                         return -ENOMEM;
 441         }
 442
 443         if (!p->cell2) {
 444                 p->cell2 = alloc_prison_cell(cache);
 445                 if (!p->cell2)
 446                         return -ENOMEM;
 447         }
 448
 449         return 0;
 450 }
 451
 452 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
 453 {
 454         if (p->cell2)
 455                 free_prison_cell(cache, p->cell2);
 456
 457         if (p->cell1)
 458                 free_prison_cell(cache, p->cell1);
 459
 460         if (p->mg)
 461                 free_migration(p->mg);
 462 }
 463
 464 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
 465 {
 466         struct dm_cache_migration *mg = p->mg;
 467
 468         BUG_ON(!mg);
 469         p->mg = NULL;
 470
 471         return mg;
 472 }
 473
 474 /*
 475  * You must have a cell within the prealloc struct to return.  If not this
 476  * function will BUG() rather than returning NULL.
 477  */
 478 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
 479 {
 480         struct dm_bio_prison_cell *r = NULL;
 481
 482         if (p->cell1) {
 483                 r = p->cell1;
 484                 p->cell1 = NULL;
 485
 486         } else if (p->cell2) {
 487                 r = p->cell2;
 488                 p->cell2 = NULL;
 489         } else
 490                 BUG();
 491
 492         return r;
 493 }
 494
 495 /*
 496  * You can't have more than two cells in a prealloc struct.  BUG() will be
 497  * called if you try and overfill.
 498  */
 499 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
 500 {
 501         if (!p->cell2)
 502                 p->cell2 = cell;
 503
 504         else if (!p->cell1)
 505                 p->cell1 = cell;
 506
 507         else
 508                 BUG();
 509 }
 510
 511 /*----------------------------------------------------------------*/
 512
 513 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key *key)
 514 {
 515         key->virtual = 0;
 516         key->dev = 0;
 517         key->block_begin = from_oblock(begin);
 518         key->block_end = from_oblock(end);
 519 }
 520
 521 /*
 522  * The caller hands in a preallocated cell, and a free function for it.
 523  * The cell will be freed if there's an error, or if it wasn't used because
 524  * a cell with that key already exists.
 525  */
 526 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
 527
 528 static int bio_detain_range(struct cache *cache, dm_oblock_t oblock_begin, dm_oblock_t oblock_end,
 529                             struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 530                             cell_free_fn free_fn, void *free_context,
 531                             struct dm_bio_prison_cell **cell_result)
 532 {
 533         int r;
 534         struct dm_cell_key key;
 535
 536         build_key(oblock_begin, oblock_end, &key);
 537         r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
 538         if (r)
 539                 free_fn(free_context, cell_prealloc);
 540
 541         return r;
 542 }
 543
 544 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
 545                       struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 546                       cell_free_fn free_fn, void *free_context,
 547                       struct dm_bio_prison_cell **cell_result)
 548 {
 549         dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 550         return bio_detain_range(cache, oblock, end, bio,
 551                                 cell_prealloc, free_fn, free_context, cell_result);
 552 }
 553
 554 static int get_cell(struct cache *cache,
 555                     dm_oblock_t oblock,
 556                     struct prealloc *structs,
 557                     struct dm_bio_prison_cell **cell_result)
 558 {
 559         int r;
 560         struct dm_cell_key key;
 561         struct dm_bio_prison_cell *cell_prealloc;
 562
 563         cell_prealloc = prealloc_get_cell(structs);
 564
 565         build_key(oblock, to_oblock(from_oblock(oblock) + 1ULL), &key);
 566         r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
 567         if (r)
 568                 prealloc_put_cell(structs, cell_prealloc);
 569
 570         return r;
 571 }
 572
 573 /*----------------------------------------------------------------*/
 574
 575 static bool is_dirty(struct cache *cache, dm_cblock_t b)
 576 {
 577         return test_bit(from_cblock(b), cache->dirty_bitset);
 578 }
 579
 580 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 581 {
 582         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 583                 atomic_inc(&cache->nr_dirty);
 584                 policy_set_dirty(cache->policy, oblock);
 585         }
 586 }
 587
 588 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 589 {
 590         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 591                 policy_clear_dirty(cache->policy, oblock);
 592                 if (atomic_dec_return(&cache->nr_dirty) == 0)
 593                         dm_table_event(cache->ti->table);
 594         }
 595 }
 596
 597 /*----------------------------------------------------------------*/
 598
 599 static bool block_size_is_power_of_two(struct cache *cache)
 600 {
 601         return cache->sectors_per_block_shift >= 0;
 602 }
 603
 604 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
 605 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
 606 __always_inline
 607 #endif
 608 static dm_block_t block_div(dm_block_t b, uint32_t n)
 609 {
 610         do_div(b, n);
 611
 612         return b;
 613 }
 614
 615 static dm_block_t oblocks_per_dblock(struct cache *cache)
 616 {
 617         dm_block_t oblocks = cache->discard_block_size;
 618
 619         if (block_size_is_power_of_two(cache))
 620                 oblocks >>= cache->sectors_per_block_shift;
 621         else
 622                 oblocks = block_div(oblocks, cache->sectors_per_block);
 623
 624         return oblocks;
 625 }
 626
 627 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 628 {
 629         return to_dblock(block_div(from_oblock(oblock),
 630                                    oblocks_per_dblock(cache)));
 631 }
 632
 633 static dm_oblock_t dblock_to_oblock(struct cache *cache, dm_dblock_t dblock)
 634 {
 635         return to_oblock(from_dblock(dblock) * oblocks_per_dblock(cache));
 636 }
 637
 638 static void set_discard(struct cache *cache, dm_dblock_t b)
 639 {
 640         unsigned long flags;
 641
 642         BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 643         atomic_inc(&cache->stats.discard_count);
 644
 645         spin_lock_irqsave(&cache->lock, flags);
 646         set_bit(from_dblock(b), cache->discard_bitset);
 647         spin_unlock_irqrestore(&cache->lock, flags);
 648 }
 649
 650 static void clear_discard(struct cache *cache, dm_dblock_t b)
 651 {
 652         unsigned long flags;
 653
 654         spin_lock_irqsave(&cache->lock, flags);
 655         clear_bit(from_dblock(b), cache->discard_bitset);
 656         spin_unlock_irqrestore(&cache->lock, flags);
 657 }
 658
 659 static bool is_discarded(struct cache *cache, dm_dblock_t b)
 660 {
 661         int r;
 662         unsigned long flags;
 663
 664         spin_lock_irqsave(&cache->lock, flags);
 665         r = test_bit(from_dblock(b), cache->discard_bitset);
 666         spin_unlock_irqrestore(&cache->lock, flags);
 667
 668         return r;
 669 }
 670
 671 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 672 {
 673         int r;
 674         unsigned long flags;
 675
 676         spin_lock_irqsave(&cache->lock, flags);
 677         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 678                      cache->discard_bitset);
 679         spin_unlock_irqrestore(&cache->lock, flags);
 680
 681         return r;
 682 }
 683
 684 /*----------------------------------------------------------------*/
 685
 686 static void load_stats(struct cache *cache)
 687 {
 688         struct dm_cache_statistics stats;
 689
 690         dm_cache_metadata_get_stats(cache->cmd, &stats);
 691         atomic_set(&cache->stats.read_hit, stats.read_hits);
 692         atomic_set(&cache->stats.read_miss, stats.read_misses);
 693         atomic_set(&cache->stats.write_hit, stats.write_hits);
 694         atomic_set(&cache->stats.write_miss, stats.write_misses);
 695 }
 696
 697 static void save_stats(struct cache *cache)
 698 {
 699         struct dm_cache_statistics stats;
 700
 701         if (get_cache_mode(cache) >= CM_READ_ONLY)
 702                 return;
 703
 704         stats.read_hits = atomic_read(&cache->stats.read_hit);
 705         stats.read_misses = atomic_read(&cache->stats.read_miss);
 706         stats.write_hits = atomic_read(&cache->stats.write_hit);
 707         stats.write_misses = atomic_read(&cache->stats.write_miss);
 708
 709         dm_cache_metadata_set_stats(cache->cmd, &stats);
 710 }
 711
 712 /*----------------------------------------------------------------
 713  * Per bio data
 714  *--------------------------------------------------------------*/
 715
 716 /*
 717  * If using writeback, leave out struct per_bio_data's writethrough fields.
 718  */
 719 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
 720 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
 721
 722 static bool writethrough_mode(struct cache_features *f)
 723 {
 724         return f->io_mode == CM_IO_WRITETHROUGH;
 725 }
 726
 727 static bool writeback_mode(struct cache_features *f)
 728 {
 729         return f->io_mode == CM_IO_WRITEBACK;
 730 }
 731
 732 static bool passthrough_mode(struct cache_features *f)
 733 {
 734         return f->io_mode == CM_IO_PASSTHROUGH;
 735 }
 736
 737 static size_t get_per_bio_data_size(struct cache *cache)
 738 {
 739         return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
 740 }
 741
 742 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
 743 {
 744         struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
 745         BUG_ON(!pb);
 746         return pb;
 747 }
 748
 749 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
 750 {
 751         struct per_bio_data *pb = get_per_bio_data(bio, data_size);
 752
 753         pb->tick = false;
 754         pb->req_nr = dm_bio_get_target_bio_nr(bio);
 755         pb->all_io_entry = NULL;
 756         pb->len = 0;
 757
 758         return pb;
 759 }
 760
 761 /*----------------------------------------------------------------
 762  * Remapping
 763  *--------------------------------------------------------------*/
 764 static void remap_to_origin(struct cache *cache, struct bio *bio)
 765 {
 766         bio->bi_bdev = cache->origin_dev->bdev;
 767 }
 768
 769 static void remap_to_cache(struct cache *cache, struct bio *bio,
 770                            dm_cblock_t cblock)
 771 {
 772         sector_t bi_sector = bio->bi_iter.bi_sector;
 773         sector_t block = from_cblock(cblock);
 774
 775         bio->bi_bdev = cache->cache_dev->bdev;
 776         if (!block_size_is_power_of_two(cache))
 777                 bio->bi_iter.bi_sector =
 778                         (block * cache->sectors_per_block) +
 779                         sector_div(bi_sector, cache->sectors_per_block);
 780         else
 781                 bio->bi_iter.bi_sector =
 782                         (block << cache->sectors_per_block_shift) |
 783                         (bi_sector & (cache->sectors_per_block - 1));
 784 }
 785
 786 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 787 {
 788         unsigned long flags;
 789         size_t pb_data_size = get_per_bio_data_size(cache);
 790         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 791
 792         spin_lock_irqsave(&cache->lock, flags);
 793         if (cache->need_tick_bio &&
 794             !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
 795                 pb->tick = true;
 796                 cache->need_tick_bio = false;
 797         }
 798         spin_unlock_irqrestore(&cache->lock, flags);
 799 }
 800
 801 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 802                                   dm_oblock_t oblock)
 803 {
 804         check_if_tick_bio_needed(cache, bio);
 805         remap_to_origin(cache, bio);
 806         if (bio_data_dir(bio) == WRITE)
 807                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 808 }
 809
 810 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 811                                  dm_oblock_t oblock, dm_cblock_t cblock)
 812 {
 813         check_if_tick_bio_needed(cache, bio);
 814         remap_to_cache(cache, bio, cblock);
 815         if (bio_data_dir(bio) == WRITE) {
 816                 set_dirty(cache, oblock, cblock);
 817                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 818         }
 819 }
 820
 821 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 822 {
 823         sector_t block_nr = bio->bi_iter.bi_sector;
 824
 825         if (!block_size_is_power_of_two(cache))
 826                 (void) sector_div(block_nr, cache->sectors_per_block);
 827         else
 828                 block_nr >>= cache->sectors_per_block_shift;
 829
 830         return to_oblock(block_nr);
 831 }
 832
 833 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
 834 {
 835         return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
 836 }
 837
 838 /*
 839  * You must increment the deferred set whilst the prison cell is held.  To
 840  * encourage this, we ask for 'cell' to be passed in.
 841  */
 842 static void inc_ds(struct cache *cache, struct bio *bio,
 843                    struct dm_bio_prison_cell *cell)
 844 {
 845         size_t pb_data_size = get_per_bio_data_size(cache);
 846         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 847
 848         BUG_ON(!cell);
 849         BUG_ON(pb->all_io_entry);
 850
 851         pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
 852 }
 853
 854 static bool accountable_bio(struct cache *cache, struct bio *bio)
 855 {
 856         return ((bio->bi_bdev == cache->origin_dev->bdev) &&
 857                 !(bio->bi_rw & REQ_DISCARD));
 858 }
 859
 860 static void accounted_begin(struct cache *cache, struct bio *bio)
 861 {
 862         size_t pb_data_size = get_per_bio_data_size(cache);
 863         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 864
 865         if (accountable_bio(cache, bio)) {
 866                 pb->len = bio_sectors(bio);
 867                 iot_io_begin(&cache->origin_tracker, pb->len);
 868         }
 869 }
 870
 871 static void accounted_complete(struct cache *cache, struct bio *bio)
 872 {
 873         size_t pb_data_size = get_per_bio_data_size(cache);
 874         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 875
 876         iot_io_end(&cache->origin_tracker, pb->len);
 877 }
 878
 879 static void accounted_request(struct cache *cache, struct bio *bio)
 880 {
 881         accounted_begin(cache, bio);
 882         generic_make_request(bio);
 883 }
 884
 885 static void issue(struct cache *cache, struct bio *bio)
 886 {
 887         unsigned long flags;
 888
 889         if (!bio_triggers_commit(cache, bio)) {
 890                 accounted_request(cache, bio);
 891                 return;
 892         }
 893
 894         /*
 895          * Batch together any bios that trigger commits and then issue a
 896          * single commit for them in do_worker().
 897          */
 898         spin_lock_irqsave(&cache->lock, flags);
 899         cache->commit_requested = true;
 900         bio_list_add(&cache->deferred_flush_bios, bio);
 901         spin_unlock_irqrestore(&cache->lock, flags);
 902 }
 903
 904 static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
 905 {
 906         inc_ds(cache, bio, cell);
 907         issue(cache, bio);
 908 }
 909
 910 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
 911 {
 912         unsigned long flags;
 913
 914         spin_lock_irqsave(&cache->lock, flags);
 915         bio_list_add(&cache->deferred_writethrough_bios, bio);
 916         spin_unlock_irqrestore(&cache->lock, flags);
 917
 918         wake_worker(cache);
 919 }
 920
 921 static void writethrough_endio(struct bio *bio, int err)
 922 {
 923         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 924
 925         dm_unhook_bio(&pb->hook_info, bio);
 926
 927         if (err) {
 928                 bio_endio(bio, err);
 929                 return;
 930         }
 931
 932         dm_bio_restore(&pb->bio_details, bio);
 933         remap_to_cache(pb->cache, bio, pb->cblock);
 934
 935         /*
 936          * We can't issue this bio directly, since we're in interrupt
 937          * context.  So it gets put on a bio list for processing by the
 938          * worker thread.
 939          */
 940         defer_writethrough_bio(pb->cache, bio);
 941 }
 942
 943 /*
 944  * When running in writethrough mode we need to send writes to clean blocks
 945  * to both the cache and origin devices.  In future we'd like to clone the
 946  * bio and send them in parallel, but for now we're doing them in
 947  * series as this is easier.
 948  */
 949 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
 950                                        dm_oblock_t oblock, dm_cblock_t cblock)
 951 {
 952         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 953
 954         pb->cache = cache;
 955         pb->cblock = cblock;
 956         dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
 957         dm_bio_record(&pb->bio_details, bio);
 958
 959         remap_to_origin_clear_discard(pb->cache, bio, oblock);
 960 }
 961
 962 /*----------------------------------------------------------------
 963  * Failure modes
 964  *--------------------------------------------------------------*/
 965 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 966 {
 967         return cache->features.mode;
 968 }
 969
 970 static const char *cache_device_name(struct cache *cache)
 971 {
 972         return dm_device_name(dm_table_get_md(cache->ti->table));
 973 }
 974
 975 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 976 {
 977         const char *descs[] = {
 978                 "write",
 979                 "read-only",
 980                 "fail"
 981         };
 982
 983         dm_table_event(cache->ti->table);
 984         DMINFO("%s: switching cache to %s mode",
 985                cache_device_name(cache), descs[(int)mode]);
 986 }
 987
 988 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 989 {
 990         bool needs_check = dm_cache_metadata_needs_check(cache->cmd);
 991         enum cache_metadata_mode old_mode = get_cache_mode(cache);
 992
 993         if (new_mode == CM_WRITE && needs_check) {
 994                 DMERR("%s: unable to switch cache to write mode until repaired.",
 995                       cache_device_name(cache));
 996                 if (old_mode != new_mode)
 997                         new_mode = old_mode;
 998                 else
 999                         new_mode = CM_READ_ONLY;
1000         }
1001
1002         /* Never move out of fail mode */
1003         if (old_mode == CM_FAIL)
1004                 new_mode = CM_FAIL;
1005
1006         switch (new_mode) {
1007         case CM_FAIL:
1008         case CM_READ_ONLY:
1009                 dm_cache_metadata_set_read_only(cache->cmd);
1010                 break;
1011
1012         case CM_WRITE:
1013                 dm_cache_metadata_set_read_write(cache->cmd);
1014                 break;
1015         }
1016
1017         cache->features.mode = new_mode;
1018
1019         if (new_mode != old_mode)
1020                 notify_mode_switch(cache, new_mode);
1021 }
1022
1023 static void abort_transaction(struct cache *cache)
1024 {
1025         const char *dev_name = cache_device_name(cache);
1026
1027         if (get_cache_mode(cache) >= CM_READ_ONLY)
1028                 return;
1029
1030         if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1031                 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1032                 set_cache_mode(cache, CM_FAIL);
1033         }
1034
1035         DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1036         if (dm_cache_metadata_abort(cache->cmd)) {
1037                 DMERR("%s: failed to abort metadata transaction", dev_name);
1038                 set_cache_mode(cache, CM_FAIL);
1039         }
1040 }
1041
1042 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1043 {
1044         DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1045                     cache_device_name(cache), op, r);
1046         abort_transaction(cache);
1047         set_cache_mode(cache, CM_READ_ONLY);
1048 }
1049
1050 /*----------------------------------------------------------------
1051  * Migration processing
1052  *
1053  * Migration covers moving data from the origin device to the cache, or
1054  * vice versa.
1055  *--------------------------------------------------------------*/
1056 static void inc_io_migrations(struct cache *cache)
1057 {
1058         atomic_inc(&cache->nr_io_migrations);
1059 }
1060
1061 static void dec_io_migrations(struct cache *cache)
1062 {
1063         atomic_dec(&cache->nr_io_migrations);
1064 }
1065
1066 static bool discard_or_flush(struct bio *bio)
1067 {
1068         return bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD);
1069 }
1070
1071 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1072 {
1073         if (discard_or_flush(cell->holder)) {
1074                 /*
1075                  * We have to handle these bios individually.
1076                  */
1077                 dm_cell_release(cache->prison, cell, &cache->deferred_bios);
1078                 free_prison_cell(cache, cell);
1079         } else
1080                 list_add_tail(&cell->user_list, &cache->deferred_cells);
1081 }
1082
1083 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1084 {
1085         unsigned long flags;
1086
1087         if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1088                 /*
1089                  * There was no prisoner to promote to holder, the
1090                  * cell has been released.
1091                  */
1092                 free_prison_cell(cache, cell);
1093                 return;
1094         }
1095
1096         spin_lock_irqsave(&cache->lock, flags);
1097         __cell_defer(cache, cell);
1098         spin_unlock_irqrestore(&cache->lock, flags);
1099
1100         wake_worker(cache);
1101 }
1102
1103 static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1104 {
1105         dm_cell_error(cache->prison, cell, err);
1106         free_prison_cell(cache, cell);
1107 }
1108
1109 static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1110 {
1111         cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1112 }
1113
1114 static void free_io_migration(struct dm_cache_migration *mg)
1115 {
1116         dec_io_migrations(mg->cache);
1117         free_migration(mg);
1118         wake_worker(mg->cache);
1119 }
1120
1121 static void migration_failure(struct dm_cache_migration *mg)
1122 {
1123         struct cache *cache = mg->cache;
1124         const char *dev_name = cache_device_name(cache);
1125
1126         if (mg->writeback) {
1127                 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1128                 set_dirty(cache, mg->old_oblock, mg->cblock);
1129                 cell_defer(cache, mg->old_ocell, false);
1130
1131         } else if (mg->demote) {
1132                 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1133                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1134
1135                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1136                 if (mg->promote)
1137                         cell_defer(cache, mg->new_ocell, true);
1138         } else {
1139                 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1140                 policy_remove_mapping(cache->policy, mg->new_oblock);
1141                 cell_defer(cache, mg->new_ocell, true);
1142         }
1143
1144         free_io_migration(mg);
1145 }
1146
1147 static void migration_success_pre_commit(struct dm_cache_migration *mg)
1148 {
1149         int r;
1150         unsigned long flags;
1151         struct cache *cache = mg->cache;
1152
1153         if (mg->writeback) {
1154                 clear_dirty(cache, mg->old_oblock, mg->cblock);
1155                 cell_defer(cache, mg->old_ocell, false);
1156                 free_io_migration(mg);
1157                 return;
1158
1159         } else if (mg->demote) {
1160                 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1161                 if (r) {
1162                         DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1163                                     cache_device_name(cache));
1164                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1165                         policy_force_mapping(cache->policy, mg->new_oblock,
1166                                              mg->old_oblock);
1167                         if (mg->promote)
1168                                 cell_defer(cache, mg->new_ocell, true);
1169                         free_io_migration(mg);
1170                         return;
1171                 }
1172         } else {
1173                 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1174                 if (r) {
1175                         DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1176                                     cache_device_name(cache));
1177                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1178                         policy_remove_mapping(cache->policy, mg->new_oblock);
1179                         free_io_migration(mg);
1180                         return;
1181                 }
1182         }
1183
1184         spin_lock_irqsave(&cache->lock, flags);
1185         list_add_tail(&mg->list, &cache->need_commit_migrations);
1186         cache->commit_requested = true;
1187         spin_unlock_irqrestore(&cache->lock, flags);
1188 }
1189
1190 static void migration_success_post_commit(struct dm_cache_migration *mg)
1191 {
1192         unsigned long flags;
1193         struct cache *cache = mg->cache;
1194
1195         if (mg->writeback) {
1196                 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1197                              cache_device_name(cache));
1198                 return;
1199
1200         } else if (mg->demote) {
1201                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1202
1203                 if (mg->promote) {
1204                         mg->demote = false;
1205
1206                         spin_lock_irqsave(&cache->lock, flags);
1207                         list_add_tail(&mg->list, &cache->quiesced_migrations);
1208                         spin_unlock_irqrestore(&cache->lock, flags);
1209
1210                 } else {
1211                         if (mg->invalidate)
1212                                 policy_remove_mapping(cache->policy, mg->old_oblock);
1213                         free_io_migration(mg);
1214                 }
1215
1216         } else {
1217                 if (mg->requeue_holder) {
1218                         clear_dirty(cache, mg->new_oblock, mg->cblock);
1219                         cell_defer(cache, mg->new_ocell, true);
1220                 } else {
1221                         /*
1222                          * The block was promoted via an overwrite, so it's dirty.
1223                          */
1224                         set_dirty(cache, mg->new_oblock, mg->cblock);
1225                         bio_endio(mg->new_ocell->holder, 0);
1226                         cell_defer(cache, mg->new_ocell, false);
1227                 }
1228                 free_io_migration(mg);
1229         }
1230 }
1231
1232 static void copy_complete(int read_err, unsigned long write_err, void *context)
1233 {
1234         unsigned long flags;
1235         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1236         struct cache *cache = mg->cache;
1237
1238         if (read_err || write_err)
1239                 mg->err = true;
1240
1241         spin_lock_irqsave(&cache->lock, flags);
1242         list_add_tail(&mg->list, &cache->completed_migrations);
1243         spin_unlock_irqrestore(&cache->lock, flags);
1244
1245         wake_worker(cache);
1246 }
1247
1248 static void issue_copy(struct dm_cache_migration *mg)
1249 {
1250         int r;
1251         struct dm_io_region o_region, c_region;
1252         struct cache *cache = mg->cache;
1253         sector_t cblock = from_cblock(mg->cblock);
1254
1255         o_region.bdev = cache->origin_dev->bdev;
1256         o_region.count = cache->sectors_per_block;
1257
1258         c_region.bdev = cache->cache_dev->bdev;
1259         c_region.sector = cblock * cache->sectors_per_block;
1260         c_region.count = cache->sectors_per_block;
1261
1262         if (mg->writeback || mg->demote) {
1263                 /* demote */
1264                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1265                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1266         } else {
1267                 /* promote */
1268                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1269                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1270         }
1271
1272         if (r < 0) {
1273                 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1274                 migration_failure(mg);
1275         }
1276 }
1277
1278 static void overwrite_endio(struct bio *bio, int err)
1279 {
1280         struct dm_cache_migration *mg = bio->bi_private;
1281         struct cache *cache = mg->cache;
1282         size_t pb_data_size = get_per_bio_data_size(cache);
1283         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1284         unsigned long flags;
1285
1286         dm_unhook_bio(&pb->hook_info, bio);
1287
1288         if (err)
1289                 mg->err = true;
1290
1291         mg->requeue_holder = false;
1292
1293         spin_lock_irqsave(&cache->lock, flags);
1294         list_add_tail(&mg->list, &cache->completed_migrations);
1295         spin_unlock_irqrestore(&cache->lock, flags);
1296
1297         wake_worker(cache);
1298 }
1299
1300 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1301 {
1302         size_t pb_data_size = get_per_bio_data_size(mg->cache);
1303         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1304
1305         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1306         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1307
1308         /*
1309          * No need to inc_ds() here, since the cell will be held for the
1310          * duration of the io.
1311          */
1312         accounted_request(mg->cache, bio);
1313 }
1314
1315 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1316 {
1317         return (bio_data_dir(bio) == WRITE) &&
1318                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1319 }
1320
1321 static void avoid_copy(struct dm_cache_migration *mg)
1322 {
1323         atomic_inc(&mg->cache->stats.copies_avoided);
1324         migration_success_pre_commit(mg);
1325 }
1326
1327 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1328                                      dm_dblock_t *b, dm_dblock_t *e)
1329 {
1330         sector_t sb = bio->bi_iter.bi_sector;
1331         sector_t se = bio_end_sector(bio);
1332
1333         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1334
1335         if (se - sb < cache->discard_block_size)
1336                 *e = *b;
1337         else
1338                 *e = to_dblock(block_div(se, cache->discard_block_size));
1339 }
1340
1341 static void issue_discard(struct dm_cache_migration *mg)
1342 {
1343         dm_dblock_t b, e;
1344         struct bio *bio = mg->new_ocell->holder;
1345
1346         calc_discard_block_range(mg->cache, bio, &b, &e);
1347         while (b != e) {
1348                 set_discard(mg->cache, b);
1349                 b = to_dblock(from_dblock(b) + 1);
1350         }
1351
1352         bio_endio(bio, 0);
1353         cell_defer(mg->cache, mg->new_ocell, false);
1354         free_migration(mg);
1355         wake_worker(mg->cache);
1356 }
1357
1358 static void issue_copy_or_discard(struct dm_cache_migration *mg)
1359 {
1360         bool avoid;
1361         struct cache *cache = mg->cache;
1362
1363         if (mg->discard) {
1364                 issue_discard(mg);
1365                 return;
1366         }
1367
1368         if (mg->writeback || mg->demote)
1369                 avoid = !is_dirty(cache, mg->cblock) ||
1370                         is_discarded_oblock(cache, mg->old_oblock);
1371         else {
1372                 struct bio *bio = mg->new_ocell->holder;
1373
1374                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1375
1376                 if (writeback_mode(&cache->features) &&
1377                     !avoid && bio_writes_complete_block(cache, bio)) {
1378                         issue_overwrite(mg, bio);
1379                         return;
1380                 }
1381         }
1382
1383         avoid ? avoid_copy(mg) : issue_copy(mg);
1384 }
1385
1386 static void complete_migration(struct dm_cache_migration *mg)
1387 {
1388         if (mg->err)
1389                 migration_failure(mg);
1390         else
1391                 migration_success_pre_commit(mg);
1392 }
1393
1394 static void process_migrations(struct cache *cache, struct list_head *head,
1395                                void (*fn)(struct dm_cache_migration *))
1396 {
1397         unsigned long flags;
1398         struct list_head list;
1399         struct dm_cache_migration *mg, *tmp;
1400
1401         INIT_LIST_HEAD(&list);
1402         spin_lock_irqsave(&cache->lock, flags);
1403         list_splice_init(head, &list);
1404         spin_unlock_irqrestore(&cache->lock, flags);
1405
1406         list_for_each_entry_safe(mg, tmp, &list, list)
1407                 fn(mg);
1408 }
1409
1410 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1411 {
1412         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1413 }
1414
1415 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1416 {
1417         unsigned long flags;
1418         struct cache *cache = mg->cache;
1419
1420         spin_lock_irqsave(&cache->lock, flags);
1421         __queue_quiesced_migration(mg);
1422         spin_unlock_irqrestore(&cache->lock, flags);
1423
1424         wake_worker(cache);
1425 }
1426
1427 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1428 {
1429         unsigned long flags;
1430         struct dm_cache_migration *mg, *tmp;
1431
1432         spin_lock_irqsave(&cache->lock, flags);
1433         list_for_each_entry_safe(mg, tmp, work, list)
1434                 __queue_quiesced_migration(mg);
1435         spin_unlock_irqrestore(&cache->lock, flags);
1436
1437         wake_worker(cache);
1438 }
1439
1440 static void check_for_quiesced_migrations(struct cache *cache,
1441                                           struct per_bio_data *pb)
1442 {
1443         struct list_head work;
1444
1445         if (!pb->all_io_entry)
1446                 return;
1447
1448         INIT_LIST_HEAD(&work);
1449         dm_deferred_entry_dec(pb->all_io_entry, &work);
1450
1451         if (!list_empty(&work))
1452                 queue_quiesced_migrations(cache, &work);
1453 }
1454
1455 static void quiesce_migration(struct dm_cache_migration *mg)
1456 {
1457         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1458                 queue_quiesced_migration(mg);
1459 }
1460
1461 static void promote(struct cache *cache, struct prealloc *structs,
1462                     dm_oblock_t oblock, dm_cblock_t cblock,
1463                     struct dm_bio_prison_cell *cell)
1464 {
1465         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1466
1467         mg->err = false;
1468         mg->discard = false;
1469         mg->writeback = false;
1470         mg->demote = false;
1471         mg->promote = true;
1472         mg->requeue_holder = true;
1473         mg->invalidate = false;
1474         mg->cache = cache;
1475         mg->new_oblock = oblock;
1476         mg->cblock = cblock;
1477         mg->old_ocell = NULL;
1478         mg->new_ocell = cell;
1479         mg->start_jiffies = jiffies;
1480
1481         inc_io_migrations(cache);
1482         quiesce_migration(mg);
1483 }
1484
1485 static void writeback(struct cache *cache, struct prealloc *structs,
1486                       dm_oblock_t oblock, dm_cblock_t cblock,
1487                       struct dm_bio_prison_cell *cell)
1488 {
1489         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1490
1491         mg->err = false;
1492         mg->discard = false;
1493         mg->writeback = true;
1494         mg->demote = false;
1495         mg->promote = false;
1496         mg->requeue_holder = true;
1497         mg->invalidate = false;
1498         mg->cache = cache;
1499         mg->old_oblock = oblock;
1500         mg->cblock = cblock;
1501         mg->old_ocell = cell;
1502         mg->new_ocell = NULL;
1503         mg->start_jiffies = jiffies;
1504
1505         inc_io_migrations(cache);
1506         quiesce_migration(mg);
1507 }
1508
1509 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1510                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1511                                 dm_cblock_t cblock,
1512                                 struct dm_bio_prison_cell *old_ocell,
1513                                 struct dm_bio_prison_cell *new_ocell)
1514 {
1515         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1516
1517         mg->err = false;
1518         mg->discard = false;
1519         mg->writeback = false;
1520         mg->demote = true;
1521         mg->promote = true;
1522         mg->requeue_holder = true;
1523         mg->invalidate = false;
1524         mg->cache = cache;
1525         mg->old_oblock = old_oblock;
1526         mg->new_oblock = new_oblock;
1527         mg->cblock = cblock;
1528         mg->old_ocell = old_ocell;
1529         mg->new_ocell = new_ocell;
1530         mg->start_jiffies = jiffies;
1531
1532         inc_io_migrations(cache);
1533         quiesce_migration(mg);
1534 }
1535
1536 /*
1537  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1538  * block are thrown away.
1539  */
1540 static void invalidate(struct cache *cache, struct prealloc *structs,
1541                        dm_oblock_t oblock, dm_cblock_t cblock,
1542                        struct dm_bio_prison_cell *cell)
1543 {
1544         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1545
1546         mg->err = false;
1547         mg->discard = false;
1548         mg->writeback = false;
1549         mg->demote = true;
1550         mg->promote = false;
1551         mg->requeue_holder = true;
1552         mg->invalidate = true;
1553         mg->cache = cache;
1554         mg->old_oblock = oblock;
1555         mg->cblock = cblock;
1556         mg->old_ocell = cell;
1557         mg->new_ocell = NULL;
1558         mg->start_jiffies = jiffies;
1559
1560         inc_io_migrations(cache);
1561         quiesce_migration(mg);
1562 }
1563
1564 static void discard(struct cache *cache, struct prealloc *structs,
1565                     struct dm_bio_prison_cell *cell)
1566 {
1567         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1568
1569         mg->err = false;
1570         mg->discard = true;
1571         mg->writeback = false;
1572         mg->demote = false;
1573         mg->promote = false;
1574         mg->requeue_holder = false;
1575         mg->invalidate = false;
1576         mg->cache = cache;
1577         mg->old_ocell = NULL;
1578         mg->new_ocell = cell;
1579         mg->start_jiffies = jiffies;
1580
1581         quiesce_migration(mg);
1582 }
1583
1584 /*----------------------------------------------------------------
1585  * bio processing
1586  *--------------------------------------------------------------*/
1587 static void defer_bio(struct cache *cache, struct bio *bio)
1588 {
1589         unsigned long flags;
1590
1591         spin_lock_irqsave(&cache->lock, flags);
1592         bio_list_add(&cache->deferred_bios, bio);
1593         spin_unlock_irqrestore(&cache->lock, flags);
1594
1595         wake_worker(cache);
1596 }
1597
1598 static void process_flush_bio(struct cache *cache, struct bio *bio)
1599 {
1600         size_t pb_data_size = get_per_bio_data_size(cache);
1601         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1602
1603         BUG_ON(bio->bi_iter.bi_size);
1604         if (!pb->req_nr)
1605                 remap_to_origin(cache, bio);
1606         else
1607                 remap_to_cache(cache, bio, 0);
1608
1609         /*
1610          * REQ_FLUSH is not directed at any particular block so we don't
1611          * need to inc_ds().  REQ_FUA's are split into a write + REQ_FLUSH
1612          * by dm-core.
1613          */
1614         issue(cache, bio);
1615 }
1616
1617 static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1618                                 struct bio *bio)
1619 {
1620         int r;
1621         dm_dblock_t b, e;
1622         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1623
1624         calc_discard_block_range(cache, bio, &b, &e);
1625         if (b == e) {
1626                 bio_endio(bio, 0);
1627                 return;
1628         }
1629
1630         cell_prealloc = prealloc_get_cell(structs);
1631         r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1632                              (cell_free_fn) prealloc_put_cell,
1633                              structs, &new_ocell);
1634         if (r > 0)
1635                 return;
1636
1637         discard(cache, structs, new_ocell);
1638 }
1639
1640 static bool spare_migration_bandwidth(struct cache *cache)
1641 {
1642         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1643                 cache->sectors_per_block;
1644         return current_volume < cache->migration_threshold;
1645 }
1646
1647 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1648 {
1649         atomic_inc(bio_data_dir(bio) == READ ?
1650                    &cache->stats.read_hit : &cache->stats.write_hit);
1651 }
1652
1653 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1654 {
1655         atomic_inc(bio_data_dir(bio) == READ ?
1656                    &cache->stats.read_miss : &cache->stats.write_miss);
1657 }
1658
1659 /*----------------------------------------------------------------*/
1660
1661 struct inc_detail {
1662         struct cache *cache;
1663         struct bio_list bios_for_issue;
1664         struct bio_list unhandled_bios;
1665         bool any_writes;
1666 };
1667
1668 static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1669 {
1670         struct bio *bio;
1671         struct inc_detail *detail = context;
1672         struct cache *cache = detail->cache;
1673
1674         inc_ds(cache, cell->holder, cell);
1675         if (bio_data_dir(cell->holder) == WRITE)
1676                 detail->any_writes = true;
1677
1678         while ((bio = bio_list_pop(&cell->bios))) {
1679                 if (discard_or_flush(bio)) {
1680                         bio_list_add(&detail->unhandled_bios, bio);
1681                         continue;
1682                 }
1683
1684                 if (bio_data_dir(bio) == WRITE)
1685                         detail->any_writes = true;
1686
1687                 bio_list_add(&detail->bios_for_issue, bio);
1688                 inc_ds(cache, bio, cell);
1689         }
1690 }
1691
1692 // FIXME: refactor these two
1693 static void remap_cell_to_origin_clear_discard(struct cache *cache,
1694                                                struct dm_bio_prison_cell *cell,
1695                                                dm_oblock_t oblock, bool issue_holder)
1696 {
1697         struct bio *bio;
1698         unsigned long flags;
1699         struct inc_detail detail;
1700
1701         detail.cache = cache;
1702         bio_list_init(&detail.bios_for_issue);
1703         bio_list_init(&detail.unhandled_bios);
1704         detail.any_writes = false;
1705
1706         spin_lock_irqsave(&cache->lock, flags);
1707         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1708         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1709         spin_unlock_irqrestore(&cache->lock, flags);
1710
1711         remap_to_origin(cache, cell->holder);
1712         if (issue_holder)
1713                 issue(cache, cell->holder);
1714         else
1715                 accounted_begin(cache, cell->holder);
1716
1717         if (detail.any_writes)
1718                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1719
1720         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1721                 remap_to_origin(cache, bio);
1722                 issue(cache, bio);
1723         }
1724
1725         free_prison_cell(cache, cell);
1726 }
1727
1728 static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1729                                       dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1730 {
1731         struct bio *bio;
1732         unsigned long flags;
1733         struct inc_detail detail;
1734
1735         detail.cache = cache;
1736         bio_list_init(&detail.bios_for_issue);
1737         bio_list_init(&detail.unhandled_bios);
1738         detail.any_writes = false;
1739
1740         spin_lock_irqsave(&cache->lock, flags);
1741         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1742         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1743         spin_unlock_irqrestore(&cache->lock, flags);
1744
1745         remap_to_cache(cache, cell->holder, cblock);
1746         if (issue_holder)
1747                 issue(cache, cell->holder);
1748         else
1749                 accounted_begin(cache, cell->holder);
1750
1751         if (detail.any_writes) {
1752                 set_dirty(cache, oblock, cblock);
1753                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1754         }
1755
1756         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1757                 remap_to_cache(cache, bio, cblock);
1758                 issue(cache, bio);
1759         }
1760
1761         free_prison_cell(cache, cell);
1762 }
1763
1764 /*----------------------------------------------------------------*/
1765
1766 struct old_oblock_lock {
1767         struct policy_locker locker;
1768         struct cache *cache;
1769         struct prealloc *structs;
1770         struct dm_bio_prison_cell *cell;
1771 };
1772
1773 static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1774 {
1775         /* This should never be called */
1776         BUG();
1777         return 0;
1778 }
1779
1780 static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1781 {
1782         struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1783         struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1784
1785         return bio_detain(l->cache, b, NULL, cell_prealloc,
1786                           (cell_free_fn) prealloc_put_cell,
1787                           l->structs, &l->cell);
1788 }
1789
1790 static void process_cell(struct cache *cache, struct prealloc *structs,
1791                          struct dm_bio_prison_cell *new_ocell)
1792 {
1793         int r;
1794         bool release_cell = true;
1795         struct bio *bio = new_ocell->holder;
1796         dm_oblock_t block = get_bio_block(cache, bio);
1797         struct policy_result lookup_result;
1798         bool passthrough = passthrough_mode(&cache->features);
1799         bool fast_promotion, can_migrate;
1800         struct old_oblock_lock ool;
1801
1802         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1803         can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1804
1805         ool.locker.fn = cell_locker;
1806         ool.cache = cache;
1807         ool.structs = structs;
1808         ool.cell = NULL;
1809         r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1810                        bio, &ool.locker, &lookup_result);
1811
1812         if (r == -EWOULDBLOCK)
1813                 /* migration has been denied */
1814                 lookup_result.op = POLICY_MISS;
1815
1816         switch (lookup_result.op) {
1817         case POLICY_HIT:
1818                 if (passthrough) {
1819                         inc_miss_counter(cache, bio);
1820
1821                         /*
1822                          * Passthrough always maps to the origin,
1823                          * invalidating any cache blocks that are written
1824                          * to.
1825                          */
1826
1827                         if (bio_data_dir(bio) == WRITE) {
1828                                 atomic_inc(&cache->stats.demotion);
1829                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1830                                 release_cell = false;
1831
1832                         } else {
1833                                 /* FIXME: factor out issue_origin() */
1834                                 remap_to_origin_clear_discard(cache, bio, block);
1835                                 inc_and_issue(cache, bio, new_ocell);
1836                         }
1837                 } else {
1838                         inc_hit_counter(cache, bio);
1839
1840                         if (bio_data_dir(bio) == WRITE &&
1841                             writethrough_mode(&cache->features) &&
1842                             !is_dirty(cache, lookup_result.cblock)) {
1843                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1844                                 inc_and_issue(cache, bio, new_ocell);
1845
1846                         } else {
1847                                 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1848                                 release_cell = false;
1849                         }
1850                 }
1851
1852                 break;
1853
1854         case POLICY_MISS:
1855                 inc_miss_counter(cache, bio);
1856                 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1857                 release_cell = false;
1858                 break;
1859
1860         case POLICY_NEW:
1861                 atomic_inc(&cache->stats.promotion);
1862                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1863                 release_cell = false;
1864                 break;
1865
1866         case POLICY_REPLACE:
1867                 atomic_inc(&cache->stats.demotion);
1868                 atomic_inc(&cache->stats.promotion);
1869                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1870                                     block, lookup_result.cblock,
1871                                     ool.cell, new_ocell);
1872                 release_cell = false;
1873                 break;
1874
1875         default:
1876                 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1877                             cache_device_name(cache), __func__,
1878                             (unsigned) lookup_result.op);
1879                 bio_io_error(bio);
1880         }
1881
1882         if (release_cell)
1883                 cell_defer(cache, new_ocell, false);
1884 }
1885
1886 static void process_bio(struct cache *cache, struct prealloc *structs,
1887                         struct bio *bio)
1888 {
1889         int r;
1890         dm_oblock_t block = get_bio_block(cache, bio);
1891         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1892
1893         /*
1894          * Check to see if that block is currently migrating.
1895          */
1896         cell_prealloc = prealloc_get_cell(structs);
1897         r = bio_detain(cache, block, bio, cell_prealloc,
1898                        (cell_free_fn) prealloc_put_cell,
1899                        structs, &new_ocell);
1900         if (r > 0)
1901                 return;
1902
1903         process_cell(cache, structs, new_ocell);
1904 }
1905
1906 static int need_commit_due_to_time(struct cache *cache)
1907 {
1908         return jiffies < cache->last_commit_jiffies ||
1909                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1910 }
1911
1912 /*
1913  * A non-zero return indicates read_only or fail_io mode.
1914  */
1915 static int commit(struct cache *cache, bool clean_shutdown)
1916 {
1917         int r;
1918
1919         if (get_cache_mode(cache) >= CM_READ_ONLY)
1920                 return -EINVAL;
1921
1922         atomic_inc(&cache->stats.commit_count);
1923         r = dm_cache_commit(cache->cmd, clean_shutdown);
1924         if (r)
1925                 metadata_operation_failed(cache, "dm_cache_commit", r);
1926
1927         return r;
1928 }
1929
1930 static int commit_if_needed(struct cache *cache)
1931 {
1932         int r = 0;
1933
1934         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1935             dm_cache_changed_this_transaction(cache->cmd)) {
1936                 r = commit(cache, false);
1937                 cache->commit_requested = false;
1938                 cache->last_commit_jiffies = jiffies;
1939         }
1940
1941         return r;
1942 }
1943
1944 static void process_deferred_bios(struct cache *cache)
1945 {
1946         bool prealloc_used = false;
1947         unsigned long flags;
1948         struct bio_list bios;
1949         struct bio *bio;
1950         struct prealloc structs;
1951
1952         memset(&structs, 0, sizeof(structs));
1953         bio_list_init(&bios);
1954
1955         spin_lock_irqsave(&cache->lock, flags);
1956         bio_list_merge(&bios, &cache->deferred_bios);
1957         bio_list_init(&cache->deferred_bios);
1958         spin_unlock_irqrestore(&cache->lock, flags);
1959
1960         while (!bio_list_empty(&bios)) {
1961                 /*
1962                  * If we've got no free migration structs, and processing
1963                  * this bio might require one, we pause until there are some
1964                  * prepared mappings to process.
1965                  */
1966                 prealloc_used = true;
1967                 if (prealloc_data_structs(cache, &structs)) {
1968                         spin_lock_irqsave(&cache->lock, flags);
1969                         bio_list_merge(&cache->deferred_bios, &bios);
1970                         spin_unlock_irqrestore(&cache->lock, flags);
1971                         break;
1972                 }
1973
1974                 bio = bio_list_pop(&bios);
1975
1976                 if (bio->bi_rw & REQ_FLUSH)
1977                         process_flush_bio(cache, bio);
1978                 else if (bio->bi_rw & REQ_DISCARD)
1979                         process_discard_bio(cache, &structs, bio);
1980                 else
1981                         process_bio(cache, &structs, bio);
1982         }
1983
1984         if (prealloc_used)
1985                 prealloc_free_structs(cache, &structs);
1986 }
1987
1988 static void process_deferred_cells(struct cache *cache)
1989 {
1990         bool prealloc_used = false;
1991         unsigned long flags;
1992         struct dm_bio_prison_cell *cell, *tmp;
1993         struct list_head cells;
1994         struct prealloc structs;
1995
1996         memset(&structs, 0, sizeof(structs));
1997
1998         INIT_LIST_HEAD(&cells);
1999
2000         spin_lock_irqsave(&cache->lock, flags);
2001         list_splice_init(&cache->deferred_cells, &cells);
2002         spin_unlock_irqrestore(&cache->lock, flags);
2003
2004         list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2005                 /*
2006                  * If we've got no free migration structs, and processing
2007                  * this bio might require one, we pause until there are some
2008                  * prepared mappings to process.
2009                  */
2010                 prealloc_used = true;
2011                 if (prealloc_data_structs(cache, &structs)) {
2012                         spin_lock_irqsave(&cache->lock, flags);
2013                         list_splice(&cells, &cache->deferred_cells);
2014                         spin_unlock_irqrestore(&cache->lock, flags);
2015                         break;
2016                 }
2017
2018                 process_cell(cache, &structs, cell);
2019         }
2020
2021         if (prealloc_used)
2022                 prealloc_free_structs(cache, &structs);
2023 }
2024
2025 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2026 {
2027         unsigned long flags;
2028         struct bio_list bios;
2029         struct bio *bio;
2030
2031         bio_list_init(&bios);
2032
2033         spin_lock_irqsave(&cache->lock, flags);
2034         bio_list_merge(&bios, &cache->deferred_flush_bios);
2035         bio_list_init(&cache->deferred_flush_bios);
2036         spin_unlock_irqrestore(&cache->lock, flags);
2037
2038         /*
2039          * These bios have already been through inc_ds()
2040          */
2041         while ((bio = bio_list_pop(&bios)))
2042                 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2043 }
2044
2045 static void process_deferred_writethrough_bios(struct cache *cache)
2046 {
2047         unsigned long flags;
2048         struct bio_list bios;
2049         struct bio *bio;
2050
2051         bio_list_init(&bios);
2052
2053         spin_lock_irqsave(&cache->lock, flags);
2054         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2055         bio_list_init(&cache->deferred_writethrough_bios);
2056         spin_unlock_irqrestore(&cache->lock, flags);
2057
2058         /*
2059          * These bios have already been through inc_ds()
2060          */
2061         while ((bio = bio_list_pop(&bios)))
2062                 accounted_request(cache, bio);
2063 }
2064
2065 static void writeback_some_dirty_blocks(struct cache *cache)
2066 {
2067         bool prealloc_used = false;
2068         dm_oblock_t oblock;
2069         dm_cblock_t cblock;
2070         struct prealloc structs;
2071         struct dm_bio_prison_cell *old_ocell;
2072         bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2073
2074         memset(&structs, 0, sizeof(structs));
2075
2076         while (spare_migration_bandwidth(cache)) {
2077                 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2078                         break; /* no work to do */
2079
2080                 prealloc_used = true;
2081                 if (prealloc_data_structs(cache, &structs) ||
2082                     get_cell(cache, oblock, &structs, &old_ocell)) {
2083                         policy_set_dirty(cache->policy, oblock);
2084                         break;
2085                 }
2086
2087                 writeback(cache, &structs, oblock, cblock, old_ocell);
2088         }
2089
2090         if (prealloc_used)
2091                 prealloc_free_structs(cache, &structs);
2092 }
2093
2094 /*----------------------------------------------------------------
2095  * Invalidations.
2096  * Dropping something from the cache *without* writing back.
2097  *--------------------------------------------------------------*/
2098
2099 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2100 {
2101         int r = 0;
2102         uint64_t begin = from_cblock(req->cblocks->begin);
2103         uint64_t end = from_cblock(req->cblocks->end);
2104
2105         while (begin != end) {
2106                 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2107                 if (!r) {
2108                         r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2109                         if (r) {
2110                                 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2111                                 break;
2112                         }
2113
2114                 } else if (r == -ENODATA) {
2115                         /* harmless, already unmapped */
2116                         r = 0;
2117
2118                 } else {
2119                         DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2120                         break;
2121                 }
2122
2123                 begin++;
2124         }
2125
2126         cache->commit_requested = true;
2127
2128         req->err = r;
2129         atomic_set(&req->complete, 1);
2130
2131         wake_up(&req->result_wait);
2132 }
2133
2134 static void process_invalidation_requests(struct cache *cache)
2135 {
2136         struct list_head list;
2137         struct invalidation_request *req, *tmp;
2138
2139         INIT_LIST_HEAD(&list);
2140         spin_lock(&cache->invalidation_lock);
2141         list_splice_init(&cache->invalidation_requests, &list);
2142         spin_unlock(&cache->invalidation_lock);
2143
2144         list_for_each_entry_safe (req, tmp, &list, list)
2145                 process_invalidation_request(cache, req);
2146 }
2147
2148 /*----------------------------------------------------------------
2149  * Main worker loop
2150  *--------------------------------------------------------------*/
2151 static bool is_quiescing(struct cache *cache)
2152 {
2153         return atomic_read(&cache->quiescing);
2154 }
2155
2156 static void ack_quiescing(struct cache *cache)
2157 {
2158         if (is_quiescing(cache)) {
2159                 atomic_inc(&cache->quiescing_ack);
2160                 wake_up(&cache->quiescing_wait);
2161         }
2162 }
2163
2164 static void wait_for_quiescing_ack(struct cache *cache)
2165 {
2166         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2167 }
2168
2169 static void start_quiescing(struct cache *cache)
2170 {
2171         atomic_inc(&cache->quiescing);
2172         wait_for_quiescing_ack(cache);
2173 }
2174
2175 static void stop_quiescing(struct cache *cache)
2176 {
2177         atomic_set(&cache->quiescing, 0);
2178         atomic_set(&cache->quiescing_ack, 0);
2179 }
2180
2181 static void wait_for_migrations(struct cache *cache)
2182 {
2183         wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2184 }
2185
2186 static void stop_worker(struct cache *cache)
2187 {
2188         cancel_delayed_work(&cache->waker);
2189         flush_workqueue(cache->wq);
2190 }
2191
2192 static void requeue_deferred_cells(struct cache *cache)
2193 {
2194         unsigned long flags;
2195         struct list_head cells;
2196         struct dm_bio_prison_cell *cell, *tmp;
2197
2198         INIT_LIST_HEAD(&cells);
2199         spin_lock_irqsave(&cache->lock, flags);
2200         list_splice_init(&cache->deferred_cells, &cells);
2201         spin_unlock_irqrestore(&cache->lock, flags);
2202
2203         list_for_each_entry_safe(cell, tmp, &cells, user_list)
2204                 cell_requeue(cache, cell);
2205 }
2206
2207 static void requeue_deferred_bios(struct cache *cache)
2208 {
2209         struct bio *bio;
2210         struct bio_list bios;
2211
2212         bio_list_init(&bios);
2213         bio_list_merge(&bios, &cache->deferred_bios);
2214         bio_list_init(&cache->deferred_bios);
2215
2216         while ((bio = bio_list_pop(&bios)))
2217                 bio_endio(bio, DM_ENDIO_REQUEUE);
2218 }
2219
2220 static int more_work(struct cache *cache)
2221 {
2222         if (is_quiescing(cache))
2223                 return !list_empty(&cache->quiesced_migrations) ||
2224                         !list_empty(&cache->completed_migrations) ||
2225                         !list_empty(&cache->need_commit_migrations);
2226         else
2227                 return !bio_list_empty(&cache->deferred_bios) ||
2228                         !list_empty(&cache->deferred_cells) ||
2229                         !bio_list_empty(&cache->deferred_flush_bios) ||
2230                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
2231                         !list_empty(&cache->quiesced_migrations) ||
2232                         !list_empty(&cache->completed_migrations) ||
2233                         !list_empty(&cache->need_commit_migrations) ||
2234                         cache->invalidate;
2235 }
2236
2237 static void do_worker(struct work_struct *ws)
2238 {
2239         struct cache *cache = container_of(ws, struct cache, worker);
2240
2241         do {
2242                 if (!is_quiescing(cache)) {
2243                         writeback_some_dirty_blocks(cache);
2244                         process_deferred_writethrough_bios(cache);
2245                         process_deferred_bios(cache);
2246                         process_deferred_cells(cache);
2247                         process_invalidation_requests(cache);
2248                 }
2249
2250                 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2251                 process_migrations(cache, &cache->completed_migrations, complete_migration);
2252
2253                 if (commit_if_needed(cache)) {
2254                         process_deferred_flush_bios(cache, false);
2255                         process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2256                 } else {
2257                         process_deferred_flush_bios(cache, true);
2258                         process_migrations(cache, &cache->need_commit_migrations,
2259                                            migration_success_post_commit);
2260                 }
2261
2262                 ack_quiescing(cache);
2263
2264         } while (more_work(cache));
2265 }
2266
2267 /*
2268  * We want to commit periodically so that not too much
2269  * unwritten metadata builds up.
2270  */
2271 static void do_waker(struct work_struct *ws)
2272 {
2273         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2274         policy_tick(cache->policy, true);
2275         wake_worker(cache);
2276         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2277 }
2278
2279 /*----------------------------------------------------------------*/
2280
2281 static int is_congested(struct dm_dev *dev, int bdi_bits)
2282 {
2283         struct request_queue *q = bdev_get_queue(dev->bdev);
2284         return bdi_congested(&q->backing_dev_info, bdi_bits);
2285 }
2286
2287 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2288 {
2289         struct cache *cache = container_of(cb, struct cache, callbacks);
2290
2291         return is_congested(cache->origin_dev, bdi_bits) ||
2292                 is_congested(cache->cache_dev, bdi_bits);
2293 }
2294
2295 /*----------------------------------------------------------------
2296  * Target methods
2297  *--------------------------------------------------------------*/
2298
2299 /*
2300  * This function gets called on the error paths of the constructor, so we
2301  * have to cope with a partially initialised struct.
2302  */
2303 static void destroy(struct cache *cache)
2304 {
2305         unsigned i;
2306
2307         if (cache->migration_pool)
2308                 mempool_destroy(cache->migration_pool);
2309
2310         if (cache->all_io_ds)
2311                 dm_deferred_set_destroy(cache->all_io_ds);
2312
2313         if (cache->prison)
2314                 dm_bio_prison_destroy(cache->prison);
2315
2316         if (cache->wq)
2317                 destroy_workqueue(cache->wq);
2318
2319         if (cache->dirty_bitset)
2320                 free_bitset(cache->dirty_bitset);
2321
2322         if (cache->discard_bitset)
2323                 free_bitset(cache->discard_bitset);
2324
2325         if (cache->copier)
2326                 dm_kcopyd_client_destroy(cache->copier);
2327
2328         if (cache->cmd)
2329                 dm_cache_metadata_close(cache->cmd);
2330
2331         if (cache->metadata_dev)
2332                 dm_put_device(cache->ti, cache->metadata_dev);
2333
2334         if (cache->origin_dev)
2335                 dm_put_device(cache->ti, cache->origin_dev);
2336
2337         if (cache->cache_dev)
2338                 dm_put_device(cache->ti, cache->cache_dev);
2339
2340         if (cache->policy)
2341                 dm_cache_policy_destroy(cache->policy);
2342
2343         for (i = 0; i < cache->nr_ctr_args ; i++)
2344                 kfree(cache->ctr_args[i]);
2345         kfree(cache->ctr_args);
2346
2347         kfree(cache);
2348 }
2349
2350 static void cache_dtr(struct dm_target *ti)
2351 {
2352         struct cache *cache = ti->private;
2353
2354         destroy(cache);
2355 }
2356
2357 static sector_t get_dev_size(struct dm_dev *dev)
2358 {
2359         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2360 }
2361
2362 /*----------------------------------------------------------------*/
2363
2364 /*
2365  * Construct a cache device mapping.
2366  *
2367  * cache <metadata dev> <cache dev> <origin dev> <block size>
2368  *       <#feature args> [<feature arg>]*
2369  *       <policy> <#policy args> [<policy arg>]*
2370  *
2371  * metadata dev    : fast device holding the persistent metadata
2372  * cache dev       : fast device holding cached data blocks
2373  * origin dev      : slow device holding original data blocks
2374  * block size      : cache unit size in sectors
2375  *
2376  * #feature args   : number of feature arguments passed
2377  * feature args    : writethrough.  (The default is writeback.)
2378  *
2379  * policy          : the replacement policy to use
2380  * #policy args    : an even number of policy arguments corresponding
2381  *                   to key/value pairs passed to the policy
2382  * policy args     : key/value pairs passed to the policy
2383  *                   E.g. 'sequential_threshold 1024'
2384  *                   See cache-policies.txt for details.
2385  *
2386  * Optional feature arguments are:
2387  *   writethrough  : write through caching that prohibits cache block
2388  *                   content from being different from origin block content.
2389  *                   Without this argument, the default behaviour is to write
2390  *                   back cache block contents later for performance reasons,
2391  *                   so they may differ from the corresponding origin blocks.
2392  */
2393 struct cache_args {
2394         struct dm_target *ti;
2395
2396         struct dm_dev *metadata_dev;
2397
2398         struct dm_dev *cache_dev;
2399         sector_t cache_sectors;
2400
2401         struct dm_dev *origin_dev;
2402         sector_t origin_sectors;
2403
2404         uint32_t block_size;
2405
2406         const char *policy_name;
2407         int policy_argc;
2408         const char **policy_argv;
2409
2410         struct cache_features features;
2411 };
2412
2413 static void destroy_cache_args(struct cache_args *ca)
2414 {
2415         if (ca->metadata_dev)
2416                 dm_put_device(ca->ti, ca->metadata_dev);
2417
2418         if (ca->cache_dev)
2419                 dm_put_device(ca->ti, ca->cache_dev);
2420
2421         if (ca->origin_dev)
2422                 dm_put_device(ca->ti, ca->origin_dev);
2423
2424         kfree(ca);
2425 }
2426
2427 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2428 {
2429         if (!as->argc) {
2430                 *error = "Insufficient args";
2431                 return false;
2432         }
2433
2434         return true;
2435 }
2436
2437 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2438                               char **error)
2439 {
2440         int r;
2441         sector_t metadata_dev_size;
2442         char b[BDEVNAME_SIZE];
2443
2444         if (!at_least_one_arg(as, error))
2445                 return -EINVAL;
2446
2447         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2448                           &ca->metadata_dev);
2449         if (r) {
2450                 *error = "Error opening metadata device";
2451                 return r;
2452         }
2453
2454         metadata_dev_size = get_dev_size(ca->metadata_dev);
2455         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2456                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2457                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2458
2459         return 0;
2460 }
2461
2462 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2463                            char **error)
2464 {
2465         int r;
2466
2467         if (!at_least_one_arg(as, error))
2468                 return -EINVAL;
2469
2470         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2471                           &ca->cache_dev);
2472         if (r) {
2473                 *error = "Error opening cache device";
2474                 return r;
2475         }
2476         ca->cache_sectors = get_dev_size(ca->cache_dev);
2477
2478         return 0;
2479 }
2480
2481 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2482                             char **error)
2483 {
2484         int r;
2485
2486         if (!at_least_one_arg(as, error))
2487                 return -EINVAL;
2488
2489         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2490                           &ca->origin_dev);
2491         if (r) {
2492                 *error = "Error opening origin device";
2493                 return r;
2494         }
2495
2496         ca->origin_sectors = get_dev_size(ca->origin_dev);
2497         if (ca->ti->len > ca->origin_sectors) {
2498                 *error = "Device size larger than cached device";
2499                 return -EINVAL;
2500         }
2501
2502         return 0;
2503 }
2504
2505 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2506                             char **error)
2507 {
2508         unsigned long block_size;
2509
2510         if (!at_least_one_arg(as, error))
2511                 return -EINVAL;
2512
2513         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2514             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2515             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2516             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2517                 *error = "Invalid data block size";
2518                 return -EINVAL;
2519         }
2520
2521         if (block_size > ca->cache_sectors) {
2522                 *error = "Data block size is larger than the cache device";
2523                 return -EINVAL;
2524         }
2525
2526         ca->block_size = block_size;
2527
2528         return 0;
2529 }
2530
2531 static void init_features(struct cache_features *cf)
2532 {
2533         cf->mode = CM_WRITE;
2534         cf->io_mode = CM_IO_WRITEBACK;
2535 }
2536
2537 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2538                           char **error)
2539 {
2540         static struct dm_arg _args[] = {
2541                 {0, 1, "Invalid number of cache feature arguments"},
2542         };
2543
2544         int r;
2545         unsigned argc;
2546         const char *arg;
2547         struct cache_features *cf = &ca->features;
2548
2549         init_features(cf);
2550
2551         r = dm_read_arg_group(_args, as, &argc, error);
2552         if (r)
2553                 return -EINVAL;
2554
2555         while (argc--) {
2556                 arg = dm_shift_arg(as);
2557
2558                 if (!strcasecmp(arg, "writeback"))
2559                         cf->io_mode = CM_IO_WRITEBACK;
2560
2561                 else if (!strcasecmp(arg, "writethrough"))
2562                         cf->io_mode = CM_IO_WRITETHROUGH;
2563
2564                 else if (!strcasecmp(arg, "passthrough"))
2565                         cf->io_mode = CM_IO_PASSTHROUGH;
2566
2567                 else {
2568                         *error = "Unrecognised cache feature requested";
2569                         return -EINVAL;
2570                 }
2571         }
2572
2573         return 0;
2574 }
2575
2576 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2577                         char **error)
2578 {
2579         static struct dm_arg _args[] = {
2580                 {0, 1024, "Invalid number of policy arguments"},
2581         };
2582
2583         int r;
2584
2585         if (!at_least_one_arg(as, error))
2586                 return -EINVAL;
2587
2588         ca->policy_name = dm_shift_arg(as);
2589
2590         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2591         if (r)
2592                 return -EINVAL;
2593
2594         ca->policy_argv = (const char **)as->argv;
2595         dm_consume_args(as, ca->policy_argc);
2596
2597         return 0;
2598 }
2599
2600 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2601                             char **error)
2602 {
2603         int r;
2604         struct dm_arg_set as;
2605
2606         as.argc = argc;
2607         as.argv = argv;
2608
2609         r = parse_metadata_dev(ca, &as, error);
2610         if (r)
2611                 return r;
2612
2613         r = parse_cache_dev(ca, &as, error);
2614         if (r)
2615                 return r;
2616
2617         r = parse_origin_dev(ca, &as, error);
2618         if (r)
2619                 return r;
2620
2621         r = parse_block_size(ca, &as, error);
2622         if (r)
2623                 return r;
2624
2625         r = parse_features(ca, &as, error);
2626         if (r)
2627                 return r;
2628
2629         r = parse_policy(ca, &as, error);
2630         if (r)
2631                 return r;
2632
2633         return 0;
2634 }
2635
2636 /*----------------------------------------------------------------*/
2637
2638 static struct kmem_cache *migration_cache;
2639
2640 #define NOT_CORE_OPTION 1
2641
2642 static int process_config_option(struct cache *cache, const char *key, const char *value)
2643 {
2644         unsigned long tmp;
2645
2646         if (!strcasecmp(key, "migration_threshold")) {
2647                 if (kstrtoul(value, 10, &tmp))
2648                         return -EINVAL;
2649
2650                 cache->migration_threshold = tmp;
2651                 return 0;
2652         }
2653
2654         return NOT_CORE_OPTION;
2655 }
2656
2657 static int set_config_value(struct cache *cache, const char *key, const char *value)
2658 {
2659         int r = process_config_option(cache, key, value);
2660
2661         if (r == NOT_CORE_OPTION)
2662                 r = policy_set_config_value(cache->policy, key, value);
2663
2664         if (r)
2665                 DMWARN("bad config value for %s: %s", key, value);
2666
2667         return r;
2668 }
2669
2670 static int set_config_values(struct cache *cache, int argc, const char **argv)
2671 {
2672         int r = 0;
2673
2674         if (argc & 1) {
2675                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2676                 return -EINVAL;
2677         }
2678
2679         while (argc) {
2680                 r = set_config_value(cache, argv[0], argv[1]);
2681                 if (r)
2682                         break;
2683
2684                 argc -= 2;
2685                 argv += 2;
2686         }
2687
2688         return r;
2689 }
2690
2691 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2692                                char **error)
2693 {
2694         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2695                                                            cache->cache_size,
2696                                                            cache->origin_sectors,
2697                                                            cache->sectors_per_block);
2698         if (IS_ERR(p)) {
2699                 *error = "Error creating cache's policy";
2700                 return PTR_ERR(p);
2701         }
2702         cache->policy = p;
2703
2704         return 0;
2705 }
2706
2707 /*
2708  * We want the discard block size to be at least the size of the cache
2709  * block size and have no more than 2^14 discard blocks across the origin.
2710  */
2711 #define MAX_DISCARD_BLOCKS (1 << 14)
2712
2713 static bool too_many_discard_blocks(sector_t discard_block_size,
2714                                     sector_t origin_size)
2715 {
2716         (void) sector_div(origin_size, discard_block_size);
2717
2718         return origin_size > MAX_DISCARD_BLOCKS;
2719 }
2720
2721 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2722                                              sector_t origin_size)
2723 {
2724         sector_t discard_block_size = cache_block_size;
2725
2726         if (origin_size)
2727                 while (too_many_discard_blocks(discard_block_size, origin_size))
2728                         discard_block_size *= 2;
2729
2730         return discard_block_size;
2731 }
2732
2733 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2734 {
2735         dm_block_t nr_blocks = from_cblock(size);
2736
2737         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2738                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2739                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2740                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2741                              (unsigned long long) nr_blocks);
2742
2743         cache->cache_size = size;
2744 }
2745
2746 #define DEFAULT_MIGRATION_THRESHOLD 2048
2747
2748 static int cache_create(struct cache_args *ca, struct cache **result)
2749 {
2750         int r = 0;
2751         char **error = &ca->ti->error;
2752         struct cache *cache;
2753         struct dm_target *ti = ca->ti;
2754         dm_block_t origin_blocks;
2755         struct dm_cache_metadata *cmd;
2756         bool may_format = ca->features.mode == CM_WRITE;
2757
2758         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2759         if (!cache)
2760                 return -ENOMEM;
2761
2762         cache->ti = ca->ti;
2763         ti->private = cache;
2764         ti->num_flush_bios = 2;
2765         ti->flush_supported = true;
2766
2767         ti->num_discard_bios = 1;
2768         ti->discards_supported = true;
2769         ti->discard_zeroes_data_unsupported = true;
2770         ti->split_discard_bios = false;
2771
2772         cache->features = ca->features;
2773         ti->per_bio_data_size = get_per_bio_data_size(cache);
2774
2775         cache->callbacks.congested_fn = cache_is_congested;
2776         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2777
2778         cache->metadata_dev = ca->metadata_dev;
2779         cache->origin_dev = ca->origin_dev;
2780         cache->cache_dev = ca->cache_dev;
2781
2782         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2783
2784         /* FIXME: factor out this whole section */
2785         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2786         origin_blocks = block_div(origin_blocks, ca->block_size);
2787         cache->origin_blocks = to_oblock(origin_blocks);
2788
2789         cache->sectors_per_block = ca->block_size;
2790         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2791                 r = -EINVAL;
2792                 goto bad;
2793         }
2794
2795         if (ca->block_size & (ca->block_size - 1)) {
2796                 dm_block_t cache_size = ca->cache_sectors;
2797
2798                 cache->sectors_per_block_shift = -1;
2799                 cache_size = block_div(cache_size, ca->block_size);
2800                 set_cache_size(cache, to_cblock(cache_size));
2801         } else {
2802                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2803                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2804         }
2805
2806         r = create_cache_policy(cache, ca, error);
2807         if (r)
2808                 goto bad;
2809
2810         cache->policy_nr_args = ca->policy_argc;
2811         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2812
2813         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2814         if (r) {
2815                 *error = "Error setting cache policy's config values";
2816                 goto bad;
2817         }
2818
2819         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2820                                      ca->block_size, may_format,
2821                                      dm_cache_policy_get_hint_size(cache->policy));
2822         if (IS_ERR(cmd)) {
2823                 *error = "Error creating metadata object";
2824                 r = PTR_ERR(cmd);
2825                 goto bad;
2826         }
2827         cache->cmd = cmd;
2828         set_cache_mode(cache, CM_WRITE);
2829         if (get_cache_mode(cache) != CM_WRITE) {
2830                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2831                 r = -EINVAL;
2832                 goto bad;
2833         }
2834
2835         if (passthrough_mode(&cache->features)) {
2836                 bool all_clean;
2837
2838                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2839                 if (r) {
2840                         *error = "dm_cache_metadata_all_clean() failed";
2841                         goto bad;
2842                 }
2843
2844                 if (!all_clean) {
2845                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2846                         r = -EINVAL;
2847                         goto bad;
2848                 }
2849         }
2850
2851         spin_lock_init(&cache->lock);
2852         INIT_LIST_HEAD(&cache->deferred_cells);
2853         bio_list_init(&cache->deferred_bios);
2854         bio_list_init(&cache->deferred_flush_bios);
2855         bio_list_init(&cache->deferred_writethrough_bios);
2856         INIT_LIST_HEAD(&cache->quiesced_migrations);
2857         INIT_LIST_HEAD(&cache->completed_migrations);
2858         INIT_LIST_HEAD(&cache->need_commit_migrations);
2859         atomic_set(&cache->nr_allocated_migrations, 0);
2860         atomic_set(&cache->nr_io_migrations, 0);
2861         init_waitqueue_head(&cache->migration_wait);
2862
2863         init_waitqueue_head(&cache->quiescing_wait);
2864         atomic_set(&cache->quiescing, 0);
2865         atomic_set(&cache->quiescing_ack, 0);
2866
2867         r = -ENOMEM;
2868         atomic_set(&cache->nr_dirty, 0);
2869         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2870         if (!cache->dirty_bitset) {
2871                 *error = "could not allocate dirty bitset";
2872                 goto bad;
2873         }
2874         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2875
2876         cache->discard_block_size =
2877                 calculate_discard_block_size(cache->sectors_per_block,
2878                                              cache->origin_sectors);
2879         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2880                                                               cache->discard_block_size));
2881         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2882         if (!cache->discard_bitset) {
2883                 *error = "could not allocate discard bitset";
2884                 goto bad;
2885         }
2886         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2887
2888         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2889         if (IS_ERR(cache->copier)) {
2890                 *error = "could not create kcopyd client";
2891                 r = PTR_ERR(cache->copier);
2892                 goto bad;
2893         }
2894
2895         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2896         if (!cache->wq) {
2897                 *error = "could not create workqueue for metadata object";
2898                 goto bad;
2899         }
2900         INIT_WORK(&cache->worker, do_worker);
2901         INIT_DELAYED_WORK(&cache->waker, do_waker);
2902         cache->last_commit_jiffies = jiffies;
2903
2904         cache->prison = dm_bio_prison_create();
2905         if (!cache->prison) {
2906                 *error = "could not create bio prison";
2907                 goto bad;
2908         }
2909
2910         cache->all_io_ds = dm_deferred_set_create();
2911         if (!cache->all_io_ds) {
2912                 *error = "could not create all_io deferred set";
2913                 goto bad;
2914         }
2915
2916         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2917                                                          migration_cache);
2918         if (!cache->migration_pool) {
2919                 *error = "Error creating cache's migration mempool";
2920                 goto bad;
2921         }
2922
2923         cache->need_tick_bio = true;
2924         cache->sized = false;
2925         cache->invalidate = false;
2926         cache->commit_requested = false;
2927         cache->loaded_mappings = false;
2928         cache->loaded_discards = false;
2929
2930         load_stats(cache);
2931
2932         atomic_set(&cache->stats.demotion, 0);
2933         atomic_set(&cache->stats.promotion, 0);
2934         atomic_set(&cache->stats.copies_avoided, 0);
2935         atomic_set(&cache->stats.cache_cell_clash, 0);
2936         atomic_set(&cache->stats.commit_count, 0);
2937         atomic_set(&cache->stats.discard_count, 0);
2938
2939         spin_lock_init(&cache->invalidation_lock);
2940         INIT_LIST_HEAD(&cache->invalidation_requests);
2941
2942         iot_init(&cache->origin_tracker);
2943
2944         *result = cache;
2945         return 0;
2946
2947 bad:
2948         destroy(cache);
2949         return r;
2950 }
2951
2952 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2953 {
2954         unsigned i;
2955         const char **copy;
2956
2957         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2958         if (!copy)
2959                 return -ENOMEM;
2960         for (i = 0; i < argc; i++) {
2961                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2962                 if (!copy[i]) {
2963                         while (i--)
2964                                 kfree(copy[i]);
2965                         kfree(copy);
2966                         return -ENOMEM;
2967                 }
2968         }
2969
2970         cache->nr_ctr_args = argc;
2971         cache->ctr_args = copy;
2972
2973         return 0;
2974 }
2975
2976 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2977 {
2978         int r = -EINVAL;
2979         struct cache_args *ca;
2980         struct cache *cache = NULL;
2981
2982         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2983         if (!ca) {
2984                 ti->error = "Error allocating memory for cache";
2985                 return -ENOMEM;
2986         }
2987         ca->ti = ti;
2988
2989         r = parse_cache_args(ca, argc, argv, &ti->error);
2990         if (r)
2991                 goto out;
2992
2993         r = cache_create(ca, &cache);
2994         if (r)
2995                 goto out;
2996
2997         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2998         if (r) {
2999                 destroy(cache);
3000                 goto out;
3001         }
3002
3003         ti->private = cache;
3004
3005 out:
3006         destroy_cache_args(ca);
3007         return r;
3008 }
3009
3010 /*----------------------------------------------------------------*/
3011
3012 static int cache_map(struct dm_target *ti, struct bio *bio)
3013 {
3014         struct cache *cache = ti->private;
3015
3016         int r;
3017         struct dm_bio_prison_cell *cell = NULL;
3018         dm_oblock_t block = get_bio_block(cache, bio);
3019         size_t pb_data_size = get_per_bio_data_size(cache);
3020         bool can_migrate = false;
3021         bool fast_promotion;
3022         struct policy_result lookup_result;
3023         struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3024         struct old_oblock_lock ool;
3025
3026         ool.locker.fn = null_locker;
3027
3028         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3029                 /*
3030                  * This can only occur if the io goes to a partial block at
3031                  * the end of the origin device.  We don't cache these.
3032                  * Just remap to the origin and carry on.
3033                  */
3034                 remap_to_origin(cache, bio);
3035                 accounted_begin(cache, bio);
3036                 return DM_MAPIO_REMAPPED;
3037         }
3038
3039         if (discard_or_flush(bio)) {
3040                 defer_bio(cache, bio);
3041                 return DM_MAPIO_SUBMITTED;
3042         }
3043
3044         /*
3045          * Check to see if that block is currently migrating.
3046          */
3047         cell = alloc_prison_cell(cache);
3048         if (!cell) {
3049                 defer_bio(cache, bio);
3050                 return DM_MAPIO_SUBMITTED;
3051         }
3052
3053         r = bio_detain(cache, block, bio, cell,
3054                        (cell_free_fn) free_prison_cell,
3055                        cache, &cell);
3056         if (r) {
3057                 if (r < 0)
3058                         defer_bio(cache, bio);
3059
3060                 return DM_MAPIO_SUBMITTED;
3061         }
3062
3063         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3064
3065         r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3066                        bio, &ool.locker, &lookup_result);
3067         if (r == -EWOULDBLOCK) {
3068                 cell_defer(cache, cell, true);
3069                 return DM_MAPIO_SUBMITTED;
3070
3071         } else if (r) {
3072                 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3073                             cache_device_name(cache), r);
3074                 cell_defer(cache, cell, false);
3075                 bio_io_error(bio);
3076                 return DM_MAPIO_SUBMITTED;
3077         }
3078
3079         r = DM_MAPIO_REMAPPED;
3080         switch (lookup_result.op) {
3081         case POLICY_HIT:
3082                 if (passthrough_mode(&cache->features)) {
3083                         if (bio_data_dir(bio) == WRITE) {
3084                                 /*
3085                                  * We need to invalidate this block, so
3086                                  * defer for the worker thread.
3087                                  */
3088                                 cell_defer(cache, cell, true);
3089                                 r = DM_MAPIO_SUBMITTED;
3090
3091                         } else {
3092                                 inc_miss_counter(cache, bio);
3093                                 remap_to_origin_clear_discard(cache, bio, block);
3094                                 accounted_begin(cache, bio);
3095                                 inc_ds(cache, bio, cell);
3096                                 // FIXME: we want to remap hits or misses straight
3097                                 // away rather than passing over to the worker.
3098                                 cell_defer(cache, cell, false);
3099                         }
3100
3101                 } else {
3102                         inc_hit_counter(cache, bio);
3103                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3104                             !is_dirty(cache, lookup_result.cblock)) {
3105                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3106                                 accounted_begin(cache, bio);
3107                                 inc_ds(cache, bio, cell);
3108                                 cell_defer(cache, cell, false);
3109
3110                         } else
3111                                 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3112                 }
3113                 break;
3114
3115         case POLICY_MISS:
3116                 inc_miss_counter(cache, bio);
3117                 if (pb->req_nr != 0) {
3118                         /*
3119                          * This is a duplicate writethrough io that is no
3120                          * longer needed because the block has been demoted.
3121                          */
3122                         bio_endio(bio, 0);
3123                         // FIXME: remap everything as a miss
3124                         cell_defer(cache, cell, false);
3125                         r = DM_MAPIO_SUBMITTED;
3126
3127                 } else
3128                         remap_cell_to_origin_clear_discard(cache, cell, block, false);
3129                 break;
3130
3131         default:
3132                 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3133                             cache_device_name(cache), __func__,
3134                             (unsigned) lookup_result.op);
3135                 cell_defer(cache, cell, false);
3136                 bio_io_error(bio);
3137                 r = DM_MAPIO_SUBMITTED;
3138         }
3139
3140         return r;
3141 }
3142
3143 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3144 {
3145         struct cache *cache = ti->private;
3146         unsigned long flags;
3147         size_t pb_data_size = get_per_bio_data_size(cache);
3148         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3149
3150         if (pb->tick) {
3151                 policy_tick(cache->policy, false);
3152
3153                 spin_lock_irqsave(&cache->lock, flags);
3154                 cache->need_tick_bio = true;
3155                 spin_unlock_irqrestore(&cache->lock, flags);
3156         }
3157
3158         check_for_quiesced_migrations(cache, pb);
3159         accounted_complete(cache, bio);
3160
3161         return 0;
3162 }
3163
3164 static int write_dirty_bitset(struct cache *cache)
3165 {
3166         unsigned i, r;
3167
3168         if (get_cache_mode(cache) >= CM_READ_ONLY)
3169                 return -EINVAL;
3170
3171         for (i = 0; i < from_cblock(cache->cache_size); i++) {
3172                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3173                                        is_dirty(cache, to_cblock(i)));
3174                 if (r) {
3175                         metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3176                         return r;
3177                 }
3178         }
3179
3180         return 0;
3181 }
3182
3183 static int write_discard_bitset(struct cache *cache)
3184 {
3185         unsigned i, r;
3186
3187         if (get_cache_mode(cache) >= CM_READ_ONLY)
3188                 return -EINVAL;
3189
3190         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3191                                            cache->discard_nr_blocks);
3192         if (r) {
3193                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3194                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3195                 return r;
3196         }
3197
3198         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3199                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3200                                          is_discarded(cache, to_dblock(i)));
3201                 if (r) {
3202                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
3203                         return r;
3204                 }
3205         }
3206
3207         return 0;
3208 }
3209
3210 static int write_hints(struct cache *cache)
3211 {
3212         int r;
3213
3214         if (get_cache_mode(cache) >= CM_READ_ONLY)
3215                 return -EINVAL;
3216
3217         r = dm_cache_write_hints(cache->cmd, cache->policy);
3218         if (r) {
3219                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3220                 return r;
3221         }
3222
3223         return 0;
3224 }
3225
3226 /*
3227  * returns true on success
3228  */
3229 static bool sync_metadata(struct cache *cache)
3230 {
3231         int r1, r2, r3, r4;
3232
3233         r1 = write_dirty_bitset(cache);
3234         if (r1)
3235                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3236
3237         r2 = write_discard_bitset(cache);
3238         if (r2)
3239                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3240
3241         save_stats(cache);
3242
3243         r3 = write_hints(cache);
3244         if (r3)
3245                 DMERR("%s: could not write hints", cache_device_name(cache));
3246
3247         /*
3248          * If writing the above metadata failed, we still commit, but don't
3249          * set the clean shutdown flag.  This will effectively force every
3250          * dirty bit to be set on reload.
3251          */
3252         r4 = commit(cache, !r1 && !r2 && !r3);
3253         if (r4)
3254                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3255
3256         return !r1 && !r2 && !r3 && !r4;
3257 }
3258
3259 static void cache_postsuspend(struct dm_target *ti)
3260 {
3261         struct cache *cache = ti->private;
3262
3263         start_quiescing(cache);
3264         wait_for_migrations(cache);
3265         stop_worker(cache);
3266         requeue_deferred_bios(cache);
3267         requeue_deferred_cells(cache);
3268         stop_quiescing(cache);
3269
3270         if (get_cache_mode(cache) == CM_WRITE)
3271                 (void) sync_metadata(cache);
3272 }
3273
3274 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3275                         bool dirty, uint32_t hint, bool hint_valid)
3276 {
3277         int r;
3278         struct cache *cache = context;
3279
3280         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3281         if (r)
3282                 return r;
3283
3284         if (dirty)
3285                 set_dirty(cache, oblock, cblock);
3286         else
3287                 clear_dirty(cache, oblock, cblock);
3288
3289         return 0;
3290 }
3291
3292 /*
3293  * The discard block size in the on disk metadata is not
3294  * neccessarily the same as we're currently using.  So we have to
3295  * be careful to only set the discarded attribute if we know it
3296  * covers a complete block of the new size.
3297  */
3298 struct discard_load_info {
3299         struct cache *cache;
3300
3301         /*
3302          * These blocks are sized using the on disk dblock size, rather
3303          * than the current one.
3304          */
3305         dm_block_t block_size;
3306         dm_block_t discard_begin, discard_end;
3307 };
3308
3309 static void discard_load_info_init(struct cache *cache,
3310                                    struct discard_load_info *li)
3311 {
3312         li->cache = cache;
3313         li->discard_begin = li->discard_end = 0;
3314 }
3315
3316 static void set_discard_range(struct discard_load_info *li)
3317 {
3318         sector_t b, e;
3319
3320         if (li->discard_begin == li->discard_end)
3321                 return;
3322
3323         /*
3324          * Convert to sectors.
3325          */
3326         b = li->discard_begin * li->block_size;
3327         e = li->discard_end * li->block_size;
3328
3329         /*
3330          * Then convert back to the current dblock size.
3331          */
3332         b = dm_sector_div_up(b, li->cache->discard_block_size);
3333         sector_div(e, li->cache->discard_block_size);
3334
3335         /*
3336          * The origin may have shrunk, so we need to check we're still in
3337          * bounds.
3338          */
3339         if (e > from_dblock(li->cache->discard_nr_blocks))
3340                 e = from_dblock(li->cache->discard_nr_blocks);
3341
3342         for (; b < e; b++)
3343                 set_discard(li->cache, to_dblock(b));
3344 }
3345
3346 static int load_discard(void *context, sector_t discard_block_size,
3347                         dm_dblock_t dblock, bool discard)
3348 {
3349         struct discard_load_info *li = context;
3350
3351         li->block_size = discard_block_size;
3352
3353         if (discard) {
3354                 if (from_dblock(dblock) == li->discard_end)
3355                         /*
3356                          * We're already in a discard range, just extend it.
3357                          */
3358                         li->discard_end = li->discard_end + 1ULL;
3359
3360                 else {
3361                         /*
3362                          * Emit the old range and start a new one.
3363                          */
3364                         set_discard_range(li);
3365                         li->discard_begin = from_dblock(dblock);
3366                         li->discard_end = li->discard_begin + 1ULL;
3367                 }
3368         } else {
3369                 set_discard_range(li);
3370                 li->discard_begin = li->discard_end = 0;
3371         }
3372
3373         return 0;
3374 }
3375
3376 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3377 {
3378         sector_t size = get_dev_size(cache->cache_dev);
3379         (void) sector_div(size, cache->sectors_per_block);
3380         return to_cblock(size);
3381 }
3382
3383 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3384 {
3385         if (from_cblock(new_size) > from_cblock(cache->cache_size))
3386                 return true;
3387
3388         /*
3389          * We can't drop a dirty block when shrinking the cache.
3390          */
3391         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3392                 new_size = to_cblock(from_cblock(new_size) + 1);
3393                 if (is_dirty(cache, new_size)) {
3394                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3395                               cache_device_name(cache),
3396                               (unsigned long long) from_cblock(new_size));
3397                         return false;
3398                 }
3399         }
3400
3401         return true;
3402 }
3403
3404 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3405 {
3406         int r;
3407
3408         r = dm_cache_resize(cache->cmd, new_size);
3409         if (r) {
3410                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3411                 metadata_operation_failed(cache, "dm_cache_resize", r);
3412                 return r;
3413         }
3414
3415         set_cache_size(cache, new_size);
3416
3417         return 0;
3418 }
3419
3420 static int cache_preresume(struct dm_target *ti)
3421 {
3422         int r = 0;
3423         struct cache *cache = ti->private;
3424         dm_cblock_t csize = get_cache_dev_size(cache);
3425
3426         /*
3427          * Check to see if the cache has resized.
3428          */
3429         if (!cache->sized) {
3430                 r = resize_cache_dev(cache, csize);
3431                 if (r)
3432                         return r;
3433
3434                 cache->sized = true;
3435
3436         } else if (csize != cache->cache_size) {
3437                 if (!can_resize(cache, csize))
3438                         return -EINVAL;
3439
3440                 r = resize_cache_dev(cache, csize);
3441                 if (r)
3442                         return r;
3443         }
3444
3445         if (!cache->loaded_mappings) {
3446                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3447                                            load_mapping, cache);
3448                 if (r) {
3449                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3450                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3451                         return r;
3452                 }
3453
3454                 cache->loaded_mappings = true;
3455         }
3456
3457         if (!cache->loaded_discards) {
3458                 struct discard_load_info li;
3459
3460                 /*
3461                  * The discard bitset could have been resized, or the
3462                  * discard block size changed.  To be safe we start by
3463                  * setting every dblock to not discarded.
3464                  */
3465                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3466
3467                 discard_load_info_init(cache, &li);
3468                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3469                 if (r) {
3470                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3471                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3472                         return r;
3473                 }
3474                 set_discard_range(&li);
3475
3476                 cache->loaded_discards = true;
3477         }
3478
3479         return r;
3480 }
3481
3482 static void cache_resume(struct dm_target *ti)
3483 {
3484         struct cache *cache = ti->private;
3485
3486         cache->need_tick_bio = true;
3487         do_waker(&cache->waker.work);
3488 }
3489
3490 /*
3491  * Status format:
3492  *
3493  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3494  * <cache block size> <#used cache blocks>/<#total cache blocks>
3495  * <#read hits> <#read misses> <#write hits> <#write misses>
3496  * <#demotions> <#promotions> <#dirty>
3497  * <#features> <features>*
3498  * <#core args> <core args>
3499  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3500  */
3501 static void cache_status(struct dm_target *ti, status_type_t type,
3502                          unsigned status_flags, char *result, unsigned maxlen)
3503 {
3504         int r = 0;
3505         unsigned i;
3506         ssize_t sz = 0;
3507         dm_block_t nr_free_blocks_metadata = 0;
3508         dm_block_t nr_blocks_metadata = 0;
3509         char buf[BDEVNAME_SIZE];
3510         struct cache *cache = ti->private;
3511         dm_cblock_t residency;
3512
3513         switch (type) {
3514         case STATUSTYPE_INFO:
3515                 if (get_cache_mode(cache) == CM_FAIL) {
3516                         DMEMIT("Fail");
3517                         break;
3518                 }
3519
3520                 /* Commit to ensure statistics aren't out-of-date */
3521                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3522                         (void) commit(cache, false);
3523
3524                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3525                 if (r) {
3526                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3527                               cache_device_name(cache), r);
3528                         goto err;
3529                 }
3530
3531                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3532                 if (r) {
3533                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3534                               cache_device_name(cache), r);
3535                         goto err;
3536                 }
3537
3538                 residency = policy_residency(cache->policy);
3539
3540                 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3541                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3542                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3543                        (unsigned long long)nr_blocks_metadata,
3544                        cache->sectors_per_block,
3545                        (unsigned long long) from_cblock(residency),
3546                        (unsigned long long) from_cblock(cache->cache_size),
3547                        (unsigned) atomic_read(&cache->stats.read_hit),
3548                        (unsigned) atomic_read(&cache->stats.read_miss),
3549                        (unsigned) atomic_read(&cache->stats.write_hit),
3550                        (unsigned) atomic_read(&cache->stats.write_miss),
3551                        (unsigned) atomic_read(&cache->stats.demotion),
3552                        (unsigned) atomic_read(&cache->stats.promotion),
3553                        (unsigned long) atomic_read(&cache->nr_dirty));
3554
3555                 if (writethrough_mode(&cache->features))
3556                         DMEMIT("1 writethrough ");
3557
3558                 else if (passthrough_mode(&cache->features))
3559                         DMEMIT("1 passthrough ");
3560
3561                 else if (writeback_mode(&cache->features))
3562                         DMEMIT("1 writeback ");
3563
3564                 else {
3565                         DMERR("%s: internal error: unknown io mode: %d",
3566                               cache_device_name(cache), (int) cache->features.io_mode);
3567                         goto err;
3568                 }
3569
3570                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3571
3572                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3573                 if (sz < maxlen) {
3574                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3575                         if (r)
3576                                 DMERR("%s: policy_emit_config_values returned %d",
3577                                       cache_device_name(cache), r);
3578                 }
3579
3580                 if (get_cache_mode(cache) == CM_READ_ONLY)
3581                         DMEMIT("ro ");
3582                 else
3583                         DMEMIT("rw ");
3584
3585                 if (dm_cache_metadata_needs_check(cache->cmd))
3586                         DMEMIT("needs_check ");
3587                 else
3588                         DMEMIT("- ");
3589
3590                 break;
3591
3592         case STATUSTYPE_TABLE:
3593                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3594                 DMEMIT("%s ", buf);
3595                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3596                 DMEMIT("%s ", buf);
3597                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3598                 DMEMIT("%s", buf);
3599
3600                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3601                         DMEMIT(" %s", cache->ctr_args[i]);
3602                 if (cache->nr_ctr_args)
3603                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3604         }
3605
3606         return;
3607
3608 err:
3609         DMEMIT("Error");
3610 }
3611
3612 /*
3613  * A cache block range can take two forms:
3614  *
3615  * i) A single cblock, eg. '3456'
3616  * ii) A begin and end cblock with dots between, eg. 123-234
3617  */
3618 static int parse_cblock_range(struct cache *cache, const char *str,
3619                               struct cblock_range *result)
3620 {
3621         char dummy;
3622         uint64_t b, e;
3623         int r;
3624
3625         /*
3626          * Try and parse form (ii) first.
3627          */
3628         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3629         if (r < 0)
3630                 return r;
3631
3632         if (r == 2) {
3633                 result->begin = to_cblock(b);
3634                 result->end = to_cblock(e);
3635                 return 0;
3636         }
3637
3638         /*
3639          * That didn't work, try form (i).
3640          */
3641         r = sscanf(str, "%llu%c", &b, &dummy);
3642         if (r < 0)
3643                 return r;
3644
3645         if (r == 1) {
3646                 result->begin = to_cblock(b);
3647                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3648                 return 0;
3649         }
3650
3651         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3652         return -EINVAL;
3653 }
3654
3655 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3656 {
3657         uint64_t b = from_cblock(range->begin);
3658         uint64_t e = from_cblock(range->end);
3659         uint64_t n = from_cblock(cache->cache_size);
3660
3661         if (b >= n) {
3662                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3663                       cache_device_name(cache), b, n);
3664                 return -EINVAL;
3665         }
3666
3667         if (e > n) {
3668                 DMERR("%s: end cblock out of range: %llu > %llu",
3669                       cache_device_name(cache), e, n);
3670                 return -EINVAL;
3671         }
3672
3673         if (b >= e) {
3674                 DMERR("%s: invalid cblock range: %llu >= %llu",
3675                       cache_device_name(cache), b, e);
3676                 return -EINVAL;
3677         }
3678
3679         return 0;
3680 }
3681
3682 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3683 {
3684         struct invalidation_request req;
3685
3686         INIT_LIST_HEAD(&req.list);
3687         req.cblocks = range;
3688         atomic_set(&req.complete, 0);
3689         req.err = 0;
3690         init_waitqueue_head(&req.result_wait);
3691
3692         spin_lock(&cache->invalidation_lock);
3693         list_add(&req.list, &cache->invalidation_requests);
3694         spin_unlock(&cache->invalidation_lock);
3695         wake_worker(cache);
3696
3697         wait_event(req.result_wait, atomic_read(&req.complete));
3698         return req.err;
3699 }
3700
3701 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3702                                               const char **cblock_ranges)
3703 {
3704         int r = 0;
3705         unsigned i;
3706         struct cblock_range range;
3707
3708         if (!passthrough_mode(&cache->features)) {
3709                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3710                       cache_device_name(cache));
3711                 return -EPERM;
3712         }
3713
3714         for (i = 0; i < count; i++) {
3715                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3716                 if (r)
3717                         break;
3718
3719                 r = validate_cblock_range(cache, &range);
3720                 if (r)
3721                         break;
3722
3723                 /*
3724                  * Pass begin and end origin blocks to the worker and wake it.
3725                  */
3726                 r = request_invalidation(cache, &range);
3727                 if (r)
3728                         break;
3729         }
3730
3731         return r;
3732 }
3733
3734 /*
3735  * Supports
3736  *      "<key> <value>"
3737  * and
3738  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3739  *
3740  * The key migration_threshold is supported by the cache target core.
3741  */
3742 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3743 {
3744         struct cache *cache = ti->private;
3745
3746         if (!argc)
3747                 return -EINVAL;
3748
3749         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3750                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3751                       cache_device_name(cache));
3752                 return -EOPNOTSUPP;
3753         }
3754
3755         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3756                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3757
3758         if (argc != 2)
3759                 return -EINVAL;
3760
3761         return set_config_value(cache, argv[0], argv[1]);
3762 }
3763
3764 static int cache_iterate_devices(struct dm_target *ti,
3765                                  iterate_devices_callout_fn fn, void *data)
3766 {
3767         int r = 0;
3768         struct cache *cache = ti->private;
3769
3770         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3771         if (!r)
3772                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3773
3774         return r;
3775 }
3776
3777 /*
3778  * We assume I/O is going to the origin (which is the volume
3779  * more likely to have restrictions e.g. by being striped).
3780  * (Looking up the exact location of the data would be expensive
3781  * and could always be out of date by the time the bio is submitted.)
3782  */
3783 static int cache_bvec_merge(struct dm_target *ti,
3784                             struct bvec_merge_data *bvm,
3785                             struct bio_vec *biovec, int max_size)
3786 {
3787         struct cache *cache = ti->private;
3788         struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3789
3790         if (!q->merge_bvec_fn)
3791                 return max_size;
3792
3793         bvm->bi_bdev = cache->origin_dev->bdev;
3794         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3795 }
3796
3797 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3798 {
3799         /*
3800          * FIXME: these limits may be incompatible with the cache device
3801          */
3802         limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3803                                             cache->origin_sectors);
3804         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3805 }
3806
3807 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3808 {
3809         struct cache *cache = ti->private;
3810         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3811
3812         /*
3813          * If the system-determined stacked limits are compatible with the
3814          * cache's blocksize (io_opt is a factor) do not override them.
3815          */
3816         if (io_opt_sectors < cache->sectors_per_block ||
3817             do_div(io_opt_sectors, cache->sectors_per_block)) {
3818                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3819                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3820         }
3821         set_discard_limits(cache, limits);
3822 }
3823
3824 /*----------------------------------------------------------------*/
3825
3826 static struct target_type cache_target = {
3827         .name = "cache",
3828         .version = {1, 8, 0},
3829         .module = THIS_MODULE,
3830         .ctr = cache_ctr,
3831         .dtr = cache_dtr,
3832         .map = cache_map,
3833         .end_io = cache_end_io,
3834         .postsuspend = cache_postsuspend,
3835         .preresume = cache_preresume,
3836         .resume = cache_resume,
3837         .status = cache_status,
3838         .message = cache_message,
3839         .iterate_devices = cache_iterate_devices,
3840         .merge = cache_bvec_merge,
3841         .io_hints = cache_io_hints,
3842 };
3843
3844 static int __init dm_cache_init(void)
3845 {
3846         int r;
3847
3848         r = dm_register_target(&cache_target);
3849         if (r) {
3850                 DMERR("cache target registration failed: %d", r);
3851                 return r;
3852         }
3853
3854         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3855         if (!migration_cache) {
3856                 dm_unregister_target(&cache_target);
3857                 return -ENOMEM;
3858         }
3859
3860         return 0;
3861 }
3862
3863 static void __exit dm_cache_exit(void)
3864 {
3865         dm_unregister_target(&cache_target);
3866         kmem_cache_destroy(migration_cache);
3867 }
3868
3869 module_init(dm_cache_init);
3870 module_exit(dm_cache_exit);
3871
3872 MODULE_DESCRIPTION(DM_NAME " cache target");
3873 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3874 MODULE_LICENSE("GPL");