kernel/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)
 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 (bio->bi_error) {
 928                 bio_endio(bio);
 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         struct cache *cache = mg->cache;
1117
1118         dec_io_migrations(cache);
1119         free_migration(mg);
1120         wake_worker(cache);
1121 }
1122
1123 static void migration_failure(struct dm_cache_migration *mg)
1124 {
1125         struct cache *cache = mg->cache;
1126         const char *dev_name = cache_device_name(cache);
1127
1128         if (mg->writeback) {
1129                 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1130                 set_dirty(cache, mg->old_oblock, mg->cblock);
1131                 cell_defer(cache, mg->old_ocell, false);
1132
1133         } else if (mg->demote) {
1134                 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1135                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1136
1137                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1138                 if (mg->promote)
1139                         cell_defer(cache, mg->new_ocell, true);
1140         } else {
1141                 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1142                 policy_remove_mapping(cache->policy, mg->new_oblock);
1143                 cell_defer(cache, mg->new_ocell, true);
1144         }
1145
1146         free_io_migration(mg);
1147 }
1148
1149 static void migration_success_pre_commit(struct dm_cache_migration *mg)
1150 {
1151         int r;
1152         unsigned long flags;
1153         struct cache *cache = mg->cache;
1154
1155         if (mg->writeback) {
1156                 clear_dirty(cache, mg->old_oblock, mg->cblock);
1157                 cell_defer(cache, mg->old_ocell, false);
1158                 free_io_migration(mg);
1159                 return;
1160
1161         } else if (mg->demote) {
1162                 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1163                 if (r) {
1164                         DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1165                                     cache_device_name(cache));
1166                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1167                         policy_force_mapping(cache->policy, mg->new_oblock,
1168                                              mg->old_oblock);
1169                         if (mg->promote)
1170                                 cell_defer(cache, mg->new_ocell, true);
1171                         free_io_migration(mg);
1172                         return;
1173                 }
1174         } else {
1175                 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1176                 if (r) {
1177                         DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1178                                     cache_device_name(cache));
1179                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1180                         policy_remove_mapping(cache->policy, mg->new_oblock);
1181                         free_io_migration(mg);
1182                         return;
1183                 }
1184         }
1185
1186         spin_lock_irqsave(&cache->lock, flags);
1187         list_add_tail(&mg->list, &cache->need_commit_migrations);
1188         cache->commit_requested = true;
1189         spin_unlock_irqrestore(&cache->lock, flags);
1190 }
1191
1192 static void migration_success_post_commit(struct dm_cache_migration *mg)
1193 {
1194         unsigned long flags;
1195         struct cache *cache = mg->cache;
1196
1197         if (mg->writeback) {
1198                 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1199                              cache_device_name(cache));
1200                 return;
1201
1202         } else if (mg->demote) {
1203                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1204
1205                 if (mg->promote) {
1206                         mg->demote = false;
1207
1208                         spin_lock_irqsave(&cache->lock, flags);
1209                         list_add_tail(&mg->list, &cache->quiesced_migrations);
1210                         spin_unlock_irqrestore(&cache->lock, flags);
1211
1212                 } else {
1213                         if (mg->invalidate)
1214                                 policy_remove_mapping(cache->policy, mg->old_oblock);
1215                         free_io_migration(mg);
1216                 }
1217
1218         } else {
1219                 if (mg->requeue_holder) {
1220                         clear_dirty(cache, mg->new_oblock, mg->cblock);
1221                         cell_defer(cache, mg->new_ocell, true);
1222                 } else {
1223                         /*
1224                          * The block was promoted via an overwrite, so it's dirty.
1225                          */
1226                         set_dirty(cache, mg->new_oblock, mg->cblock);
1227                         bio_endio(mg->new_ocell->holder);
1228                         cell_defer(cache, mg->new_ocell, false);
1229                 }
1230                 free_io_migration(mg);
1231         }
1232 }
1233
1234 static void copy_complete(int read_err, unsigned long write_err, void *context)
1235 {
1236         unsigned long flags;
1237         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1238         struct cache *cache = mg->cache;
1239
1240         if (read_err || write_err)
1241                 mg->err = true;
1242
1243         spin_lock_irqsave(&cache->lock, flags);
1244         list_add_tail(&mg->list, &cache->completed_migrations);
1245         spin_unlock_irqrestore(&cache->lock, flags);
1246
1247         wake_worker(cache);
1248 }
1249
1250 static void issue_copy(struct dm_cache_migration *mg)
1251 {
1252         int r;
1253         struct dm_io_region o_region, c_region;
1254         struct cache *cache = mg->cache;
1255         sector_t cblock = from_cblock(mg->cblock);
1256
1257         o_region.bdev = cache->origin_dev->bdev;
1258         o_region.count = cache->sectors_per_block;
1259
1260         c_region.bdev = cache->cache_dev->bdev;
1261         c_region.sector = cblock * cache->sectors_per_block;
1262         c_region.count = cache->sectors_per_block;
1263
1264         if (mg->writeback || mg->demote) {
1265                 /* demote */
1266                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1267                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1268         } else {
1269                 /* promote */
1270                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1271                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1272         }
1273
1274         if (r < 0) {
1275                 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1276                 migration_failure(mg);
1277         }
1278 }
1279
1280 static void overwrite_endio(struct bio *bio)
1281 {
1282         struct dm_cache_migration *mg = bio->bi_private;
1283         struct cache *cache = mg->cache;
1284         size_t pb_data_size = get_per_bio_data_size(cache);
1285         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1286         unsigned long flags;
1287
1288         dm_unhook_bio(&pb->hook_info, bio);
1289
1290         if (bio->bi_error)
1291                 mg->err = true;
1292
1293         mg->requeue_holder = false;
1294
1295         spin_lock_irqsave(&cache->lock, flags);
1296         list_add_tail(&mg->list, &cache->completed_migrations);
1297         spin_unlock_irqrestore(&cache->lock, flags);
1298
1299         wake_worker(cache);
1300 }
1301
1302 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1303 {
1304         size_t pb_data_size = get_per_bio_data_size(mg->cache);
1305         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1306
1307         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1308         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1309
1310         /*
1311          * No need to inc_ds() here, since the cell will be held for the
1312          * duration of the io.
1313          */
1314         accounted_request(mg->cache, bio);
1315 }
1316
1317 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1318 {
1319         return (bio_data_dir(bio) == WRITE) &&
1320                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1321 }
1322
1323 static void avoid_copy(struct dm_cache_migration *mg)
1324 {
1325         atomic_inc(&mg->cache->stats.copies_avoided);
1326         migration_success_pre_commit(mg);
1327 }
1328
1329 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1330                                      dm_dblock_t *b, dm_dblock_t *e)
1331 {
1332         sector_t sb = bio->bi_iter.bi_sector;
1333         sector_t se = bio_end_sector(bio);
1334
1335         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1336
1337         if (se - sb < cache->discard_block_size)
1338                 *e = *b;
1339         else
1340                 *e = to_dblock(block_div(se, cache->discard_block_size));
1341 }
1342
1343 static void issue_discard(struct dm_cache_migration *mg)
1344 {
1345         dm_dblock_t b, e;
1346         struct bio *bio = mg->new_ocell->holder;
1347         struct cache *cache = mg->cache;
1348
1349         calc_discard_block_range(cache, bio, &b, &e);
1350         while (b != e) {
1351                 set_discard(cache, b);
1352                 b = to_dblock(from_dblock(b) + 1);
1353         }
1354
1355         bio_endio(bio);
1356         cell_defer(cache, mg->new_ocell, false);
1357         free_migration(mg);
1358         wake_worker(cache);
1359 }
1360
1361 static void issue_copy_or_discard(struct dm_cache_migration *mg)
1362 {
1363         bool avoid;
1364         struct cache *cache = mg->cache;
1365
1366         if (mg->discard) {
1367                 issue_discard(mg);
1368                 return;
1369         }
1370
1371         if (mg->writeback || mg->demote)
1372                 avoid = !is_dirty(cache, mg->cblock) ||
1373                         is_discarded_oblock(cache, mg->old_oblock);
1374         else {
1375                 struct bio *bio = mg->new_ocell->holder;
1376
1377                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1378
1379                 if (writeback_mode(&cache->features) &&
1380                     !avoid && bio_writes_complete_block(cache, bio)) {
1381                         issue_overwrite(mg, bio);
1382                         return;
1383                 }
1384         }
1385
1386         avoid ? avoid_copy(mg) : issue_copy(mg);
1387 }
1388
1389 static void complete_migration(struct dm_cache_migration *mg)
1390 {
1391         if (mg->err)
1392                 migration_failure(mg);
1393         else
1394                 migration_success_pre_commit(mg);
1395 }
1396
1397 static void process_migrations(struct cache *cache, struct list_head *head,
1398                                void (*fn)(struct dm_cache_migration *))
1399 {
1400         unsigned long flags;
1401         struct list_head list;
1402         struct dm_cache_migration *mg, *tmp;
1403
1404         INIT_LIST_HEAD(&list);
1405         spin_lock_irqsave(&cache->lock, flags);
1406         list_splice_init(head, &list);
1407         spin_unlock_irqrestore(&cache->lock, flags);
1408
1409         list_for_each_entry_safe(mg, tmp, &list, list)
1410                 fn(mg);
1411 }
1412
1413 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1414 {
1415         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1416 }
1417
1418 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1419 {
1420         unsigned long flags;
1421         struct cache *cache = mg->cache;
1422
1423         spin_lock_irqsave(&cache->lock, flags);
1424         __queue_quiesced_migration(mg);
1425         spin_unlock_irqrestore(&cache->lock, flags);
1426
1427         wake_worker(cache);
1428 }
1429
1430 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1431 {
1432         unsigned long flags;
1433         struct dm_cache_migration *mg, *tmp;
1434
1435         spin_lock_irqsave(&cache->lock, flags);
1436         list_for_each_entry_safe(mg, tmp, work, list)
1437                 __queue_quiesced_migration(mg);
1438         spin_unlock_irqrestore(&cache->lock, flags);
1439
1440         wake_worker(cache);
1441 }
1442
1443 static void check_for_quiesced_migrations(struct cache *cache,
1444                                           struct per_bio_data *pb)
1445 {
1446         struct list_head work;
1447
1448         if (!pb->all_io_entry)
1449                 return;
1450
1451         INIT_LIST_HEAD(&work);
1452         dm_deferred_entry_dec(pb->all_io_entry, &work);
1453
1454         if (!list_empty(&work))
1455                 queue_quiesced_migrations(cache, &work);
1456 }
1457
1458 static void quiesce_migration(struct dm_cache_migration *mg)
1459 {
1460         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1461                 queue_quiesced_migration(mg);
1462 }
1463
1464 static void promote(struct cache *cache, struct prealloc *structs,
1465                     dm_oblock_t oblock, dm_cblock_t cblock,
1466                     struct dm_bio_prison_cell *cell)
1467 {
1468         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1469
1470         mg->err = false;
1471         mg->discard = false;
1472         mg->writeback = false;
1473         mg->demote = false;
1474         mg->promote = true;
1475         mg->requeue_holder = true;
1476         mg->invalidate = false;
1477         mg->cache = cache;
1478         mg->new_oblock = oblock;
1479         mg->cblock = cblock;
1480         mg->old_ocell = NULL;
1481         mg->new_ocell = cell;
1482         mg->start_jiffies = jiffies;
1483
1484         inc_io_migrations(cache);
1485         quiesce_migration(mg);
1486 }
1487
1488 static void writeback(struct cache *cache, struct prealloc *structs,
1489                       dm_oblock_t oblock, dm_cblock_t cblock,
1490                       struct dm_bio_prison_cell *cell)
1491 {
1492         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1493
1494         mg->err = false;
1495         mg->discard = false;
1496         mg->writeback = true;
1497         mg->demote = false;
1498         mg->promote = false;
1499         mg->requeue_holder = true;
1500         mg->invalidate = false;
1501         mg->cache = cache;
1502         mg->old_oblock = oblock;
1503         mg->cblock = cblock;
1504         mg->old_ocell = cell;
1505         mg->new_ocell = NULL;
1506         mg->start_jiffies = jiffies;
1507
1508         inc_io_migrations(cache);
1509         quiesce_migration(mg);
1510 }
1511
1512 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1513                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1514                                 dm_cblock_t cblock,
1515                                 struct dm_bio_prison_cell *old_ocell,
1516                                 struct dm_bio_prison_cell *new_ocell)
1517 {
1518         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1519
1520         mg->err = false;
1521         mg->discard = false;
1522         mg->writeback = false;
1523         mg->demote = true;
1524         mg->promote = true;
1525         mg->requeue_holder = true;
1526         mg->invalidate = false;
1527         mg->cache = cache;
1528         mg->old_oblock = old_oblock;
1529         mg->new_oblock = new_oblock;
1530         mg->cblock = cblock;
1531         mg->old_ocell = old_ocell;
1532         mg->new_ocell = new_ocell;
1533         mg->start_jiffies = jiffies;
1534
1535         inc_io_migrations(cache);
1536         quiesce_migration(mg);
1537 }
1538
1539 /*
1540  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1541  * block are thrown away.
1542  */
1543 static void invalidate(struct cache *cache, struct prealloc *structs,
1544                        dm_oblock_t oblock, dm_cblock_t cblock,
1545                        struct dm_bio_prison_cell *cell)
1546 {
1547         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1548
1549         mg->err = false;
1550         mg->discard = false;
1551         mg->writeback = false;
1552         mg->demote = true;
1553         mg->promote = false;
1554         mg->requeue_holder = true;
1555         mg->invalidate = true;
1556         mg->cache = cache;
1557         mg->old_oblock = oblock;
1558         mg->cblock = cblock;
1559         mg->old_ocell = cell;
1560         mg->new_ocell = NULL;
1561         mg->start_jiffies = jiffies;
1562
1563         inc_io_migrations(cache);
1564         quiesce_migration(mg);
1565 }
1566
1567 static void discard(struct cache *cache, struct prealloc *structs,
1568                     struct dm_bio_prison_cell *cell)
1569 {
1570         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1571
1572         mg->err = false;
1573         mg->discard = true;
1574         mg->writeback = false;
1575         mg->demote = false;
1576         mg->promote = false;
1577         mg->requeue_holder = false;
1578         mg->invalidate = false;
1579         mg->cache = cache;
1580         mg->old_ocell = NULL;
1581         mg->new_ocell = cell;
1582         mg->start_jiffies = jiffies;
1583
1584         quiesce_migration(mg);
1585 }
1586
1587 /*----------------------------------------------------------------
1588  * bio processing
1589  *--------------------------------------------------------------*/
1590 static void defer_bio(struct cache *cache, struct bio *bio)
1591 {
1592         unsigned long flags;
1593
1594         spin_lock_irqsave(&cache->lock, flags);
1595         bio_list_add(&cache->deferred_bios, bio);
1596         spin_unlock_irqrestore(&cache->lock, flags);
1597
1598         wake_worker(cache);
1599 }
1600
1601 static void process_flush_bio(struct cache *cache, struct bio *bio)
1602 {
1603         size_t pb_data_size = get_per_bio_data_size(cache);
1604         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1605
1606         BUG_ON(bio->bi_iter.bi_size);
1607         if (!pb->req_nr)
1608                 remap_to_origin(cache, bio);
1609         else
1610                 remap_to_cache(cache, bio, 0);
1611
1612         /*
1613          * REQ_FLUSH is not directed at any particular block so we don't
1614          * need to inc_ds().  REQ_FUA's are split into a write + REQ_FLUSH
1615          * by dm-core.
1616          */
1617         issue(cache, bio);
1618 }
1619
1620 static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1621                                 struct bio *bio)
1622 {
1623         int r;
1624         dm_dblock_t b, e;
1625         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1626
1627         calc_discard_block_range(cache, bio, &b, &e);
1628         if (b == e) {
1629                 bio_endio(bio);
1630                 return;
1631         }
1632
1633         cell_prealloc = prealloc_get_cell(structs);
1634         r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1635                              (cell_free_fn) prealloc_put_cell,
1636                              structs, &new_ocell);
1637         if (r > 0)
1638                 return;
1639
1640         discard(cache, structs, new_ocell);
1641 }
1642
1643 static bool spare_migration_bandwidth(struct cache *cache)
1644 {
1645         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1646                 cache->sectors_per_block;
1647         return current_volume < cache->migration_threshold;
1648 }
1649
1650 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1651 {
1652         atomic_inc(bio_data_dir(bio) == READ ?
1653                    &cache->stats.read_hit : &cache->stats.write_hit);
1654 }
1655
1656 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1657 {
1658         atomic_inc(bio_data_dir(bio) == READ ?
1659                    &cache->stats.read_miss : &cache->stats.write_miss);
1660 }
1661
1662 /*----------------------------------------------------------------*/
1663
1664 struct inc_detail {
1665         struct cache *cache;
1666         struct bio_list bios_for_issue;
1667         struct bio_list unhandled_bios;
1668         bool any_writes;
1669 };
1670
1671 static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1672 {
1673         struct bio *bio;
1674         struct inc_detail *detail = context;
1675         struct cache *cache = detail->cache;
1676
1677         inc_ds(cache, cell->holder, cell);
1678         if (bio_data_dir(cell->holder) == WRITE)
1679                 detail->any_writes = true;
1680
1681         while ((bio = bio_list_pop(&cell->bios))) {
1682                 if (discard_or_flush(bio)) {
1683                         bio_list_add(&detail->unhandled_bios, bio);
1684                         continue;
1685                 }
1686
1687                 if (bio_data_dir(bio) == WRITE)
1688                         detail->any_writes = true;
1689
1690                 bio_list_add(&detail->bios_for_issue, bio);
1691                 inc_ds(cache, bio, cell);
1692         }
1693 }
1694
1695 // FIXME: refactor these two
1696 static void remap_cell_to_origin_clear_discard(struct cache *cache,
1697                                                struct dm_bio_prison_cell *cell,
1698                                                dm_oblock_t oblock, bool issue_holder)
1699 {
1700         struct bio *bio;
1701         unsigned long flags;
1702         struct inc_detail detail;
1703
1704         detail.cache = cache;
1705         bio_list_init(&detail.bios_for_issue);
1706         bio_list_init(&detail.unhandled_bios);
1707         detail.any_writes = false;
1708
1709         spin_lock_irqsave(&cache->lock, flags);
1710         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1711         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1712         spin_unlock_irqrestore(&cache->lock, flags);
1713
1714         remap_to_origin(cache, cell->holder);
1715         if (issue_holder)
1716                 issue(cache, cell->holder);
1717         else
1718                 accounted_begin(cache, cell->holder);
1719
1720         if (detail.any_writes)
1721                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1722
1723         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1724                 remap_to_origin(cache, bio);
1725                 issue(cache, bio);
1726         }
1727
1728         free_prison_cell(cache, cell);
1729 }
1730
1731 static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1732                                       dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1733 {
1734         struct bio *bio;
1735         unsigned long flags;
1736         struct inc_detail detail;
1737
1738         detail.cache = cache;
1739         bio_list_init(&detail.bios_for_issue);
1740         bio_list_init(&detail.unhandled_bios);
1741         detail.any_writes = false;
1742
1743         spin_lock_irqsave(&cache->lock, flags);
1744         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1745         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1746         spin_unlock_irqrestore(&cache->lock, flags);
1747
1748         remap_to_cache(cache, cell->holder, cblock);
1749         if (issue_holder)
1750                 issue(cache, cell->holder);
1751         else
1752                 accounted_begin(cache, cell->holder);
1753
1754         if (detail.any_writes) {
1755                 set_dirty(cache, oblock, cblock);
1756                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1757         }
1758
1759         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1760                 remap_to_cache(cache, bio, cblock);
1761                 issue(cache, bio);
1762         }
1763
1764         free_prison_cell(cache, cell);
1765 }
1766
1767 /*----------------------------------------------------------------*/
1768
1769 struct old_oblock_lock {
1770         struct policy_locker locker;
1771         struct cache *cache;
1772         struct prealloc *structs;
1773         struct dm_bio_prison_cell *cell;
1774 };
1775
1776 static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1777 {
1778         /* This should never be called */
1779         BUG();
1780         return 0;
1781 }
1782
1783 static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1784 {
1785         struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1786         struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1787
1788         return bio_detain(l->cache, b, NULL, cell_prealloc,
1789                           (cell_free_fn) prealloc_put_cell,
1790                           l->structs, &l->cell);
1791 }
1792
1793 static void process_cell(struct cache *cache, struct prealloc *structs,
1794                          struct dm_bio_prison_cell *new_ocell)
1795 {
1796         int r;
1797         bool release_cell = true;
1798         struct bio *bio = new_ocell->holder;
1799         dm_oblock_t block = get_bio_block(cache, bio);
1800         struct policy_result lookup_result;
1801         bool passthrough = passthrough_mode(&cache->features);
1802         bool fast_promotion, can_migrate;
1803         struct old_oblock_lock ool;
1804
1805         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1806         can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1807
1808         ool.locker.fn = cell_locker;
1809         ool.cache = cache;
1810         ool.structs = structs;
1811         ool.cell = NULL;
1812         r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1813                        bio, &ool.locker, &lookup_result);
1814
1815         if (r == -EWOULDBLOCK)
1816                 /* migration has been denied */
1817                 lookup_result.op = POLICY_MISS;
1818
1819         switch (lookup_result.op) {
1820         case POLICY_HIT:
1821                 if (passthrough) {
1822                         inc_miss_counter(cache, bio);
1823
1824                         /*
1825                          * Passthrough always maps to the origin,
1826                          * invalidating any cache blocks that are written
1827                          * to.
1828                          */
1829
1830                         if (bio_data_dir(bio) == WRITE) {
1831                                 atomic_inc(&cache->stats.demotion);
1832                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1833                                 release_cell = false;
1834
1835                         } else {
1836                                 /* FIXME: factor out issue_origin() */
1837                                 remap_to_origin_clear_discard(cache, bio, block);
1838                                 inc_and_issue(cache, bio, new_ocell);
1839                         }
1840                 } else {
1841                         inc_hit_counter(cache, bio);
1842
1843                         if (bio_data_dir(bio) == WRITE &&
1844                             writethrough_mode(&cache->features) &&
1845                             !is_dirty(cache, lookup_result.cblock)) {
1846                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1847                                 inc_and_issue(cache, bio, new_ocell);
1848
1849                         } else {
1850                                 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1851                                 release_cell = false;
1852                         }
1853                 }
1854
1855                 break;
1856
1857         case POLICY_MISS:
1858                 inc_miss_counter(cache, bio);
1859                 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1860                 release_cell = false;
1861                 break;
1862
1863         case POLICY_NEW:
1864                 atomic_inc(&cache->stats.promotion);
1865                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1866                 release_cell = false;
1867                 break;
1868
1869         case POLICY_REPLACE:
1870                 atomic_inc(&cache->stats.demotion);
1871                 atomic_inc(&cache->stats.promotion);
1872                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1873                                     block, lookup_result.cblock,
1874                                     ool.cell, new_ocell);
1875                 release_cell = false;
1876                 break;
1877
1878         default:
1879                 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1880                             cache_device_name(cache), __func__,
1881                             (unsigned) lookup_result.op);
1882                 bio_io_error(bio);
1883         }
1884
1885         if (release_cell)
1886                 cell_defer(cache, new_ocell, false);
1887 }
1888
1889 static void process_bio(struct cache *cache, struct prealloc *structs,
1890                         struct bio *bio)
1891 {
1892         int r;
1893         dm_oblock_t block = get_bio_block(cache, bio);
1894         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1895
1896         /*
1897          * Check to see if that block is currently migrating.
1898          */
1899         cell_prealloc = prealloc_get_cell(structs);
1900         r = bio_detain(cache, block, bio, cell_prealloc,
1901                        (cell_free_fn) prealloc_put_cell,
1902                        structs, &new_ocell);
1903         if (r > 0)
1904                 return;
1905
1906         process_cell(cache, structs, new_ocell);
1907 }
1908
1909 static int need_commit_due_to_time(struct cache *cache)
1910 {
1911         return jiffies < cache->last_commit_jiffies ||
1912                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1913 }
1914
1915 /*
1916  * A non-zero return indicates read_only or fail_io mode.
1917  */
1918 static int commit(struct cache *cache, bool clean_shutdown)
1919 {
1920         int r;
1921
1922         if (get_cache_mode(cache) >= CM_READ_ONLY)
1923                 return -EINVAL;
1924
1925         atomic_inc(&cache->stats.commit_count);
1926         r = dm_cache_commit(cache->cmd, clean_shutdown);
1927         if (r)
1928                 metadata_operation_failed(cache, "dm_cache_commit", r);
1929
1930         return r;
1931 }
1932
1933 static int commit_if_needed(struct cache *cache)
1934 {
1935         int r = 0;
1936
1937         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1938             dm_cache_changed_this_transaction(cache->cmd)) {
1939                 r = commit(cache, false);
1940                 cache->commit_requested = false;
1941                 cache->last_commit_jiffies = jiffies;
1942         }
1943
1944         return r;
1945 }
1946
1947 static void process_deferred_bios(struct cache *cache)
1948 {
1949         bool prealloc_used = false;
1950         unsigned long flags;
1951         struct bio_list bios;
1952         struct bio *bio;
1953         struct prealloc structs;
1954
1955         memset(&structs, 0, sizeof(structs));
1956         bio_list_init(&bios);
1957
1958         spin_lock_irqsave(&cache->lock, flags);
1959         bio_list_merge(&bios, &cache->deferred_bios);
1960         bio_list_init(&cache->deferred_bios);
1961         spin_unlock_irqrestore(&cache->lock, flags);
1962
1963         while (!bio_list_empty(&bios)) {
1964                 /*
1965                  * If we've got no free migration structs, and processing
1966                  * this bio might require one, we pause until there are some
1967                  * prepared mappings to process.
1968                  */
1969                 prealloc_used = true;
1970                 if (prealloc_data_structs(cache, &structs)) {
1971                         spin_lock_irqsave(&cache->lock, flags);
1972                         bio_list_merge(&cache->deferred_bios, &bios);
1973                         spin_unlock_irqrestore(&cache->lock, flags);
1974                         break;
1975                 }
1976
1977                 bio = bio_list_pop(&bios);
1978
1979                 if (bio->bi_rw & REQ_FLUSH)
1980                         process_flush_bio(cache, bio);
1981                 else if (bio->bi_rw & REQ_DISCARD)
1982                         process_discard_bio(cache, &structs, bio);
1983                 else
1984                         process_bio(cache, &structs, bio);
1985         }
1986
1987         if (prealloc_used)
1988                 prealloc_free_structs(cache, &structs);
1989 }
1990
1991 static void process_deferred_cells(struct cache *cache)
1992 {
1993         bool prealloc_used = false;
1994         unsigned long flags;
1995         struct dm_bio_prison_cell *cell, *tmp;
1996         struct list_head cells;
1997         struct prealloc structs;
1998
1999         memset(&structs, 0, sizeof(structs));
2000
2001         INIT_LIST_HEAD(&cells);
2002
2003         spin_lock_irqsave(&cache->lock, flags);
2004         list_splice_init(&cache->deferred_cells, &cells);
2005         spin_unlock_irqrestore(&cache->lock, flags);
2006
2007         list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2008                 /*
2009                  * If we've got no free migration structs, and processing
2010                  * this bio might require one, we pause until there are some
2011                  * prepared mappings to process.
2012                  */
2013                 prealloc_used = true;
2014                 if (prealloc_data_structs(cache, &structs)) {
2015                         spin_lock_irqsave(&cache->lock, flags);
2016                         list_splice(&cells, &cache->deferred_cells);
2017                         spin_unlock_irqrestore(&cache->lock, flags);
2018                         break;
2019                 }
2020
2021                 process_cell(cache, &structs, cell);
2022         }
2023
2024         if (prealloc_used)
2025                 prealloc_free_structs(cache, &structs);
2026 }
2027
2028 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2029 {
2030         unsigned long flags;
2031         struct bio_list bios;
2032         struct bio *bio;
2033
2034         bio_list_init(&bios);
2035
2036         spin_lock_irqsave(&cache->lock, flags);
2037         bio_list_merge(&bios, &cache->deferred_flush_bios);
2038         bio_list_init(&cache->deferred_flush_bios);
2039         spin_unlock_irqrestore(&cache->lock, flags);
2040
2041         /*
2042          * These bios have already been through inc_ds()
2043          */
2044         while ((bio = bio_list_pop(&bios)))
2045                 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2046 }
2047
2048 static void process_deferred_writethrough_bios(struct cache *cache)
2049 {
2050         unsigned long flags;
2051         struct bio_list bios;
2052         struct bio *bio;
2053
2054         bio_list_init(&bios);
2055
2056         spin_lock_irqsave(&cache->lock, flags);
2057         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2058         bio_list_init(&cache->deferred_writethrough_bios);
2059         spin_unlock_irqrestore(&cache->lock, flags);
2060
2061         /*
2062          * These bios have already been through inc_ds()
2063          */
2064         while ((bio = bio_list_pop(&bios)))
2065                 accounted_request(cache, bio);
2066 }
2067
2068 static void writeback_some_dirty_blocks(struct cache *cache)
2069 {
2070         bool prealloc_used = false;
2071         dm_oblock_t oblock;
2072         dm_cblock_t cblock;
2073         struct prealloc structs;
2074         struct dm_bio_prison_cell *old_ocell;
2075         bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2076
2077         memset(&structs, 0, sizeof(structs));
2078
2079         while (spare_migration_bandwidth(cache)) {
2080                 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2081                         break; /* no work to do */
2082
2083                 prealloc_used = true;
2084                 if (prealloc_data_structs(cache, &structs) ||
2085                     get_cell(cache, oblock, &structs, &old_ocell)) {
2086                         policy_set_dirty(cache->policy, oblock);
2087                         break;
2088                 }
2089
2090                 writeback(cache, &structs, oblock, cblock, old_ocell);
2091         }
2092
2093         if (prealloc_used)
2094                 prealloc_free_structs(cache, &structs);
2095 }
2096
2097 /*----------------------------------------------------------------
2098  * Invalidations.
2099  * Dropping something from the cache *without* writing back.
2100  *--------------------------------------------------------------*/
2101
2102 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2103 {
2104         int r = 0;
2105         uint64_t begin = from_cblock(req->cblocks->begin);
2106         uint64_t end = from_cblock(req->cblocks->end);
2107
2108         while (begin != end) {
2109                 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2110                 if (!r) {
2111                         r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2112                         if (r) {
2113                                 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2114                                 break;
2115                         }
2116
2117                 } else if (r == -ENODATA) {
2118                         /* harmless, already unmapped */
2119                         r = 0;
2120
2121                 } else {
2122                         DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2123                         break;
2124                 }
2125
2126                 begin++;
2127         }
2128
2129         cache->commit_requested = true;
2130
2131         req->err = r;
2132         atomic_set(&req->complete, 1);
2133
2134         wake_up(&req->result_wait);
2135 }
2136
2137 static void process_invalidation_requests(struct cache *cache)
2138 {
2139         struct list_head list;
2140         struct invalidation_request *req, *tmp;
2141
2142         INIT_LIST_HEAD(&list);
2143         spin_lock(&cache->invalidation_lock);
2144         list_splice_init(&cache->invalidation_requests, &list);
2145         spin_unlock(&cache->invalidation_lock);
2146
2147         list_for_each_entry_safe (req, tmp, &list, list)
2148                 process_invalidation_request(cache, req);
2149 }
2150
2151 /*----------------------------------------------------------------
2152  * Main worker loop
2153  *--------------------------------------------------------------*/
2154 static bool is_quiescing(struct cache *cache)
2155 {
2156         return atomic_read(&cache->quiescing);
2157 }
2158
2159 static void ack_quiescing(struct cache *cache)
2160 {
2161         if (is_quiescing(cache)) {
2162                 atomic_inc(&cache->quiescing_ack);
2163                 wake_up(&cache->quiescing_wait);
2164         }
2165 }
2166
2167 static void wait_for_quiescing_ack(struct cache *cache)
2168 {
2169         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2170 }
2171
2172 static void start_quiescing(struct cache *cache)
2173 {
2174         atomic_inc(&cache->quiescing);
2175         wait_for_quiescing_ack(cache);
2176 }
2177
2178 static void stop_quiescing(struct cache *cache)
2179 {
2180         atomic_set(&cache->quiescing, 0);
2181         atomic_set(&cache->quiescing_ack, 0);
2182 }
2183
2184 static void wait_for_migrations(struct cache *cache)
2185 {
2186         wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2187 }
2188
2189 static void stop_worker(struct cache *cache)
2190 {
2191         cancel_delayed_work(&cache->waker);
2192         flush_workqueue(cache->wq);
2193 }
2194
2195 static void requeue_deferred_cells(struct cache *cache)
2196 {
2197         unsigned long flags;
2198         struct list_head cells;
2199         struct dm_bio_prison_cell *cell, *tmp;
2200
2201         INIT_LIST_HEAD(&cells);
2202         spin_lock_irqsave(&cache->lock, flags);
2203         list_splice_init(&cache->deferred_cells, &cells);
2204         spin_unlock_irqrestore(&cache->lock, flags);
2205
2206         list_for_each_entry_safe(cell, tmp, &cells, user_list)
2207                 cell_requeue(cache, cell);
2208 }
2209
2210 static void requeue_deferred_bios(struct cache *cache)
2211 {
2212         struct bio *bio;
2213         struct bio_list bios;
2214
2215         bio_list_init(&bios);
2216         bio_list_merge(&bios, &cache->deferred_bios);
2217         bio_list_init(&cache->deferred_bios);
2218
2219         while ((bio = bio_list_pop(&bios))) {
2220                 bio->bi_error = DM_ENDIO_REQUEUE;
2221                 bio_endio(bio);
2222         }
2223 }
2224
2225 static int more_work(struct cache *cache)
2226 {
2227         if (is_quiescing(cache))
2228                 return !list_empty(&cache->quiesced_migrations) ||
2229                         !list_empty(&cache->completed_migrations) ||
2230                         !list_empty(&cache->need_commit_migrations);
2231         else
2232                 return !bio_list_empty(&cache->deferred_bios) ||
2233                         !list_empty(&cache->deferred_cells) ||
2234                         !bio_list_empty(&cache->deferred_flush_bios) ||
2235                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
2236                         !list_empty(&cache->quiesced_migrations) ||
2237                         !list_empty(&cache->completed_migrations) ||
2238                         !list_empty(&cache->need_commit_migrations) ||
2239                         cache->invalidate;
2240 }
2241
2242 static void do_worker(struct work_struct *ws)
2243 {
2244         struct cache *cache = container_of(ws, struct cache, worker);
2245
2246         do {
2247                 if (!is_quiescing(cache)) {
2248                         writeback_some_dirty_blocks(cache);
2249                         process_deferred_writethrough_bios(cache);
2250                         process_deferred_bios(cache);
2251                         process_deferred_cells(cache);
2252                         process_invalidation_requests(cache);
2253                 }
2254
2255                 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2256                 process_migrations(cache, &cache->completed_migrations, complete_migration);
2257
2258                 if (commit_if_needed(cache)) {
2259                         process_deferred_flush_bios(cache, false);
2260                         process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2261                 } else {
2262                         process_deferred_flush_bios(cache, true);
2263                         process_migrations(cache, &cache->need_commit_migrations,
2264                                            migration_success_post_commit);
2265                 }
2266
2267                 ack_quiescing(cache);
2268
2269         } while (more_work(cache));
2270 }
2271
2272 /*
2273  * We want to commit periodically so that not too much
2274  * unwritten metadata builds up.
2275  */
2276 static void do_waker(struct work_struct *ws)
2277 {
2278         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2279         policy_tick(cache->policy, true);
2280         wake_worker(cache);
2281         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2282 }
2283
2284 /*----------------------------------------------------------------*/
2285
2286 static int is_congested(struct dm_dev *dev, int bdi_bits)
2287 {
2288         struct request_queue *q = bdev_get_queue(dev->bdev);
2289         return bdi_congested(&q->backing_dev_info, bdi_bits);
2290 }
2291
2292 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2293 {
2294         struct cache *cache = container_of(cb, struct cache, callbacks);
2295
2296         return is_congested(cache->origin_dev, bdi_bits) ||
2297                 is_congested(cache->cache_dev, bdi_bits);
2298 }
2299
2300 /*----------------------------------------------------------------
2301  * Target methods
2302  *--------------------------------------------------------------*/
2303
2304 /*
2305  * This function gets called on the error paths of the constructor, so we
2306  * have to cope with a partially initialised struct.
2307  */
2308 static void destroy(struct cache *cache)
2309 {
2310         unsigned i;
2311
2312         mempool_destroy(cache->migration_pool);
2313
2314         if (cache->all_io_ds)
2315                 dm_deferred_set_destroy(cache->all_io_ds);
2316
2317         if (cache->prison)
2318                 dm_bio_prison_destroy(cache->prison);
2319
2320         if (cache->wq)
2321                 destroy_workqueue(cache->wq);
2322
2323         if (cache->dirty_bitset)
2324                 free_bitset(cache->dirty_bitset);
2325
2326         if (cache->discard_bitset)
2327                 free_bitset(cache->discard_bitset);
2328
2329         if (cache->copier)
2330                 dm_kcopyd_client_destroy(cache->copier);
2331
2332         if (cache->cmd)
2333                 dm_cache_metadata_close(cache->cmd);
2334
2335         if (cache->metadata_dev)
2336                 dm_put_device(cache->ti, cache->metadata_dev);
2337
2338         if (cache->origin_dev)
2339                 dm_put_device(cache->ti, cache->origin_dev);
2340
2341         if (cache->cache_dev)
2342                 dm_put_device(cache->ti, cache->cache_dev);
2343
2344         if (cache->policy)
2345                 dm_cache_policy_destroy(cache->policy);
2346
2347         for (i = 0; i < cache->nr_ctr_args ; i++)
2348                 kfree(cache->ctr_args[i]);
2349         kfree(cache->ctr_args);
2350
2351         kfree(cache);
2352 }
2353
2354 static void cache_dtr(struct dm_target *ti)
2355 {
2356         struct cache *cache = ti->private;
2357
2358         destroy(cache);
2359 }
2360
2361 static sector_t get_dev_size(struct dm_dev *dev)
2362 {
2363         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2364 }
2365
2366 /*----------------------------------------------------------------*/
2367
2368 /*
2369  * Construct a cache device mapping.
2370  *
2371  * cache <metadata dev> <cache dev> <origin dev> <block size>
2372  *       <#feature args> [<feature arg>]*
2373  *       <policy> <#policy args> [<policy arg>]*
2374  *
2375  * metadata dev    : fast device holding the persistent metadata
2376  * cache dev       : fast device holding cached data blocks
2377  * origin dev      : slow device holding original data blocks
2378  * block size      : cache unit size in sectors
2379  *
2380  * #feature args   : number of feature arguments passed
2381  * feature args    : writethrough.  (The default is writeback.)
2382  *
2383  * policy          : the replacement policy to use
2384  * #policy args    : an even number of policy arguments corresponding
2385  *                   to key/value pairs passed to the policy
2386  * policy args     : key/value pairs passed to the policy
2387  *                   E.g. 'sequential_threshold 1024'
2388  *                   See cache-policies.txt for details.
2389  *
2390  * Optional feature arguments are:
2391  *   writethrough  : write through caching that prohibits cache block
2392  *                   content from being different from origin block content.
2393  *                   Without this argument, the default behaviour is to write
2394  *                   back cache block contents later for performance reasons,
2395  *                   so they may differ from the corresponding origin blocks.
2396  */
2397 struct cache_args {
2398         struct dm_target *ti;
2399
2400         struct dm_dev *metadata_dev;
2401
2402         struct dm_dev *cache_dev;
2403         sector_t cache_sectors;
2404
2405         struct dm_dev *origin_dev;
2406         sector_t origin_sectors;
2407
2408         uint32_t block_size;
2409
2410         const char *policy_name;
2411         int policy_argc;
2412         const char **policy_argv;
2413
2414         struct cache_features features;
2415 };
2416
2417 static void destroy_cache_args(struct cache_args *ca)
2418 {
2419         if (ca->metadata_dev)
2420                 dm_put_device(ca->ti, ca->metadata_dev);
2421
2422         if (ca->cache_dev)
2423                 dm_put_device(ca->ti, ca->cache_dev);
2424
2425         if (ca->origin_dev)
2426                 dm_put_device(ca->ti, ca->origin_dev);
2427
2428         kfree(ca);
2429 }
2430
2431 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2432 {
2433         if (!as->argc) {
2434                 *error = "Insufficient args";
2435                 return false;
2436         }
2437
2438         return true;
2439 }
2440
2441 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2442                               char **error)
2443 {
2444         int r;
2445         sector_t metadata_dev_size;
2446         char b[BDEVNAME_SIZE];
2447
2448         if (!at_least_one_arg(as, error))
2449                 return -EINVAL;
2450
2451         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2452                           &ca->metadata_dev);
2453         if (r) {
2454                 *error = "Error opening metadata device";
2455                 return r;
2456         }
2457
2458         metadata_dev_size = get_dev_size(ca->metadata_dev);
2459         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2460                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2461                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2462
2463         return 0;
2464 }
2465
2466 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2467                            char **error)
2468 {
2469         int r;
2470
2471         if (!at_least_one_arg(as, error))
2472                 return -EINVAL;
2473
2474         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2475                           &ca->cache_dev);
2476         if (r) {
2477                 *error = "Error opening cache device";
2478                 return r;
2479         }
2480         ca->cache_sectors = get_dev_size(ca->cache_dev);
2481
2482         return 0;
2483 }
2484
2485 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2486                             char **error)
2487 {
2488         int r;
2489
2490         if (!at_least_one_arg(as, error))
2491                 return -EINVAL;
2492
2493         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2494                           &ca->origin_dev);
2495         if (r) {
2496                 *error = "Error opening origin device";
2497                 return r;
2498         }
2499
2500         ca->origin_sectors = get_dev_size(ca->origin_dev);
2501         if (ca->ti->len > ca->origin_sectors) {
2502                 *error = "Device size larger than cached device";
2503                 return -EINVAL;
2504         }
2505
2506         return 0;
2507 }
2508
2509 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2510                             char **error)
2511 {
2512         unsigned long block_size;
2513
2514         if (!at_least_one_arg(as, error))
2515                 return -EINVAL;
2516
2517         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2518             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2519             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2520             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2521                 *error = "Invalid data block size";
2522                 return -EINVAL;
2523         }
2524
2525         if (block_size > ca->cache_sectors) {
2526                 *error = "Data block size is larger than the cache device";
2527                 return -EINVAL;
2528         }
2529
2530         ca->block_size = block_size;
2531
2532         return 0;
2533 }
2534
2535 static void init_features(struct cache_features *cf)
2536 {
2537         cf->mode = CM_WRITE;
2538         cf->io_mode = CM_IO_WRITEBACK;
2539 }
2540
2541 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2542                           char **error)
2543 {
2544         static struct dm_arg _args[] = {
2545                 {0, 1, "Invalid number of cache feature arguments"},
2546         };
2547
2548         int r;
2549         unsigned argc;
2550         const char *arg;
2551         struct cache_features *cf = &ca->features;
2552
2553         init_features(cf);
2554
2555         r = dm_read_arg_group(_args, as, &argc, error);
2556         if (r)
2557                 return -EINVAL;
2558
2559         while (argc--) {
2560                 arg = dm_shift_arg(as);
2561
2562                 if (!strcasecmp(arg, "writeback"))
2563                         cf->io_mode = CM_IO_WRITEBACK;
2564
2565                 else if (!strcasecmp(arg, "writethrough"))
2566                         cf->io_mode = CM_IO_WRITETHROUGH;
2567
2568                 else if (!strcasecmp(arg, "passthrough"))
2569                         cf->io_mode = CM_IO_PASSTHROUGH;
2570
2571                 else {
2572                         *error = "Unrecognised cache feature requested";
2573                         return -EINVAL;
2574                 }
2575         }
2576
2577         return 0;
2578 }
2579
2580 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2581                         char **error)
2582 {
2583         static struct dm_arg _args[] = {
2584                 {0, 1024, "Invalid number of policy arguments"},
2585         };
2586
2587         int r;
2588
2589         if (!at_least_one_arg(as, error))
2590                 return -EINVAL;
2591
2592         ca->policy_name = dm_shift_arg(as);
2593
2594         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2595         if (r)
2596                 return -EINVAL;
2597
2598         ca->policy_argv = (const char **)as->argv;
2599         dm_consume_args(as, ca->policy_argc);
2600
2601         return 0;
2602 }
2603
2604 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2605                             char **error)
2606 {
2607         int r;
2608         struct dm_arg_set as;
2609
2610         as.argc = argc;
2611         as.argv = argv;
2612
2613         r = parse_metadata_dev(ca, &as, error);
2614         if (r)
2615                 return r;
2616
2617         r = parse_cache_dev(ca, &as, error);
2618         if (r)
2619                 return r;
2620
2621         r = parse_origin_dev(ca, &as, error);
2622         if (r)
2623                 return r;
2624
2625         r = parse_block_size(ca, &as, error);
2626         if (r)
2627                 return r;
2628
2629         r = parse_features(ca, &as, error);
2630         if (r)
2631                 return r;
2632
2633         r = parse_policy(ca, &as, error);
2634         if (r)
2635                 return r;
2636
2637         return 0;
2638 }
2639
2640 /*----------------------------------------------------------------*/
2641
2642 static struct kmem_cache *migration_cache;
2643
2644 #define NOT_CORE_OPTION 1
2645
2646 static int process_config_option(struct cache *cache, const char *key, const char *value)
2647 {
2648         unsigned long tmp;
2649
2650         if (!strcasecmp(key, "migration_threshold")) {
2651                 if (kstrtoul(value, 10, &tmp))
2652                         return -EINVAL;
2653
2654                 cache->migration_threshold = tmp;
2655                 return 0;
2656         }
2657
2658         return NOT_CORE_OPTION;
2659 }
2660
2661 static int set_config_value(struct cache *cache, const char *key, const char *value)
2662 {
2663         int r = process_config_option(cache, key, value);
2664
2665         if (r == NOT_CORE_OPTION)
2666                 r = policy_set_config_value(cache->policy, key, value);
2667
2668         if (r)
2669                 DMWARN("bad config value for %s: %s", key, value);
2670
2671         return r;
2672 }
2673
2674 static int set_config_values(struct cache *cache, int argc, const char **argv)
2675 {
2676         int r = 0;
2677
2678         if (argc & 1) {
2679                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2680                 return -EINVAL;
2681         }
2682
2683         while (argc) {
2684                 r = set_config_value(cache, argv[0], argv[1]);
2685                 if (r)
2686                         break;
2687
2688                 argc -= 2;
2689                 argv += 2;
2690         }
2691
2692         return r;
2693 }
2694
2695 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2696                                char **error)
2697 {
2698         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2699                                                            cache->cache_size,
2700                                                            cache->origin_sectors,
2701                                                            cache->sectors_per_block);
2702         if (IS_ERR(p)) {
2703                 *error = "Error creating cache's policy";
2704                 return PTR_ERR(p);
2705         }
2706         cache->policy = p;
2707
2708         return 0;
2709 }
2710
2711 /*
2712  * We want the discard block size to be at least the size of the cache
2713  * block size and have no more than 2^14 discard blocks across the origin.
2714  */
2715 #define MAX_DISCARD_BLOCKS (1 << 14)
2716
2717 static bool too_many_discard_blocks(sector_t discard_block_size,
2718                                     sector_t origin_size)
2719 {
2720         (void) sector_div(origin_size, discard_block_size);
2721
2722         return origin_size > MAX_DISCARD_BLOCKS;
2723 }
2724
2725 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2726                                              sector_t origin_size)
2727 {
2728         sector_t discard_block_size = cache_block_size;
2729
2730         if (origin_size)
2731                 while (too_many_discard_blocks(discard_block_size, origin_size))
2732                         discard_block_size *= 2;
2733
2734         return discard_block_size;
2735 }
2736
2737 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2738 {
2739         dm_block_t nr_blocks = from_cblock(size);
2740
2741         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2742                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2743                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2744                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2745                              (unsigned long long) nr_blocks);
2746
2747         cache->cache_size = size;
2748 }
2749
2750 #define DEFAULT_MIGRATION_THRESHOLD 2048
2751
2752 static int cache_create(struct cache_args *ca, struct cache **result)
2753 {
2754         int r = 0;
2755         char **error = &ca->ti->error;
2756         struct cache *cache;
2757         struct dm_target *ti = ca->ti;
2758         dm_block_t origin_blocks;
2759         struct dm_cache_metadata *cmd;
2760         bool may_format = ca->features.mode == CM_WRITE;
2761
2762         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2763         if (!cache)
2764                 return -ENOMEM;
2765
2766         cache->ti = ca->ti;
2767         ti->private = cache;
2768         ti->num_flush_bios = 2;
2769         ti->flush_supported = true;
2770
2771         ti->num_discard_bios = 1;
2772         ti->discards_supported = true;
2773         ti->discard_zeroes_data_unsupported = true;
2774         ti->split_discard_bios = false;
2775
2776         cache->features = ca->features;
2777         ti->per_bio_data_size = get_per_bio_data_size(cache);
2778
2779         cache->callbacks.congested_fn = cache_is_congested;
2780         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2781
2782         cache->metadata_dev = ca->metadata_dev;
2783         cache->origin_dev = ca->origin_dev;
2784         cache->cache_dev = ca->cache_dev;
2785
2786         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2787
2788         /* FIXME: factor out this whole section */
2789         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2790         origin_blocks = block_div(origin_blocks, ca->block_size);
2791         cache->origin_blocks = to_oblock(origin_blocks);
2792
2793         cache->sectors_per_block = ca->block_size;
2794         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2795                 r = -EINVAL;
2796                 goto bad;
2797         }
2798
2799         if (ca->block_size & (ca->block_size - 1)) {
2800                 dm_block_t cache_size = ca->cache_sectors;
2801
2802                 cache->sectors_per_block_shift = -1;
2803                 cache_size = block_div(cache_size, ca->block_size);
2804                 set_cache_size(cache, to_cblock(cache_size));
2805         } else {
2806                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2807                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2808         }
2809
2810         r = create_cache_policy(cache, ca, error);
2811         if (r)
2812                 goto bad;
2813
2814         cache->policy_nr_args = ca->policy_argc;
2815         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2816
2817         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2818         if (r) {
2819                 *error = "Error setting cache policy's config values";
2820                 goto bad;
2821         }
2822
2823         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2824                                      ca->block_size, may_format,
2825                                      dm_cache_policy_get_hint_size(cache->policy));
2826         if (IS_ERR(cmd)) {
2827                 *error = "Error creating metadata object";
2828                 r = PTR_ERR(cmd);
2829                 goto bad;
2830         }
2831         cache->cmd = cmd;
2832         set_cache_mode(cache, CM_WRITE);
2833         if (get_cache_mode(cache) != CM_WRITE) {
2834                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2835                 r = -EINVAL;
2836                 goto bad;
2837         }
2838
2839         if (passthrough_mode(&cache->features)) {
2840                 bool all_clean;
2841
2842                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2843                 if (r) {
2844                         *error = "dm_cache_metadata_all_clean() failed";
2845                         goto bad;
2846                 }
2847
2848                 if (!all_clean) {
2849                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2850                         r = -EINVAL;
2851                         goto bad;
2852                 }
2853         }
2854
2855         spin_lock_init(&cache->lock);
2856         INIT_LIST_HEAD(&cache->deferred_cells);
2857         bio_list_init(&cache->deferred_bios);
2858         bio_list_init(&cache->deferred_flush_bios);
2859         bio_list_init(&cache->deferred_writethrough_bios);
2860         INIT_LIST_HEAD(&cache->quiesced_migrations);
2861         INIT_LIST_HEAD(&cache->completed_migrations);
2862         INIT_LIST_HEAD(&cache->need_commit_migrations);
2863         atomic_set(&cache->nr_allocated_migrations, 0);
2864         atomic_set(&cache->nr_io_migrations, 0);
2865         init_waitqueue_head(&cache->migration_wait);
2866
2867         init_waitqueue_head(&cache->quiescing_wait);
2868         atomic_set(&cache->quiescing, 0);
2869         atomic_set(&cache->quiescing_ack, 0);
2870
2871         r = -ENOMEM;
2872         atomic_set(&cache->nr_dirty, 0);
2873         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2874         if (!cache->dirty_bitset) {
2875                 *error = "could not allocate dirty bitset";
2876                 goto bad;
2877         }
2878         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2879
2880         cache->discard_block_size =
2881                 calculate_discard_block_size(cache->sectors_per_block,
2882                                              cache->origin_sectors);
2883         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2884                                                               cache->discard_block_size));
2885         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2886         if (!cache->discard_bitset) {
2887                 *error = "could not allocate discard bitset";
2888                 goto bad;
2889         }
2890         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2891
2892         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2893         if (IS_ERR(cache->copier)) {
2894                 *error = "could not create kcopyd client";
2895                 r = PTR_ERR(cache->copier);
2896                 goto bad;
2897         }
2898
2899         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2900         if (!cache->wq) {
2901                 *error = "could not create workqueue for metadata object";
2902                 goto bad;
2903         }
2904         INIT_WORK(&cache->worker, do_worker);
2905         INIT_DELAYED_WORK(&cache->waker, do_waker);
2906         cache->last_commit_jiffies = jiffies;
2907
2908         cache->prison = dm_bio_prison_create();
2909         if (!cache->prison) {
2910                 *error = "could not create bio prison";
2911                 goto bad;
2912         }
2913
2914         cache->all_io_ds = dm_deferred_set_create();
2915         if (!cache->all_io_ds) {
2916                 *error = "could not create all_io deferred set";
2917                 goto bad;
2918         }
2919
2920         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2921                                                          migration_cache);
2922         if (!cache->migration_pool) {
2923                 *error = "Error creating cache's migration mempool";
2924                 goto bad;
2925         }
2926
2927         cache->need_tick_bio = true;
2928         cache->sized = false;
2929         cache->invalidate = false;
2930         cache->commit_requested = false;
2931         cache->loaded_mappings = false;
2932         cache->loaded_discards = false;
2933
2934         load_stats(cache);
2935
2936         atomic_set(&cache->stats.demotion, 0);
2937         atomic_set(&cache->stats.promotion, 0);
2938         atomic_set(&cache->stats.copies_avoided, 0);
2939         atomic_set(&cache->stats.cache_cell_clash, 0);
2940         atomic_set(&cache->stats.commit_count, 0);
2941         atomic_set(&cache->stats.discard_count, 0);
2942
2943         spin_lock_init(&cache->invalidation_lock);
2944         INIT_LIST_HEAD(&cache->invalidation_requests);
2945
2946         iot_init(&cache->origin_tracker);
2947
2948         *result = cache;
2949         return 0;
2950
2951 bad:
2952         destroy(cache);
2953         return r;
2954 }
2955
2956 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2957 {
2958         unsigned i;
2959         const char **copy;
2960
2961         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2962         if (!copy)
2963                 return -ENOMEM;
2964         for (i = 0; i < argc; i++) {
2965                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2966                 if (!copy[i]) {
2967                         while (i--)
2968                                 kfree(copy[i]);
2969                         kfree(copy);
2970                         return -ENOMEM;
2971                 }
2972         }
2973
2974         cache->nr_ctr_args = argc;
2975         cache->ctr_args = copy;
2976
2977         return 0;
2978 }
2979
2980 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2981 {
2982         int r = -EINVAL;
2983         struct cache_args *ca;
2984         struct cache *cache = NULL;
2985
2986         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2987         if (!ca) {
2988                 ti->error = "Error allocating memory for cache";
2989                 return -ENOMEM;
2990         }
2991         ca->ti = ti;
2992
2993         r = parse_cache_args(ca, argc, argv, &ti->error);
2994         if (r)
2995                 goto out;
2996
2997         r = cache_create(ca, &cache);
2998         if (r)
2999                 goto out;
3000
3001         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
3002         if (r) {
3003                 destroy(cache);
3004                 goto out;
3005         }
3006
3007         ti->private = cache;
3008
3009 out:
3010         destroy_cache_args(ca);
3011         return r;
3012 }
3013
3014 /*----------------------------------------------------------------*/
3015
3016 static int cache_map(struct dm_target *ti, struct bio *bio)
3017 {
3018         struct cache *cache = ti->private;
3019
3020         int r;
3021         struct dm_bio_prison_cell *cell = NULL;
3022         dm_oblock_t block = get_bio_block(cache, bio);
3023         size_t pb_data_size = get_per_bio_data_size(cache);
3024         bool can_migrate = false;
3025         bool fast_promotion;
3026         struct policy_result lookup_result;
3027         struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3028         struct old_oblock_lock ool;
3029
3030         ool.locker.fn = null_locker;
3031
3032         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3033                 /*
3034                  * This can only occur if the io goes to a partial block at
3035                  * the end of the origin device.  We don't cache these.
3036                  * Just remap to the origin and carry on.
3037                  */
3038                 remap_to_origin(cache, bio);
3039                 accounted_begin(cache, bio);
3040                 return DM_MAPIO_REMAPPED;
3041         }
3042
3043         if (discard_or_flush(bio)) {
3044                 defer_bio(cache, bio);
3045                 return DM_MAPIO_SUBMITTED;
3046         }
3047
3048         /*
3049          * Check to see if that block is currently migrating.
3050          */
3051         cell = alloc_prison_cell(cache);
3052         if (!cell) {
3053                 defer_bio(cache, bio);
3054                 return DM_MAPIO_SUBMITTED;
3055         }
3056
3057         r = bio_detain(cache, block, bio, cell,
3058                        (cell_free_fn) free_prison_cell,
3059                        cache, &cell);
3060         if (r) {
3061                 if (r < 0)
3062                         defer_bio(cache, bio);
3063
3064                 return DM_MAPIO_SUBMITTED;
3065         }
3066
3067         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3068
3069         r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3070                        bio, &ool.locker, &lookup_result);
3071         if (r == -EWOULDBLOCK) {
3072                 cell_defer(cache, cell, true);
3073                 return DM_MAPIO_SUBMITTED;
3074
3075         } else if (r) {
3076                 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3077                             cache_device_name(cache), r);
3078                 cell_defer(cache, cell, false);
3079                 bio_io_error(bio);
3080                 return DM_MAPIO_SUBMITTED;
3081         }
3082
3083         r = DM_MAPIO_REMAPPED;
3084         switch (lookup_result.op) {
3085         case POLICY_HIT:
3086                 if (passthrough_mode(&cache->features)) {
3087                         if (bio_data_dir(bio) == WRITE) {
3088                                 /*
3089                                  * We need to invalidate this block, so
3090                                  * defer for the worker thread.
3091                                  */
3092                                 cell_defer(cache, cell, true);
3093                                 r = DM_MAPIO_SUBMITTED;
3094
3095                         } else {
3096                                 inc_miss_counter(cache, bio);
3097                                 remap_to_origin_clear_discard(cache, bio, block);
3098                                 accounted_begin(cache, bio);
3099                                 inc_ds(cache, bio, cell);
3100                                 // FIXME: we want to remap hits or misses straight
3101                                 // away rather than passing over to the worker.
3102                                 cell_defer(cache, cell, false);
3103                         }
3104
3105                 } else {
3106                         inc_hit_counter(cache, bio);
3107                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3108                             !is_dirty(cache, lookup_result.cblock)) {
3109                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3110                                 accounted_begin(cache, bio);
3111                                 inc_ds(cache, bio, cell);
3112                                 cell_defer(cache, cell, false);
3113
3114                         } else
3115                                 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3116                 }
3117                 break;
3118
3119         case POLICY_MISS:
3120                 inc_miss_counter(cache, bio);
3121                 if (pb->req_nr != 0) {
3122                         /*
3123                          * This is a duplicate writethrough io that is no
3124                          * longer needed because the block has been demoted.
3125                          */
3126                         bio_endio(bio);
3127                         // FIXME: remap everything as a miss
3128                         cell_defer(cache, cell, false);
3129                         r = DM_MAPIO_SUBMITTED;
3130
3131                 } else
3132                         remap_cell_to_origin_clear_discard(cache, cell, block, false);
3133                 break;
3134
3135         default:
3136                 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3137                             cache_device_name(cache), __func__,
3138                             (unsigned) lookup_result.op);
3139                 cell_defer(cache, cell, false);
3140                 bio_io_error(bio);
3141                 r = DM_MAPIO_SUBMITTED;
3142         }
3143
3144         return r;
3145 }
3146
3147 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3148 {
3149         struct cache *cache = ti->private;
3150         unsigned long flags;
3151         size_t pb_data_size = get_per_bio_data_size(cache);
3152         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3153
3154         if (pb->tick) {
3155                 policy_tick(cache->policy, false);
3156
3157                 spin_lock_irqsave(&cache->lock, flags);
3158                 cache->need_tick_bio = true;
3159                 spin_unlock_irqrestore(&cache->lock, flags);
3160         }
3161
3162         check_for_quiesced_migrations(cache, pb);
3163         accounted_complete(cache, bio);
3164
3165         return 0;
3166 }
3167
3168 static int write_dirty_bitset(struct cache *cache)
3169 {
3170         unsigned i, r;
3171
3172         if (get_cache_mode(cache) >= CM_READ_ONLY)
3173                 return -EINVAL;
3174
3175         for (i = 0; i < from_cblock(cache->cache_size); i++) {
3176                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3177                                        is_dirty(cache, to_cblock(i)));
3178                 if (r) {
3179                         metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3180                         return r;
3181                 }
3182         }
3183
3184         return 0;
3185 }
3186
3187 static int write_discard_bitset(struct cache *cache)
3188 {
3189         unsigned i, r;
3190
3191         if (get_cache_mode(cache) >= CM_READ_ONLY)
3192                 return -EINVAL;
3193
3194         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3195                                            cache->discard_nr_blocks);
3196         if (r) {
3197                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3198                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3199                 return r;
3200         }
3201
3202         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3203                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3204                                          is_discarded(cache, to_dblock(i)));
3205                 if (r) {
3206                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
3207                         return r;
3208                 }
3209         }
3210
3211         return 0;
3212 }
3213
3214 static int write_hints(struct cache *cache)
3215 {
3216         int r;
3217
3218         if (get_cache_mode(cache) >= CM_READ_ONLY)
3219                 return -EINVAL;
3220
3221         r = dm_cache_write_hints(cache->cmd, cache->policy);
3222         if (r) {
3223                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3224                 return r;
3225         }
3226
3227         return 0;
3228 }
3229
3230 /*
3231  * returns true on success
3232  */
3233 static bool sync_metadata(struct cache *cache)
3234 {
3235         int r1, r2, r3, r4;
3236
3237         r1 = write_dirty_bitset(cache);
3238         if (r1)
3239                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3240
3241         r2 = write_discard_bitset(cache);
3242         if (r2)
3243                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3244
3245         save_stats(cache);
3246
3247         r3 = write_hints(cache);
3248         if (r3)
3249                 DMERR("%s: could not write hints", cache_device_name(cache));
3250
3251         /*
3252          * If writing the above metadata failed, we still commit, but don't
3253          * set the clean shutdown flag.  This will effectively force every
3254          * dirty bit to be set on reload.
3255          */
3256         r4 = commit(cache, !r1 && !r2 && !r3);
3257         if (r4)
3258                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3259
3260         return !r1 && !r2 && !r3 && !r4;
3261 }
3262
3263 static void cache_postsuspend(struct dm_target *ti)
3264 {
3265         struct cache *cache = ti->private;
3266
3267         start_quiescing(cache);
3268         wait_for_migrations(cache);
3269         stop_worker(cache);
3270         requeue_deferred_bios(cache);
3271         requeue_deferred_cells(cache);
3272         stop_quiescing(cache);
3273
3274         if (get_cache_mode(cache) == CM_WRITE)
3275                 (void) sync_metadata(cache);
3276 }
3277
3278 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3279                         bool dirty, uint32_t hint, bool hint_valid)
3280 {
3281         int r;
3282         struct cache *cache = context;
3283
3284         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3285         if (r)
3286                 return r;
3287
3288         if (dirty)
3289                 set_dirty(cache, oblock, cblock);
3290         else
3291                 clear_dirty(cache, oblock, cblock);
3292
3293         return 0;
3294 }
3295
3296 /*
3297  * The discard block size in the on disk metadata is not
3298  * neccessarily the same as we're currently using.  So we have to
3299  * be careful to only set the discarded attribute if we know it
3300  * covers a complete block of the new size.
3301  */
3302 struct discard_load_info {
3303         struct cache *cache;
3304
3305         /*
3306          * These blocks are sized using the on disk dblock size, rather
3307          * than the current one.
3308          */
3309         dm_block_t block_size;
3310         dm_block_t discard_begin, discard_end;
3311 };
3312
3313 static void discard_load_info_init(struct cache *cache,
3314                                    struct discard_load_info *li)
3315 {
3316         li->cache = cache;
3317         li->discard_begin = li->discard_end = 0;
3318 }
3319
3320 static void set_discard_range(struct discard_load_info *li)
3321 {
3322         sector_t b, e;
3323
3324         if (li->discard_begin == li->discard_end)
3325                 return;
3326
3327         /*
3328          * Convert to sectors.
3329          */
3330         b = li->discard_begin * li->block_size;
3331         e = li->discard_end * li->block_size;
3332
3333         /*
3334          * Then convert back to the current dblock size.
3335          */
3336         b = dm_sector_div_up(b, li->cache->discard_block_size);
3337         sector_div(e, li->cache->discard_block_size);
3338
3339         /*
3340          * The origin may have shrunk, so we need to check we're still in
3341          * bounds.
3342          */
3343         if (e > from_dblock(li->cache->discard_nr_blocks))
3344                 e = from_dblock(li->cache->discard_nr_blocks);
3345
3346         for (; b < e; b++)
3347                 set_discard(li->cache, to_dblock(b));
3348 }
3349
3350 static int load_discard(void *context, sector_t discard_block_size,
3351                         dm_dblock_t dblock, bool discard)
3352 {
3353         struct discard_load_info *li = context;
3354
3355         li->block_size = discard_block_size;
3356
3357         if (discard) {
3358                 if (from_dblock(dblock) == li->discard_end)
3359                         /*
3360                          * We're already in a discard range, just extend it.
3361                          */
3362                         li->discard_end = li->discard_end + 1ULL;
3363
3364                 else {
3365                         /*
3366                          * Emit the old range and start a new one.
3367                          */
3368                         set_discard_range(li);
3369                         li->discard_begin = from_dblock(dblock);
3370                         li->discard_end = li->discard_begin + 1ULL;
3371                 }
3372         } else {
3373                 set_discard_range(li);
3374                 li->discard_begin = li->discard_end = 0;
3375         }
3376
3377         return 0;
3378 }
3379
3380 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3381 {
3382         sector_t size = get_dev_size(cache->cache_dev);
3383         (void) sector_div(size, cache->sectors_per_block);
3384         return to_cblock(size);
3385 }
3386
3387 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3388 {
3389         if (from_cblock(new_size) > from_cblock(cache->cache_size))
3390                 return true;
3391
3392         /*
3393          * We can't drop a dirty block when shrinking the cache.
3394          */
3395         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3396                 new_size = to_cblock(from_cblock(new_size) + 1);
3397                 if (is_dirty(cache, new_size)) {
3398                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3399                               cache_device_name(cache),
3400                               (unsigned long long) from_cblock(new_size));
3401                         return false;
3402                 }
3403         }
3404
3405         return true;
3406 }
3407
3408 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3409 {
3410         int r;
3411
3412         r = dm_cache_resize(cache->cmd, new_size);
3413         if (r) {
3414                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3415                 metadata_operation_failed(cache, "dm_cache_resize", r);
3416                 return r;
3417         }
3418
3419         set_cache_size(cache, new_size);
3420
3421         return 0;
3422 }
3423
3424 static int cache_preresume(struct dm_target *ti)
3425 {
3426         int r = 0;
3427         struct cache *cache = ti->private;
3428         dm_cblock_t csize = get_cache_dev_size(cache);
3429
3430         /*
3431          * Check to see if the cache has resized.
3432          */
3433         if (!cache->sized) {
3434                 r = resize_cache_dev(cache, csize);
3435                 if (r)
3436                         return r;
3437
3438                 cache->sized = true;
3439
3440         } else if (csize != cache->cache_size) {
3441                 if (!can_resize(cache, csize))
3442                         return -EINVAL;
3443
3444                 r = resize_cache_dev(cache, csize);
3445                 if (r)
3446                         return r;
3447         }
3448
3449         if (!cache->loaded_mappings) {
3450                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3451                                            load_mapping, cache);
3452                 if (r) {
3453                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3454                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3455                         return r;
3456                 }
3457
3458                 cache->loaded_mappings = true;
3459         }
3460
3461         if (!cache->loaded_discards) {
3462                 struct discard_load_info li;
3463
3464                 /*
3465                  * The discard bitset could have been resized, or the
3466                  * discard block size changed.  To be safe we start by
3467                  * setting every dblock to not discarded.
3468                  */
3469                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3470
3471                 discard_load_info_init(cache, &li);
3472                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3473                 if (r) {
3474                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3475                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3476                         return r;
3477                 }
3478                 set_discard_range(&li);
3479
3480                 cache->loaded_discards = true;
3481         }
3482
3483         return r;
3484 }
3485
3486 static void cache_resume(struct dm_target *ti)
3487 {
3488         struct cache *cache = ti->private;
3489
3490         cache->need_tick_bio = true;
3491         do_waker(&cache->waker.work);
3492 }
3493
3494 /*
3495  * Status format:
3496  *
3497  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3498  * <cache block size> <#used cache blocks>/<#total cache blocks>
3499  * <#read hits> <#read misses> <#write hits> <#write misses>
3500  * <#demotions> <#promotions> <#dirty>
3501  * <#features> <features>*
3502  * <#core args> <core args>
3503  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3504  */
3505 static void cache_status(struct dm_target *ti, status_type_t type,
3506                          unsigned status_flags, char *result, unsigned maxlen)
3507 {
3508         int r = 0;
3509         unsigned i;
3510         ssize_t sz = 0;
3511         dm_block_t nr_free_blocks_metadata = 0;
3512         dm_block_t nr_blocks_metadata = 0;
3513         char buf[BDEVNAME_SIZE];
3514         struct cache *cache = ti->private;
3515         dm_cblock_t residency;
3516
3517         switch (type) {
3518         case STATUSTYPE_INFO:
3519                 if (get_cache_mode(cache) == CM_FAIL) {
3520                         DMEMIT("Fail");
3521                         break;
3522                 }
3523
3524                 /* Commit to ensure statistics aren't out-of-date */
3525                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3526                         (void) commit(cache, false);
3527
3528                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3529                 if (r) {
3530                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3531                               cache_device_name(cache), r);
3532                         goto err;
3533                 }
3534
3535                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3536                 if (r) {
3537                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3538                               cache_device_name(cache), r);
3539                         goto err;
3540                 }
3541
3542                 residency = policy_residency(cache->policy);
3543
3544                 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3545                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3546                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3547                        (unsigned long long)nr_blocks_metadata,
3548                        cache->sectors_per_block,
3549                        (unsigned long long) from_cblock(residency),
3550                        (unsigned long long) from_cblock(cache->cache_size),
3551                        (unsigned) atomic_read(&cache->stats.read_hit),
3552                        (unsigned) atomic_read(&cache->stats.read_miss),
3553                        (unsigned) atomic_read(&cache->stats.write_hit),
3554                        (unsigned) atomic_read(&cache->stats.write_miss),
3555                        (unsigned) atomic_read(&cache->stats.demotion),
3556                        (unsigned) atomic_read(&cache->stats.promotion),
3557                        (unsigned long) atomic_read(&cache->nr_dirty));
3558
3559                 if (writethrough_mode(&cache->features))
3560                         DMEMIT("1 writethrough ");
3561
3562                 else if (passthrough_mode(&cache->features))
3563                         DMEMIT("1 passthrough ");
3564
3565                 else if (writeback_mode(&cache->features))
3566                         DMEMIT("1 writeback ");
3567
3568                 else {
3569                         DMERR("%s: internal error: unknown io mode: %d",
3570                               cache_device_name(cache), (int) cache->features.io_mode);
3571                         goto err;
3572                 }
3573
3574                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3575
3576                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3577                 if (sz < maxlen) {
3578                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3579                         if (r)
3580                                 DMERR("%s: policy_emit_config_values returned %d",
3581                                       cache_device_name(cache), r);
3582                 }
3583
3584                 if (get_cache_mode(cache) == CM_READ_ONLY)
3585                         DMEMIT("ro ");
3586                 else
3587                         DMEMIT("rw ");
3588
3589                 if (dm_cache_metadata_needs_check(cache->cmd))
3590                         DMEMIT("needs_check ");
3591                 else
3592                         DMEMIT("- ");
3593
3594                 break;
3595
3596         case STATUSTYPE_TABLE:
3597                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3598                 DMEMIT("%s ", buf);
3599                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3600                 DMEMIT("%s ", buf);
3601                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3602                 DMEMIT("%s", buf);
3603
3604                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3605                         DMEMIT(" %s", cache->ctr_args[i]);
3606                 if (cache->nr_ctr_args)
3607                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3608         }
3609
3610         return;
3611
3612 err:
3613         DMEMIT("Error");
3614 }
3615
3616 /*
3617  * A cache block range can take two forms:
3618  *
3619  * i) A single cblock, eg. '3456'
3620  * ii) A begin and end cblock with dots between, eg. 123-234
3621  */
3622 static int parse_cblock_range(struct cache *cache, const char *str,
3623                               struct cblock_range *result)
3624 {
3625         char dummy;
3626         uint64_t b, e;
3627         int r;
3628
3629         /*
3630          * Try and parse form (ii) first.
3631          */
3632         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3633         if (r < 0)
3634                 return r;
3635
3636         if (r == 2) {
3637                 result->begin = to_cblock(b);
3638                 result->end = to_cblock(e);
3639                 return 0;
3640         }
3641
3642         /*
3643          * That didn't work, try form (i).
3644          */
3645         r = sscanf(str, "%llu%c", &b, &dummy);
3646         if (r < 0)
3647                 return r;
3648
3649         if (r == 1) {
3650                 result->begin = to_cblock(b);
3651                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3652                 return 0;
3653         }
3654
3655         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3656         return -EINVAL;
3657 }
3658
3659 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3660 {
3661         uint64_t b = from_cblock(range->begin);
3662         uint64_t e = from_cblock(range->end);
3663         uint64_t n = from_cblock(cache->cache_size);
3664
3665         if (b >= n) {
3666                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3667                       cache_device_name(cache), b, n);
3668                 return -EINVAL;
3669         }
3670
3671         if (e > n) {
3672                 DMERR("%s: end cblock out of range: %llu > %llu",
3673                       cache_device_name(cache), e, n);
3674                 return -EINVAL;
3675         }
3676
3677         if (b >= e) {
3678                 DMERR("%s: invalid cblock range: %llu >= %llu",
3679                       cache_device_name(cache), b, e);
3680                 return -EINVAL;
3681         }
3682
3683         return 0;
3684 }
3685
3686 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3687 {
3688         struct invalidation_request req;
3689
3690         INIT_LIST_HEAD(&req.list);
3691         req.cblocks = range;
3692         atomic_set(&req.complete, 0);
3693         req.err = 0;
3694         init_waitqueue_head(&req.result_wait);
3695
3696         spin_lock(&cache->invalidation_lock);
3697         list_add(&req.list, &cache->invalidation_requests);
3698         spin_unlock(&cache->invalidation_lock);
3699         wake_worker(cache);
3700
3701         wait_event(req.result_wait, atomic_read(&req.complete));
3702         return req.err;
3703 }
3704
3705 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3706                                               const char **cblock_ranges)
3707 {
3708         int r = 0;
3709         unsigned i;
3710         struct cblock_range range;
3711
3712         if (!passthrough_mode(&cache->features)) {
3713                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3714                       cache_device_name(cache));
3715                 return -EPERM;
3716         }
3717
3718         for (i = 0; i < count; i++) {
3719                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3720                 if (r)
3721                         break;
3722
3723                 r = validate_cblock_range(cache, &range);
3724                 if (r)
3725                         break;
3726
3727                 /*
3728                  * Pass begin and end origin blocks to the worker and wake it.
3729                  */
3730                 r = request_invalidation(cache, &range);
3731                 if (r)
3732                         break;
3733         }
3734
3735         return r;
3736 }
3737
3738 /*
3739  * Supports
3740  *      "<key> <value>"
3741  * and
3742  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3743  *
3744  * The key migration_threshold is supported by the cache target core.
3745  */
3746 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3747 {
3748         struct cache *cache = ti->private;
3749
3750         if (!argc)
3751                 return -EINVAL;
3752
3753         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3754                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3755                       cache_device_name(cache));
3756                 return -EOPNOTSUPP;
3757         }
3758
3759         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3760                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3761
3762         if (argc != 2)
3763                 return -EINVAL;
3764
3765         return set_config_value(cache, argv[0], argv[1]);
3766 }
3767
3768 static int cache_iterate_devices(struct dm_target *ti,
3769                                  iterate_devices_callout_fn fn, void *data)
3770 {
3771         int r = 0;
3772         struct cache *cache = ti->private;
3773
3774         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3775         if (!r)
3776                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3777
3778         return r;
3779 }
3780
3781 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3782 {
3783         /*
3784          * FIXME: these limits may be incompatible with the cache device
3785          */
3786         limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3787                                             cache->origin_sectors);
3788         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3789 }
3790
3791 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3792 {
3793         struct cache *cache = ti->private;
3794         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3795
3796         /*
3797          * If the system-determined stacked limits are compatible with the
3798          * cache's blocksize (io_opt is a factor) do not override them.
3799          */
3800         if (io_opt_sectors < cache->sectors_per_block ||
3801             do_div(io_opt_sectors, cache->sectors_per_block)) {
3802                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3803                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3804         }
3805         set_discard_limits(cache, limits);
3806 }
3807
3808 /*----------------------------------------------------------------*/
3809
3810 static struct target_type cache_target = {
3811         .name = "cache",
3812         .version = {1, 8, 0},
3813         .module = THIS_MODULE,
3814         .ctr = cache_ctr,
3815         .dtr = cache_dtr,
3816         .map = cache_map,
3817         .end_io = cache_end_io,
3818         .postsuspend = cache_postsuspend,
3819         .preresume = cache_preresume,
3820         .resume = cache_resume,
3821         .status = cache_status,
3822         .message = cache_message,
3823         .iterate_devices = cache_iterate_devices,
3824         .io_hints = cache_io_hints,
3825 };
3826
3827 static int __init dm_cache_init(void)
3828 {
3829         int r;
3830
3831         r = dm_register_target(&cache_target);
3832         if (r) {
3833                 DMERR("cache target registration failed: %d", r);
3834                 return r;
3835         }
3836
3837         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3838         if (!migration_cache) {
3839                 dm_unregister_target(&cache_target);
3840                 return -ENOMEM;
3841         }
3842
3843         return 0;
3844 }
3845
3846 static void __exit dm_cache_exit(void)
3847 {
3848         dm_unregister_target(&cache_target);
3849         kmem_cache_destroy(migration_cache);
3850 }
3851
3852 module_init(dm_cache_init);
3853 module_exit(dm_cache_exit);
3854
3855 MODULE_DESCRIPTION(DM_NAME " cache target");
3856 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3857 MODULE_LICENSE("GPL");