drivers/md/dm-cache-target.c

   1 /*
   2  * Copyright (C) 2012 Red Hat. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm.h"
   8 #include "dm-bio-prison.h"
   9 #include "dm-bio-record.h"
  10 #include "dm-cache-metadata.h"
  11
  12 #include <linux/dm-io.h>
  13 #include <linux/dm-kcopyd.h>
  14 #include <linux/jiffies.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/vmalloc.h>
  20
  21 #define DM_MSG_PREFIX "cache"
  22
  23 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  24         "A percentage of time allocated for copying to and/or from cache");
  25
  26 /*----------------------------------------------------------------*/
  27
  28 #define IOT_RESOLUTION 4
  29
  30 struct io_tracker {
  31         spinlock_t lock;
  32
  33         /*
  34          * Sectors of in-flight IO.
  35          */
  36         sector_t in_flight;
  37
  38         /*
  39          * The time, in jiffies, when this device became idle (if it is
  40          * indeed idle).
  41          */
  42         unsigned long idle_time;
  43         unsigned long last_update_time;
  44 };
  45
  46 static void iot_init(struct io_tracker *iot)
  47 {
  48         spin_lock_init(&iot->lock);
  49         iot->in_flight = 0ul;
  50         iot->idle_time = 0ul;
  51         iot->last_update_time = jiffies;
  52 }
  53
  54 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  55 {
  56         if (iot->in_flight)
  57                 return false;
  58
  59         return time_after(jiffies, iot->idle_time + jifs);
  60 }
  61
  62 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  63 {
  64         bool r;
  65         unsigned long flags;
  66
  67         spin_lock_irqsave(&iot->lock, flags);
  68         r = __iot_idle_for(iot, jifs);
  69         spin_unlock_irqrestore(&iot->lock, flags);
  70
  71         return r;
  72 }
  73
  74 static void iot_io_begin(struct io_tracker *iot, sector_t len)
  75 {
  76         unsigned long flags;
  77
  78         spin_lock_irqsave(&iot->lock, flags);
  79         iot->in_flight += len;
  80         spin_unlock_irqrestore(&iot->lock, flags);
  81 }
  82
  83 static void __iot_io_end(struct io_tracker *iot, sector_t len)
  84 {
  85         iot->in_flight -= len;
  86         if (!iot->in_flight)
  87                 iot->idle_time = jiffies;
  88 }
  89
  90 static void iot_io_end(struct io_tracker *iot, sector_t len)
  91 {
  92         unsigned long flags;
  93
  94         spin_lock_irqsave(&iot->lock, flags);
  95         __iot_io_end(iot, len);
  96         spin_unlock_irqrestore(&iot->lock, flags);
  97 }
  98
  99 /*----------------------------------------------------------------*/
 100
 101 /*
 102  * Glossary:
 103  *
 104  * oblock: index of an origin block
 105  * cblock: index of a cache block
 106  * promotion: movement of a block from origin to cache
 107  * demotion: movement of a block from cache to origin
 108  * migration: movement of a block between the origin and cache device,
 109  *            either direction
 110  */
 111
 112 /*----------------------------------------------------------------*/
 113
 114 /*
 115  * There are a couple of places where we let a bio run, but want to do some
 116  * work before calling its endio function.  We do this by temporarily
 117  * changing the endio fn.
 118  */
 119 struct dm_hook_info {
 120         bio_end_io_t *bi_end_io;
 121         void *bi_private;
 122 };
 123
 124 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 125                         bio_end_io_t *bi_end_io, void *bi_private)
 126 {
 127         h->bi_end_io = bio->bi_end_io;
 128         h->bi_private = bio->bi_private;
 129
 130         bio->bi_end_io = bi_end_io;
 131         bio->bi_private = bi_private;
 132 }
 133
 134 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 135 {
 136         bio->bi_end_io = h->bi_end_io;
 137         bio->bi_private = h->bi_private;
 138 }
 139
 140 /*----------------------------------------------------------------*/
 141
 142 #define MIGRATION_POOL_SIZE 128
 143 #define COMMIT_PERIOD HZ
 144 #define MIGRATION_COUNT_WINDOW 10
 145
 146 /*
 147  * The block size of the device holding cache data must be
 148  * between 32KB and 1GB.
 149  */
 150 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 151 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 152
 153 enum cache_metadata_mode {
 154         CM_WRITE,               /* metadata may be changed */
 155         CM_READ_ONLY,           /* metadata may not be changed */
 156         CM_FAIL
 157 };
 158
 159 enum cache_io_mode {
 160         /*
 161          * Data is written to cached blocks only.  These blocks are marked
 162          * dirty.  If you lose the cache device you will lose data.
 163          * Potential performance increase for both reads and writes.
 164          */
 165         CM_IO_WRITEBACK,
 166
 167         /*
 168          * Data is written to both cache and origin.  Blocks are never
 169          * dirty.  Potential performance benfit for reads only.
 170          */
 171         CM_IO_WRITETHROUGH,
 172
 173         /*
 174          * A degraded mode useful for various cache coherency situations
 175          * (eg, rolling back snapshots).  Reads and writes always go to the
 176          * origin.  If a write goes to a cached oblock, then the cache
 177          * block is invalidated.
 178          */
 179         CM_IO_PASSTHROUGH
 180 };
 181
 182 struct cache_features {
 183         enum cache_metadata_mode mode;
 184         enum cache_io_mode io_mode;
 185 };
 186
 187 struct cache_stats {
 188         atomic_t read_hit;
 189         atomic_t read_miss;
 190         atomic_t write_hit;
 191         atomic_t write_miss;
 192         atomic_t demotion;
 193         atomic_t promotion;
 194         atomic_t copies_avoided;
 195         atomic_t cache_cell_clash;
 196         atomic_t commit_count;
 197         atomic_t discard_count;
 198 };
 199
 200 /*
 201  * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
 202  * the one-past-the-end value.
 203  */
 204 struct cblock_range {
 205         dm_cblock_t begin;
 206         dm_cblock_t end;
 207 };
 208
 209 struct invalidation_request {
 210         struct list_head list;
 211         struct cblock_range *cblocks;
 212
 213         atomic_t complete;
 214         int err;
 215
 216         wait_queue_head_t result_wait;
 217 };
 218
 219 struct cache {
 220         struct dm_target *ti;
 221         struct dm_target_callbacks callbacks;
 222
 223         struct dm_cache_metadata *cmd;
 224
 225         /*
 226          * Metadata is written to this device.
 227          */
 228         struct dm_dev *metadata_dev;
 229
 230         /*
 231          * The slower of the two data devices.  Typically a spindle.
 232          */
 233         struct dm_dev *origin_dev;
 234
 235         /*
 236          * The faster of the two data devices.  Typically an SSD.
 237          */
 238         struct dm_dev *cache_dev;
 239
 240         /*
 241          * Size of the origin device in _complete_ blocks and native sectors.
 242          */
 243         dm_oblock_t origin_blocks;
 244         sector_t origin_sectors;
 245
 246         /*
 247          * Size of the cache device in blocks.
 248          */
 249         dm_cblock_t cache_size;
 250
 251         /*
 252          * Fields for converting from sectors to blocks.
 253          */
 254         uint32_t sectors_per_block;
 255         int sectors_per_block_shift;
 256
 257         spinlock_t lock;
 258         struct list_head deferred_cells;
 259         struct bio_list deferred_bios;
 260         struct bio_list deferred_flush_bios;
 261         struct bio_list deferred_writethrough_bios;
 262         struct list_head quiesced_migrations;
 263         struct list_head completed_migrations;
 264         struct list_head need_commit_migrations;
 265         sector_t migration_threshold;
 266         wait_queue_head_t migration_wait;
 267         atomic_t nr_allocated_migrations;
 268
 269         /*
 270          * The number of in flight migrations that are performing
 271          * background io. eg, promotion, writeback.
 272          */
 273         atomic_t nr_io_migrations;
 274
 275         wait_queue_head_t quiescing_wait;
 276         atomic_t quiescing;
 277         atomic_t quiescing_ack;
 278
 279         /*
 280          * cache_size entries, dirty if set
 281          */
 282         atomic_t nr_dirty;
 283         unsigned long *dirty_bitset;
 284
 285         /*
 286          * origin_blocks entries, discarded if set.
 287          */
 288         dm_dblock_t discard_nr_blocks;
 289         unsigned long *discard_bitset;
 290         uint32_t discard_block_size; /* a power of 2 times sectors per block */
 291
 292         /*
 293          * Rather than reconstructing the table line for the status we just
 294          * save it and regurgitate.
 295          */
 296         unsigned nr_ctr_args;
 297         const char **ctr_args;
 298
 299         struct dm_kcopyd_client *copier;
 300         struct workqueue_struct *wq;
 301         struct work_struct worker;
 302
 303         struct delayed_work waker;
 304         unsigned long last_commit_jiffies;
 305
 306         struct dm_bio_prison *prison;
 307         struct dm_deferred_set *all_io_ds;
 308
 309         mempool_t *migration_pool;
 310
 311         struct dm_cache_policy *policy;
 312         unsigned policy_nr_args;
 313
 314         bool need_tick_bio:1;
 315         bool sized:1;
 316         bool invalidate:1;
 317         bool commit_requested:1;
 318         bool loaded_mappings:1;
 319         bool loaded_discards:1;
 320
 321         /*
 322          * Cache features such as write-through.
 323          */
 324         struct cache_features features;
 325
 326         struct cache_stats stats;
 327
 328         /*
 329          * Invalidation fields.
 330          */
 331         spinlock_t invalidation_lock;
 332         struct list_head invalidation_requests;
 333
 334         struct io_tracker origin_tracker;
 335 };
 336
 337 struct per_bio_data {
 338         bool tick:1;
 339         unsigned req_nr:2;
 340         struct dm_deferred_entry *all_io_entry;
 341         struct dm_hook_info hook_info;
 342         sector_t len;
 343
 344         /*
 345          * writethrough fields.  These MUST remain at the end of this
 346          * structure and the 'cache' member must be the first as it
 347          * is used to determine the offset of the writethrough fields.
 348          */
 349         struct cache *cache;
 350         dm_cblock_t cblock;
 351         struct dm_bio_details bio_details;
 352 };
 353
 354 struct dm_cache_migration {
 355         struct list_head list;
 356         struct cache *cache;
 357
 358         unsigned long start_jiffies;
 359         dm_oblock_t old_oblock;
 360         dm_oblock_t new_oblock;
 361         dm_cblock_t cblock;
 362
 363         bool err:1;
 364         bool discard:1;
 365         bool writeback:1;
 366         bool demote:1;
 367         bool promote:1;
 368         bool requeue_holder:1;
 369         bool invalidate:1;
 370
 371         struct dm_bio_prison_cell *old_ocell;
 372         struct dm_bio_prison_cell *new_ocell;
 373 };
 374
 375 /*
 376  * Processing a bio in the worker thread may require these memory
 377  * allocations.  We prealloc to avoid deadlocks (the same worker thread
 378  * frees them back to the mempool).
 379  */
 380 struct prealloc {
 381         struct dm_cache_migration *mg;
 382         struct dm_bio_prison_cell *cell1;
 383         struct dm_bio_prison_cell *cell2;
 384 };
 385
 386 static enum cache_metadata_mode get_cache_mode(struct cache *cache);
 387
 388 static void wake_worker(struct cache *cache)
 389 {
 390         queue_work(cache->wq, &cache->worker);
 391 }
 392
 393 /*----------------------------------------------------------------*/
 394
 395 static struct dm_bio_prison_cell *alloc_prison_cell(struct cache *cache)
 396 {
 397         /* FIXME: change to use a local slab. */
 398         return dm_bio_prison_alloc_cell(cache->prison, GFP_NOWAIT);
 399 }
 400
 401 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell *cell)
 402 {
 403         dm_bio_prison_free_cell(cache->prison, cell);
 404 }
 405
 406 static struct dm_cache_migration *alloc_migration(struct cache *cache)
 407 {
 408         struct dm_cache_migration *mg;
 409
 410         mg = mempool_alloc(cache->migration_pool, GFP_NOWAIT);
 411         if (mg) {
 412                 mg->cache = cache;
 413                 atomic_inc(&mg->cache->nr_allocated_migrations);
 414         }
 415
 416         return mg;
 417 }
 418
 419 static void free_migration(struct dm_cache_migration *mg)
 420 {
 421         struct cache *cache = mg->cache;
 422
 423         if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 424                 wake_up(&cache->migration_wait);
 425
 426         mempool_free(mg, cache->migration_pool);
 427 }
 428
 429 static int prealloc_data_structs(struct cache *cache, struct prealloc *p)
 430 {
 431         if (!p->mg) {
 432                 p->mg = alloc_migration(cache);
 433                 if (!p->mg)
 434                         return -ENOMEM;
 435         }
 436
 437         if (!p->cell1) {
 438                 p->cell1 = alloc_prison_cell(cache);
 439                 if (!p->cell1)
 440                         return -ENOMEM;
 441         }
 442
 443         if (!p->cell2) {
 444                 p->cell2 = alloc_prison_cell(cache);
 445                 if (!p->cell2)
 446                         return -ENOMEM;
 447         }
 448
 449         return 0;
 450 }
 451
 452 static void prealloc_free_structs(struct cache *cache, struct prealloc *p)
 453 {
 454         if (p->cell2)
 455                 free_prison_cell(cache, p->cell2);
 456
 457         if (p->cell1)
 458                 free_prison_cell(cache, p->cell1);
 459
 460         if (p->mg)
 461                 free_migration(p->mg);
 462 }
 463
 464 static struct dm_cache_migration *prealloc_get_migration(struct prealloc *p)
 465 {
 466         struct dm_cache_migration *mg = p->mg;
 467
 468         BUG_ON(!mg);
 469         p->mg = NULL;
 470
 471         return mg;
 472 }
 473
 474 /*
 475  * You must have a cell within the prealloc struct to return.  If not this
 476  * function will BUG() rather than returning NULL.
 477  */
 478 static struct dm_bio_prison_cell *prealloc_get_cell(struct prealloc *p)
 479 {
 480         struct dm_bio_prison_cell *r = NULL;
 481
 482         if (p->cell1) {
 483                 r = p->cell1;
 484                 p->cell1 = NULL;
 485
 486         } else if (p->cell2) {
 487                 r = p->cell2;
 488                 p->cell2 = NULL;
 489         } else
 490                 BUG();
 491
 492         return r;
 493 }
 494
 495 /*
 496  * You can't have more than two cells in a prealloc struct.  BUG() will be
 497  * called if you try and overfill.
 498  */
 499 static void prealloc_put_cell(struct prealloc *p, struct dm_bio_prison_cell *cell)
 500 {
 501         if (!p->cell2)
 502                 p->cell2 = cell;
 503
 504         else if (!p->cell1)
 505                 p->cell1 = cell;
 506
 507         else
 508                 BUG();
 509 }
 510
 511 /*----------------------------------------------------------------*/
 512
 513 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key *key)
 514 {
 515         key->virtual = 0;
 516         key->dev = 0;
 517         key->block_begin = from_oblock(begin);
 518         key->block_end = from_oblock(end);
 519 }
 520
 521 /*
 522  * The caller hands in a preallocated cell, and a free function for it.
 523  * The cell will be freed if there's an error, or if it wasn't used because
 524  * a cell with that key already exists.
 525  */
 526 typedef void (*cell_free_fn)(void *context, struct dm_bio_prison_cell *cell);
 527
 528 static int bio_detain_range(struct cache *cache, dm_oblock_t oblock_begin, dm_oblock_t oblock_end,
 529                             struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 530                             cell_free_fn free_fn, void *free_context,
 531                             struct dm_bio_prison_cell **cell_result)
 532 {
 533         int r;
 534         struct dm_cell_key key;
 535
 536         build_key(oblock_begin, oblock_end, &key);
 537         r = dm_bio_detain(cache->prison, &key, bio, cell_prealloc, cell_result);
 538         if (r)
 539                 free_fn(free_context, cell_prealloc);
 540
 541         return r;
 542 }
 543
 544 static int bio_detain(struct cache *cache, dm_oblock_t oblock,
 545                       struct bio *bio, struct dm_bio_prison_cell *cell_prealloc,
 546                       cell_free_fn free_fn, void *free_context,
 547                       struct dm_bio_prison_cell **cell_result)
 548 {
 549         dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 550         return bio_detain_range(cache, oblock, end, bio,
 551                                 cell_prealloc, free_fn, free_context, cell_result);
 552 }
 553
 554 static int get_cell(struct cache *cache,
 555                     dm_oblock_t oblock,
 556                     struct prealloc *structs,
 557                     struct dm_bio_prison_cell **cell_result)
 558 {
 559         int r;
 560         struct dm_cell_key key;
 561         struct dm_bio_prison_cell *cell_prealloc;
 562
 563         cell_prealloc = prealloc_get_cell(structs);
 564
 565         build_key(oblock, to_oblock(from_oblock(oblock) + 1ULL), &key);
 566         r = dm_get_cell(cache->prison, &key, cell_prealloc, cell_result);
 567         if (r)
 568                 prealloc_put_cell(structs, cell_prealloc);
 569
 570         return r;
 571 }
 572
 573 /*----------------------------------------------------------------*/
 574
 575 static bool is_dirty(struct cache *cache, dm_cblock_t b)
 576 {
 577         return test_bit(from_cblock(b), cache->dirty_bitset);
 578 }
 579
 580 static void set_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 581 {
 582         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 583                 atomic_inc(&cache->nr_dirty);
 584                 policy_set_dirty(cache->policy, oblock);
 585         }
 586 }
 587
 588 static void clear_dirty(struct cache *cache, dm_oblock_t oblock, dm_cblock_t cblock)
 589 {
 590         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 591                 policy_clear_dirty(cache->policy, oblock);
 592                 if (atomic_dec_return(&cache->nr_dirty) == 0)
 593                         dm_table_event(cache->ti->table);
 594         }
 595 }
 596
 597 /*----------------------------------------------------------------*/
 598
 599 static bool block_size_is_power_of_two(struct cache *cache)
 600 {
 601         return cache->sectors_per_block_shift >= 0;
 602 }
 603
 604 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
 605 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
 606 __always_inline
 607 #endif
 608 static dm_block_t block_div(dm_block_t b, uint32_t n)
 609 {
 610         do_div(b, n);
 611
 612         return b;
 613 }
 614
 615 static dm_block_t oblocks_per_dblock(struct cache *cache)
 616 {
 617         dm_block_t oblocks = cache->discard_block_size;
 618
 619         if (block_size_is_power_of_two(cache))
 620                 oblocks >>= cache->sectors_per_block_shift;
 621         else
 622                 oblocks = block_div(oblocks, cache->sectors_per_block);
 623
 624         return oblocks;
 625 }
 626
 627 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 628 {
 629         return to_dblock(block_div(from_oblock(oblock),
 630                                    oblocks_per_dblock(cache)));
 631 }
 632
 633 static dm_oblock_t dblock_to_oblock(struct cache *cache, dm_dblock_t dblock)
 634 {
 635         return to_oblock(from_dblock(dblock) * oblocks_per_dblock(cache));
 636 }
 637
 638 static void set_discard(struct cache *cache, dm_dblock_t b)
 639 {
 640         unsigned long flags;
 641
 642         BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 643         atomic_inc(&cache->stats.discard_count);
 644
 645         spin_lock_irqsave(&cache->lock, flags);
 646         set_bit(from_dblock(b), cache->discard_bitset);
 647         spin_unlock_irqrestore(&cache->lock, flags);
 648 }
 649
 650 static void clear_discard(struct cache *cache, dm_dblock_t b)
 651 {
 652         unsigned long flags;
 653
 654         spin_lock_irqsave(&cache->lock, flags);
 655         clear_bit(from_dblock(b), cache->discard_bitset);
 656         spin_unlock_irqrestore(&cache->lock, flags);
 657 }
 658
 659 static bool is_discarded(struct cache *cache, dm_dblock_t b)
 660 {
 661         int r;
 662         unsigned long flags;
 663
 664         spin_lock_irqsave(&cache->lock, flags);
 665         r = test_bit(from_dblock(b), cache->discard_bitset);
 666         spin_unlock_irqrestore(&cache->lock, flags);
 667
 668         return r;
 669 }
 670
 671 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 672 {
 673         int r;
 674         unsigned long flags;
 675
 676         spin_lock_irqsave(&cache->lock, flags);
 677         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 678                      cache->discard_bitset);
 679         spin_unlock_irqrestore(&cache->lock, flags);
 680
 681         return r;
 682 }
 683
 684 /*----------------------------------------------------------------*/
 685
 686 static void load_stats(struct cache *cache)
 687 {
 688         struct dm_cache_statistics stats;
 689
 690         dm_cache_metadata_get_stats(cache->cmd, &stats);
 691         atomic_set(&cache->stats.read_hit, stats.read_hits);
 692         atomic_set(&cache->stats.read_miss, stats.read_misses);
 693         atomic_set(&cache->stats.write_hit, stats.write_hits);
 694         atomic_set(&cache->stats.write_miss, stats.write_misses);
 695 }
 696
 697 static void save_stats(struct cache *cache)
 698 {
 699         struct dm_cache_statistics stats;
 700
 701         if (get_cache_mode(cache) >= CM_READ_ONLY)
 702                 return;
 703
 704         stats.read_hits = atomic_read(&cache->stats.read_hit);
 705         stats.read_misses = atomic_read(&cache->stats.read_miss);
 706         stats.write_hits = atomic_read(&cache->stats.write_hit);
 707         stats.write_misses = atomic_read(&cache->stats.write_miss);
 708
 709         dm_cache_metadata_set_stats(cache->cmd, &stats);
 710 }
 711
 712 /*----------------------------------------------------------------
 713  * Per bio data
 714  *--------------------------------------------------------------*/
 715
 716 /*
 717  * If using writeback, leave out struct per_bio_data's writethrough fields.
 718  */
 719 #define PB_DATA_SIZE_WB (offsetof(struct per_bio_data, cache))
 720 #define PB_DATA_SIZE_WT (sizeof(struct per_bio_data))
 721
 722 static bool writethrough_mode(struct cache_features *f)
 723 {
 724         return f->io_mode == CM_IO_WRITETHROUGH;
 725 }
 726
 727 static bool writeback_mode(struct cache_features *f)
 728 {
 729         return f->io_mode == CM_IO_WRITEBACK;
 730 }
 731
 732 static bool passthrough_mode(struct cache_features *f)
 733 {
 734         return f->io_mode == CM_IO_PASSTHROUGH;
 735 }
 736
 737 static size_t get_per_bio_data_size(struct cache *cache)
 738 {
 739         return writethrough_mode(&cache->features) ? PB_DATA_SIZE_WT : PB_DATA_SIZE_WB;
 740 }
 741
 742 static struct per_bio_data *get_per_bio_data(struct bio *bio, size_t data_size)
 743 {
 744         struct per_bio_data *pb = dm_per_bio_data(bio, data_size);
 745         BUG_ON(!pb);
 746         return pb;
 747 }
 748
 749 static struct per_bio_data *init_per_bio_data(struct bio *bio, size_t data_size)
 750 {
 751         struct per_bio_data *pb = get_per_bio_data(bio, data_size);
 752
 753         pb->tick = false;
 754         pb->req_nr = dm_bio_get_target_bio_nr(bio);
 755         pb->all_io_entry = NULL;
 756         pb->len = 0;
 757
 758         return pb;
 759 }
 760
 761 /*----------------------------------------------------------------
 762  * Remapping
 763  *--------------------------------------------------------------*/
 764 static void remap_to_origin(struct cache *cache, struct bio *bio)
 765 {
 766         bio->bi_bdev = cache->origin_dev->bdev;
 767 }
 768
 769 static void remap_to_cache(struct cache *cache, struct bio *bio,
 770                            dm_cblock_t cblock)
 771 {
 772         sector_t bi_sector = bio->bi_iter.bi_sector;
 773         sector_t block = from_cblock(cblock);
 774
 775         bio->bi_bdev = cache->cache_dev->bdev;
 776         if (!block_size_is_power_of_two(cache))
 777                 bio->bi_iter.bi_sector =
 778                         (block * cache->sectors_per_block) +
 779                         sector_div(bi_sector, cache->sectors_per_block);
 780         else
 781                 bio->bi_iter.bi_sector =
 782                         (block << cache->sectors_per_block_shift) |
 783                         (bi_sector & (cache->sectors_per_block - 1));
 784 }
 785
 786 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 787 {
 788         unsigned long flags;
 789         size_t pb_data_size = get_per_bio_data_size(cache);
 790         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 791
 792         spin_lock_irqsave(&cache->lock, flags);
 793         if (cache->need_tick_bio &&
 794             !(bio->bi_rw & (REQ_FUA | REQ_FLUSH | REQ_DISCARD))) {
 795                 pb->tick = true;
 796                 cache->need_tick_bio = false;
 797         }
 798         spin_unlock_irqrestore(&cache->lock, flags);
 799 }
 800
 801 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 802                                   dm_oblock_t oblock)
 803 {
 804         check_if_tick_bio_needed(cache, bio);
 805         remap_to_origin(cache, bio);
 806         if (bio_data_dir(bio) == WRITE)
 807                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 808 }
 809
 810 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 811                                  dm_oblock_t oblock, dm_cblock_t cblock)
 812 {
 813         check_if_tick_bio_needed(cache, bio);
 814         remap_to_cache(cache, bio, cblock);
 815         if (bio_data_dir(bio) == WRITE) {
 816                 set_dirty(cache, oblock, cblock);
 817                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 818         }
 819 }
 820
 821 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 822 {
 823         sector_t block_nr = bio->bi_iter.bi_sector;
 824
 825         if (!block_size_is_power_of_two(cache))
 826                 (void) sector_div(block_nr, cache->sectors_per_block);
 827         else
 828                 block_nr >>= cache->sectors_per_block_shift;
 829
 830         return to_oblock(block_nr);
 831 }
 832
 833 static int bio_triggers_commit(struct cache *cache, struct bio *bio)
 834 {
 835         return bio->bi_rw & (REQ_FLUSH | REQ_FUA);
 836 }
 837
 838 /*
 839  * You must increment the deferred set whilst the prison cell is held.  To
 840  * encourage this, we ask for 'cell' to be passed in.
 841  */
 842 static void inc_ds(struct cache *cache, struct bio *bio,
 843                    struct dm_bio_prison_cell *cell)
 844 {
 845         size_t pb_data_size = get_per_bio_data_size(cache);
 846         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 847
 848         BUG_ON(!cell);
 849         BUG_ON(pb->all_io_entry);
 850
 851         pb->all_io_entry = dm_deferred_entry_inc(cache->all_io_ds);
 852 }
 853
 854 static bool accountable_bio(struct cache *cache, struct bio *bio)
 855 {
 856         return ((bio->bi_bdev == cache->origin_dev->bdev) &&
 857                 !(bio->bi_rw & REQ_DISCARD));
 858 }
 859
 860 static void accounted_begin(struct cache *cache, struct bio *bio)
 861 {
 862         size_t pb_data_size = get_per_bio_data_size(cache);
 863         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 864
 865         if (accountable_bio(cache, bio)) {
 866                 pb->len = bio_sectors(bio);
 867                 iot_io_begin(&cache->origin_tracker, pb->len);
 868         }
 869 }
 870
 871 static void accounted_complete(struct cache *cache, struct bio *bio)
 872 {
 873         size_t pb_data_size = get_per_bio_data_size(cache);
 874         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
 875
 876         iot_io_end(&cache->origin_tracker, pb->len);
 877 }
 878
 879 static void accounted_request(struct cache *cache, struct bio *bio)
 880 {
 881         accounted_begin(cache, bio);
 882         generic_make_request(bio);
 883 }
 884
 885 static void issue(struct cache *cache, struct bio *bio)
 886 {
 887         unsigned long flags;
 888
 889         if (!bio_triggers_commit(cache, bio)) {
 890                 accounted_request(cache, bio);
 891                 return;
 892         }
 893
 894         /*
 895          * Batch together any bios that trigger commits and then issue a
 896          * single commit for them in do_worker().
 897          */
 898         spin_lock_irqsave(&cache->lock, flags);
 899         cache->commit_requested = true;
 900         bio_list_add(&cache->deferred_flush_bios, bio);
 901         spin_unlock_irqrestore(&cache->lock, flags);
 902 }
 903
 904 static void inc_and_issue(struct cache *cache, struct bio *bio, struct dm_bio_prison_cell *cell)
 905 {
 906         inc_ds(cache, bio, cell);
 907         issue(cache, bio);
 908 }
 909
 910 static void defer_writethrough_bio(struct cache *cache, struct bio *bio)
 911 {
 912         unsigned long flags;
 913
 914         spin_lock_irqsave(&cache->lock, flags);
 915         bio_list_add(&cache->deferred_writethrough_bios, bio);
 916         spin_unlock_irqrestore(&cache->lock, flags);
 917
 918         wake_worker(cache);
 919 }
 920
 921 static void writethrough_endio(struct bio *bio, int err)
 922 {
 923         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 924
 925         dm_unhook_bio(&pb->hook_info, bio);
 926
 927         if (err) {
 928                 bio_endio(bio, err);
 929                 return;
 930         }
 931
 932         dm_bio_restore(&pb->bio_details, bio);
 933         remap_to_cache(pb->cache, bio, pb->cblock);
 934
 935         /*
 936          * We can't issue this bio directly, since we're in interrupt
 937          * context.  So it gets put on a bio list for processing by the
 938          * worker thread.
 939          */
 940         defer_writethrough_bio(pb->cache, bio);
 941 }
 942
 943 /*
 944  * When running in writethrough mode we need to send writes to clean blocks
 945  * to both the cache and origin devices.  In future we'd like to clone the
 946  * bio and send them in parallel, but for now we're doing them in
 947  * series as this is easier.
 948  */
 949 static void remap_to_origin_then_cache(struct cache *cache, struct bio *bio,
 950                                        dm_oblock_t oblock, dm_cblock_t cblock)
 951 {
 952         struct per_bio_data *pb = get_per_bio_data(bio, PB_DATA_SIZE_WT);
 953
 954         pb->cache = cache;
 955         pb->cblock = cblock;
 956         dm_hook_bio(&pb->hook_info, bio, writethrough_endio, NULL);
 957         dm_bio_record(&pb->bio_details, bio);
 958
 959         remap_to_origin_clear_discard(pb->cache, bio, oblock);
 960 }
 961
 962 /*----------------------------------------------------------------
 963  * Failure modes
 964  *--------------------------------------------------------------*/
 965 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 966 {
 967         return cache->features.mode;
 968 }
 969
 970 static const char *cache_device_name(struct cache *cache)
 971 {
 972         return dm_device_name(dm_table_get_md(cache->ti->table));
 973 }
 974
 975 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 976 {
 977         const char *descs[] = {
 978                 "write",
 979                 "read-only",
 980                 "fail"
 981         };
 982
 983         dm_table_event(cache->ti->table);
 984         DMINFO("%s: switching cache to %s mode",
 985                cache_device_name(cache), descs[(int)mode]);
 986 }
 987
 988 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 989 {
 990         bool needs_check = dm_cache_metadata_needs_check(cache->cmd);
 991         enum cache_metadata_mode old_mode = get_cache_mode(cache);
 992
 993         if (new_mode == CM_WRITE && needs_check) {
 994                 DMERR("%s: unable to switch cache to write mode until repaired.",
 995                       cache_device_name(cache));
 996                 if (old_mode != new_mode)
 997                         new_mode = old_mode;
 998                 else
 999                         new_mode = CM_READ_ONLY;
1000         }
1001
1002         /* Never move out of fail mode */
1003         if (old_mode == CM_FAIL)
1004                 new_mode = CM_FAIL;
1005
1006         switch (new_mode) {
1007         case CM_FAIL:
1008         case CM_READ_ONLY:
1009                 dm_cache_metadata_set_read_only(cache->cmd);
1010                 break;
1011
1012         case CM_WRITE:
1013                 dm_cache_metadata_set_read_write(cache->cmd);
1014                 break;
1015         }
1016
1017         cache->features.mode = new_mode;
1018
1019         if (new_mode != old_mode)
1020                 notify_mode_switch(cache, new_mode);
1021 }
1022
1023 static void abort_transaction(struct cache *cache)
1024 {
1025         const char *dev_name = cache_device_name(cache);
1026
1027         if (get_cache_mode(cache) >= CM_READ_ONLY)
1028                 return;
1029
1030         if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1031                 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1032                 set_cache_mode(cache, CM_FAIL);
1033         }
1034
1035         DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1036         if (dm_cache_metadata_abort(cache->cmd)) {
1037                 DMERR("%s: failed to abort metadata transaction", dev_name);
1038                 set_cache_mode(cache, CM_FAIL);
1039         }
1040 }
1041
1042 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1043 {
1044         DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1045                     cache_device_name(cache), op, r);
1046         abort_transaction(cache);
1047         set_cache_mode(cache, CM_READ_ONLY);
1048 }
1049
1050 /*----------------------------------------------------------------
1051  * Migration processing
1052  *
1053  * Migration covers moving data from the origin device to the cache, or
1054  * vice versa.
1055  *--------------------------------------------------------------*/
1056 static void inc_io_migrations(struct cache *cache)
1057 {
1058         atomic_inc(&cache->nr_io_migrations);
1059 }
1060
1061 static void dec_io_migrations(struct cache *cache)
1062 {
1063         atomic_dec(&cache->nr_io_migrations);
1064 }
1065
1066 static void __cell_release(struct cache *cache, struct dm_bio_prison_cell *cell,
1067                            bool holder, struct bio_list *bios)
1068 {
1069         (holder ? dm_cell_release : dm_cell_release_no_holder)
1070                 (cache->prison, cell, bios);
1071         free_prison_cell(cache, cell);
1072 }
1073
1074 static bool discard_or_flush(struct bio *bio)
1075 {
1076         return bio->bi_rw & (REQ_FLUSH | REQ_FUA | REQ_DISCARD);
1077 }
1078
1079 static void __cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell)
1080 {
1081         if (discard_or_flush(cell->holder))
1082                 /*
1083                  * We have to handle these bios
1084                  * individually.
1085                  */
1086                 __cell_release(cache, cell, true, &cache->deferred_bios);
1087
1088         else
1089                 list_add_tail(&cell->user_list, &cache->deferred_cells);
1090 }
1091
1092 static void cell_defer(struct cache *cache, struct dm_bio_prison_cell *cell, bool holder)
1093 {
1094         unsigned long flags;
1095
1096         if (!holder && dm_cell_promote_or_release(cache->prison, cell)) {
1097                 /*
1098                  * There was no prisoner to promote to holder, the
1099                  * cell has been released.
1100                  */
1101                 free_prison_cell(cache, cell);
1102                 return;
1103         }
1104
1105         spin_lock_irqsave(&cache->lock, flags);
1106         __cell_defer(cache, cell);
1107         spin_unlock_irqrestore(&cache->lock, flags);
1108
1109         wake_worker(cache);
1110 }
1111
1112 static void cell_error_with_code(struct cache *cache, struct dm_bio_prison_cell *cell, int err)
1113 {
1114         dm_cell_error(cache->prison, cell, err);
1115         dm_bio_prison_free_cell(cache->prison, cell);
1116 }
1117
1118 static void cell_requeue(struct cache *cache, struct dm_bio_prison_cell *cell)
1119 {
1120         cell_error_with_code(cache, cell, DM_ENDIO_REQUEUE);
1121 }
1122
1123 static void free_io_migration(struct dm_cache_migration *mg)
1124 {
1125         dec_io_migrations(mg->cache);
1126         free_migration(mg);
1127         wake_worker(mg->cache);
1128 }
1129
1130 static void migration_failure(struct dm_cache_migration *mg)
1131 {
1132         struct cache *cache = mg->cache;
1133         const char *dev_name = cache_device_name(cache);
1134
1135         if (mg->writeback) {
1136                 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1137                 set_dirty(cache, mg->old_oblock, mg->cblock);
1138                 cell_defer(cache, mg->old_ocell, false);
1139
1140         } else if (mg->demote) {
1141                 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1142                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1143
1144                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1145                 if (mg->promote)
1146                         cell_defer(cache, mg->new_ocell, true);
1147         } else {
1148                 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1149                 policy_remove_mapping(cache->policy, mg->new_oblock);
1150                 cell_defer(cache, mg->new_ocell, true);
1151         }
1152
1153         free_io_migration(mg);
1154 }
1155
1156 static void migration_success_pre_commit(struct dm_cache_migration *mg)
1157 {
1158         int r;
1159         unsigned long flags;
1160         struct cache *cache = mg->cache;
1161
1162         if (mg->writeback) {
1163                 clear_dirty(cache, mg->old_oblock, mg->cblock);
1164                 cell_defer(cache, mg->old_ocell, false);
1165                 free_io_migration(mg);
1166                 return;
1167
1168         } else if (mg->demote) {
1169                 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1170                 if (r) {
1171                         DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1172                                     cache_device_name(cache));
1173                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1174                         policy_force_mapping(cache->policy, mg->new_oblock,
1175                                              mg->old_oblock);
1176                         if (mg->promote)
1177                                 cell_defer(cache, mg->new_ocell, true);
1178                         free_io_migration(mg);
1179                         return;
1180                 }
1181         } else {
1182                 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1183                 if (r) {
1184                         DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1185                                     cache_device_name(cache));
1186                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1187                         policy_remove_mapping(cache->policy, mg->new_oblock);
1188                         free_io_migration(mg);
1189                         return;
1190                 }
1191         }
1192
1193         spin_lock_irqsave(&cache->lock, flags);
1194         list_add_tail(&mg->list, &cache->need_commit_migrations);
1195         cache->commit_requested = true;
1196         spin_unlock_irqrestore(&cache->lock, flags);
1197 }
1198
1199 static void migration_success_post_commit(struct dm_cache_migration *mg)
1200 {
1201         unsigned long flags;
1202         struct cache *cache = mg->cache;
1203
1204         if (mg->writeback) {
1205                 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1206                              cache_device_name(cache));
1207                 return;
1208
1209         } else if (mg->demote) {
1210                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1211
1212                 if (mg->promote) {
1213                         mg->demote = false;
1214
1215                         spin_lock_irqsave(&cache->lock, flags);
1216                         list_add_tail(&mg->list, &cache->quiesced_migrations);
1217                         spin_unlock_irqrestore(&cache->lock, flags);
1218
1219                 } else {
1220                         if (mg->invalidate)
1221                                 policy_remove_mapping(cache->policy, mg->old_oblock);
1222                         free_io_migration(mg);
1223                 }
1224
1225         } else {
1226                 if (mg->requeue_holder) {
1227                         clear_dirty(cache, mg->new_oblock, mg->cblock);
1228                         cell_defer(cache, mg->new_ocell, true);
1229                 } else {
1230                         /*
1231                          * The block was promoted via an overwrite, so it's dirty.
1232                          */
1233                         set_dirty(cache, mg->new_oblock, mg->cblock);
1234                         bio_endio(mg->new_ocell->holder, 0);
1235                         cell_defer(cache, mg->new_ocell, false);
1236                 }
1237                 free_io_migration(mg);
1238         }
1239 }
1240
1241 static void copy_complete(int read_err, unsigned long write_err, void *context)
1242 {
1243         unsigned long flags;
1244         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1245         struct cache *cache = mg->cache;
1246
1247         if (read_err || write_err)
1248                 mg->err = true;
1249
1250         spin_lock_irqsave(&cache->lock, flags);
1251         list_add_tail(&mg->list, &cache->completed_migrations);
1252         spin_unlock_irqrestore(&cache->lock, flags);
1253
1254         wake_worker(cache);
1255 }
1256
1257 static void issue_copy(struct dm_cache_migration *mg)
1258 {
1259         int r;
1260         struct dm_io_region o_region, c_region;
1261         struct cache *cache = mg->cache;
1262         sector_t cblock = from_cblock(mg->cblock);
1263
1264         o_region.bdev = cache->origin_dev->bdev;
1265         o_region.count = cache->sectors_per_block;
1266
1267         c_region.bdev = cache->cache_dev->bdev;
1268         c_region.sector = cblock * cache->sectors_per_block;
1269         c_region.count = cache->sectors_per_block;
1270
1271         if (mg->writeback || mg->demote) {
1272                 /* demote */
1273                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1274                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1275         } else {
1276                 /* promote */
1277                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1278                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1279         }
1280
1281         if (r < 0) {
1282                 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1283                 migration_failure(mg);
1284         }
1285 }
1286
1287 static void overwrite_endio(struct bio *bio, int err)
1288 {
1289         struct dm_cache_migration *mg = bio->bi_private;
1290         struct cache *cache = mg->cache;
1291         size_t pb_data_size = get_per_bio_data_size(cache);
1292         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1293         unsigned long flags;
1294
1295         dm_unhook_bio(&pb->hook_info, bio);
1296
1297         if (err)
1298                 mg->err = true;
1299
1300         mg->requeue_holder = false;
1301
1302         spin_lock_irqsave(&cache->lock, flags);
1303         list_add_tail(&mg->list, &cache->completed_migrations);
1304         spin_unlock_irqrestore(&cache->lock, flags);
1305
1306         wake_worker(cache);
1307 }
1308
1309 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1310 {
1311         size_t pb_data_size = get_per_bio_data_size(mg->cache);
1312         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1313
1314         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1315         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1316
1317         /*
1318          * No need to inc_ds() here, since the cell will be held for the
1319          * duration of the io.
1320          */
1321         accounted_request(mg->cache, bio);
1322 }
1323
1324 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1325 {
1326         return (bio_data_dir(bio) == WRITE) &&
1327                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1328 }
1329
1330 static void avoid_copy(struct dm_cache_migration *mg)
1331 {
1332         atomic_inc(&mg->cache->stats.copies_avoided);
1333         migration_success_pre_commit(mg);
1334 }
1335
1336 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1337                                      dm_dblock_t *b, dm_dblock_t *e)
1338 {
1339         sector_t sb = bio->bi_iter.bi_sector;
1340         sector_t se = bio_end_sector(bio);
1341
1342         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1343
1344         if (se - sb < cache->discard_block_size)
1345                 *e = *b;
1346         else
1347                 *e = to_dblock(block_div(se, cache->discard_block_size));
1348 }
1349
1350 static void issue_discard(struct dm_cache_migration *mg)
1351 {
1352         dm_dblock_t b, e;
1353         struct bio *bio = mg->new_ocell->holder;
1354
1355         calc_discard_block_range(mg->cache, bio, &b, &e);
1356         while (b != e) {
1357                 set_discard(mg->cache, b);
1358                 b = to_dblock(from_dblock(b) + 1);
1359         }
1360
1361         bio_endio(bio, 0);
1362         cell_defer(mg->cache, mg->new_ocell, false);
1363         free_migration(mg);
1364         wake_worker(mg->cache);
1365 }
1366
1367 static void issue_copy_or_discard(struct dm_cache_migration *mg)
1368 {
1369         bool avoid;
1370         struct cache *cache = mg->cache;
1371
1372         if (mg->discard) {
1373                 issue_discard(mg);
1374                 return;
1375         }
1376
1377         if (mg->writeback || mg->demote)
1378                 avoid = !is_dirty(cache, mg->cblock) ||
1379                         is_discarded_oblock(cache, mg->old_oblock);
1380         else {
1381                 struct bio *bio = mg->new_ocell->holder;
1382
1383                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1384
1385                 if (writeback_mode(&cache->features) &&
1386                     !avoid && bio_writes_complete_block(cache, bio)) {
1387                         issue_overwrite(mg, bio);
1388                         return;
1389                 }
1390         }
1391
1392         avoid ? avoid_copy(mg) : issue_copy(mg);
1393 }
1394
1395 static void complete_migration(struct dm_cache_migration *mg)
1396 {
1397         if (mg->err)
1398                 migration_failure(mg);
1399         else
1400                 migration_success_pre_commit(mg);
1401 }
1402
1403 static void process_migrations(struct cache *cache, struct list_head *head,
1404                                void (*fn)(struct dm_cache_migration *))
1405 {
1406         unsigned long flags;
1407         struct list_head list;
1408         struct dm_cache_migration *mg, *tmp;
1409
1410         INIT_LIST_HEAD(&list);
1411         spin_lock_irqsave(&cache->lock, flags);
1412         list_splice_init(head, &list);
1413         spin_unlock_irqrestore(&cache->lock, flags);
1414
1415         list_for_each_entry_safe(mg, tmp, &list, list)
1416                 fn(mg);
1417 }
1418
1419 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1420 {
1421         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1422 }
1423
1424 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1425 {
1426         unsigned long flags;
1427         struct cache *cache = mg->cache;
1428
1429         spin_lock_irqsave(&cache->lock, flags);
1430         __queue_quiesced_migration(mg);
1431         spin_unlock_irqrestore(&cache->lock, flags);
1432
1433         wake_worker(cache);
1434 }
1435
1436 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1437 {
1438         unsigned long flags;
1439         struct dm_cache_migration *mg, *tmp;
1440
1441         spin_lock_irqsave(&cache->lock, flags);
1442         list_for_each_entry_safe(mg, tmp, work, list)
1443                 __queue_quiesced_migration(mg);
1444         spin_unlock_irqrestore(&cache->lock, flags);
1445
1446         wake_worker(cache);
1447 }
1448
1449 static void check_for_quiesced_migrations(struct cache *cache,
1450                                           struct per_bio_data *pb)
1451 {
1452         struct list_head work;
1453
1454         if (!pb->all_io_entry)
1455                 return;
1456
1457         INIT_LIST_HEAD(&work);
1458         dm_deferred_entry_dec(pb->all_io_entry, &work);
1459
1460         if (!list_empty(&work))
1461                 queue_quiesced_migrations(cache, &work);
1462 }
1463
1464 static void quiesce_migration(struct dm_cache_migration *mg)
1465 {
1466         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1467                 queue_quiesced_migration(mg);
1468 }
1469
1470 static void promote(struct cache *cache, struct prealloc *structs,
1471                     dm_oblock_t oblock, dm_cblock_t cblock,
1472                     struct dm_bio_prison_cell *cell)
1473 {
1474         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1475
1476         mg->err = false;
1477         mg->discard = false;
1478         mg->writeback = false;
1479         mg->demote = false;
1480         mg->promote = true;
1481         mg->requeue_holder = true;
1482         mg->invalidate = false;
1483         mg->cache = cache;
1484         mg->new_oblock = oblock;
1485         mg->cblock = cblock;
1486         mg->old_ocell = NULL;
1487         mg->new_ocell = cell;
1488         mg->start_jiffies = jiffies;
1489
1490         inc_io_migrations(cache);
1491         quiesce_migration(mg);
1492 }
1493
1494 static void writeback(struct cache *cache, struct prealloc *structs,
1495                       dm_oblock_t oblock, dm_cblock_t cblock,
1496                       struct dm_bio_prison_cell *cell)
1497 {
1498         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1499
1500         mg->err = false;
1501         mg->discard = false;
1502         mg->writeback = true;
1503         mg->demote = false;
1504         mg->promote = false;
1505         mg->requeue_holder = true;
1506         mg->invalidate = false;
1507         mg->cache = cache;
1508         mg->old_oblock = oblock;
1509         mg->cblock = cblock;
1510         mg->old_ocell = cell;
1511         mg->new_ocell = NULL;
1512         mg->start_jiffies = jiffies;
1513
1514         inc_io_migrations(cache);
1515         quiesce_migration(mg);
1516 }
1517
1518 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1519                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1520                                 dm_cblock_t cblock,
1521                                 struct dm_bio_prison_cell *old_ocell,
1522                                 struct dm_bio_prison_cell *new_ocell)
1523 {
1524         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1525
1526         mg->err = false;
1527         mg->discard = false;
1528         mg->writeback = false;
1529         mg->demote = true;
1530         mg->promote = true;
1531         mg->requeue_holder = true;
1532         mg->invalidate = false;
1533         mg->cache = cache;
1534         mg->old_oblock = old_oblock;
1535         mg->new_oblock = new_oblock;
1536         mg->cblock = cblock;
1537         mg->old_ocell = old_ocell;
1538         mg->new_ocell = new_ocell;
1539         mg->start_jiffies = jiffies;
1540
1541         inc_io_migrations(cache);
1542         quiesce_migration(mg);
1543 }
1544
1545 /*
1546  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1547  * block are thrown away.
1548  */
1549 static void invalidate(struct cache *cache, struct prealloc *structs,
1550                        dm_oblock_t oblock, dm_cblock_t cblock,
1551                        struct dm_bio_prison_cell *cell)
1552 {
1553         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1554
1555         mg->err = false;
1556         mg->discard = false;
1557         mg->writeback = false;
1558         mg->demote = true;
1559         mg->promote = false;
1560         mg->requeue_holder = true;
1561         mg->invalidate = true;
1562         mg->cache = cache;
1563         mg->old_oblock = oblock;
1564         mg->cblock = cblock;
1565         mg->old_ocell = cell;
1566         mg->new_ocell = NULL;
1567         mg->start_jiffies = jiffies;
1568
1569         inc_io_migrations(cache);
1570         quiesce_migration(mg);
1571 }
1572
1573 static void discard(struct cache *cache, struct prealloc *structs,
1574                     struct dm_bio_prison_cell *cell)
1575 {
1576         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1577
1578         mg->err = false;
1579         mg->discard = true;
1580         mg->writeback = false;
1581         mg->demote = false;
1582         mg->promote = false;
1583         mg->requeue_holder = false;
1584         mg->invalidate = false;
1585         mg->cache = cache;
1586         mg->old_ocell = NULL;
1587         mg->new_ocell = cell;
1588         mg->start_jiffies = jiffies;
1589
1590         quiesce_migration(mg);
1591 }
1592
1593 /*----------------------------------------------------------------
1594  * bio processing
1595  *--------------------------------------------------------------*/
1596 static void defer_bio(struct cache *cache, struct bio *bio)
1597 {
1598         unsigned long flags;
1599
1600         spin_lock_irqsave(&cache->lock, flags);
1601         bio_list_add(&cache->deferred_bios, bio);
1602         spin_unlock_irqrestore(&cache->lock, flags);
1603
1604         wake_worker(cache);
1605 }
1606
1607 static void process_flush_bio(struct cache *cache, struct bio *bio)
1608 {
1609         size_t pb_data_size = get_per_bio_data_size(cache);
1610         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1611
1612         BUG_ON(bio->bi_iter.bi_size);
1613         if (!pb->req_nr)
1614                 remap_to_origin(cache, bio);
1615         else
1616                 remap_to_cache(cache, bio, 0);
1617
1618         /*
1619          * REQ_FLUSH is not directed at any particular block so we don't
1620          * need to inc_ds().  REQ_FUA's are split into a write + REQ_FLUSH
1621          * by dm-core.
1622          */
1623         issue(cache, bio);
1624 }
1625
1626 static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1627                                 struct bio *bio)
1628 {
1629         int r;
1630         dm_dblock_t b, e;
1631         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1632
1633         calc_discard_block_range(cache, bio, &b, &e);
1634         if (b == e) {
1635                 bio_endio(bio, 0);
1636                 return;
1637         }
1638
1639         cell_prealloc = prealloc_get_cell(structs);
1640         r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1641                              (cell_free_fn) prealloc_put_cell,
1642                              structs, &new_ocell);
1643         if (r > 0)
1644                 return;
1645
1646         discard(cache, structs, new_ocell);
1647 }
1648
1649 static bool spare_migration_bandwidth(struct cache *cache)
1650 {
1651         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1652                 cache->sectors_per_block;
1653         return current_volume < cache->migration_threshold;
1654 }
1655
1656 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1657 {
1658         atomic_inc(bio_data_dir(bio) == READ ?
1659                    &cache->stats.read_hit : &cache->stats.write_hit);
1660 }
1661
1662 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1663 {
1664         atomic_inc(bio_data_dir(bio) == READ ?
1665                    &cache->stats.read_miss : &cache->stats.write_miss);
1666 }
1667
1668 /*----------------------------------------------------------------*/
1669
1670 struct inc_detail {
1671         struct cache *cache;
1672         struct bio_list bios_for_issue;
1673         struct bio_list unhandled_bios;
1674         bool any_writes;
1675 };
1676
1677 static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1678 {
1679         struct bio *bio;
1680         struct inc_detail *detail = context;
1681         struct cache *cache = detail->cache;
1682
1683         inc_ds(cache, cell->holder, cell);
1684         if (bio_data_dir(cell->holder) == WRITE)
1685                 detail->any_writes = true;
1686
1687         while ((bio = bio_list_pop(&cell->bios))) {
1688                 if (discard_or_flush(bio)) {
1689                         bio_list_add(&detail->unhandled_bios, bio);
1690                         continue;
1691                 }
1692
1693                 if (bio_data_dir(bio) == WRITE)
1694                         detail->any_writes = true;
1695
1696                 bio_list_add(&detail->bios_for_issue, bio);
1697                 inc_ds(cache, bio, cell);
1698         }
1699 }
1700
1701 // FIXME: refactor these two
1702 static void remap_cell_to_origin_clear_discard(struct cache *cache,
1703                                                struct dm_bio_prison_cell *cell,
1704                                                dm_oblock_t oblock, bool issue_holder)
1705 {
1706         struct bio *bio;
1707         unsigned long flags;
1708         struct inc_detail detail;
1709
1710         detail.cache = cache;
1711         bio_list_init(&detail.bios_for_issue);
1712         bio_list_init(&detail.unhandled_bios);
1713         detail.any_writes = false;
1714
1715         spin_lock_irqsave(&cache->lock, flags);
1716         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1717         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1718         spin_unlock_irqrestore(&cache->lock, flags);
1719
1720         remap_to_origin(cache, cell->holder);
1721         if (issue_holder)
1722                 issue(cache, cell->holder);
1723         else
1724                 accounted_begin(cache, cell->holder);
1725
1726         if (detail.any_writes)
1727                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1728
1729         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1730                 remap_to_origin(cache, bio);
1731                 issue(cache, bio);
1732         }
1733
1734         free_prison_cell(cache, cell);
1735 }
1736
1737 static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1738                                       dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1739 {
1740         struct bio *bio;
1741         unsigned long flags;
1742         struct inc_detail detail;
1743
1744         detail.cache = cache;
1745         bio_list_init(&detail.bios_for_issue);
1746         bio_list_init(&detail.unhandled_bios);
1747         detail.any_writes = false;
1748
1749         spin_lock_irqsave(&cache->lock, flags);
1750         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1751         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1752         spin_unlock_irqrestore(&cache->lock, flags);
1753
1754         remap_to_cache(cache, cell->holder, cblock);
1755         if (issue_holder)
1756                 issue(cache, cell->holder);
1757         else
1758                 accounted_begin(cache, cell->holder);
1759
1760         if (detail.any_writes) {
1761                 set_dirty(cache, oblock, cblock);
1762                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1763         }
1764
1765         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1766                 remap_to_cache(cache, bio, cblock);
1767                 issue(cache, bio);
1768         }
1769
1770         free_prison_cell(cache, cell);
1771 }
1772
1773 /*----------------------------------------------------------------*/
1774
1775 struct old_oblock_lock {
1776         struct policy_locker locker;
1777         struct cache *cache;
1778         struct prealloc *structs;
1779         struct dm_bio_prison_cell *cell;
1780 };
1781
1782 static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1783 {
1784         /* This should never be called */
1785         BUG();
1786         return 0;
1787 }
1788
1789 static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1790 {
1791         struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1792         struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1793
1794         return bio_detain(l->cache, b, NULL, cell_prealloc,
1795                           (cell_free_fn) prealloc_put_cell,
1796                           l->structs, &l->cell);
1797 }
1798
1799 static void process_cell(struct cache *cache, struct prealloc *structs,
1800                          struct dm_bio_prison_cell *new_ocell)
1801 {
1802         int r;
1803         bool release_cell = true;
1804         struct bio *bio = new_ocell->holder;
1805         dm_oblock_t block = get_bio_block(cache, bio);
1806         struct policy_result lookup_result;
1807         bool passthrough = passthrough_mode(&cache->features);
1808         bool fast_promotion, can_migrate;
1809         struct old_oblock_lock ool;
1810
1811         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1812         can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1813
1814         ool.locker.fn = cell_locker;
1815         ool.cache = cache;
1816         ool.structs = structs;
1817         ool.cell = NULL;
1818         r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1819                        bio, &ool.locker, &lookup_result);
1820
1821         if (r == -EWOULDBLOCK)
1822                 /* migration has been denied */
1823                 lookup_result.op = POLICY_MISS;
1824
1825         switch (lookup_result.op) {
1826         case POLICY_HIT:
1827                 if (passthrough) {
1828                         inc_miss_counter(cache, bio);
1829
1830                         /*
1831                          * Passthrough always maps to the origin,
1832                          * invalidating any cache blocks that are written
1833                          * to.
1834                          */
1835
1836                         if (bio_data_dir(bio) == WRITE) {
1837                                 atomic_inc(&cache->stats.demotion);
1838                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1839                                 release_cell = false;
1840
1841                         } else {
1842                                 /* FIXME: factor out issue_origin() */
1843                                 remap_to_origin_clear_discard(cache, bio, block);
1844                                 inc_and_issue(cache, bio, new_ocell);
1845                         }
1846                 } else {
1847                         inc_hit_counter(cache, bio);
1848
1849                         if (bio_data_dir(bio) == WRITE &&
1850                             writethrough_mode(&cache->features) &&
1851                             !is_dirty(cache, lookup_result.cblock)) {
1852                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1853                                 inc_and_issue(cache, bio, new_ocell);
1854
1855                         } else {
1856                                 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1857                                 release_cell = false;
1858                         }
1859                 }
1860
1861                 break;
1862
1863         case POLICY_MISS:
1864                 inc_miss_counter(cache, bio);
1865                 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1866                 release_cell = false;
1867                 break;
1868
1869         case POLICY_NEW:
1870                 atomic_inc(&cache->stats.promotion);
1871                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1872                 release_cell = false;
1873                 break;
1874
1875         case POLICY_REPLACE:
1876                 atomic_inc(&cache->stats.demotion);
1877                 atomic_inc(&cache->stats.promotion);
1878                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1879                                     block, lookup_result.cblock,
1880                                     ool.cell, new_ocell);
1881                 release_cell = false;
1882                 break;
1883
1884         default:
1885                 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1886                             cache_device_name(cache), __func__,
1887                             (unsigned) lookup_result.op);
1888                 bio_io_error(bio);
1889         }
1890
1891         if (release_cell)
1892                 cell_defer(cache, new_ocell, false);
1893 }
1894
1895 static void process_bio(struct cache *cache, struct prealloc *structs,
1896                         struct bio *bio)
1897 {
1898         int r;
1899         dm_oblock_t block = get_bio_block(cache, bio);
1900         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1901
1902         /*
1903          * Check to see if that block is currently migrating.
1904          */
1905         cell_prealloc = prealloc_get_cell(structs);
1906         r = bio_detain(cache, block, bio, cell_prealloc,
1907                        (cell_free_fn) prealloc_put_cell,
1908                        structs, &new_ocell);
1909         if (r > 0)
1910                 return;
1911
1912         process_cell(cache, structs, new_ocell);
1913 }
1914
1915 static int need_commit_due_to_time(struct cache *cache)
1916 {
1917         return jiffies < cache->last_commit_jiffies ||
1918                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1919 }
1920
1921 /*
1922  * A non-zero return indicates read_only or fail_io mode.
1923  */
1924 static int commit(struct cache *cache, bool clean_shutdown)
1925 {
1926         int r;
1927
1928         if (get_cache_mode(cache) >= CM_READ_ONLY)
1929                 return -EINVAL;
1930
1931         atomic_inc(&cache->stats.commit_count);
1932         r = dm_cache_commit(cache->cmd, clean_shutdown);
1933         if (r)
1934                 metadata_operation_failed(cache, "dm_cache_commit", r);
1935
1936         return r;
1937 }
1938
1939 static int commit_if_needed(struct cache *cache)
1940 {
1941         int r = 0;
1942
1943         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1944             dm_cache_changed_this_transaction(cache->cmd)) {
1945                 r = commit(cache, false);
1946                 cache->commit_requested = false;
1947                 cache->last_commit_jiffies = jiffies;
1948         }
1949
1950         return r;
1951 }
1952
1953 static void process_deferred_bios(struct cache *cache)
1954 {
1955         bool prealloc_used = false;
1956         unsigned long flags;
1957         struct bio_list bios;
1958         struct bio *bio;
1959         struct prealloc structs;
1960
1961         memset(&structs, 0, sizeof(structs));
1962         bio_list_init(&bios);
1963
1964         spin_lock_irqsave(&cache->lock, flags);
1965         bio_list_merge(&bios, &cache->deferred_bios);
1966         bio_list_init(&cache->deferred_bios);
1967         spin_unlock_irqrestore(&cache->lock, flags);
1968
1969         while (!bio_list_empty(&bios)) {
1970                 /*
1971                  * If we've got no free migration structs, and processing
1972                  * this bio might require one, we pause until there are some
1973                  * prepared mappings to process.
1974                  */
1975                 prealloc_used = true;
1976                 if (prealloc_data_structs(cache, &structs)) {
1977                         spin_lock_irqsave(&cache->lock, flags);
1978                         bio_list_merge(&cache->deferred_bios, &bios);
1979                         spin_unlock_irqrestore(&cache->lock, flags);
1980                         break;
1981                 }
1982
1983                 bio = bio_list_pop(&bios);
1984
1985                 if (bio->bi_rw & REQ_FLUSH)
1986                         process_flush_bio(cache, bio);
1987                 else if (bio->bi_rw & REQ_DISCARD)
1988                         process_discard_bio(cache, &structs, bio);
1989                 else
1990                         process_bio(cache, &structs, bio);
1991         }
1992
1993         if (prealloc_used)
1994                 prealloc_free_structs(cache, &structs);
1995 }
1996
1997 static void process_deferred_cells(struct cache *cache)
1998 {
1999         bool prealloc_used = false;
2000         unsigned long flags;
2001         struct dm_bio_prison_cell *cell, *tmp;
2002         struct list_head cells;
2003         struct prealloc structs;
2004
2005         memset(&structs, 0, sizeof(structs));
2006
2007         INIT_LIST_HEAD(&cells);
2008
2009         spin_lock_irqsave(&cache->lock, flags);
2010         list_splice_init(&cache->deferred_cells, &cells);
2011         spin_unlock_irqrestore(&cache->lock, flags);
2012
2013         list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2014                 /*
2015                  * If we've got no free migration structs, and processing
2016                  * this bio might require one, we pause until there are some
2017                  * prepared mappings to process.
2018                  */
2019                 prealloc_used = true;
2020                 if (prealloc_data_structs(cache, &structs)) {
2021                         spin_lock_irqsave(&cache->lock, flags);
2022                         list_splice(&cells, &cache->deferred_cells);
2023                         spin_unlock_irqrestore(&cache->lock, flags);
2024                         break;
2025                 }
2026
2027                 process_cell(cache, &structs, cell);
2028         }
2029
2030         if (prealloc_used)
2031                 prealloc_free_structs(cache, &structs);
2032 }
2033
2034 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2035 {
2036         unsigned long flags;
2037         struct bio_list bios;
2038         struct bio *bio;
2039
2040         bio_list_init(&bios);
2041
2042         spin_lock_irqsave(&cache->lock, flags);
2043         bio_list_merge(&bios, &cache->deferred_flush_bios);
2044         bio_list_init(&cache->deferred_flush_bios);
2045         spin_unlock_irqrestore(&cache->lock, flags);
2046
2047         /*
2048          * These bios have already been through inc_ds()
2049          */
2050         while ((bio = bio_list_pop(&bios)))
2051                 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2052 }
2053
2054 static void process_deferred_writethrough_bios(struct cache *cache)
2055 {
2056         unsigned long flags;
2057         struct bio_list bios;
2058         struct bio *bio;
2059
2060         bio_list_init(&bios);
2061
2062         spin_lock_irqsave(&cache->lock, flags);
2063         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2064         bio_list_init(&cache->deferred_writethrough_bios);
2065         spin_unlock_irqrestore(&cache->lock, flags);
2066
2067         /*
2068          * These bios have already been through inc_ds()
2069          */
2070         while ((bio = bio_list_pop(&bios)))
2071                 accounted_request(cache, bio);
2072 }
2073
2074 static void writeback_some_dirty_blocks(struct cache *cache)
2075 {
2076         bool prealloc_used = false;
2077         dm_oblock_t oblock;
2078         dm_cblock_t cblock;
2079         struct prealloc structs;
2080         struct dm_bio_prison_cell *old_ocell;
2081         bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2082
2083         memset(&structs, 0, sizeof(structs));
2084
2085         while (spare_migration_bandwidth(cache)) {
2086                 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2087                         break; /* no work to do */
2088
2089                 prealloc_used = true;
2090                 if (prealloc_data_structs(cache, &structs) ||
2091                     get_cell(cache, oblock, &structs, &old_ocell)) {
2092                         policy_set_dirty(cache->policy, oblock);
2093                         break;
2094                 }
2095
2096                 writeback(cache, &structs, oblock, cblock, old_ocell);
2097         }
2098
2099         if (prealloc_used)
2100                 prealloc_free_structs(cache, &structs);
2101 }
2102
2103 /*----------------------------------------------------------------
2104  * Invalidations.
2105  * Dropping something from the cache *without* writing back.
2106  *--------------------------------------------------------------*/
2107
2108 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2109 {
2110         int r = 0;
2111         uint64_t begin = from_cblock(req->cblocks->begin);
2112         uint64_t end = from_cblock(req->cblocks->end);
2113
2114         while (begin != end) {
2115                 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2116                 if (!r) {
2117                         r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2118                         if (r) {
2119                                 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2120                                 break;
2121                         }
2122
2123                 } else if (r == -ENODATA) {
2124                         /* harmless, already unmapped */
2125                         r = 0;
2126
2127                 } else {
2128                         DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2129                         break;
2130                 }
2131
2132                 begin++;
2133         }
2134
2135         cache->commit_requested = true;
2136
2137         req->err = r;
2138         atomic_set(&req->complete, 1);
2139
2140         wake_up(&req->result_wait);
2141 }
2142
2143 static void process_invalidation_requests(struct cache *cache)
2144 {
2145         struct list_head list;
2146         struct invalidation_request *req, *tmp;
2147
2148         INIT_LIST_HEAD(&list);
2149         spin_lock(&cache->invalidation_lock);
2150         list_splice_init(&cache->invalidation_requests, &list);
2151         spin_unlock(&cache->invalidation_lock);
2152
2153         list_for_each_entry_safe (req, tmp, &list, list)
2154                 process_invalidation_request(cache, req);
2155 }
2156
2157 /*----------------------------------------------------------------
2158  * Main worker loop
2159  *--------------------------------------------------------------*/
2160 static bool is_quiescing(struct cache *cache)
2161 {
2162         return atomic_read(&cache->quiescing);
2163 }
2164
2165 static void ack_quiescing(struct cache *cache)
2166 {
2167         if (is_quiescing(cache)) {
2168                 atomic_inc(&cache->quiescing_ack);
2169                 wake_up(&cache->quiescing_wait);
2170         }
2171 }
2172
2173 static void wait_for_quiescing_ack(struct cache *cache)
2174 {
2175         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2176 }
2177
2178 static void start_quiescing(struct cache *cache)
2179 {
2180         atomic_inc(&cache->quiescing);
2181         wait_for_quiescing_ack(cache);
2182 }
2183
2184 static void stop_quiescing(struct cache *cache)
2185 {
2186         atomic_set(&cache->quiescing, 0);
2187         atomic_set(&cache->quiescing_ack, 0);
2188 }
2189
2190 static void wait_for_migrations(struct cache *cache)
2191 {
2192         wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2193 }
2194
2195 static void stop_worker(struct cache *cache)
2196 {
2197         cancel_delayed_work(&cache->waker);
2198         flush_workqueue(cache->wq);
2199 }
2200
2201 static void requeue_deferred_cells(struct cache *cache)
2202 {
2203         unsigned long flags;
2204         struct list_head cells;
2205         struct dm_bio_prison_cell *cell, *tmp;
2206
2207         INIT_LIST_HEAD(&cells);
2208         spin_lock_irqsave(&cache->lock, flags);
2209         list_splice_init(&cache->deferred_cells, &cells);
2210         spin_unlock_irqrestore(&cache->lock, flags);
2211
2212         list_for_each_entry_safe(cell, tmp, &cells, user_list)
2213                 cell_requeue(cache, cell);
2214 }
2215
2216 static void requeue_deferred_bios(struct cache *cache)
2217 {
2218         struct bio *bio;
2219         struct bio_list bios;
2220
2221         bio_list_init(&bios);
2222         bio_list_merge(&bios, &cache->deferred_bios);
2223         bio_list_init(&cache->deferred_bios);
2224
2225         while ((bio = bio_list_pop(&bios)))
2226                 bio_endio(bio, DM_ENDIO_REQUEUE);
2227 }
2228
2229 static int more_work(struct cache *cache)
2230 {
2231         if (is_quiescing(cache))
2232                 return !list_empty(&cache->quiesced_migrations) ||
2233                         !list_empty(&cache->completed_migrations) ||
2234                         !list_empty(&cache->need_commit_migrations);
2235         else
2236                 return !bio_list_empty(&cache->deferred_bios) ||
2237                         !list_empty(&cache->deferred_cells) ||
2238                         !bio_list_empty(&cache->deferred_flush_bios) ||
2239                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
2240                         !list_empty(&cache->quiesced_migrations) ||
2241                         !list_empty(&cache->completed_migrations) ||
2242                         !list_empty(&cache->need_commit_migrations) ||
2243                         cache->invalidate;
2244 }
2245
2246 static void do_worker(struct work_struct *ws)
2247 {
2248         struct cache *cache = container_of(ws, struct cache, worker);
2249
2250         do {
2251                 if (!is_quiescing(cache)) {
2252                         writeback_some_dirty_blocks(cache);
2253                         process_deferred_writethrough_bios(cache);
2254                         process_deferred_bios(cache);
2255                         process_deferred_cells(cache);
2256                         process_invalidation_requests(cache);
2257                 }
2258
2259                 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2260                 process_migrations(cache, &cache->completed_migrations, complete_migration);
2261
2262                 if (commit_if_needed(cache)) {
2263                         process_deferred_flush_bios(cache, false);
2264                         process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2265                 } else {
2266                         process_deferred_flush_bios(cache, true);
2267                         process_migrations(cache, &cache->need_commit_migrations,
2268                                            migration_success_post_commit);
2269                 }
2270
2271                 ack_quiescing(cache);
2272
2273         } while (more_work(cache));
2274 }
2275
2276 /*
2277  * We want to commit periodically so that not too much
2278  * unwritten metadata builds up.
2279  */
2280 static void do_waker(struct work_struct *ws)
2281 {
2282         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2283         policy_tick(cache->policy, true);
2284         wake_worker(cache);
2285         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2286 }
2287
2288 /*----------------------------------------------------------------*/
2289
2290 static int is_congested(struct dm_dev *dev, int bdi_bits)
2291 {
2292         struct request_queue *q = bdev_get_queue(dev->bdev);
2293         return bdi_congested(&q->backing_dev_info, bdi_bits);
2294 }
2295
2296 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2297 {
2298         struct cache *cache = container_of(cb, struct cache, callbacks);
2299
2300         return is_congested(cache->origin_dev, bdi_bits) ||
2301                 is_congested(cache->cache_dev, bdi_bits);
2302 }
2303
2304 /*----------------------------------------------------------------
2305  * Target methods
2306  *--------------------------------------------------------------*/
2307
2308 /*
2309  * This function gets called on the error paths of the constructor, so we
2310  * have to cope with a partially initialised struct.
2311  */
2312 static void destroy(struct cache *cache)
2313 {
2314         unsigned i;
2315
2316         if (cache->migration_pool)
2317                 mempool_destroy(cache->migration_pool);
2318
2319         if (cache->all_io_ds)
2320                 dm_deferred_set_destroy(cache->all_io_ds);
2321
2322         if (cache->prison)
2323                 dm_bio_prison_destroy(cache->prison);
2324
2325         if (cache->wq)
2326                 destroy_workqueue(cache->wq);
2327
2328         if (cache->dirty_bitset)
2329                 free_bitset(cache->dirty_bitset);
2330
2331         if (cache->discard_bitset)
2332                 free_bitset(cache->discard_bitset);
2333
2334         if (cache->copier)
2335                 dm_kcopyd_client_destroy(cache->copier);
2336
2337         if (cache->cmd)
2338                 dm_cache_metadata_close(cache->cmd);
2339
2340         if (cache->metadata_dev)
2341                 dm_put_device(cache->ti, cache->metadata_dev);
2342
2343         if (cache->origin_dev)
2344                 dm_put_device(cache->ti, cache->origin_dev);
2345
2346         if (cache->cache_dev)
2347                 dm_put_device(cache->ti, cache->cache_dev);
2348
2349         if (cache->policy)
2350                 dm_cache_policy_destroy(cache->policy);
2351
2352         for (i = 0; i < cache->nr_ctr_args ; i++)
2353                 kfree(cache->ctr_args[i]);
2354         kfree(cache->ctr_args);
2355
2356         kfree(cache);
2357 }
2358
2359 static void cache_dtr(struct dm_target *ti)
2360 {
2361         struct cache *cache = ti->private;
2362
2363         destroy(cache);
2364 }
2365
2366 static sector_t get_dev_size(struct dm_dev *dev)
2367 {
2368         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2369 }
2370
2371 /*----------------------------------------------------------------*/
2372
2373 /*
2374  * Construct a cache device mapping.
2375  *
2376  * cache <metadata dev> <cache dev> <origin dev> <block size>
2377  *       <#feature args> [<feature arg>]*
2378  *       <policy> <#policy args> [<policy arg>]*
2379  *
2380  * metadata dev    : fast device holding the persistent metadata
2381  * cache dev       : fast device holding cached data blocks
2382  * origin dev      : slow device holding original data blocks
2383  * block size      : cache unit size in sectors
2384  *
2385  * #feature args   : number of feature arguments passed
2386  * feature args    : writethrough.  (The default is writeback.)
2387  *
2388  * policy          : the replacement policy to use
2389  * #policy args    : an even number of policy arguments corresponding
2390  *                   to key/value pairs passed to the policy
2391  * policy args     : key/value pairs passed to the policy
2392  *                   E.g. 'sequential_threshold 1024'
2393  *                   See cache-policies.txt for details.
2394  *
2395  * Optional feature arguments are:
2396  *   writethrough  : write through caching that prohibits cache block
2397  *                   content from being different from origin block content.
2398  *                   Without this argument, the default behaviour is to write
2399  *                   back cache block contents later for performance reasons,
2400  *                   so they may differ from the corresponding origin blocks.
2401  */
2402 struct cache_args {
2403         struct dm_target *ti;
2404
2405         struct dm_dev *metadata_dev;
2406
2407         struct dm_dev *cache_dev;
2408         sector_t cache_sectors;
2409
2410         struct dm_dev *origin_dev;
2411         sector_t origin_sectors;
2412
2413         uint32_t block_size;
2414
2415         const char *policy_name;
2416         int policy_argc;
2417         const char **policy_argv;
2418
2419         struct cache_features features;
2420 };
2421
2422 static void destroy_cache_args(struct cache_args *ca)
2423 {
2424         if (ca->metadata_dev)
2425                 dm_put_device(ca->ti, ca->metadata_dev);
2426
2427         if (ca->cache_dev)
2428                 dm_put_device(ca->ti, ca->cache_dev);
2429
2430         if (ca->origin_dev)
2431                 dm_put_device(ca->ti, ca->origin_dev);
2432
2433         kfree(ca);
2434 }
2435
2436 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2437 {
2438         if (!as->argc) {
2439                 *error = "Insufficient args";
2440                 return false;
2441         }
2442
2443         return true;
2444 }
2445
2446 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2447                               char **error)
2448 {
2449         int r;
2450         sector_t metadata_dev_size;
2451         char b[BDEVNAME_SIZE];
2452
2453         if (!at_least_one_arg(as, error))
2454                 return -EINVAL;
2455
2456         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2457                           &ca->metadata_dev);
2458         if (r) {
2459                 *error = "Error opening metadata device";
2460                 return r;
2461         }
2462
2463         metadata_dev_size = get_dev_size(ca->metadata_dev);
2464         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2465                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2466                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2467
2468         return 0;
2469 }
2470
2471 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2472                            char **error)
2473 {
2474         int r;
2475
2476         if (!at_least_one_arg(as, error))
2477                 return -EINVAL;
2478
2479         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2480                           &ca->cache_dev);
2481         if (r) {
2482                 *error = "Error opening cache device";
2483                 return r;
2484         }
2485         ca->cache_sectors = get_dev_size(ca->cache_dev);
2486
2487         return 0;
2488 }
2489
2490 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2491                             char **error)
2492 {
2493         int r;
2494
2495         if (!at_least_one_arg(as, error))
2496                 return -EINVAL;
2497
2498         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2499                           &ca->origin_dev);
2500         if (r) {
2501                 *error = "Error opening origin device";
2502                 return r;
2503         }
2504
2505         ca->origin_sectors = get_dev_size(ca->origin_dev);
2506         if (ca->ti->len > ca->origin_sectors) {
2507                 *error = "Device size larger than cached device";
2508                 return -EINVAL;
2509         }
2510
2511         return 0;
2512 }
2513
2514 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2515                             char **error)
2516 {
2517         unsigned long block_size;
2518
2519         if (!at_least_one_arg(as, error))
2520                 return -EINVAL;
2521
2522         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2523             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2524             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2525             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2526                 *error = "Invalid data block size";
2527                 return -EINVAL;
2528         }
2529
2530         if (block_size > ca->cache_sectors) {
2531                 *error = "Data block size is larger than the cache device";
2532                 return -EINVAL;
2533         }
2534
2535         ca->block_size = block_size;
2536
2537         return 0;
2538 }
2539
2540 static void init_features(struct cache_features *cf)
2541 {
2542         cf->mode = CM_WRITE;
2543         cf->io_mode = CM_IO_WRITEBACK;
2544 }
2545
2546 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2547                           char **error)
2548 {
2549         static struct dm_arg _args[] = {
2550                 {0, 1, "Invalid number of cache feature arguments"},
2551         };
2552
2553         int r;
2554         unsigned argc;
2555         const char *arg;
2556         struct cache_features *cf = &ca->features;
2557
2558         init_features(cf);
2559
2560         r = dm_read_arg_group(_args, as, &argc, error);
2561         if (r)
2562                 return -EINVAL;
2563
2564         while (argc--) {
2565                 arg = dm_shift_arg(as);
2566
2567                 if (!strcasecmp(arg, "writeback"))
2568                         cf->io_mode = CM_IO_WRITEBACK;
2569
2570                 else if (!strcasecmp(arg, "writethrough"))
2571                         cf->io_mode = CM_IO_WRITETHROUGH;
2572
2573                 else if (!strcasecmp(arg, "passthrough"))
2574                         cf->io_mode = CM_IO_PASSTHROUGH;
2575
2576                 else {
2577                         *error = "Unrecognised cache feature requested";
2578                         return -EINVAL;
2579                 }
2580         }
2581
2582         return 0;
2583 }
2584
2585 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2586                         char **error)
2587 {
2588         static struct dm_arg _args[] = {
2589                 {0, 1024, "Invalid number of policy arguments"},
2590         };
2591
2592         int r;
2593
2594         if (!at_least_one_arg(as, error))
2595                 return -EINVAL;
2596
2597         ca->policy_name = dm_shift_arg(as);
2598
2599         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2600         if (r)
2601                 return -EINVAL;
2602
2603         ca->policy_argv = (const char **)as->argv;
2604         dm_consume_args(as, ca->policy_argc);
2605
2606         return 0;
2607 }
2608
2609 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2610                             char **error)
2611 {
2612         int r;
2613         struct dm_arg_set as;
2614
2615         as.argc = argc;
2616         as.argv = argv;
2617
2618         r = parse_metadata_dev(ca, &as, error);
2619         if (r)
2620                 return r;
2621
2622         r = parse_cache_dev(ca, &as, error);
2623         if (r)
2624                 return r;
2625
2626         r = parse_origin_dev(ca, &as, error);
2627         if (r)
2628                 return r;
2629
2630         r = parse_block_size(ca, &as, error);
2631         if (r)
2632                 return r;
2633
2634         r = parse_features(ca, &as, error);
2635         if (r)
2636                 return r;
2637
2638         r = parse_policy(ca, &as, error);
2639         if (r)
2640                 return r;
2641
2642         return 0;
2643 }
2644
2645 /*----------------------------------------------------------------*/
2646
2647 static struct kmem_cache *migration_cache;
2648
2649 #define NOT_CORE_OPTION 1
2650
2651 static int process_config_option(struct cache *cache, const char *key, const char *value)
2652 {
2653         unsigned long tmp;
2654
2655         if (!strcasecmp(key, "migration_threshold")) {
2656                 if (kstrtoul(value, 10, &tmp))
2657                         return -EINVAL;
2658
2659                 cache->migration_threshold = tmp;
2660                 return 0;
2661         }
2662
2663         return NOT_CORE_OPTION;
2664 }
2665
2666 static int set_config_value(struct cache *cache, const char *key, const char *value)
2667 {
2668         int r = process_config_option(cache, key, value);
2669
2670         if (r == NOT_CORE_OPTION)
2671                 r = policy_set_config_value(cache->policy, key, value);
2672
2673         if (r)
2674                 DMWARN("bad config value for %s: %s", key, value);
2675
2676         return r;
2677 }
2678
2679 static int set_config_values(struct cache *cache, int argc, const char **argv)
2680 {
2681         int r = 0;
2682
2683         if (argc & 1) {
2684                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2685                 return -EINVAL;
2686         }
2687
2688         while (argc) {
2689                 r = set_config_value(cache, argv[0], argv[1]);
2690                 if (r)
2691                         break;
2692
2693                 argc -= 2;
2694                 argv += 2;
2695         }
2696
2697         return r;
2698 }
2699
2700 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2701                                char **error)
2702 {
2703         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2704                                                            cache->cache_size,
2705                                                            cache->origin_sectors,
2706                                                            cache->sectors_per_block);
2707         if (IS_ERR(p)) {
2708                 *error = "Error creating cache's policy";
2709                 return PTR_ERR(p);
2710         }
2711         cache->policy = p;
2712
2713         return 0;
2714 }
2715
2716 /*
2717  * We want the discard block size to be at least the size of the cache
2718  * block size and have no more than 2^14 discard blocks across the origin.
2719  */
2720 #define MAX_DISCARD_BLOCKS (1 << 14)
2721
2722 static bool too_many_discard_blocks(sector_t discard_block_size,
2723                                     sector_t origin_size)
2724 {
2725         (void) sector_div(origin_size, discard_block_size);
2726
2727         return origin_size > MAX_DISCARD_BLOCKS;
2728 }
2729
2730 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2731                                              sector_t origin_size)
2732 {
2733         sector_t discard_block_size = cache_block_size;
2734
2735         if (origin_size)
2736                 while (too_many_discard_blocks(discard_block_size, origin_size))
2737                         discard_block_size *= 2;
2738
2739         return discard_block_size;
2740 }
2741
2742 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2743 {
2744         dm_block_t nr_blocks = from_cblock(size);
2745
2746         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2747                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2748                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2749                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2750                              (unsigned long long) nr_blocks);
2751
2752         cache->cache_size = size;
2753 }
2754
2755 #define DEFAULT_MIGRATION_THRESHOLD 2048
2756
2757 static int cache_create(struct cache_args *ca, struct cache **result)
2758 {
2759         int r = 0;
2760         char **error = &ca->ti->error;
2761         struct cache *cache;
2762         struct dm_target *ti = ca->ti;
2763         dm_block_t origin_blocks;
2764         struct dm_cache_metadata *cmd;
2765         bool may_format = ca->features.mode == CM_WRITE;
2766
2767         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2768         if (!cache)
2769                 return -ENOMEM;
2770
2771         cache->ti = ca->ti;
2772         ti->private = cache;
2773         ti->num_flush_bios = 2;
2774         ti->flush_supported = true;
2775
2776         ti->num_discard_bios = 1;
2777         ti->discards_supported = true;
2778         ti->discard_zeroes_data_unsupported = true;
2779         ti->split_discard_bios = false;
2780
2781         cache->features = ca->features;
2782         ti->per_bio_data_size = get_per_bio_data_size(cache);
2783
2784         cache->callbacks.congested_fn = cache_is_congested;
2785         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2786
2787         cache->metadata_dev = ca->metadata_dev;
2788         cache->origin_dev = ca->origin_dev;
2789         cache->cache_dev = ca->cache_dev;
2790
2791         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2792
2793         /* FIXME: factor out this whole section */
2794         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2795         origin_blocks = block_div(origin_blocks, ca->block_size);
2796         cache->origin_blocks = to_oblock(origin_blocks);
2797
2798         cache->sectors_per_block = ca->block_size;
2799         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2800                 r = -EINVAL;
2801                 goto bad;
2802         }
2803
2804         if (ca->block_size & (ca->block_size - 1)) {
2805                 dm_block_t cache_size = ca->cache_sectors;
2806
2807                 cache->sectors_per_block_shift = -1;
2808                 cache_size = block_div(cache_size, ca->block_size);
2809                 set_cache_size(cache, to_cblock(cache_size));
2810         } else {
2811                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2812                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2813         }
2814
2815         r = create_cache_policy(cache, ca, error);
2816         if (r)
2817                 goto bad;
2818
2819         cache->policy_nr_args = ca->policy_argc;
2820         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2821
2822         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2823         if (r) {
2824                 *error = "Error setting cache policy's config values";
2825                 goto bad;
2826         }
2827
2828         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2829                                      ca->block_size, may_format,
2830                                      dm_cache_policy_get_hint_size(cache->policy));
2831         if (IS_ERR(cmd)) {
2832                 *error = "Error creating metadata object";
2833                 r = PTR_ERR(cmd);
2834                 goto bad;
2835         }
2836         cache->cmd = cmd;
2837         set_cache_mode(cache, CM_WRITE);
2838         if (get_cache_mode(cache) != CM_WRITE) {
2839                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2840                 r = -EINVAL;
2841                 goto bad;
2842         }
2843
2844         if (passthrough_mode(&cache->features)) {
2845                 bool all_clean;
2846
2847                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2848                 if (r) {
2849                         *error = "dm_cache_metadata_all_clean() failed";
2850                         goto bad;
2851                 }
2852
2853                 if (!all_clean) {
2854                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2855                         r = -EINVAL;
2856                         goto bad;
2857                 }
2858         }
2859
2860         spin_lock_init(&cache->lock);
2861         INIT_LIST_HEAD(&cache->deferred_cells);
2862         bio_list_init(&cache->deferred_bios);
2863         bio_list_init(&cache->deferred_flush_bios);
2864         bio_list_init(&cache->deferred_writethrough_bios);
2865         INIT_LIST_HEAD(&cache->quiesced_migrations);
2866         INIT_LIST_HEAD(&cache->completed_migrations);
2867         INIT_LIST_HEAD(&cache->need_commit_migrations);
2868         atomic_set(&cache->nr_allocated_migrations, 0);
2869         atomic_set(&cache->nr_io_migrations, 0);
2870         init_waitqueue_head(&cache->migration_wait);
2871
2872         init_waitqueue_head(&cache->quiescing_wait);
2873         atomic_set(&cache->quiescing, 0);
2874         atomic_set(&cache->quiescing_ack, 0);
2875
2876         r = -ENOMEM;
2877         atomic_set(&cache->nr_dirty, 0);
2878         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2879         if (!cache->dirty_bitset) {
2880                 *error = "could not allocate dirty bitset";
2881                 goto bad;
2882         }
2883         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2884
2885         cache->discard_block_size =
2886                 calculate_discard_block_size(cache->sectors_per_block,
2887                                              cache->origin_sectors);
2888         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2889                                                               cache->discard_block_size));
2890         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2891         if (!cache->discard_bitset) {
2892                 *error = "could not allocate discard bitset";
2893                 goto bad;
2894         }
2895         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2896
2897         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2898         if (IS_ERR(cache->copier)) {
2899                 *error = "could not create kcopyd client";
2900                 r = PTR_ERR(cache->copier);
2901                 goto bad;
2902         }
2903
2904         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2905         if (!cache->wq) {
2906                 *error = "could not create workqueue for metadata object";
2907                 goto bad;
2908         }
2909         INIT_WORK(&cache->worker, do_worker);
2910         INIT_DELAYED_WORK(&cache->waker, do_waker);
2911         cache->last_commit_jiffies = jiffies;
2912
2913         cache->prison = dm_bio_prison_create();
2914         if (!cache->prison) {
2915                 *error = "could not create bio prison";
2916                 goto bad;
2917         }
2918
2919         cache->all_io_ds = dm_deferred_set_create();
2920         if (!cache->all_io_ds) {
2921                 *error = "could not create all_io deferred set";
2922                 goto bad;
2923         }
2924
2925         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2926                                                          migration_cache);
2927         if (!cache->migration_pool) {
2928                 *error = "Error creating cache's migration mempool";
2929                 goto bad;
2930         }
2931
2932         cache->need_tick_bio = true;
2933         cache->sized = false;
2934         cache->invalidate = false;
2935         cache->commit_requested = false;
2936         cache->loaded_mappings = false;
2937         cache->loaded_discards = false;
2938
2939         load_stats(cache);
2940
2941         atomic_set(&cache->stats.demotion, 0);
2942         atomic_set(&cache->stats.promotion, 0);
2943         atomic_set(&cache->stats.copies_avoided, 0);
2944         atomic_set(&cache->stats.cache_cell_clash, 0);
2945         atomic_set(&cache->stats.commit_count, 0);
2946         atomic_set(&cache->stats.discard_count, 0);
2947
2948         spin_lock_init(&cache->invalidation_lock);
2949         INIT_LIST_HEAD(&cache->invalidation_requests);
2950
2951         iot_init(&cache->origin_tracker);
2952
2953         *result = cache;
2954         return 0;
2955
2956 bad:
2957         destroy(cache);
2958         return r;
2959 }
2960
2961 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2962 {
2963         unsigned i;
2964         const char **copy;
2965
2966         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2967         if (!copy)
2968                 return -ENOMEM;
2969         for (i = 0; i < argc; i++) {
2970                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2971                 if (!copy[i]) {
2972                         while (i--)
2973                                 kfree(copy[i]);
2974                         kfree(copy);
2975                         return -ENOMEM;
2976                 }
2977         }
2978
2979         cache->nr_ctr_args = argc;
2980         cache->ctr_args = copy;
2981
2982         return 0;
2983 }
2984
2985 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2986 {
2987         int r = -EINVAL;
2988         struct cache_args *ca;
2989         struct cache *cache = NULL;
2990
2991         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2992         if (!ca) {
2993                 ti->error = "Error allocating memory for cache";
2994                 return -ENOMEM;
2995         }
2996         ca->ti = ti;
2997
2998         r = parse_cache_args(ca, argc, argv, &ti->error);
2999         if (r)
3000                 goto out;
3001
3002         r = cache_create(ca, &cache);
3003         if (r)
3004                 goto out;
3005
3006         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
3007         if (r) {
3008                 destroy(cache);
3009                 goto out;
3010         }
3011
3012         ti->private = cache;
3013
3014 out:
3015         destroy_cache_args(ca);
3016         return r;
3017 }
3018
3019 /*----------------------------------------------------------------*/
3020
3021 static int cache_map(struct dm_target *ti, struct bio *bio)
3022 {
3023         struct cache *cache = ti->private;
3024
3025         int r;
3026         struct dm_bio_prison_cell *cell = NULL;
3027         dm_oblock_t block = get_bio_block(cache, bio);
3028         size_t pb_data_size = get_per_bio_data_size(cache);
3029         bool can_migrate = false;
3030         bool fast_promotion;
3031         struct policy_result lookup_result;
3032         struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3033         struct old_oblock_lock ool;
3034
3035         ool.locker.fn = null_locker;
3036
3037         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3038                 /*
3039                  * This can only occur if the io goes to a partial block at
3040                  * the end of the origin device.  We don't cache these.
3041                  * Just remap to the origin and carry on.
3042                  */
3043                 remap_to_origin(cache, bio);
3044                 accounted_begin(cache, bio);
3045                 return DM_MAPIO_REMAPPED;
3046         }
3047
3048         if (discard_or_flush(bio)) {
3049                 defer_bio(cache, bio);
3050                 return DM_MAPIO_SUBMITTED;
3051         }
3052
3053         /*
3054          * Check to see if that block is currently migrating.
3055          */
3056         cell = alloc_prison_cell(cache);
3057         if (!cell) {
3058                 defer_bio(cache, bio);
3059                 return DM_MAPIO_SUBMITTED;
3060         }
3061
3062         r = bio_detain(cache, block, bio, cell,
3063                        (cell_free_fn) free_prison_cell,
3064                        cache, &cell);
3065         if (r) {
3066                 if (r < 0)
3067                         defer_bio(cache, bio);
3068
3069                 return DM_MAPIO_SUBMITTED;
3070         }
3071
3072         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3073
3074         r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3075                        bio, &ool.locker, &lookup_result);
3076         if (r == -EWOULDBLOCK) {
3077                 cell_defer(cache, cell, true);
3078                 return DM_MAPIO_SUBMITTED;
3079
3080         } else if (r) {
3081                 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3082                             cache_device_name(cache), r);
3083                 cell_defer(cache, cell, false);
3084                 bio_io_error(bio);
3085                 return DM_MAPIO_SUBMITTED;
3086         }
3087
3088         r = DM_MAPIO_REMAPPED;
3089         switch (lookup_result.op) {
3090         case POLICY_HIT:
3091                 if (passthrough_mode(&cache->features)) {
3092                         if (bio_data_dir(bio) == WRITE) {
3093                                 /*
3094                                  * We need to invalidate this block, so
3095                                  * defer for the worker thread.
3096                                  */
3097                                 cell_defer(cache, cell, true);
3098                                 r = DM_MAPIO_SUBMITTED;
3099
3100                         } else {
3101                                 inc_miss_counter(cache, bio);
3102                                 remap_to_origin_clear_discard(cache, bio, block);
3103                                 accounted_begin(cache, bio);
3104                                 inc_ds(cache, bio, cell);
3105                                 // FIXME: we want to remap hits or misses straight
3106                                 // away rather than passing over to the worker.
3107                                 cell_defer(cache, cell, false);
3108                         }
3109
3110                 } else {
3111                         inc_hit_counter(cache, bio);
3112                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3113                             !is_dirty(cache, lookup_result.cblock)) {
3114                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3115                                 accounted_begin(cache, bio);
3116                                 inc_ds(cache, bio, cell);
3117                                 cell_defer(cache, cell, false);
3118
3119                         } else
3120                                 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3121                 }
3122                 break;
3123
3124         case POLICY_MISS:
3125                 inc_miss_counter(cache, bio);
3126                 if (pb->req_nr != 0) {
3127                         /*
3128                          * This is a duplicate writethrough io that is no
3129                          * longer needed because the block has been demoted.
3130                          */
3131                         bio_endio(bio, 0);
3132                         // FIXME: remap everything as a miss
3133                         cell_defer(cache, cell, false);
3134                         r = DM_MAPIO_SUBMITTED;
3135
3136                 } else
3137                         remap_cell_to_origin_clear_discard(cache, cell, block, false);
3138                 break;
3139
3140         default:
3141                 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3142                             cache_device_name(cache), __func__,
3143                             (unsigned) lookup_result.op);
3144                 cell_defer(cache, cell, false);
3145                 bio_io_error(bio);
3146                 r = DM_MAPIO_SUBMITTED;
3147         }
3148
3149         return r;
3150 }
3151
3152 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3153 {
3154         struct cache *cache = ti->private;
3155         unsigned long flags;
3156         size_t pb_data_size = get_per_bio_data_size(cache);
3157         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3158
3159         if (pb->tick) {
3160                 policy_tick(cache->policy, false);
3161
3162                 spin_lock_irqsave(&cache->lock, flags);
3163                 cache->need_tick_bio = true;
3164                 spin_unlock_irqrestore(&cache->lock, flags);
3165         }
3166
3167         check_for_quiesced_migrations(cache, pb);
3168         accounted_complete(cache, bio);
3169
3170         return 0;
3171 }
3172
3173 static int write_dirty_bitset(struct cache *cache)
3174 {
3175         unsigned i, r;
3176
3177         if (get_cache_mode(cache) >= CM_READ_ONLY)
3178                 return -EINVAL;
3179
3180         for (i = 0; i < from_cblock(cache->cache_size); i++) {
3181                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3182                                        is_dirty(cache, to_cblock(i)));
3183                 if (r) {
3184                         metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3185                         return r;
3186                 }
3187         }
3188
3189         return 0;
3190 }
3191
3192 static int write_discard_bitset(struct cache *cache)
3193 {
3194         unsigned i, r;
3195
3196         if (get_cache_mode(cache) >= CM_READ_ONLY)
3197                 return -EINVAL;
3198
3199         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3200                                            cache->discard_nr_blocks);
3201         if (r) {
3202                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3203                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3204                 return r;
3205         }
3206
3207         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3208                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3209                                          is_discarded(cache, to_dblock(i)));
3210                 if (r) {
3211                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
3212                         return r;
3213                 }
3214         }
3215
3216         return 0;
3217 }
3218
3219 static int write_hints(struct cache *cache)
3220 {
3221         int r;
3222
3223         if (get_cache_mode(cache) >= CM_READ_ONLY)
3224                 return -EINVAL;
3225
3226         r = dm_cache_write_hints(cache->cmd, cache->policy);
3227         if (r) {
3228                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3229                 return r;
3230         }
3231
3232         return 0;
3233 }
3234
3235 /*
3236  * returns true on success
3237  */
3238 static bool sync_metadata(struct cache *cache)
3239 {
3240         int r1, r2, r3, r4;
3241
3242         r1 = write_dirty_bitset(cache);
3243         if (r1)
3244                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3245
3246         r2 = write_discard_bitset(cache);
3247         if (r2)
3248                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3249
3250         save_stats(cache);
3251
3252         r3 = write_hints(cache);
3253         if (r3)
3254                 DMERR("%s: could not write hints", cache_device_name(cache));
3255
3256         /*
3257          * If writing the above metadata failed, we still commit, but don't
3258          * set the clean shutdown flag.  This will effectively force every
3259          * dirty bit to be set on reload.
3260          */
3261         r4 = commit(cache, !r1 && !r2 && !r3);
3262         if (r4)
3263                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3264
3265         return !r1 && !r2 && !r3 && !r4;
3266 }
3267
3268 static void cache_postsuspend(struct dm_target *ti)
3269 {
3270         struct cache *cache = ti->private;
3271
3272         start_quiescing(cache);
3273         wait_for_migrations(cache);
3274         stop_worker(cache);
3275         requeue_deferred_bios(cache);
3276         requeue_deferred_cells(cache);
3277         stop_quiescing(cache);
3278
3279         if (get_cache_mode(cache) == CM_WRITE)
3280                 (void) sync_metadata(cache);
3281 }
3282
3283 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3284                         bool dirty, uint32_t hint, bool hint_valid)
3285 {
3286         int r;
3287         struct cache *cache = context;
3288
3289         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3290         if (r)
3291                 return r;
3292
3293         if (dirty)
3294                 set_dirty(cache, oblock, cblock);
3295         else
3296                 clear_dirty(cache, oblock, cblock);
3297
3298         return 0;
3299 }
3300
3301 /*
3302  * The discard block size in the on disk metadata is not
3303  * neccessarily the same as we're currently using.  So we have to
3304  * be careful to only set the discarded attribute if we know it
3305  * covers a complete block of the new size.
3306  */
3307 struct discard_load_info {
3308         struct cache *cache;
3309
3310         /*
3311          * These blocks are sized using the on disk dblock size, rather
3312          * than the current one.
3313          */
3314         dm_block_t block_size;
3315         dm_block_t discard_begin, discard_end;
3316 };
3317
3318 static void discard_load_info_init(struct cache *cache,
3319                                    struct discard_load_info *li)
3320 {
3321         li->cache = cache;
3322         li->discard_begin = li->discard_end = 0;
3323 }
3324
3325 static void set_discard_range(struct discard_load_info *li)
3326 {
3327         sector_t b, e;
3328
3329         if (li->discard_begin == li->discard_end)
3330                 return;
3331
3332         /*
3333          * Convert to sectors.
3334          */
3335         b = li->discard_begin * li->block_size;
3336         e = li->discard_end * li->block_size;
3337
3338         /*
3339          * Then convert back to the current dblock size.
3340          */
3341         b = dm_sector_div_up(b, li->cache->discard_block_size);
3342         sector_div(e, li->cache->discard_block_size);
3343
3344         /*
3345          * The origin may have shrunk, so we need to check we're still in
3346          * bounds.
3347          */
3348         if (e > from_dblock(li->cache->discard_nr_blocks))
3349                 e = from_dblock(li->cache->discard_nr_blocks);
3350
3351         for (; b < e; b++)
3352                 set_discard(li->cache, to_dblock(b));
3353 }
3354
3355 static int load_discard(void *context, sector_t discard_block_size,
3356                         dm_dblock_t dblock, bool discard)
3357 {
3358         struct discard_load_info *li = context;
3359
3360         li->block_size = discard_block_size;
3361
3362         if (discard) {
3363                 if (from_dblock(dblock) == li->discard_end)
3364                         /*
3365                          * We're already in a discard range, just extend it.
3366                          */
3367                         li->discard_end = li->discard_end + 1ULL;
3368
3369                 else {
3370                         /*
3371                          * Emit the old range and start a new one.
3372                          */
3373                         set_discard_range(li);
3374                         li->discard_begin = from_dblock(dblock);
3375                         li->discard_end = li->discard_begin + 1ULL;
3376                 }
3377         } else {
3378                 set_discard_range(li);
3379                 li->discard_begin = li->discard_end = 0;
3380         }
3381
3382         return 0;
3383 }
3384
3385 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3386 {
3387         sector_t size = get_dev_size(cache->cache_dev);
3388         (void) sector_div(size, cache->sectors_per_block);
3389         return to_cblock(size);
3390 }
3391
3392 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3393 {
3394         if (from_cblock(new_size) > from_cblock(cache->cache_size))
3395                 return true;
3396
3397         /*
3398          * We can't drop a dirty block when shrinking the cache.
3399          */
3400         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3401                 new_size = to_cblock(from_cblock(new_size) + 1);
3402                 if (is_dirty(cache, new_size)) {
3403                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3404                               cache_device_name(cache),
3405                               (unsigned long long) from_cblock(new_size));
3406                         return false;
3407                 }
3408         }
3409
3410         return true;
3411 }
3412
3413 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3414 {
3415         int r;
3416
3417         r = dm_cache_resize(cache->cmd, new_size);
3418         if (r) {
3419                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3420                 metadata_operation_failed(cache, "dm_cache_resize", r);
3421                 return r;
3422         }
3423
3424         set_cache_size(cache, new_size);
3425
3426         return 0;
3427 }
3428
3429 static int cache_preresume(struct dm_target *ti)
3430 {
3431         int r = 0;
3432         struct cache *cache = ti->private;
3433         dm_cblock_t csize = get_cache_dev_size(cache);
3434
3435         /*
3436          * Check to see if the cache has resized.
3437          */
3438         if (!cache->sized) {
3439                 r = resize_cache_dev(cache, csize);
3440                 if (r)
3441                         return r;
3442
3443                 cache->sized = true;
3444
3445         } else if (csize != cache->cache_size) {
3446                 if (!can_resize(cache, csize))
3447                         return -EINVAL;
3448
3449                 r = resize_cache_dev(cache, csize);
3450                 if (r)
3451                         return r;
3452         }
3453
3454         if (!cache->loaded_mappings) {
3455                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3456                                            load_mapping, cache);
3457                 if (r) {
3458                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3459                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3460                         return r;
3461                 }
3462
3463                 cache->loaded_mappings = true;
3464         }
3465
3466         if (!cache->loaded_discards) {
3467                 struct discard_load_info li;
3468
3469                 /*
3470                  * The discard bitset could have been resized, or the
3471                  * discard block size changed.  To be safe we start by
3472                  * setting every dblock to not discarded.
3473                  */
3474                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3475
3476                 discard_load_info_init(cache, &li);
3477                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3478                 if (r) {
3479                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3480                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3481                         return r;
3482                 }
3483                 set_discard_range(&li);
3484
3485                 cache->loaded_discards = true;
3486         }
3487
3488         return r;
3489 }
3490
3491 static void cache_resume(struct dm_target *ti)
3492 {
3493         struct cache *cache = ti->private;
3494
3495         cache->need_tick_bio = true;
3496         do_waker(&cache->waker.work);
3497 }
3498
3499 /*
3500  * Status format:
3501  *
3502  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3503  * <cache block size> <#used cache blocks>/<#total cache blocks>
3504  * <#read hits> <#read misses> <#write hits> <#write misses>
3505  * <#demotions> <#promotions> <#dirty>
3506  * <#features> <features>*
3507  * <#core args> <core args>
3508  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3509  */
3510 static void cache_status(struct dm_target *ti, status_type_t type,
3511                          unsigned status_flags, char *result, unsigned maxlen)
3512 {
3513         int r = 0;
3514         unsigned i;
3515         ssize_t sz = 0;
3516         dm_block_t nr_free_blocks_metadata = 0;
3517         dm_block_t nr_blocks_metadata = 0;
3518         char buf[BDEVNAME_SIZE];
3519         struct cache *cache = ti->private;
3520         dm_cblock_t residency;
3521
3522         switch (type) {
3523         case STATUSTYPE_INFO:
3524                 if (get_cache_mode(cache) == CM_FAIL) {
3525                         DMEMIT("Fail");
3526                         break;
3527                 }
3528
3529                 /* Commit to ensure statistics aren't out-of-date */
3530                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3531                         (void) commit(cache, false);
3532
3533                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3534                 if (r) {
3535                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3536                               cache_device_name(cache), r);
3537                         goto err;
3538                 }
3539
3540                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3541                 if (r) {
3542                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3543                               cache_device_name(cache), r);
3544                         goto err;
3545                 }
3546
3547                 residency = policy_residency(cache->policy);
3548
3549                 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3550                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3551                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3552                        (unsigned long long)nr_blocks_metadata,
3553                        cache->sectors_per_block,
3554                        (unsigned long long) from_cblock(residency),
3555                        (unsigned long long) from_cblock(cache->cache_size),
3556                        (unsigned) atomic_read(&cache->stats.read_hit),
3557                        (unsigned) atomic_read(&cache->stats.read_miss),
3558                        (unsigned) atomic_read(&cache->stats.write_hit),
3559                        (unsigned) atomic_read(&cache->stats.write_miss),
3560                        (unsigned) atomic_read(&cache->stats.demotion),
3561                        (unsigned) atomic_read(&cache->stats.promotion),
3562                        (unsigned long) atomic_read(&cache->nr_dirty));
3563
3564                 if (writethrough_mode(&cache->features))
3565                         DMEMIT("1 writethrough ");
3566
3567                 else if (passthrough_mode(&cache->features))
3568                         DMEMIT("1 passthrough ");
3569
3570                 else if (writeback_mode(&cache->features))
3571                         DMEMIT("1 writeback ");
3572
3573                 else {
3574                         DMERR("%s: internal error: unknown io mode: %d",
3575                               cache_device_name(cache), (int) cache->features.io_mode);
3576                         goto err;
3577                 }
3578
3579                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3580
3581                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3582                 if (sz < maxlen) {
3583                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3584                         if (r)
3585                                 DMERR("%s: policy_emit_config_values returned %d",
3586                                       cache_device_name(cache), r);
3587                 }
3588
3589                 if (get_cache_mode(cache) == CM_READ_ONLY)
3590                         DMEMIT("ro ");
3591                 else
3592                         DMEMIT("rw ");
3593
3594                 if (dm_cache_metadata_needs_check(cache->cmd))
3595                         DMEMIT("needs_check ");
3596                 else
3597                         DMEMIT("- ");
3598
3599                 break;
3600
3601         case STATUSTYPE_TABLE:
3602                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3603                 DMEMIT("%s ", buf);
3604                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3605                 DMEMIT("%s ", buf);
3606                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3607                 DMEMIT("%s", buf);
3608
3609                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3610                         DMEMIT(" %s", cache->ctr_args[i]);
3611                 if (cache->nr_ctr_args)
3612                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3613         }
3614
3615         return;
3616
3617 err:
3618         DMEMIT("Error");
3619 }
3620
3621 /*
3622  * A cache block range can take two forms:
3623  *
3624  * i) A single cblock, eg. '3456'
3625  * ii) A begin and end cblock with dots between, eg. 123-234
3626  */
3627 static int parse_cblock_range(struct cache *cache, const char *str,
3628                               struct cblock_range *result)
3629 {
3630         char dummy;
3631         uint64_t b, e;
3632         int r;
3633
3634         /*
3635          * Try and parse form (ii) first.
3636          */
3637         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3638         if (r < 0)
3639                 return r;
3640
3641         if (r == 2) {
3642                 result->begin = to_cblock(b);
3643                 result->end = to_cblock(e);
3644                 return 0;
3645         }
3646
3647         /*
3648          * That didn't work, try form (i).
3649          */
3650         r = sscanf(str, "%llu%c", &b, &dummy);
3651         if (r < 0)
3652                 return r;
3653
3654         if (r == 1) {
3655                 result->begin = to_cblock(b);
3656                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3657                 return 0;
3658         }
3659
3660         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3661         return -EINVAL;
3662 }
3663
3664 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3665 {
3666         uint64_t b = from_cblock(range->begin);
3667         uint64_t e = from_cblock(range->end);
3668         uint64_t n = from_cblock(cache->cache_size);
3669
3670         if (b >= n) {
3671                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3672                       cache_device_name(cache), b, n);
3673                 return -EINVAL;
3674         }
3675
3676         if (e > n) {
3677                 DMERR("%s: end cblock out of range: %llu > %llu",
3678                       cache_device_name(cache), e, n);
3679                 return -EINVAL;
3680         }
3681
3682         if (b >= e) {
3683                 DMERR("%s: invalid cblock range: %llu >= %llu",
3684                       cache_device_name(cache), b, e);
3685                 return -EINVAL;
3686         }
3687
3688         return 0;
3689 }
3690
3691 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3692 {
3693         struct invalidation_request req;
3694
3695         INIT_LIST_HEAD(&req.list);
3696         req.cblocks = range;
3697         atomic_set(&req.complete, 0);
3698         req.err = 0;
3699         init_waitqueue_head(&req.result_wait);
3700
3701         spin_lock(&cache->invalidation_lock);
3702         list_add(&req.list, &cache->invalidation_requests);
3703         spin_unlock(&cache->invalidation_lock);
3704         wake_worker(cache);
3705
3706         wait_event(req.result_wait, atomic_read(&req.complete));
3707         return req.err;
3708 }
3709
3710 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3711                                               const char **cblock_ranges)
3712 {
3713         int r = 0;
3714         unsigned i;
3715         struct cblock_range range;
3716
3717         if (!passthrough_mode(&cache->features)) {
3718                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3719                       cache_device_name(cache));
3720                 return -EPERM;
3721         }
3722
3723         for (i = 0; i < count; i++) {
3724                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3725                 if (r)
3726                         break;
3727
3728                 r = validate_cblock_range(cache, &range);
3729                 if (r)
3730                         break;
3731
3732                 /*
3733                  * Pass begin and end origin blocks to the worker and wake it.
3734                  */
3735                 r = request_invalidation(cache, &range);
3736                 if (r)
3737                         break;
3738         }
3739
3740         return r;
3741 }
3742
3743 /*
3744  * Supports
3745  *      "<key> <value>"
3746  * and
3747  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3748  *
3749  * The key migration_threshold is supported by the cache target core.
3750  */
3751 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3752 {
3753         struct cache *cache = ti->private;
3754
3755         if (!argc)
3756                 return -EINVAL;
3757
3758         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3759                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3760                       cache_device_name(cache));
3761                 return -EOPNOTSUPP;
3762         }
3763
3764         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3765                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3766
3767         if (argc != 2)
3768                 return -EINVAL;
3769
3770         return set_config_value(cache, argv[0], argv[1]);
3771 }
3772
3773 static int cache_iterate_devices(struct dm_target *ti,
3774                                  iterate_devices_callout_fn fn, void *data)
3775 {
3776         int r = 0;
3777         struct cache *cache = ti->private;
3778
3779         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3780         if (!r)
3781                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3782
3783         return r;
3784 }
3785
3786 /*
3787  * We assume I/O is going to the origin (which is the volume
3788  * more likely to have restrictions e.g. by being striped).
3789  * (Looking up the exact location of the data would be expensive
3790  * and could always be out of date by the time the bio is submitted.)
3791  */
3792 static int cache_bvec_merge(struct dm_target *ti,
3793                             struct bvec_merge_data *bvm,
3794                             struct bio_vec *biovec, int max_size)
3795 {
3796         struct cache *cache = ti->private;
3797         struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3798
3799         if (!q->merge_bvec_fn)
3800                 return max_size;
3801
3802         bvm->bi_bdev = cache->origin_dev->bdev;
3803         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3804 }
3805
3806 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3807 {
3808         /*
3809          * FIXME: these limits may be incompatible with the cache device
3810          */
3811         limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3812                                             cache->origin_sectors);
3813         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3814 }
3815
3816 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3817 {
3818         struct cache *cache = ti->private;
3819         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3820
3821         /*
3822          * If the system-determined stacked limits are compatible with the
3823          * cache's blocksize (io_opt is a factor) do not override them.
3824          */
3825         if (io_opt_sectors < cache->sectors_per_block ||
3826             do_div(io_opt_sectors, cache->sectors_per_block)) {
3827                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3828                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3829         }
3830         set_discard_limits(cache, limits);
3831 }
3832
3833 /*----------------------------------------------------------------*/
3834
3835 static struct target_type cache_target = {
3836         .name = "cache",
3837         .version = {1, 8, 0},
3838         .module = THIS_MODULE,
3839         .ctr = cache_ctr,
3840         .dtr = cache_dtr,
3841         .map = cache_map,
3842         .end_io = cache_end_io,
3843         .postsuspend = cache_postsuspend,
3844         .preresume = cache_preresume,
3845         .resume = cache_resume,
3846         .status = cache_status,
3847         .message = cache_message,
3848         .iterate_devices = cache_iterate_devices,
3849         .merge = cache_bvec_merge,
3850         .io_hints = cache_io_hints,
3851 };
3852
3853 static int __init dm_cache_init(void)
3854 {
3855         int r;
3856
3857         r = dm_register_target(&cache_target);
3858         if (r) {
3859                 DMERR("cache target registration failed: %d", r);
3860                 return r;
3861         }
3862
3863         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3864         if (!migration_cache) {
3865                 dm_unregister_target(&cache_target);
3866                 return -ENOMEM;
3867         }
3868
3869         return 0;
3870 }
3871
3872 static void __exit dm_cache_exit(void)
3873 {
3874         dm_unregister_target(&cache_target);
3875         kmem_cache_destroy(migration_cache);
3876 }
3877
3878 module_init(dm_cache_init);
3879 module_exit(dm_cache_exit);
3880
3881 MODULE_DESCRIPTION(DM_NAME " cache target");
3882 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3883 MODULE_LICENSE("GPL");