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 }
1128
1129 static void migration_failure(struct dm_cache_migration *mg)
1130 {
1131         struct cache *cache = mg->cache;
1132         const char *dev_name = cache_device_name(cache);
1133
1134         if (mg->writeback) {
1135                 DMERR_LIMIT("%s: writeback failed; couldn't copy block", dev_name);
1136                 set_dirty(cache, mg->old_oblock, mg->cblock);
1137                 cell_defer(cache, mg->old_ocell, false);
1138
1139         } else if (mg->demote) {
1140                 DMERR_LIMIT("%s: demotion failed; couldn't copy block", dev_name);
1141                 policy_force_mapping(cache->policy, mg->new_oblock, mg->old_oblock);
1142
1143                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1144                 if (mg->promote)
1145                         cell_defer(cache, mg->new_ocell, true);
1146         } else {
1147                 DMERR_LIMIT("%s: promotion failed; couldn't copy block", dev_name);
1148                 policy_remove_mapping(cache->policy, mg->new_oblock);
1149                 cell_defer(cache, mg->new_ocell, true);
1150         }
1151
1152         free_io_migration(mg);
1153 }
1154
1155 static void migration_success_pre_commit(struct dm_cache_migration *mg)
1156 {
1157         int r;
1158         unsigned long flags;
1159         struct cache *cache = mg->cache;
1160
1161         if (mg->writeback) {
1162                 clear_dirty(cache, mg->old_oblock, mg->cblock);
1163                 cell_defer(cache, mg->old_ocell, false);
1164                 free_io_migration(mg);
1165                 return;
1166
1167         } else if (mg->demote) {
1168                 r = dm_cache_remove_mapping(cache->cmd, mg->cblock);
1169                 if (r) {
1170                         DMERR_LIMIT("%s: demotion failed; couldn't update on disk metadata",
1171                                     cache_device_name(cache));
1172                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1173                         policy_force_mapping(cache->policy, mg->new_oblock,
1174                                              mg->old_oblock);
1175                         if (mg->promote)
1176                                 cell_defer(cache, mg->new_ocell, true);
1177                         free_io_migration(mg);
1178                         return;
1179                 }
1180         } else {
1181                 r = dm_cache_insert_mapping(cache->cmd, mg->cblock, mg->new_oblock);
1182                 if (r) {
1183                         DMERR_LIMIT("%s: promotion failed; couldn't update on disk metadata",
1184                                     cache_device_name(cache));
1185                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1186                         policy_remove_mapping(cache->policy, mg->new_oblock);
1187                         free_io_migration(mg);
1188                         return;
1189                 }
1190         }
1191
1192         spin_lock_irqsave(&cache->lock, flags);
1193         list_add_tail(&mg->list, &cache->need_commit_migrations);
1194         cache->commit_requested = true;
1195         spin_unlock_irqrestore(&cache->lock, flags);
1196 }
1197
1198 static void migration_success_post_commit(struct dm_cache_migration *mg)
1199 {
1200         unsigned long flags;
1201         struct cache *cache = mg->cache;
1202
1203         if (mg->writeback) {
1204                 DMWARN_LIMIT("%s: writeback unexpectedly triggered commit",
1205                              cache_device_name(cache));
1206                 return;
1207
1208         } else if (mg->demote) {
1209                 cell_defer(cache, mg->old_ocell, mg->promote ? false : true);
1210
1211                 if (mg->promote) {
1212                         mg->demote = false;
1213
1214                         spin_lock_irqsave(&cache->lock, flags);
1215                         list_add_tail(&mg->list, &cache->quiesced_migrations);
1216                         spin_unlock_irqrestore(&cache->lock, flags);
1217
1218                 } else {
1219                         if (mg->invalidate)
1220                                 policy_remove_mapping(cache->policy, mg->old_oblock);
1221                         free_io_migration(mg);
1222                 }
1223
1224         } else {
1225                 if (mg->requeue_holder) {
1226                         clear_dirty(cache, mg->new_oblock, mg->cblock);
1227                         cell_defer(cache, mg->new_ocell, true);
1228                 } else {
1229                         /*
1230                          * The block was promoted via an overwrite, so it's dirty.
1231                          */
1232                         set_dirty(cache, mg->new_oblock, mg->cblock);
1233                         bio_endio(mg->new_ocell->holder, 0);
1234                         cell_defer(cache, mg->new_ocell, false);
1235                 }
1236                 free_io_migration(mg);
1237         }
1238 }
1239
1240 static void copy_complete(int read_err, unsigned long write_err, void *context)
1241 {
1242         unsigned long flags;
1243         struct dm_cache_migration *mg = (struct dm_cache_migration *) context;
1244         struct cache *cache = mg->cache;
1245
1246         if (read_err || write_err)
1247                 mg->err = true;
1248
1249         spin_lock_irqsave(&cache->lock, flags);
1250         list_add_tail(&mg->list, &cache->completed_migrations);
1251         spin_unlock_irqrestore(&cache->lock, flags);
1252
1253         wake_worker(cache);
1254 }
1255
1256 static void issue_copy(struct dm_cache_migration *mg)
1257 {
1258         int r;
1259         struct dm_io_region o_region, c_region;
1260         struct cache *cache = mg->cache;
1261         sector_t cblock = from_cblock(mg->cblock);
1262
1263         o_region.bdev = cache->origin_dev->bdev;
1264         o_region.count = cache->sectors_per_block;
1265
1266         c_region.bdev = cache->cache_dev->bdev;
1267         c_region.sector = cblock * cache->sectors_per_block;
1268         c_region.count = cache->sectors_per_block;
1269
1270         if (mg->writeback || mg->demote) {
1271                 /* demote */
1272                 o_region.sector = from_oblock(mg->old_oblock) * cache->sectors_per_block;
1273                 r = dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, mg);
1274         } else {
1275                 /* promote */
1276                 o_region.sector = from_oblock(mg->new_oblock) * cache->sectors_per_block;
1277                 r = dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, mg);
1278         }
1279
1280         if (r < 0) {
1281                 DMERR_LIMIT("%s: issuing migration failed", cache_device_name(cache));
1282                 migration_failure(mg);
1283         }
1284 }
1285
1286 static void overwrite_endio(struct bio *bio, int err)
1287 {
1288         struct dm_cache_migration *mg = bio->bi_private;
1289         struct cache *cache = mg->cache;
1290         size_t pb_data_size = get_per_bio_data_size(cache);
1291         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1292         unsigned long flags;
1293
1294         dm_unhook_bio(&pb->hook_info, bio);
1295
1296         if (err)
1297                 mg->err = true;
1298
1299         mg->requeue_holder = false;
1300
1301         spin_lock_irqsave(&cache->lock, flags);
1302         list_add_tail(&mg->list, &cache->completed_migrations);
1303         spin_unlock_irqrestore(&cache->lock, flags);
1304
1305         wake_worker(cache);
1306 }
1307
1308 static void issue_overwrite(struct dm_cache_migration *mg, struct bio *bio)
1309 {
1310         size_t pb_data_size = get_per_bio_data_size(mg->cache);
1311         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1312
1313         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1314         remap_to_cache_dirty(mg->cache, bio, mg->new_oblock, mg->cblock);
1315
1316         /*
1317          * No need to inc_ds() here, since the cell will be held for the
1318          * duration of the io.
1319          */
1320         accounted_request(mg->cache, bio);
1321 }
1322
1323 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1324 {
1325         return (bio_data_dir(bio) == WRITE) &&
1326                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1327 }
1328
1329 static void avoid_copy(struct dm_cache_migration *mg)
1330 {
1331         atomic_inc(&mg->cache->stats.copies_avoided);
1332         migration_success_pre_commit(mg);
1333 }
1334
1335 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1336                                      dm_dblock_t *b, dm_dblock_t *e)
1337 {
1338         sector_t sb = bio->bi_iter.bi_sector;
1339         sector_t se = bio_end_sector(bio);
1340
1341         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1342
1343         if (se - sb < cache->discard_block_size)
1344                 *e = *b;
1345         else
1346                 *e = to_dblock(block_div(se, cache->discard_block_size));
1347 }
1348
1349 static void issue_discard(struct dm_cache_migration *mg)
1350 {
1351         dm_dblock_t b, e;
1352         struct bio *bio = mg->new_ocell->holder;
1353
1354         calc_discard_block_range(mg->cache, bio, &b, &e);
1355         while (b != e) {
1356                 set_discard(mg->cache, b);
1357                 b = to_dblock(from_dblock(b) + 1);
1358         }
1359
1360         bio_endio(bio, 0);
1361         cell_defer(mg->cache, mg->new_ocell, false);
1362         free_migration(mg);
1363 }
1364
1365 static void issue_copy_or_discard(struct dm_cache_migration *mg)
1366 {
1367         bool avoid;
1368         struct cache *cache = mg->cache;
1369
1370         if (mg->discard) {
1371                 issue_discard(mg);
1372                 return;
1373         }
1374
1375         if (mg->writeback || mg->demote)
1376                 avoid = !is_dirty(cache, mg->cblock) ||
1377                         is_discarded_oblock(cache, mg->old_oblock);
1378         else {
1379                 struct bio *bio = mg->new_ocell->holder;
1380
1381                 avoid = is_discarded_oblock(cache, mg->new_oblock);
1382
1383                 if (writeback_mode(&cache->features) &&
1384                     !avoid && bio_writes_complete_block(cache, bio)) {
1385                         issue_overwrite(mg, bio);
1386                         return;
1387                 }
1388         }
1389
1390         avoid ? avoid_copy(mg) : issue_copy(mg);
1391 }
1392
1393 static void complete_migration(struct dm_cache_migration *mg)
1394 {
1395         if (mg->err)
1396                 migration_failure(mg);
1397         else
1398                 migration_success_pre_commit(mg);
1399 }
1400
1401 static void process_migrations(struct cache *cache, struct list_head *head,
1402                                void (*fn)(struct dm_cache_migration *))
1403 {
1404         unsigned long flags;
1405         struct list_head list;
1406         struct dm_cache_migration *mg, *tmp;
1407
1408         INIT_LIST_HEAD(&list);
1409         spin_lock_irqsave(&cache->lock, flags);
1410         list_splice_init(head, &list);
1411         spin_unlock_irqrestore(&cache->lock, flags);
1412
1413         list_for_each_entry_safe(mg, tmp, &list, list)
1414                 fn(mg);
1415 }
1416
1417 static void __queue_quiesced_migration(struct dm_cache_migration *mg)
1418 {
1419         list_add_tail(&mg->list, &mg->cache->quiesced_migrations);
1420 }
1421
1422 static void queue_quiesced_migration(struct dm_cache_migration *mg)
1423 {
1424         unsigned long flags;
1425         struct cache *cache = mg->cache;
1426
1427         spin_lock_irqsave(&cache->lock, flags);
1428         __queue_quiesced_migration(mg);
1429         spin_unlock_irqrestore(&cache->lock, flags);
1430
1431         wake_worker(cache);
1432 }
1433
1434 static void queue_quiesced_migrations(struct cache *cache, struct list_head *work)
1435 {
1436         unsigned long flags;
1437         struct dm_cache_migration *mg, *tmp;
1438
1439         spin_lock_irqsave(&cache->lock, flags);
1440         list_for_each_entry_safe(mg, tmp, work, list)
1441                 __queue_quiesced_migration(mg);
1442         spin_unlock_irqrestore(&cache->lock, flags);
1443
1444         wake_worker(cache);
1445 }
1446
1447 static void check_for_quiesced_migrations(struct cache *cache,
1448                                           struct per_bio_data *pb)
1449 {
1450         struct list_head work;
1451
1452         if (!pb->all_io_entry)
1453                 return;
1454
1455         INIT_LIST_HEAD(&work);
1456         dm_deferred_entry_dec(pb->all_io_entry, &work);
1457
1458         if (!list_empty(&work))
1459                 queue_quiesced_migrations(cache, &work);
1460 }
1461
1462 static void quiesce_migration(struct dm_cache_migration *mg)
1463 {
1464         if (!dm_deferred_set_add_work(mg->cache->all_io_ds, &mg->list))
1465                 queue_quiesced_migration(mg);
1466 }
1467
1468 static void promote(struct cache *cache, struct prealloc *structs,
1469                     dm_oblock_t oblock, dm_cblock_t cblock,
1470                     struct dm_bio_prison_cell *cell)
1471 {
1472         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1473
1474         mg->err = false;
1475         mg->discard = false;
1476         mg->writeback = false;
1477         mg->demote = false;
1478         mg->promote = true;
1479         mg->requeue_holder = true;
1480         mg->invalidate = false;
1481         mg->cache = cache;
1482         mg->new_oblock = oblock;
1483         mg->cblock = cblock;
1484         mg->old_ocell = NULL;
1485         mg->new_ocell = cell;
1486         mg->start_jiffies = jiffies;
1487
1488         inc_io_migrations(cache);
1489         quiesce_migration(mg);
1490 }
1491
1492 static void writeback(struct cache *cache, struct prealloc *structs,
1493                       dm_oblock_t oblock, dm_cblock_t cblock,
1494                       struct dm_bio_prison_cell *cell)
1495 {
1496         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1497
1498         mg->err = false;
1499         mg->discard = false;
1500         mg->writeback = true;
1501         mg->demote = false;
1502         mg->promote = false;
1503         mg->requeue_holder = true;
1504         mg->invalidate = false;
1505         mg->cache = cache;
1506         mg->old_oblock = oblock;
1507         mg->cblock = cblock;
1508         mg->old_ocell = cell;
1509         mg->new_ocell = NULL;
1510         mg->start_jiffies = jiffies;
1511
1512         inc_io_migrations(cache);
1513         quiesce_migration(mg);
1514 }
1515
1516 static void demote_then_promote(struct cache *cache, struct prealloc *structs,
1517                                 dm_oblock_t old_oblock, dm_oblock_t new_oblock,
1518                                 dm_cblock_t cblock,
1519                                 struct dm_bio_prison_cell *old_ocell,
1520                                 struct dm_bio_prison_cell *new_ocell)
1521 {
1522         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1523
1524         mg->err = false;
1525         mg->discard = false;
1526         mg->writeback = false;
1527         mg->demote = true;
1528         mg->promote = true;
1529         mg->requeue_holder = true;
1530         mg->invalidate = false;
1531         mg->cache = cache;
1532         mg->old_oblock = old_oblock;
1533         mg->new_oblock = new_oblock;
1534         mg->cblock = cblock;
1535         mg->old_ocell = old_ocell;
1536         mg->new_ocell = new_ocell;
1537         mg->start_jiffies = jiffies;
1538
1539         inc_io_migrations(cache);
1540         quiesce_migration(mg);
1541 }
1542
1543 /*
1544  * Invalidate a cache entry.  No writeback occurs; any changes in the cache
1545  * block are thrown away.
1546  */
1547 static void invalidate(struct cache *cache, struct prealloc *structs,
1548                        dm_oblock_t oblock, dm_cblock_t cblock,
1549                        struct dm_bio_prison_cell *cell)
1550 {
1551         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1552
1553         mg->err = false;
1554         mg->discard = false;
1555         mg->writeback = false;
1556         mg->demote = true;
1557         mg->promote = false;
1558         mg->requeue_holder = true;
1559         mg->invalidate = true;
1560         mg->cache = cache;
1561         mg->old_oblock = oblock;
1562         mg->cblock = cblock;
1563         mg->old_ocell = cell;
1564         mg->new_ocell = NULL;
1565         mg->start_jiffies = jiffies;
1566
1567         inc_io_migrations(cache);
1568         quiesce_migration(mg);
1569 }
1570
1571 static void discard(struct cache *cache, struct prealloc *structs,
1572                     struct dm_bio_prison_cell *cell)
1573 {
1574         struct dm_cache_migration *mg = prealloc_get_migration(structs);
1575
1576         mg->err = false;
1577         mg->discard = true;
1578         mg->writeback = false;
1579         mg->demote = false;
1580         mg->promote = false;
1581         mg->requeue_holder = false;
1582         mg->invalidate = false;
1583         mg->cache = cache;
1584         mg->old_ocell = NULL;
1585         mg->new_ocell = cell;
1586         mg->start_jiffies = jiffies;
1587
1588         quiesce_migration(mg);
1589 }
1590
1591 /*----------------------------------------------------------------
1592  * bio processing
1593  *--------------------------------------------------------------*/
1594 static void defer_bio(struct cache *cache, struct bio *bio)
1595 {
1596         unsigned long flags;
1597
1598         spin_lock_irqsave(&cache->lock, flags);
1599         bio_list_add(&cache->deferred_bios, bio);
1600         spin_unlock_irqrestore(&cache->lock, flags);
1601
1602         wake_worker(cache);
1603 }
1604
1605 static void process_flush_bio(struct cache *cache, struct bio *bio)
1606 {
1607         size_t pb_data_size = get_per_bio_data_size(cache);
1608         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
1609
1610         BUG_ON(bio->bi_iter.bi_size);
1611         if (!pb->req_nr)
1612                 remap_to_origin(cache, bio);
1613         else
1614                 remap_to_cache(cache, bio, 0);
1615
1616         /*
1617          * REQ_FLUSH is not directed at any particular block so we don't
1618          * need to inc_ds().  REQ_FUA's are split into a write + REQ_FLUSH
1619          * by dm-core.
1620          */
1621         issue(cache, bio);
1622 }
1623
1624 static void process_discard_bio(struct cache *cache, struct prealloc *structs,
1625                                 struct bio *bio)
1626 {
1627         int r;
1628         dm_dblock_t b, e;
1629         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1630
1631         calc_discard_block_range(cache, bio, &b, &e);
1632         if (b == e) {
1633                 bio_endio(bio, 0);
1634                 return;
1635         }
1636
1637         cell_prealloc = prealloc_get_cell(structs);
1638         r = bio_detain_range(cache, dblock_to_oblock(cache, b), dblock_to_oblock(cache, e), bio, cell_prealloc,
1639                              (cell_free_fn) prealloc_put_cell,
1640                              structs, &new_ocell);
1641         if (r > 0)
1642                 return;
1643
1644         discard(cache, structs, new_ocell);
1645 }
1646
1647 static bool spare_migration_bandwidth(struct cache *cache)
1648 {
1649         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1650                 cache->sectors_per_block;
1651         return current_volume < cache->migration_threshold;
1652 }
1653
1654 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1655 {
1656         atomic_inc(bio_data_dir(bio) == READ ?
1657                    &cache->stats.read_hit : &cache->stats.write_hit);
1658 }
1659
1660 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1661 {
1662         atomic_inc(bio_data_dir(bio) == READ ?
1663                    &cache->stats.read_miss : &cache->stats.write_miss);
1664 }
1665
1666 /*----------------------------------------------------------------*/
1667
1668 struct inc_detail {
1669         struct cache *cache;
1670         struct bio_list bios_for_issue;
1671         struct bio_list unhandled_bios;
1672         bool any_writes;
1673 };
1674
1675 static void inc_fn(void *context, struct dm_bio_prison_cell *cell)
1676 {
1677         struct bio *bio;
1678         struct inc_detail *detail = context;
1679         struct cache *cache = detail->cache;
1680
1681         inc_ds(cache, cell->holder, cell);
1682         if (bio_data_dir(cell->holder) == WRITE)
1683                 detail->any_writes = true;
1684
1685         while ((bio = bio_list_pop(&cell->bios))) {
1686                 if (discard_or_flush(bio)) {
1687                         bio_list_add(&detail->unhandled_bios, bio);
1688                         continue;
1689                 }
1690
1691                 if (bio_data_dir(bio) == WRITE)
1692                         detail->any_writes = true;
1693
1694                 bio_list_add(&detail->bios_for_issue, bio);
1695                 inc_ds(cache, bio, cell);
1696         }
1697 }
1698
1699 // FIXME: refactor these two
1700 static void remap_cell_to_origin_clear_discard(struct cache *cache,
1701                                                struct dm_bio_prison_cell *cell,
1702                                                dm_oblock_t oblock, bool issue_holder)
1703 {
1704         struct bio *bio;
1705         unsigned long flags;
1706         struct inc_detail detail;
1707
1708         detail.cache = cache;
1709         bio_list_init(&detail.bios_for_issue);
1710         bio_list_init(&detail.unhandled_bios);
1711         detail.any_writes = false;
1712
1713         spin_lock_irqsave(&cache->lock, flags);
1714         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1715         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1716         spin_unlock_irqrestore(&cache->lock, flags);
1717
1718         remap_to_origin(cache, cell->holder);
1719         if (issue_holder)
1720                 issue(cache, cell->holder);
1721         else
1722                 accounted_begin(cache, cell->holder);
1723
1724         if (detail.any_writes)
1725                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1726
1727         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1728                 remap_to_origin(cache, bio);
1729                 issue(cache, bio);
1730         }
1731 }
1732
1733 static void remap_cell_to_cache_dirty(struct cache *cache, struct dm_bio_prison_cell *cell,
1734                                       dm_oblock_t oblock, dm_cblock_t cblock, bool issue_holder)
1735 {
1736         struct bio *bio;
1737         unsigned long flags;
1738         struct inc_detail detail;
1739
1740         detail.cache = cache;
1741         bio_list_init(&detail.bios_for_issue);
1742         bio_list_init(&detail.unhandled_bios);
1743         detail.any_writes = false;
1744
1745         spin_lock_irqsave(&cache->lock, flags);
1746         dm_cell_visit_release(cache->prison, inc_fn, &detail, cell);
1747         bio_list_merge(&cache->deferred_bios, &detail.unhandled_bios);
1748         spin_unlock_irqrestore(&cache->lock, flags);
1749
1750         remap_to_cache(cache, cell->holder, cblock);
1751         if (issue_holder)
1752                 issue(cache, cell->holder);
1753         else
1754                 accounted_begin(cache, cell->holder);
1755
1756         if (detail.any_writes) {
1757                 set_dirty(cache, oblock, cblock);
1758                 clear_discard(cache, oblock_to_dblock(cache, oblock));
1759         }
1760
1761         while ((bio = bio_list_pop(&detail.bios_for_issue))) {
1762                 remap_to_cache(cache, bio, cblock);
1763                 issue(cache, bio);
1764         }
1765 }
1766
1767 /*----------------------------------------------------------------*/
1768
1769 struct old_oblock_lock {
1770         struct policy_locker locker;
1771         struct cache *cache;
1772         struct prealloc *structs;
1773         struct dm_bio_prison_cell *cell;
1774 };
1775
1776 static int null_locker(struct policy_locker *locker, dm_oblock_t b)
1777 {
1778         /* This should never be called */
1779         BUG();
1780         return 0;
1781 }
1782
1783 static int cell_locker(struct policy_locker *locker, dm_oblock_t b)
1784 {
1785         struct old_oblock_lock *l = container_of(locker, struct old_oblock_lock, locker);
1786         struct dm_bio_prison_cell *cell_prealloc = prealloc_get_cell(l->structs);
1787
1788         return bio_detain(l->cache, b, NULL, cell_prealloc,
1789                           (cell_free_fn) prealloc_put_cell,
1790                           l->structs, &l->cell);
1791 }
1792
1793 static void process_cell(struct cache *cache, struct prealloc *structs,
1794                          struct dm_bio_prison_cell *new_ocell)
1795 {
1796         int r;
1797         bool release_cell = true;
1798         struct bio *bio = new_ocell->holder;
1799         dm_oblock_t block = get_bio_block(cache, bio);
1800         struct policy_result lookup_result;
1801         bool passthrough = passthrough_mode(&cache->features);
1802         bool fast_promotion, can_migrate;
1803         struct old_oblock_lock ool;
1804
1805         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
1806         can_migrate = !passthrough && (fast_promotion || spare_migration_bandwidth(cache));
1807
1808         ool.locker.fn = cell_locker;
1809         ool.cache = cache;
1810         ool.structs = structs;
1811         ool.cell = NULL;
1812         r = policy_map(cache->policy, block, true, can_migrate, fast_promotion,
1813                        bio, &ool.locker, &lookup_result);
1814
1815         if (r == -EWOULDBLOCK)
1816                 /* migration has been denied */
1817                 lookup_result.op = POLICY_MISS;
1818
1819         switch (lookup_result.op) {
1820         case POLICY_HIT:
1821                 if (passthrough) {
1822                         inc_miss_counter(cache, bio);
1823
1824                         /*
1825                          * Passthrough always maps to the origin,
1826                          * invalidating any cache blocks that are written
1827                          * to.
1828                          */
1829
1830                         if (bio_data_dir(bio) == WRITE) {
1831                                 atomic_inc(&cache->stats.demotion);
1832                                 invalidate(cache, structs, block, lookup_result.cblock, new_ocell);
1833                                 release_cell = false;
1834
1835                         } else {
1836                                 /* FIXME: factor out issue_origin() */
1837                                 remap_to_origin_clear_discard(cache, bio, block);
1838                                 inc_and_issue(cache, bio, new_ocell);
1839                         }
1840                 } else {
1841                         inc_hit_counter(cache, bio);
1842
1843                         if (bio_data_dir(bio) == WRITE &&
1844                             writethrough_mode(&cache->features) &&
1845                             !is_dirty(cache, lookup_result.cblock)) {
1846                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
1847                                 inc_and_issue(cache, bio, new_ocell);
1848
1849                         } else {
1850                                 remap_cell_to_cache_dirty(cache, new_ocell, block, lookup_result.cblock, true);
1851                                 release_cell = false;
1852                         }
1853                 }
1854
1855                 break;
1856
1857         case POLICY_MISS:
1858                 inc_miss_counter(cache, bio);
1859                 remap_cell_to_origin_clear_discard(cache, new_ocell, block, true);
1860                 release_cell = false;
1861                 break;
1862
1863         case POLICY_NEW:
1864                 atomic_inc(&cache->stats.promotion);
1865                 promote(cache, structs, block, lookup_result.cblock, new_ocell);
1866                 release_cell = false;
1867                 break;
1868
1869         case POLICY_REPLACE:
1870                 atomic_inc(&cache->stats.demotion);
1871                 atomic_inc(&cache->stats.promotion);
1872                 demote_then_promote(cache, structs, lookup_result.old_oblock,
1873                                     block, lookup_result.cblock,
1874                                     ool.cell, new_ocell);
1875                 release_cell = false;
1876                 break;
1877
1878         default:
1879                 DMERR_LIMIT("%s: %s: erroring bio, unknown policy op: %u",
1880                             cache_device_name(cache), __func__,
1881                             (unsigned) lookup_result.op);
1882                 bio_io_error(bio);
1883         }
1884
1885         if (release_cell)
1886                 cell_defer(cache, new_ocell, false);
1887 }
1888
1889 static void process_bio(struct cache *cache, struct prealloc *structs,
1890                         struct bio *bio)
1891 {
1892         int r;
1893         dm_oblock_t block = get_bio_block(cache, bio);
1894         struct dm_bio_prison_cell *cell_prealloc, *new_ocell;
1895
1896         /*
1897          * Check to see if that block is currently migrating.
1898          */
1899         cell_prealloc = prealloc_get_cell(structs);
1900         r = bio_detain(cache, block, bio, cell_prealloc,
1901                        (cell_free_fn) prealloc_put_cell,
1902                        structs, &new_ocell);
1903         if (r > 0)
1904                 return;
1905
1906         process_cell(cache, structs, new_ocell);
1907 }
1908
1909 static int need_commit_due_to_time(struct cache *cache)
1910 {
1911         return jiffies < cache->last_commit_jiffies ||
1912                jiffies > cache->last_commit_jiffies + COMMIT_PERIOD;
1913 }
1914
1915 /*
1916  * A non-zero return indicates read_only or fail_io mode.
1917  */
1918 static int commit(struct cache *cache, bool clean_shutdown)
1919 {
1920         int r;
1921
1922         if (get_cache_mode(cache) >= CM_READ_ONLY)
1923                 return -EINVAL;
1924
1925         atomic_inc(&cache->stats.commit_count);
1926         r = dm_cache_commit(cache->cmd, clean_shutdown);
1927         if (r)
1928                 metadata_operation_failed(cache, "dm_cache_commit", r);
1929
1930         return r;
1931 }
1932
1933 static int commit_if_needed(struct cache *cache)
1934 {
1935         int r = 0;
1936
1937         if ((cache->commit_requested || need_commit_due_to_time(cache)) &&
1938             dm_cache_changed_this_transaction(cache->cmd)) {
1939                 r = commit(cache, false);
1940                 cache->commit_requested = false;
1941                 cache->last_commit_jiffies = jiffies;
1942         }
1943
1944         return r;
1945 }
1946
1947 static void process_deferred_bios(struct cache *cache)
1948 {
1949         unsigned long flags;
1950         struct bio_list bios;
1951         struct bio *bio;
1952         struct prealloc structs;
1953
1954         memset(&structs, 0, sizeof(structs));
1955         bio_list_init(&bios);
1956
1957         spin_lock_irqsave(&cache->lock, flags);
1958         bio_list_merge(&bios, &cache->deferred_bios);
1959         bio_list_init(&cache->deferred_bios);
1960         spin_unlock_irqrestore(&cache->lock, flags);
1961
1962         while (!bio_list_empty(&bios)) {
1963                 /*
1964                  * If we've got no free migration structs, and processing
1965                  * this bio might require one, we pause until there are some
1966                  * prepared mappings to process.
1967                  */
1968                 if (prealloc_data_structs(cache, &structs)) {
1969                         spin_lock_irqsave(&cache->lock, flags);
1970                         bio_list_merge(&cache->deferred_bios, &bios);
1971                         spin_unlock_irqrestore(&cache->lock, flags);
1972                         break;
1973                 }
1974
1975                 bio = bio_list_pop(&bios);
1976
1977                 if (bio->bi_rw & REQ_FLUSH)
1978                         process_flush_bio(cache, bio);
1979                 else if (bio->bi_rw & REQ_DISCARD)
1980                         process_discard_bio(cache, &structs, bio);
1981                 else
1982                         process_bio(cache, &structs, bio);
1983         }
1984
1985         prealloc_free_structs(cache, &structs);
1986 }
1987
1988 static void process_deferred_cells(struct cache *cache)
1989 {
1990         unsigned long flags;
1991         struct dm_bio_prison_cell *cell, *tmp;
1992         struct list_head cells;
1993         struct prealloc structs;
1994
1995         memset(&structs, 0, sizeof(structs));
1996
1997         INIT_LIST_HEAD(&cells);
1998
1999         spin_lock_irqsave(&cache->lock, flags);
2000         list_splice_init(&cache->deferred_cells, &cells);
2001         spin_unlock_irqrestore(&cache->lock, flags);
2002
2003         list_for_each_entry_safe(cell, tmp, &cells, user_list) {
2004                 /*
2005                  * If we've got no free migration structs, and processing
2006                  * this bio might require one, we pause until there are some
2007                  * prepared mappings to process.
2008                  */
2009                 if (prealloc_data_structs(cache, &structs)) {
2010                         spin_lock_irqsave(&cache->lock, flags);
2011                         list_splice(&cells, &cache->deferred_cells);
2012                         spin_unlock_irqrestore(&cache->lock, flags);
2013                         break;
2014                 }
2015
2016                 process_cell(cache, &structs, cell);
2017         }
2018
2019         prealloc_free_structs(cache, &structs);
2020 }
2021
2022 static void process_deferred_flush_bios(struct cache *cache, bool submit_bios)
2023 {
2024         unsigned long flags;
2025         struct bio_list bios;
2026         struct bio *bio;
2027
2028         bio_list_init(&bios);
2029
2030         spin_lock_irqsave(&cache->lock, flags);
2031         bio_list_merge(&bios, &cache->deferred_flush_bios);
2032         bio_list_init(&cache->deferred_flush_bios);
2033         spin_unlock_irqrestore(&cache->lock, flags);
2034
2035         /*
2036          * These bios have already been through inc_ds()
2037          */
2038         while ((bio = bio_list_pop(&bios)))
2039                 submit_bios ? accounted_request(cache, bio) : bio_io_error(bio);
2040 }
2041
2042 static void process_deferred_writethrough_bios(struct cache *cache)
2043 {
2044         unsigned long flags;
2045         struct bio_list bios;
2046         struct bio *bio;
2047
2048         bio_list_init(&bios);
2049
2050         spin_lock_irqsave(&cache->lock, flags);
2051         bio_list_merge(&bios, &cache->deferred_writethrough_bios);
2052         bio_list_init(&cache->deferred_writethrough_bios);
2053         spin_unlock_irqrestore(&cache->lock, flags);
2054
2055         /*
2056          * These bios have already been through inc_ds()
2057          */
2058         while ((bio = bio_list_pop(&bios)))
2059                 accounted_request(cache, bio);
2060 }
2061
2062 static void writeback_some_dirty_blocks(struct cache *cache)
2063 {
2064         dm_oblock_t oblock;
2065         dm_cblock_t cblock;
2066         struct prealloc structs;
2067         struct dm_bio_prison_cell *old_ocell;
2068         bool busy = !iot_idle_for(&cache->origin_tracker, HZ);
2069
2070         memset(&structs, 0, sizeof(structs));
2071
2072         while (spare_migration_bandwidth(cache)) {
2073                 if (policy_writeback_work(cache->policy, &oblock, &cblock, busy))
2074                         break; /* no work to do */
2075
2076                 if (prealloc_data_structs(cache, &structs) ||
2077                     get_cell(cache, oblock, &structs, &old_ocell)) {
2078                         policy_set_dirty(cache->policy, oblock);
2079                         break;
2080                 }
2081
2082                 writeback(cache, &structs, oblock, cblock, old_ocell);
2083         }
2084
2085         prealloc_free_structs(cache, &structs);
2086 }
2087
2088 /*----------------------------------------------------------------
2089  * Invalidations.
2090  * Dropping something from the cache *without* writing back.
2091  *--------------------------------------------------------------*/
2092
2093 static void process_invalidation_request(struct cache *cache, struct invalidation_request *req)
2094 {
2095         int r = 0;
2096         uint64_t begin = from_cblock(req->cblocks->begin);
2097         uint64_t end = from_cblock(req->cblocks->end);
2098
2099         while (begin != end) {
2100                 r = policy_remove_cblock(cache->policy, to_cblock(begin));
2101                 if (!r) {
2102                         r = dm_cache_remove_mapping(cache->cmd, to_cblock(begin));
2103                         if (r) {
2104                                 metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
2105                                 break;
2106                         }
2107
2108                 } else if (r == -ENODATA) {
2109                         /* harmless, already unmapped */
2110                         r = 0;
2111
2112                 } else {
2113                         DMERR("%s: policy_remove_cblock failed", cache_device_name(cache));
2114                         break;
2115                 }
2116
2117                 begin++;
2118         }
2119
2120         cache->commit_requested = true;
2121
2122         req->err = r;
2123         atomic_set(&req->complete, 1);
2124
2125         wake_up(&req->result_wait);
2126 }
2127
2128 static void process_invalidation_requests(struct cache *cache)
2129 {
2130         struct list_head list;
2131         struct invalidation_request *req, *tmp;
2132
2133         INIT_LIST_HEAD(&list);
2134         spin_lock(&cache->invalidation_lock);
2135         list_splice_init(&cache->invalidation_requests, &list);
2136         spin_unlock(&cache->invalidation_lock);
2137
2138         list_for_each_entry_safe (req, tmp, &list, list)
2139                 process_invalidation_request(cache, req);
2140 }
2141
2142 /*----------------------------------------------------------------
2143  * Main worker loop
2144  *--------------------------------------------------------------*/
2145 static bool is_quiescing(struct cache *cache)
2146 {
2147         return atomic_read(&cache->quiescing);
2148 }
2149
2150 static void ack_quiescing(struct cache *cache)
2151 {
2152         if (is_quiescing(cache)) {
2153                 atomic_inc(&cache->quiescing_ack);
2154                 wake_up(&cache->quiescing_wait);
2155         }
2156 }
2157
2158 static void wait_for_quiescing_ack(struct cache *cache)
2159 {
2160         wait_event(cache->quiescing_wait, atomic_read(&cache->quiescing_ack));
2161 }
2162
2163 static void start_quiescing(struct cache *cache)
2164 {
2165         atomic_inc(&cache->quiescing);
2166         wait_for_quiescing_ack(cache);
2167 }
2168
2169 static void stop_quiescing(struct cache *cache)
2170 {
2171         atomic_set(&cache->quiescing, 0);
2172         atomic_set(&cache->quiescing_ack, 0);
2173 }
2174
2175 static void wait_for_migrations(struct cache *cache)
2176 {
2177         wait_event(cache->migration_wait, !atomic_read(&cache->nr_allocated_migrations));
2178 }
2179
2180 static void stop_worker(struct cache *cache)
2181 {
2182         cancel_delayed_work(&cache->waker);
2183         flush_workqueue(cache->wq);
2184 }
2185
2186 static void requeue_deferred_cells(struct cache *cache)
2187 {
2188         unsigned long flags;
2189         struct list_head cells;
2190         struct dm_bio_prison_cell *cell, *tmp;
2191
2192         INIT_LIST_HEAD(&cells);
2193         spin_lock_irqsave(&cache->lock, flags);
2194         list_splice_init(&cache->deferred_cells, &cells);
2195         spin_unlock_irqrestore(&cache->lock, flags);
2196
2197         list_for_each_entry_safe(cell, tmp, &cells, user_list)
2198                 cell_requeue(cache, cell);
2199 }
2200
2201 static void requeue_deferred_bios(struct cache *cache)
2202 {
2203         struct bio *bio;
2204         struct bio_list bios;
2205
2206         bio_list_init(&bios);
2207         bio_list_merge(&bios, &cache->deferred_bios);
2208         bio_list_init(&cache->deferred_bios);
2209
2210         while ((bio = bio_list_pop(&bios)))
2211                 bio_endio(bio, DM_ENDIO_REQUEUE);
2212 }
2213
2214 static int more_work(struct cache *cache)
2215 {
2216         if (is_quiescing(cache))
2217                 return !list_empty(&cache->quiesced_migrations) ||
2218                         !list_empty(&cache->completed_migrations) ||
2219                         !list_empty(&cache->need_commit_migrations);
2220         else
2221                 return !bio_list_empty(&cache->deferred_bios) ||
2222                         !list_empty(&cache->deferred_cells) ||
2223                         !bio_list_empty(&cache->deferred_flush_bios) ||
2224                         !bio_list_empty(&cache->deferred_writethrough_bios) ||
2225                         !list_empty(&cache->quiesced_migrations) ||
2226                         !list_empty(&cache->completed_migrations) ||
2227                         !list_empty(&cache->need_commit_migrations) ||
2228                         cache->invalidate;
2229 }
2230
2231 static void do_worker(struct work_struct *ws)
2232 {
2233         struct cache *cache = container_of(ws, struct cache, worker);
2234
2235         do {
2236                 if (!is_quiescing(cache)) {
2237                         writeback_some_dirty_blocks(cache);
2238                         process_deferred_writethrough_bios(cache);
2239                         process_deferred_bios(cache);
2240                         process_deferred_cells(cache);
2241                         process_invalidation_requests(cache);
2242                 }
2243
2244                 process_migrations(cache, &cache->quiesced_migrations, issue_copy_or_discard);
2245                 process_migrations(cache, &cache->completed_migrations, complete_migration);
2246
2247                 if (commit_if_needed(cache)) {
2248                         process_deferred_flush_bios(cache, false);
2249                         process_migrations(cache, &cache->need_commit_migrations, migration_failure);
2250                 } else {
2251                         process_deferred_flush_bios(cache, true);
2252                         process_migrations(cache, &cache->need_commit_migrations,
2253                                            migration_success_post_commit);
2254                 }
2255
2256                 ack_quiescing(cache);
2257
2258         } while (more_work(cache));
2259 }
2260
2261 /*
2262  * We want to commit periodically so that not too much
2263  * unwritten metadata builds up.
2264  */
2265 static void do_waker(struct work_struct *ws)
2266 {
2267         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
2268         policy_tick(cache->policy, true);
2269         wake_worker(cache);
2270         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
2271 }
2272
2273 /*----------------------------------------------------------------*/
2274
2275 static int is_congested(struct dm_dev *dev, int bdi_bits)
2276 {
2277         struct request_queue *q = bdev_get_queue(dev->bdev);
2278         return bdi_congested(&q->backing_dev_info, bdi_bits);
2279 }
2280
2281 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2282 {
2283         struct cache *cache = container_of(cb, struct cache, callbacks);
2284
2285         return is_congested(cache->origin_dev, bdi_bits) ||
2286                 is_congested(cache->cache_dev, bdi_bits);
2287 }
2288
2289 /*----------------------------------------------------------------
2290  * Target methods
2291  *--------------------------------------------------------------*/
2292
2293 /*
2294  * This function gets called on the error paths of the constructor, so we
2295  * have to cope with a partially initialised struct.
2296  */
2297 static void destroy(struct cache *cache)
2298 {
2299         unsigned i;
2300
2301         if (cache->migration_pool)
2302                 mempool_destroy(cache->migration_pool);
2303
2304         if (cache->all_io_ds)
2305                 dm_deferred_set_destroy(cache->all_io_ds);
2306
2307         if (cache->prison)
2308                 dm_bio_prison_destroy(cache->prison);
2309
2310         if (cache->wq)
2311                 destroy_workqueue(cache->wq);
2312
2313         if (cache->dirty_bitset)
2314                 free_bitset(cache->dirty_bitset);
2315
2316         if (cache->discard_bitset)
2317                 free_bitset(cache->discard_bitset);
2318
2319         if (cache->copier)
2320                 dm_kcopyd_client_destroy(cache->copier);
2321
2322         if (cache->cmd)
2323                 dm_cache_metadata_close(cache->cmd);
2324
2325         if (cache->metadata_dev)
2326                 dm_put_device(cache->ti, cache->metadata_dev);
2327
2328         if (cache->origin_dev)
2329                 dm_put_device(cache->ti, cache->origin_dev);
2330
2331         if (cache->cache_dev)
2332                 dm_put_device(cache->ti, cache->cache_dev);
2333
2334         if (cache->policy)
2335                 dm_cache_policy_destroy(cache->policy);
2336
2337         for (i = 0; i < cache->nr_ctr_args ; i++)
2338                 kfree(cache->ctr_args[i]);
2339         kfree(cache->ctr_args);
2340
2341         kfree(cache);
2342 }
2343
2344 static void cache_dtr(struct dm_target *ti)
2345 {
2346         struct cache *cache = ti->private;
2347
2348         destroy(cache);
2349 }
2350
2351 static sector_t get_dev_size(struct dm_dev *dev)
2352 {
2353         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2354 }
2355
2356 /*----------------------------------------------------------------*/
2357
2358 /*
2359  * Construct a cache device mapping.
2360  *
2361  * cache <metadata dev> <cache dev> <origin dev> <block size>
2362  *       <#feature args> [<feature arg>]*
2363  *       <policy> <#policy args> [<policy arg>]*
2364  *
2365  * metadata dev    : fast device holding the persistent metadata
2366  * cache dev       : fast device holding cached data blocks
2367  * origin dev      : slow device holding original data blocks
2368  * block size      : cache unit size in sectors
2369  *
2370  * #feature args   : number of feature arguments passed
2371  * feature args    : writethrough.  (The default is writeback.)
2372  *
2373  * policy          : the replacement policy to use
2374  * #policy args    : an even number of policy arguments corresponding
2375  *                   to key/value pairs passed to the policy
2376  * policy args     : key/value pairs passed to the policy
2377  *                   E.g. 'sequential_threshold 1024'
2378  *                   See cache-policies.txt for details.
2379  *
2380  * Optional feature arguments are:
2381  *   writethrough  : write through caching that prohibits cache block
2382  *                   content from being different from origin block content.
2383  *                   Without this argument, the default behaviour is to write
2384  *                   back cache block contents later for performance reasons,
2385  *                   so they may differ from the corresponding origin blocks.
2386  */
2387 struct cache_args {
2388         struct dm_target *ti;
2389
2390         struct dm_dev *metadata_dev;
2391
2392         struct dm_dev *cache_dev;
2393         sector_t cache_sectors;
2394
2395         struct dm_dev *origin_dev;
2396         sector_t origin_sectors;
2397
2398         uint32_t block_size;
2399
2400         const char *policy_name;
2401         int policy_argc;
2402         const char **policy_argv;
2403
2404         struct cache_features features;
2405 };
2406
2407 static void destroy_cache_args(struct cache_args *ca)
2408 {
2409         if (ca->metadata_dev)
2410                 dm_put_device(ca->ti, ca->metadata_dev);
2411
2412         if (ca->cache_dev)
2413                 dm_put_device(ca->ti, ca->cache_dev);
2414
2415         if (ca->origin_dev)
2416                 dm_put_device(ca->ti, ca->origin_dev);
2417
2418         kfree(ca);
2419 }
2420
2421 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2422 {
2423         if (!as->argc) {
2424                 *error = "Insufficient args";
2425                 return false;
2426         }
2427
2428         return true;
2429 }
2430
2431 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2432                               char **error)
2433 {
2434         int r;
2435         sector_t metadata_dev_size;
2436         char b[BDEVNAME_SIZE];
2437
2438         if (!at_least_one_arg(as, error))
2439                 return -EINVAL;
2440
2441         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2442                           &ca->metadata_dev);
2443         if (r) {
2444                 *error = "Error opening metadata device";
2445                 return r;
2446         }
2447
2448         metadata_dev_size = get_dev_size(ca->metadata_dev);
2449         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2450                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2451                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2452
2453         return 0;
2454 }
2455
2456 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2457                            char **error)
2458 {
2459         int r;
2460
2461         if (!at_least_one_arg(as, error))
2462                 return -EINVAL;
2463
2464         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2465                           &ca->cache_dev);
2466         if (r) {
2467                 *error = "Error opening cache device";
2468                 return r;
2469         }
2470         ca->cache_sectors = get_dev_size(ca->cache_dev);
2471
2472         return 0;
2473 }
2474
2475 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2476                             char **error)
2477 {
2478         int r;
2479
2480         if (!at_least_one_arg(as, error))
2481                 return -EINVAL;
2482
2483         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2484                           &ca->origin_dev);
2485         if (r) {
2486                 *error = "Error opening origin device";
2487                 return r;
2488         }
2489
2490         ca->origin_sectors = get_dev_size(ca->origin_dev);
2491         if (ca->ti->len > ca->origin_sectors) {
2492                 *error = "Device size larger than cached device";
2493                 return -EINVAL;
2494         }
2495
2496         return 0;
2497 }
2498
2499 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2500                             char **error)
2501 {
2502         unsigned long block_size;
2503
2504         if (!at_least_one_arg(as, error))
2505                 return -EINVAL;
2506
2507         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2508             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2509             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2510             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2511                 *error = "Invalid data block size";
2512                 return -EINVAL;
2513         }
2514
2515         if (block_size > ca->cache_sectors) {
2516                 *error = "Data block size is larger than the cache device";
2517                 return -EINVAL;
2518         }
2519
2520         ca->block_size = block_size;
2521
2522         return 0;
2523 }
2524
2525 static void init_features(struct cache_features *cf)
2526 {
2527         cf->mode = CM_WRITE;
2528         cf->io_mode = CM_IO_WRITEBACK;
2529 }
2530
2531 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2532                           char **error)
2533 {
2534         static struct dm_arg _args[] = {
2535                 {0, 1, "Invalid number of cache feature arguments"},
2536         };
2537
2538         int r;
2539         unsigned argc;
2540         const char *arg;
2541         struct cache_features *cf = &ca->features;
2542
2543         init_features(cf);
2544
2545         r = dm_read_arg_group(_args, as, &argc, error);
2546         if (r)
2547                 return -EINVAL;
2548
2549         while (argc--) {
2550                 arg = dm_shift_arg(as);
2551
2552                 if (!strcasecmp(arg, "writeback"))
2553                         cf->io_mode = CM_IO_WRITEBACK;
2554
2555                 else if (!strcasecmp(arg, "writethrough"))
2556                         cf->io_mode = CM_IO_WRITETHROUGH;
2557
2558                 else if (!strcasecmp(arg, "passthrough"))
2559                         cf->io_mode = CM_IO_PASSTHROUGH;
2560
2561                 else {
2562                         *error = "Unrecognised cache feature requested";
2563                         return -EINVAL;
2564                 }
2565         }
2566
2567         return 0;
2568 }
2569
2570 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2571                         char **error)
2572 {
2573         static struct dm_arg _args[] = {
2574                 {0, 1024, "Invalid number of policy arguments"},
2575         };
2576
2577         int r;
2578
2579         if (!at_least_one_arg(as, error))
2580                 return -EINVAL;
2581
2582         ca->policy_name = dm_shift_arg(as);
2583
2584         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2585         if (r)
2586                 return -EINVAL;
2587
2588         ca->policy_argv = (const char **)as->argv;
2589         dm_consume_args(as, ca->policy_argc);
2590
2591         return 0;
2592 }
2593
2594 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2595                             char **error)
2596 {
2597         int r;
2598         struct dm_arg_set as;
2599
2600         as.argc = argc;
2601         as.argv = argv;
2602
2603         r = parse_metadata_dev(ca, &as, error);
2604         if (r)
2605                 return r;
2606
2607         r = parse_cache_dev(ca, &as, error);
2608         if (r)
2609                 return r;
2610
2611         r = parse_origin_dev(ca, &as, error);
2612         if (r)
2613                 return r;
2614
2615         r = parse_block_size(ca, &as, error);
2616         if (r)
2617                 return r;
2618
2619         r = parse_features(ca, &as, error);
2620         if (r)
2621                 return r;
2622
2623         r = parse_policy(ca, &as, error);
2624         if (r)
2625                 return r;
2626
2627         return 0;
2628 }
2629
2630 /*----------------------------------------------------------------*/
2631
2632 static struct kmem_cache *migration_cache;
2633
2634 #define NOT_CORE_OPTION 1
2635
2636 static int process_config_option(struct cache *cache, const char *key, const char *value)
2637 {
2638         unsigned long tmp;
2639
2640         if (!strcasecmp(key, "migration_threshold")) {
2641                 if (kstrtoul(value, 10, &tmp))
2642                         return -EINVAL;
2643
2644                 cache->migration_threshold = tmp;
2645                 return 0;
2646         }
2647
2648         return NOT_CORE_OPTION;
2649 }
2650
2651 static int set_config_value(struct cache *cache, const char *key, const char *value)
2652 {
2653         int r = process_config_option(cache, key, value);
2654
2655         if (r == NOT_CORE_OPTION)
2656                 r = policy_set_config_value(cache->policy, key, value);
2657
2658         if (r)
2659                 DMWARN("bad config value for %s: %s", key, value);
2660
2661         return r;
2662 }
2663
2664 static int set_config_values(struct cache *cache, int argc, const char **argv)
2665 {
2666         int r = 0;
2667
2668         if (argc & 1) {
2669                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2670                 return -EINVAL;
2671         }
2672
2673         while (argc) {
2674                 r = set_config_value(cache, argv[0], argv[1]);
2675                 if (r)
2676                         break;
2677
2678                 argc -= 2;
2679                 argv += 2;
2680         }
2681
2682         return r;
2683 }
2684
2685 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2686                                char **error)
2687 {
2688         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2689                                                            cache->cache_size,
2690                                                            cache->origin_sectors,
2691                                                            cache->sectors_per_block);
2692         if (IS_ERR(p)) {
2693                 *error = "Error creating cache's policy";
2694                 return PTR_ERR(p);
2695         }
2696         cache->policy = p;
2697
2698         return 0;
2699 }
2700
2701 /*
2702  * We want the discard block size to be at least the size of the cache
2703  * block size and have no more than 2^14 discard blocks across the origin.
2704  */
2705 #define MAX_DISCARD_BLOCKS (1 << 14)
2706
2707 static bool too_many_discard_blocks(sector_t discard_block_size,
2708                                     sector_t origin_size)
2709 {
2710         (void) sector_div(origin_size, discard_block_size);
2711
2712         return origin_size > MAX_DISCARD_BLOCKS;
2713 }
2714
2715 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2716                                              sector_t origin_size)
2717 {
2718         sector_t discard_block_size = cache_block_size;
2719
2720         if (origin_size)
2721                 while (too_many_discard_blocks(discard_block_size, origin_size))
2722                         discard_block_size *= 2;
2723
2724         return discard_block_size;
2725 }
2726
2727 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2728 {
2729         dm_block_t nr_blocks = from_cblock(size);
2730
2731         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2732                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2733                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2734                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2735                              (unsigned long long) nr_blocks);
2736
2737         cache->cache_size = size;
2738 }
2739
2740 #define DEFAULT_MIGRATION_THRESHOLD 2048
2741
2742 static int cache_create(struct cache_args *ca, struct cache **result)
2743 {
2744         int r = 0;
2745         char **error = &ca->ti->error;
2746         struct cache *cache;
2747         struct dm_target *ti = ca->ti;
2748         dm_block_t origin_blocks;
2749         struct dm_cache_metadata *cmd;
2750         bool may_format = ca->features.mode == CM_WRITE;
2751
2752         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2753         if (!cache)
2754                 return -ENOMEM;
2755
2756         cache->ti = ca->ti;
2757         ti->private = cache;
2758         ti->num_flush_bios = 2;
2759         ti->flush_supported = true;
2760
2761         ti->num_discard_bios = 1;
2762         ti->discards_supported = true;
2763         ti->discard_zeroes_data_unsupported = true;
2764         ti->split_discard_bios = false;
2765
2766         cache->features = ca->features;
2767         ti->per_bio_data_size = get_per_bio_data_size(cache);
2768
2769         cache->callbacks.congested_fn = cache_is_congested;
2770         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2771
2772         cache->metadata_dev = ca->metadata_dev;
2773         cache->origin_dev = ca->origin_dev;
2774         cache->cache_dev = ca->cache_dev;
2775
2776         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2777
2778         /* FIXME: factor out this whole section */
2779         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2780         origin_blocks = block_div(origin_blocks, ca->block_size);
2781         cache->origin_blocks = to_oblock(origin_blocks);
2782
2783         cache->sectors_per_block = ca->block_size;
2784         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2785                 r = -EINVAL;
2786                 goto bad;
2787         }
2788
2789         if (ca->block_size & (ca->block_size - 1)) {
2790                 dm_block_t cache_size = ca->cache_sectors;
2791
2792                 cache->sectors_per_block_shift = -1;
2793                 cache_size = block_div(cache_size, ca->block_size);
2794                 set_cache_size(cache, to_cblock(cache_size));
2795         } else {
2796                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2797                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2798         }
2799
2800         r = create_cache_policy(cache, ca, error);
2801         if (r)
2802                 goto bad;
2803
2804         cache->policy_nr_args = ca->policy_argc;
2805         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2806
2807         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2808         if (r) {
2809                 *error = "Error setting cache policy's config values";
2810                 goto bad;
2811         }
2812
2813         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2814                                      ca->block_size, may_format,
2815                                      dm_cache_policy_get_hint_size(cache->policy));
2816         if (IS_ERR(cmd)) {
2817                 *error = "Error creating metadata object";
2818                 r = PTR_ERR(cmd);
2819                 goto bad;
2820         }
2821         cache->cmd = cmd;
2822         set_cache_mode(cache, CM_WRITE);
2823         if (get_cache_mode(cache) != CM_WRITE) {
2824                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2825                 r = -EINVAL;
2826                 goto bad;
2827         }
2828
2829         if (passthrough_mode(&cache->features)) {
2830                 bool all_clean;
2831
2832                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2833                 if (r) {
2834                         *error = "dm_cache_metadata_all_clean() failed";
2835                         goto bad;
2836                 }
2837
2838                 if (!all_clean) {
2839                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2840                         r = -EINVAL;
2841                         goto bad;
2842                 }
2843         }
2844
2845         spin_lock_init(&cache->lock);
2846         INIT_LIST_HEAD(&cache->deferred_cells);
2847         bio_list_init(&cache->deferred_bios);
2848         bio_list_init(&cache->deferred_flush_bios);
2849         bio_list_init(&cache->deferred_writethrough_bios);
2850         INIT_LIST_HEAD(&cache->quiesced_migrations);
2851         INIT_LIST_HEAD(&cache->completed_migrations);
2852         INIT_LIST_HEAD(&cache->need_commit_migrations);
2853         atomic_set(&cache->nr_allocated_migrations, 0);
2854         atomic_set(&cache->nr_io_migrations, 0);
2855         init_waitqueue_head(&cache->migration_wait);
2856
2857         init_waitqueue_head(&cache->quiescing_wait);
2858         atomic_set(&cache->quiescing, 0);
2859         atomic_set(&cache->quiescing_ack, 0);
2860
2861         r = -ENOMEM;
2862         atomic_set(&cache->nr_dirty, 0);
2863         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2864         if (!cache->dirty_bitset) {
2865                 *error = "could not allocate dirty bitset";
2866                 goto bad;
2867         }
2868         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2869
2870         cache->discard_block_size =
2871                 calculate_discard_block_size(cache->sectors_per_block,
2872                                              cache->origin_sectors);
2873         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2874                                                               cache->discard_block_size));
2875         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2876         if (!cache->discard_bitset) {
2877                 *error = "could not allocate discard bitset";
2878                 goto bad;
2879         }
2880         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2881
2882         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2883         if (IS_ERR(cache->copier)) {
2884                 *error = "could not create kcopyd client";
2885                 r = PTR_ERR(cache->copier);
2886                 goto bad;
2887         }
2888
2889         cache->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
2890         if (!cache->wq) {
2891                 *error = "could not create workqueue for metadata object";
2892                 goto bad;
2893         }
2894         INIT_WORK(&cache->worker, do_worker);
2895         INIT_DELAYED_WORK(&cache->waker, do_waker);
2896         cache->last_commit_jiffies = jiffies;
2897
2898         cache->prison = dm_bio_prison_create();
2899         if (!cache->prison) {
2900                 *error = "could not create bio prison";
2901                 goto bad;
2902         }
2903
2904         cache->all_io_ds = dm_deferred_set_create();
2905         if (!cache->all_io_ds) {
2906                 *error = "could not create all_io deferred set";
2907                 goto bad;
2908         }
2909
2910         cache->migration_pool = mempool_create_slab_pool(MIGRATION_POOL_SIZE,
2911                                                          migration_cache);
2912         if (!cache->migration_pool) {
2913                 *error = "Error creating cache's migration mempool";
2914                 goto bad;
2915         }
2916
2917         cache->need_tick_bio = true;
2918         cache->sized = false;
2919         cache->invalidate = false;
2920         cache->commit_requested = false;
2921         cache->loaded_mappings = false;
2922         cache->loaded_discards = false;
2923
2924         load_stats(cache);
2925
2926         atomic_set(&cache->stats.demotion, 0);
2927         atomic_set(&cache->stats.promotion, 0);
2928         atomic_set(&cache->stats.copies_avoided, 0);
2929         atomic_set(&cache->stats.cache_cell_clash, 0);
2930         atomic_set(&cache->stats.commit_count, 0);
2931         atomic_set(&cache->stats.discard_count, 0);
2932
2933         spin_lock_init(&cache->invalidation_lock);
2934         INIT_LIST_HEAD(&cache->invalidation_requests);
2935
2936         iot_init(&cache->origin_tracker);
2937
2938         *result = cache;
2939         return 0;
2940
2941 bad:
2942         destroy(cache);
2943         return r;
2944 }
2945
2946 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2947 {
2948         unsigned i;
2949         const char **copy;
2950
2951         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2952         if (!copy)
2953                 return -ENOMEM;
2954         for (i = 0; i < argc; i++) {
2955                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2956                 if (!copy[i]) {
2957                         while (i--)
2958                                 kfree(copy[i]);
2959                         kfree(copy);
2960                         return -ENOMEM;
2961                 }
2962         }
2963
2964         cache->nr_ctr_args = argc;
2965         cache->ctr_args = copy;
2966
2967         return 0;
2968 }
2969
2970 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2971 {
2972         int r = -EINVAL;
2973         struct cache_args *ca;
2974         struct cache *cache = NULL;
2975
2976         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2977         if (!ca) {
2978                 ti->error = "Error allocating memory for cache";
2979                 return -ENOMEM;
2980         }
2981         ca->ti = ti;
2982
2983         r = parse_cache_args(ca, argc, argv, &ti->error);
2984         if (r)
2985                 goto out;
2986
2987         r = cache_create(ca, &cache);
2988         if (r)
2989                 goto out;
2990
2991         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2992         if (r) {
2993                 destroy(cache);
2994                 goto out;
2995         }
2996
2997         ti->private = cache;
2998
2999 out:
3000         destroy_cache_args(ca);
3001         return r;
3002 }
3003
3004 /*----------------------------------------------------------------*/
3005
3006 static int cache_map(struct dm_target *ti, struct bio *bio)
3007 {
3008         struct cache *cache = ti->private;
3009
3010         int r;
3011         struct dm_bio_prison_cell *cell = NULL;
3012         dm_oblock_t block = get_bio_block(cache, bio);
3013         size_t pb_data_size = get_per_bio_data_size(cache);
3014         bool can_migrate = false;
3015         bool fast_promotion;
3016         struct policy_result lookup_result;
3017         struct per_bio_data *pb = init_per_bio_data(bio, pb_data_size);
3018         struct old_oblock_lock ool;
3019
3020         ool.locker.fn = null_locker;
3021
3022         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
3023                 /*
3024                  * This can only occur if the io goes to a partial block at
3025                  * the end of the origin device.  We don't cache these.
3026                  * Just remap to the origin and carry on.
3027                  */
3028                 remap_to_origin(cache, bio);
3029                 accounted_begin(cache, bio);
3030                 return DM_MAPIO_REMAPPED;
3031         }
3032
3033         if (discard_or_flush(bio)) {
3034                 defer_bio(cache, bio);
3035                 return DM_MAPIO_SUBMITTED;
3036         }
3037
3038         /*
3039          * Check to see if that block is currently migrating.
3040          */
3041         cell = alloc_prison_cell(cache);
3042         if (!cell) {
3043                 defer_bio(cache, bio);
3044                 return DM_MAPIO_SUBMITTED;
3045         }
3046
3047         r = bio_detain(cache, block, bio, cell,
3048                        (cell_free_fn) free_prison_cell,
3049                        cache, &cell);
3050         if (r) {
3051                 if (r < 0)
3052                         defer_bio(cache, bio);
3053
3054                 return DM_MAPIO_SUBMITTED;
3055         }
3056
3057         fast_promotion = is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio);
3058
3059         r = policy_map(cache->policy, block, false, can_migrate, fast_promotion,
3060                        bio, &ool.locker, &lookup_result);
3061         if (r == -EWOULDBLOCK) {
3062                 cell_defer(cache, cell, true);
3063                 return DM_MAPIO_SUBMITTED;
3064
3065         } else if (r) {
3066                 DMERR_LIMIT("%s: Unexpected return from cache replacement policy: %d",
3067                             cache_device_name(cache), r);
3068                 cell_defer(cache, cell, false);
3069                 bio_io_error(bio);
3070                 return DM_MAPIO_SUBMITTED;
3071         }
3072
3073         r = DM_MAPIO_REMAPPED;
3074         switch (lookup_result.op) {
3075         case POLICY_HIT:
3076                 if (passthrough_mode(&cache->features)) {
3077                         if (bio_data_dir(bio) == WRITE) {
3078                                 /*
3079                                  * We need to invalidate this block, so
3080                                  * defer for the worker thread.
3081                                  */
3082                                 cell_defer(cache, cell, true);
3083                                 r = DM_MAPIO_SUBMITTED;
3084
3085                         } else {
3086                                 inc_miss_counter(cache, bio);
3087                                 remap_to_origin_clear_discard(cache, bio, block);
3088                                 accounted_begin(cache, bio);
3089                                 inc_ds(cache, bio, cell);
3090                                 // FIXME: we want to remap hits or misses straight
3091                                 // away rather than passing over to the worker.
3092                                 cell_defer(cache, cell, false);
3093                         }
3094
3095                 } else {
3096                         inc_hit_counter(cache, bio);
3097                         if (bio_data_dir(bio) == WRITE && writethrough_mode(&cache->features) &&
3098                             !is_dirty(cache, lookup_result.cblock)) {
3099                                 remap_to_origin_then_cache(cache, bio, block, lookup_result.cblock);
3100                                 accounted_begin(cache, bio);
3101                                 inc_ds(cache, bio, cell);
3102                                 cell_defer(cache, cell, false);
3103
3104                         } else
3105                                 remap_cell_to_cache_dirty(cache, cell, block, lookup_result.cblock, false);
3106                 }
3107                 break;
3108
3109         case POLICY_MISS:
3110                 inc_miss_counter(cache, bio);
3111                 if (pb->req_nr != 0) {
3112                         /*
3113                          * This is a duplicate writethrough io that is no
3114                          * longer needed because the block has been demoted.
3115                          */
3116                         bio_endio(bio, 0);
3117                         // FIXME: remap everything as a miss
3118                         cell_defer(cache, cell, false);
3119                         r = DM_MAPIO_SUBMITTED;
3120
3121                 } else
3122                         remap_cell_to_origin_clear_discard(cache, cell, block, false);
3123                 break;
3124
3125         default:
3126                 DMERR_LIMIT("%s: %s: erroring bio: unknown policy op: %u",
3127                             cache_device_name(cache), __func__,
3128                             (unsigned) lookup_result.op);
3129                 cell_defer(cache, cell, false);
3130                 bio_io_error(bio);
3131                 r = DM_MAPIO_SUBMITTED;
3132         }
3133
3134         return r;
3135 }
3136
3137 static int cache_end_io(struct dm_target *ti, struct bio *bio, int error)
3138 {
3139         struct cache *cache = ti->private;
3140         unsigned long flags;
3141         size_t pb_data_size = get_per_bio_data_size(cache);
3142         struct per_bio_data *pb = get_per_bio_data(bio, pb_data_size);
3143
3144         if (pb->tick) {
3145                 policy_tick(cache->policy, false);
3146
3147                 spin_lock_irqsave(&cache->lock, flags);
3148                 cache->need_tick_bio = true;
3149                 spin_unlock_irqrestore(&cache->lock, flags);
3150         }
3151
3152         check_for_quiesced_migrations(cache, pb);
3153         accounted_complete(cache, bio);
3154
3155         return 0;
3156 }
3157
3158 static int write_dirty_bitset(struct cache *cache)
3159 {
3160         unsigned i, r;
3161
3162         if (get_cache_mode(cache) >= CM_READ_ONLY)
3163                 return -EINVAL;
3164
3165         for (i = 0; i < from_cblock(cache->cache_size); i++) {
3166                 r = dm_cache_set_dirty(cache->cmd, to_cblock(i),
3167                                        is_dirty(cache, to_cblock(i)));
3168                 if (r) {
3169                         metadata_operation_failed(cache, "dm_cache_set_dirty", r);
3170                         return r;
3171                 }
3172         }
3173
3174         return 0;
3175 }
3176
3177 static int write_discard_bitset(struct cache *cache)
3178 {
3179         unsigned i, r;
3180
3181         if (get_cache_mode(cache) >= CM_READ_ONLY)
3182                 return -EINVAL;
3183
3184         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
3185                                            cache->discard_nr_blocks);
3186         if (r) {
3187                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
3188                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
3189                 return r;
3190         }
3191
3192         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
3193                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
3194                                          is_discarded(cache, to_dblock(i)));
3195                 if (r) {
3196                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
3197                         return r;
3198                 }
3199         }
3200
3201         return 0;
3202 }
3203
3204 static int write_hints(struct cache *cache)
3205 {
3206         int r;
3207
3208         if (get_cache_mode(cache) >= CM_READ_ONLY)
3209                 return -EINVAL;
3210
3211         r = dm_cache_write_hints(cache->cmd, cache->policy);
3212         if (r) {
3213                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
3214                 return r;
3215         }
3216
3217         return 0;
3218 }
3219
3220 /*
3221  * returns true on success
3222  */
3223 static bool sync_metadata(struct cache *cache)
3224 {
3225         int r1, r2, r3, r4;
3226
3227         r1 = write_dirty_bitset(cache);
3228         if (r1)
3229                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
3230
3231         r2 = write_discard_bitset(cache);
3232         if (r2)
3233                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
3234
3235         save_stats(cache);
3236
3237         r3 = write_hints(cache);
3238         if (r3)
3239                 DMERR("%s: could not write hints", cache_device_name(cache));
3240
3241         /*
3242          * If writing the above metadata failed, we still commit, but don't
3243          * set the clean shutdown flag.  This will effectively force every
3244          * dirty bit to be set on reload.
3245          */
3246         r4 = commit(cache, !r1 && !r2 && !r3);
3247         if (r4)
3248                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
3249
3250         return !r1 && !r2 && !r3 && !r4;
3251 }
3252
3253 static void cache_postsuspend(struct dm_target *ti)
3254 {
3255         struct cache *cache = ti->private;
3256
3257         start_quiescing(cache);
3258         wait_for_migrations(cache);
3259         stop_worker(cache);
3260         requeue_deferred_bios(cache);
3261         requeue_deferred_cells(cache);
3262         stop_quiescing(cache);
3263
3264         if (get_cache_mode(cache) == CM_WRITE)
3265                 (void) sync_metadata(cache);
3266 }
3267
3268 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
3269                         bool dirty, uint32_t hint, bool hint_valid)
3270 {
3271         int r;
3272         struct cache *cache = context;
3273
3274         r = policy_load_mapping(cache->policy, oblock, cblock, hint, hint_valid);
3275         if (r)
3276                 return r;
3277
3278         if (dirty)
3279                 set_dirty(cache, oblock, cblock);
3280         else
3281                 clear_dirty(cache, oblock, cblock);
3282
3283         return 0;
3284 }
3285
3286 /*
3287  * The discard block size in the on disk metadata is not
3288  * neccessarily the same as we're currently using.  So we have to
3289  * be careful to only set the discarded attribute if we know it
3290  * covers a complete block of the new size.
3291  */
3292 struct discard_load_info {
3293         struct cache *cache;
3294
3295         /*
3296          * These blocks are sized using the on disk dblock size, rather
3297          * than the current one.
3298          */
3299         dm_block_t block_size;
3300         dm_block_t discard_begin, discard_end;
3301 };
3302
3303 static void discard_load_info_init(struct cache *cache,
3304                                    struct discard_load_info *li)
3305 {
3306         li->cache = cache;
3307         li->discard_begin = li->discard_end = 0;
3308 }
3309
3310 static void set_discard_range(struct discard_load_info *li)
3311 {
3312         sector_t b, e;
3313
3314         if (li->discard_begin == li->discard_end)
3315                 return;
3316
3317         /*
3318          * Convert to sectors.
3319          */
3320         b = li->discard_begin * li->block_size;
3321         e = li->discard_end * li->block_size;
3322
3323         /*
3324          * Then convert back to the current dblock size.
3325          */
3326         b = dm_sector_div_up(b, li->cache->discard_block_size);
3327         sector_div(e, li->cache->discard_block_size);
3328
3329         /*
3330          * The origin may have shrunk, so we need to check we're still in
3331          * bounds.
3332          */
3333         if (e > from_dblock(li->cache->discard_nr_blocks))
3334                 e = from_dblock(li->cache->discard_nr_blocks);
3335
3336         for (; b < e; b++)
3337                 set_discard(li->cache, to_dblock(b));
3338 }
3339
3340 static int load_discard(void *context, sector_t discard_block_size,
3341                         dm_dblock_t dblock, bool discard)
3342 {
3343         struct discard_load_info *li = context;
3344
3345         li->block_size = discard_block_size;
3346
3347         if (discard) {
3348                 if (from_dblock(dblock) == li->discard_end)
3349                         /*
3350                          * We're already in a discard range, just extend it.
3351                          */
3352                         li->discard_end = li->discard_end + 1ULL;
3353
3354                 else {
3355                         /*
3356                          * Emit the old range and start a new one.
3357                          */
3358                         set_discard_range(li);
3359                         li->discard_begin = from_dblock(dblock);
3360                         li->discard_end = li->discard_begin + 1ULL;
3361                 }
3362         } else {
3363                 set_discard_range(li);
3364                 li->discard_begin = li->discard_end = 0;
3365         }
3366
3367         return 0;
3368 }
3369
3370 static dm_cblock_t get_cache_dev_size(struct cache *cache)
3371 {
3372         sector_t size = get_dev_size(cache->cache_dev);
3373         (void) sector_div(size, cache->sectors_per_block);
3374         return to_cblock(size);
3375 }
3376
3377 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
3378 {
3379         if (from_cblock(new_size) > from_cblock(cache->cache_size))
3380                 return true;
3381
3382         /*
3383          * We can't drop a dirty block when shrinking the cache.
3384          */
3385         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3386                 new_size = to_cblock(from_cblock(new_size) + 1);
3387                 if (is_dirty(cache, new_size)) {
3388                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3389                               cache_device_name(cache),
3390                               (unsigned long long) from_cblock(new_size));
3391                         return false;
3392                 }
3393         }
3394
3395         return true;
3396 }
3397
3398 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3399 {
3400         int r;
3401
3402         r = dm_cache_resize(cache->cmd, new_size);
3403         if (r) {
3404                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3405                 metadata_operation_failed(cache, "dm_cache_resize", r);
3406                 return r;
3407         }
3408
3409         set_cache_size(cache, new_size);
3410
3411         return 0;
3412 }
3413
3414 static int cache_preresume(struct dm_target *ti)
3415 {
3416         int r = 0;
3417         struct cache *cache = ti->private;
3418         dm_cblock_t csize = get_cache_dev_size(cache);
3419
3420         /*
3421          * Check to see if the cache has resized.
3422          */
3423         if (!cache->sized) {
3424                 r = resize_cache_dev(cache, csize);
3425                 if (r)
3426                         return r;
3427
3428                 cache->sized = true;
3429
3430         } else if (csize != cache->cache_size) {
3431                 if (!can_resize(cache, csize))
3432                         return -EINVAL;
3433
3434                 r = resize_cache_dev(cache, csize);
3435                 if (r)
3436                         return r;
3437         }
3438
3439         if (!cache->loaded_mappings) {
3440                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3441                                            load_mapping, cache);
3442                 if (r) {
3443                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3444                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3445                         return r;
3446                 }
3447
3448                 cache->loaded_mappings = true;
3449         }
3450
3451         if (!cache->loaded_discards) {
3452                 struct discard_load_info li;
3453
3454                 /*
3455                  * The discard bitset could have been resized, or the
3456                  * discard block size changed.  To be safe we start by
3457                  * setting every dblock to not discarded.
3458                  */
3459                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3460
3461                 discard_load_info_init(cache, &li);
3462                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3463                 if (r) {
3464                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3465                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3466                         return r;
3467                 }
3468                 set_discard_range(&li);
3469
3470                 cache->loaded_discards = true;
3471         }
3472
3473         return r;
3474 }
3475
3476 static void cache_resume(struct dm_target *ti)
3477 {
3478         struct cache *cache = ti->private;
3479
3480         cache->need_tick_bio = true;
3481         do_waker(&cache->waker.work);
3482 }
3483
3484 /*
3485  * Status format:
3486  *
3487  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3488  * <cache block size> <#used cache blocks>/<#total cache blocks>
3489  * <#read hits> <#read misses> <#write hits> <#write misses>
3490  * <#demotions> <#promotions> <#dirty>
3491  * <#features> <features>*
3492  * <#core args> <core args>
3493  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3494  */
3495 static void cache_status(struct dm_target *ti, status_type_t type,
3496                          unsigned status_flags, char *result, unsigned maxlen)
3497 {
3498         int r = 0;
3499         unsigned i;
3500         ssize_t sz = 0;
3501         dm_block_t nr_free_blocks_metadata = 0;
3502         dm_block_t nr_blocks_metadata = 0;
3503         char buf[BDEVNAME_SIZE];
3504         struct cache *cache = ti->private;
3505         dm_cblock_t residency;
3506
3507         switch (type) {
3508         case STATUSTYPE_INFO:
3509                 if (get_cache_mode(cache) == CM_FAIL) {
3510                         DMEMIT("Fail");
3511                         break;
3512                 }
3513
3514                 /* Commit to ensure statistics aren't out-of-date */
3515                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3516                         (void) commit(cache, false);
3517
3518                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3519                 if (r) {
3520                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3521                               cache_device_name(cache), r);
3522                         goto err;
3523                 }
3524
3525                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3526                 if (r) {
3527                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3528                               cache_device_name(cache), r);
3529                         goto err;
3530                 }
3531
3532                 residency = policy_residency(cache->policy);
3533
3534                 DMEMIT("%u %llu/%llu %u %llu/%llu %u %u %u %u %u %u %lu ",
3535                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3536                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3537                        (unsigned long long)nr_blocks_metadata,
3538                        cache->sectors_per_block,
3539                        (unsigned long long) from_cblock(residency),
3540                        (unsigned long long) from_cblock(cache->cache_size),
3541                        (unsigned) atomic_read(&cache->stats.read_hit),
3542                        (unsigned) atomic_read(&cache->stats.read_miss),
3543                        (unsigned) atomic_read(&cache->stats.write_hit),
3544                        (unsigned) atomic_read(&cache->stats.write_miss),
3545                        (unsigned) atomic_read(&cache->stats.demotion),
3546                        (unsigned) atomic_read(&cache->stats.promotion),
3547                        (unsigned long) atomic_read(&cache->nr_dirty));
3548
3549                 if (writethrough_mode(&cache->features))
3550                         DMEMIT("1 writethrough ");
3551
3552                 else if (passthrough_mode(&cache->features))
3553                         DMEMIT("1 passthrough ");
3554
3555                 else if (writeback_mode(&cache->features))
3556                         DMEMIT("1 writeback ");
3557
3558                 else {
3559                         DMERR("%s: internal error: unknown io mode: %d",
3560                               cache_device_name(cache), (int) cache->features.io_mode);
3561                         goto err;
3562                 }
3563
3564                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3565
3566                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3567                 if (sz < maxlen) {
3568                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3569                         if (r)
3570                                 DMERR("%s: policy_emit_config_values returned %d",
3571                                       cache_device_name(cache), r);
3572                 }
3573
3574                 if (get_cache_mode(cache) == CM_READ_ONLY)
3575                         DMEMIT("ro ");
3576                 else
3577                         DMEMIT("rw ");
3578
3579                 if (dm_cache_metadata_needs_check(cache->cmd))
3580                         DMEMIT("needs_check ");
3581                 else
3582                         DMEMIT("- ");
3583
3584                 break;
3585
3586         case STATUSTYPE_TABLE:
3587                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3588                 DMEMIT("%s ", buf);
3589                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3590                 DMEMIT("%s ", buf);
3591                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3592                 DMEMIT("%s", buf);
3593
3594                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3595                         DMEMIT(" %s", cache->ctr_args[i]);
3596                 if (cache->nr_ctr_args)
3597                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3598         }
3599
3600         return;
3601
3602 err:
3603         DMEMIT("Error");
3604 }
3605
3606 /*
3607  * A cache block range can take two forms:
3608  *
3609  * i) A single cblock, eg. '3456'
3610  * ii) A begin and end cblock with dots between, eg. 123-234
3611  */
3612 static int parse_cblock_range(struct cache *cache, const char *str,
3613                               struct cblock_range *result)
3614 {
3615         char dummy;
3616         uint64_t b, e;
3617         int r;
3618
3619         /*
3620          * Try and parse form (ii) first.
3621          */
3622         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3623         if (r < 0)
3624                 return r;
3625
3626         if (r == 2) {
3627                 result->begin = to_cblock(b);
3628                 result->end = to_cblock(e);
3629                 return 0;
3630         }
3631
3632         /*
3633          * That didn't work, try form (i).
3634          */
3635         r = sscanf(str, "%llu%c", &b, &dummy);
3636         if (r < 0)
3637                 return r;
3638
3639         if (r == 1) {
3640                 result->begin = to_cblock(b);
3641                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3642                 return 0;
3643         }
3644
3645         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3646         return -EINVAL;
3647 }
3648
3649 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3650 {
3651         uint64_t b = from_cblock(range->begin);
3652         uint64_t e = from_cblock(range->end);
3653         uint64_t n = from_cblock(cache->cache_size);
3654
3655         if (b >= n) {
3656                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3657                       cache_device_name(cache), b, n);
3658                 return -EINVAL;
3659         }
3660
3661         if (e > n) {
3662                 DMERR("%s: end cblock out of range: %llu > %llu",
3663                       cache_device_name(cache), e, n);
3664                 return -EINVAL;
3665         }
3666
3667         if (b >= e) {
3668                 DMERR("%s: invalid cblock range: %llu >= %llu",
3669                       cache_device_name(cache), b, e);
3670                 return -EINVAL;
3671         }
3672
3673         return 0;
3674 }
3675
3676 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3677 {
3678         struct invalidation_request req;
3679
3680         INIT_LIST_HEAD(&req.list);
3681         req.cblocks = range;
3682         atomic_set(&req.complete, 0);
3683         req.err = 0;
3684         init_waitqueue_head(&req.result_wait);
3685
3686         spin_lock(&cache->invalidation_lock);
3687         list_add(&req.list, &cache->invalidation_requests);
3688         spin_unlock(&cache->invalidation_lock);
3689         wake_worker(cache);
3690
3691         wait_event(req.result_wait, atomic_read(&req.complete));
3692         return req.err;
3693 }
3694
3695 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3696                                               const char **cblock_ranges)
3697 {
3698         int r = 0;
3699         unsigned i;
3700         struct cblock_range range;
3701
3702         if (!passthrough_mode(&cache->features)) {
3703                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3704                       cache_device_name(cache));
3705                 return -EPERM;
3706         }
3707
3708         for (i = 0; i < count; i++) {
3709                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3710                 if (r)
3711                         break;
3712
3713                 r = validate_cblock_range(cache, &range);
3714                 if (r)
3715                         break;
3716
3717                 /*
3718                  * Pass begin and end origin blocks to the worker and wake it.
3719                  */
3720                 r = request_invalidation(cache, &range);
3721                 if (r)
3722                         break;
3723         }
3724
3725         return r;
3726 }
3727
3728 /*
3729  * Supports
3730  *      "<key> <value>"
3731  * and
3732  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3733  *
3734  * The key migration_threshold is supported by the cache target core.
3735  */
3736 static int cache_message(struct dm_target *ti, unsigned argc, char **argv)
3737 {
3738         struct cache *cache = ti->private;
3739
3740         if (!argc)
3741                 return -EINVAL;
3742
3743         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3744                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3745                       cache_device_name(cache));
3746                 return -EOPNOTSUPP;
3747         }
3748
3749         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3750                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3751
3752         if (argc != 2)
3753                 return -EINVAL;
3754
3755         return set_config_value(cache, argv[0], argv[1]);
3756 }
3757
3758 static int cache_iterate_devices(struct dm_target *ti,
3759                                  iterate_devices_callout_fn fn, void *data)
3760 {
3761         int r = 0;
3762         struct cache *cache = ti->private;
3763
3764         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3765         if (!r)
3766                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3767
3768         return r;
3769 }
3770
3771 /*
3772  * We assume I/O is going to the origin (which is the volume
3773  * more likely to have restrictions e.g. by being striped).
3774  * (Looking up the exact location of the data would be expensive
3775  * and could always be out of date by the time the bio is submitted.)
3776  */
3777 static int cache_bvec_merge(struct dm_target *ti,
3778                             struct bvec_merge_data *bvm,
3779                             struct bio_vec *biovec, int max_size)
3780 {
3781         struct cache *cache = ti->private;
3782         struct request_queue *q = bdev_get_queue(cache->origin_dev->bdev);
3783
3784         if (!q->merge_bvec_fn)
3785                 return max_size;
3786
3787         bvm->bi_bdev = cache->origin_dev->bdev;
3788         return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
3789 }
3790
3791 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3792 {
3793         /*
3794          * FIXME: these limits may be incompatible with the cache device
3795          */
3796         limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3797                                             cache->origin_sectors);
3798         limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3799 }
3800
3801 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3802 {
3803         struct cache *cache = ti->private;
3804         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3805
3806         /*
3807          * If the system-determined stacked limits are compatible with the
3808          * cache's blocksize (io_opt is a factor) do not override them.
3809          */
3810         if (io_opt_sectors < cache->sectors_per_block ||
3811             do_div(io_opt_sectors, cache->sectors_per_block)) {
3812                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3813                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3814         }
3815         set_discard_limits(cache, limits);
3816 }
3817
3818 /*----------------------------------------------------------------*/
3819
3820 static struct target_type cache_target = {
3821         .name = "cache",
3822         .version = {1, 8, 0},
3823         .module = THIS_MODULE,
3824         .ctr = cache_ctr,
3825         .dtr = cache_dtr,
3826         .map = cache_map,
3827         .end_io = cache_end_io,
3828         .postsuspend = cache_postsuspend,
3829         .preresume = cache_preresume,
3830         .resume = cache_resume,
3831         .status = cache_status,
3832         .message = cache_message,
3833         .iterate_devices = cache_iterate_devices,
3834         .merge = cache_bvec_merge,
3835         .io_hints = cache_io_hints,
3836 };
3837
3838 static int __init dm_cache_init(void)
3839 {
3840         int r;
3841
3842         r = dm_register_target(&cache_target);
3843         if (r) {
3844                 DMERR("cache target registration failed: %d", r);
3845                 return r;
3846         }
3847
3848         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3849         if (!migration_cache) {
3850                 dm_unregister_target(&cache_target);
3851                 return -ENOMEM;
3852         }
3853
3854         return 0;
3855 }
3856
3857 static void __exit dm_cache_exit(void)
3858 {
3859         dm_unregister_target(&cache_target);
3860         kmem_cache_destroy(migration_cache);
3861 }
3862
3863 module_init(dm_cache_init);
3864 module_exit(dm_cache_exit);
3865
3866 MODULE_DESCRIPTION(DM_NAME " cache target");
3867 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3868 MODULE_LICENSE("GPL");