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