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-v2.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/rwsem.h>
  19 #include <linux/slab.h>
  20 #include <linux/vmalloc.h>
  21
  22 #define DM_MSG_PREFIX "cache"
  23
  24 DECLARE_DM_KCOPYD_THROTTLE_WITH_MODULE_PARM(cache_copy_throttle,
  25         "A percentage of time allocated for copying to and/or from cache");
  26
  27 /*----------------------------------------------------------------*/
  28
  29 /*
  30  * Glossary:
  31  *
  32  * oblock: index of an origin block
  33  * cblock: index of a cache block
  34  * promotion: movement of a block from origin to cache
  35  * demotion: movement of a block from cache to origin
  36  * migration: movement of a block between the origin and cache device,
  37  *            either direction
  38  */
  39
  40 /*----------------------------------------------------------------*/
  41
  42 struct io_tracker {
  43         spinlock_t lock;
  44
  45         /*
  46          * Sectors of in-flight IO.
  47          */
  48         sector_t in_flight;
  49
  50         /*
  51          * The time, in jiffies, when this device became idle (if it is
  52          * indeed idle).
  53          */
  54         unsigned long idle_time;
  55         unsigned long last_update_time;
  56 };
  57
  58 static void iot_init(struct io_tracker *iot)
  59 {
  60         spin_lock_init(&iot->lock);
  61         iot->in_flight = 0ul;
  62         iot->idle_time = 0ul;
  63         iot->last_update_time = jiffies;
  64 }
  65
  66 static bool __iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  67 {
  68         if (iot->in_flight)
  69                 return false;
  70
  71         return time_after(jiffies, iot->idle_time + jifs);
  72 }
  73
  74 static bool iot_idle_for(struct io_tracker *iot, unsigned long jifs)
  75 {
  76         bool r;
  77         unsigned long flags;
  78
  79         spin_lock_irqsave(&iot->lock, flags);
  80         r = __iot_idle_for(iot, jifs);
  81         spin_unlock_irqrestore(&iot->lock, flags);
  82
  83         return r;
  84 }
  85
  86 static void iot_io_begin(struct io_tracker *iot, sector_t len)
  87 {
  88         unsigned long flags;
  89
  90         spin_lock_irqsave(&iot->lock, flags);
  91         iot->in_flight += len;
  92         spin_unlock_irqrestore(&iot->lock, flags);
  93 }
  94
  95 static void __iot_io_end(struct io_tracker *iot, sector_t len)
  96 {
  97         if (!len)
  98                 return;
  99
 100         iot->in_flight -= len;
 101         if (!iot->in_flight)
 102                 iot->idle_time = jiffies;
 103 }
 104
 105 static void iot_io_end(struct io_tracker *iot, sector_t len)
 106 {
 107         unsigned long flags;
 108
 109         spin_lock_irqsave(&iot->lock, flags);
 110         __iot_io_end(iot, len);
 111         spin_unlock_irqrestore(&iot->lock, flags);
 112 }
 113
 114 /*----------------------------------------------------------------*/
 115
 116 /*
 117  * Represents a chunk of future work.  'input' allows continuations to pass
 118  * values between themselves, typically error values.
 119  */
 120 struct continuation {
 121         struct work_struct ws;
 122         blk_status_t input;
 123 };
 124
 125 static inline void init_continuation(struct continuation *k,
 126                                      void (*fn)(struct work_struct *))
 127 {
 128         INIT_WORK(&k->ws, fn);
 129         k->input = 0;
 130 }
 131
 132 static inline void queue_continuation(struct workqueue_struct *wq,
 133                                       struct continuation *k)
 134 {
 135         queue_work(wq, &k->ws);
 136 }
 137
 138 /*----------------------------------------------------------------*/
 139
 140 /*
 141  * The batcher collects together pieces of work that need a particular
 142  * operation to occur before they can proceed (typically a commit).
 143  */
 144 struct batcher {
 145         /*
 146          * The operation that everyone is waiting for.
 147          */
 148         blk_status_t (*commit_op)(void *context);
 149         void *commit_context;
 150
 151         /*
 152          * This is how bios should be issued once the commit op is complete
 153          * (accounted_request).
 154          */
 155         void (*issue_op)(struct bio *bio, void *context);
 156         void *issue_context;
 157
 158         /*
 159          * Queued work gets put on here after commit.
 160          */
 161         struct workqueue_struct *wq;
 162
 163         spinlock_t lock;
 164         struct list_head work_items;
 165         struct bio_list bios;
 166         struct work_struct commit_work;
 167
 168         bool commit_scheduled;
 169 };
 170
 171 static void __commit(struct work_struct *_ws)
 172 {
 173         struct batcher *b = container_of(_ws, struct batcher, commit_work);
 174         blk_status_t r;
 175         unsigned long flags;
 176         struct list_head work_items;
 177         struct work_struct *ws, *tmp;
 178         struct continuation *k;
 179         struct bio *bio;
 180         struct bio_list bios;
 181
 182         INIT_LIST_HEAD(&work_items);
 183         bio_list_init(&bios);
 184
 185         /*
 186          * We have to grab these before the commit_op to avoid a race
 187          * condition.
 188          */
 189         spin_lock_irqsave(&b->lock, flags);
 190         list_splice_init(&b->work_items, &work_items);
 191         bio_list_merge(&bios, &b->bios);
 192         bio_list_init(&b->bios);
 193         b->commit_scheduled = false;
 194         spin_unlock_irqrestore(&b->lock, flags);
 195
 196         r = b->commit_op(b->commit_context);
 197
 198         list_for_each_entry_safe(ws, tmp, &work_items, entry) {
 199                 k = container_of(ws, struct continuation, ws);
 200                 k->input = r;
 201                 INIT_LIST_HEAD(&ws->entry); /* to avoid a WARN_ON */
 202                 queue_work(b->wq, ws);
 203         }
 204
 205         while ((bio = bio_list_pop(&bios))) {
 206                 if (r) {
 207                         bio->bi_status = r;
 208                         bio_endio(bio);
 209                 } else
 210                         b->issue_op(bio, b->issue_context);
 211         }
 212 }
 213
 214 static void batcher_init(struct batcher *b,
 215                          blk_status_t (*commit_op)(void *),
 216                          void *commit_context,
 217                          void (*issue_op)(struct bio *bio, void *),
 218                          void *issue_context,
 219                          struct workqueue_struct *wq)
 220 {
 221         b->commit_op = commit_op;
 222         b->commit_context = commit_context;
 223         b->issue_op = issue_op;
 224         b->issue_context = issue_context;
 225         b->wq = wq;
 226
 227         spin_lock_init(&b->lock);
 228         INIT_LIST_HEAD(&b->work_items);
 229         bio_list_init(&b->bios);
 230         INIT_WORK(&b->commit_work, __commit);
 231         b->commit_scheduled = false;
 232 }
 233
 234 static void async_commit(struct batcher *b)
 235 {
 236         queue_work(b->wq, &b->commit_work);
 237 }
 238
 239 static void continue_after_commit(struct batcher *b, struct continuation *k)
 240 {
 241         unsigned long flags;
 242         bool commit_scheduled;
 243
 244         spin_lock_irqsave(&b->lock, flags);
 245         commit_scheduled = b->commit_scheduled;
 246         list_add_tail(&k->ws.entry, &b->work_items);
 247         spin_unlock_irqrestore(&b->lock, flags);
 248
 249         if (commit_scheduled)
 250                 async_commit(b);
 251 }
 252
 253 /*
 254  * Bios are errored if commit failed.
 255  */
 256 static void issue_after_commit(struct batcher *b, struct bio *bio)
 257 {
 258        unsigned long flags;
 259        bool commit_scheduled;
 260
 261        spin_lock_irqsave(&b->lock, flags);
 262        commit_scheduled = b->commit_scheduled;
 263        bio_list_add(&b->bios, bio);
 264        spin_unlock_irqrestore(&b->lock, flags);
 265
 266        if (commit_scheduled)
 267                async_commit(b);
 268 }
 269
 270 /*
 271  * Call this if some urgent work is waiting for the commit to complete.
 272  */
 273 static void schedule_commit(struct batcher *b)
 274 {
 275         bool immediate;
 276         unsigned long flags;
 277
 278         spin_lock_irqsave(&b->lock, flags);
 279         immediate = !list_empty(&b->work_items) || !bio_list_empty(&b->bios);
 280         b->commit_scheduled = true;
 281         spin_unlock_irqrestore(&b->lock, flags);
 282
 283         if (immediate)
 284                 async_commit(b);
 285 }
 286
 287 /*
 288  * There are a couple of places where we let a bio run, but want to do some
 289  * work before calling its endio function.  We do this by temporarily
 290  * changing the endio fn.
 291  */
 292 struct dm_hook_info {
 293         bio_end_io_t *bi_end_io;
 294 };
 295
 296 static void dm_hook_bio(struct dm_hook_info *h, struct bio *bio,
 297                         bio_end_io_t *bi_end_io, void *bi_private)
 298 {
 299         h->bi_end_io = bio->bi_end_io;
 300
 301         bio->bi_end_io = bi_end_io;
 302         bio->bi_private = bi_private;
 303 }
 304
 305 static void dm_unhook_bio(struct dm_hook_info *h, struct bio *bio)
 306 {
 307         bio->bi_end_io = h->bi_end_io;
 308 }
 309
 310 /*----------------------------------------------------------------*/
 311
 312 #define MIGRATION_POOL_SIZE 128
 313 #define COMMIT_PERIOD HZ
 314 #define MIGRATION_COUNT_WINDOW 10
 315
 316 /*
 317  * The block size of the device holding cache data must be
 318  * between 32KB and 1GB.
 319  */
 320 #define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (32 * 1024 >> SECTOR_SHIFT)
 321 #define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
 322
 323 enum cache_metadata_mode {
 324         CM_WRITE,               /* metadata may be changed */
 325         CM_READ_ONLY,           /* metadata may not be changed */
 326         CM_FAIL
 327 };
 328
 329 enum cache_io_mode {
 330         /*
 331          * Data is written to cached blocks only.  These blocks are marked
 332          * dirty.  If you lose the cache device you will lose data.
 333          * Potential performance increase for both reads and writes.
 334          */
 335         CM_IO_WRITEBACK,
 336
 337         /*
 338          * Data is written to both cache and origin.  Blocks are never
 339          * dirty.  Potential performance benfit for reads only.
 340          */
 341         CM_IO_WRITETHROUGH,
 342
 343         /*
 344          * A degraded mode useful for various cache coherency situations
 345          * (eg, rolling back snapshots).  Reads and writes always go to the
 346          * origin.  If a write goes to a cached oblock, then the cache
 347          * block is invalidated.
 348          */
 349         CM_IO_PASSTHROUGH
 350 };
 351
 352 struct cache_features {
 353         enum cache_metadata_mode mode;
 354         enum cache_io_mode io_mode;
 355         unsigned metadata_version;
 356         bool discard_passdown:1;
 357 };
 358
 359 struct cache_stats {
 360         atomic_t read_hit;
 361         atomic_t read_miss;
 362         atomic_t write_hit;
 363         atomic_t write_miss;
 364         atomic_t demotion;
 365         atomic_t promotion;
 366         atomic_t writeback;
 367         atomic_t copies_avoided;
 368         atomic_t cache_cell_clash;
 369         atomic_t commit_count;
 370         atomic_t discard_count;
 371 };
 372
 373 struct cache {
 374         struct dm_target *ti;
 375         spinlock_t lock;
 376
 377         /*
 378          * Fields for converting from sectors to blocks.
 379          */
 380         int sectors_per_block_shift;
 381         sector_t sectors_per_block;
 382
 383         struct dm_cache_metadata *cmd;
 384
 385         /*
 386          * Metadata is written to this device.
 387          */
 388         struct dm_dev *metadata_dev;
 389
 390         /*
 391          * The slower of the two data devices.  Typically a spindle.
 392          */
 393         struct dm_dev *origin_dev;
 394
 395         /*
 396          * The faster of the two data devices.  Typically an SSD.
 397          */
 398         struct dm_dev *cache_dev;
 399
 400         /*
 401          * Size of the origin device in _complete_ blocks and native sectors.
 402          */
 403         dm_oblock_t origin_blocks;
 404         sector_t origin_sectors;
 405
 406         /*
 407          * Size of the cache device in blocks.
 408          */
 409         dm_cblock_t cache_size;
 410
 411         /*
 412          * Invalidation fields.
 413          */
 414         spinlock_t invalidation_lock;
 415         struct list_head invalidation_requests;
 416
 417         sector_t migration_threshold;
 418         wait_queue_head_t migration_wait;
 419         atomic_t nr_allocated_migrations;
 420
 421         /*
 422          * The number of in flight migrations that are performing
 423          * background io. eg, promotion, writeback.
 424          */
 425         atomic_t nr_io_migrations;
 426
 427         struct bio_list deferred_bios;
 428
 429         struct rw_semaphore quiesce_lock;
 430
 431         struct dm_target_callbacks callbacks;
 432
 433         /*
 434          * origin_blocks entries, discarded if set.
 435          */
 436         dm_dblock_t discard_nr_blocks;
 437         unsigned long *discard_bitset;
 438         uint32_t discard_block_size; /* a power of 2 times sectors per block */
 439
 440         /*
 441          * Rather than reconstructing the table line for the status we just
 442          * save it and regurgitate.
 443          */
 444         unsigned nr_ctr_args;
 445         const char **ctr_args;
 446
 447         struct dm_kcopyd_client *copier;
 448         struct work_struct deferred_bio_worker;
 449         struct work_struct migration_worker;
 450         struct workqueue_struct *wq;
 451         struct delayed_work waker;
 452         struct dm_bio_prison_v2 *prison;
 453
 454         /*
 455          * cache_size entries, dirty if set
 456          */
 457         unsigned long *dirty_bitset;
 458         atomic_t nr_dirty;
 459
 460         unsigned policy_nr_args;
 461         struct dm_cache_policy *policy;
 462
 463         /*
 464          * Cache features such as write-through.
 465          */
 466         struct cache_features features;
 467
 468         struct cache_stats stats;
 469
 470         bool need_tick_bio:1;
 471         bool sized:1;
 472         bool invalidate:1;
 473         bool commit_requested:1;
 474         bool loaded_mappings:1;
 475         bool loaded_discards:1;
 476
 477         struct rw_semaphore background_work_lock;
 478
 479         struct batcher committer;
 480         struct work_struct commit_ws;
 481
 482         struct io_tracker tracker;
 483
 484         mempool_t migration_pool;
 485
 486         struct bio_set bs;
 487 };
 488
 489 struct per_bio_data {
 490         bool tick:1;
 491         unsigned req_nr:2;
 492         struct dm_bio_prison_cell_v2 *cell;
 493         struct dm_hook_info hook_info;
 494         sector_t len;
 495 };
 496
 497 struct dm_cache_migration {
 498         struct continuation k;
 499         struct cache *cache;
 500
 501         struct policy_work *op;
 502         struct bio *overwrite_bio;
 503         struct dm_bio_prison_cell_v2 *cell;
 504
 505         dm_cblock_t invalidate_cblock;
 506         dm_oblock_t invalidate_oblock;
 507 };
 508
 509 /*----------------------------------------------------------------*/
 510
 511 static bool writethrough_mode(struct cache *cache)
 512 {
 513         return cache->features.io_mode == CM_IO_WRITETHROUGH;
 514 }
 515
 516 static bool writeback_mode(struct cache *cache)
 517 {
 518         return cache->features.io_mode == CM_IO_WRITEBACK;
 519 }
 520
 521 static inline bool passthrough_mode(struct cache *cache)
 522 {
 523         return unlikely(cache->features.io_mode == CM_IO_PASSTHROUGH);
 524 }
 525
 526 /*----------------------------------------------------------------*/
 527
 528 static void wake_deferred_bio_worker(struct cache *cache)
 529 {
 530         queue_work(cache->wq, &cache->deferred_bio_worker);
 531 }
 532
 533 static void wake_migration_worker(struct cache *cache)
 534 {
 535         if (passthrough_mode(cache))
 536                 return;
 537
 538         queue_work(cache->wq, &cache->migration_worker);
 539 }
 540
 541 /*----------------------------------------------------------------*/
 542
 543 static struct dm_bio_prison_cell_v2 *alloc_prison_cell(struct cache *cache)
 544 {
 545         return dm_bio_prison_alloc_cell_v2(cache->prison, GFP_NOIO);
 546 }
 547
 548 static void free_prison_cell(struct cache *cache, struct dm_bio_prison_cell_v2 *cell)
 549 {
 550         dm_bio_prison_free_cell_v2(cache->prison, cell);
 551 }
 552
 553 static struct dm_cache_migration *alloc_migration(struct cache *cache)
 554 {
 555         struct dm_cache_migration *mg;
 556
 557         mg = mempool_alloc(&cache->migration_pool, GFP_NOIO);
 558
 559         memset(mg, 0, sizeof(*mg));
 560
 561         mg->cache = cache;
 562         atomic_inc(&cache->nr_allocated_migrations);
 563
 564         return mg;
 565 }
 566
 567 static void free_migration(struct dm_cache_migration *mg)
 568 {
 569         struct cache *cache = mg->cache;
 570
 571         if (atomic_dec_and_test(&cache->nr_allocated_migrations))
 572                 wake_up(&cache->migration_wait);
 573
 574         mempool_free(mg, &cache->migration_pool);
 575 }
 576
 577 /*----------------------------------------------------------------*/
 578
 579 static inline dm_oblock_t oblock_succ(dm_oblock_t b)
 580 {
 581         return to_oblock(from_oblock(b) + 1ull);
 582 }
 583
 584 static void build_key(dm_oblock_t begin, dm_oblock_t end, struct dm_cell_key_v2 *key)
 585 {
 586         key->virtual = 0;
 587         key->dev = 0;
 588         key->block_begin = from_oblock(begin);
 589         key->block_end = from_oblock(end);
 590 }
 591
 592 /*
 593  * We have two lock levels.  Level 0, which is used to prevent WRITEs, and
 594  * level 1 which prevents *both* READs and WRITEs.
 595  */
 596 #define WRITE_LOCK_LEVEL 0
 597 #define READ_WRITE_LOCK_LEVEL 1
 598
 599 static unsigned lock_level(struct bio *bio)
 600 {
 601         return bio_data_dir(bio) == WRITE ?
 602                 WRITE_LOCK_LEVEL :
 603                 READ_WRITE_LOCK_LEVEL;
 604 }
 605
 606 /*----------------------------------------------------------------
 607  * Per bio data
 608  *--------------------------------------------------------------*/
 609
 610 static struct per_bio_data *get_per_bio_data(struct bio *bio)
 611 {
 612         struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 613         BUG_ON(!pb);
 614         return pb;
 615 }
 616
 617 static struct per_bio_data *init_per_bio_data(struct bio *bio)
 618 {
 619         struct per_bio_data *pb = get_per_bio_data(bio);
 620
 621         pb->tick = false;
 622         pb->req_nr = dm_bio_get_target_bio_nr(bio);
 623         pb->cell = NULL;
 624         pb->len = 0;
 625
 626         return pb;
 627 }
 628
 629 /*----------------------------------------------------------------*/
 630
 631 static void defer_bio(struct cache *cache, struct bio *bio)
 632 {
 633         unsigned long flags;
 634
 635         spin_lock_irqsave(&cache->lock, flags);
 636         bio_list_add(&cache->deferred_bios, bio);
 637         spin_unlock_irqrestore(&cache->lock, flags);
 638
 639         wake_deferred_bio_worker(cache);
 640 }
 641
 642 static void defer_bios(struct cache *cache, struct bio_list *bios)
 643 {
 644         unsigned long flags;
 645
 646         spin_lock_irqsave(&cache->lock, flags);
 647         bio_list_merge(&cache->deferred_bios, bios);
 648         bio_list_init(bios);
 649         spin_unlock_irqrestore(&cache->lock, flags);
 650
 651         wake_deferred_bio_worker(cache);
 652 }
 653
 654 /*----------------------------------------------------------------*/
 655
 656 static bool bio_detain_shared(struct cache *cache, dm_oblock_t oblock, struct bio *bio)
 657 {
 658         bool r;
 659         struct per_bio_data *pb;
 660         struct dm_cell_key_v2 key;
 661         dm_oblock_t end = to_oblock(from_oblock(oblock) + 1ULL);
 662         struct dm_bio_prison_cell_v2 *cell_prealloc, *cell;
 663
 664         cell_prealloc = alloc_prison_cell(cache); /* FIXME: allow wait if calling from worker */
 665
 666         build_key(oblock, end, &key);
 667         r = dm_cell_get_v2(cache->prison, &key, lock_level(bio), bio, cell_prealloc, &cell);
 668         if (!r) {
 669                 /*
 670                  * Failed to get the lock.
 671                  */
 672                 free_prison_cell(cache, cell_prealloc);
 673                 return r;
 674         }
 675
 676         if (cell != cell_prealloc)
 677                 free_prison_cell(cache, cell_prealloc);
 678
 679         pb = get_per_bio_data(bio);
 680         pb->cell = cell;
 681
 682         return r;
 683 }
 684
 685 /*----------------------------------------------------------------*/
 686
 687 static bool is_dirty(struct cache *cache, dm_cblock_t b)
 688 {
 689         return test_bit(from_cblock(b), cache->dirty_bitset);
 690 }
 691
 692 static void set_dirty(struct cache *cache, dm_cblock_t cblock)
 693 {
 694         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset)) {
 695                 atomic_inc(&cache->nr_dirty);
 696                 policy_set_dirty(cache->policy, cblock);
 697         }
 698 }
 699
 700 /*
 701  * These two are called when setting after migrations to force the policy
 702  * and dirty bitset to be in sync.
 703  */
 704 static void force_set_dirty(struct cache *cache, dm_cblock_t cblock)
 705 {
 706         if (!test_and_set_bit(from_cblock(cblock), cache->dirty_bitset))
 707                 atomic_inc(&cache->nr_dirty);
 708         policy_set_dirty(cache->policy, cblock);
 709 }
 710
 711 static void force_clear_dirty(struct cache *cache, dm_cblock_t cblock)
 712 {
 713         if (test_and_clear_bit(from_cblock(cblock), cache->dirty_bitset)) {
 714                 if (atomic_dec_return(&cache->nr_dirty) == 0)
 715                         dm_table_event(cache->ti->table);
 716         }
 717
 718         policy_clear_dirty(cache->policy, cblock);
 719 }
 720
 721 /*----------------------------------------------------------------*/
 722
 723 static bool block_size_is_power_of_two(struct cache *cache)
 724 {
 725         return cache->sectors_per_block_shift >= 0;
 726 }
 727
 728 /* gcc on ARM generates spurious references to __udivdi3 and __umoddi3 */
 729 #if defined(CONFIG_ARM) && __GNUC__ == 4 && __GNUC_MINOR__ <= 6
 730 __always_inline
 731 #endif
 732 static dm_block_t block_div(dm_block_t b, uint32_t n)
 733 {
 734         do_div(b, n);
 735
 736         return b;
 737 }
 738
 739 static dm_block_t oblocks_per_dblock(struct cache *cache)
 740 {
 741         dm_block_t oblocks = cache->discard_block_size;
 742
 743         if (block_size_is_power_of_two(cache))
 744                 oblocks >>= cache->sectors_per_block_shift;
 745         else
 746                 oblocks = block_div(oblocks, cache->sectors_per_block);
 747
 748         return oblocks;
 749 }
 750
 751 static dm_dblock_t oblock_to_dblock(struct cache *cache, dm_oblock_t oblock)
 752 {
 753         return to_dblock(block_div(from_oblock(oblock),
 754                                    oblocks_per_dblock(cache)));
 755 }
 756
 757 static void set_discard(struct cache *cache, dm_dblock_t b)
 758 {
 759         unsigned long flags;
 760
 761         BUG_ON(from_dblock(b) >= from_dblock(cache->discard_nr_blocks));
 762         atomic_inc(&cache->stats.discard_count);
 763
 764         spin_lock_irqsave(&cache->lock, flags);
 765         set_bit(from_dblock(b), cache->discard_bitset);
 766         spin_unlock_irqrestore(&cache->lock, flags);
 767 }
 768
 769 static void clear_discard(struct cache *cache, dm_dblock_t b)
 770 {
 771         unsigned long flags;
 772
 773         spin_lock_irqsave(&cache->lock, flags);
 774         clear_bit(from_dblock(b), cache->discard_bitset);
 775         spin_unlock_irqrestore(&cache->lock, flags);
 776 }
 777
 778 static bool is_discarded(struct cache *cache, dm_dblock_t b)
 779 {
 780         int r;
 781         unsigned long flags;
 782
 783         spin_lock_irqsave(&cache->lock, flags);
 784         r = test_bit(from_dblock(b), cache->discard_bitset);
 785         spin_unlock_irqrestore(&cache->lock, flags);
 786
 787         return r;
 788 }
 789
 790 static bool is_discarded_oblock(struct cache *cache, dm_oblock_t b)
 791 {
 792         int r;
 793         unsigned long flags;
 794
 795         spin_lock_irqsave(&cache->lock, flags);
 796         r = test_bit(from_dblock(oblock_to_dblock(cache, b)),
 797                      cache->discard_bitset);
 798         spin_unlock_irqrestore(&cache->lock, flags);
 799
 800         return r;
 801 }
 802
 803 /*----------------------------------------------------------------
 804  * Remapping
 805  *--------------------------------------------------------------*/
 806 static void remap_to_origin(struct cache *cache, struct bio *bio)
 807 {
 808         bio_set_dev(bio, cache->origin_dev->bdev);
 809 }
 810
 811 static void remap_to_cache(struct cache *cache, struct bio *bio,
 812                            dm_cblock_t cblock)
 813 {
 814         sector_t bi_sector = bio->bi_iter.bi_sector;
 815         sector_t block = from_cblock(cblock);
 816
 817         bio_set_dev(bio, cache->cache_dev->bdev);
 818         if (!block_size_is_power_of_two(cache))
 819                 bio->bi_iter.bi_sector =
 820                         (block * cache->sectors_per_block) +
 821                         sector_div(bi_sector, cache->sectors_per_block);
 822         else
 823                 bio->bi_iter.bi_sector =
 824                         (block << cache->sectors_per_block_shift) |
 825                         (bi_sector & (cache->sectors_per_block - 1));
 826 }
 827
 828 static void check_if_tick_bio_needed(struct cache *cache, struct bio *bio)
 829 {
 830         unsigned long flags;
 831         struct per_bio_data *pb;
 832
 833         spin_lock_irqsave(&cache->lock, flags);
 834         if (cache->need_tick_bio && !op_is_flush(bio->bi_opf) &&
 835             bio_op(bio) != REQ_OP_DISCARD) {
 836                 pb = get_per_bio_data(bio);
 837                 pb->tick = true;
 838                 cache->need_tick_bio = false;
 839         }
 840         spin_unlock_irqrestore(&cache->lock, flags);
 841 }
 842
 843 static void __remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 844                                             dm_oblock_t oblock, bool bio_has_pbd)
 845 {
 846         if (bio_has_pbd)
 847                 check_if_tick_bio_needed(cache, bio);
 848         remap_to_origin(cache, bio);
 849         if (bio_data_dir(bio) == WRITE)
 850                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 851 }
 852
 853 static void remap_to_origin_clear_discard(struct cache *cache, struct bio *bio,
 854                                           dm_oblock_t oblock)
 855 {
 856         // FIXME: check_if_tick_bio_needed() is called way too much through this interface
 857         __remap_to_origin_clear_discard(cache, bio, oblock, true);
 858 }
 859
 860 static void remap_to_cache_dirty(struct cache *cache, struct bio *bio,
 861                                  dm_oblock_t oblock, dm_cblock_t cblock)
 862 {
 863         check_if_tick_bio_needed(cache, bio);
 864         remap_to_cache(cache, bio, cblock);
 865         if (bio_data_dir(bio) == WRITE) {
 866                 set_dirty(cache, cblock);
 867                 clear_discard(cache, oblock_to_dblock(cache, oblock));
 868         }
 869 }
 870
 871 static dm_oblock_t get_bio_block(struct cache *cache, struct bio *bio)
 872 {
 873         sector_t block_nr = bio->bi_iter.bi_sector;
 874
 875         if (!block_size_is_power_of_two(cache))
 876                 (void) sector_div(block_nr, cache->sectors_per_block);
 877         else
 878                 block_nr >>= cache->sectors_per_block_shift;
 879
 880         return to_oblock(block_nr);
 881 }
 882
 883 static bool accountable_bio(struct cache *cache, struct bio *bio)
 884 {
 885         return bio_op(bio) != REQ_OP_DISCARD;
 886 }
 887
 888 static void accounted_begin(struct cache *cache, struct bio *bio)
 889 {
 890         struct per_bio_data *pb;
 891
 892         if (accountable_bio(cache, bio)) {
 893                 pb = get_per_bio_data(bio);
 894                 pb->len = bio_sectors(bio);
 895                 iot_io_begin(&cache->tracker, pb->len);
 896         }
 897 }
 898
 899 static void accounted_complete(struct cache *cache, struct bio *bio)
 900 {
 901         struct per_bio_data *pb = get_per_bio_data(bio);
 902
 903         iot_io_end(&cache->tracker, pb->len);
 904 }
 905
 906 static void accounted_request(struct cache *cache, struct bio *bio)
 907 {
 908         accounted_begin(cache, bio);
 909         generic_make_request(bio);
 910 }
 911
 912 static void issue_op(struct bio *bio, void *context)
 913 {
 914         struct cache *cache = context;
 915         accounted_request(cache, bio);
 916 }
 917
 918 /*
 919  * When running in writethrough mode we need to send writes to clean blocks
 920  * to both the cache and origin devices.  Clone the bio and send them in parallel.
 921  */
 922 static void remap_to_origin_and_cache(struct cache *cache, struct bio *bio,
 923                                       dm_oblock_t oblock, dm_cblock_t cblock)
 924 {
 925         struct bio *origin_bio = bio_clone_fast(bio, GFP_NOIO, &cache->bs);
 926
 927         BUG_ON(!origin_bio);
 928
 929         bio_chain(origin_bio, bio);
 930         /*
 931          * Passing false to __remap_to_origin_clear_discard() skips
 932          * all code that might use per_bio_data (since clone doesn't have it)
 933          */
 934         __remap_to_origin_clear_discard(cache, origin_bio, oblock, false);
 935         submit_bio(origin_bio);
 936
 937         remap_to_cache(cache, bio, cblock);
 938 }
 939
 940 /*----------------------------------------------------------------
 941  * Failure modes
 942  *--------------------------------------------------------------*/
 943 static enum cache_metadata_mode get_cache_mode(struct cache *cache)
 944 {
 945         return cache->features.mode;
 946 }
 947
 948 static const char *cache_device_name(struct cache *cache)
 949 {
 950         return dm_device_name(dm_table_get_md(cache->ti->table));
 951 }
 952
 953 static void notify_mode_switch(struct cache *cache, enum cache_metadata_mode mode)
 954 {
 955         const char *descs[] = {
 956                 "write",
 957                 "read-only",
 958                 "fail"
 959         };
 960
 961         dm_table_event(cache->ti->table);
 962         DMINFO("%s: switching cache to %s mode",
 963                cache_device_name(cache), descs[(int)mode]);
 964 }
 965
 966 static void set_cache_mode(struct cache *cache, enum cache_metadata_mode new_mode)
 967 {
 968         bool needs_check;
 969         enum cache_metadata_mode old_mode = get_cache_mode(cache);
 970
 971         if (dm_cache_metadata_needs_check(cache->cmd, &needs_check)) {
 972                 DMERR("%s: unable to read needs_check flag, setting failure mode.",
 973                       cache_device_name(cache));
 974                 new_mode = CM_FAIL;
 975         }
 976
 977         if (new_mode == CM_WRITE && needs_check) {
 978                 DMERR("%s: unable to switch cache to write mode until repaired.",
 979                       cache_device_name(cache));
 980                 if (old_mode != new_mode)
 981                         new_mode = old_mode;
 982                 else
 983                         new_mode = CM_READ_ONLY;
 984         }
 985
 986         /* Never move out of fail mode */
 987         if (old_mode == CM_FAIL)
 988                 new_mode = CM_FAIL;
 989
 990         switch (new_mode) {
 991         case CM_FAIL:
 992         case CM_READ_ONLY:
 993                 dm_cache_metadata_set_read_only(cache->cmd);
 994                 break;
 995
 996         case CM_WRITE:
 997                 dm_cache_metadata_set_read_write(cache->cmd);
 998                 break;
 999         }
1000
1001         cache->features.mode = new_mode;
1002
1003         if (new_mode != old_mode)
1004                 notify_mode_switch(cache, new_mode);
1005 }
1006
1007 static void abort_transaction(struct cache *cache)
1008 {
1009         const char *dev_name = cache_device_name(cache);
1010
1011         if (get_cache_mode(cache) >= CM_READ_ONLY)
1012                 return;
1013
1014         if (dm_cache_metadata_set_needs_check(cache->cmd)) {
1015                 DMERR("%s: failed to set 'needs_check' flag in metadata", dev_name);
1016                 set_cache_mode(cache, CM_FAIL);
1017         }
1018
1019         DMERR_LIMIT("%s: aborting current metadata transaction", dev_name);
1020         if (dm_cache_metadata_abort(cache->cmd)) {
1021                 DMERR("%s: failed to abort metadata transaction", dev_name);
1022                 set_cache_mode(cache, CM_FAIL);
1023         }
1024 }
1025
1026 static void metadata_operation_failed(struct cache *cache, const char *op, int r)
1027 {
1028         DMERR_LIMIT("%s: metadata operation '%s' failed: error = %d",
1029                     cache_device_name(cache), op, r);
1030         abort_transaction(cache);
1031         set_cache_mode(cache, CM_READ_ONLY);
1032 }
1033
1034 /*----------------------------------------------------------------*/
1035
1036 static void load_stats(struct cache *cache)
1037 {
1038         struct dm_cache_statistics stats;
1039
1040         dm_cache_metadata_get_stats(cache->cmd, &stats);
1041         atomic_set(&cache->stats.read_hit, stats.read_hits);
1042         atomic_set(&cache->stats.read_miss, stats.read_misses);
1043         atomic_set(&cache->stats.write_hit, stats.write_hits);
1044         atomic_set(&cache->stats.write_miss, stats.write_misses);
1045 }
1046
1047 static void save_stats(struct cache *cache)
1048 {
1049         struct dm_cache_statistics stats;
1050
1051         if (get_cache_mode(cache) >= CM_READ_ONLY)
1052                 return;
1053
1054         stats.read_hits = atomic_read(&cache->stats.read_hit);
1055         stats.read_misses = atomic_read(&cache->stats.read_miss);
1056         stats.write_hits = atomic_read(&cache->stats.write_hit);
1057         stats.write_misses = atomic_read(&cache->stats.write_miss);
1058
1059         dm_cache_metadata_set_stats(cache->cmd, &stats);
1060 }
1061
1062 static void update_stats(struct cache_stats *stats, enum policy_operation op)
1063 {
1064         switch (op) {
1065         case POLICY_PROMOTE:
1066                 atomic_inc(&stats->promotion);
1067                 break;
1068
1069         case POLICY_DEMOTE:
1070                 atomic_inc(&stats->demotion);
1071                 break;
1072
1073         case POLICY_WRITEBACK:
1074                 atomic_inc(&stats->writeback);
1075                 break;
1076         }
1077 }
1078
1079 /*----------------------------------------------------------------
1080  * Migration processing
1081  *
1082  * Migration covers moving data from the origin device to the cache, or
1083  * vice versa.
1084  *--------------------------------------------------------------*/
1085
1086 static void inc_io_migrations(struct cache *cache)
1087 {
1088         atomic_inc(&cache->nr_io_migrations);
1089 }
1090
1091 static void dec_io_migrations(struct cache *cache)
1092 {
1093         atomic_dec(&cache->nr_io_migrations);
1094 }
1095
1096 static bool discard_or_flush(struct bio *bio)
1097 {
1098         return bio_op(bio) == REQ_OP_DISCARD || op_is_flush(bio->bi_opf);
1099 }
1100
1101 static void calc_discard_block_range(struct cache *cache, struct bio *bio,
1102                                      dm_dblock_t *b, dm_dblock_t *e)
1103 {
1104         sector_t sb = bio->bi_iter.bi_sector;
1105         sector_t se = bio_end_sector(bio);
1106
1107         *b = to_dblock(dm_sector_div_up(sb, cache->discard_block_size));
1108
1109         if (se - sb < cache->discard_block_size)
1110                 *e = *b;
1111         else
1112                 *e = to_dblock(block_div(se, cache->discard_block_size));
1113 }
1114
1115 /*----------------------------------------------------------------*/
1116
1117 static void prevent_background_work(struct cache *cache)
1118 {
1119         lockdep_off();
1120         down_write(&cache->background_work_lock);
1121         lockdep_on();
1122 }
1123
1124 static void allow_background_work(struct cache *cache)
1125 {
1126         lockdep_off();
1127         up_write(&cache->background_work_lock);
1128         lockdep_on();
1129 }
1130
1131 static bool background_work_begin(struct cache *cache)
1132 {
1133         bool r;
1134
1135         lockdep_off();
1136         r = down_read_trylock(&cache->background_work_lock);
1137         lockdep_on();
1138
1139         return r;
1140 }
1141
1142 static void background_work_end(struct cache *cache)
1143 {
1144         lockdep_off();
1145         up_read(&cache->background_work_lock);
1146         lockdep_on();
1147 }
1148
1149 /*----------------------------------------------------------------*/
1150
1151 static bool bio_writes_complete_block(struct cache *cache, struct bio *bio)
1152 {
1153         return (bio_data_dir(bio) == WRITE) &&
1154                 (bio->bi_iter.bi_size == (cache->sectors_per_block << SECTOR_SHIFT));
1155 }
1156
1157 static bool optimisable_bio(struct cache *cache, struct bio *bio, dm_oblock_t block)
1158 {
1159         return writeback_mode(cache) &&
1160                 (is_discarded_oblock(cache, block) || bio_writes_complete_block(cache, bio));
1161 }
1162
1163 static void quiesce(struct dm_cache_migration *mg,
1164                     void (*continuation)(struct work_struct *))
1165 {
1166         init_continuation(&mg->k, continuation);
1167         dm_cell_quiesce_v2(mg->cache->prison, mg->cell, &mg->k.ws);
1168 }
1169
1170 static struct dm_cache_migration *ws_to_mg(struct work_struct *ws)
1171 {
1172         struct continuation *k = container_of(ws, struct continuation, ws);
1173         return container_of(k, struct dm_cache_migration, k);
1174 }
1175
1176 static void copy_complete(int read_err, unsigned long write_err, void *context)
1177 {
1178         struct dm_cache_migration *mg = container_of(context, struct dm_cache_migration, k);
1179
1180         if (read_err || write_err)
1181                 mg->k.input = BLK_STS_IOERR;
1182
1183         queue_continuation(mg->cache->wq, &mg->k);
1184 }
1185
1186 static void copy(struct dm_cache_migration *mg, bool promote)
1187 {
1188         struct dm_io_region o_region, c_region;
1189         struct cache *cache = mg->cache;
1190
1191         o_region.bdev = cache->origin_dev->bdev;
1192         o_region.sector = from_oblock(mg->op->oblock) * cache->sectors_per_block;
1193         o_region.count = cache->sectors_per_block;
1194
1195         c_region.bdev = cache->cache_dev->bdev;
1196         c_region.sector = from_cblock(mg->op->cblock) * cache->sectors_per_block;
1197         c_region.count = cache->sectors_per_block;
1198
1199         if (promote)
1200                 dm_kcopyd_copy(cache->copier, &o_region, 1, &c_region, 0, copy_complete, &mg->k);
1201         else
1202                 dm_kcopyd_copy(cache->copier, &c_region, 1, &o_region, 0, copy_complete, &mg->k);
1203 }
1204
1205 static void bio_drop_shared_lock(struct cache *cache, struct bio *bio)
1206 {
1207         struct per_bio_data *pb = get_per_bio_data(bio);
1208
1209         if (pb->cell && dm_cell_put_v2(cache->prison, pb->cell))
1210                 free_prison_cell(cache, pb->cell);
1211         pb->cell = NULL;
1212 }
1213
1214 static void overwrite_endio(struct bio *bio)
1215 {
1216         struct dm_cache_migration *mg = bio->bi_private;
1217         struct cache *cache = mg->cache;
1218         struct per_bio_data *pb = get_per_bio_data(bio);
1219
1220         dm_unhook_bio(&pb->hook_info, bio);
1221
1222         if (bio->bi_status)
1223                 mg->k.input = bio->bi_status;
1224
1225         queue_continuation(cache->wq, &mg->k);
1226 }
1227
1228 static void overwrite(struct dm_cache_migration *mg,
1229                       void (*continuation)(struct work_struct *))
1230 {
1231         struct bio *bio = mg->overwrite_bio;
1232         struct per_bio_data *pb = get_per_bio_data(bio);
1233
1234         dm_hook_bio(&pb->hook_info, bio, overwrite_endio, mg);
1235
1236         /*
1237          * The overwrite bio is part of the copy operation, as such it does
1238          * not set/clear discard or dirty flags.
1239          */
1240         if (mg->op->op == POLICY_PROMOTE)
1241                 remap_to_cache(mg->cache, bio, mg->op->cblock);
1242         else
1243                 remap_to_origin(mg->cache, bio);
1244
1245         init_continuation(&mg->k, continuation);
1246         accounted_request(mg->cache, bio);
1247 }
1248
1249 /*
1250  * Migration steps:
1251  *
1252  * 1) exclusive lock preventing WRITEs
1253  * 2) quiesce
1254  * 3) copy or issue overwrite bio
1255  * 4) upgrade to exclusive lock preventing READs and WRITEs
1256  * 5) quiesce
1257  * 6) update metadata and commit
1258  * 7) unlock
1259  */
1260 static void mg_complete(struct dm_cache_migration *mg, bool success)
1261 {
1262         struct bio_list bios;
1263         struct cache *cache = mg->cache;
1264         struct policy_work *op = mg->op;
1265         dm_cblock_t cblock = op->cblock;
1266
1267         if (success)
1268                 update_stats(&cache->stats, op->op);
1269
1270         switch (op->op) {
1271         case POLICY_PROMOTE:
1272                 clear_discard(cache, oblock_to_dblock(cache, op->oblock));
1273                 policy_complete_background_work(cache->policy, op, success);
1274
1275                 if (mg->overwrite_bio) {
1276                         if (success)
1277                                 force_set_dirty(cache, cblock);
1278                         else if (mg->k.input)
1279                                 mg->overwrite_bio->bi_status = mg->k.input;
1280                         else
1281                                 mg->overwrite_bio->bi_status = BLK_STS_IOERR;
1282                         bio_endio(mg->overwrite_bio);
1283                 } else {
1284                         if (success)
1285                                 force_clear_dirty(cache, cblock);
1286                         dec_io_migrations(cache);
1287                 }
1288                 break;
1289
1290         case POLICY_DEMOTE:
1291                 /*
1292                  * We clear dirty here to update the nr_dirty counter.
1293                  */
1294                 if (success)
1295                         force_clear_dirty(cache, cblock);
1296                 policy_complete_background_work(cache->policy, op, success);
1297                 dec_io_migrations(cache);
1298                 break;
1299
1300         case POLICY_WRITEBACK:
1301                 if (success)
1302                         force_clear_dirty(cache, cblock);
1303                 policy_complete_background_work(cache->policy, op, success);
1304                 dec_io_migrations(cache);
1305                 break;
1306         }
1307
1308         bio_list_init(&bios);
1309         if (mg->cell) {
1310                 if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1311                         free_prison_cell(cache, mg->cell);
1312         }
1313
1314         free_migration(mg);
1315         defer_bios(cache, &bios);
1316         wake_migration_worker(cache);
1317
1318         background_work_end(cache);
1319 }
1320
1321 static void mg_success(struct work_struct *ws)
1322 {
1323         struct dm_cache_migration *mg = ws_to_mg(ws);
1324         mg_complete(mg, mg->k.input == 0);
1325 }
1326
1327 static void mg_update_metadata(struct work_struct *ws)
1328 {
1329         int r;
1330         struct dm_cache_migration *mg = ws_to_mg(ws);
1331         struct cache *cache = mg->cache;
1332         struct policy_work *op = mg->op;
1333
1334         switch (op->op) {
1335         case POLICY_PROMOTE:
1336                 r = dm_cache_insert_mapping(cache->cmd, op->cblock, op->oblock);
1337                 if (r) {
1338                         DMERR_LIMIT("%s: migration failed; couldn't insert mapping",
1339                                     cache_device_name(cache));
1340                         metadata_operation_failed(cache, "dm_cache_insert_mapping", r);
1341
1342                         mg_complete(mg, false);
1343                         return;
1344                 }
1345                 mg_complete(mg, true);
1346                 break;
1347
1348         case POLICY_DEMOTE:
1349                 r = dm_cache_remove_mapping(cache->cmd, op->cblock);
1350                 if (r) {
1351                         DMERR_LIMIT("%s: migration failed; couldn't update on disk metadata",
1352                                     cache_device_name(cache));
1353                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1354
1355                         mg_complete(mg, false);
1356                         return;
1357                 }
1358
1359                 /*
1360                  * It would be nice if we only had to commit when a REQ_FLUSH
1361                  * comes through.  But there's one scenario that we have to
1362                  * look out for:
1363                  *
1364                  * - vblock x in a cache block
1365                  * - domotion occurs
1366                  * - cache block gets reallocated and over written
1367                  * - crash
1368                  *
1369                  * When we recover, because there was no commit the cache will
1370                  * rollback to having the data for vblock x in the cache block.
1371                  * But the cache block has since been overwritten, so it'll end
1372                  * up pointing to data that was never in 'x' during the history
1373                  * of the device.
1374                  *
1375                  * To avoid this issue we require a commit as part of the
1376                  * demotion operation.
1377                  */
1378                 init_continuation(&mg->k, mg_success);
1379                 continue_after_commit(&cache->committer, &mg->k);
1380                 schedule_commit(&cache->committer);
1381                 break;
1382
1383         case POLICY_WRITEBACK:
1384                 mg_complete(mg, true);
1385                 break;
1386         }
1387 }
1388
1389 static void mg_update_metadata_after_copy(struct work_struct *ws)
1390 {
1391         struct dm_cache_migration *mg = ws_to_mg(ws);
1392
1393         /*
1394          * Did the copy succeed?
1395          */
1396         if (mg->k.input)
1397                 mg_complete(mg, false);
1398         else
1399                 mg_update_metadata(ws);
1400 }
1401
1402 static void mg_upgrade_lock(struct work_struct *ws)
1403 {
1404         int r;
1405         struct dm_cache_migration *mg = ws_to_mg(ws);
1406
1407         /*
1408          * Did the copy succeed?
1409          */
1410         if (mg->k.input)
1411                 mg_complete(mg, false);
1412
1413         else {
1414                 /*
1415                  * Now we want the lock to prevent both reads and writes.
1416                  */
1417                 r = dm_cell_lock_promote_v2(mg->cache->prison, mg->cell,
1418                                             READ_WRITE_LOCK_LEVEL);
1419                 if (r < 0)
1420                         mg_complete(mg, false);
1421
1422                 else if (r)
1423                         quiesce(mg, mg_update_metadata);
1424
1425                 else
1426                         mg_update_metadata(ws);
1427         }
1428 }
1429
1430 static void mg_full_copy(struct work_struct *ws)
1431 {
1432         struct dm_cache_migration *mg = ws_to_mg(ws);
1433         struct cache *cache = mg->cache;
1434         struct policy_work *op = mg->op;
1435         bool is_policy_promote = (op->op == POLICY_PROMOTE);
1436
1437         if ((!is_policy_promote && !is_dirty(cache, op->cblock)) ||
1438             is_discarded_oblock(cache, op->oblock)) {
1439                 mg_upgrade_lock(ws);
1440                 return;
1441         }
1442
1443         init_continuation(&mg->k, mg_upgrade_lock);
1444         copy(mg, is_policy_promote);
1445 }
1446
1447 static void mg_copy(struct work_struct *ws)
1448 {
1449         struct dm_cache_migration *mg = ws_to_mg(ws);
1450
1451         if (mg->overwrite_bio) {
1452                 /*
1453                  * No exclusive lock was held when we last checked if the bio
1454                  * was optimisable.  So we have to check again in case things
1455                  * have changed (eg, the block may no longer be discarded).
1456                  */
1457                 if (!optimisable_bio(mg->cache, mg->overwrite_bio, mg->op->oblock)) {
1458                         /*
1459                          * Fallback to a real full copy after doing some tidying up.
1460                          */
1461                         bool rb = bio_detain_shared(mg->cache, mg->op->oblock, mg->overwrite_bio);
1462                         BUG_ON(rb); /* An exclussive lock must _not_ be held for this block */
1463                         mg->overwrite_bio = NULL;
1464                         inc_io_migrations(mg->cache);
1465                         mg_full_copy(ws);
1466                         return;
1467                 }
1468
1469                 /*
1470                  * It's safe to do this here, even though it's new data
1471                  * because all IO has been locked out of the block.
1472                  *
1473                  * mg_lock_writes() already took READ_WRITE_LOCK_LEVEL
1474                  * so _not_ using mg_upgrade_lock() as continutation.
1475                  */
1476                 overwrite(mg, mg_update_metadata_after_copy);
1477
1478         } else
1479                 mg_full_copy(ws);
1480 }
1481
1482 static int mg_lock_writes(struct dm_cache_migration *mg)
1483 {
1484         int r;
1485         struct dm_cell_key_v2 key;
1486         struct cache *cache = mg->cache;
1487         struct dm_bio_prison_cell_v2 *prealloc;
1488
1489         prealloc = alloc_prison_cell(cache);
1490
1491         /*
1492          * Prevent writes to the block, but allow reads to continue.
1493          * Unless we're using an overwrite bio, in which case we lock
1494          * everything.
1495          */
1496         build_key(mg->op->oblock, oblock_succ(mg->op->oblock), &key);
1497         r = dm_cell_lock_v2(cache->prison, &key,
1498                             mg->overwrite_bio ?  READ_WRITE_LOCK_LEVEL : WRITE_LOCK_LEVEL,
1499                             prealloc, &mg->cell);
1500         if (r < 0) {
1501                 free_prison_cell(cache, prealloc);
1502                 mg_complete(mg, false);
1503                 return r;
1504         }
1505
1506         if (mg->cell != prealloc)
1507                 free_prison_cell(cache, prealloc);
1508
1509         if (r == 0)
1510                 mg_copy(&mg->k.ws);
1511         else
1512                 quiesce(mg, mg_copy);
1513
1514         return 0;
1515 }
1516
1517 static int mg_start(struct cache *cache, struct policy_work *op, struct bio *bio)
1518 {
1519         struct dm_cache_migration *mg;
1520
1521         if (!background_work_begin(cache)) {
1522                 policy_complete_background_work(cache->policy, op, false);
1523                 return -EPERM;
1524         }
1525
1526         mg = alloc_migration(cache);
1527
1528         mg->op = op;
1529         mg->overwrite_bio = bio;
1530
1531         if (!bio)
1532                 inc_io_migrations(cache);
1533
1534         return mg_lock_writes(mg);
1535 }
1536
1537 /*----------------------------------------------------------------
1538  * invalidation processing
1539  *--------------------------------------------------------------*/
1540
1541 static void invalidate_complete(struct dm_cache_migration *mg, bool success)
1542 {
1543         struct bio_list bios;
1544         struct cache *cache = mg->cache;
1545
1546         bio_list_init(&bios);
1547         if (dm_cell_unlock_v2(cache->prison, mg->cell, &bios))
1548                 free_prison_cell(cache, mg->cell);
1549
1550         if (!success && mg->overwrite_bio)
1551                 bio_io_error(mg->overwrite_bio);
1552
1553         free_migration(mg);
1554         defer_bios(cache, &bios);
1555
1556         background_work_end(cache);
1557 }
1558
1559 static void invalidate_completed(struct work_struct *ws)
1560 {
1561         struct dm_cache_migration *mg = ws_to_mg(ws);
1562         invalidate_complete(mg, !mg->k.input);
1563 }
1564
1565 static int invalidate_cblock(struct cache *cache, dm_cblock_t cblock)
1566 {
1567         int r = policy_invalidate_mapping(cache->policy, cblock);
1568         if (!r) {
1569                 r = dm_cache_remove_mapping(cache->cmd, cblock);
1570                 if (r) {
1571                         DMERR_LIMIT("%s: invalidation failed; couldn't update on disk metadata",
1572                                     cache_device_name(cache));
1573                         metadata_operation_failed(cache, "dm_cache_remove_mapping", r);
1574                 }
1575
1576         } else if (r == -ENODATA) {
1577                 /*
1578                  * Harmless, already unmapped.
1579                  */
1580                 r = 0;
1581
1582         } else
1583                 DMERR("%s: policy_invalidate_mapping failed", cache_device_name(cache));
1584
1585         return r;
1586 }
1587
1588 static void invalidate_remove(struct work_struct *ws)
1589 {
1590         int r;
1591         struct dm_cache_migration *mg = ws_to_mg(ws);
1592         struct cache *cache = mg->cache;
1593
1594         r = invalidate_cblock(cache, mg->invalidate_cblock);
1595         if (r) {
1596                 invalidate_complete(mg, false);
1597                 return;
1598         }
1599
1600         init_continuation(&mg->k, invalidate_completed);
1601         continue_after_commit(&cache->committer, &mg->k);
1602         remap_to_origin_clear_discard(cache, mg->overwrite_bio, mg->invalidate_oblock);
1603         mg->overwrite_bio = NULL;
1604         schedule_commit(&cache->committer);
1605 }
1606
1607 static int invalidate_lock(struct dm_cache_migration *mg)
1608 {
1609         int r;
1610         struct dm_cell_key_v2 key;
1611         struct cache *cache = mg->cache;
1612         struct dm_bio_prison_cell_v2 *prealloc;
1613
1614         prealloc = alloc_prison_cell(cache);
1615
1616         build_key(mg->invalidate_oblock, oblock_succ(mg->invalidate_oblock), &key);
1617         r = dm_cell_lock_v2(cache->prison, &key,
1618                             READ_WRITE_LOCK_LEVEL, prealloc, &mg->cell);
1619         if (r < 0) {
1620                 free_prison_cell(cache, prealloc);
1621                 invalidate_complete(mg, false);
1622                 return r;
1623         }
1624
1625         if (mg->cell != prealloc)
1626                 free_prison_cell(cache, prealloc);
1627
1628         if (r)
1629                 quiesce(mg, invalidate_remove);
1630
1631         else {
1632                 /*
1633                  * We can't call invalidate_remove() directly here because we
1634                  * might still be in request context.
1635                  */
1636                 init_continuation(&mg->k, invalidate_remove);
1637                 queue_work(cache->wq, &mg->k.ws);
1638         }
1639
1640         return 0;
1641 }
1642
1643 static int invalidate_start(struct cache *cache, dm_cblock_t cblock,
1644                             dm_oblock_t oblock, struct bio *bio)
1645 {
1646         struct dm_cache_migration *mg;
1647
1648         if (!background_work_begin(cache))
1649                 return -EPERM;
1650
1651         mg = alloc_migration(cache);
1652
1653         mg->overwrite_bio = bio;
1654         mg->invalidate_cblock = cblock;
1655         mg->invalidate_oblock = oblock;
1656
1657         return invalidate_lock(mg);
1658 }
1659
1660 /*----------------------------------------------------------------
1661  * bio processing
1662  *--------------------------------------------------------------*/
1663
1664 enum busy {
1665         IDLE,
1666         BUSY
1667 };
1668
1669 static enum busy spare_migration_bandwidth(struct cache *cache)
1670 {
1671         bool idle = iot_idle_for(&cache->tracker, HZ);
1672         sector_t current_volume = (atomic_read(&cache->nr_io_migrations) + 1) *
1673                 cache->sectors_per_block;
1674
1675         if (idle && current_volume <= cache->migration_threshold)
1676                 return IDLE;
1677         else
1678                 return BUSY;
1679 }
1680
1681 static void inc_hit_counter(struct cache *cache, struct bio *bio)
1682 {
1683         atomic_inc(bio_data_dir(bio) == READ ?
1684                    &cache->stats.read_hit : &cache->stats.write_hit);
1685 }
1686
1687 static void inc_miss_counter(struct cache *cache, struct bio *bio)
1688 {
1689         atomic_inc(bio_data_dir(bio) == READ ?
1690                    &cache->stats.read_miss : &cache->stats.write_miss);
1691 }
1692
1693 /*----------------------------------------------------------------*/
1694
1695 static int map_bio(struct cache *cache, struct bio *bio, dm_oblock_t block,
1696                    bool *commit_needed)
1697 {
1698         int r, data_dir;
1699         bool rb, background_queued;
1700         dm_cblock_t cblock;
1701
1702         *commit_needed = false;
1703
1704         rb = bio_detain_shared(cache, block, bio);
1705         if (!rb) {
1706                 /*
1707                  * An exclusive lock is held for this block, so we have to
1708                  * wait.  We set the commit_needed flag so the current
1709                  * transaction will be committed asap, allowing this lock
1710                  * to be dropped.
1711                  */
1712                 *commit_needed = true;
1713                 return DM_MAPIO_SUBMITTED;
1714         }
1715
1716         data_dir = bio_data_dir(bio);
1717
1718         if (optimisable_bio(cache, bio, block)) {
1719                 struct policy_work *op = NULL;
1720
1721                 r = policy_lookup_with_work(cache->policy, block, &cblock, data_dir, true, &op);
1722                 if (unlikely(r && r != -ENOENT)) {
1723                         DMERR_LIMIT("%s: policy_lookup_with_work() failed with r = %d",
1724                                     cache_device_name(cache), r);
1725                         bio_io_error(bio);
1726                         return DM_MAPIO_SUBMITTED;
1727                 }
1728
1729                 if (r == -ENOENT && op) {
1730                         bio_drop_shared_lock(cache, bio);
1731                         BUG_ON(op->op != POLICY_PROMOTE);
1732                         mg_start(cache, op, bio);
1733                         return DM_MAPIO_SUBMITTED;
1734                 }
1735         } else {
1736                 r = policy_lookup(cache->policy, block, &cblock, data_dir, false, &background_queued);
1737                 if (unlikely(r && r != -ENOENT)) {
1738                         DMERR_LIMIT("%s: policy_lookup() failed with r = %d",
1739                                     cache_device_name(cache), r);
1740                         bio_io_error(bio);
1741                         return DM_MAPIO_SUBMITTED;
1742                 }
1743
1744                 if (background_queued)
1745                         wake_migration_worker(cache);
1746         }
1747
1748         if (r == -ENOENT) {
1749                 struct per_bio_data *pb = get_per_bio_data(bio);
1750
1751                 /*
1752                  * Miss.
1753                  */
1754                 inc_miss_counter(cache, bio);
1755                 if (pb->req_nr == 0) {
1756                         accounted_begin(cache, bio);
1757                         remap_to_origin_clear_discard(cache, bio, block);
1758                 } else {
1759                         /*
1760                          * This is a duplicate writethrough io that is no
1761                          * longer needed because the block has been demoted.
1762                          */
1763                         bio_endio(bio);
1764                         return DM_MAPIO_SUBMITTED;
1765                 }
1766         } else {
1767                 /*
1768                  * Hit.
1769                  */
1770                 inc_hit_counter(cache, bio);
1771
1772                 /*
1773                  * Passthrough always maps to the origin, invalidating any
1774                  * cache blocks that are written to.
1775                  */
1776                 if (passthrough_mode(cache)) {
1777                         if (bio_data_dir(bio) == WRITE) {
1778                                 bio_drop_shared_lock(cache, bio);
1779                                 atomic_inc(&cache->stats.demotion);
1780                                 invalidate_start(cache, cblock, block, bio);
1781                         } else
1782                                 remap_to_origin_clear_discard(cache, bio, block);
1783                 } else {
1784                         if (bio_data_dir(bio) == WRITE && writethrough_mode(cache) &&
1785                             !is_dirty(cache, cblock)) {
1786                                 remap_to_origin_and_cache(cache, bio, block, cblock);
1787                                 accounted_begin(cache, bio);
1788                         } else
1789                                 remap_to_cache_dirty(cache, bio, block, cblock);
1790                 }
1791         }
1792
1793         /*
1794          * dm core turns FUA requests into a separate payload and FLUSH req.
1795          */
1796         if (bio->bi_opf & REQ_FUA) {
1797                 /*
1798                  * issue_after_commit will call accounted_begin a second time.  So
1799                  * we call accounted_complete() to avoid double accounting.
1800                  */
1801                 accounted_complete(cache, bio);
1802                 issue_after_commit(&cache->committer, bio);
1803                 *commit_needed = true;
1804                 return DM_MAPIO_SUBMITTED;
1805         }
1806
1807         return DM_MAPIO_REMAPPED;
1808 }
1809
1810 static bool process_bio(struct cache *cache, struct bio *bio)
1811 {
1812         bool commit_needed;
1813
1814         if (map_bio(cache, bio, get_bio_block(cache, bio), &commit_needed) == DM_MAPIO_REMAPPED)
1815                 generic_make_request(bio);
1816
1817         return commit_needed;
1818 }
1819
1820 /*
1821  * A non-zero return indicates read_only or fail_io mode.
1822  */
1823 static int commit(struct cache *cache, bool clean_shutdown)
1824 {
1825         int r;
1826
1827         if (get_cache_mode(cache) >= CM_READ_ONLY)
1828                 return -EINVAL;
1829
1830         atomic_inc(&cache->stats.commit_count);
1831         r = dm_cache_commit(cache->cmd, clean_shutdown);
1832         if (r)
1833                 metadata_operation_failed(cache, "dm_cache_commit", r);
1834
1835         return r;
1836 }
1837
1838 /*
1839  * Used by the batcher.
1840  */
1841 static blk_status_t commit_op(void *context)
1842 {
1843         struct cache *cache = context;
1844
1845         if (dm_cache_changed_this_transaction(cache->cmd))
1846                 return errno_to_blk_status(commit(cache, false));
1847
1848         return 0;
1849 }
1850
1851 /*----------------------------------------------------------------*/
1852
1853 static bool process_flush_bio(struct cache *cache, struct bio *bio)
1854 {
1855         struct per_bio_data *pb = get_per_bio_data(bio);
1856
1857         if (!pb->req_nr)
1858                 remap_to_origin(cache, bio);
1859         else
1860                 remap_to_cache(cache, bio, 0);
1861
1862         issue_after_commit(&cache->committer, bio);
1863         return true;
1864 }
1865
1866 static bool process_discard_bio(struct cache *cache, struct bio *bio)
1867 {
1868         dm_dblock_t b, e;
1869
1870         // FIXME: do we need to lock the region?  Or can we just assume the
1871         // user wont be so foolish as to issue discard concurrently with
1872         // other IO?
1873         calc_discard_block_range(cache, bio, &b, &e);
1874         while (b != e) {
1875                 set_discard(cache, b);
1876                 b = to_dblock(from_dblock(b) + 1);
1877         }
1878
1879         if (cache->features.discard_passdown) {
1880                 remap_to_origin(cache, bio);
1881                 generic_make_request(bio);
1882         } else
1883                 bio_endio(bio);
1884
1885         return false;
1886 }
1887
1888 static void process_deferred_bios(struct work_struct *ws)
1889 {
1890         struct cache *cache = container_of(ws, struct cache, deferred_bio_worker);
1891
1892         unsigned long flags;
1893         bool commit_needed = false;
1894         struct bio_list bios;
1895         struct bio *bio;
1896
1897         bio_list_init(&bios);
1898
1899         spin_lock_irqsave(&cache->lock, flags);
1900         bio_list_merge(&bios, &cache->deferred_bios);
1901         bio_list_init(&cache->deferred_bios);
1902         spin_unlock_irqrestore(&cache->lock, flags);
1903
1904         while ((bio = bio_list_pop(&bios))) {
1905                 if (bio->bi_opf & REQ_PREFLUSH)
1906                         commit_needed = process_flush_bio(cache, bio) || commit_needed;
1907
1908                 else if (bio_op(bio) == REQ_OP_DISCARD)
1909                         commit_needed = process_discard_bio(cache, bio) || commit_needed;
1910
1911                 else
1912                         commit_needed = process_bio(cache, bio) || commit_needed;
1913         }
1914
1915         if (commit_needed)
1916                 schedule_commit(&cache->committer);
1917 }
1918
1919 /*----------------------------------------------------------------
1920  * Main worker loop
1921  *--------------------------------------------------------------*/
1922
1923 static void requeue_deferred_bios(struct cache *cache)
1924 {
1925         struct bio *bio;
1926         struct bio_list bios;
1927
1928         bio_list_init(&bios);
1929         bio_list_merge(&bios, &cache->deferred_bios);
1930         bio_list_init(&cache->deferred_bios);
1931
1932         while ((bio = bio_list_pop(&bios))) {
1933                 bio->bi_status = BLK_STS_DM_REQUEUE;
1934                 bio_endio(bio);
1935         }
1936 }
1937
1938 /*
1939  * We want to commit periodically so that not too much
1940  * unwritten metadata builds up.
1941  */
1942 static void do_waker(struct work_struct *ws)
1943 {
1944         struct cache *cache = container_of(to_delayed_work(ws), struct cache, waker);
1945
1946         policy_tick(cache->policy, true);
1947         wake_migration_worker(cache);
1948         schedule_commit(&cache->committer);
1949         queue_delayed_work(cache->wq, &cache->waker, COMMIT_PERIOD);
1950 }
1951
1952 static void check_migrations(struct work_struct *ws)
1953 {
1954         int r;
1955         struct policy_work *op;
1956         struct cache *cache = container_of(ws, struct cache, migration_worker);
1957         enum busy b;
1958
1959         for (;;) {
1960                 b = spare_migration_bandwidth(cache);
1961
1962                 r = policy_get_background_work(cache->policy, b == IDLE, &op);
1963                 if (r == -ENODATA)
1964                         break;
1965
1966                 if (r) {
1967                         DMERR_LIMIT("%s: policy_background_work failed",
1968                                     cache_device_name(cache));
1969                         break;
1970                 }
1971
1972                 r = mg_start(cache, op, NULL);
1973                 if (r)
1974                         break;
1975         }
1976 }
1977
1978 /*----------------------------------------------------------------
1979  * Target methods
1980  *--------------------------------------------------------------*/
1981
1982 /*
1983  * This function gets called on the error paths of the constructor, so we
1984  * have to cope with a partially initialised struct.
1985  */
1986 static void destroy(struct cache *cache)
1987 {
1988         unsigned i;
1989
1990         mempool_exit(&cache->migration_pool);
1991
1992         if (cache->prison)
1993                 dm_bio_prison_destroy_v2(cache->prison);
1994
1995         if (cache->wq)
1996                 destroy_workqueue(cache->wq);
1997
1998         if (cache->dirty_bitset)
1999                 free_bitset(cache->dirty_bitset);
2000
2001         if (cache->discard_bitset)
2002                 free_bitset(cache->discard_bitset);
2003
2004         if (cache->copier)
2005                 dm_kcopyd_client_destroy(cache->copier);
2006
2007         if (cache->cmd)
2008                 dm_cache_metadata_close(cache->cmd);
2009
2010         if (cache->metadata_dev)
2011                 dm_put_device(cache->ti, cache->metadata_dev);
2012
2013         if (cache->origin_dev)
2014                 dm_put_device(cache->ti, cache->origin_dev);
2015
2016         if (cache->cache_dev)
2017                 dm_put_device(cache->ti, cache->cache_dev);
2018
2019         if (cache->policy)
2020                 dm_cache_policy_destroy(cache->policy);
2021
2022         for (i = 0; i < cache->nr_ctr_args ; i++)
2023                 kfree(cache->ctr_args[i]);
2024         kfree(cache->ctr_args);
2025
2026         bioset_exit(&cache->bs);
2027
2028         kfree(cache);
2029 }
2030
2031 static void cache_dtr(struct dm_target *ti)
2032 {
2033         struct cache *cache = ti->private;
2034
2035         destroy(cache);
2036 }
2037
2038 static sector_t get_dev_size(struct dm_dev *dev)
2039 {
2040         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
2041 }
2042
2043 /*----------------------------------------------------------------*/
2044
2045 /*
2046  * Construct a cache device mapping.
2047  *
2048  * cache <metadata dev> <cache dev> <origin dev> <block size>
2049  *       <#feature args> [<feature arg>]*
2050  *       <policy> <#policy args> [<policy arg>]*
2051  *
2052  * metadata dev    : fast device holding the persistent metadata
2053  * cache dev       : fast device holding cached data blocks
2054  * origin dev      : slow device holding original data blocks
2055  * block size      : cache unit size in sectors
2056  *
2057  * #feature args   : number of feature arguments passed
2058  * feature args    : writethrough.  (The default is writeback.)
2059  *
2060  * policy          : the replacement policy to use
2061  * #policy args    : an even number of policy arguments corresponding
2062  *                   to key/value pairs passed to the policy
2063  * policy args     : key/value pairs passed to the policy
2064  *                   E.g. 'sequential_threshold 1024'
2065  *                   See cache-policies.txt for details.
2066  *
2067  * Optional feature arguments are:
2068  *   writethrough  : write through caching that prohibits cache block
2069  *                   content from being different from origin block content.
2070  *                   Without this argument, the default behaviour is to write
2071  *                   back cache block contents later for performance reasons,
2072  *                   so they may differ from the corresponding origin blocks.
2073  */
2074 struct cache_args {
2075         struct dm_target *ti;
2076
2077         struct dm_dev *metadata_dev;
2078
2079         struct dm_dev *cache_dev;
2080         sector_t cache_sectors;
2081
2082         struct dm_dev *origin_dev;
2083         sector_t origin_sectors;
2084
2085         uint32_t block_size;
2086
2087         const char *policy_name;
2088         int policy_argc;
2089         const char **policy_argv;
2090
2091         struct cache_features features;
2092 };
2093
2094 static void destroy_cache_args(struct cache_args *ca)
2095 {
2096         if (ca->metadata_dev)
2097                 dm_put_device(ca->ti, ca->metadata_dev);
2098
2099         if (ca->cache_dev)
2100                 dm_put_device(ca->ti, ca->cache_dev);
2101
2102         if (ca->origin_dev)
2103                 dm_put_device(ca->ti, ca->origin_dev);
2104
2105         kfree(ca);
2106 }
2107
2108 static bool at_least_one_arg(struct dm_arg_set *as, char **error)
2109 {
2110         if (!as->argc) {
2111                 *error = "Insufficient args";
2112                 return false;
2113         }
2114
2115         return true;
2116 }
2117
2118 static int parse_metadata_dev(struct cache_args *ca, struct dm_arg_set *as,
2119                               char **error)
2120 {
2121         int r;
2122         sector_t metadata_dev_size;
2123         char b[BDEVNAME_SIZE];
2124
2125         if (!at_least_one_arg(as, error))
2126                 return -EINVAL;
2127
2128         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2129                           &ca->metadata_dev);
2130         if (r) {
2131                 *error = "Error opening metadata device";
2132                 return r;
2133         }
2134
2135         metadata_dev_size = get_dev_size(ca->metadata_dev);
2136         if (metadata_dev_size > DM_CACHE_METADATA_MAX_SECTORS_WARNING)
2137                 DMWARN("Metadata device %s is larger than %u sectors: excess space will not be used.",
2138                        bdevname(ca->metadata_dev->bdev, b), THIN_METADATA_MAX_SECTORS);
2139
2140         return 0;
2141 }
2142
2143 static int parse_cache_dev(struct cache_args *ca, struct dm_arg_set *as,
2144                            char **error)
2145 {
2146         int r;
2147
2148         if (!at_least_one_arg(as, error))
2149                 return -EINVAL;
2150
2151         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2152                           &ca->cache_dev);
2153         if (r) {
2154                 *error = "Error opening cache device";
2155                 return r;
2156         }
2157         ca->cache_sectors = get_dev_size(ca->cache_dev);
2158
2159         return 0;
2160 }
2161
2162 static int parse_origin_dev(struct cache_args *ca, struct dm_arg_set *as,
2163                             char **error)
2164 {
2165         int r;
2166
2167         if (!at_least_one_arg(as, error))
2168                 return -EINVAL;
2169
2170         r = dm_get_device(ca->ti, dm_shift_arg(as), FMODE_READ | FMODE_WRITE,
2171                           &ca->origin_dev);
2172         if (r) {
2173                 *error = "Error opening origin device";
2174                 return r;
2175         }
2176
2177         ca->origin_sectors = get_dev_size(ca->origin_dev);
2178         if (ca->ti->len > ca->origin_sectors) {
2179                 *error = "Device size larger than cached device";
2180                 return -EINVAL;
2181         }
2182
2183         return 0;
2184 }
2185
2186 static int parse_block_size(struct cache_args *ca, struct dm_arg_set *as,
2187                             char **error)
2188 {
2189         unsigned long block_size;
2190
2191         if (!at_least_one_arg(as, error))
2192                 return -EINVAL;
2193
2194         if (kstrtoul(dm_shift_arg(as), 10, &block_size) || !block_size ||
2195             block_size < DATA_DEV_BLOCK_SIZE_MIN_SECTORS ||
2196             block_size > DATA_DEV_BLOCK_SIZE_MAX_SECTORS ||
2197             block_size & (DATA_DEV_BLOCK_SIZE_MIN_SECTORS - 1)) {
2198                 *error = "Invalid data block size";
2199                 return -EINVAL;
2200         }
2201
2202         if (block_size > ca->cache_sectors) {
2203                 *error = "Data block size is larger than the cache device";
2204                 return -EINVAL;
2205         }
2206
2207         ca->block_size = block_size;
2208
2209         return 0;
2210 }
2211
2212 static void init_features(struct cache_features *cf)
2213 {
2214         cf->mode = CM_WRITE;
2215         cf->io_mode = CM_IO_WRITEBACK;
2216         cf->metadata_version = 1;
2217         cf->discard_passdown = true;
2218 }
2219
2220 static int parse_features(struct cache_args *ca, struct dm_arg_set *as,
2221                           char **error)
2222 {
2223         static const struct dm_arg _args[] = {
2224                 {0, 3, "Invalid number of cache feature arguments"},
2225         };
2226
2227         int r, mode_ctr = 0;
2228         unsigned argc;
2229         const char *arg;
2230         struct cache_features *cf = &ca->features;
2231
2232         init_features(cf);
2233
2234         r = dm_read_arg_group(_args, as, &argc, error);
2235         if (r)
2236                 return -EINVAL;
2237
2238         while (argc--) {
2239                 arg = dm_shift_arg(as);
2240
2241                 if (!strcasecmp(arg, "writeback")) {
2242                         cf->io_mode = CM_IO_WRITEBACK;
2243                         mode_ctr++;
2244                 }
2245
2246                 else if (!strcasecmp(arg, "writethrough")) {
2247                         cf->io_mode = CM_IO_WRITETHROUGH;
2248                         mode_ctr++;
2249                 }
2250
2251                 else if (!strcasecmp(arg, "passthrough")) {
2252                         cf->io_mode = CM_IO_PASSTHROUGH;
2253                         mode_ctr++;
2254                 }
2255
2256                 else if (!strcasecmp(arg, "metadata2"))
2257                         cf->metadata_version = 2;
2258
2259                 else if (!strcasecmp(arg, "no_discard_passdown"))
2260                         cf->discard_passdown = false;
2261
2262                 else {
2263                         *error = "Unrecognised cache feature requested";
2264                         return -EINVAL;
2265                 }
2266         }
2267
2268         if (mode_ctr > 1) {
2269                 *error = "Duplicate cache io_mode features requested";
2270                 return -EINVAL;
2271         }
2272
2273         return 0;
2274 }
2275
2276 static int parse_policy(struct cache_args *ca, struct dm_arg_set *as,
2277                         char **error)
2278 {
2279         static const struct dm_arg _args[] = {
2280                 {0, 1024, "Invalid number of policy arguments"},
2281         };
2282
2283         int r;
2284
2285         if (!at_least_one_arg(as, error))
2286                 return -EINVAL;
2287
2288         ca->policy_name = dm_shift_arg(as);
2289
2290         r = dm_read_arg_group(_args, as, &ca->policy_argc, error);
2291         if (r)
2292                 return -EINVAL;
2293
2294         ca->policy_argv = (const char **)as->argv;
2295         dm_consume_args(as, ca->policy_argc);
2296
2297         return 0;
2298 }
2299
2300 static int parse_cache_args(struct cache_args *ca, int argc, char **argv,
2301                             char **error)
2302 {
2303         int r;
2304         struct dm_arg_set as;
2305
2306         as.argc = argc;
2307         as.argv = argv;
2308
2309         r = parse_metadata_dev(ca, &as, error);
2310         if (r)
2311                 return r;
2312
2313         r = parse_cache_dev(ca, &as, error);
2314         if (r)
2315                 return r;
2316
2317         r = parse_origin_dev(ca, &as, error);
2318         if (r)
2319                 return r;
2320
2321         r = parse_block_size(ca, &as, error);
2322         if (r)
2323                 return r;
2324
2325         r = parse_features(ca, &as, error);
2326         if (r)
2327                 return r;
2328
2329         r = parse_policy(ca, &as, error);
2330         if (r)
2331                 return r;
2332
2333         return 0;
2334 }
2335
2336 /*----------------------------------------------------------------*/
2337
2338 static struct kmem_cache *migration_cache;
2339
2340 #define NOT_CORE_OPTION 1
2341
2342 static int process_config_option(struct cache *cache, const char *key, const char *value)
2343 {
2344         unsigned long tmp;
2345
2346         if (!strcasecmp(key, "migration_threshold")) {
2347                 if (kstrtoul(value, 10, &tmp))
2348                         return -EINVAL;
2349
2350                 cache->migration_threshold = tmp;
2351                 return 0;
2352         }
2353
2354         return NOT_CORE_OPTION;
2355 }
2356
2357 static int set_config_value(struct cache *cache, const char *key, const char *value)
2358 {
2359         int r = process_config_option(cache, key, value);
2360
2361         if (r == NOT_CORE_OPTION)
2362                 r = policy_set_config_value(cache->policy, key, value);
2363
2364         if (r)
2365                 DMWARN("bad config value for %s: %s", key, value);
2366
2367         return r;
2368 }
2369
2370 static int set_config_values(struct cache *cache, int argc, const char **argv)
2371 {
2372         int r = 0;
2373
2374         if (argc & 1) {
2375                 DMWARN("Odd number of policy arguments given but they should be <key> <value> pairs.");
2376                 return -EINVAL;
2377         }
2378
2379         while (argc) {
2380                 r = set_config_value(cache, argv[0], argv[1]);
2381                 if (r)
2382                         break;
2383
2384                 argc -= 2;
2385                 argv += 2;
2386         }
2387
2388         return r;
2389 }
2390
2391 static int create_cache_policy(struct cache *cache, struct cache_args *ca,
2392                                char **error)
2393 {
2394         struct dm_cache_policy *p = dm_cache_policy_create(ca->policy_name,
2395                                                            cache->cache_size,
2396                                                            cache->origin_sectors,
2397                                                            cache->sectors_per_block);
2398         if (IS_ERR(p)) {
2399                 *error = "Error creating cache's policy";
2400                 return PTR_ERR(p);
2401         }
2402         cache->policy = p;
2403         BUG_ON(!cache->policy);
2404
2405         return 0;
2406 }
2407
2408 /*
2409  * We want the discard block size to be at least the size of the cache
2410  * block size and have no more than 2^14 discard blocks across the origin.
2411  */
2412 #define MAX_DISCARD_BLOCKS (1 << 14)
2413
2414 static bool too_many_discard_blocks(sector_t discard_block_size,
2415                                     sector_t origin_size)
2416 {
2417         (void) sector_div(origin_size, discard_block_size);
2418
2419         return origin_size > MAX_DISCARD_BLOCKS;
2420 }
2421
2422 static sector_t calculate_discard_block_size(sector_t cache_block_size,
2423                                              sector_t origin_size)
2424 {
2425         sector_t discard_block_size = cache_block_size;
2426
2427         if (origin_size)
2428                 while (too_many_discard_blocks(discard_block_size, origin_size))
2429                         discard_block_size *= 2;
2430
2431         return discard_block_size;
2432 }
2433
2434 static void set_cache_size(struct cache *cache, dm_cblock_t size)
2435 {
2436         dm_block_t nr_blocks = from_cblock(size);
2437
2438         if (nr_blocks > (1 << 20) && cache->cache_size != size)
2439                 DMWARN_LIMIT("You have created a cache device with a lot of individual cache blocks (%llu)\n"
2440                              "All these mappings can consume a lot of kernel memory, and take some time to read/write.\n"
2441                              "Please consider increasing the cache block size to reduce the overall cache block count.",
2442                              (unsigned long long) nr_blocks);
2443
2444         cache->cache_size = size;
2445 }
2446
2447 static int is_congested(struct dm_dev *dev, int bdi_bits)
2448 {
2449         struct request_queue *q = bdev_get_queue(dev->bdev);
2450         return bdi_congested(q->backing_dev_info, bdi_bits);
2451 }
2452
2453 static int cache_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
2454 {
2455         struct cache *cache = container_of(cb, struct cache, callbacks);
2456
2457         return is_congested(cache->origin_dev, bdi_bits) ||
2458                 is_congested(cache->cache_dev, bdi_bits);
2459 }
2460
2461 #define DEFAULT_MIGRATION_THRESHOLD 2048
2462
2463 static int cache_create(struct cache_args *ca, struct cache **result)
2464 {
2465         int r = 0;
2466         char **error = &ca->ti->error;
2467         struct cache *cache;
2468         struct dm_target *ti = ca->ti;
2469         dm_block_t origin_blocks;
2470         struct dm_cache_metadata *cmd;
2471         bool may_format = ca->features.mode == CM_WRITE;
2472
2473         cache = kzalloc(sizeof(*cache), GFP_KERNEL);
2474         if (!cache)
2475                 return -ENOMEM;
2476
2477         cache->ti = ca->ti;
2478         ti->private = cache;
2479         ti->num_flush_bios = 2;
2480         ti->flush_supported = true;
2481
2482         ti->num_discard_bios = 1;
2483         ti->discards_supported = true;
2484
2485         ti->per_io_data_size = sizeof(struct per_bio_data);
2486
2487         cache->features = ca->features;
2488         if (writethrough_mode(cache)) {
2489                 /* Create bioset for writethrough bios issued to origin */
2490                 r = bioset_init(&cache->bs, BIO_POOL_SIZE, 0, 0);
2491                 if (r)
2492                         goto bad;
2493         }
2494
2495         cache->callbacks.congested_fn = cache_is_congested;
2496         dm_table_add_target_callbacks(ti->table, &cache->callbacks);
2497
2498         cache->metadata_dev = ca->metadata_dev;
2499         cache->origin_dev = ca->origin_dev;
2500         cache->cache_dev = ca->cache_dev;
2501
2502         ca->metadata_dev = ca->origin_dev = ca->cache_dev = NULL;
2503
2504         origin_blocks = cache->origin_sectors = ca->origin_sectors;
2505         origin_blocks = block_div(origin_blocks, ca->block_size);
2506         cache->origin_blocks = to_oblock(origin_blocks);
2507
2508         cache->sectors_per_block = ca->block_size;
2509         if (dm_set_target_max_io_len(ti, cache->sectors_per_block)) {
2510                 r = -EINVAL;
2511                 goto bad;
2512         }
2513
2514         if (ca->block_size & (ca->block_size - 1)) {
2515                 dm_block_t cache_size = ca->cache_sectors;
2516
2517                 cache->sectors_per_block_shift = -1;
2518                 cache_size = block_div(cache_size, ca->block_size);
2519                 set_cache_size(cache, to_cblock(cache_size));
2520         } else {
2521                 cache->sectors_per_block_shift = __ffs(ca->block_size);
2522                 set_cache_size(cache, to_cblock(ca->cache_sectors >> cache->sectors_per_block_shift));
2523         }
2524
2525         r = create_cache_policy(cache, ca, error);
2526         if (r)
2527                 goto bad;
2528
2529         cache->policy_nr_args = ca->policy_argc;
2530         cache->migration_threshold = DEFAULT_MIGRATION_THRESHOLD;
2531
2532         r = set_config_values(cache, ca->policy_argc, ca->policy_argv);
2533         if (r) {
2534                 *error = "Error setting cache policy's config values";
2535                 goto bad;
2536         }
2537
2538         cmd = dm_cache_metadata_open(cache->metadata_dev->bdev,
2539                                      ca->block_size, may_format,
2540                                      dm_cache_policy_get_hint_size(cache->policy),
2541                                      ca->features.metadata_version);
2542         if (IS_ERR(cmd)) {
2543                 *error = "Error creating metadata object";
2544                 r = PTR_ERR(cmd);
2545                 goto bad;
2546         }
2547         cache->cmd = cmd;
2548         set_cache_mode(cache, CM_WRITE);
2549         if (get_cache_mode(cache) != CM_WRITE) {
2550                 *error = "Unable to get write access to metadata, please check/repair metadata.";
2551                 r = -EINVAL;
2552                 goto bad;
2553         }
2554
2555         if (passthrough_mode(cache)) {
2556                 bool all_clean;
2557
2558                 r = dm_cache_metadata_all_clean(cache->cmd, &all_clean);
2559                 if (r) {
2560                         *error = "dm_cache_metadata_all_clean() failed";
2561                         goto bad;
2562                 }
2563
2564                 if (!all_clean) {
2565                         *error = "Cannot enter passthrough mode unless all blocks are clean";
2566                         r = -EINVAL;
2567                         goto bad;
2568                 }
2569
2570                 policy_allow_migrations(cache->policy, false);
2571         }
2572
2573         spin_lock_init(&cache->lock);
2574         bio_list_init(&cache->deferred_bios);
2575         atomic_set(&cache->nr_allocated_migrations, 0);
2576         atomic_set(&cache->nr_io_migrations, 0);
2577         init_waitqueue_head(&cache->migration_wait);
2578
2579         r = -ENOMEM;
2580         atomic_set(&cache->nr_dirty, 0);
2581         cache->dirty_bitset = alloc_bitset(from_cblock(cache->cache_size));
2582         if (!cache->dirty_bitset) {
2583                 *error = "could not allocate dirty bitset";
2584                 goto bad;
2585         }
2586         clear_bitset(cache->dirty_bitset, from_cblock(cache->cache_size));
2587
2588         cache->discard_block_size =
2589                 calculate_discard_block_size(cache->sectors_per_block,
2590                                              cache->origin_sectors);
2591         cache->discard_nr_blocks = to_dblock(dm_sector_div_up(cache->origin_sectors,
2592                                                               cache->discard_block_size));
2593         cache->discard_bitset = alloc_bitset(from_dblock(cache->discard_nr_blocks));
2594         if (!cache->discard_bitset) {
2595                 *error = "could not allocate discard bitset";
2596                 goto bad;
2597         }
2598         clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
2599
2600         cache->copier = dm_kcopyd_client_create(&dm_kcopyd_throttle);
2601         if (IS_ERR(cache->copier)) {
2602                 *error = "could not create kcopyd client";
2603                 r = PTR_ERR(cache->copier);
2604                 goto bad;
2605         }
2606
2607         cache->wq = alloc_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM, 0);
2608         if (!cache->wq) {
2609                 *error = "could not create workqueue for metadata object";
2610                 goto bad;
2611         }
2612         INIT_WORK(&cache->deferred_bio_worker, process_deferred_bios);
2613         INIT_WORK(&cache->migration_worker, check_migrations);
2614         INIT_DELAYED_WORK(&cache->waker, do_waker);
2615
2616         cache->prison = dm_bio_prison_create_v2(cache->wq);
2617         if (!cache->prison) {
2618                 *error = "could not create bio prison";
2619                 goto bad;
2620         }
2621
2622         r = mempool_init_slab_pool(&cache->migration_pool, MIGRATION_POOL_SIZE,
2623                                    migration_cache);
2624         if (r) {
2625                 *error = "Error creating cache's migration mempool";
2626                 goto bad;
2627         }
2628
2629         cache->need_tick_bio = true;
2630         cache->sized = false;
2631         cache->invalidate = false;
2632         cache->commit_requested = false;
2633         cache->loaded_mappings = false;
2634         cache->loaded_discards = false;
2635
2636         load_stats(cache);
2637
2638         atomic_set(&cache->stats.demotion, 0);
2639         atomic_set(&cache->stats.promotion, 0);
2640         atomic_set(&cache->stats.copies_avoided, 0);
2641         atomic_set(&cache->stats.cache_cell_clash, 0);
2642         atomic_set(&cache->stats.commit_count, 0);
2643         atomic_set(&cache->stats.discard_count, 0);
2644
2645         spin_lock_init(&cache->invalidation_lock);
2646         INIT_LIST_HEAD(&cache->invalidation_requests);
2647
2648         batcher_init(&cache->committer, commit_op, cache,
2649                      issue_op, cache, cache->wq);
2650         iot_init(&cache->tracker);
2651
2652         init_rwsem(&cache->background_work_lock);
2653         prevent_background_work(cache);
2654
2655         *result = cache;
2656         return 0;
2657 bad:
2658         destroy(cache);
2659         return r;
2660 }
2661
2662 static int copy_ctr_args(struct cache *cache, int argc, const char **argv)
2663 {
2664         unsigned i;
2665         const char **copy;
2666
2667         copy = kcalloc(argc, sizeof(*copy), GFP_KERNEL);
2668         if (!copy)
2669                 return -ENOMEM;
2670         for (i = 0; i < argc; i++) {
2671                 copy[i] = kstrdup(argv[i], GFP_KERNEL);
2672                 if (!copy[i]) {
2673                         while (i--)
2674                                 kfree(copy[i]);
2675                         kfree(copy);
2676                         return -ENOMEM;
2677                 }
2678         }
2679
2680         cache->nr_ctr_args = argc;
2681         cache->ctr_args = copy;
2682
2683         return 0;
2684 }
2685
2686 static int cache_ctr(struct dm_target *ti, unsigned argc, char **argv)
2687 {
2688         int r = -EINVAL;
2689         struct cache_args *ca;
2690         struct cache *cache = NULL;
2691
2692         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2693         if (!ca) {
2694                 ti->error = "Error allocating memory for cache";
2695                 return -ENOMEM;
2696         }
2697         ca->ti = ti;
2698
2699         r = parse_cache_args(ca, argc, argv, &ti->error);
2700         if (r)
2701                 goto out;
2702
2703         r = cache_create(ca, &cache);
2704         if (r)
2705                 goto out;
2706
2707         r = copy_ctr_args(cache, argc - 3, (const char **)argv + 3);
2708         if (r) {
2709                 destroy(cache);
2710                 goto out;
2711         }
2712
2713         ti->private = cache;
2714 out:
2715         destroy_cache_args(ca);
2716         return r;
2717 }
2718
2719 /*----------------------------------------------------------------*/
2720
2721 static int cache_map(struct dm_target *ti, struct bio *bio)
2722 {
2723         struct cache *cache = ti->private;
2724
2725         int r;
2726         bool commit_needed;
2727         dm_oblock_t block = get_bio_block(cache, bio);
2728
2729         init_per_bio_data(bio);
2730         if (unlikely(from_oblock(block) >= from_oblock(cache->origin_blocks))) {
2731                 /*
2732                  * This can only occur if the io goes to a partial block at
2733                  * the end of the origin device.  We don't cache these.
2734                  * Just remap to the origin and carry on.
2735                  */
2736                 remap_to_origin(cache, bio);
2737                 accounted_begin(cache, bio);
2738                 return DM_MAPIO_REMAPPED;
2739         }
2740
2741         if (discard_or_flush(bio)) {
2742                 defer_bio(cache, bio);
2743                 return DM_MAPIO_SUBMITTED;
2744         }
2745
2746         r = map_bio(cache, bio, block, &commit_needed);
2747         if (commit_needed)
2748                 schedule_commit(&cache->committer);
2749
2750         return r;
2751 }
2752
2753 static int cache_end_io(struct dm_target *ti, struct bio *bio, blk_status_t *error)
2754 {
2755         struct cache *cache = ti->private;
2756         unsigned long flags;
2757         struct per_bio_data *pb = get_per_bio_data(bio);
2758
2759         if (pb->tick) {
2760                 policy_tick(cache->policy, false);
2761
2762                 spin_lock_irqsave(&cache->lock, flags);
2763                 cache->need_tick_bio = true;
2764                 spin_unlock_irqrestore(&cache->lock, flags);
2765         }
2766
2767         bio_drop_shared_lock(cache, bio);
2768         accounted_complete(cache, bio);
2769
2770         return DM_ENDIO_DONE;
2771 }
2772
2773 static int write_dirty_bitset(struct cache *cache)
2774 {
2775         int r;
2776
2777         if (get_cache_mode(cache) >= CM_READ_ONLY)
2778                 return -EINVAL;
2779
2780         r = dm_cache_set_dirty_bits(cache->cmd, from_cblock(cache->cache_size), cache->dirty_bitset);
2781         if (r)
2782                 metadata_operation_failed(cache, "dm_cache_set_dirty_bits", r);
2783
2784         return r;
2785 }
2786
2787 static int write_discard_bitset(struct cache *cache)
2788 {
2789         unsigned i, r;
2790
2791         if (get_cache_mode(cache) >= CM_READ_ONLY)
2792                 return -EINVAL;
2793
2794         r = dm_cache_discard_bitset_resize(cache->cmd, cache->discard_block_size,
2795                                            cache->discard_nr_blocks);
2796         if (r) {
2797                 DMERR("%s: could not resize on-disk discard bitset", cache_device_name(cache));
2798                 metadata_operation_failed(cache, "dm_cache_discard_bitset_resize", r);
2799                 return r;
2800         }
2801
2802         for (i = 0; i < from_dblock(cache->discard_nr_blocks); i++) {
2803                 r = dm_cache_set_discard(cache->cmd, to_dblock(i),
2804                                          is_discarded(cache, to_dblock(i)));
2805                 if (r) {
2806                         metadata_operation_failed(cache, "dm_cache_set_discard", r);
2807                         return r;
2808                 }
2809         }
2810
2811         return 0;
2812 }
2813
2814 static int write_hints(struct cache *cache)
2815 {
2816         int r;
2817
2818         if (get_cache_mode(cache) >= CM_READ_ONLY)
2819                 return -EINVAL;
2820
2821         r = dm_cache_write_hints(cache->cmd, cache->policy);
2822         if (r) {
2823                 metadata_operation_failed(cache, "dm_cache_write_hints", r);
2824                 return r;
2825         }
2826
2827         return 0;
2828 }
2829
2830 /*
2831  * returns true on success
2832  */
2833 static bool sync_metadata(struct cache *cache)
2834 {
2835         int r1, r2, r3, r4;
2836
2837         r1 = write_dirty_bitset(cache);
2838         if (r1)
2839                 DMERR("%s: could not write dirty bitset", cache_device_name(cache));
2840
2841         r2 = write_discard_bitset(cache);
2842         if (r2)
2843                 DMERR("%s: could not write discard bitset", cache_device_name(cache));
2844
2845         save_stats(cache);
2846
2847         r3 = write_hints(cache);
2848         if (r3)
2849                 DMERR("%s: could not write hints", cache_device_name(cache));
2850
2851         /*
2852          * If writing the above metadata failed, we still commit, but don't
2853          * set the clean shutdown flag.  This will effectively force every
2854          * dirty bit to be set on reload.
2855          */
2856         r4 = commit(cache, !r1 && !r2 && !r3);
2857         if (r4)
2858                 DMERR("%s: could not write cache metadata", cache_device_name(cache));
2859
2860         return !r1 && !r2 && !r3 && !r4;
2861 }
2862
2863 static void cache_postsuspend(struct dm_target *ti)
2864 {
2865         struct cache *cache = ti->private;
2866
2867         prevent_background_work(cache);
2868         BUG_ON(atomic_read(&cache->nr_io_migrations));
2869
2870         cancel_delayed_work(&cache->waker);
2871         flush_workqueue(cache->wq);
2872         WARN_ON(cache->tracker.in_flight);
2873
2874         /*
2875          * If it's a flush suspend there won't be any deferred bios, so this
2876          * call is harmless.
2877          */
2878         requeue_deferred_bios(cache);
2879
2880         if (get_cache_mode(cache) == CM_WRITE)
2881                 (void) sync_metadata(cache);
2882 }
2883
2884 static int load_mapping(void *context, dm_oblock_t oblock, dm_cblock_t cblock,
2885                         bool dirty, uint32_t hint, bool hint_valid)
2886 {
2887         int r;
2888         struct cache *cache = context;
2889
2890         if (dirty) {
2891                 set_bit(from_cblock(cblock), cache->dirty_bitset);
2892                 atomic_inc(&cache->nr_dirty);
2893         } else
2894                 clear_bit(from_cblock(cblock), cache->dirty_bitset);
2895
2896         r = policy_load_mapping(cache->policy, oblock, cblock, dirty, hint, hint_valid);
2897         if (r)
2898                 return r;
2899
2900         return 0;
2901 }
2902
2903 /*
2904  * The discard block size in the on disk metadata is not
2905  * neccessarily the same as we're currently using.  So we have to
2906  * be careful to only set the discarded attribute if we know it
2907  * covers a complete block of the new size.
2908  */
2909 struct discard_load_info {
2910         struct cache *cache;
2911
2912         /*
2913          * These blocks are sized using the on disk dblock size, rather
2914          * than the current one.
2915          */
2916         dm_block_t block_size;
2917         dm_block_t discard_begin, discard_end;
2918 };
2919
2920 static void discard_load_info_init(struct cache *cache,
2921                                    struct discard_load_info *li)
2922 {
2923         li->cache = cache;
2924         li->discard_begin = li->discard_end = 0;
2925 }
2926
2927 static void set_discard_range(struct discard_load_info *li)
2928 {
2929         sector_t b, e;
2930
2931         if (li->discard_begin == li->discard_end)
2932                 return;
2933
2934         /*
2935          * Convert to sectors.
2936          */
2937         b = li->discard_begin * li->block_size;
2938         e = li->discard_end * li->block_size;
2939
2940         /*
2941          * Then convert back to the current dblock size.
2942          */
2943         b = dm_sector_div_up(b, li->cache->discard_block_size);
2944         sector_div(e, li->cache->discard_block_size);
2945
2946         /*
2947          * The origin may have shrunk, so we need to check we're still in
2948          * bounds.
2949          */
2950         if (e > from_dblock(li->cache->discard_nr_blocks))
2951                 e = from_dblock(li->cache->discard_nr_blocks);
2952
2953         for (; b < e; b++)
2954                 set_discard(li->cache, to_dblock(b));
2955 }
2956
2957 static int load_discard(void *context, sector_t discard_block_size,
2958                         dm_dblock_t dblock, bool discard)
2959 {
2960         struct discard_load_info *li = context;
2961
2962         li->block_size = discard_block_size;
2963
2964         if (discard) {
2965                 if (from_dblock(dblock) == li->discard_end)
2966                         /*
2967                          * We're already in a discard range, just extend it.
2968                          */
2969                         li->discard_end = li->discard_end + 1ULL;
2970
2971                 else {
2972                         /*
2973                          * Emit the old range and start a new one.
2974                          */
2975                         set_discard_range(li);
2976                         li->discard_begin = from_dblock(dblock);
2977                         li->discard_end = li->discard_begin + 1ULL;
2978                 }
2979         } else {
2980                 set_discard_range(li);
2981                 li->discard_begin = li->discard_end = 0;
2982         }
2983
2984         return 0;
2985 }
2986
2987 static dm_cblock_t get_cache_dev_size(struct cache *cache)
2988 {
2989         sector_t size = get_dev_size(cache->cache_dev);
2990         (void) sector_div(size, cache->sectors_per_block);
2991         return to_cblock(size);
2992 }
2993
2994 static bool can_resize(struct cache *cache, dm_cblock_t new_size)
2995 {
2996         if (from_cblock(new_size) > from_cblock(cache->cache_size)) {
2997                 if (cache->sized) {
2998                         DMERR("%s: unable to extend cache due to missing cache table reload",
2999                               cache_device_name(cache));
3000                         return false;
3001                 }
3002         }
3003
3004         /*
3005          * We can't drop a dirty block when shrinking the cache.
3006          */
3007         while (from_cblock(new_size) < from_cblock(cache->cache_size)) {
3008                 new_size = to_cblock(from_cblock(new_size) + 1);
3009                 if (is_dirty(cache, new_size)) {
3010                         DMERR("%s: unable to shrink cache; cache block %llu is dirty",
3011                               cache_device_name(cache),
3012                               (unsigned long long) from_cblock(new_size));
3013                         return false;
3014                 }
3015         }
3016
3017         return true;
3018 }
3019
3020 static int resize_cache_dev(struct cache *cache, dm_cblock_t new_size)
3021 {
3022         int r;
3023
3024         r = dm_cache_resize(cache->cmd, new_size);
3025         if (r) {
3026                 DMERR("%s: could not resize cache metadata", cache_device_name(cache));
3027                 metadata_operation_failed(cache, "dm_cache_resize", r);
3028                 return r;
3029         }
3030
3031         set_cache_size(cache, new_size);
3032
3033         return 0;
3034 }
3035
3036 static int cache_preresume(struct dm_target *ti)
3037 {
3038         int r = 0;
3039         struct cache *cache = ti->private;
3040         dm_cblock_t csize = get_cache_dev_size(cache);
3041
3042         /*
3043          * Check to see if the cache has resized.
3044          */
3045         if (!cache->sized) {
3046                 r = resize_cache_dev(cache, csize);
3047                 if (r)
3048                         return r;
3049
3050                 cache->sized = true;
3051
3052         } else if (csize != cache->cache_size) {
3053                 if (!can_resize(cache, csize))
3054                         return -EINVAL;
3055
3056                 r = resize_cache_dev(cache, csize);
3057                 if (r)
3058                         return r;
3059         }
3060
3061         if (!cache->loaded_mappings) {
3062                 r = dm_cache_load_mappings(cache->cmd, cache->policy,
3063                                            load_mapping, cache);
3064                 if (r) {
3065                         DMERR("%s: could not load cache mappings", cache_device_name(cache));
3066                         metadata_operation_failed(cache, "dm_cache_load_mappings", r);
3067                         return r;
3068                 }
3069
3070                 cache->loaded_mappings = true;
3071         }
3072
3073         if (!cache->loaded_discards) {
3074                 struct discard_load_info li;
3075
3076                 /*
3077                  * The discard bitset could have been resized, or the
3078                  * discard block size changed.  To be safe we start by
3079                  * setting every dblock to not discarded.
3080                  */
3081                 clear_bitset(cache->discard_bitset, from_dblock(cache->discard_nr_blocks));
3082
3083                 discard_load_info_init(cache, &li);
3084                 r = dm_cache_load_discards(cache->cmd, load_discard, &li);
3085                 if (r) {
3086                         DMERR("%s: could not load origin discards", cache_device_name(cache));
3087                         metadata_operation_failed(cache, "dm_cache_load_discards", r);
3088                         return r;
3089                 }
3090                 set_discard_range(&li);
3091
3092                 cache->loaded_discards = true;
3093         }
3094
3095         return r;
3096 }
3097
3098 static void cache_resume(struct dm_target *ti)
3099 {
3100         struct cache *cache = ti->private;
3101
3102         cache->need_tick_bio = true;
3103         allow_background_work(cache);
3104         do_waker(&cache->waker.work);
3105 }
3106
3107 static void emit_flags(struct cache *cache, char *result,
3108                        unsigned maxlen, ssize_t *sz_ptr)
3109 {
3110         ssize_t sz = *sz_ptr;
3111         struct cache_features *cf = &cache->features;
3112         unsigned count = (cf->metadata_version == 2) + !cf->discard_passdown + 1;
3113
3114         DMEMIT("%u ", count);
3115
3116         if (cf->metadata_version == 2)
3117                 DMEMIT("metadata2 ");
3118
3119         if (writethrough_mode(cache))
3120                 DMEMIT("writethrough ");
3121
3122         else if (passthrough_mode(cache))
3123                 DMEMIT("passthrough ");
3124
3125         else if (writeback_mode(cache))
3126                 DMEMIT("writeback ");
3127
3128         else {
3129                 DMEMIT("unknown ");
3130                 DMERR("%s: internal error: unknown io mode: %d",
3131                       cache_device_name(cache), (int) cf->io_mode);
3132         }
3133
3134         if (!cf->discard_passdown)
3135                 DMEMIT("no_discard_passdown ");
3136
3137         *sz_ptr = sz;
3138 }
3139
3140 /*
3141  * Status format:
3142  *
3143  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
3144  * <cache block size> <#used cache blocks>/<#total cache blocks>
3145  * <#read hits> <#read misses> <#write hits> <#write misses>
3146  * <#demotions> <#promotions> <#dirty>
3147  * <#features> <features>*
3148  * <#core args> <core args>
3149  * <policy name> <#policy args> <policy args>* <cache metadata mode> <needs_check>
3150  */
3151 static void cache_status(struct dm_target *ti, status_type_t type,
3152                          unsigned status_flags, char *result, unsigned maxlen)
3153 {
3154         int r = 0;
3155         unsigned i;
3156         ssize_t sz = 0;
3157         dm_block_t nr_free_blocks_metadata = 0;
3158         dm_block_t nr_blocks_metadata = 0;
3159         char buf[BDEVNAME_SIZE];
3160         struct cache *cache = ti->private;
3161         dm_cblock_t residency;
3162         bool needs_check;
3163
3164         switch (type) {
3165         case STATUSTYPE_INFO:
3166                 if (get_cache_mode(cache) == CM_FAIL) {
3167                         DMEMIT("Fail");
3168                         break;
3169                 }
3170
3171                 /* Commit to ensure statistics aren't out-of-date */
3172                 if (!(status_flags & DM_STATUS_NOFLUSH_FLAG) && !dm_suspended(ti))
3173                         (void) commit(cache, false);
3174
3175                 r = dm_cache_get_free_metadata_block_count(cache->cmd, &nr_free_blocks_metadata);
3176                 if (r) {
3177                         DMERR("%s: dm_cache_get_free_metadata_block_count returned %d",
3178                               cache_device_name(cache), r);
3179                         goto err;
3180                 }
3181
3182                 r = dm_cache_get_metadata_dev_size(cache->cmd, &nr_blocks_metadata);
3183                 if (r) {
3184                         DMERR("%s: dm_cache_get_metadata_dev_size returned %d",
3185                               cache_device_name(cache), r);
3186                         goto err;
3187                 }
3188
3189                 residency = policy_residency(cache->policy);
3190
3191                 DMEMIT("%u %llu/%llu %llu %llu/%llu %u %u %u %u %u %u %lu ",
3192                        (unsigned)DM_CACHE_METADATA_BLOCK_SIZE,
3193                        (unsigned long long)(nr_blocks_metadata - nr_free_blocks_metadata),
3194                        (unsigned long long)nr_blocks_metadata,
3195                        (unsigned long long)cache->sectors_per_block,
3196                        (unsigned long long) from_cblock(residency),
3197                        (unsigned long long) from_cblock(cache->cache_size),
3198                        (unsigned) atomic_read(&cache->stats.read_hit),
3199                        (unsigned) atomic_read(&cache->stats.read_miss),
3200                        (unsigned) atomic_read(&cache->stats.write_hit),
3201                        (unsigned) atomic_read(&cache->stats.write_miss),
3202                        (unsigned) atomic_read(&cache->stats.demotion),
3203                        (unsigned) atomic_read(&cache->stats.promotion),
3204                        (unsigned long) atomic_read(&cache->nr_dirty));
3205
3206                 emit_flags(cache, result, maxlen, &sz);
3207
3208                 DMEMIT("2 migration_threshold %llu ", (unsigned long long) cache->migration_threshold);
3209
3210                 DMEMIT("%s ", dm_cache_policy_get_name(cache->policy));
3211                 if (sz < maxlen) {
3212                         r = policy_emit_config_values(cache->policy, result, maxlen, &sz);
3213                         if (r)
3214                                 DMERR("%s: policy_emit_config_values returned %d",
3215                                       cache_device_name(cache), r);
3216                 }
3217
3218                 if (get_cache_mode(cache) == CM_READ_ONLY)
3219                         DMEMIT("ro ");
3220                 else
3221                         DMEMIT("rw ");
3222
3223                 r = dm_cache_metadata_needs_check(cache->cmd, &needs_check);
3224
3225                 if (r || needs_check)
3226                         DMEMIT("needs_check ");
3227                 else
3228                         DMEMIT("- ");
3229
3230                 break;
3231
3232         case STATUSTYPE_TABLE:
3233                 format_dev_t(buf, cache->metadata_dev->bdev->bd_dev);
3234                 DMEMIT("%s ", buf);
3235                 format_dev_t(buf, cache->cache_dev->bdev->bd_dev);
3236                 DMEMIT("%s ", buf);
3237                 format_dev_t(buf, cache->origin_dev->bdev->bd_dev);
3238                 DMEMIT("%s", buf);
3239
3240                 for (i = 0; i < cache->nr_ctr_args - 1; i++)
3241                         DMEMIT(" %s", cache->ctr_args[i]);
3242                 if (cache->nr_ctr_args)
3243                         DMEMIT(" %s", cache->ctr_args[cache->nr_ctr_args - 1]);
3244         }
3245
3246         return;
3247
3248 err:
3249         DMEMIT("Error");
3250 }
3251
3252 /*
3253  * Defines a range of cblocks, begin to (end - 1) are in the range.  end is
3254  * the one-past-the-end value.
3255  */
3256 struct cblock_range {
3257         dm_cblock_t begin;
3258         dm_cblock_t end;
3259 };
3260
3261 /*
3262  * A cache block range can take two forms:
3263  *
3264  * i) A single cblock, eg. '3456'
3265  * ii) A begin and end cblock with a dash between, eg. 123-234
3266  */
3267 static int parse_cblock_range(struct cache *cache, const char *str,
3268                               struct cblock_range *result)
3269 {
3270         char dummy;
3271         uint64_t b, e;
3272         int r;
3273
3274         /*
3275          * Try and parse form (ii) first.
3276          */
3277         r = sscanf(str, "%llu-%llu%c", &b, &e, &dummy);
3278         if (r < 0)
3279                 return r;
3280
3281         if (r == 2) {
3282                 result->begin = to_cblock(b);
3283                 result->end = to_cblock(e);
3284                 return 0;
3285         }
3286
3287         /*
3288          * That didn't work, try form (i).
3289          */
3290         r = sscanf(str, "%llu%c", &b, &dummy);
3291         if (r < 0)
3292                 return r;
3293
3294         if (r == 1) {
3295                 result->begin = to_cblock(b);
3296                 result->end = to_cblock(from_cblock(result->begin) + 1u);
3297                 return 0;
3298         }
3299
3300         DMERR("%s: invalid cblock range '%s'", cache_device_name(cache), str);
3301         return -EINVAL;
3302 }
3303
3304 static int validate_cblock_range(struct cache *cache, struct cblock_range *range)
3305 {
3306         uint64_t b = from_cblock(range->begin);
3307         uint64_t e = from_cblock(range->end);
3308         uint64_t n = from_cblock(cache->cache_size);
3309
3310         if (b >= n) {
3311                 DMERR("%s: begin cblock out of range: %llu >= %llu",
3312                       cache_device_name(cache), b, n);
3313                 return -EINVAL;
3314         }
3315
3316         if (e > n) {
3317                 DMERR("%s: end cblock out of range: %llu > %llu",
3318                       cache_device_name(cache), e, n);
3319                 return -EINVAL;
3320         }
3321
3322         if (b >= e) {
3323                 DMERR("%s: invalid cblock range: %llu >= %llu",
3324                       cache_device_name(cache), b, e);
3325                 return -EINVAL;
3326         }
3327
3328         return 0;
3329 }
3330
3331 static inline dm_cblock_t cblock_succ(dm_cblock_t b)
3332 {
3333         return to_cblock(from_cblock(b) + 1);
3334 }
3335
3336 static int request_invalidation(struct cache *cache, struct cblock_range *range)
3337 {
3338         int r = 0;
3339
3340         /*
3341          * We don't need to do any locking here because we know we're in
3342          * passthrough mode.  There's is potential for a race between an
3343          * invalidation triggered by an io and an invalidation message.  This
3344          * is harmless, we must not worry if the policy call fails.
3345          */
3346         while (range->begin != range->end) {
3347                 r = invalidate_cblock(cache, range->begin);
3348                 if (r)
3349                         return r;
3350
3351                 range->begin = cblock_succ(range->begin);
3352         }
3353
3354         cache->commit_requested = true;
3355         return r;
3356 }
3357
3358 static int process_invalidate_cblocks_message(struct cache *cache, unsigned count,
3359                                               const char **cblock_ranges)
3360 {
3361         int r = 0;
3362         unsigned i;
3363         struct cblock_range range;
3364
3365         if (!passthrough_mode(cache)) {
3366                 DMERR("%s: cache has to be in passthrough mode for invalidation",
3367                       cache_device_name(cache));
3368                 return -EPERM;
3369         }
3370
3371         for (i = 0; i < count; i++) {
3372                 r = parse_cblock_range(cache, cblock_ranges[i], &range);
3373                 if (r)
3374                         break;
3375
3376                 r = validate_cblock_range(cache, &range);
3377                 if (r)
3378                         break;
3379
3380                 /*
3381                  * Pass begin and end origin blocks to the worker and wake it.
3382                  */
3383                 r = request_invalidation(cache, &range);
3384                 if (r)
3385                         break;
3386         }
3387
3388         return r;
3389 }
3390
3391 /*
3392  * Supports
3393  *      "<key> <value>"
3394  * and
3395  *     "invalidate_cblocks [(<begin>)|(<begin>-<end>)]*
3396  *
3397  * The key migration_threshold is supported by the cache target core.
3398  */
3399 static int cache_message(struct dm_target *ti, unsigned argc, char **argv,
3400                          char *result, unsigned maxlen)
3401 {
3402         struct cache *cache = ti->private;
3403
3404         if (!argc)
3405                 return -EINVAL;
3406
3407         if (get_cache_mode(cache) >= CM_READ_ONLY) {
3408                 DMERR("%s: unable to service cache target messages in READ_ONLY or FAIL mode",
3409                       cache_device_name(cache));
3410                 return -EOPNOTSUPP;
3411         }
3412
3413         if (!strcasecmp(argv[0], "invalidate_cblocks"))
3414                 return process_invalidate_cblocks_message(cache, argc - 1, (const char **) argv + 1);
3415
3416         if (argc != 2)
3417                 return -EINVAL;
3418
3419         return set_config_value(cache, argv[0], argv[1]);
3420 }
3421
3422 static int cache_iterate_devices(struct dm_target *ti,
3423                                  iterate_devices_callout_fn fn, void *data)
3424 {
3425         int r = 0;
3426         struct cache *cache = ti->private;
3427
3428         r = fn(ti, cache->cache_dev, 0, get_dev_size(cache->cache_dev), data);
3429         if (!r)
3430                 r = fn(ti, cache->origin_dev, 0, ti->len, data);
3431
3432         return r;
3433 }
3434
3435 static bool origin_dev_supports_discard(struct block_device *origin_bdev)
3436 {
3437         struct request_queue *q = bdev_get_queue(origin_bdev);
3438
3439         return q && blk_queue_discard(q);
3440 }
3441
3442 /*
3443  * If discard_passdown was enabled verify that the origin device
3444  * supports discards.  Disable discard_passdown if not.
3445  */
3446 static void disable_passdown_if_not_supported(struct cache *cache)
3447 {
3448         struct block_device *origin_bdev = cache->origin_dev->bdev;
3449         struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3450         const char *reason = NULL;
3451         char buf[BDEVNAME_SIZE];
3452
3453         if (!cache->features.discard_passdown)
3454                 return;
3455
3456         if (!origin_dev_supports_discard(origin_bdev))
3457                 reason = "discard unsupported";
3458
3459         else if (origin_limits->max_discard_sectors < cache->sectors_per_block)
3460                 reason = "max discard sectors smaller than a block";
3461
3462         if (reason) {
3463                 DMWARN("Origin device (%s) %s: Disabling discard passdown.",
3464                        bdevname(origin_bdev, buf), reason);
3465                 cache->features.discard_passdown = false;
3466         }
3467 }
3468
3469 static void set_discard_limits(struct cache *cache, struct queue_limits *limits)
3470 {
3471         struct block_device *origin_bdev = cache->origin_dev->bdev;
3472         struct queue_limits *origin_limits = &bdev_get_queue(origin_bdev)->limits;
3473
3474         if (!cache->features.discard_passdown) {
3475                 /* No passdown is done so setting own virtual limits */
3476                 limits->max_discard_sectors = min_t(sector_t, cache->discard_block_size * 1024,
3477                                                     cache->origin_sectors);
3478                 limits->discard_granularity = cache->discard_block_size << SECTOR_SHIFT;
3479                 return;
3480         }
3481
3482         /*
3483          * cache_iterate_devices() is stacking both origin and fast device limits
3484          * but discards aren't passed to fast device, so inherit origin's limits.
3485          */
3486         limits->max_discard_sectors = origin_limits->max_discard_sectors;
3487         limits->max_hw_discard_sectors = origin_limits->max_hw_discard_sectors;
3488         limits->discard_granularity = origin_limits->discard_granularity;
3489         limits->discard_alignment = origin_limits->discard_alignment;
3490         limits->discard_misaligned = origin_limits->discard_misaligned;
3491 }
3492
3493 static void cache_io_hints(struct dm_target *ti, struct queue_limits *limits)
3494 {
3495         struct cache *cache = ti->private;
3496         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
3497
3498         /*
3499          * If the system-determined stacked limits are compatible with the
3500          * cache's blocksize (io_opt is a factor) do not override them.
3501          */
3502         if (io_opt_sectors < cache->sectors_per_block ||
3503             do_div(io_opt_sectors, cache->sectors_per_block)) {
3504                 blk_limits_io_min(limits, cache->sectors_per_block << SECTOR_SHIFT);
3505                 blk_limits_io_opt(limits, cache->sectors_per_block << SECTOR_SHIFT);
3506         }
3507
3508         disable_passdown_if_not_supported(cache);
3509         set_discard_limits(cache, limits);
3510 }
3511
3512 /*----------------------------------------------------------------*/
3513
3514 static struct target_type cache_target = {
3515         .name = "cache",
3516         .version = {2, 1, 0},
3517         .module = THIS_MODULE,
3518         .ctr = cache_ctr,
3519         .dtr = cache_dtr,
3520         .map = cache_map,
3521         .end_io = cache_end_io,
3522         .postsuspend = cache_postsuspend,
3523         .preresume = cache_preresume,
3524         .resume = cache_resume,
3525         .status = cache_status,
3526         .message = cache_message,
3527         .iterate_devices = cache_iterate_devices,
3528         .io_hints = cache_io_hints,
3529 };
3530
3531 static int __init dm_cache_init(void)
3532 {
3533         int r;
3534
3535         migration_cache = KMEM_CACHE(dm_cache_migration, 0);
3536         if (!migration_cache)
3537                 return -ENOMEM;
3538
3539         r = dm_register_target(&cache_target);
3540         if (r) {
3541                 DMERR("cache target registration failed: %d", r);
3542                 kmem_cache_destroy(migration_cache);
3543                 return r;
3544         }
3545
3546         return 0;
3547 }
3548
3549 static void __exit dm_cache_exit(void)
3550 {
3551         dm_unregister_target(&cache_target);
3552         kmem_cache_destroy(migration_cache);
3553 }
3554
3555 module_init(dm_cache_init);
3556 module_exit(dm_cache_exit);
3557
3558 MODULE_DESCRIPTION(DM_NAME " cache target");
3559 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
3560 MODULE_LICENSE("GPL");