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