drivers/md/dm-cache-target.c

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