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