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