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