drivers/md/dm-cache-target.c

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