drivers/md/dm-cache-metadata.c

   1 /*
   2  * Copyright (C) 2012 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-cache-metadata.h"
   8
   9 #include "persistent-data/dm-array.h"
  10 #include "persistent-data/dm-bitset.h"
  11 #include "persistent-data/dm-space-map.h"
  12 #include "persistent-data/dm-space-map-disk.h"
  13 #include "persistent-data/dm-transaction-manager.h"
  14
  15 #include <linux/device-mapper.h>
  16
  17 /*----------------------------------------------------------------*/
  18
  19 #define DM_MSG_PREFIX   "cache metadata"
  20
  21 #define CACHE_SUPERBLOCK_MAGIC 06142003
  22 #define CACHE_SUPERBLOCK_LOCATION 0
  23
  24 /*
  25  * defines a range of metadata versions that this module can handle.
  26  */
  27 #define MIN_CACHE_VERSION 1
  28 #define MAX_CACHE_VERSION 2
  29
  30 #define CACHE_METADATA_CACHE_SIZE 64
  31
  32 /*
  33  *  3 for btree insert +
  34  *  2 for btree lookup used within space map
  35  */
  36 #define CACHE_MAX_CONCURRENT_LOCKS 5
  37 #define SPACE_MAP_ROOT_SIZE 128
  38
  39 enum superblock_flag_bits {
  40         /* for spotting crashes that would invalidate the dirty bitset */
  41         CLEAN_SHUTDOWN,
  42         /* metadata must be checked using the tools */
  43         NEEDS_CHECK,
  44 };
  45
  46 /*
  47  * Each mapping from cache block -> origin block carries a set of flags.
  48  */
  49 enum mapping_bits {
  50         /*
  51          * A valid mapping.  Because we're using an array we clear this
  52          * flag for an non existant mapping.
  53          */
  54         M_VALID = 1,
  55
  56         /*
  57          * The data on the cache is different from that on the origin.
  58          * This flag is only used by metadata format 1.
  59          */
  60         M_DIRTY = 2
  61 };
  62
  63 struct cache_disk_superblock {
  64         __le32 csum;
  65         __le32 flags;
  66         __le64 blocknr;
  67
  68         __u8 uuid[16];
  69         __le64 magic;
  70         __le32 version;
  71
  72         __u8 policy_name[CACHE_POLICY_NAME_SIZE];
  73         __le32 policy_hint_size;
  74
  75         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
  76         __le64 mapping_root;
  77         __le64 hint_root;
  78
  79         __le64 discard_root;
  80         __le64 discard_block_size;
  81         __le64 discard_nr_blocks;
  82
  83         __le32 data_block_size;
  84         __le32 metadata_block_size;
  85         __le32 cache_blocks;
  86
  87         __le32 compat_flags;
  88         __le32 compat_ro_flags;
  89         __le32 incompat_flags;
  90
  91         __le32 read_hits;
  92         __le32 read_misses;
  93         __le32 write_hits;
  94         __le32 write_misses;
  95
  96         __le32 policy_version[CACHE_POLICY_VERSION_SIZE];
  97
  98         /*
  99          * Metadata format 2 fields.
 100          */
 101         __le64 dirty_root;
 102 } __packed;
 103
 104 struct dm_cache_metadata {
 105         atomic_t ref_count;
 106         struct list_head list;
 107
 108         unsigned version;
 109         struct block_device *bdev;
 110         struct dm_block_manager *bm;
 111         struct dm_space_map *metadata_sm;
 112         struct dm_transaction_manager *tm;
 113
 114         struct dm_array_info info;
 115         struct dm_array_info hint_info;
 116         struct dm_disk_bitset discard_info;
 117
 118         struct rw_semaphore root_lock;
 119         unsigned long flags;
 120         dm_block_t root;
 121         dm_block_t hint_root;
 122         dm_block_t discard_root;
 123
 124         sector_t discard_block_size;
 125         dm_dblock_t discard_nr_blocks;
 126
 127         sector_t data_block_size;
 128         dm_cblock_t cache_blocks;
 129         bool changed:1;
 130         bool clean_when_opened:1;
 131
 132         char policy_name[CACHE_POLICY_NAME_SIZE];
 133         unsigned policy_version[CACHE_POLICY_VERSION_SIZE];
 134         size_t policy_hint_size;
 135         struct dm_cache_statistics stats;
 136
 137         /*
 138          * Reading the space map root can fail, so we read it into this
 139          * buffer before the superblock is locked and updated.
 140          */
 141         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 142
 143         /*
 144          * Set if a transaction has to be aborted but the attempt to roll
 145          * back to the previous (good) transaction failed.  The only
 146          * metadata operation permissible in this state is the closing of
 147          * the device.
 148          */
 149         bool fail_io:1;
 150
 151         /*
 152          * Metadata format 2 fields.
 153          */
 154         dm_block_t dirty_root;
 155         struct dm_disk_bitset dirty_info;
 156
 157         /*
 158          * These structures are used when loading metadata.  They're too
 159          * big to put on the stack.
 160          */
 161         struct dm_array_cursor mapping_cursor;
 162         struct dm_array_cursor hint_cursor;
 163         struct dm_bitset_cursor dirty_cursor;
 164 };
 165
 166 /*-------------------------------------------------------------------
 167  * superblock validator
 168  *-----------------------------------------------------------------*/
 169
 170 #define SUPERBLOCK_CSUM_XOR 9031977
 171
 172 static void sb_prepare_for_write(struct dm_block_validator *v,
 173                                  struct dm_block *b,
 174                                  size_t sb_block_size)
 175 {
 176         struct cache_disk_superblock *disk_super = dm_block_data(b);
 177
 178         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 179         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 180                                                       sb_block_size - sizeof(__le32),
 181                                                       SUPERBLOCK_CSUM_XOR));
 182 }
 183
 184 static int check_metadata_version(struct cache_disk_superblock *disk_super)
 185 {
 186         uint32_t metadata_version = le32_to_cpu(disk_super->version);
 187
 188         if (metadata_version < MIN_CACHE_VERSION || metadata_version > MAX_CACHE_VERSION) {
 189                 DMERR("Cache metadata version %u found, but only versions between %u and %u supported.",
 190                       metadata_version, MIN_CACHE_VERSION, MAX_CACHE_VERSION);
 191                 return -EINVAL;
 192         }
 193
 194         return 0;
 195 }
 196
 197 static int sb_check(struct dm_block_validator *v,
 198                     struct dm_block *b,
 199                     size_t sb_block_size)
 200 {
 201         struct cache_disk_superblock *disk_super = dm_block_data(b);
 202         __le32 csum_le;
 203
 204         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 205                 DMERR("sb_check failed: blocknr %llu: wanted %llu",
 206                       le64_to_cpu(disk_super->blocknr),
 207                       (unsigned long long)dm_block_location(b));
 208                 return -ENOTBLK;
 209         }
 210
 211         if (le64_to_cpu(disk_super->magic) != CACHE_SUPERBLOCK_MAGIC) {
 212                 DMERR("sb_check failed: magic %llu: wanted %llu",
 213                       le64_to_cpu(disk_super->magic),
 214                       (unsigned long long)CACHE_SUPERBLOCK_MAGIC);
 215                 return -EILSEQ;
 216         }
 217
 218         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 219                                              sb_block_size - sizeof(__le32),
 220                                              SUPERBLOCK_CSUM_XOR));
 221         if (csum_le != disk_super->csum) {
 222                 DMERR("sb_check failed: csum %u: wanted %u",
 223                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 224                 return -EILSEQ;
 225         }
 226
 227         return check_metadata_version(disk_super);
 228 }
 229
 230 static struct dm_block_validator sb_validator = {
 231         .name = "superblock",
 232         .prepare_for_write = sb_prepare_for_write,
 233         .check = sb_check
 234 };
 235
 236 /*----------------------------------------------------------------*/
 237
 238 static int superblock_read_lock(struct dm_cache_metadata *cmd,
 239                                 struct dm_block **sblock)
 240 {
 241         return dm_bm_read_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 242                                &sb_validator, sblock);
 243 }
 244
 245 static int superblock_lock_zero(struct dm_cache_metadata *cmd,
 246                                 struct dm_block **sblock)
 247 {
 248         return dm_bm_write_lock_zero(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 249                                      &sb_validator, sblock);
 250 }
 251
 252 static int superblock_lock(struct dm_cache_metadata *cmd,
 253                            struct dm_block **sblock)
 254 {
 255         return dm_bm_write_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 256                                 &sb_validator, sblock);
 257 }
 258
 259 /*----------------------------------------------------------------*/
 260
 261 static int __superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
 262 {
 263         int r;
 264         unsigned i;
 265         struct dm_block *b;
 266         __le64 *data_le, zero = cpu_to_le64(0);
 267         unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
 268
 269         /*
 270          * We can't use a validator here - it may be all zeroes.
 271          */
 272         r = dm_bm_read_lock(bm, CACHE_SUPERBLOCK_LOCATION, NULL, &b);
 273         if (r)
 274                 return r;
 275
 276         data_le = dm_block_data(b);
 277         *result = true;
 278         for (i = 0; i < sb_block_size; i++) {
 279                 if (data_le[i] != zero) {
 280                         *result = false;
 281                         break;
 282                 }
 283         }
 284
 285         dm_bm_unlock(b);
 286
 287         return 0;
 288 }
 289
 290 static void __setup_mapping_info(struct dm_cache_metadata *cmd)
 291 {
 292         struct dm_btree_value_type vt;
 293
 294         vt.context = NULL;
 295         vt.size = sizeof(__le64);
 296         vt.inc = NULL;
 297         vt.dec = NULL;
 298         vt.equal = NULL;
 299         dm_array_info_init(&cmd->info, cmd->tm, &vt);
 300
 301         if (cmd->policy_hint_size) {
 302                 vt.size = sizeof(__le32);
 303                 dm_array_info_init(&cmd->hint_info, cmd->tm, &vt);
 304         }
 305 }
 306
 307 static int __save_sm_root(struct dm_cache_metadata *cmd)
 308 {
 309         int r;
 310         size_t metadata_len;
 311
 312         r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
 313         if (r < 0)
 314                 return r;
 315
 316         return dm_sm_copy_root(cmd->metadata_sm, &cmd->metadata_space_map_root,
 317                                metadata_len);
 318 }
 319
 320 static void __copy_sm_root(struct dm_cache_metadata *cmd,
 321                            struct cache_disk_superblock *disk_super)
 322 {
 323         memcpy(&disk_super->metadata_space_map_root,
 324                &cmd->metadata_space_map_root,
 325                sizeof(cmd->metadata_space_map_root));
 326 }
 327
 328 static bool separate_dirty_bits(struct dm_cache_metadata *cmd)
 329 {
 330         return cmd->version >= 2;
 331 }
 332
 333 static int __write_initial_superblock(struct dm_cache_metadata *cmd)
 334 {
 335         int r;
 336         struct dm_block *sblock;
 337         struct cache_disk_superblock *disk_super;
 338         sector_t bdev_size = i_size_read(cmd->bdev->bd_inode) >> SECTOR_SHIFT;
 339
 340         /* FIXME: see if we can lose the max sectors limit */
 341         if (bdev_size > DM_CACHE_METADATA_MAX_SECTORS)
 342                 bdev_size = DM_CACHE_METADATA_MAX_SECTORS;
 343
 344         r = dm_tm_pre_commit(cmd->tm);
 345         if (r < 0)
 346                 return r;
 347
 348         /*
 349          * dm_sm_copy_root() can fail.  So we need to do it before we start
 350          * updating the superblock.
 351          */
 352         r = __save_sm_root(cmd);
 353         if (r)
 354                 return r;
 355
 356         r = superblock_lock_zero(cmd, &sblock);
 357         if (r)
 358                 return r;
 359
 360         disk_super = dm_block_data(sblock);
 361         disk_super->flags = 0;
 362         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 363         disk_super->magic = cpu_to_le64(CACHE_SUPERBLOCK_MAGIC);
 364         disk_super->version = cpu_to_le32(cmd->version);
 365         memset(disk_super->policy_name, 0, sizeof(disk_super->policy_name));
 366         memset(disk_super->policy_version, 0, sizeof(disk_super->policy_version));
 367         disk_super->policy_hint_size = 0;
 368
 369         __copy_sm_root(cmd, disk_super);
 370
 371         disk_super->mapping_root = cpu_to_le64(cmd->root);
 372         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 373         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 374         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 375         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 376         disk_super->metadata_block_size = cpu_to_le32(DM_CACHE_METADATA_BLOCK_SIZE);
 377         disk_super->data_block_size = cpu_to_le32(cmd->data_block_size);
 378         disk_super->cache_blocks = cpu_to_le32(0);
 379
 380         disk_super->read_hits = cpu_to_le32(0);
 381         disk_super->read_misses = cpu_to_le32(0);
 382         disk_super->write_hits = cpu_to_le32(0);
 383         disk_super->write_misses = cpu_to_le32(0);
 384
 385         if (separate_dirty_bits(cmd))
 386                 disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
 387
 388         return dm_tm_commit(cmd->tm, sblock);
 389 }
 390
 391 static int __format_metadata(struct dm_cache_metadata *cmd)
 392 {
 393         int r;
 394
 395         r = dm_tm_create_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 396                                  &cmd->tm, &cmd->metadata_sm);
 397         if (r < 0) {
 398                 DMERR("tm_create_with_sm failed");
 399                 return r;
 400         }
 401
 402         __setup_mapping_info(cmd);
 403
 404         r = dm_array_empty(&cmd->info, &cmd->root);
 405         if (r < 0)
 406                 goto bad;
 407
 408         if (separate_dirty_bits(cmd)) {
 409                 dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
 410                 r = dm_bitset_empty(&cmd->dirty_info, &cmd->dirty_root);
 411                 if (r < 0)
 412                         goto bad;
 413         }
 414
 415         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 416         r = dm_bitset_empty(&cmd->discard_info, &cmd->discard_root);
 417         if (r < 0)
 418                 goto bad;
 419
 420         cmd->discard_block_size = 0;
 421         cmd->discard_nr_blocks = 0;
 422
 423         r = __write_initial_superblock(cmd);
 424         if (r)
 425                 goto bad;
 426
 427         cmd->clean_when_opened = true;
 428         return 0;
 429
 430 bad:
 431         dm_tm_destroy(cmd->tm);
 432         dm_sm_destroy(cmd->metadata_sm);
 433
 434         return r;
 435 }
 436
 437 static int __check_incompat_features(struct cache_disk_superblock *disk_super,
 438                                      struct dm_cache_metadata *cmd)
 439 {
 440         uint32_t incompat_flags, features;
 441
 442         incompat_flags = le32_to_cpu(disk_super->incompat_flags);
 443         features = incompat_flags & ~DM_CACHE_FEATURE_INCOMPAT_SUPP;
 444         if (features) {
 445                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 446                       (unsigned long)features);
 447                 return -EINVAL;
 448         }
 449
 450         /*
 451          * Check for read-only metadata to skip the following RDWR checks.
 452          */
 453         if (get_disk_ro(cmd->bdev->bd_disk))
 454                 return 0;
 455
 456         features = le32_to_cpu(disk_super->compat_ro_flags) & ~DM_CACHE_FEATURE_COMPAT_RO_SUPP;
 457         if (features) {
 458                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 459                       (unsigned long)features);
 460                 return -EINVAL;
 461         }
 462
 463         return 0;
 464 }
 465
 466 static int __open_metadata(struct dm_cache_metadata *cmd)
 467 {
 468         int r;
 469         struct dm_block *sblock;
 470         struct cache_disk_superblock *disk_super;
 471         unsigned long sb_flags;
 472
 473         r = superblock_read_lock(cmd, &sblock);
 474         if (r < 0) {
 475                 DMERR("couldn't read lock superblock");
 476                 return r;
 477         }
 478
 479         disk_super = dm_block_data(sblock);
 480
 481         /* Verify the data block size hasn't changed */
 482         if (le32_to_cpu(disk_super->data_block_size) != cmd->data_block_size) {
 483                 DMERR("changing the data block size (from %u to %llu) is not supported",
 484                       le32_to_cpu(disk_super->data_block_size),
 485                       (unsigned long long)cmd->data_block_size);
 486                 r = -EINVAL;
 487                 goto bad;
 488         }
 489
 490         r = __check_incompat_features(disk_super, cmd);
 491         if (r < 0)
 492                 goto bad;
 493
 494         r = dm_tm_open_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 495                                disk_super->metadata_space_map_root,
 496                                sizeof(disk_super->metadata_space_map_root),
 497                                &cmd->tm, &cmd->metadata_sm);
 498         if (r < 0) {
 499                 DMERR("tm_open_with_sm failed");
 500                 goto bad;
 501         }
 502
 503         __setup_mapping_info(cmd);
 504         dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
 505         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 506         sb_flags = le32_to_cpu(disk_super->flags);
 507         cmd->clean_when_opened = test_bit(CLEAN_SHUTDOWN, &sb_flags);
 508         dm_bm_unlock(sblock);
 509
 510         return 0;
 511
 512 bad:
 513         dm_bm_unlock(sblock);
 514         return r;
 515 }
 516
 517 static int __open_or_format_metadata(struct dm_cache_metadata *cmd,
 518                                      bool format_device)
 519 {
 520         int r;
 521         bool unformatted = false;
 522
 523         r = __superblock_all_zeroes(cmd->bm, &unformatted);
 524         if (r)
 525                 return r;
 526
 527         if (unformatted)
 528                 return format_device ? __format_metadata(cmd) : -EPERM;
 529
 530         return __open_metadata(cmd);
 531 }
 532
 533 static int __create_persistent_data_objects(struct dm_cache_metadata *cmd,
 534                                             bool may_format_device)
 535 {
 536         int r;
 537         cmd->bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 538                                           CACHE_METADATA_CACHE_SIZE,
 539                                           CACHE_MAX_CONCURRENT_LOCKS);
 540         if (IS_ERR(cmd->bm)) {
 541                 DMERR("could not create block manager");
 542                 return PTR_ERR(cmd->bm);
 543         }
 544
 545         r = __open_or_format_metadata(cmd, may_format_device);
 546         if (r)
 547                 dm_block_manager_destroy(cmd->bm);
 548
 549         return r;
 550 }
 551
 552 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd)
 553 {
 554         dm_sm_destroy(cmd->metadata_sm);
 555         dm_tm_destroy(cmd->tm);
 556         dm_block_manager_destroy(cmd->bm);
 557 }
 558
 559 typedef unsigned long (*flags_mutator)(unsigned long);
 560
 561 static void update_flags(struct cache_disk_superblock *disk_super,
 562                          flags_mutator mutator)
 563 {
 564         uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
 565         disk_super->flags = cpu_to_le32(sb_flags);
 566 }
 567
 568 static unsigned long set_clean_shutdown(unsigned long flags)
 569 {
 570         set_bit(CLEAN_SHUTDOWN, &flags);
 571         return flags;
 572 }
 573
 574 static unsigned long clear_clean_shutdown(unsigned long flags)
 575 {
 576         clear_bit(CLEAN_SHUTDOWN, &flags);
 577         return flags;
 578 }
 579
 580 static void read_superblock_fields(struct dm_cache_metadata *cmd,
 581                                    struct cache_disk_superblock *disk_super)
 582 {
 583         cmd->version = le32_to_cpu(disk_super->version);
 584         cmd->flags = le32_to_cpu(disk_super->flags);
 585         cmd->root = le64_to_cpu(disk_super->mapping_root);
 586         cmd->hint_root = le64_to_cpu(disk_super->hint_root);
 587         cmd->discard_root = le64_to_cpu(disk_super->discard_root);
 588         cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
 589         cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
 590         cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 591         cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
 592         strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
 593         cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
 594         cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
 595         cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
 596         cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
 597
 598         cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
 599         cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
 600         cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
 601         cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
 602
 603         if (separate_dirty_bits(cmd))
 604                 cmd->dirty_root = le64_to_cpu(disk_super->dirty_root);
 605
 606         cmd->changed = false;
 607 }
 608
 609 /*
 610  * The mutator updates the superblock flags.
 611  */
 612 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
 613                                      flags_mutator mutator)
 614 {
 615         int r;
 616         struct cache_disk_superblock *disk_super;
 617         struct dm_block *sblock;
 618
 619         r = superblock_lock(cmd, &sblock);
 620         if (r)
 621                 return r;
 622
 623         disk_super = dm_block_data(sblock);
 624         update_flags(disk_super, mutator);
 625         read_superblock_fields(cmd, disk_super);
 626         dm_bm_unlock(sblock);
 627
 628         return dm_bm_flush(cmd->bm);
 629 }
 630
 631 static int __begin_transaction(struct dm_cache_metadata *cmd)
 632 {
 633         int r;
 634         struct cache_disk_superblock *disk_super;
 635         struct dm_block *sblock;
 636
 637         /*
 638          * We re-read the superblock every time.  Shouldn't need to do this
 639          * really.
 640          */
 641         r = superblock_read_lock(cmd, &sblock);
 642         if (r)
 643                 return r;
 644
 645         disk_super = dm_block_data(sblock);
 646         read_superblock_fields(cmd, disk_super);
 647         dm_bm_unlock(sblock);
 648
 649         return 0;
 650 }
 651
 652 static int __commit_transaction(struct dm_cache_metadata *cmd,
 653                                 flags_mutator mutator)
 654 {
 655         int r;
 656         struct cache_disk_superblock *disk_super;
 657         struct dm_block *sblock;
 658
 659         /*
 660          * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
 661          */
 662         BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
 663
 664         if (separate_dirty_bits(cmd)) {
 665                 r = dm_bitset_flush(&cmd->dirty_info, cmd->dirty_root,
 666                                     &cmd->dirty_root);
 667                 if (r)
 668                         return r;
 669         }
 670
 671         r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
 672                             &cmd->discard_root);
 673         if (r)
 674                 return r;
 675
 676         r = dm_tm_pre_commit(cmd->tm);
 677         if (r < 0)
 678                 return r;
 679
 680         r = __save_sm_root(cmd);
 681         if (r)
 682                 return r;
 683
 684         r = superblock_lock(cmd, &sblock);
 685         if (r)
 686                 return r;
 687
 688         disk_super = dm_block_data(sblock);
 689
 690         disk_super->flags = cpu_to_le32(cmd->flags);
 691         if (mutator)
 692                 update_flags(disk_super, mutator);
 693
 694         disk_super->mapping_root = cpu_to_le64(cmd->root);
 695         if (separate_dirty_bits(cmd))
 696                 disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
 697         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 698         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 699         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 700         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 701         disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
 702         strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
 703         disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
 704         disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
 705         disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
 706
 707         disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
 708         disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
 709         disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
 710         disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
 711         __copy_sm_root(cmd, disk_super);
 712
 713         return dm_tm_commit(cmd->tm, sblock);
 714 }
 715
 716 /*----------------------------------------------------------------*/
 717
 718 /*
 719  * The mappings are held in a dm-array that has 64-bit values stored in
 720  * little-endian format.  The index is the cblock, the high 48bits of the
 721  * value are the oblock and the low 16 bit the flags.
 722  */
 723 #define FLAGS_MASK ((1 << 16) - 1)
 724
 725 static __le64 pack_value(dm_oblock_t block, unsigned flags)
 726 {
 727         uint64_t value = from_oblock(block);
 728         value <<= 16;
 729         value = value | (flags & FLAGS_MASK);
 730         return cpu_to_le64(value);
 731 }
 732
 733 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
 734 {
 735         uint64_t value = le64_to_cpu(value_le);
 736         uint64_t b = value >> 16;
 737         *block = to_oblock(b);
 738         *flags = value & FLAGS_MASK;
 739 }
 740
 741 /*----------------------------------------------------------------*/
 742
 743 static struct dm_cache_metadata *metadata_open(struct block_device *bdev,
 744                                                sector_t data_block_size,
 745                                                bool may_format_device,
 746                                                size_t policy_hint_size,
 747                                                unsigned metadata_version)
 748 {
 749         int r;
 750         struct dm_cache_metadata *cmd;
 751
 752         cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
 753         if (!cmd) {
 754                 DMERR("could not allocate metadata struct");
 755                 return ERR_PTR(-ENOMEM);
 756         }
 757
 758         cmd->version = metadata_version;
 759         atomic_set(&cmd->ref_count, 1);
 760         init_rwsem(&cmd->root_lock);
 761         cmd->bdev = bdev;
 762         cmd->data_block_size = data_block_size;
 763         cmd->cache_blocks = 0;
 764         cmd->policy_hint_size = policy_hint_size;
 765         cmd->changed = true;
 766         cmd->fail_io = false;
 767
 768         r = __create_persistent_data_objects(cmd, may_format_device);
 769         if (r) {
 770                 kfree(cmd);
 771                 return ERR_PTR(r);
 772         }
 773
 774         r = __begin_transaction_flags(cmd, clear_clean_shutdown);
 775         if (r < 0) {
 776                 dm_cache_metadata_close(cmd);
 777                 return ERR_PTR(r);
 778         }
 779
 780         return cmd;
 781 }
 782
 783 /*
 784  * We keep a little list of ref counted metadata objects to prevent two
 785  * different target instances creating separate bufio instances.  This is
 786  * an issue if a table is reloaded before the suspend.
 787  */
 788 static DEFINE_MUTEX(table_lock);
 789 static LIST_HEAD(table);
 790
 791 static struct dm_cache_metadata *lookup(struct block_device *bdev)
 792 {
 793         struct dm_cache_metadata *cmd;
 794
 795         list_for_each_entry(cmd, &table, list)
 796                 if (cmd->bdev == bdev) {
 797                         atomic_inc(&cmd->ref_count);
 798                         return cmd;
 799                 }
 800
 801         return NULL;
 802 }
 803
 804 static struct dm_cache_metadata *lookup_or_open(struct block_device *bdev,
 805                                                 sector_t data_block_size,
 806                                                 bool may_format_device,
 807                                                 size_t policy_hint_size,
 808                                                 unsigned metadata_version)
 809 {
 810         struct dm_cache_metadata *cmd, *cmd2;
 811
 812         mutex_lock(&table_lock);
 813         cmd = lookup(bdev);
 814         mutex_unlock(&table_lock);
 815
 816         if (cmd)
 817                 return cmd;
 818
 819         cmd = metadata_open(bdev, data_block_size, may_format_device,
 820                             policy_hint_size, metadata_version);
 821         if (!IS_ERR(cmd)) {
 822                 mutex_lock(&table_lock);
 823                 cmd2 = lookup(bdev);
 824                 if (cmd2) {
 825                         mutex_unlock(&table_lock);
 826                         __destroy_persistent_data_objects(cmd);
 827                         kfree(cmd);
 828                         return cmd2;
 829                 }
 830                 list_add(&cmd->list, &table);
 831                 mutex_unlock(&table_lock);
 832         }
 833
 834         return cmd;
 835 }
 836
 837 static bool same_params(struct dm_cache_metadata *cmd, sector_t data_block_size)
 838 {
 839         if (cmd->data_block_size != data_block_size) {
 840                 DMERR("data_block_size (%llu) different from that in metadata (%llu)",
 841                       (unsigned long long) data_block_size,
 842                       (unsigned long long) cmd->data_block_size);
 843                 return false;
 844         }
 845
 846         return true;
 847 }
 848
 849 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
 850                                                  sector_t data_block_size,
 851                                                  bool may_format_device,
 852                                                  size_t policy_hint_size,
 853                                                  unsigned metadata_version)
 854 {
 855         struct dm_cache_metadata *cmd = lookup_or_open(bdev, data_block_size, may_format_device,
 856                                                        policy_hint_size, metadata_version);
 857
 858         if (!IS_ERR(cmd) && !same_params(cmd, data_block_size)) {
 859                 dm_cache_metadata_close(cmd);
 860                 return ERR_PTR(-EINVAL);
 861         }
 862
 863         return cmd;
 864 }
 865
 866 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
 867 {
 868         if (atomic_dec_and_test(&cmd->ref_count)) {
 869                 mutex_lock(&table_lock);
 870                 list_del(&cmd->list);
 871                 mutex_unlock(&table_lock);
 872
 873                 if (!cmd->fail_io)
 874                         __destroy_persistent_data_objects(cmd);
 875                 kfree(cmd);
 876         }
 877 }
 878
 879 /*
 880  * Checks that the given cache block is either unmapped or clean.
 881  */
 882 static int block_clean_combined_dirty(struct dm_cache_metadata *cmd, dm_cblock_t b,
 883                                       bool *result)
 884 {
 885         int r;
 886         __le64 value;
 887         dm_oblock_t ob;
 888         unsigned flags;
 889
 890         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(b), &value);
 891         if (r)
 892                 return r;
 893
 894         unpack_value(value, &ob, &flags);
 895         *result = !((flags & M_VALID) && (flags & M_DIRTY));
 896
 897         return 0;
 898 }
 899
 900 static int blocks_are_clean_combined_dirty(struct dm_cache_metadata *cmd,
 901                                            dm_cblock_t begin, dm_cblock_t end,
 902                                            bool *result)
 903 {
 904         int r;
 905         *result = true;
 906
 907         while (begin != end) {
 908                 r = block_clean_combined_dirty(cmd, begin, result);
 909                 if (r) {
 910                         DMERR("block_clean_combined_dirty failed");
 911                         return r;
 912                 }
 913
 914                 if (!*result) {
 915                         DMERR("cache block %llu is dirty",
 916                               (unsigned long long) from_cblock(begin));
 917                         return 0;
 918                 }
 919
 920                 begin = to_cblock(from_cblock(begin) + 1);
 921         }
 922
 923         return 0;
 924 }
 925
 926 static int blocks_are_clean_separate_dirty(struct dm_cache_metadata *cmd,
 927                                            dm_cblock_t begin, dm_cblock_t end,
 928                                            bool *result)
 929 {
 930         int r;
 931         bool dirty_flag;
 932         *result = true;
 933
 934         r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
 935                                    from_cblock(begin), &cmd->dirty_cursor);
 936         if (r) {
 937                 DMERR("%s: dm_bitset_cursor_begin for dirty failed", __func__);
 938                 return r;
 939         }
 940
 941         r = dm_bitset_cursor_skip(&cmd->dirty_cursor, from_cblock(begin));
 942         if (r) {
 943                 DMERR("%s: dm_bitset_cursor_skip for dirty failed", __func__);
 944                 dm_bitset_cursor_end(&cmd->dirty_cursor);
 945                 return r;
 946         }
 947
 948         while (begin != end) {
 949                 /*
 950                  * We assume that unmapped blocks have their dirty bit
 951                  * cleared.
 952                  */
 953                 dirty_flag = dm_bitset_cursor_get_value(&cmd->dirty_cursor);
 954                 if (dirty_flag) {
 955                         DMERR("%s: cache block %llu is dirty", __func__,
 956                               (unsigned long long) from_cblock(begin));
 957                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 958                         *result = false;
 959                         return 0;
 960                 }
 961
 962                 r = dm_bitset_cursor_next(&cmd->dirty_cursor);
 963                 if (r) {
 964                         DMERR("%s: dm_bitset_cursor_next for dirty failed", __func__);
 965                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 966                         return r;
 967                 }
 968
 969                 begin = to_cblock(from_cblock(begin) + 1);
 970         }
 971
 972         dm_bitset_cursor_end(&cmd->dirty_cursor);
 973
 974         return 0;
 975 }
 976
 977 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
 978                                         dm_cblock_t begin, dm_cblock_t end,
 979                                         bool *result)
 980 {
 981         if (separate_dirty_bits(cmd))
 982                 return blocks_are_clean_separate_dirty(cmd, begin, end, result);
 983         else
 984                 return blocks_are_clean_combined_dirty(cmd, begin, end, result);
 985 }
 986
 987 static bool cmd_write_lock(struct dm_cache_metadata *cmd)
 988 {
 989         down_write(&cmd->root_lock);
 990         if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
 991                 up_write(&cmd->root_lock);
 992                 return false;
 993         }
 994         return true;
 995 }
 996
 997 #define WRITE_LOCK(cmd)                         \
 998         do {                                    \
 999                 if (!cmd_write_lock((cmd)))     \
1000                         return -EINVAL;         \
1001         } while(0)
1002
1003 #define WRITE_LOCK_VOID(cmd)                    \
1004         do {                                    \
1005                 if (!cmd_write_lock((cmd)))     \
1006                         return;                 \
1007         } while(0)
1008
1009 #define WRITE_UNLOCK(cmd) \
1010         up_write(&(cmd)->root_lock)
1011
1012 static bool cmd_read_lock(struct dm_cache_metadata *cmd)
1013 {
1014         down_read(&cmd->root_lock);
1015         if (cmd->fail_io) {
1016                 up_read(&cmd->root_lock);
1017                 return false;
1018         }
1019         return true;
1020 }
1021
1022 #define READ_LOCK(cmd)                          \
1023         do {                                    \
1024                 if (!cmd_read_lock((cmd)))      \
1025                         return -EINVAL;         \
1026         } while(0)
1027
1028 #define READ_LOCK_VOID(cmd)                     \
1029         do {                                    \
1030                 if (!cmd_read_lock((cmd)))      \
1031                         return;                 \
1032         } while(0)
1033
1034 #define READ_UNLOCK(cmd) \
1035         up_read(&(cmd)->root_lock)
1036
1037 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
1038 {
1039         int r;
1040         bool clean;
1041         __le64 null_mapping = pack_value(0, 0);
1042
1043         WRITE_LOCK(cmd);
1044         __dm_bless_for_disk(&null_mapping);
1045
1046         if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
1047                 r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
1048                 if (r) {
1049                         __dm_unbless_for_disk(&null_mapping);
1050                         goto out;
1051                 }
1052
1053                 if (!clean) {
1054                         DMERR("unable to shrink cache due to dirty blocks");
1055                         r = -EINVAL;
1056                         __dm_unbless_for_disk(&null_mapping);
1057                         goto out;
1058                 }
1059         }
1060
1061         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
1062                             from_cblock(new_cache_size),
1063                             &null_mapping, &cmd->root);
1064         if (r)
1065                 goto out;
1066
1067         if (separate_dirty_bits(cmd)) {
1068                 r = dm_bitset_resize(&cmd->dirty_info, cmd->dirty_root,
1069                                      from_cblock(cmd->cache_blocks), from_cblock(new_cache_size),
1070                                      false, &cmd->dirty_root);
1071                 if (r)
1072                         goto out;
1073         }
1074
1075         cmd->cache_blocks = new_cache_size;
1076         cmd->changed = true;
1077
1078 out:
1079         WRITE_UNLOCK(cmd);
1080
1081         return r;
1082 }
1083
1084 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
1085                                    sector_t discard_block_size,
1086                                    dm_dblock_t new_nr_entries)
1087 {
1088         int r;
1089
1090         WRITE_LOCK(cmd);
1091         r = dm_bitset_resize(&cmd->discard_info,
1092                              cmd->discard_root,
1093                              from_dblock(cmd->discard_nr_blocks),
1094                              from_dblock(new_nr_entries),
1095                              false, &cmd->discard_root);
1096         if (!r) {
1097                 cmd->discard_block_size = discard_block_size;
1098                 cmd->discard_nr_blocks = new_nr_entries;
1099         }
1100
1101         cmd->changed = true;
1102         WRITE_UNLOCK(cmd);
1103
1104         return r;
1105 }
1106
1107 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1108 {
1109         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
1110                                  from_dblock(b), &cmd->discard_root);
1111 }
1112
1113 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1114 {
1115         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
1116                                    from_dblock(b), &cmd->discard_root);
1117 }
1118
1119 static int __discard(struct dm_cache_metadata *cmd,
1120                      dm_dblock_t dblock, bool discard)
1121 {
1122         int r;
1123
1124         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
1125         if (r)
1126                 return r;
1127
1128         cmd->changed = true;
1129         return 0;
1130 }
1131
1132 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
1133                          dm_dblock_t dblock, bool discard)
1134 {
1135         int r;
1136
1137         WRITE_LOCK(cmd);
1138         r = __discard(cmd, dblock, discard);
1139         WRITE_UNLOCK(cmd);
1140
1141         return r;
1142 }
1143
1144 static int __load_discards(struct dm_cache_metadata *cmd,
1145                            load_discard_fn fn, void *context)
1146 {
1147         int r = 0;
1148         uint32_t b;
1149         struct dm_bitset_cursor c;
1150
1151         if (from_dblock(cmd->discard_nr_blocks) == 0)
1152                 /* nothing to do */
1153                 return 0;
1154
1155         if (cmd->clean_when_opened) {
1156                 r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root, &cmd->discard_root);
1157                 if (r)
1158                         return r;
1159
1160                 r = dm_bitset_cursor_begin(&cmd->discard_info, cmd->discard_root,
1161                                            from_dblock(cmd->discard_nr_blocks), &c);
1162                 if (r)
1163                         return r;
1164
1165                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1166                         r = fn(context, cmd->discard_block_size, to_dblock(b),
1167                                dm_bitset_cursor_get_value(&c));
1168                         if (r)
1169                                 break;
1170                 }
1171
1172                 dm_bitset_cursor_end(&c);
1173
1174         } else {
1175                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1176                         r = fn(context, cmd->discard_block_size, to_dblock(b), false);
1177                         if (r)
1178                                 return r;
1179                 }
1180         }
1181
1182         return r;
1183 }
1184
1185 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
1186                            load_discard_fn fn, void *context)
1187 {
1188         int r;
1189
1190         READ_LOCK(cmd);
1191         r = __load_discards(cmd, fn, context);
1192         READ_UNLOCK(cmd);
1193
1194         return r;
1195 }
1196
1197 int dm_cache_size(struct dm_cache_metadata *cmd, dm_cblock_t *result)
1198 {
1199         READ_LOCK(cmd);
1200         *result = cmd->cache_blocks;
1201         READ_UNLOCK(cmd);
1202
1203         return 0;
1204 }
1205
1206 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1207 {
1208         int r;
1209         __le64 value = pack_value(0, 0);
1210
1211         __dm_bless_for_disk(&value);
1212         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1213                                &value, &cmd->root);
1214         if (r)
1215                 return r;
1216
1217         cmd->changed = true;
1218         return 0;
1219 }
1220
1221 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1222 {
1223         int r;
1224
1225         WRITE_LOCK(cmd);
1226         r = __remove(cmd, cblock);
1227         WRITE_UNLOCK(cmd);
1228
1229         return r;
1230 }
1231
1232 static int __insert(struct dm_cache_metadata *cmd,
1233                     dm_cblock_t cblock, dm_oblock_t oblock)
1234 {
1235         int r;
1236         __le64 value = pack_value(oblock, M_VALID);
1237         __dm_bless_for_disk(&value);
1238
1239         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1240                                &value, &cmd->root);
1241         if (r)
1242                 return r;
1243
1244         cmd->changed = true;
1245         return 0;
1246 }
1247
1248 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
1249                             dm_cblock_t cblock, dm_oblock_t oblock)
1250 {
1251         int r;
1252
1253         WRITE_LOCK(cmd);
1254         r = __insert(cmd, cblock, oblock);
1255         WRITE_UNLOCK(cmd);
1256
1257         return r;
1258 }
1259
1260 struct thunk {
1261         load_mapping_fn fn;
1262         void *context;
1263
1264         struct dm_cache_metadata *cmd;
1265         bool respect_dirty_flags;
1266         bool hints_valid;
1267 };
1268
1269 static bool policy_unchanged(struct dm_cache_metadata *cmd,
1270                              struct dm_cache_policy *policy)
1271 {
1272         const char *policy_name = dm_cache_policy_get_name(policy);
1273         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1274         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
1275
1276         /*
1277          * Ensure policy names match.
1278          */
1279         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
1280                 return false;
1281
1282         /*
1283          * Ensure policy major versions match.
1284          */
1285         if (cmd->policy_version[0] != policy_version[0])
1286                 return false;
1287
1288         /*
1289          * Ensure policy hint sizes match.
1290          */
1291         if (cmd->policy_hint_size != policy_hint_size)
1292                 return false;
1293
1294         return true;
1295 }
1296
1297 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
1298 {
1299         return cmd->hint_root && cmd->policy_hint_size;
1300 }
1301
1302 static bool hints_array_available(struct dm_cache_metadata *cmd,
1303                                   struct dm_cache_policy *policy)
1304 {
1305         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
1306                 hints_array_initialized(cmd);
1307 }
1308
1309 static int __load_mapping_v1(struct dm_cache_metadata *cmd,
1310                              uint64_t cb, bool hints_valid,
1311                              struct dm_array_cursor *mapping_cursor,
1312                              struct dm_array_cursor *hint_cursor,
1313                              load_mapping_fn fn, void *context)
1314 {
1315         int r = 0;
1316
1317         __le64 mapping;
1318         __le32 hint = 0;
1319
1320         __le64 *mapping_value_le;
1321         __le32 *hint_value_le;
1322
1323         dm_oblock_t oblock;
1324         unsigned flags;
1325
1326         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1327         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1328         unpack_value(mapping, &oblock, &flags);
1329
1330         if (flags & M_VALID) {
1331                 if (hints_valid) {
1332                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1333                         memcpy(&hint, hint_value_le, sizeof(hint));
1334                 }
1335
1336                 r = fn(context, oblock, to_cblock(cb), flags & M_DIRTY,
1337                        le32_to_cpu(hint), hints_valid);
1338                 if (r) {
1339                         DMERR("policy couldn't load cache block %llu",
1340                               (unsigned long long) from_cblock(to_cblock(cb)));
1341                 }
1342         }
1343
1344         return r;
1345 }
1346
1347 static int __load_mapping_v2(struct dm_cache_metadata *cmd,
1348                              uint64_t cb, bool hints_valid,
1349                              struct dm_array_cursor *mapping_cursor,
1350                              struct dm_array_cursor *hint_cursor,
1351                              struct dm_bitset_cursor *dirty_cursor,
1352                              load_mapping_fn fn, void *context)
1353 {
1354         int r = 0;
1355
1356         __le64 mapping;
1357         __le32 hint = 0;
1358
1359         __le64 *mapping_value_le;
1360         __le32 *hint_value_le;
1361
1362         dm_oblock_t oblock;
1363         unsigned flags;
1364         bool dirty;
1365
1366         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1367         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1368         unpack_value(mapping, &oblock, &flags);
1369
1370         if (flags & M_VALID) {
1371                 if (hints_valid) {
1372                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1373                         memcpy(&hint, hint_value_le, sizeof(hint));
1374                 }
1375
1376                 dirty = dm_bitset_cursor_get_value(dirty_cursor);
1377                 r = fn(context, oblock, to_cblock(cb), dirty,
1378                        le32_to_cpu(hint), hints_valid);
1379                 if (r) {
1380                         DMERR("policy couldn't load cache block %llu",
1381                               (unsigned long long) from_cblock(to_cblock(cb)));
1382                 }
1383         }
1384
1385         return r;
1386 }
1387
1388 static int __load_mappings(struct dm_cache_metadata *cmd,
1389                            struct dm_cache_policy *policy,
1390                            load_mapping_fn fn, void *context)
1391 {
1392         int r;
1393         uint64_t cb;
1394
1395         bool hints_valid = hints_array_available(cmd, policy);
1396
1397         if (from_cblock(cmd->cache_blocks) == 0)
1398                 /* Nothing to do */
1399                 return 0;
1400
1401         r = dm_array_cursor_begin(&cmd->info, cmd->root, &cmd->mapping_cursor);
1402         if (r)
1403                 return r;
1404
1405         if (hints_valid) {
1406                 r = dm_array_cursor_begin(&cmd->hint_info, cmd->hint_root, &cmd->hint_cursor);
1407                 if (r) {
1408                         dm_array_cursor_end(&cmd->mapping_cursor);
1409                         return r;
1410                 }
1411         }
1412
1413         if (separate_dirty_bits(cmd)) {
1414                 r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
1415                                            from_cblock(cmd->cache_blocks),
1416                                            &cmd->dirty_cursor);
1417                 if (r) {
1418                         dm_array_cursor_end(&cmd->hint_cursor);
1419                         dm_array_cursor_end(&cmd->mapping_cursor);
1420                         return r;
1421                 }
1422         }
1423
1424         for (cb = 0; ; cb++) {
1425                 if (separate_dirty_bits(cmd))
1426                         r = __load_mapping_v2(cmd, cb, hints_valid,
1427                                               &cmd->mapping_cursor,
1428                                               &cmd->hint_cursor,
1429                                               &cmd->dirty_cursor,
1430                                               fn, context);
1431                 else
1432                         r = __load_mapping_v1(cmd, cb, hints_valid,
1433                                               &cmd->mapping_cursor, &cmd->hint_cursor,
1434                                               fn, context);
1435                 if (r)
1436                         goto out;
1437
1438                 /*
1439                  * We need to break out before we move the cursors.
1440                  */
1441                 if (cb >= (from_cblock(cmd->cache_blocks) - 1))
1442                         break;
1443
1444                 r = dm_array_cursor_next(&cmd->mapping_cursor);
1445                 if (r) {
1446                         DMERR("dm_array_cursor_next for mapping failed");
1447                         goto out;
1448                 }
1449
1450                 if (hints_valid) {
1451                         r = dm_array_cursor_next(&cmd->hint_cursor);
1452                         if (r) {
1453                                 DMERR("dm_array_cursor_next for hint failed");
1454                                 goto out;
1455                         }
1456                 }
1457
1458                 if (separate_dirty_bits(cmd)) {
1459                         r = dm_bitset_cursor_next(&cmd->dirty_cursor);
1460                         if (r) {
1461                                 DMERR("dm_bitset_cursor_next for dirty failed");
1462                                 goto out;
1463                         }
1464                 }
1465         }
1466 out:
1467         dm_array_cursor_end(&cmd->mapping_cursor);
1468         if (hints_valid)
1469                 dm_array_cursor_end(&cmd->hint_cursor);
1470
1471         if (separate_dirty_bits(cmd))
1472                 dm_bitset_cursor_end(&cmd->dirty_cursor);
1473
1474         return r;
1475 }
1476
1477 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1478                            struct dm_cache_policy *policy,
1479                            load_mapping_fn fn, void *context)
1480 {
1481         int r;
1482
1483         READ_LOCK(cmd);
1484         r = __load_mappings(cmd, policy, fn, context);
1485         READ_UNLOCK(cmd);
1486
1487         return r;
1488 }
1489
1490 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1491 {
1492         int r = 0;
1493         __le64 value;
1494         dm_oblock_t oblock;
1495         unsigned flags;
1496
1497         memcpy(&value, leaf, sizeof(value));
1498         unpack_value(value, &oblock, &flags);
1499
1500         return r;
1501 }
1502
1503 static int __dump_mappings(struct dm_cache_metadata *cmd)
1504 {
1505         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1506 }
1507
1508 void dm_cache_dump(struct dm_cache_metadata *cmd)
1509 {
1510         READ_LOCK_VOID(cmd);
1511         __dump_mappings(cmd);
1512         READ_UNLOCK(cmd);
1513 }
1514
1515 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1516 {
1517         int r;
1518
1519         READ_LOCK(cmd);
1520         r = cmd->changed;
1521         READ_UNLOCK(cmd);
1522
1523         return r;
1524 }
1525
1526 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1527 {
1528         int r;
1529         unsigned flags;
1530         dm_oblock_t oblock;
1531         __le64 value;
1532
1533         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1534         if (r)
1535                 return r;
1536
1537         unpack_value(value, &oblock, &flags);
1538
1539         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1540                 /* nothing to be done */
1541                 return 0;
1542
1543         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1544         __dm_bless_for_disk(&value);
1545
1546         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1547                                &value, &cmd->root);
1548         if (r)
1549                 return r;
1550
1551         cmd->changed = true;
1552         return 0;
1553
1554 }
1555
1556 static int __set_dirty_bits_v1(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1557 {
1558         int r;
1559         unsigned i;
1560         for (i = 0; i < nr_bits; i++) {
1561                 r = __dirty(cmd, to_cblock(i), test_bit(i, bits));
1562                 if (r)
1563                         return r;
1564         }
1565
1566         return 0;
1567 }
1568
1569 static int is_dirty_callback(uint32_t index, bool *value, void *context)
1570 {
1571         unsigned long *bits = context;
1572         *value = test_bit(index, bits);
1573         return 0;
1574 }
1575
1576 static int __set_dirty_bits_v2(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1577 {
1578         int r = 0;
1579
1580         /* nr_bits is really just a sanity check */
1581         if (nr_bits != from_cblock(cmd->cache_blocks)) {
1582                 DMERR("dirty bitset is wrong size");
1583                 return -EINVAL;
1584         }
1585
1586         r = dm_bitset_del(&cmd->dirty_info, cmd->dirty_root);
1587         if (r)
1588                 return r;
1589
1590         cmd->changed = true;
1591         return dm_bitset_new(&cmd->dirty_info, &cmd->dirty_root, nr_bits, is_dirty_callback, bits);
1592 }
1593
1594 int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
1595                             unsigned nr_bits,
1596                             unsigned long *bits)
1597 {
1598         int r;
1599
1600         WRITE_LOCK(cmd);
1601         if (separate_dirty_bits(cmd))
1602                 r = __set_dirty_bits_v2(cmd, nr_bits, bits);
1603         else
1604                 r = __set_dirty_bits_v1(cmd, nr_bits, bits);
1605         WRITE_UNLOCK(cmd);
1606
1607         return r;
1608 }
1609
1610 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1611                                  struct dm_cache_statistics *stats)
1612 {
1613         READ_LOCK_VOID(cmd);
1614         *stats = cmd->stats;
1615         READ_UNLOCK(cmd);
1616 }
1617
1618 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1619                                  struct dm_cache_statistics *stats)
1620 {
1621         WRITE_LOCK_VOID(cmd);
1622         cmd->stats = *stats;
1623         WRITE_UNLOCK(cmd);
1624 }
1625
1626 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1627 {
1628         int r;
1629         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1630                                  clear_clean_shutdown);
1631
1632         WRITE_LOCK(cmd);
1633         r = __commit_transaction(cmd, mutator);
1634         if (r)
1635                 goto out;
1636
1637         r = __begin_transaction(cmd);
1638
1639 out:
1640         WRITE_UNLOCK(cmd);
1641         return r;
1642 }
1643
1644 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1645                                            dm_block_t *result)
1646 {
1647         int r = -EINVAL;
1648
1649         READ_LOCK(cmd);
1650         r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1651         READ_UNLOCK(cmd);
1652
1653         return r;
1654 }
1655
1656 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1657                                    dm_block_t *result)
1658 {
1659         int r = -EINVAL;
1660
1661         READ_LOCK(cmd);
1662         r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1663         READ_UNLOCK(cmd);
1664
1665         return r;
1666 }
1667
1668 /*----------------------------------------------------------------*/
1669
1670 static int get_hint(uint32_t index, void *value_le, void *context)
1671 {
1672         uint32_t value;
1673         struct dm_cache_policy *policy = context;
1674
1675         value = policy_get_hint(policy, to_cblock(index));
1676         *((__le32 *) value_le) = cpu_to_le32(value);
1677
1678         return 0;
1679 }
1680
1681 /*
1682  * It's quicker to always delete the hint array, and recreate with
1683  * dm_array_new().
1684  */
1685 static int write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1686 {
1687         int r;
1688         size_t hint_size;
1689         const char *policy_name = dm_cache_policy_get_name(policy);
1690         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1691
1692         if (!policy_name[0] ||
1693             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1694                 return -EINVAL;
1695
1696         strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1697         memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1698
1699         hint_size = dm_cache_policy_get_hint_size(policy);
1700         if (!hint_size)
1701                 return 0; /* short-circuit hints initialization */
1702         cmd->policy_hint_size = hint_size;
1703
1704         if (cmd->hint_root) {
1705                 r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1706                 if (r)
1707                         return r;
1708         }
1709
1710         return dm_array_new(&cmd->hint_info, &cmd->hint_root,
1711                             from_cblock(cmd->cache_blocks),
1712                             get_hint, policy);
1713 }
1714
1715 int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1716 {
1717         int r;
1718
1719         WRITE_LOCK(cmd);
1720         r = write_hints(cmd, policy);
1721         WRITE_UNLOCK(cmd);
1722
1723         return r;
1724 }
1725
1726 int dm_cache_metadata_all_clean(struct dm_cache_metadata *cmd, bool *result)
1727 {
1728         int r;
1729
1730         READ_LOCK(cmd);
1731         r = blocks_are_unmapped_or_clean(cmd, 0, cmd->cache_blocks, result);
1732         READ_UNLOCK(cmd);
1733
1734         return r;
1735 }
1736
1737 void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd)
1738 {
1739         WRITE_LOCK_VOID(cmd);
1740         dm_bm_set_read_only(cmd->bm);
1741         WRITE_UNLOCK(cmd);
1742 }
1743
1744 void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd)
1745 {
1746         WRITE_LOCK_VOID(cmd);
1747         dm_bm_set_read_write(cmd->bm);
1748         WRITE_UNLOCK(cmd);
1749 }
1750
1751 int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd)
1752 {
1753         int r;
1754         struct dm_block *sblock;
1755         struct cache_disk_superblock *disk_super;
1756
1757         WRITE_LOCK(cmd);
1758         set_bit(NEEDS_CHECK, &cmd->flags);
1759
1760         r = superblock_lock(cmd, &sblock);
1761         if (r) {
1762                 DMERR("couldn't read superblock");
1763                 goto out;
1764         }
1765
1766         disk_super = dm_block_data(sblock);
1767         disk_super->flags = cpu_to_le32(cmd->flags);
1768
1769         dm_bm_unlock(sblock);
1770
1771 out:
1772         WRITE_UNLOCK(cmd);
1773         return r;
1774 }
1775
1776 int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result)
1777 {
1778         READ_LOCK(cmd);
1779         *result = !!test_bit(NEEDS_CHECK, &cmd->flags);
1780         READ_UNLOCK(cmd);
1781
1782         return 0;
1783 }
1784
1785 int dm_cache_metadata_abort(struct dm_cache_metadata *cmd)
1786 {
1787         int r;
1788
1789         WRITE_LOCK(cmd);
1790         __destroy_persistent_data_objects(cmd);
1791         r = __create_persistent_data_objects(cmd, false);
1792         if (r)
1793                 cmd->fail_io = true;
1794         WRITE_UNLOCK(cmd);
1795
1796         return r;
1797 }