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(cmd->cache_blocks), &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                 begin = to_cblock(from_cblock(begin) + 1);
 963                 if (begin == end)
 964                         break;
 965
 966                 r = dm_bitset_cursor_next(&cmd->dirty_cursor);
 967                 if (r) {
 968                         DMERR("%s: dm_bitset_cursor_next for dirty failed", __func__);
 969                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 970                         return r;
 971                 }
 972         }
 973
 974         dm_bitset_cursor_end(&cmd->dirty_cursor);
 975
 976         return 0;
 977 }
 978
 979 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
 980                                         dm_cblock_t begin, dm_cblock_t end,
 981                                         bool *result)
 982 {
 983         if (separate_dirty_bits(cmd))
 984                 return blocks_are_clean_separate_dirty(cmd, begin, end, result);
 985         else
 986                 return blocks_are_clean_combined_dirty(cmd, begin, end, result);
 987 }
 988
 989 static bool cmd_write_lock(struct dm_cache_metadata *cmd)
 990 {
 991         down_write(&cmd->root_lock);
 992         if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
 993                 up_write(&cmd->root_lock);
 994                 return false;
 995         }
 996         return true;
 997 }
 998
 999 #define WRITE_LOCK(cmd)                         \
1000         do {                                    \
1001                 if (!cmd_write_lock((cmd)))     \
1002                         return -EINVAL;         \
1003         } while(0)
1004
1005 #define WRITE_LOCK_VOID(cmd)                    \
1006         do {                                    \
1007                 if (!cmd_write_lock((cmd)))     \
1008                         return;                 \
1009         } while(0)
1010
1011 #define WRITE_UNLOCK(cmd) \
1012         up_write(&(cmd)->root_lock)
1013
1014 static bool cmd_read_lock(struct dm_cache_metadata *cmd)
1015 {
1016         down_read(&cmd->root_lock);
1017         if (cmd->fail_io) {
1018                 up_read(&cmd->root_lock);
1019                 return false;
1020         }
1021         return true;
1022 }
1023
1024 #define READ_LOCK(cmd)                          \
1025         do {                                    \
1026                 if (!cmd_read_lock((cmd)))      \
1027                         return -EINVAL;         \
1028         } while(0)
1029
1030 #define READ_LOCK_VOID(cmd)                     \
1031         do {                                    \
1032                 if (!cmd_read_lock((cmd)))      \
1033                         return;                 \
1034         } while(0)
1035
1036 #define READ_UNLOCK(cmd) \
1037         up_read(&(cmd)->root_lock)
1038
1039 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
1040 {
1041         int r;
1042         bool clean;
1043         __le64 null_mapping = pack_value(0, 0);
1044
1045         WRITE_LOCK(cmd);
1046         __dm_bless_for_disk(&null_mapping);
1047
1048         if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
1049                 r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
1050                 if (r) {
1051                         __dm_unbless_for_disk(&null_mapping);
1052                         goto out;
1053                 }
1054
1055                 if (!clean) {
1056                         DMERR("unable to shrink cache due to dirty blocks");
1057                         r = -EINVAL;
1058                         __dm_unbless_for_disk(&null_mapping);
1059                         goto out;
1060                 }
1061         }
1062
1063         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
1064                             from_cblock(new_cache_size),
1065                             &null_mapping, &cmd->root);
1066         if (r)
1067                 goto out;
1068
1069         if (separate_dirty_bits(cmd)) {
1070                 r = dm_bitset_resize(&cmd->dirty_info, cmd->dirty_root,
1071                                      from_cblock(cmd->cache_blocks), from_cblock(new_cache_size),
1072                                      false, &cmd->dirty_root);
1073                 if (r)
1074                         goto out;
1075         }
1076
1077         cmd->cache_blocks = new_cache_size;
1078         cmd->changed = true;
1079
1080 out:
1081         WRITE_UNLOCK(cmd);
1082
1083         return r;
1084 }
1085
1086 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
1087                                    sector_t discard_block_size,
1088                                    dm_dblock_t new_nr_entries)
1089 {
1090         int r;
1091
1092         WRITE_LOCK(cmd);
1093         r = dm_bitset_resize(&cmd->discard_info,
1094                              cmd->discard_root,
1095                              from_dblock(cmd->discard_nr_blocks),
1096                              from_dblock(new_nr_entries),
1097                              false, &cmd->discard_root);
1098         if (!r) {
1099                 cmd->discard_block_size = discard_block_size;
1100                 cmd->discard_nr_blocks = new_nr_entries;
1101         }
1102
1103         cmd->changed = true;
1104         WRITE_UNLOCK(cmd);
1105
1106         return r;
1107 }
1108
1109 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1110 {
1111         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
1112                                  from_dblock(b), &cmd->discard_root);
1113 }
1114
1115 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1116 {
1117         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
1118                                    from_dblock(b), &cmd->discard_root);
1119 }
1120
1121 static int __discard(struct dm_cache_metadata *cmd,
1122                      dm_dblock_t dblock, bool discard)
1123 {
1124         int r;
1125
1126         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
1127         if (r)
1128                 return r;
1129
1130         cmd->changed = true;
1131         return 0;
1132 }
1133
1134 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
1135                          dm_dblock_t dblock, bool discard)
1136 {
1137         int r;
1138
1139         WRITE_LOCK(cmd);
1140         r = __discard(cmd, dblock, discard);
1141         WRITE_UNLOCK(cmd);
1142
1143         return r;
1144 }
1145
1146 static int __load_discards(struct dm_cache_metadata *cmd,
1147                            load_discard_fn fn, void *context)
1148 {
1149         int r = 0;
1150         uint32_t b;
1151         struct dm_bitset_cursor c;
1152
1153         if (from_dblock(cmd->discard_nr_blocks) == 0)
1154                 /* nothing to do */
1155                 return 0;
1156
1157         if (cmd->clean_when_opened) {
1158                 r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root, &cmd->discard_root);
1159                 if (r)
1160                         return r;
1161
1162                 r = dm_bitset_cursor_begin(&cmd->discard_info, cmd->discard_root,
1163                                            from_dblock(cmd->discard_nr_blocks), &c);
1164                 if (r)
1165                         return r;
1166
1167                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1168                         r = fn(context, cmd->discard_block_size, to_dblock(b),
1169                                dm_bitset_cursor_get_value(&c));
1170                         if (r)
1171                                 break;
1172                 }
1173
1174                 dm_bitset_cursor_end(&c);
1175
1176         } else {
1177                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1178                         r = fn(context, cmd->discard_block_size, to_dblock(b), false);
1179                         if (r)
1180                                 return r;
1181                 }
1182         }
1183
1184         return r;
1185 }
1186
1187 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
1188                            load_discard_fn fn, void *context)
1189 {
1190         int r;
1191
1192         READ_LOCK(cmd);
1193         r = __load_discards(cmd, fn, context);
1194         READ_UNLOCK(cmd);
1195
1196         return r;
1197 }
1198
1199 int dm_cache_size(struct dm_cache_metadata *cmd, dm_cblock_t *result)
1200 {
1201         READ_LOCK(cmd);
1202         *result = cmd->cache_blocks;
1203         READ_UNLOCK(cmd);
1204
1205         return 0;
1206 }
1207
1208 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1209 {
1210         int r;
1211         __le64 value = pack_value(0, 0);
1212
1213         __dm_bless_for_disk(&value);
1214         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1215                                &value, &cmd->root);
1216         if (r)
1217                 return r;
1218
1219         cmd->changed = true;
1220         return 0;
1221 }
1222
1223 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1224 {
1225         int r;
1226
1227         WRITE_LOCK(cmd);
1228         r = __remove(cmd, cblock);
1229         WRITE_UNLOCK(cmd);
1230
1231         return r;
1232 }
1233
1234 static int __insert(struct dm_cache_metadata *cmd,
1235                     dm_cblock_t cblock, dm_oblock_t oblock)
1236 {
1237         int r;
1238         __le64 value = pack_value(oblock, M_VALID);
1239         __dm_bless_for_disk(&value);
1240
1241         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1242                                &value, &cmd->root);
1243         if (r)
1244                 return r;
1245
1246         cmd->changed = true;
1247         return 0;
1248 }
1249
1250 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
1251                             dm_cblock_t cblock, dm_oblock_t oblock)
1252 {
1253         int r;
1254
1255         WRITE_LOCK(cmd);
1256         r = __insert(cmd, cblock, oblock);
1257         WRITE_UNLOCK(cmd);
1258
1259         return r;
1260 }
1261
1262 struct thunk {
1263         load_mapping_fn fn;
1264         void *context;
1265
1266         struct dm_cache_metadata *cmd;
1267         bool respect_dirty_flags;
1268         bool hints_valid;
1269 };
1270
1271 static bool policy_unchanged(struct dm_cache_metadata *cmd,
1272                              struct dm_cache_policy *policy)
1273 {
1274         const char *policy_name = dm_cache_policy_get_name(policy);
1275         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1276         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
1277
1278         /*
1279          * Ensure policy names match.
1280          */
1281         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
1282                 return false;
1283
1284         /*
1285          * Ensure policy major versions match.
1286          */
1287         if (cmd->policy_version[0] != policy_version[0])
1288                 return false;
1289
1290         /*
1291          * Ensure policy hint sizes match.
1292          */
1293         if (cmd->policy_hint_size != policy_hint_size)
1294                 return false;
1295
1296         return true;
1297 }
1298
1299 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
1300 {
1301         return cmd->hint_root && cmd->policy_hint_size;
1302 }
1303
1304 static bool hints_array_available(struct dm_cache_metadata *cmd,
1305                                   struct dm_cache_policy *policy)
1306 {
1307         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
1308                 hints_array_initialized(cmd);
1309 }
1310
1311 static int __load_mapping_v1(struct dm_cache_metadata *cmd,
1312                              uint64_t cb, bool hints_valid,
1313                              struct dm_array_cursor *mapping_cursor,
1314                              struct dm_array_cursor *hint_cursor,
1315                              load_mapping_fn fn, void *context)
1316 {
1317         int r = 0;
1318
1319         __le64 mapping;
1320         __le32 hint = 0;
1321
1322         __le64 *mapping_value_le;
1323         __le32 *hint_value_le;
1324
1325         dm_oblock_t oblock;
1326         unsigned flags;
1327
1328         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1329         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1330         unpack_value(mapping, &oblock, &flags);
1331
1332         if (flags & M_VALID) {
1333                 if (hints_valid) {
1334                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1335                         memcpy(&hint, hint_value_le, sizeof(hint));
1336                 }
1337
1338                 r = fn(context, oblock, to_cblock(cb), flags & M_DIRTY,
1339                        le32_to_cpu(hint), hints_valid);
1340                 if (r) {
1341                         DMERR("policy couldn't load cache block %llu",
1342                               (unsigned long long) from_cblock(to_cblock(cb)));
1343                 }
1344         }
1345
1346         return r;
1347 }
1348
1349 static int __load_mapping_v2(struct dm_cache_metadata *cmd,
1350                              uint64_t cb, bool hints_valid,
1351                              struct dm_array_cursor *mapping_cursor,
1352                              struct dm_array_cursor *hint_cursor,
1353                              struct dm_bitset_cursor *dirty_cursor,
1354                              load_mapping_fn fn, void *context)
1355 {
1356         int r = 0;
1357
1358         __le64 mapping;
1359         __le32 hint = 0;
1360
1361         __le64 *mapping_value_le;
1362         __le32 *hint_value_le;
1363
1364         dm_oblock_t oblock;
1365         unsigned flags;
1366         bool dirty;
1367
1368         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1369         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1370         unpack_value(mapping, &oblock, &flags);
1371
1372         if (flags & M_VALID) {
1373                 if (hints_valid) {
1374                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1375                         memcpy(&hint, hint_value_le, sizeof(hint));
1376                 }
1377
1378                 dirty = dm_bitset_cursor_get_value(dirty_cursor);
1379                 r = fn(context, oblock, to_cblock(cb), dirty,
1380                        le32_to_cpu(hint), hints_valid);
1381                 if (r) {
1382                         DMERR("policy couldn't load cache block %llu",
1383                               (unsigned long long) from_cblock(to_cblock(cb)));
1384                 }
1385         }
1386
1387         return r;
1388 }
1389
1390 static int __load_mappings(struct dm_cache_metadata *cmd,
1391                            struct dm_cache_policy *policy,
1392                            load_mapping_fn fn, void *context)
1393 {
1394         int r;
1395         uint64_t cb;
1396
1397         bool hints_valid = hints_array_available(cmd, policy);
1398
1399         if (from_cblock(cmd->cache_blocks) == 0)
1400                 /* Nothing to do */
1401                 return 0;
1402
1403         r = dm_array_cursor_begin(&cmd->info, cmd->root, &cmd->mapping_cursor);
1404         if (r)
1405                 return r;
1406
1407         if (hints_valid) {
1408                 r = dm_array_cursor_begin(&cmd->hint_info, cmd->hint_root, &cmd->hint_cursor);
1409                 if (r) {
1410                         dm_array_cursor_end(&cmd->mapping_cursor);
1411                         return r;
1412                 }
1413         }
1414
1415         if (separate_dirty_bits(cmd)) {
1416                 r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
1417                                            from_cblock(cmd->cache_blocks),
1418                                            &cmd->dirty_cursor);
1419                 if (r) {
1420                         dm_array_cursor_end(&cmd->hint_cursor);
1421                         dm_array_cursor_end(&cmd->mapping_cursor);
1422                         return r;
1423                 }
1424         }
1425
1426         for (cb = 0; ; cb++) {
1427                 if (separate_dirty_bits(cmd))
1428                         r = __load_mapping_v2(cmd, cb, hints_valid,
1429                                               &cmd->mapping_cursor,
1430                                               &cmd->hint_cursor,
1431                                               &cmd->dirty_cursor,
1432                                               fn, context);
1433                 else
1434                         r = __load_mapping_v1(cmd, cb, hints_valid,
1435                                               &cmd->mapping_cursor, &cmd->hint_cursor,
1436                                               fn, context);
1437                 if (r)
1438                         goto out;
1439
1440                 /*
1441                  * We need to break out before we move the cursors.
1442                  */
1443                 if (cb >= (from_cblock(cmd->cache_blocks) - 1))
1444                         break;
1445
1446                 r = dm_array_cursor_next(&cmd->mapping_cursor);
1447                 if (r) {
1448                         DMERR("dm_array_cursor_next for mapping failed");
1449                         goto out;
1450                 }
1451
1452                 if (hints_valid) {
1453                         r = dm_array_cursor_next(&cmd->hint_cursor);
1454                         if (r) {
1455                                 DMERR("dm_array_cursor_next for hint failed");
1456                                 goto out;
1457                         }
1458                 }
1459
1460                 if (separate_dirty_bits(cmd)) {
1461                         r = dm_bitset_cursor_next(&cmd->dirty_cursor);
1462                         if (r) {
1463                                 DMERR("dm_bitset_cursor_next for dirty failed");
1464                                 goto out;
1465                         }
1466                 }
1467         }
1468 out:
1469         dm_array_cursor_end(&cmd->mapping_cursor);
1470         if (hints_valid)
1471                 dm_array_cursor_end(&cmd->hint_cursor);
1472
1473         if (separate_dirty_bits(cmd))
1474                 dm_bitset_cursor_end(&cmd->dirty_cursor);
1475
1476         return r;
1477 }
1478
1479 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1480                            struct dm_cache_policy *policy,
1481                            load_mapping_fn fn, void *context)
1482 {
1483         int r;
1484
1485         READ_LOCK(cmd);
1486         r = __load_mappings(cmd, policy, fn, context);
1487         READ_UNLOCK(cmd);
1488
1489         return r;
1490 }
1491
1492 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1493 {
1494         int r = 0;
1495         __le64 value;
1496         dm_oblock_t oblock;
1497         unsigned flags;
1498
1499         memcpy(&value, leaf, sizeof(value));
1500         unpack_value(value, &oblock, &flags);
1501
1502         return r;
1503 }
1504
1505 static int __dump_mappings(struct dm_cache_metadata *cmd)
1506 {
1507         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1508 }
1509
1510 void dm_cache_dump(struct dm_cache_metadata *cmd)
1511 {
1512         READ_LOCK_VOID(cmd);
1513         __dump_mappings(cmd);
1514         READ_UNLOCK(cmd);
1515 }
1516
1517 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1518 {
1519         int r;
1520
1521         READ_LOCK(cmd);
1522         r = cmd->changed;
1523         READ_UNLOCK(cmd);
1524
1525         return r;
1526 }
1527
1528 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1529 {
1530         int r;
1531         unsigned flags;
1532         dm_oblock_t oblock;
1533         __le64 value;
1534
1535         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1536         if (r)
1537                 return r;
1538
1539         unpack_value(value, &oblock, &flags);
1540
1541         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1542                 /* nothing to be done */
1543                 return 0;
1544
1545         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1546         __dm_bless_for_disk(&value);
1547
1548         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1549                                &value, &cmd->root);
1550         if (r)
1551                 return r;
1552
1553         cmd->changed = true;
1554         return 0;
1555
1556 }
1557
1558 static int __set_dirty_bits_v1(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1559 {
1560         int r;
1561         unsigned i;
1562         for (i = 0; i < nr_bits; i++) {
1563                 r = __dirty(cmd, to_cblock(i), test_bit(i, bits));
1564                 if (r)
1565                         return r;
1566         }
1567
1568         return 0;
1569 }
1570
1571 static int is_dirty_callback(uint32_t index, bool *value, void *context)
1572 {
1573         unsigned long *bits = context;
1574         *value = test_bit(index, bits);
1575         return 0;
1576 }
1577
1578 static int __set_dirty_bits_v2(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1579 {
1580         int r = 0;
1581
1582         /* nr_bits is really just a sanity check */
1583         if (nr_bits != from_cblock(cmd->cache_blocks)) {
1584                 DMERR("dirty bitset is wrong size");
1585                 return -EINVAL;
1586         }
1587
1588         r = dm_bitset_del(&cmd->dirty_info, cmd->dirty_root);
1589         if (r)
1590                 return r;
1591
1592         cmd->changed = true;
1593         return dm_bitset_new(&cmd->dirty_info, &cmd->dirty_root, nr_bits, is_dirty_callback, bits);
1594 }
1595
1596 int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
1597                             unsigned nr_bits,
1598                             unsigned long *bits)
1599 {
1600         int r;
1601
1602         WRITE_LOCK(cmd);
1603         if (separate_dirty_bits(cmd))
1604                 r = __set_dirty_bits_v2(cmd, nr_bits, bits);
1605         else
1606                 r = __set_dirty_bits_v1(cmd, nr_bits, bits);
1607         WRITE_UNLOCK(cmd);
1608
1609         return r;
1610 }
1611
1612 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1613                                  struct dm_cache_statistics *stats)
1614 {
1615         READ_LOCK_VOID(cmd);
1616         *stats = cmd->stats;
1617         READ_UNLOCK(cmd);
1618 }
1619
1620 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1621                                  struct dm_cache_statistics *stats)
1622 {
1623         WRITE_LOCK_VOID(cmd);
1624         cmd->stats = *stats;
1625         WRITE_UNLOCK(cmd);
1626 }
1627
1628 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1629 {
1630         int r;
1631         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1632                                  clear_clean_shutdown);
1633
1634         WRITE_LOCK(cmd);
1635         r = __commit_transaction(cmd, mutator);
1636         if (r)
1637                 goto out;
1638
1639         r = __begin_transaction(cmd);
1640
1641 out:
1642         WRITE_UNLOCK(cmd);
1643         return r;
1644 }
1645
1646 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1647                                            dm_block_t *result)
1648 {
1649         int r = -EINVAL;
1650
1651         READ_LOCK(cmd);
1652         r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1653         READ_UNLOCK(cmd);
1654
1655         return r;
1656 }
1657
1658 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1659                                    dm_block_t *result)
1660 {
1661         int r = -EINVAL;
1662
1663         READ_LOCK(cmd);
1664         r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1665         READ_UNLOCK(cmd);
1666
1667         return r;
1668 }
1669
1670 /*----------------------------------------------------------------*/
1671
1672 static int get_hint(uint32_t index, void *value_le, void *context)
1673 {
1674         uint32_t value;
1675         struct dm_cache_policy *policy = context;
1676
1677         value = policy_get_hint(policy, to_cblock(index));
1678         *((__le32 *) value_le) = cpu_to_le32(value);
1679
1680         return 0;
1681 }
1682
1683 /*
1684  * It's quicker to always delete the hint array, and recreate with
1685  * dm_array_new().
1686  */
1687 static int write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1688 {
1689         int r;
1690         size_t hint_size;
1691         const char *policy_name = dm_cache_policy_get_name(policy);
1692         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1693
1694         if (!policy_name[0] ||
1695             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1696                 return -EINVAL;
1697
1698         strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1699         memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1700
1701         hint_size = dm_cache_policy_get_hint_size(policy);
1702         if (!hint_size)
1703                 return 0; /* short-circuit hints initialization */
1704         cmd->policy_hint_size = hint_size;
1705
1706         if (cmd->hint_root) {
1707                 r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1708                 if (r)
1709                         return r;
1710         }
1711
1712         return dm_array_new(&cmd->hint_info, &cmd->hint_root,
1713                             from_cblock(cmd->cache_blocks),
1714                             get_hint, policy);
1715 }
1716
1717 int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1718 {
1719         int r;
1720
1721         WRITE_LOCK(cmd);
1722         r = write_hints(cmd, policy);
1723         WRITE_UNLOCK(cmd);
1724
1725         return r;
1726 }
1727
1728 int dm_cache_metadata_all_clean(struct dm_cache_metadata *cmd, bool *result)
1729 {
1730         int r;
1731
1732         READ_LOCK(cmd);
1733         r = blocks_are_unmapped_or_clean(cmd, 0, cmd->cache_blocks, result);
1734         READ_UNLOCK(cmd);
1735
1736         return r;
1737 }
1738
1739 void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd)
1740 {
1741         WRITE_LOCK_VOID(cmd);
1742         dm_bm_set_read_only(cmd->bm);
1743         WRITE_UNLOCK(cmd);
1744 }
1745
1746 void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd)
1747 {
1748         WRITE_LOCK_VOID(cmd);
1749         dm_bm_set_read_write(cmd->bm);
1750         WRITE_UNLOCK(cmd);
1751 }
1752
1753 int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd)
1754 {
1755         int r;
1756         struct dm_block *sblock;
1757         struct cache_disk_superblock *disk_super;
1758
1759         WRITE_LOCK(cmd);
1760         set_bit(NEEDS_CHECK, &cmd->flags);
1761
1762         r = superblock_lock(cmd, &sblock);
1763         if (r) {
1764                 DMERR("couldn't read superblock");
1765                 goto out;
1766         }
1767
1768         disk_super = dm_block_data(sblock);
1769         disk_super->flags = cpu_to_le32(cmd->flags);
1770
1771         dm_bm_unlock(sblock);
1772
1773 out:
1774         WRITE_UNLOCK(cmd);
1775         return r;
1776 }
1777
1778 int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result)
1779 {
1780         READ_LOCK(cmd);
1781         *result = !!test_bit(NEEDS_CHECK, &cmd->flags);
1782         READ_UNLOCK(cmd);
1783
1784         return 0;
1785 }
1786
1787 int dm_cache_metadata_abort(struct dm_cache_metadata *cmd)
1788 {
1789         int r;
1790
1791         WRITE_LOCK(cmd);
1792         __destroy_persistent_data_objects(cmd);
1793         r = __create_persistent_data_objects(cmd, false);
1794         if (r)
1795                 cmd->fail_io = true;
1796         WRITE_UNLOCK(cmd);
1797
1798         return r;
1799 }