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 /*
  31  *  3 for btree insert +
  32  *  2 for btree lookup used within space map
  33  */
  34 #define CACHE_MAX_CONCURRENT_LOCKS 5
  35 #define SPACE_MAP_ROOT_SIZE 128
  36
  37 enum superblock_flag_bits {
  38         /* for spotting crashes that would invalidate the dirty bitset */
  39         CLEAN_SHUTDOWN,
  40         /* metadata must be checked using the tools */
  41         NEEDS_CHECK,
  42 };
  43
  44 /*
  45  * Each mapping from cache block -> origin block carries a set of flags.
  46  */
  47 enum mapping_bits {
  48         /*
  49          * A valid mapping.  Because we're using an array we clear this
  50          * flag for an non existant mapping.
  51          */
  52         M_VALID = 1,
  53
  54         /*
  55          * The data on the cache is different from that on the origin.
  56          * This flag is only used by metadata format 1.
  57          */
  58         M_DIRTY = 2
  59 };
  60
  61 struct cache_disk_superblock {
  62         __le32 csum;
  63         __le32 flags;
  64         __le64 blocknr;
  65
  66         __u8 uuid[16];
  67         __le64 magic;
  68         __le32 version;
  69
  70         __u8 policy_name[CACHE_POLICY_NAME_SIZE];
  71         __le32 policy_hint_size;
  72
  73         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
  74         __le64 mapping_root;
  75         __le64 hint_root;
  76
  77         __le64 discard_root;
  78         __le64 discard_block_size;
  79         __le64 discard_nr_blocks;
  80
  81         __le32 data_block_size;
  82         __le32 metadata_block_size;
  83         __le32 cache_blocks;
  84
  85         __le32 compat_flags;
  86         __le32 compat_ro_flags;
  87         __le32 incompat_flags;
  88
  89         __le32 read_hits;
  90         __le32 read_misses;
  91         __le32 write_hits;
  92         __le32 write_misses;
  93
  94         __le32 policy_version[CACHE_POLICY_VERSION_SIZE];
  95
  96         /*
  97          * Metadata format 2 fields.
  98          */
  99         __le64 dirty_root;
 100 } __packed;
 101
 102 struct dm_cache_metadata {
 103         atomic_t ref_count;
 104         struct list_head list;
 105
 106         unsigned version;
 107         struct block_device *bdev;
 108         struct dm_block_manager *bm;
 109         struct dm_space_map *metadata_sm;
 110         struct dm_transaction_manager *tm;
 111
 112         struct dm_array_info info;
 113         struct dm_array_info hint_info;
 114         struct dm_disk_bitset discard_info;
 115
 116         struct rw_semaphore root_lock;
 117         unsigned long flags;
 118         dm_block_t root;
 119         dm_block_t hint_root;
 120         dm_block_t discard_root;
 121
 122         sector_t discard_block_size;
 123         dm_dblock_t discard_nr_blocks;
 124
 125         sector_t data_block_size;
 126         dm_cblock_t cache_blocks;
 127         bool changed:1;
 128         bool clean_when_opened:1;
 129
 130         char policy_name[CACHE_POLICY_NAME_SIZE];
 131         unsigned policy_version[CACHE_POLICY_VERSION_SIZE];
 132         size_t policy_hint_size;
 133         struct dm_cache_statistics stats;
 134
 135         /*
 136          * Reading the space map root can fail, so we read it into this
 137          * buffer before the superblock is locked and updated.
 138          */
 139         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 140
 141         /*
 142          * Set if a transaction has to be aborted but the attempt to roll
 143          * back to the previous (good) transaction failed.  The only
 144          * metadata operation permissible in this state is the closing of
 145          * the device.
 146          */
 147         bool fail_io:1;
 148
 149         /*
 150          * Metadata format 2 fields.
 151          */
 152         dm_block_t dirty_root;
 153         struct dm_disk_bitset dirty_info;
 154
 155         /*
 156          * These structures are used when loading metadata.  They're too
 157          * big to put on the stack.
 158          */
 159         struct dm_array_cursor mapping_cursor;
 160         struct dm_array_cursor hint_cursor;
 161         struct dm_bitset_cursor dirty_cursor;
 162 };
 163
 164 /*-------------------------------------------------------------------
 165  * superblock validator
 166  *-----------------------------------------------------------------*/
 167
 168 #define SUPERBLOCK_CSUM_XOR 9031977
 169
 170 static void sb_prepare_for_write(struct dm_block_validator *v,
 171                                  struct dm_block *b,
 172                                  size_t sb_block_size)
 173 {
 174         struct cache_disk_superblock *disk_super = dm_block_data(b);
 175
 176         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 177         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 178                                                       sb_block_size - sizeof(__le32),
 179                                                       SUPERBLOCK_CSUM_XOR));
 180 }
 181
 182 static int check_metadata_version(struct cache_disk_superblock *disk_super)
 183 {
 184         uint32_t metadata_version = le32_to_cpu(disk_super->version);
 185
 186         if (metadata_version < MIN_CACHE_VERSION || metadata_version > MAX_CACHE_VERSION) {
 187                 DMERR("Cache metadata version %u found, but only versions between %u and %u supported.",
 188                       metadata_version, MIN_CACHE_VERSION, MAX_CACHE_VERSION);
 189                 return -EINVAL;
 190         }
 191
 192         return 0;
 193 }
 194
 195 static int sb_check(struct dm_block_validator *v,
 196                     struct dm_block *b,
 197                     size_t sb_block_size)
 198 {
 199         struct cache_disk_superblock *disk_super = dm_block_data(b);
 200         __le32 csum_le;
 201
 202         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 203                 DMERR("sb_check failed: blocknr %llu: wanted %llu",
 204                       le64_to_cpu(disk_super->blocknr),
 205                       (unsigned long long)dm_block_location(b));
 206                 return -ENOTBLK;
 207         }
 208
 209         if (le64_to_cpu(disk_super->magic) != CACHE_SUPERBLOCK_MAGIC) {
 210                 DMERR("sb_check failed: magic %llu: wanted %llu",
 211                       le64_to_cpu(disk_super->magic),
 212                       (unsigned long long)CACHE_SUPERBLOCK_MAGIC);
 213                 return -EILSEQ;
 214         }
 215
 216         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 217                                              sb_block_size - sizeof(__le32),
 218                                              SUPERBLOCK_CSUM_XOR));
 219         if (csum_le != disk_super->csum) {
 220                 DMERR("sb_check failed: csum %u: wanted %u",
 221                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 222                 return -EILSEQ;
 223         }
 224
 225         return check_metadata_version(disk_super);
 226 }
 227
 228 static struct dm_block_validator sb_validator = {
 229         .name = "superblock",
 230         .prepare_for_write = sb_prepare_for_write,
 231         .check = sb_check
 232 };
 233
 234 /*----------------------------------------------------------------*/
 235
 236 static int superblock_read_lock(struct dm_cache_metadata *cmd,
 237                                 struct dm_block **sblock)
 238 {
 239         return dm_bm_read_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 240                                &sb_validator, sblock);
 241 }
 242
 243 static int superblock_lock_zero(struct dm_cache_metadata *cmd,
 244                                 struct dm_block **sblock)
 245 {
 246         return dm_bm_write_lock_zero(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 247                                      &sb_validator, sblock);
 248 }
 249
 250 static int superblock_lock(struct dm_cache_metadata *cmd,
 251                            struct dm_block **sblock)
 252 {
 253         return dm_bm_write_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 254                                 &sb_validator, sblock);
 255 }
 256
 257 /*----------------------------------------------------------------*/
 258
 259 static int __superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
 260 {
 261         int r;
 262         unsigned i;
 263         struct dm_block *b;
 264         __le64 *data_le, zero = cpu_to_le64(0);
 265         unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
 266
 267         /*
 268          * We can't use a validator here - it may be all zeroes.
 269          */
 270         r = dm_bm_read_lock(bm, CACHE_SUPERBLOCK_LOCATION, NULL, &b);
 271         if (r)
 272                 return r;
 273
 274         data_le = dm_block_data(b);
 275         *result = true;
 276         for (i = 0; i < sb_block_size; i++) {
 277                 if (data_le[i] != zero) {
 278                         *result = false;
 279                         break;
 280                 }
 281         }
 282
 283         dm_bm_unlock(b);
 284
 285         return 0;
 286 }
 287
 288 static void __setup_mapping_info(struct dm_cache_metadata *cmd)
 289 {
 290         struct dm_btree_value_type vt;
 291
 292         vt.context = NULL;
 293         vt.size = sizeof(__le64);
 294         vt.inc = NULL;
 295         vt.dec = NULL;
 296         vt.equal = NULL;
 297         dm_array_info_init(&cmd->info, cmd->tm, &vt);
 298
 299         if (cmd->policy_hint_size) {
 300                 vt.size = sizeof(__le32);
 301                 dm_array_info_init(&cmd->hint_info, cmd->tm, &vt);
 302         }
 303 }
 304
 305 static int __save_sm_root(struct dm_cache_metadata *cmd)
 306 {
 307         int r;
 308         size_t metadata_len;
 309
 310         r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
 311         if (r < 0)
 312                 return r;
 313
 314         return dm_sm_copy_root(cmd->metadata_sm, &cmd->metadata_space_map_root,
 315                                metadata_len);
 316 }
 317
 318 static void __copy_sm_root(struct dm_cache_metadata *cmd,
 319                            struct cache_disk_superblock *disk_super)
 320 {
 321         memcpy(&disk_super->metadata_space_map_root,
 322                &cmd->metadata_space_map_root,
 323                sizeof(cmd->metadata_space_map_root));
 324 }
 325
 326 static bool separate_dirty_bits(struct dm_cache_metadata *cmd)
 327 {
 328         return cmd->version >= 2;
 329 }
 330
 331 static int __write_initial_superblock(struct dm_cache_metadata *cmd)
 332 {
 333         int r;
 334         struct dm_block *sblock;
 335         struct cache_disk_superblock *disk_super;
 336         sector_t bdev_size = i_size_read(cmd->bdev->bd_inode) >> SECTOR_SHIFT;
 337
 338         /* FIXME: see if we can lose the max sectors limit */
 339         if (bdev_size > DM_CACHE_METADATA_MAX_SECTORS)
 340                 bdev_size = DM_CACHE_METADATA_MAX_SECTORS;
 341
 342         r = dm_tm_pre_commit(cmd->tm);
 343         if (r < 0)
 344                 return r;
 345
 346         /*
 347          * dm_sm_copy_root() can fail.  So we need to do it before we start
 348          * updating the superblock.
 349          */
 350         r = __save_sm_root(cmd);
 351         if (r)
 352                 return r;
 353
 354         r = superblock_lock_zero(cmd, &sblock);
 355         if (r)
 356                 return r;
 357
 358         disk_super = dm_block_data(sblock);
 359         disk_super->flags = 0;
 360         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 361         disk_super->magic = cpu_to_le64(CACHE_SUPERBLOCK_MAGIC);
 362         disk_super->version = cpu_to_le32(cmd->version);
 363         memset(disk_super->policy_name, 0, sizeof(disk_super->policy_name));
 364         memset(disk_super->policy_version, 0, sizeof(disk_super->policy_version));
 365         disk_super->policy_hint_size = cpu_to_le32(0);
 366
 367         __copy_sm_root(cmd, disk_super);
 368
 369         disk_super->mapping_root = cpu_to_le64(cmd->root);
 370         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 371         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 372         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 373         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 374         disk_super->metadata_block_size = cpu_to_le32(DM_CACHE_METADATA_BLOCK_SIZE);
 375         disk_super->data_block_size = cpu_to_le32(cmd->data_block_size);
 376         disk_super->cache_blocks = cpu_to_le32(0);
 377
 378         disk_super->read_hits = cpu_to_le32(0);
 379         disk_super->read_misses = cpu_to_le32(0);
 380         disk_super->write_hits = cpu_to_le32(0);
 381         disk_super->write_misses = cpu_to_le32(0);
 382
 383         if (separate_dirty_bits(cmd))
 384                 disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
 385
 386         return dm_tm_commit(cmd->tm, sblock);
 387 }
 388
 389 static int __format_metadata(struct dm_cache_metadata *cmd)
 390 {
 391         int r;
 392
 393         r = dm_tm_create_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 394                                  &cmd->tm, &cmd->metadata_sm);
 395         if (r < 0) {
 396                 DMERR("tm_create_with_sm failed");
 397                 return r;
 398         }
 399
 400         __setup_mapping_info(cmd);
 401
 402         r = dm_array_empty(&cmd->info, &cmd->root);
 403         if (r < 0)
 404                 goto bad;
 405
 406         if (separate_dirty_bits(cmd)) {
 407                 dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
 408                 r = dm_bitset_empty(&cmd->dirty_info, &cmd->dirty_root);
 409                 if (r < 0)
 410                         goto bad;
 411         }
 412
 413         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 414         r = dm_bitset_empty(&cmd->discard_info, &cmd->discard_root);
 415         if (r < 0)
 416                 goto bad;
 417
 418         cmd->discard_block_size = 0;
 419         cmd->discard_nr_blocks = 0;
 420
 421         r = __write_initial_superblock(cmd);
 422         if (r)
 423                 goto bad;
 424
 425         cmd->clean_when_opened = true;
 426         return 0;
 427
 428 bad:
 429         dm_tm_destroy(cmd->tm);
 430         dm_sm_destroy(cmd->metadata_sm);
 431
 432         return r;
 433 }
 434
 435 static int __check_incompat_features(struct cache_disk_superblock *disk_super,
 436                                      struct dm_cache_metadata *cmd)
 437 {
 438         uint32_t incompat_flags, features;
 439
 440         incompat_flags = le32_to_cpu(disk_super->incompat_flags);
 441         features = incompat_flags & ~DM_CACHE_FEATURE_INCOMPAT_SUPP;
 442         if (features) {
 443                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 444                       (unsigned long)features);
 445                 return -EINVAL;
 446         }
 447
 448         /*
 449          * Check for read-only metadata to skip the following RDWR checks.
 450          */
 451         if (get_disk_ro(cmd->bdev->bd_disk))
 452                 return 0;
 453
 454         features = le32_to_cpu(disk_super->compat_ro_flags) & ~DM_CACHE_FEATURE_COMPAT_RO_SUPP;
 455         if (features) {
 456                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 457                       (unsigned long)features);
 458                 return -EINVAL;
 459         }
 460
 461         return 0;
 462 }
 463
 464 static int __open_metadata(struct dm_cache_metadata *cmd)
 465 {
 466         int r;
 467         struct dm_block *sblock;
 468         struct cache_disk_superblock *disk_super;
 469         unsigned long sb_flags;
 470
 471         r = superblock_read_lock(cmd, &sblock);
 472         if (r < 0) {
 473                 DMERR("couldn't read lock superblock");
 474                 return r;
 475         }
 476
 477         disk_super = dm_block_data(sblock);
 478
 479         /* Verify the data block size hasn't changed */
 480         if (le32_to_cpu(disk_super->data_block_size) != cmd->data_block_size) {
 481                 DMERR("changing the data block size (from %u to %llu) is not supported",
 482                       le32_to_cpu(disk_super->data_block_size),
 483                       (unsigned long long)cmd->data_block_size);
 484                 r = -EINVAL;
 485                 goto bad;
 486         }
 487
 488         r = __check_incompat_features(disk_super, cmd);
 489         if (r < 0)
 490                 goto bad;
 491
 492         r = dm_tm_open_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 493                                disk_super->metadata_space_map_root,
 494                                sizeof(disk_super->metadata_space_map_root),
 495                                &cmd->tm, &cmd->metadata_sm);
 496         if (r < 0) {
 497                 DMERR("tm_open_with_sm failed");
 498                 goto bad;
 499         }
 500
 501         __setup_mapping_info(cmd);
 502         dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
 503         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 504         sb_flags = le32_to_cpu(disk_super->flags);
 505         cmd->clean_when_opened = test_bit(CLEAN_SHUTDOWN, &sb_flags);
 506         dm_bm_unlock(sblock);
 507
 508         return 0;
 509
 510 bad:
 511         dm_bm_unlock(sblock);
 512         return r;
 513 }
 514
 515 static int __open_or_format_metadata(struct dm_cache_metadata *cmd,
 516                                      bool format_device)
 517 {
 518         int r;
 519         bool unformatted = false;
 520
 521         r = __superblock_all_zeroes(cmd->bm, &unformatted);
 522         if (r)
 523                 return r;
 524
 525         if (unformatted)
 526                 return format_device ? __format_metadata(cmd) : -EPERM;
 527
 528         return __open_metadata(cmd);
 529 }
 530
 531 static int __create_persistent_data_objects(struct dm_cache_metadata *cmd,
 532                                             bool may_format_device)
 533 {
 534         int r;
 535         cmd->bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 536                                           CACHE_MAX_CONCURRENT_LOCKS);
 537         if (IS_ERR(cmd->bm)) {
 538                 DMERR("could not create block manager");
 539                 return PTR_ERR(cmd->bm);
 540         }
 541
 542         r = __open_or_format_metadata(cmd, may_format_device);
 543         if (r)
 544                 dm_block_manager_destroy(cmd->bm);
 545
 546         return r;
 547 }
 548
 549 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd)
 550 {
 551         dm_sm_destroy(cmd->metadata_sm);
 552         dm_tm_destroy(cmd->tm);
 553         dm_block_manager_destroy(cmd->bm);
 554 }
 555
 556 typedef unsigned long (*flags_mutator)(unsigned long);
 557
 558 static void update_flags(struct cache_disk_superblock *disk_super,
 559                          flags_mutator mutator)
 560 {
 561         uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
 562         disk_super->flags = cpu_to_le32(sb_flags);
 563 }
 564
 565 static unsigned long set_clean_shutdown(unsigned long flags)
 566 {
 567         set_bit(CLEAN_SHUTDOWN, &flags);
 568         return flags;
 569 }
 570
 571 static unsigned long clear_clean_shutdown(unsigned long flags)
 572 {
 573         clear_bit(CLEAN_SHUTDOWN, &flags);
 574         return flags;
 575 }
 576
 577 static void read_superblock_fields(struct dm_cache_metadata *cmd,
 578                                    struct cache_disk_superblock *disk_super)
 579 {
 580         cmd->version = le32_to_cpu(disk_super->version);
 581         cmd->flags = le32_to_cpu(disk_super->flags);
 582         cmd->root = le64_to_cpu(disk_super->mapping_root);
 583         cmd->hint_root = le64_to_cpu(disk_super->hint_root);
 584         cmd->discard_root = le64_to_cpu(disk_super->discard_root);
 585         cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
 586         cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
 587         cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 588         cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
 589         strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
 590         cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
 591         cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
 592         cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
 593         cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
 594
 595         cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
 596         cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
 597         cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
 598         cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
 599
 600         if (separate_dirty_bits(cmd))
 601                 cmd->dirty_root = le64_to_cpu(disk_super->dirty_root);
 602
 603         cmd->changed = false;
 604 }
 605
 606 /*
 607  * The mutator updates the superblock flags.
 608  */
 609 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
 610                                      flags_mutator mutator)
 611 {
 612         int r;
 613         struct cache_disk_superblock *disk_super;
 614         struct dm_block *sblock;
 615
 616         r = superblock_lock(cmd, &sblock);
 617         if (r)
 618                 return r;
 619
 620         disk_super = dm_block_data(sblock);
 621         update_flags(disk_super, mutator);
 622         read_superblock_fields(cmd, disk_super);
 623         dm_bm_unlock(sblock);
 624
 625         return dm_bm_flush(cmd->bm);
 626 }
 627
 628 static int __begin_transaction(struct dm_cache_metadata *cmd)
 629 {
 630         int r;
 631         struct cache_disk_superblock *disk_super;
 632         struct dm_block *sblock;
 633
 634         /*
 635          * We re-read the superblock every time.  Shouldn't need to do this
 636          * really.
 637          */
 638         r = superblock_read_lock(cmd, &sblock);
 639         if (r)
 640                 return r;
 641
 642         disk_super = dm_block_data(sblock);
 643         read_superblock_fields(cmd, disk_super);
 644         dm_bm_unlock(sblock);
 645
 646         return 0;
 647 }
 648
 649 static int __commit_transaction(struct dm_cache_metadata *cmd,
 650                                 flags_mutator mutator)
 651 {
 652         int r;
 653         struct cache_disk_superblock *disk_super;
 654         struct dm_block *sblock;
 655
 656         /*
 657          * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
 658          */
 659         BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
 660
 661         if (separate_dirty_bits(cmd)) {
 662                 r = dm_bitset_flush(&cmd->dirty_info, cmd->dirty_root,
 663                                     &cmd->dirty_root);
 664                 if (r)
 665                         return r;
 666         }
 667
 668         r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
 669                             &cmd->discard_root);
 670         if (r)
 671                 return r;
 672
 673         r = dm_tm_pre_commit(cmd->tm);
 674         if (r < 0)
 675                 return r;
 676
 677         r = __save_sm_root(cmd);
 678         if (r)
 679                 return r;
 680
 681         r = superblock_lock(cmd, &sblock);
 682         if (r)
 683                 return r;
 684
 685         disk_super = dm_block_data(sblock);
 686
 687         disk_super->flags = cpu_to_le32(cmd->flags);
 688         if (mutator)
 689                 update_flags(disk_super, mutator);
 690
 691         disk_super->mapping_root = cpu_to_le64(cmd->root);
 692         if (separate_dirty_bits(cmd))
 693                 disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
 694         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 695         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 696         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 697         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 698         disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
 699         strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
 700         disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
 701         disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
 702         disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
 703         disk_super->policy_hint_size = cpu_to_le32(cmd->policy_hint_size);
 704
 705         disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
 706         disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
 707         disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
 708         disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
 709         __copy_sm_root(cmd, disk_super);
 710
 711         return dm_tm_commit(cmd->tm, sblock);
 712 }
 713
 714 /*----------------------------------------------------------------*/
 715
 716 /*
 717  * The mappings are held in a dm-array that has 64-bit values stored in
 718  * little-endian format.  The index is the cblock, the high 48bits of the
 719  * value are the oblock and the low 16 bit the flags.
 720  */
 721 #define FLAGS_MASK ((1 << 16) - 1)
 722
 723 static __le64 pack_value(dm_oblock_t block, unsigned flags)
 724 {
 725         uint64_t value = from_oblock(block);
 726         value <<= 16;
 727         value = value | (flags & FLAGS_MASK);
 728         return cpu_to_le64(value);
 729 }
 730
 731 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
 732 {
 733         uint64_t value = le64_to_cpu(value_le);
 734         uint64_t b = value >> 16;
 735         *block = to_oblock(b);
 736         *flags = value & FLAGS_MASK;
 737 }
 738
 739 /*----------------------------------------------------------------*/
 740
 741 static struct dm_cache_metadata *metadata_open(struct block_device *bdev,
 742                                                sector_t data_block_size,
 743                                                bool may_format_device,
 744                                                size_t policy_hint_size,
 745                                                unsigned metadata_version)
 746 {
 747         int r;
 748         struct dm_cache_metadata *cmd;
 749
 750         cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
 751         if (!cmd) {
 752                 DMERR("could not allocate metadata struct");
 753                 return ERR_PTR(-ENOMEM);
 754         }
 755
 756         cmd->version = metadata_version;
 757         atomic_set(&cmd->ref_count, 1);
 758         init_rwsem(&cmd->root_lock);
 759         cmd->bdev = bdev;
 760         cmd->data_block_size = data_block_size;
 761         cmd->cache_blocks = 0;
 762         cmd->policy_hint_size = policy_hint_size;
 763         cmd->changed = true;
 764         cmd->fail_io = false;
 765
 766         r = __create_persistent_data_objects(cmd, may_format_device);
 767         if (r) {
 768                 kfree(cmd);
 769                 return ERR_PTR(r);
 770         }
 771
 772         r = __begin_transaction_flags(cmd, clear_clean_shutdown);
 773         if (r < 0) {
 774                 dm_cache_metadata_close(cmd);
 775                 return ERR_PTR(r);
 776         }
 777
 778         return cmd;
 779 }
 780
 781 /*
 782  * We keep a little list of ref counted metadata objects to prevent two
 783  * different target instances creating separate bufio instances.  This is
 784  * an issue if a table is reloaded before the suspend.
 785  */
 786 static DEFINE_MUTEX(table_lock);
 787 static LIST_HEAD(table);
 788
 789 static struct dm_cache_metadata *lookup(struct block_device *bdev)
 790 {
 791         struct dm_cache_metadata *cmd;
 792
 793         list_for_each_entry(cmd, &table, list)
 794                 if (cmd->bdev == bdev) {
 795                         atomic_inc(&cmd->ref_count);
 796                         return cmd;
 797                 }
 798
 799         return NULL;
 800 }
 801
 802 static struct dm_cache_metadata *lookup_or_open(struct block_device *bdev,
 803                                                 sector_t data_block_size,
 804                                                 bool may_format_device,
 805                                                 size_t policy_hint_size,
 806                                                 unsigned metadata_version)
 807 {
 808         struct dm_cache_metadata *cmd, *cmd2;
 809
 810         mutex_lock(&table_lock);
 811         cmd = lookup(bdev);
 812         mutex_unlock(&table_lock);
 813
 814         if (cmd)
 815                 return cmd;
 816
 817         cmd = metadata_open(bdev, data_block_size, may_format_device,
 818                             policy_hint_size, metadata_version);
 819         if (!IS_ERR(cmd)) {
 820                 mutex_lock(&table_lock);
 821                 cmd2 = lookup(bdev);
 822                 if (cmd2) {
 823                         mutex_unlock(&table_lock);
 824                         __destroy_persistent_data_objects(cmd);
 825                         kfree(cmd);
 826                         return cmd2;
 827                 }
 828                 list_add(&cmd->list, &table);
 829                 mutex_unlock(&table_lock);
 830         }
 831
 832         return cmd;
 833 }
 834
 835 static bool same_params(struct dm_cache_metadata *cmd, sector_t data_block_size)
 836 {
 837         if (cmd->data_block_size != data_block_size) {
 838                 DMERR("data_block_size (%llu) different from that in metadata (%llu)",
 839                       (unsigned long long) data_block_size,
 840                       (unsigned long long) cmd->data_block_size);
 841                 return false;
 842         }
 843
 844         return true;
 845 }
 846
 847 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
 848                                                  sector_t data_block_size,
 849                                                  bool may_format_device,
 850                                                  size_t policy_hint_size,
 851                                                  unsigned metadata_version)
 852 {
 853         struct dm_cache_metadata *cmd = lookup_or_open(bdev, data_block_size, may_format_device,
 854                                                        policy_hint_size, metadata_version);
 855
 856         if (!IS_ERR(cmd) && !same_params(cmd, data_block_size)) {
 857                 dm_cache_metadata_close(cmd);
 858                 return ERR_PTR(-EINVAL);
 859         }
 860
 861         return cmd;
 862 }
 863
 864 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
 865 {
 866         if (atomic_dec_and_test(&cmd->ref_count)) {
 867                 mutex_lock(&table_lock);
 868                 list_del(&cmd->list);
 869                 mutex_unlock(&table_lock);
 870
 871                 if (!cmd->fail_io)
 872                         __destroy_persistent_data_objects(cmd);
 873                 kfree(cmd);
 874         }
 875 }
 876
 877 /*
 878  * Checks that the given cache block is either unmapped or clean.
 879  */
 880 static int block_clean_combined_dirty(struct dm_cache_metadata *cmd, dm_cblock_t b,
 881                                       bool *result)
 882 {
 883         int r;
 884         __le64 value;
 885         dm_oblock_t ob;
 886         unsigned flags;
 887
 888         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(b), &value);
 889         if (r)
 890                 return r;
 891
 892         unpack_value(value, &ob, &flags);
 893         *result = !((flags & M_VALID) && (flags & M_DIRTY));
 894
 895         return 0;
 896 }
 897
 898 static int blocks_are_clean_combined_dirty(struct dm_cache_metadata *cmd,
 899                                            dm_cblock_t begin, dm_cblock_t end,
 900                                            bool *result)
 901 {
 902         int r;
 903         *result = true;
 904
 905         while (begin != end) {
 906                 r = block_clean_combined_dirty(cmd, begin, result);
 907                 if (r) {
 908                         DMERR("block_clean_combined_dirty failed");
 909                         return r;
 910                 }
 911
 912                 if (!*result) {
 913                         DMERR("cache block %llu is dirty",
 914                               (unsigned long long) from_cblock(begin));
 915                         return 0;
 916                 }
 917
 918                 begin = to_cblock(from_cblock(begin) + 1);
 919         }
 920
 921         return 0;
 922 }
 923
 924 static int blocks_are_clean_separate_dirty(struct dm_cache_metadata *cmd,
 925                                            dm_cblock_t begin, dm_cblock_t end,
 926                                            bool *result)
 927 {
 928         int r;
 929         bool dirty_flag;
 930         *result = true;
 931
 932         if (from_cblock(cmd->cache_blocks) == 0)
 933                 /* Nothing to do */
 934                 return 0;
 935
 936         r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
 937                                    from_cblock(cmd->cache_blocks), &cmd->dirty_cursor);
 938         if (r) {
 939                 DMERR("%s: dm_bitset_cursor_begin for dirty failed", __func__);
 940                 return r;
 941         }
 942
 943         r = dm_bitset_cursor_skip(&cmd->dirty_cursor, from_cblock(begin));
 944         if (r) {
 945                 DMERR("%s: dm_bitset_cursor_skip for dirty failed", __func__);
 946                 dm_bitset_cursor_end(&cmd->dirty_cursor);
 947                 return r;
 948         }
 949
 950         while (begin != end) {
 951                 /*
 952                  * We assume that unmapped blocks have their dirty bit
 953                  * cleared.
 954                  */
 955                 dirty_flag = dm_bitset_cursor_get_value(&cmd->dirty_cursor);
 956                 if (dirty_flag) {
 957                         DMERR("%s: cache block %llu is dirty", __func__,
 958                               (unsigned long long) from_cblock(begin));
 959                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 960                         *result = false;
 961                         return 0;
 962                 }
 963
 964                 begin = to_cblock(from_cblock(begin) + 1);
 965                 if (begin == end)
 966                         break;
 967
 968                 r = dm_bitset_cursor_next(&cmd->dirty_cursor);
 969                 if (r) {
 970                         DMERR("%s: dm_bitset_cursor_next for dirty failed", __func__);
 971                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 972                         return r;
 973                 }
 974         }
 975
 976         dm_bitset_cursor_end(&cmd->dirty_cursor);
 977
 978         return 0;
 979 }
 980
 981 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
 982                                         dm_cblock_t begin, dm_cblock_t end,
 983                                         bool *result)
 984 {
 985         if (separate_dirty_bits(cmd))
 986                 return blocks_are_clean_separate_dirty(cmd, begin, end, result);
 987         else
 988                 return blocks_are_clean_combined_dirty(cmd, begin, end, result);
 989 }
 990
 991 static bool cmd_write_lock(struct dm_cache_metadata *cmd)
 992 {
 993         down_write(&cmd->root_lock);
 994         if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
 995                 up_write(&cmd->root_lock);
 996                 return false;
 997         }
 998         return true;
 999 }
1000
1001 #define WRITE_LOCK(cmd)                         \
1002         do {                                    \
1003                 if (!cmd_write_lock((cmd)))     \
1004                         return -EINVAL;         \
1005         } while(0)
1006
1007 #define WRITE_LOCK_VOID(cmd)                    \
1008         do {                                    \
1009                 if (!cmd_write_lock((cmd)))     \
1010                         return;                 \
1011         } while(0)
1012
1013 #define WRITE_UNLOCK(cmd) \
1014         up_write(&(cmd)->root_lock)
1015
1016 static bool cmd_read_lock(struct dm_cache_metadata *cmd)
1017 {
1018         down_read(&cmd->root_lock);
1019         if (cmd->fail_io) {
1020                 up_read(&cmd->root_lock);
1021                 return false;
1022         }
1023         return true;
1024 }
1025
1026 #define READ_LOCK(cmd)                          \
1027         do {                                    \
1028                 if (!cmd_read_lock((cmd)))      \
1029                         return -EINVAL;         \
1030         } while(0)
1031
1032 #define READ_LOCK_VOID(cmd)                     \
1033         do {                                    \
1034                 if (!cmd_read_lock((cmd)))      \
1035                         return;                 \
1036         } while(0)
1037
1038 #define READ_UNLOCK(cmd) \
1039         up_read(&(cmd)->root_lock)
1040
1041 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
1042 {
1043         int r;
1044         bool clean;
1045         __le64 null_mapping = pack_value(0, 0);
1046
1047         WRITE_LOCK(cmd);
1048         __dm_bless_for_disk(&null_mapping);
1049
1050         if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
1051                 r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
1052                 if (r) {
1053                         __dm_unbless_for_disk(&null_mapping);
1054                         goto out;
1055                 }
1056
1057                 if (!clean) {
1058                         DMERR("unable to shrink cache due to dirty blocks");
1059                         r = -EINVAL;
1060                         __dm_unbless_for_disk(&null_mapping);
1061                         goto out;
1062                 }
1063         }
1064
1065         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
1066                             from_cblock(new_cache_size),
1067                             &null_mapping, &cmd->root);
1068         if (r)
1069                 goto out;
1070
1071         if (separate_dirty_bits(cmd)) {
1072                 r = dm_bitset_resize(&cmd->dirty_info, cmd->dirty_root,
1073                                      from_cblock(cmd->cache_blocks), from_cblock(new_cache_size),
1074                                      false, &cmd->dirty_root);
1075                 if (r)
1076                         goto out;
1077         }
1078
1079         cmd->cache_blocks = new_cache_size;
1080         cmd->changed = true;
1081
1082 out:
1083         WRITE_UNLOCK(cmd);
1084
1085         return r;
1086 }
1087
1088 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
1089                                    sector_t discard_block_size,
1090                                    dm_dblock_t new_nr_entries)
1091 {
1092         int r;
1093
1094         WRITE_LOCK(cmd);
1095         r = dm_bitset_resize(&cmd->discard_info,
1096                              cmd->discard_root,
1097                              from_dblock(cmd->discard_nr_blocks),
1098                              from_dblock(new_nr_entries),
1099                              false, &cmd->discard_root);
1100         if (!r) {
1101                 cmd->discard_block_size = discard_block_size;
1102                 cmd->discard_nr_blocks = new_nr_entries;
1103         }
1104
1105         cmd->changed = true;
1106         WRITE_UNLOCK(cmd);
1107
1108         return r;
1109 }
1110
1111 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1112 {
1113         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
1114                                  from_dblock(b), &cmd->discard_root);
1115 }
1116
1117 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1118 {
1119         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
1120                                    from_dblock(b), &cmd->discard_root);
1121 }
1122
1123 static int __discard(struct dm_cache_metadata *cmd,
1124                      dm_dblock_t dblock, bool discard)
1125 {
1126         int r;
1127
1128         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
1129         if (r)
1130                 return r;
1131
1132         cmd->changed = true;
1133         return 0;
1134 }
1135
1136 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
1137                          dm_dblock_t dblock, bool discard)
1138 {
1139         int r;
1140
1141         WRITE_LOCK(cmd);
1142         r = __discard(cmd, dblock, discard);
1143         WRITE_UNLOCK(cmd);
1144
1145         return r;
1146 }
1147
1148 static int __load_discards(struct dm_cache_metadata *cmd,
1149                            load_discard_fn fn, void *context)
1150 {
1151         int r = 0;
1152         uint32_t b;
1153         struct dm_bitset_cursor c;
1154
1155         if (from_dblock(cmd->discard_nr_blocks) == 0)
1156                 /* nothing to do */
1157                 return 0;
1158
1159         if (cmd->clean_when_opened) {
1160                 r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root, &cmd->discard_root);
1161                 if (r)
1162                         return r;
1163
1164                 r = dm_bitset_cursor_begin(&cmd->discard_info, cmd->discard_root,
1165                                            from_dblock(cmd->discard_nr_blocks), &c);
1166                 if (r)
1167                         return r;
1168
1169                 for (b = 0; ; b++) {
1170                         r = fn(context, cmd->discard_block_size, to_dblock(b),
1171                                dm_bitset_cursor_get_value(&c));
1172                         if (r)
1173                                 break;
1174
1175                         if (b >= (from_dblock(cmd->discard_nr_blocks) - 1))
1176                                 break;
1177
1178                         r = dm_bitset_cursor_next(&c);
1179                         if (r)
1180                                 break;
1181                 }
1182
1183                 dm_bitset_cursor_end(&c);
1184
1185         } else {
1186                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1187                         r = fn(context, cmd->discard_block_size, to_dblock(b), false);
1188                         if (r)
1189                                 return r;
1190                 }
1191         }
1192
1193         return r;
1194 }
1195
1196 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
1197                            load_discard_fn fn, void *context)
1198 {
1199         int r;
1200
1201         READ_LOCK(cmd);
1202         r = __load_discards(cmd, fn, context);
1203         READ_UNLOCK(cmd);
1204
1205         return r;
1206 }
1207
1208 int dm_cache_size(struct dm_cache_metadata *cmd, dm_cblock_t *result)
1209 {
1210         READ_LOCK(cmd);
1211         *result = cmd->cache_blocks;
1212         READ_UNLOCK(cmd);
1213
1214         return 0;
1215 }
1216
1217 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1218 {
1219         int r;
1220         __le64 value = pack_value(0, 0);
1221
1222         __dm_bless_for_disk(&value);
1223         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1224                                &value, &cmd->root);
1225         if (r)
1226                 return r;
1227
1228         cmd->changed = true;
1229         return 0;
1230 }
1231
1232 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1233 {
1234         int r;
1235
1236         WRITE_LOCK(cmd);
1237         r = __remove(cmd, cblock);
1238         WRITE_UNLOCK(cmd);
1239
1240         return r;
1241 }
1242
1243 static int __insert(struct dm_cache_metadata *cmd,
1244                     dm_cblock_t cblock, dm_oblock_t oblock)
1245 {
1246         int r;
1247         __le64 value = pack_value(oblock, M_VALID);
1248         __dm_bless_for_disk(&value);
1249
1250         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1251                                &value, &cmd->root);
1252         if (r)
1253                 return r;
1254
1255         cmd->changed = true;
1256         return 0;
1257 }
1258
1259 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
1260                             dm_cblock_t cblock, dm_oblock_t oblock)
1261 {
1262         int r;
1263
1264         WRITE_LOCK(cmd);
1265         r = __insert(cmd, cblock, oblock);
1266         WRITE_UNLOCK(cmd);
1267
1268         return r;
1269 }
1270
1271 struct thunk {
1272         load_mapping_fn fn;
1273         void *context;
1274
1275         struct dm_cache_metadata *cmd;
1276         bool respect_dirty_flags;
1277         bool hints_valid;
1278 };
1279
1280 static bool policy_unchanged(struct dm_cache_metadata *cmd,
1281                              struct dm_cache_policy *policy)
1282 {
1283         const char *policy_name = dm_cache_policy_get_name(policy);
1284         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1285         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
1286
1287         /*
1288          * Ensure policy names match.
1289          */
1290         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
1291                 return false;
1292
1293         /*
1294          * Ensure policy major versions match.
1295          */
1296         if (cmd->policy_version[0] != policy_version[0])
1297                 return false;
1298
1299         /*
1300          * Ensure policy hint sizes match.
1301          */
1302         if (cmd->policy_hint_size != policy_hint_size)
1303                 return false;
1304
1305         return true;
1306 }
1307
1308 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
1309 {
1310         return cmd->hint_root && cmd->policy_hint_size;
1311 }
1312
1313 static bool hints_array_available(struct dm_cache_metadata *cmd,
1314                                   struct dm_cache_policy *policy)
1315 {
1316         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
1317                 hints_array_initialized(cmd);
1318 }
1319
1320 static int __load_mapping_v1(struct dm_cache_metadata *cmd,
1321                              uint64_t cb, bool hints_valid,
1322                              struct dm_array_cursor *mapping_cursor,
1323                              struct dm_array_cursor *hint_cursor,
1324                              load_mapping_fn fn, void *context)
1325 {
1326         int r = 0;
1327
1328         __le64 mapping;
1329         __le32 hint = 0;
1330
1331         __le64 *mapping_value_le;
1332         __le32 *hint_value_le;
1333
1334         dm_oblock_t oblock;
1335         unsigned flags;
1336         bool dirty = true;
1337
1338         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1339         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1340         unpack_value(mapping, &oblock, &flags);
1341
1342         if (flags & M_VALID) {
1343                 if (hints_valid) {
1344                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1345                         memcpy(&hint, hint_value_le, sizeof(hint));
1346                 }
1347                 if (cmd->clean_when_opened)
1348                         dirty = flags & M_DIRTY;
1349
1350                 r = fn(context, oblock, to_cblock(cb), dirty,
1351                        le32_to_cpu(hint), hints_valid);
1352                 if (r) {
1353                         DMERR("policy couldn't load cache block %llu",
1354                               (unsigned long long) from_cblock(to_cblock(cb)));
1355                 }
1356         }
1357
1358         return r;
1359 }
1360
1361 static int __load_mapping_v2(struct dm_cache_metadata *cmd,
1362                              uint64_t cb, bool hints_valid,
1363                              struct dm_array_cursor *mapping_cursor,
1364                              struct dm_array_cursor *hint_cursor,
1365                              struct dm_bitset_cursor *dirty_cursor,
1366                              load_mapping_fn fn, void *context)
1367 {
1368         int r = 0;
1369
1370         __le64 mapping;
1371         __le32 hint = 0;
1372
1373         __le64 *mapping_value_le;
1374         __le32 *hint_value_le;
1375
1376         dm_oblock_t oblock;
1377         unsigned flags;
1378         bool dirty = true;
1379
1380         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1381         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1382         unpack_value(mapping, &oblock, &flags);
1383
1384         if (flags & M_VALID) {
1385                 if (hints_valid) {
1386                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1387                         memcpy(&hint, hint_value_le, sizeof(hint));
1388                 }
1389                 if (cmd->clean_when_opened)
1390                         dirty = dm_bitset_cursor_get_value(dirty_cursor);
1391
1392                 r = fn(context, oblock, to_cblock(cb), dirty,
1393                        le32_to_cpu(hint), hints_valid);
1394                 if (r) {
1395                         DMERR("policy couldn't load cache block %llu",
1396                               (unsigned long long) from_cblock(to_cblock(cb)));
1397                 }
1398         }
1399
1400         return r;
1401 }
1402
1403 static int __load_mappings(struct dm_cache_metadata *cmd,
1404                            struct dm_cache_policy *policy,
1405                            load_mapping_fn fn, void *context)
1406 {
1407         int r;
1408         uint64_t cb;
1409
1410         bool hints_valid = hints_array_available(cmd, policy);
1411
1412         if (from_cblock(cmd->cache_blocks) == 0)
1413                 /* Nothing to do */
1414                 return 0;
1415
1416         r = dm_array_cursor_begin(&cmd->info, cmd->root, &cmd->mapping_cursor);
1417         if (r)
1418                 return r;
1419
1420         if (hints_valid) {
1421                 r = dm_array_cursor_begin(&cmd->hint_info, cmd->hint_root, &cmd->hint_cursor);
1422                 if (r) {
1423                         dm_array_cursor_end(&cmd->mapping_cursor);
1424                         return r;
1425                 }
1426         }
1427
1428         if (separate_dirty_bits(cmd)) {
1429                 r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
1430                                            from_cblock(cmd->cache_blocks),
1431                                            &cmd->dirty_cursor);
1432                 if (r) {
1433                         dm_array_cursor_end(&cmd->hint_cursor);
1434                         dm_array_cursor_end(&cmd->mapping_cursor);
1435                         return r;
1436                 }
1437         }
1438
1439         for (cb = 0; ; cb++) {
1440                 if (separate_dirty_bits(cmd))
1441                         r = __load_mapping_v2(cmd, cb, hints_valid,
1442                                               &cmd->mapping_cursor,
1443                                               &cmd->hint_cursor,
1444                                               &cmd->dirty_cursor,
1445                                               fn, context);
1446                 else
1447                         r = __load_mapping_v1(cmd, cb, hints_valid,
1448                                               &cmd->mapping_cursor, &cmd->hint_cursor,
1449                                               fn, context);
1450                 if (r)
1451                         goto out;
1452
1453                 /*
1454                  * We need to break out before we move the cursors.
1455                  */
1456                 if (cb >= (from_cblock(cmd->cache_blocks) - 1))
1457                         break;
1458
1459                 r = dm_array_cursor_next(&cmd->mapping_cursor);
1460                 if (r) {
1461                         DMERR("dm_array_cursor_next for mapping failed");
1462                         goto out;
1463                 }
1464
1465                 if (hints_valid) {
1466                         r = dm_array_cursor_next(&cmd->hint_cursor);
1467                         if (r) {
1468                                 dm_array_cursor_end(&cmd->hint_cursor);
1469                                 hints_valid = false;
1470                         }
1471                 }
1472
1473                 if (separate_dirty_bits(cmd)) {
1474                         r = dm_bitset_cursor_next(&cmd->dirty_cursor);
1475                         if (r) {
1476                                 DMERR("dm_bitset_cursor_next for dirty failed");
1477                                 goto out;
1478                         }
1479                 }
1480         }
1481 out:
1482         dm_array_cursor_end(&cmd->mapping_cursor);
1483         if (hints_valid)
1484                 dm_array_cursor_end(&cmd->hint_cursor);
1485
1486         if (separate_dirty_bits(cmd))
1487                 dm_bitset_cursor_end(&cmd->dirty_cursor);
1488
1489         return r;
1490 }
1491
1492 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1493                            struct dm_cache_policy *policy,
1494                            load_mapping_fn fn, void *context)
1495 {
1496         int r;
1497
1498         READ_LOCK(cmd);
1499         r = __load_mappings(cmd, policy, fn, context);
1500         READ_UNLOCK(cmd);
1501
1502         return r;
1503 }
1504
1505 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1506 {
1507         int r = 0;
1508         __le64 value;
1509         dm_oblock_t oblock;
1510         unsigned flags;
1511
1512         memcpy(&value, leaf, sizeof(value));
1513         unpack_value(value, &oblock, &flags);
1514
1515         return r;
1516 }
1517
1518 static int __dump_mappings(struct dm_cache_metadata *cmd)
1519 {
1520         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1521 }
1522
1523 void dm_cache_dump(struct dm_cache_metadata *cmd)
1524 {
1525         READ_LOCK_VOID(cmd);
1526         __dump_mappings(cmd);
1527         READ_UNLOCK(cmd);
1528 }
1529
1530 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1531 {
1532         int r;
1533
1534         READ_LOCK(cmd);
1535         r = cmd->changed;
1536         READ_UNLOCK(cmd);
1537
1538         return r;
1539 }
1540
1541 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1542 {
1543         int r;
1544         unsigned flags;
1545         dm_oblock_t oblock;
1546         __le64 value;
1547
1548         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1549         if (r)
1550                 return r;
1551
1552         unpack_value(value, &oblock, &flags);
1553
1554         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1555                 /* nothing to be done */
1556                 return 0;
1557
1558         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1559         __dm_bless_for_disk(&value);
1560
1561         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1562                                &value, &cmd->root);
1563         if (r)
1564                 return r;
1565
1566         cmd->changed = true;
1567         return 0;
1568
1569 }
1570
1571 static int __set_dirty_bits_v1(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1572 {
1573         int r;
1574         unsigned i;
1575         for (i = 0; i < nr_bits; i++) {
1576                 r = __dirty(cmd, to_cblock(i), test_bit(i, bits));
1577                 if (r)
1578                         return r;
1579         }
1580
1581         return 0;
1582 }
1583
1584 static int is_dirty_callback(uint32_t index, bool *value, void *context)
1585 {
1586         unsigned long *bits = context;
1587         *value = test_bit(index, bits);
1588         return 0;
1589 }
1590
1591 static int __set_dirty_bits_v2(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1592 {
1593         int r = 0;
1594
1595         /* nr_bits is really just a sanity check */
1596         if (nr_bits != from_cblock(cmd->cache_blocks)) {
1597                 DMERR("dirty bitset is wrong size");
1598                 return -EINVAL;
1599         }
1600
1601         r = dm_bitset_del(&cmd->dirty_info, cmd->dirty_root);
1602         if (r)
1603                 return r;
1604
1605         cmd->changed = true;
1606         return dm_bitset_new(&cmd->dirty_info, &cmd->dirty_root, nr_bits, is_dirty_callback, bits);
1607 }
1608
1609 int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
1610                             unsigned nr_bits,
1611                             unsigned long *bits)
1612 {
1613         int r;
1614
1615         WRITE_LOCK(cmd);
1616         if (separate_dirty_bits(cmd))
1617                 r = __set_dirty_bits_v2(cmd, nr_bits, bits);
1618         else
1619                 r = __set_dirty_bits_v1(cmd, nr_bits, bits);
1620         WRITE_UNLOCK(cmd);
1621
1622         return r;
1623 }
1624
1625 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1626                                  struct dm_cache_statistics *stats)
1627 {
1628         READ_LOCK_VOID(cmd);
1629         *stats = cmd->stats;
1630         READ_UNLOCK(cmd);
1631 }
1632
1633 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1634                                  struct dm_cache_statistics *stats)
1635 {
1636         WRITE_LOCK_VOID(cmd);
1637         cmd->stats = *stats;
1638         WRITE_UNLOCK(cmd);
1639 }
1640
1641 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1642 {
1643         int r = -EINVAL;
1644         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1645                                  clear_clean_shutdown);
1646
1647         WRITE_LOCK(cmd);
1648         if (cmd->fail_io)
1649                 goto out;
1650
1651         r = __commit_transaction(cmd, mutator);
1652         if (r)
1653                 goto out;
1654
1655         r = __begin_transaction(cmd);
1656 out:
1657         WRITE_UNLOCK(cmd);
1658         return r;
1659 }
1660
1661 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1662                                            dm_block_t *result)
1663 {
1664         int r = -EINVAL;
1665
1666         READ_LOCK(cmd);
1667         if (!cmd->fail_io)
1668                 r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1669         READ_UNLOCK(cmd);
1670
1671         return r;
1672 }
1673
1674 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1675                                    dm_block_t *result)
1676 {
1677         int r = -EINVAL;
1678
1679         READ_LOCK(cmd);
1680         if (!cmd->fail_io)
1681                 r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1682         READ_UNLOCK(cmd);
1683
1684         return r;
1685 }
1686
1687 /*----------------------------------------------------------------*/
1688
1689 static int get_hint(uint32_t index, void *value_le, void *context)
1690 {
1691         uint32_t value;
1692         struct dm_cache_policy *policy = context;
1693
1694         value = policy_get_hint(policy, to_cblock(index));
1695         *((__le32 *) value_le) = cpu_to_le32(value);
1696
1697         return 0;
1698 }
1699
1700 /*
1701  * It's quicker to always delete the hint array, and recreate with
1702  * dm_array_new().
1703  */
1704 static int write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1705 {
1706         int r;
1707         size_t hint_size;
1708         const char *policy_name = dm_cache_policy_get_name(policy);
1709         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1710
1711         if (!policy_name[0] ||
1712             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1713                 return -EINVAL;
1714
1715         strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1716         memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1717
1718         hint_size = dm_cache_policy_get_hint_size(policy);
1719         if (!hint_size)
1720                 return 0; /* short-circuit hints initialization */
1721         cmd->policy_hint_size = hint_size;
1722
1723         if (cmd->hint_root) {
1724                 r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1725                 if (r)
1726                         return r;
1727         }
1728
1729         return dm_array_new(&cmd->hint_info, &cmd->hint_root,
1730                             from_cblock(cmd->cache_blocks),
1731                             get_hint, policy);
1732 }
1733
1734 int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1735 {
1736         int r;
1737
1738         WRITE_LOCK(cmd);
1739         r = write_hints(cmd, policy);
1740         WRITE_UNLOCK(cmd);
1741
1742         return r;
1743 }
1744
1745 int dm_cache_metadata_all_clean(struct dm_cache_metadata *cmd, bool *result)
1746 {
1747         int r;
1748
1749         READ_LOCK(cmd);
1750         r = blocks_are_unmapped_or_clean(cmd, 0, cmd->cache_blocks, result);
1751         READ_UNLOCK(cmd);
1752
1753         return r;
1754 }
1755
1756 void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd)
1757 {
1758         WRITE_LOCK_VOID(cmd);
1759         dm_bm_set_read_only(cmd->bm);
1760         WRITE_UNLOCK(cmd);
1761 }
1762
1763 void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd)
1764 {
1765         WRITE_LOCK_VOID(cmd);
1766         dm_bm_set_read_write(cmd->bm);
1767         WRITE_UNLOCK(cmd);
1768 }
1769
1770 int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd)
1771 {
1772         int r;
1773         struct dm_block *sblock;
1774         struct cache_disk_superblock *disk_super;
1775
1776         WRITE_LOCK(cmd);
1777         set_bit(NEEDS_CHECK, &cmd->flags);
1778
1779         r = superblock_lock(cmd, &sblock);
1780         if (r) {
1781                 DMERR("couldn't read superblock");
1782                 goto out;
1783         }
1784
1785         disk_super = dm_block_data(sblock);
1786         disk_super->flags = cpu_to_le32(cmd->flags);
1787
1788         dm_bm_unlock(sblock);
1789
1790 out:
1791         WRITE_UNLOCK(cmd);
1792         return r;
1793 }
1794
1795 int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result)
1796 {
1797         READ_LOCK(cmd);
1798         *result = !!test_bit(NEEDS_CHECK, &cmd->flags);
1799         READ_UNLOCK(cmd);
1800
1801         return 0;
1802 }
1803
1804 int dm_cache_metadata_abort(struct dm_cache_metadata *cmd)
1805 {
1806         int r;
1807
1808         WRITE_LOCK(cmd);
1809         __destroy_persistent_data_objects(cmd);
1810         r = __create_persistent_data_objects(cmd, false);
1811         if (r)
1812                 cmd->fail_io = true;
1813         WRITE_UNLOCK(cmd);
1814
1815         return r;
1816 }