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                 r = PTR_ERR(cmd->bm);
 540                 cmd->bm = NULL;
 541                 return r;
 542         }
 543
 544         r = __open_or_format_metadata(cmd, may_format_device);
 545         if (r) {
 546                 dm_block_manager_destroy(cmd->bm);
 547                 cmd->bm = NULL;
 548         }
 549
 550         return r;
 551 }
 552
 553 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd,
 554                                               bool destroy_bm)
 555 {
 556         dm_sm_destroy(cmd->metadata_sm);
 557         dm_tm_destroy(cmd->tm);
 558         if (destroy_bm)
 559                 dm_block_manager_destroy(cmd->bm);
 560 }
 561
 562 typedef unsigned long (*flags_mutator)(unsigned long);
 563
 564 static void update_flags(struct cache_disk_superblock *disk_super,
 565                          flags_mutator mutator)
 566 {
 567         uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
 568         disk_super->flags = cpu_to_le32(sb_flags);
 569 }
 570
 571 static unsigned long set_clean_shutdown(unsigned long flags)
 572 {
 573         set_bit(CLEAN_SHUTDOWN, &flags);
 574         return flags;
 575 }
 576
 577 static unsigned long clear_clean_shutdown(unsigned long flags)
 578 {
 579         clear_bit(CLEAN_SHUTDOWN, &flags);
 580         return flags;
 581 }
 582
 583 static void read_superblock_fields(struct dm_cache_metadata *cmd,
 584                                    struct cache_disk_superblock *disk_super)
 585 {
 586         cmd->version = le32_to_cpu(disk_super->version);
 587         cmd->flags = le32_to_cpu(disk_super->flags);
 588         cmd->root = le64_to_cpu(disk_super->mapping_root);
 589         cmd->hint_root = le64_to_cpu(disk_super->hint_root);
 590         cmd->discard_root = le64_to_cpu(disk_super->discard_root);
 591         cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
 592         cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
 593         cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 594         cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
 595         strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
 596         cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
 597         cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
 598         cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
 599         cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
 600
 601         cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
 602         cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
 603         cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
 604         cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
 605
 606         if (separate_dirty_bits(cmd))
 607                 cmd->dirty_root = le64_to_cpu(disk_super->dirty_root);
 608
 609         cmd->changed = false;
 610 }
 611
 612 /*
 613  * The mutator updates the superblock flags.
 614  */
 615 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
 616                                      flags_mutator mutator)
 617 {
 618         int r;
 619         struct cache_disk_superblock *disk_super;
 620         struct dm_block *sblock;
 621
 622         r = superblock_lock(cmd, &sblock);
 623         if (r)
 624                 return r;
 625
 626         disk_super = dm_block_data(sblock);
 627         update_flags(disk_super, mutator);
 628         read_superblock_fields(cmd, disk_super);
 629         dm_bm_unlock(sblock);
 630
 631         return dm_bm_flush(cmd->bm);
 632 }
 633
 634 static int __begin_transaction(struct dm_cache_metadata *cmd)
 635 {
 636         int r;
 637         struct cache_disk_superblock *disk_super;
 638         struct dm_block *sblock;
 639
 640         /*
 641          * We re-read the superblock every time.  Shouldn't need to do this
 642          * really.
 643          */
 644         r = superblock_read_lock(cmd, &sblock);
 645         if (r)
 646                 return r;
 647
 648         disk_super = dm_block_data(sblock);
 649         read_superblock_fields(cmd, disk_super);
 650         dm_bm_unlock(sblock);
 651
 652         return 0;
 653 }
 654
 655 static int __commit_transaction(struct dm_cache_metadata *cmd,
 656                                 flags_mutator mutator)
 657 {
 658         int r;
 659         struct cache_disk_superblock *disk_super;
 660         struct dm_block *sblock;
 661
 662         /*
 663          * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
 664          */
 665         BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
 666
 667         if (separate_dirty_bits(cmd)) {
 668                 r = dm_bitset_flush(&cmd->dirty_info, cmd->dirty_root,
 669                                     &cmd->dirty_root);
 670                 if (r)
 671                         return r;
 672         }
 673
 674         r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
 675                             &cmd->discard_root);
 676         if (r)
 677                 return r;
 678
 679         r = dm_tm_pre_commit(cmd->tm);
 680         if (r < 0)
 681                 return r;
 682
 683         r = __save_sm_root(cmd);
 684         if (r)
 685                 return r;
 686
 687         r = superblock_lock(cmd, &sblock);
 688         if (r)
 689                 return r;
 690
 691         disk_super = dm_block_data(sblock);
 692
 693         disk_super->flags = cpu_to_le32(cmd->flags);
 694         if (mutator)
 695                 update_flags(disk_super, mutator);
 696
 697         disk_super->mapping_root = cpu_to_le64(cmd->root);
 698         if (separate_dirty_bits(cmd))
 699                 disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
 700         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 701         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 702         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 703         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 704         disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
 705         strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
 706         disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
 707         disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
 708         disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
 709         disk_super->policy_hint_size = cpu_to_le32(cmd->policy_hint_size);
 710
 711         disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
 712         disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
 713         disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
 714         disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
 715         __copy_sm_root(cmd, disk_super);
 716
 717         return dm_tm_commit(cmd->tm, sblock);
 718 }
 719
 720 /*----------------------------------------------------------------*/
 721
 722 /*
 723  * The mappings are held in a dm-array that has 64-bit values stored in
 724  * little-endian format.  The index is the cblock, the high 48bits of the
 725  * value are the oblock and the low 16 bit the flags.
 726  */
 727 #define FLAGS_MASK ((1 << 16) - 1)
 728
 729 static __le64 pack_value(dm_oblock_t block, unsigned flags)
 730 {
 731         uint64_t value = from_oblock(block);
 732         value <<= 16;
 733         value = value | (flags & FLAGS_MASK);
 734         return cpu_to_le64(value);
 735 }
 736
 737 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
 738 {
 739         uint64_t value = le64_to_cpu(value_le);
 740         uint64_t b = value >> 16;
 741         *block = to_oblock(b);
 742         *flags = value & FLAGS_MASK;
 743 }
 744
 745 /*----------------------------------------------------------------*/
 746
 747 static struct dm_cache_metadata *metadata_open(struct block_device *bdev,
 748                                                sector_t data_block_size,
 749                                                bool may_format_device,
 750                                                size_t policy_hint_size,
 751                                                unsigned metadata_version)
 752 {
 753         int r;
 754         struct dm_cache_metadata *cmd;
 755
 756         cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
 757         if (!cmd) {
 758                 DMERR("could not allocate metadata struct");
 759                 return ERR_PTR(-ENOMEM);
 760         }
 761
 762         cmd->version = metadata_version;
 763         atomic_set(&cmd->ref_count, 1);
 764         init_rwsem(&cmd->root_lock);
 765         cmd->bdev = bdev;
 766         cmd->data_block_size = data_block_size;
 767         cmd->cache_blocks = 0;
 768         cmd->policy_hint_size = policy_hint_size;
 769         cmd->changed = true;
 770         cmd->fail_io = false;
 771
 772         r = __create_persistent_data_objects(cmd, may_format_device);
 773         if (r) {
 774                 kfree(cmd);
 775                 return ERR_PTR(r);
 776         }
 777
 778         r = __begin_transaction_flags(cmd, clear_clean_shutdown);
 779         if (r < 0) {
 780                 dm_cache_metadata_close(cmd);
 781                 return ERR_PTR(r);
 782         }
 783
 784         return cmd;
 785 }
 786
 787 /*
 788  * We keep a little list of ref counted metadata objects to prevent two
 789  * different target instances creating separate bufio instances.  This is
 790  * an issue if a table is reloaded before the suspend.
 791  */
 792 static DEFINE_MUTEX(table_lock);
 793 static LIST_HEAD(table);
 794
 795 static struct dm_cache_metadata *lookup(struct block_device *bdev)
 796 {
 797         struct dm_cache_metadata *cmd;
 798
 799         list_for_each_entry(cmd, &table, list)
 800                 if (cmd->bdev == bdev) {
 801                         atomic_inc(&cmd->ref_count);
 802                         return cmd;
 803                 }
 804
 805         return NULL;
 806 }
 807
 808 static struct dm_cache_metadata *lookup_or_open(struct block_device *bdev,
 809                                                 sector_t data_block_size,
 810                                                 bool may_format_device,
 811                                                 size_t policy_hint_size,
 812                                                 unsigned metadata_version)
 813 {
 814         struct dm_cache_metadata *cmd, *cmd2;
 815
 816         mutex_lock(&table_lock);
 817         cmd = lookup(bdev);
 818         mutex_unlock(&table_lock);
 819
 820         if (cmd)
 821                 return cmd;
 822
 823         cmd = metadata_open(bdev, data_block_size, may_format_device,
 824                             policy_hint_size, metadata_version);
 825         if (!IS_ERR(cmd)) {
 826                 mutex_lock(&table_lock);
 827                 cmd2 = lookup(bdev);
 828                 if (cmd2) {
 829                         mutex_unlock(&table_lock);
 830                         __destroy_persistent_data_objects(cmd, true);
 831                         kfree(cmd);
 832                         return cmd2;
 833                 }
 834                 list_add(&cmd->list, &table);
 835                 mutex_unlock(&table_lock);
 836         }
 837
 838         return cmd;
 839 }
 840
 841 static bool same_params(struct dm_cache_metadata *cmd, sector_t data_block_size)
 842 {
 843         if (cmd->data_block_size != data_block_size) {
 844                 DMERR("data_block_size (%llu) different from that in metadata (%llu)",
 845                       (unsigned long long) data_block_size,
 846                       (unsigned long long) cmd->data_block_size);
 847                 return false;
 848         }
 849
 850         return true;
 851 }
 852
 853 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
 854                                                  sector_t data_block_size,
 855                                                  bool may_format_device,
 856                                                  size_t policy_hint_size,
 857                                                  unsigned metadata_version)
 858 {
 859         struct dm_cache_metadata *cmd = lookup_or_open(bdev, data_block_size, may_format_device,
 860                                                        policy_hint_size, metadata_version);
 861
 862         if (!IS_ERR(cmd) && !same_params(cmd, data_block_size)) {
 863                 dm_cache_metadata_close(cmd);
 864                 return ERR_PTR(-EINVAL);
 865         }
 866
 867         return cmd;
 868 }
 869
 870 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
 871 {
 872         if (atomic_dec_and_test(&cmd->ref_count)) {
 873                 mutex_lock(&table_lock);
 874                 list_del(&cmd->list);
 875                 mutex_unlock(&table_lock);
 876
 877                 if (!cmd->fail_io)
 878                         __destroy_persistent_data_objects(cmd, true);
 879                 kfree(cmd);
 880         }
 881 }
 882
 883 /*
 884  * Checks that the given cache block is either unmapped or clean.
 885  */
 886 static int block_clean_combined_dirty(struct dm_cache_metadata *cmd, dm_cblock_t b,
 887                                       bool *result)
 888 {
 889         int r;
 890         __le64 value;
 891         dm_oblock_t ob;
 892         unsigned flags;
 893
 894         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(b), &value);
 895         if (r)
 896                 return r;
 897
 898         unpack_value(value, &ob, &flags);
 899         *result = !((flags & M_VALID) && (flags & M_DIRTY));
 900
 901         return 0;
 902 }
 903
 904 static int blocks_are_clean_combined_dirty(struct dm_cache_metadata *cmd,
 905                                            dm_cblock_t begin, dm_cblock_t end,
 906                                            bool *result)
 907 {
 908         int r;
 909         *result = true;
 910
 911         while (begin != end) {
 912                 r = block_clean_combined_dirty(cmd, begin, result);
 913                 if (r) {
 914                         DMERR("block_clean_combined_dirty failed");
 915                         return r;
 916                 }
 917
 918                 if (!*result) {
 919                         DMERR("cache block %llu is dirty",
 920                               (unsigned long long) from_cblock(begin));
 921                         return 0;
 922                 }
 923
 924                 begin = to_cblock(from_cblock(begin) + 1);
 925         }
 926
 927         return 0;
 928 }
 929
 930 static int blocks_are_clean_separate_dirty(struct dm_cache_metadata *cmd,
 931                                            dm_cblock_t begin, dm_cblock_t end,
 932                                            bool *result)
 933 {
 934         int r;
 935         bool dirty_flag;
 936         *result = true;
 937
 938         if (from_cblock(cmd->cache_blocks) == 0)
 939                 /* Nothing to do */
 940                 return 0;
 941
 942         r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
 943                                    from_cblock(cmd->cache_blocks), &cmd->dirty_cursor);
 944         if (r) {
 945                 DMERR("%s: dm_bitset_cursor_begin for dirty failed", __func__);
 946                 return r;
 947         }
 948
 949         r = dm_bitset_cursor_skip(&cmd->dirty_cursor, from_cblock(begin));
 950         if (r) {
 951                 DMERR("%s: dm_bitset_cursor_skip for dirty failed", __func__);
 952                 dm_bitset_cursor_end(&cmd->dirty_cursor);
 953                 return r;
 954         }
 955
 956         while (begin != end) {
 957                 /*
 958                  * We assume that unmapped blocks have their dirty bit
 959                  * cleared.
 960                  */
 961                 dirty_flag = dm_bitset_cursor_get_value(&cmd->dirty_cursor);
 962                 if (dirty_flag) {
 963                         DMERR("%s: cache block %llu is dirty", __func__,
 964                               (unsigned long long) from_cblock(begin));
 965                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 966                         *result = false;
 967                         return 0;
 968                 }
 969
 970                 begin = to_cblock(from_cblock(begin) + 1);
 971                 if (begin == end)
 972                         break;
 973
 974                 r = dm_bitset_cursor_next(&cmd->dirty_cursor);
 975                 if (r) {
 976                         DMERR("%s: dm_bitset_cursor_next for dirty failed", __func__);
 977                         dm_bitset_cursor_end(&cmd->dirty_cursor);
 978                         return r;
 979                 }
 980         }
 981
 982         dm_bitset_cursor_end(&cmd->dirty_cursor);
 983
 984         return 0;
 985 }
 986
 987 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
 988                                         dm_cblock_t begin, dm_cblock_t end,
 989                                         bool *result)
 990 {
 991         if (separate_dirty_bits(cmd))
 992                 return blocks_are_clean_separate_dirty(cmd, begin, end, result);
 993         else
 994                 return blocks_are_clean_combined_dirty(cmd, begin, end, result);
 995 }
 996
 997 static bool cmd_write_lock(struct dm_cache_metadata *cmd)
 998 {
 999         down_write(&cmd->root_lock);
1000         if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
1001                 up_write(&cmd->root_lock);
1002                 return false;
1003         }
1004         return true;
1005 }
1006
1007 #define WRITE_LOCK(cmd)                         \
1008         do {                                    \
1009                 if (!cmd_write_lock((cmd)))     \
1010                         return -EINVAL;         \
1011         } while(0)
1012
1013 #define WRITE_LOCK_VOID(cmd)                    \
1014         do {                                    \
1015                 if (!cmd_write_lock((cmd)))     \
1016                         return;                 \
1017         } while(0)
1018
1019 #define WRITE_UNLOCK(cmd) \
1020         up_write(&(cmd)->root_lock)
1021
1022 static bool cmd_read_lock(struct dm_cache_metadata *cmd)
1023 {
1024         down_read(&cmd->root_lock);
1025         if (cmd->fail_io) {
1026                 up_read(&cmd->root_lock);
1027                 return false;
1028         }
1029         return true;
1030 }
1031
1032 #define READ_LOCK(cmd)                          \
1033         do {                                    \
1034                 if (!cmd_read_lock((cmd)))      \
1035                         return -EINVAL;         \
1036         } while(0)
1037
1038 #define READ_LOCK_VOID(cmd)                     \
1039         do {                                    \
1040                 if (!cmd_read_lock((cmd)))      \
1041                         return;                 \
1042         } while(0)
1043
1044 #define READ_UNLOCK(cmd) \
1045         up_read(&(cmd)->root_lock)
1046
1047 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
1048 {
1049         int r;
1050         bool clean;
1051         __le64 null_mapping = pack_value(0, 0);
1052
1053         WRITE_LOCK(cmd);
1054         __dm_bless_for_disk(&null_mapping);
1055
1056         if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
1057                 r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
1058                 if (r) {
1059                         __dm_unbless_for_disk(&null_mapping);
1060                         goto out;
1061                 }
1062
1063                 if (!clean) {
1064                         DMERR("unable to shrink cache due to dirty blocks");
1065                         r = -EINVAL;
1066                         __dm_unbless_for_disk(&null_mapping);
1067                         goto out;
1068                 }
1069         }
1070
1071         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
1072                             from_cblock(new_cache_size),
1073                             &null_mapping, &cmd->root);
1074         if (r)
1075                 goto out;
1076
1077         if (separate_dirty_bits(cmd)) {
1078                 r = dm_bitset_resize(&cmd->dirty_info, cmd->dirty_root,
1079                                      from_cblock(cmd->cache_blocks), from_cblock(new_cache_size),
1080                                      false, &cmd->dirty_root);
1081                 if (r)
1082                         goto out;
1083         }
1084
1085         cmd->cache_blocks = new_cache_size;
1086         cmd->changed = true;
1087
1088 out:
1089         WRITE_UNLOCK(cmd);
1090
1091         return r;
1092 }
1093
1094 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
1095                                    sector_t discard_block_size,
1096                                    dm_dblock_t new_nr_entries)
1097 {
1098         int r;
1099
1100         WRITE_LOCK(cmd);
1101         r = dm_bitset_resize(&cmd->discard_info,
1102                              cmd->discard_root,
1103                              from_dblock(cmd->discard_nr_blocks),
1104                              from_dblock(new_nr_entries),
1105                              false, &cmd->discard_root);
1106         if (!r) {
1107                 cmd->discard_block_size = discard_block_size;
1108                 cmd->discard_nr_blocks = new_nr_entries;
1109         }
1110
1111         cmd->changed = true;
1112         WRITE_UNLOCK(cmd);
1113
1114         return r;
1115 }
1116
1117 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1118 {
1119         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
1120                                  from_dblock(b), &cmd->discard_root);
1121 }
1122
1123 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1124 {
1125         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
1126                                    from_dblock(b), &cmd->discard_root);
1127 }
1128
1129 static int __discard(struct dm_cache_metadata *cmd,
1130                      dm_dblock_t dblock, bool discard)
1131 {
1132         int r;
1133
1134         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
1135         if (r)
1136                 return r;
1137
1138         cmd->changed = true;
1139         return 0;
1140 }
1141
1142 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
1143                          dm_dblock_t dblock, bool discard)
1144 {
1145         int r;
1146
1147         WRITE_LOCK(cmd);
1148         r = __discard(cmd, dblock, discard);
1149         WRITE_UNLOCK(cmd);
1150
1151         return r;
1152 }
1153
1154 static int __load_discards(struct dm_cache_metadata *cmd,
1155                            load_discard_fn fn, void *context)
1156 {
1157         int r = 0;
1158         uint32_t b;
1159         struct dm_bitset_cursor c;
1160
1161         if (from_dblock(cmd->discard_nr_blocks) == 0)
1162                 /* nothing to do */
1163                 return 0;
1164
1165         if (cmd->clean_when_opened) {
1166                 r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root, &cmd->discard_root);
1167                 if (r)
1168                         return r;
1169
1170                 r = dm_bitset_cursor_begin(&cmd->discard_info, cmd->discard_root,
1171                                            from_dblock(cmd->discard_nr_blocks), &c);
1172                 if (r)
1173                         return r;
1174
1175                 for (b = 0; ; b++) {
1176                         r = fn(context, cmd->discard_block_size, to_dblock(b),
1177                                dm_bitset_cursor_get_value(&c));
1178                         if (r)
1179                                 break;
1180
1181                         if (b >= (from_dblock(cmd->discard_nr_blocks) - 1))
1182                                 break;
1183
1184                         r = dm_bitset_cursor_next(&c);
1185                         if (r)
1186                                 break;
1187                 }
1188
1189                 dm_bitset_cursor_end(&c);
1190
1191         } else {
1192                 for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1193                         r = fn(context, cmd->discard_block_size, to_dblock(b), false);
1194                         if (r)
1195                                 return r;
1196                 }
1197         }
1198
1199         return r;
1200 }
1201
1202 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
1203                            load_discard_fn fn, void *context)
1204 {
1205         int r;
1206
1207         READ_LOCK(cmd);
1208         r = __load_discards(cmd, fn, context);
1209         READ_UNLOCK(cmd);
1210
1211         return r;
1212 }
1213
1214 int dm_cache_size(struct dm_cache_metadata *cmd, dm_cblock_t *result)
1215 {
1216         READ_LOCK(cmd);
1217         *result = cmd->cache_blocks;
1218         READ_UNLOCK(cmd);
1219
1220         return 0;
1221 }
1222
1223 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1224 {
1225         int r;
1226         __le64 value = pack_value(0, 0);
1227
1228         __dm_bless_for_disk(&value);
1229         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1230                                &value, &cmd->root);
1231         if (r)
1232                 return r;
1233
1234         cmd->changed = true;
1235         return 0;
1236 }
1237
1238 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1239 {
1240         int r;
1241
1242         WRITE_LOCK(cmd);
1243         r = __remove(cmd, cblock);
1244         WRITE_UNLOCK(cmd);
1245
1246         return r;
1247 }
1248
1249 static int __insert(struct dm_cache_metadata *cmd,
1250                     dm_cblock_t cblock, dm_oblock_t oblock)
1251 {
1252         int r;
1253         __le64 value = pack_value(oblock, M_VALID);
1254         __dm_bless_for_disk(&value);
1255
1256         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1257                                &value, &cmd->root);
1258         if (r)
1259                 return r;
1260
1261         cmd->changed = true;
1262         return 0;
1263 }
1264
1265 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
1266                             dm_cblock_t cblock, dm_oblock_t oblock)
1267 {
1268         int r;
1269
1270         WRITE_LOCK(cmd);
1271         r = __insert(cmd, cblock, oblock);
1272         WRITE_UNLOCK(cmd);
1273
1274         return r;
1275 }
1276
1277 struct thunk {
1278         load_mapping_fn fn;
1279         void *context;
1280
1281         struct dm_cache_metadata *cmd;
1282         bool respect_dirty_flags;
1283         bool hints_valid;
1284 };
1285
1286 static bool policy_unchanged(struct dm_cache_metadata *cmd,
1287                              struct dm_cache_policy *policy)
1288 {
1289         const char *policy_name = dm_cache_policy_get_name(policy);
1290         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1291         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
1292
1293         /*
1294          * Ensure policy names match.
1295          */
1296         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
1297                 return false;
1298
1299         /*
1300          * Ensure policy major versions match.
1301          */
1302         if (cmd->policy_version[0] != policy_version[0])
1303                 return false;
1304
1305         /*
1306          * Ensure policy hint sizes match.
1307          */
1308         if (cmd->policy_hint_size != policy_hint_size)
1309                 return false;
1310
1311         return true;
1312 }
1313
1314 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
1315 {
1316         return cmd->hint_root && cmd->policy_hint_size;
1317 }
1318
1319 static bool hints_array_available(struct dm_cache_metadata *cmd,
1320                                   struct dm_cache_policy *policy)
1321 {
1322         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
1323                 hints_array_initialized(cmd);
1324 }
1325
1326 static int __load_mapping_v1(struct dm_cache_metadata *cmd,
1327                              uint64_t cb, bool hints_valid,
1328                              struct dm_array_cursor *mapping_cursor,
1329                              struct dm_array_cursor *hint_cursor,
1330                              load_mapping_fn fn, void *context)
1331 {
1332         int r = 0;
1333
1334         __le64 mapping;
1335         __le32 hint = 0;
1336
1337         __le64 *mapping_value_le;
1338         __le32 *hint_value_le;
1339
1340         dm_oblock_t oblock;
1341         unsigned flags;
1342         bool dirty = true;
1343
1344         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1345         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1346         unpack_value(mapping, &oblock, &flags);
1347
1348         if (flags & M_VALID) {
1349                 if (hints_valid) {
1350                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1351                         memcpy(&hint, hint_value_le, sizeof(hint));
1352                 }
1353                 if (cmd->clean_when_opened)
1354                         dirty = flags & M_DIRTY;
1355
1356                 r = fn(context, oblock, to_cblock(cb), dirty,
1357                        le32_to_cpu(hint), hints_valid);
1358                 if (r) {
1359                         DMERR("policy couldn't load cache block %llu",
1360                               (unsigned long long) from_cblock(to_cblock(cb)));
1361                 }
1362         }
1363
1364         return r;
1365 }
1366
1367 static int __load_mapping_v2(struct dm_cache_metadata *cmd,
1368                              uint64_t cb, bool hints_valid,
1369                              struct dm_array_cursor *mapping_cursor,
1370                              struct dm_array_cursor *hint_cursor,
1371                              struct dm_bitset_cursor *dirty_cursor,
1372                              load_mapping_fn fn, void *context)
1373 {
1374         int r = 0;
1375
1376         __le64 mapping;
1377         __le32 hint = 0;
1378
1379         __le64 *mapping_value_le;
1380         __le32 *hint_value_le;
1381
1382         dm_oblock_t oblock;
1383         unsigned flags;
1384         bool dirty = true;
1385
1386         dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1387         memcpy(&mapping, mapping_value_le, sizeof(mapping));
1388         unpack_value(mapping, &oblock, &flags);
1389
1390         if (flags & M_VALID) {
1391                 if (hints_valid) {
1392                         dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1393                         memcpy(&hint, hint_value_le, sizeof(hint));
1394                 }
1395                 if (cmd->clean_when_opened)
1396                         dirty = dm_bitset_cursor_get_value(dirty_cursor);
1397
1398                 r = fn(context, oblock, to_cblock(cb), dirty,
1399                        le32_to_cpu(hint), hints_valid);
1400                 if (r) {
1401                         DMERR("policy couldn't load cache block %llu",
1402                               (unsigned long long) from_cblock(to_cblock(cb)));
1403                 }
1404         }
1405
1406         return r;
1407 }
1408
1409 static int __load_mappings(struct dm_cache_metadata *cmd,
1410                            struct dm_cache_policy *policy,
1411                            load_mapping_fn fn, void *context)
1412 {
1413         int r;
1414         uint64_t cb;
1415
1416         bool hints_valid = hints_array_available(cmd, policy);
1417
1418         if (from_cblock(cmd->cache_blocks) == 0)
1419                 /* Nothing to do */
1420                 return 0;
1421
1422         r = dm_array_cursor_begin(&cmd->info, cmd->root, &cmd->mapping_cursor);
1423         if (r)
1424                 return r;
1425
1426         if (hints_valid) {
1427                 r = dm_array_cursor_begin(&cmd->hint_info, cmd->hint_root, &cmd->hint_cursor);
1428                 if (r) {
1429                         dm_array_cursor_end(&cmd->mapping_cursor);
1430                         return r;
1431                 }
1432         }
1433
1434         if (separate_dirty_bits(cmd)) {
1435                 r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
1436                                            from_cblock(cmd->cache_blocks),
1437                                            &cmd->dirty_cursor);
1438                 if (r) {
1439                         dm_array_cursor_end(&cmd->hint_cursor);
1440                         dm_array_cursor_end(&cmd->mapping_cursor);
1441                         return r;
1442                 }
1443         }
1444
1445         for (cb = 0; ; cb++) {
1446                 if (separate_dirty_bits(cmd))
1447                         r = __load_mapping_v2(cmd, cb, hints_valid,
1448                                               &cmd->mapping_cursor,
1449                                               &cmd->hint_cursor,
1450                                               &cmd->dirty_cursor,
1451                                               fn, context);
1452                 else
1453                         r = __load_mapping_v1(cmd, cb, hints_valid,
1454                                               &cmd->mapping_cursor, &cmd->hint_cursor,
1455                                               fn, context);
1456                 if (r)
1457                         goto out;
1458
1459                 /*
1460                  * We need to break out before we move the cursors.
1461                  */
1462                 if (cb >= (from_cblock(cmd->cache_blocks) - 1))
1463                         break;
1464
1465                 r = dm_array_cursor_next(&cmd->mapping_cursor);
1466                 if (r) {
1467                         DMERR("dm_array_cursor_next for mapping failed");
1468                         goto out;
1469                 }
1470
1471                 if (hints_valid) {
1472                         r = dm_array_cursor_next(&cmd->hint_cursor);
1473                         if (r) {
1474                                 dm_array_cursor_end(&cmd->hint_cursor);
1475                                 hints_valid = false;
1476                         }
1477                 }
1478
1479                 if (separate_dirty_bits(cmd)) {
1480                         r = dm_bitset_cursor_next(&cmd->dirty_cursor);
1481                         if (r) {
1482                                 DMERR("dm_bitset_cursor_next for dirty failed");
1483                                 goto out;
1484                         }
1485                 }
1486         }
1487 out:
1488         dm_array_cursor_end(&cmd->mapping_cursor);
1489         if (hints_valid)
1490                 dm_array_cursor_end(&cmd->hint_cursor);
1491
1492         if (separate_dirty_bits(cmd))
1493                 dm_bitset_cursor_end(&cmd->dirty_cursor);
1494
1495         return r;
1496 }
1497
1498 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1499                            struct dm_cache_policy *policy,
1500                            load_mapping_fn fn, void *context)
1501 {
1502         int r;
1503
1504         READ_LOCK(cmd);
1505         r = __load_mappings(cmd, policy, fn, context);
1506         READ_UNLOCK(cmd);
1507
1508         return r;
1509 }
1510
1511 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1512 {
1513         int r = 0;
1514         __le64 value;
1515         dm_oblock_t oblock;
1516         unsigned flags;
1517
1518         memcpy(&value, leaf, sizeof(value));
1519         unpack_value(value, &oblock, &flags);
1520
1521         return r;
1522 }
1523
1524 static int __dump_mappings(struct dm_cache_metadata *cmd)
1525 {
1526         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1527 }
1528
1529 void dm_cache_dump(struct dm_cache_metadata *cmd)
1530 {
1531         READ_LOCK_VOID(cmd);
1532         __dump_mappings(cmd);
1533         READ_UNLOCK(cmd);
1534 }
1535
1536 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1537 {
1538         int r;
1539
1540         READ_LOCK(cmd);
1541         r = cmd->changed;
1542         READ_UNLOCK(cmd);
1543
1544         return r;
1545 }
1546
1547 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1548 {
1549         int r;
1550         unsigned flags;
1551         dm_oblock_t oblock;
1552         __le64 value;
1553
1554         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1555         if (r)
1556                 return r;
1557
1558         unpack_value(value, &oblock, &flags);
1559
1560         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1561                 /* nothing to be done */
1562                 return 0;
1563
1564         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1565         __dm_bless_for_disk(&value);
1566
1567         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1568                                &value, &cmd->root);
1569         if (r)
1570                 return r;
1571
1572         cmd->changed = true;
1573         return 0;
1574
1575 }
1576
1577 static int __set_dirty_bits_v1(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1578 {
1579         int r;
1580         unsigned i;
1581         for (i = 0; i < nr_bits; i++) {
1582                 r = __dirty(cmd, to_cblock(i), test_bit(i, bits));
1583                 if (r)
1584                         return r;
1585         }
1586
1587         return 0;
1588 }
1589
1590 static int is_dirty_callback(uint32_t index, bool *value, void *context)
1591 {
1592         unsigned long *bits = context;
1593         *value = test_bit(index, bits);
1594         return 0;
1595 }
1596
1597 static int __set_dirty_bits_v2(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1598 {
1599         int r = 0;
1600
1601         /* nr_bits is really just a sanity check */
1602         if (nr_bits != from_cblock(cmd->cache_blocks)) {
1603                 DMERR("dirty bitset is wrong size");
1604                 return -EINVAL;
1605         }
1606
1607         r = dm_bitset_del(&cmd->dirty_info, cmd->dirty_root);
1608         if (r)
1609                 return r;
1610
1611         cmd->changed = true;
1612         return dm_bitset_new(&cmd->dirty_info, &cmd->dirty_root, nr_bits, is_dirty_callback, bits);
1613 }
1614
1615 int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
1616                             unsigned nr_bits,
1617                             unsigned long *bits)
1618 {
1619         int r;
1620
1621         WRITE_LOCK(cmd);
1622         if (separate_dirty_bits(cmd))
1623                 r = __set_dirty_bits_v2(cmd, nr_bits, bits);
1624         else
1625                 r = __set_dirty_bits_v1(cmd, nr_bits, bits);
1626         WRITE_UNLOCK(cmd);
1627
1628         return r;
1629 }
1630
1631 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1632                                  struct dm_cache_statistics *stats)
1633 {
1634         READ_LOCK_VOID(cmd);
1635         *stats = cmd->stats;
1636         READ_UNLOCK(cmd);
1637 }
1638
1639 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1640                                  struct dm_cache_statistics *stats)
1641 {
1642         WRITE_LOCK_VOID(cmd);
1643         cmd->stats = *stats;
1644         WRITE_UNLOCK(cmd);
1645 }
1646
1647 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1648 {
1649         int r = -EINVAL;
1650         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1651                                  clear_clean_shutdown);
1652
1653         WRITE_LOCK(cmd);
1654         if (cmd->fail_io)
1655                 goto out;
1656
1657         r = __commit_transaction(cmd, mutator);
1658         if (r)
1659                 goto out;
1660
1661         r = __begin_transaction(cmd);
1662 out:
1663         WRITE_UNLOCK(cmd);
1664         return r;
1665 }
1666
1667 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1668                                            dm_block_t *result)
1669 {
1670         int r = -EINVAL;
1671
1672         READ_LOCK(cmd);
1673         if (!cmd->fail_io)
1674                 r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1675         READ_UNLOCK(cmd);
1676
1677         return r;
1678 }
1679
1680 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1681                                    dm_block_t *result)
1682 {
1683         int r = -EINVAL;
1684
1685         READ_LOCK(cmd);
1686         if (!cmd->fail_io)
1687                 r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1688         READ_UNLOCK(cmd);
1689
1690         return r;
1691 }
1692
1693 /*----------------------------------------------------------------*/
1694
1695 static int get_hint(uint32_t index, void *value_le, void *context)
1696 {
1697         uint32_t value;
1698         struct dm_cache_policy *policy = context;
1699
1700         value = policy_get_hint(policy, to_cblock(index));
1701         *((__le32 *) value_le) = cpu_to_le32(value);
1702
1703         return 0;
1704 }
1705
1706 /*
1707  * It's quicker to always delete the hint array, and recreate with
1708  * dm_array_new().
1709  */
1710 static int write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1711 {
1712         int r;
1713         size_t hint_size;
1714         const char *policy_name = dm_cache_policy_get_name(policy);
1715         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1716
1717         if (!policy_name[0] ||
1718             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1719                 return -EINVAL;
1720
1721         strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1722         memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1723
1724         hint_size = dm_cache_policy_get_hint_size(policy);
1725         if (!hint_size)
1726                 return 0; /* short-circuit hints initialization */
1727         cmd->policy_hint_size = hint_size;
1728
1729         if (cmd->hint_root) {
1730                 r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1731                 if (r)
1732                         return r;
1733         }
1734
1735         return dm_array_new(&cmd->hint_info, &cmd->hint_root,
1736                             from_cblock(cmd->cache_blocks),
1737                             get_hint, policy);
1738 }
1739
1740 int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1741 {
1742         int r;
1743
1744         WRITE_LOCK(cmd);
1745         r = write_hints(cmd, policy);
1746         WRITE_UNLOCK(cmd);
1747
1748         return r;
1749 }
1750
1751 int dm_cache_metadata_all_clean(struct dm_cache_metadata *cmd, bool *result)
1752 {
1753         int r;
1754
1755         READ_LOCK(cmd);
1756         r = blocks_are_unmapped_or_clean(cmd, 0, cmd->cache_blocks, result);
1757         READ_UNLOCK(cmd);
1758
1759         return r;
1760 }
1761
1762 void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd)
1763 {
1764         WRITE_LOCK_VOID(cmd);
1765         dm_bm_set_read_only(cmd->bm);
1766         WRITE_UNLOCK(cmd);
1767 }
1768
1769 void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd)
1770 {
1771         WRITE_LOCK_VOID(cmd);
1772         dm_bm_set_read_write(cmd->bm);
1773         WRITE_UNLOCK(cmd);
1774 }
1775
1776 int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd)
1777 {
1778         int r;
1779         struct dm_block *sblock;
1780         struct cache_disk_superblock *disk_super;
1781
1782         WRITE_LOCK(cmd);
1783         set_bit(NEEDS_CHECK, &cmd->flags);
1784
1785         r = superblock_lock(cmd, &sblock);
1786         if (r) {
1787                 DMERR("couldn't read superblock");
1788                 goto out;
1789         }
1790
1791         disk_super = dm_block_data(sblock);
1792         disk_super->flags = cpu_to_le32(cmd->flags);
1793
1794         dm_bm_unlock(sblock);
1795
1796 out:
1797         WRITE_UNLOCK(cmd);
1798         return r;
1799 }
1800
1801 int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result)
1802 {
1803         READ_LOCK(cmd);
1804         *result = !!test_bit(NEEDS_CHECK, &cmd->flags);
1805         READ_UNLOCK(cmd);
1806
1807         return 0;
1808 }
1809
1810 int dm_cache_metadata_abort(struct dm_cache_metadata *cmd)
1811 {
1812         int r = -EINVAL;
1813         struct dm_block_manager *old_bm = NULL, *new_bm = NULL;
1814
1815         /* fail_io is double-checked with cmd->root_lock held below */
1816         if (unlikely(cmd->fail_io))
1817                 return r;
1818
1819         /*
1820          * Replacement block manager (new_bm) is created and old_bm destroyed outside of
1821          * cmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of
1822          * shrinker associated with the block manager's bufio client vs cmd root_lock).
1823          * - must take shrinker_rwsem without holding cmd->root_lock
1824          */
1825         new_bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
1826                                          CACHE_MAX_CONCURRENT_LOCKS);
1827
1828         WRITE_LOCK(cmd);
1829         if (cmd->fail_io) {
1830                 WRITE_UNLOCK(cmd);
1831                 goto out;
1832         }
1833
1834         __destroy_persistent_data_objects(cmd, false);
1835         old_bm = cmd->bm;
1836         if (IS_ERR(new_bm)) {
1837                 DMERR("could not create block manager during abort");
1838                 cmd->bm = NULL;
1839                 r = PTR_ERR(new_bm);
1840                 goto out_unlock;
1841         }
1842
1843         cmd->bm = new_bm;
1844         r = __open_or_format_metadata(cmd, false);
1845         if (r) {
1846                 cmd->bm = NULL;
1847                 goto out_unlock;
1848         }
1849         new_bm = NULL;
1850 out_unlock:
1851         if (r)
1852                 cmd->fail_io = true;
1853         WRITE_UNLOCK(cmd);
1854         dm_block_manager_destroy(old_bm);
1855 out:
1856         if (new_bm && !IS_ERR(new_bm))
1857                 dm_block_manager_destroy(new_bm);
1858
1859         return r;
1860 }