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 #define CACHE_VERSION 1
  24 #define CACHE_METADATA_CACHE_SIZE 64
  25
  26 /*
  27  *  3 for btree insert +
  28  *  2 for btree lookup used within space map
  29  */
  30 #define CACHE_MAX_CONCURRENT_LOCKS 5
  31 #define SPACE_MAP_ROOT_SIZE 128
  32
  33 enum superblock_flag_bits {
  34         /* for spotting crashes that would invalidate the dirty bitset */
  35         CLEAN_SHUTDOWN,
  36 };
  37
  38 /*
  39  * Each mapping from cache block -> origin block carries a set of flags.
  40  */
  41 enum mapping_bits {
  42         /*
  43          * A valid mapping.  Because we're using an array we clear this
  44          * flag for an non existant mapping.
  45          */
  46         M_VALID = 1,
  47
  48         /*
  49          * The data on the cache is different from that on the origin.
  50          */
  51         M_DIRTY = 2
  52 };
  53
  54 struct cache_disk_superblock {
  55         __le32 csum;
  56         __le32 flags;
  57         __le64 blocknr;
  58
  59         __u8 uuid[16];
  60         __le64 magic;
  61         __le32 version;
  62
  63         __u8 policy_name[CACHE_POLICY_NAME_SIZE];
  64         __le32 policy_hint_size;
  65
  66         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
  67         __le64 mapping_root;
  68         __le64 hint_root;
  69
  70         __le64 discard_root;
  71         __le64 discard_block_size;
  72         __le64 discard_nr_blocks;
  73
  74         __le32 data_block_size;
  75         __le32 metadata_block_size;
  76         __le32 cache_blocks;
  77
  78         __le32 compat_flags;
  79         __le32 compat_ro_flags;
  80         __le32 incompat_flags;
  81
  82         __le32 read_hits;
  83         __le32 read_misses;
  84         __le32 write_hits;
  85         __le32 write_misses;
  86
  87         __le32 policy_version[CACHE_POLICY_VERSION_SIZE];
  88 } __packed;
  89
  90 struct dm_cache_metadata {
  91         struct block_device *bdev;
  92         struct dm_block_manager *bm;
  93         struct dm_space_map *metadata_sm;
  94         struct dm_transaction_manager *tm;
  95
  96         struct dm_array_info info;
  97         struct dm_array_info hint_info;
  98         struct dm_disk_bitset discard_info;
  99
 100         struct rw_semaphore root_lock;
 101         dm_block_t root;
 102         dm_block_t hint_root;
 103         dm_block_t discard_root;
 104
 105         sector_t discard_block_size;
 106         dm_dblock_t discard_nr_blocks;
 107
 108         sector_t data_block_size;
 109         dm_cblock_t cache_blocks;
 110         bool changed:1;
 111         bool clean_when_opened:1;
 112
 113         char policy_name[CACHE_POLICY_NAME_SIZE];
 114         unsigned policy_version[CACHE_POLICY_VERSION_SIZE];
 115         size_t policy_hint_size;
 116         struct dm_cache_statistics stats;
 117 };
 118
 119 /*-------------------------------------------------------------------
 120  * superblock validator
 121  *-----------------------------------------------------------------*/
 122
 123 #define SUPERBLOCK_CSUM_XOR 9031977
 124
 125 static void sb_prepare_for_write(struct dm_block_validator *v,
 126                                  struct dm_block *b,
 127                                  size_t sb_block_size)
 128 {
 129         struct cache_disk_superblock *disk_super = dm_block_data(b);
 130
 131         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 132         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 133                                                       sb_block_size - sizeof(__le32),
 134                                                       SUPERBLOCK_CSUM_XOR));
 135 }
 136
 137 static int sb_check(struct dm_block_validator *v,
 138                     struct dm_block *b,
 139                     size_t sb_block_size)
 140 {
 141         struct cache_disk_superblock *disk_super = dm_block_data(b);
 142         __le32 csum_le;
 143
 144         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 145                 DMERR("sb_check failed: blocknr %llu: wanted %llu",
 146                       le64_to_cpu(disk_super->blocknr),
 147                       (unsigned long long)dm_block_location(b));
 148                 return -ENOTBLK;
 149         }
 150
 151         if (le64_to_cpu(disk_super->magic) != CACHE_SUPERBLOCK_MAGIC) {
 152                 DMERR("sb_check failed: magic %llu: wanted %llu",
 153                       le64_to_cpu(disk_super->magic),
 154                       (unsigned long long)CACHE_SUPERBLOCK_MAGIC);
 155                 return -EILSEQ;
 156         }
 157
 158         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 159                                              sb_block_size - sizeof(__le32),
 160                                              SUPERBLOCK_CSUM_XOR));
 161         if (csum_le != disk_super->csum) {
 162                 DMERR("sb_check failed: csum %u: wanted %u",
 163                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 164                 return -EILSEQ;
 165         }
 166
 167         return 0;
 168 }
 169
 170 static struct dm_block_validator sb_validator = {
 171         .name = "superblock",
 172         .prepare_for_write = sb_prepare_for_write,
 173         .check = sb_check
 174 };
 175
 176 /*----------------------------------------------------------------*/
 177
 178 static int superblock_read_lock(struct dm_cache_metadata *cmd,
 179                                 struct dm_block **sblock)
 180 {
 181         return dm_bm_read_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 182                                &sb_validator, sblock);
 183 }
 184
 185 static int superblock_lock_zero(struct dm_cache_metadata *cmd,
 186                                 struct dm_block **sblock)
 187 {
 188         return dm_bm_write_lock_zero(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 189                                      &sb_validator, sblock);
 190 }
 191
 192 static int superblock_lock(struct dm_cache_metadata *cmd,
 193                            struct dm_block **sblock)
 194 {
 195         return dm_bm_write_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 196                                 &sb_validator, sblock);
 197 }
 198
 199 /*----------------------------------------------------------------*/
 200
 201 static int __superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
 202 {
 203         int r;
 204         unsigned i;
 205         struct dm_block *b;
 206         __le64 *data_le, zero = cpu_to_le64(0);
 207         unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
 208
 209         /*
 210          * We can't use a validator here - it may be all zeroes.
 211          */
 212         r = dm_bm_read_lock(bm, CACHE_SUPERBLOCK_LOCATION, NULL, &b);
 213         if (r)
 214                 return r;
 215
 216         data_le = dm_block_data(b);
 217         *result = true;
 218         for (i = 0; i < sb_block_size; i++) {
 219                 if (data_le[i] != zero) {
 220                         *result = false;
 221                         break;
 222                 }
 223         }
 224
 225         return dm_bm_unlock(b);
 226 }
 227
 228 static void __setup_mapping_info(struct dm_cache_metadata *cmd)
 229 {
 230         struct dm_btree_value_type vt;
 231
 232         vt.context = NULL;
 233         vt.size = sizeof(__le64);
 234         vt.inc = NULL;
 235         vt.dec = NULL;
 236         vt.equal = NULL;
 237         dm_array_info_init(&cmd->info, cmd->tm, &vt);
 238
 239         if (cmd->policy_hint_size) {
 240                 vt.size = sizeof(__le32);
 241                 dm_array_info_init(&cmd->hint_info, cmd->tm, &vt);
 242         }
 243 }
 244
 245 static int __write_initial_superblock(struct dm_cache_metadata *cmd)
 246 {
 247         int r;
 248         struct dm_block *sblock;
 249         size_t metadata_len;
 250         struct cache_disk_superblock *disk_super;
 251         sector_t bdev_size = i_size_read(cmd->bdev->bd_inode) >> SECTOR_SHIFT;
 252
 253         /* FIXME: see if we can lose the max sectors limit */
 254         if (bdev_size > DM_CACHE_METADATA_MAX_SECTORS)
 255                 bdev_size = DM_CACHE_METADATA_MAX_SECTORS;
 256
 257         r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
 258         if (r < 0)
 259                 return r;
 260
 261         r = dm_tm_pre_commit(cmd->tm);
 262         if (r < 0)
 263                 return r;
 264
 265         r = superblock_lock_zero(cmd, &sblock);
 266         if (r)
 267                 return r;
 268
 269         disk_super = dm_block_data(sblock);
 270         disk_super->flags = 0;
 271         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 272         disk_super->magic = cpu_to_le64(CACHE_SUPERBLOCK_MAGIC);
 273         disk_super->version = cpu_to_le32(CACHE_VERSION);
 274         memset(disk_super->policy_name, 0, sizeof(disk_super->policy_name));
 275         memset(disk_super->policy_version, 0, sizeof(disk_super->policy_version));
 276         disk_super->policy_hint_size = 0;
 277
 278         r = dm_sm_copy_root(cmd->metadata_sm, &disk_super->metadata_space_map_root,
 279                             metadata_len);
 280         if (r < 0)
 281                 goto bad_locked;
 282
 283         disk_super->mapping_root = cpu_to_le64(cmd->root);
 284         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 285         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 286         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 287         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 288         disk_super->metadata_block_size = cpu_to_le32(DM_CACHE_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
 289         disk_super->data_block_size = cpu_to_le32(cmd->data_block_size);
 290         disk_super->cache_blocks = cpu_to_le32(0);
 291
 292         disk_super->read_hits = cpu_to_le32(0);
 293         disk_super->read_misses = cpu_to_le32(0);
 294         disk_super->write_hits = cpu_to_le32(0);
 295         disk_super->write_misses = cpu_to_le32(0);
 296
 297         return dm_tm_commit(cmd->tm, sblock);
 298
 299 bad_locked:
 300         dm_bm_unlock(sblock);
 301         return r;
 302 }
 303
 304 static int __format_metadata(struct dm_cache_metadata *cmd)
 305 {
 306         int r;
 307
 308         r = dm_tm_create_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 309                                  &cmd->tm, &cmd->metadata_sm);
 310         if (r < 0) {
 311                 DMERR("tm_create_with_sm failed");
 312                 return r;
 313         }
 314
 315         __setup_mapping_info(cmd);
 316
 317         r = dm_array_empty(&cmd->info, &cmd->root);
 318         if (r < 0)
 319                 goto bad;
 320
 321         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 322
 323         r = dm_bitset_empty(&cmd->discard_info, &cmd->discard_root);
 324         if (r < 0)
 325                 goto bad;
 326
 327         cmd->discard_block_size = 0;
 328         cmd->discard_nr_blocks = 0;
 329
 330         r = __write_initial_superblock(cmd);
 331         if (r)
 332                 goto bad;
 333
 334         cmd->clean_when_opened = true;
 335         return 0;
 336
 337 bad:
 338         dm_tm_destroy(cmd->tm);
 339         dm_sm_destroy(cmd->metadata_sm);
 340
 341         return r;
 342 }
 343
 344 static int __check_incompat_features(struct cache_disk_superblock *disk_super,
 345                                      struct dm_cache_metadata *cmd)
 346 {
 347         uint32_t features;
 348
 349         features = le32_to_cpu(disk_super->incompat_flags) & ~DM_CACHE_FEATURE_INCOMPAT_SUPP;
 350         if (features) {
 351                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 352                       (unsigned long)features);
 353                 return -EINVAL;
 354         }
 355
 356         /*
 357          * Check for read-only metadata to skip the following RDWR checks.
 358          */
 359         if (get_disk_ro(cmd->bdev->bd_disk))
 360                 return 0;
 361
 362         features = le32_to_cpu(disk_super->compat_ro_flags) & ~DM_CACHE_FEATURE_COMPAT_RO_SUPP;
 363         if (features) {
 364                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 365                       (unsigned long)features);
 366                 return -EINVAL;
 367         }
 368
 369         return 0;
 370 }
 371
 372 static int __open_metadata(struct dm_cache_metadata *cmd)
 373 {
 374         int r;
 375         struct dm_block *sblock;
 376         struct cache_disk_superblock *disk_super;
 377         unsigned long sb_flags;
 378
 379         r = superblock_read_lock(cmd, &sblock);
 380         if (r < 0) {
 381                 DMERR("couldn't read lock superblock");
 382                 return r;
 383         }
 384
 385         disk_super = dm_block_data(sblock);
 386
 387         r = __check_incompat_features(disk_super, cmd);
 388         if (r < 0)
 389                 goto bad;
 390
 391         r = dm_tm_open_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 392                                disk_super->metadata_space_map_root,
 393                                sizeof(disk_super->metadata_space_map_root),
 394                                &cmd->tm, &cmd->metadata_sm);
 395         if (r < 0) {
 396                 DMERR("tm_open_with_sm failed");
 397                 goto bad;
 398         }
 399
 400         __setup_mapping_info(cmd);
 401         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 402         sb_flags = le32_to_cpu(disk_super->flags);
 403         cmd->clean_when_opened = test_bit(CLEAN_SHUTDOWN, &sb_flags);
 404         return dm_bm_unlock(sblock);
 405
 406 bad:
 407         dm_bm_unlock(sblock);
 408         return r;
 409 }
 410
 411 static int __open_or_format_metadata(struct dm_cache_metadata *cmd,
 412                                      bool format_device)
 413 {
 414         int r;
 415         bool unformatted = false;
 416
 417         r = __superblock_all_zeroes(cmd->bm, &unformatted);
 418         if (r)
 419                 return r;
 420
 421         if (unformatted)
 422                 return format_device ? __format_metadata(cmd) : -EPERM;
 423
 424         return __open_metadata(cmd);
 425 }
 426
 427 static int __create_persistent_data_objects(struct dm_cache_metadata *cmd,
 428                                             bool may_format_device)
 429 {
 430         int r;
 431         cmd->bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE,
 432                                           CACHE_METADATA_CACHE_SIZE,
 433                                           CACHE_MAX_CONCURRENT_LOCKS);
 434         if (IS_ERR(cmd->bm)) {
 435                 DMERR("could not create block manager");
 436                 return PTR_ERR(cmd->bm);
 437         }
 438
 439         r = __open_or_format_metadata(cmd, may_format_device);
 440         if (r)
 441                 dm_block_manager_destroy(cmd->bm);
 442
 443         return r;
 444 }
 445
 446 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd)
 447 {
 448         dm_sm_destroy(cmd->metadata_sm);
 449         dm_tm_destroy(cmd->tm);
 450         dm_block_manager_destroy(cmd->bm);
 451 }
 452
 453 typedef unsigned long (*flags_mutator)(unsigned long);
 454
 455 static void update_flags(struct cache_disk_superblock *disk_super,
 456                          flags_mutator mutator)
 457 {
 458         uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
 459         disk_super->flags = cpu_to_le32(sb_flags);
 460 }
 461
 462 static unsigned long set_clean_shutdown(unsigned long flags)
 463 {
 464         set_bit(CLEAN_SHUTDOWN, &flags);
 465         return flags;
 466 }
 467
 468 static unsigned long clear_clean_shutdown(unsigned long flags)
 469 {
 470         clear_bit(CLEAN_SHUTDOWN, &flags);
 471         return flags;
 472 }
 473
 474 static void read_superblock_fields(struct dm_cache_metadata *cmd,
 475                                    struct cache_disk_superblock *disk_super)
 476 {
 477         cmd->root = le64_to_cpu(disk_super->mapping_root);
 478         cmd->hint_root = le64_to_cpu(disk_super->hint_root);
 479         cmd->discard_root = le64_to_cpu(disk_super->discard_root);
 480         cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
 481         cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
 482         cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 483         cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
 484         strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
 485         cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
 486         cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
 487         cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
 488         cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
 489
 490         cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
 491         cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
 492         cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
 493         cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
 494
 495         cmd->changed = false;
 496 }
 497
 498 /*
 499  * The mutator updates the superblock flags.
 500  */
 501 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
 502                                      flags_mutator mutator)
 503 {
 504         int r;
 505         struct cache_disk_superblock *disk_super;
 506         struct dm_block *sblock;
 507
 508         r = superblock_lock(cmd, &sblock);
 509         if (r)
 510                 return r;
 511
 512         disk_super = dm_block_data(sblock);
 513         update_flags(disk_super, mutator);
 514         read_superblock_fields(cmd, disk_super);
 515
 516         return dm_bm_flush_and_unlock(cmd->bm, sblock);
 517 }
 518
 519 static int __begin_transaction(struct dm_cache_metadata *cmd)
 520 {
 521         int r;
 522         struct cache_disk_superblock *disk_super;
 523         struct dm_block *sblock;
 524
 525         /*
 526          * We re-read the superblock every time.  Shouldn't need to do this
 527          * really.
 528          */
 529         r = superblock_read_lock(cmd, &sblock);
 530         if (r)
 531                 return r;
 532
 533         disk_super = dm_block_data(sblock);
 534         read_superblock_fields(cmd, disk_super);
 535         dm_bm_unlock(sblock);
 536
 537         return 0;
 538 }
 539
 540 static int __commit_transaction(struct dm_cache_metadata *cmd,
 541                                 flags_mutator mutator)
 542 {
 543         int r;
 544         size_t metadata_len;
 545         struct cache_disk_superblock *disk_super;
 546         struct dm_block *sblock;
 547
 548         /*
 549          * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
 550          */
 551         BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
 552
 553         r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
 554                             &cmd->discard_root);
 555         if (r)
 556                 return r;
 557
 558         r = dm_tm_pre_commit(cmd->tm);
 559         if (r < 0)
 560                 return r;
 561
 562         r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
 563         if (r < 0)
 564                 return r;
 565
 566         r = superblock_lock(cmd, &sblock);
 567         if (r)
 568                 return r;
 569
 570         disk_super = dm_block_data(sblock);
 571
 572         if (mutator)
 573                 update_flags(disk_super, mutator);
 574
 575         disk_super->mapping_root = cpu_to_le64(cmd->root);
 576         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 577         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 578         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 579         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 580         disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
 581         strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
 582         disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
 583         disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
 584         disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
 585
 586         disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
 587         disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
 588         disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
 589         disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
 590
 591         r = dm_sm_copy_root(cmd->metadata_sm, &disk_super->metadata_space_map_root,
 592                             metadata_len);
 593         if (r < 0) {
 594                 dm_bm_unlock(sblock);
 595                 return r;
 596         }
 597
 598         return dm_tm_commit(cmd->tm, sblock);
 599 }
 600
 601 /*----------------------------------------------------------------*/
 602
 603 /*
 604  * The mappings are held in a dm-array that has 64-bit values stored in
 605  * little-endian format.  The index is the cblock, the high 48bits of the
 606  * value are the oblock and the low 16 bit the flags.
 607  */
 608 #define FLAGS_MASK ((1 << 16) - 1)
 609
 610 static __le64 pack_value(dm_oblock_t block, unsigned flags)
 611 {
 612         uint64_t value = from_oblock(block);
 613         value <<= 16;
 614         value = value | (flags & FLAGS_MASK);
 615         return cpu_to_le64(value);
 616 }
 617
 618 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
 619 {
 620         uint64_t value = le64_to_cpu(value_le);
 621         uint64_t b = value >> 16;
 622         *block = to_oblock(b);
 623         *flags = value & FLAGS_MASK;
 624 }
 625
 626 /*----------------------------------------------------------------*/
 627
 628 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
 629                                                  sector_t data_block_size,
 630                                                  bool may_format_device,
 631                                                  size_t policy_hint_size)
 632 {
 633         int r;
 634         struct dm_cache_metadata *cmd;
 635
 636         cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
 637         if (!cmd) {
 638                 DMERR("could not allocate metadata struct");
 639                 return NULL;
 640         }
 641
 642         init_rwsem(&cmd->root_lock);
 643         cmd->bdev = bdev;
 644         cmd->data_block_size = data_block_size;
 645         cmd->cache_blocks = 0;
 646         cmd->policy_hint_size = policy_hint_size;
 647         cmd->changed = true;
 648
 649         r = __create_persistent_data_objects(cmd, may_format_device);
 650         if (r) {
 651                 kfree(cmd);
 652                 return ERR_PTR(r);
 653         }
 654
 655         r = __begin_transaction_flags(cmd, clear_clean_shutdown);
 656         if (r < 0) {
 657                 dm_cache_metadata_close(cmd);
 658                 return ERR_PTR(r);
 659         }
 660
 661         return cmd;
 662 }
 663
 664 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
 665 {
 666         __destroy_persistent_data_objects(cmd);
 667         kfree(cmd);
 668 }
 669
 670 /*
 671  * Checks that the given cache block is either unmapped or clean.
 672  */
 673 static int block_unmapped_or_clean(struct dm_cache_metadata *cmd, dm_cblock_t b,
 674                                    bool *result)
 675 {
 676         int r;
 677         __le64 value;
 678         dm_oblock_t ob;
 679         unsigned flags;
 680
 681         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(b), &value);
 682         if (r) {
 683                 DMERR("block_unmapped_or_clean failed");
 684                 return r;
 685         }
 686
 687         unpack_value(value, &ob, &flags);
 688         *result = !((flags & M_VALID) && (flags & M_DIRTY));
 689
 690         return 0;
 691 }
 692
 693 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
 694                                         dm_cblock_t begin, dm_cblock_t end,
 695                                         bool *result)
 696 {
 697         int r;
 698         *result = true;
 699
 700         while (begin != end) {
 701                 r = block_unmapped_or_clean(cmd, begin, result);
 702                 if (r)
 703                         return r;
 704
 705                 if (!*result) {
 706                         DMERR("cache block %llu is dirty",
 707                               (unsigned long long) from_cblock(begin));
 708                         return 0;
 709                 }
 710
 711                 begin = to_cblock(from_cblock(begin) + 1);
 712         }
 713
 714         return 0;
 715 }
 716
 717 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
 718 {
 719         int r;
 720         bool clean;
 721         __le64 null_mapping = pack_value(0, 0);
 722
 723         down_write(&cmd->root_lock);
 724         __dm_bless_for_disk(&null_mapping);
 725
 726         if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
 727                 r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
 728                 if (r) {
 729                         __dm_unbless_for_disk(&null_mapping);
 730                         goto out;
 731                 }
 732
 733                 if (!clean) {
 734                         DMERR("unable to shrink cache due to dirty blocks");
 735                         r = -EINVAL;
 736                         __dm_unbless_for_disk(&null_mapping);
 737                         goto out;
 738                 }
 739         }
 740
 741         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
 742                             from_cblock(new_cache_size),
 743                             &null_mapping, &cmd->root);
 744         if (!r)
 745                 cmd->cache_blocks = new_cache_size;
 746         cmd->changed = true;
 747
 748 out:
 749         up_write(&cmd->root_lock);
 750
 751         return r;
 752 }
 753
 754 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
 755                                    sector_t discard_block_size,
 756                                    dm_dblock_t new_nr_entries)
 757 {
 758         int r;
 759
 760         down_write(&cmd->root_lock);
 761         r = dm_bitset_resize(&cmd->discard_info,
 762                              cmd->discard_root,
 763                              from_dblock(cmd->discard_nr_blocks),
 764                              from_dblock(new_nr_entries),
 765                              false, &cmd->discard_root);
 766         if (!r) {
 767                 cmd->discard_block_size = discard_block_size;
 768                 cmd->discard_nr_blocks = new_nr_entries;
 769         }
 770
 771         cmd->changed = true;
 772         up_write(&cmd->root_lock);
 773
 774         return r;
 775 }
 776
 777 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
 778 {
 779         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
 780                                  from_dblock(b), &cmd->discard_root);
 781 }
 782
 783 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
 784 {
 785         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
 786                                    from_dblock(b), &cmd->discard_root);
 787 }
 788
 789 static int __is_discarded(struct dm_cache_metadata *cmd, dm_dblock_t b,
 790                           bool *is_discarded)
 791 {
 792         return dm_bitset_test_bit(&cmd->discard_info, cmd->discard_root,
 793                                   from_dblock(b), &cmd->discard_root,
 794                                   is_discarded);
 795 }
 796
 797 static int __discard(struct dm_cache_metadata *cmd,
 798                      dm_dblock_t dblock, bool discard)
 799 {
 800         int r;
 801
 802         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
 803         if (r)
 804                 return r;
 805
 806         cmd->changed = true;
 807         return 0;
 808 }
 809
 810 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
 811                          dm_dblock_t dblock, bool discard)
 812 {
 813         int r;
 814
 815         down_write(&cmd->root_lock);
 816         r = __discard(cmd, dblock, discard);
 817         up_write(&cmd->root_lock);
 818
 819         return r;
 820 }
 821
 822 static int __load_discards(struct dm_cache_metadata *cmd,
 823                            load_discard_fn fn, void *context)
 824 {
 825         int r = 0;
 826         dm_block_t b;
 827         bool discard;
 828
 829         for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
 830                 dm_dblock_t dblock = to_dblock(b);
 831
 832                 if (cmd->clean_when_opened) {
 833                         r = __is_discarded(cmd, dblock, &discard);
 834                         if (r)
 835                                 return r;
 836                 } else
 837                         discard = false;
 838
 839                 r = fn(context, cmd->discard_block_size, dblock, discard);
 840                 if (r)
 841                         break;
 842         }
 843
 844         return r;
 845 }
 846
 847 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
 848                            load_discard_fn fn, void *context)
 849 {
 850         int r;
 851
 852         down_read(&cmd->root_lock);
 853         r = __load_discards(cmd, fn, context);
 854         up_read(&cmd->root_lock);
 855
 856         return r;
 857 }
 858
 859 dm_cblock_t dm_cache_size(struct dm_cache_metadata *cmd)
 860 {
 861         dm_cblock_t r;
 862
 863         down_read(&cmd->root_lock);
 864         r = cmd->cache_blocks;
 865         up_read(&cmd->root_lock);
 866
 867         return r;
 868 }
 869
 870 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
 871 {
 872         int r;
 873         __le64 value = pack_value(0, 0);
 874
 875         __dm_bless_for_disk(&value);
 876         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
 877                                &value, &cmd->root);
 878         if (r)
 879                 return r;
 880
 881         cmd->changed = true;
 882         return 0;
 883 }
 884
 885 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
 886 {
 887         int r;
 888
 889         down_write(&cmd->root_lock);
 890         r = __remove(cmd, cblock);
 891         up_write(&cmd->root_lock);
 892
 893         return r;
 894 }
 895
 896 static int __insert(struct dm_cache_metadata *cmd,
 897                     dm_cblock_t cblock, dm_oblock_t oblock)
 898 {
 899         int r;
 900         __le64 value = pack_value(oblock, M_VALID);
 901         __dm_bless_for_disk(&value);
 902
 903         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
 904                                &value, &cmd->root);
 905         if (r)
 906                 return r;
 907
 908         cmd->changed = true;
 909         return 0;
 910 }
 911
 912 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
 913                             dm_cblock_t cblock, dm_oblock_t oblock)
 914 {
 915         int r;
 916
 917         down_write(&cmd->root_lock);
 918         r = __insert(cmd, cblock, oblock);
 919         up_write(&cmd->root_lock);
 920
 921         return r;
 922 }
 923
 924 struct thunk {
 925         load_mapping_fn fn;
 926         void *context;
 927
 928         struct dm_cache_metadata *cmd;
 929         bool respect_dirty_flags;
 930         bool hints_valid;
 931 };
 932
 933 static bool policy_unchanged(struct dm_cache_metadata *cmd,
 934                              struct dm_cache_policy *policy)
 935 {
 936         const char *policy_name = dm_cache_policy_get_name(policy);
 937         const unsigned *policy_version = dm_cache_policy_get_version(policy);
 938         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
 939
 940         /*
 941          * Ensure policy names match.
 942          */
 943         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
 944                 return false;
 945
 946         /*
 947          * Ensure policy major versions match.
 948          */
 949         if (cmd->policy_version[0] != policy_version[0])
 950                 return false;
 951
 952         /*
 953          * Ensure policy hint sizes match.
 954          */
 955         if (cmd->policy_hint_size != policy_hint_size)
 956                 return false;
 957
 958         return true;
 959 }
 960
 961 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
 962 {
 963         return cmd->hint_root && cmd->policy_hint_size;
 964 }
 965
 966 static bool hints_array_available(struct dm_cache_metadata *cmd,
 967                                   struct dm_cache_policy *policy)
 968 {
 969         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
 970                 hints_array_initialized(cmd);
 971 }
 972
 973 static int __load_mapping(void *context, uint64_t cblock, void *leaf)
 974 {
 975         int r = 0;
 976         bool dirty;
 977         __le64 value;
 978         __le32 hint_value = 0;
 979         dm_oblock_t oblock;
 980         unsigned flags;
 981         struct thunk *thunk = context;
 982         struct dm_cache_metadata *cmd = thunk->cmd;
 983
 984         memcpy(&value, leaf, sizeof(value));
 985         unpack_value(value, &oblock, &flags);
 986
 987         if (flags & M_VALID) {
 988                 if (thunk->hints_valid) {
 989                         r = dm_array_get_value(&cmd->hint_info, cmd->hint_root,
 990                                                cblock, &hint_value);
 991                         if (r && r != -ENODATA)
 992                                 return r;
 993                 }
 994
 995                 dirty = thunk->respect_dirty_flags ? (flags & M_DIRTY) : true;
 996                 r = thunk->fn(thunk->context, oblock, to_cblock(cblock),
 997                               dirty, le32_to_cpu(hint_value), thunk->hints_valid);
 998         }
 999
1000         return r;
1001 }
1002
1003 static int __load_mappings(struct dm_cache_metadata *cmd,
1004                            struct dm_cache_policy *policy,
1005                            load_mapping_fn fn, void *context)
1006 {
1007         struct thunk thunk;
1008
1009         thunk.fn = fn;
1010         thunk.context = context;
1011
1012         thunk.cmd = cmd;
1013         thunk.respect_dirty_flags = cmd->clean_when_opened;
1014         thunk.hints_valid = hints_array_available(cmd, policy);
1015
1016         return dm_array_walk(&cmd->info, cmd->root, __load_mapping, &thunk);
1017 }
1018
1019 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1020                            struct dm_cache_policy *policy,
1021                            load_mapping_fn fn, void *context)
1022 {
1023         int r;
1024
1025         down_read(&cmd->root_lock);
1026         r = __load_mappings(cmd, policy, fn, context);
1027         up_read(&cmd->root_lock);
1028
1029         return r;
1030 }
1031
1032 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1033 {
1034         int r = 0;
1035         __le64 value;
1036         dm_oblock_t oblock;
1037         unsigned flags;
1038
1039         memcpy(&value, leaf, sizeof(value));
1040         unpack_value(value, &oblock, &flags);
1041
1042         return r;
1043 }
1044
1045 static int __dump_mappings(struct dm_cache_metadata *cmd)
1046 {
1047         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1048 }
1049
1050 void dm_cache_dump(struct dm_cache_metadata *cmd)
1051 {
1052         down_read(&cmd->root_lock);
1053         __dump_mappings(cmd);
1054         up_read(&cmd->root_lock);
1055 }
1056
1057 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1058 {
1059         int r;
1060
1061         down_read(&cmd->root_lock);
1062         r = cmd->changed;
1063         up_read(&cmd->root_lock);
1064
1065         return r;
1066 }
1067
1068 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1069 {
1070         int r;
1071         unsigned flags;
1072         dm_oblock_t oblock;
1073         __le64 value;
1074
1075         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1076         if (r)
1077                 return r;
1078
1079         unpack_value(value, &oblock, &flags);
1080
1081         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1082                 /* nothing to be done */
1083                 return 0;
1084
1085         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1086         __dm_bless_for_disk(&value);
1087
1088         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1089                                &value, &cmd->root);
1090         if (r)
1091                 return r;
1092
1093         cmd->changed = true;
1094         return 0;
1095
1096 }
1097
1098 int dm_cache_set_dirty(struct dm_cache_metadata *cmd,
1099                        dm_cblock_t cblock, bool dirty)
1100 {
1101         int r;
1102
1103         down_write(&cmd->root_lock);
1104         r = __dirty(cmd, cblock, dirty);
1105         up_write(&cmd->root_lock);
1106
1107         return r;
1108 }
1109
1110 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1111                                  struct dm_cache_statistics *stats)
1112 {
1113         down_read(&cmd->root_lock);
1114         *stats = cmd->stats;
1115         up_read(&cmd->root_lock);
1116 }
1117
1118 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1119                                  struct dm_cache_statistics *stats)
1120 {
1121         down_write(&cmd->root_lock);
1122         cmd->stats = *stats;
1123         up_write(&cmd->root_lock);
1124 }
1125
1126 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1127 {
1128         int r;
1129         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1130                                  clear_clean_shutdown);
1131
1132         down_write(&cmd->root_lock);
1133         r = __commit_transaction(cmd, mutator);
1134         if (r)
1135                 goto out;
1136
1137         r = __begin_transaction(cmd);
1138
1139 out:
1140         up_write(&cmd->root_lock);
1141         return r;
1142 }
1143
1144 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1145                                            dm_block_t *result)
1146 {
1147         int r = -EINVAL;
1148
1149         down_read(&cmd->root_lock);
1150         r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1151         up_read(&cmd->root_lock);
1152
1153         return r;
1154 }
1155
1156 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1157                                    dm_block_t *result)
1158 {
1159         int r = -EINVAL;
1160
1161         down_read(&cmd->root_lock);
1162         r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1163         up_read(&cmd->root_lock);
1164
1165         return r;
1166 }
1167
1168 /*----------------------------------------------------------------*/
1169
1170 static int begin_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1171 {
1172         int r;
1173         __le32 value;
1174         size_t hint_size;
1175         const char *policy_name = dm_cache_policy_get_name(policy);
1176         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1177
1178         if (!policy_name[0] ||
1179             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1180                 return -EINVAL;
1181
1182         if (!policy_unchanged(cmd, policy)) {
1183                 strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1184                 memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1185
1186                 hint_size = dm_cache_policy_get_hint_size(policy);
1187                 if (!hint_size)
1188                         return 0; /* short-circuit hints initialization */
1189                 cmd->policy_hint_size = hint_size;
1190
1191                 if (cmd->hint_root) {
1192                         r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1193                         if (r)
1194                                 return r;
1195                 }
1196
1197                 r = dm_array_empty(&cmd->hint_info, &cmd->hint_root);
1198                 if (r)
1199                         return r;
1200
1201                 value = cpu_to_le32(0);
1202                 __dm_bless_for_disk(&value);
1203                 r = dm_array_resize(&cmd->hint_info, cmd->hint_root, 0,
1204                                     from_cblock(cmd->cache_blocks),
1205                                     &value, &cmd->hint_root);
1206                 if (r)
1207                         return r;
1208         }
1209
1210         return 0;
1211 }
1212
1213 int dm_cache_begin_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1214 {
1215         int r;
1216
1217         down_write(&cmd->root_lock);
1218         r = begin_hints(cmd, policy);
1219         up_write(&cmd->root_lock);
1220
1221         return r;
1222 }
1223
1224 static int save_hint(struct dm_cache_metadata *cmd, dm_cblock_t cblock,
1225                      uint32_t hint)
1226 {
1227         int r;
1228         __le32 value = cpu_to_le32(hint);
1229         __dm_bless_for_disk(&value);
1230
1231         r = dm_array_set_value(&cmd->hint_info, cmd->hint_root,
1232                                from_cblock(cblock), &value, &cmd->hint_root);
1233         cmd->changed = true;
1234
1235         return r;
1236 }
1237
1238 int dm_cache_save_hint(struct dm_cache_metadata *cmd, dm_cblock_t cblock,
1239                        uint32_t hint)
1240 {
1241         int r;
1242
1243         if (!hints_array_initialized(cmd))
1244                 return 0;
1245
1246         down_write(&cmd->root_lock);
1247         r = save_hint(cmd, cblock, hint);
1248         up_write(&cmd->root_lock);
1249
1250         return r;
1251 }