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, int *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 = 1;
 218         for (i = 0; i < sb_block_size; i++) {
 219                 if (data_le[i] != zero) {
 220                         *result = 0;
 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         /* Verify the data block size hasn't changed */
 388         if (le32_to_cpu(disk_super->data_block_size) != cmd->data_block_size) {
 389                 DMERR("changing the data block size (from %u to %llu) is not supported",
 390                       le32_to_cpu(disk_super->data_block_size),
 391                       (unsigned long long)cmd->data_block_size);
 392                 r = -EINVAL;
 393                 goto bad;
 394         }
 395
 396         r = __check_incompat_features(disk_super, cmd);
 397         if (r < 0)
 398                 goto bad;
 399
 400         r = dm_tm_open_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
 401                                disk_super->metadata_space_map_root,
 402                                sizeof(disk_super->metadata_space_map_root),
 403                                &cmd->tm, &cmd->metadata_sm);
 404         if (r < 0) {
 405                 DMERR("tm_open_with_sm failed");
 406                 goto bad;
 407         }
 408
 409         __setup_mapping_info(cmd);
 410         dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
 411         sb_flags = le32_to_cpu(disk_super->flags);
 412         cmd->clean_when_opened = test_bit(CLEAN_SHUTDOWN, &sb_flags);
 413         return dm_bm_unlock(sblock);
 414
 415 bad:
 416         dm_bm_unlock(sblock);
 417         return r;
 418 }
 419
 420 static int __open_or_format_metadata(struct dm_cache_metadata *cmd,
 421                                      bool format_device)
 422 {
 423         int r, unformatted;
 424
 425         r = __superblock_all_zeroes(cmd->bm, &unformatted);
 426         if (r)
 427                 return r;
 428
 429         if (unformatted)
 430                 return format_device ? __format_metadata(cmd) : -EPERM;
 431
 432         return __open_metadata(cmd);
 433 }
 434
 435 static int __create_persistent_data_objects(struct dm_cache_metadata *cmd,
 436                                             bool may_format_device)
 437 {
 438         int r;
 439         cmd->bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE,
 440                                           CACHE_METADATA_CACHE_SIZE,
 441                                           CACHE_MAX_CONCURRENT_LOCKS);
 442         if (IS_ERR(cmd->bm)) {
 443                 DMERR("could not create block manager");
 444                 return PTR_ERR(cmd->bm);
 445         }
 446
 447         r = __open_or_format_metadata(cmd, may_format_device);
 448         if (r)
 449                 dm_block_manager_destroy(cmd->bm);
 450
 451         return r;
 452 }
 453
 454 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd)
 455 {
 456         dm_sm_destroy(cmd->metadata_sm);
 457         dm_tm_destroy(cmd->tm);
 458         dm_block_manager_destroy(cmd->bm);
 459 }
 460
 461 typedef unsigned long (*flags_mutator)(unsigned long);
 462
 463 static void update_flags(struct cache_disk_superblock *disk_super,
 464                          flags_mutator mutator)
 465 {
 466         uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
 467         disk_super->flags = cpu_to_le32(sb_flags);
 468 }
 469
 470 static unsigned long set_clean_shutdown(unsigned long flags)
 471 {
 472         set_bit(CLEAN_SHUTDOWN, &flags);
 473         return flags;
 474 }
 475
 476 static unsigned long clear_clean_shutdown(unsigned long flags)
 477 {
 478         clear_bit(CLEAN_SHUTDOWN, &flags);
 479         return flags;
 480 }
 481
 482 static void read_superblock_fields(struct dm_cache_metadata *cmd,
 483                                    struct cache_disk_superblock *disk_super)
 484 {
 485         cmd->root = le64_to_cpu(disk_super->mapping_root);
 486         cmd->hint_root = le64_to_cpu(disk_super->hint_root);
 487         cmd->discard_root = le64_to_cpu(disk_super->discard_root);
 488         cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
 489         cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
 490         cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 491         cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
 492         strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
 493         cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
 494         cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
 495         cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
 496         cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
 497
 498         cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
 499         cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
 500         cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
 501         cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
 502
 503         cmd->changed = false;
 504 }
 505
 506 /*
 507  * The mutator updates the superblock flags.
 508  */
 509 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
 510                                      flags_mutator mutator)
 511 {
 512         int r;
 513         struct cache_disk_superblock *disk_super;
 514         struct dm_block *sblock;
 515
 516         r = superblock_lock(cmd, &sblock);
 517         if (r)
 518                 return r;
 519
 520         disk_super = dm_block_data(sblock);
 521         update_flags(disk_super, mutator);
 522         read_superblock_fields(cmd, disk_super);
 523         dm_bm_unlock(sblock);
 524
 525         return dm_bm_flush(cmd->bm);
 526 }
 527
 528 static int __begin_transaction(struct dm_cache_metadata *cmd)
 529 {
 530         int r;
 531         struct cache_disk_superblock *disk_super;
 532         struct dm_block *sblock;
 533
 534         /*
 535          * We re-read the superblock every time.  Shouldn't need to do this
 536          * really.
 537          */
 538         r = superblock_read_lock(cmd, &sblock);
 539         if (r)
 540                 return r;
 541
 542         disk_super = dm_block_data(sblock);
 543         read_superblock_fields(cmd, disk_super);
 544         dm_bm_unlock(sblock);
 545
 546         return 0;
 547 }
 548
 549 static int __commit_transaction(struct dm_cache_metadata *cmd,
 550                                 flags_mutator mutator)
 551 {
 552         int r;
 553         size_t metadata_len;
 554         struct cache_disk_superblock *disk_super;
 555         struct dm_block *sblock;
 556
 557         /*
 558          * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
 559          */
 560         BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
 561
 562         r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
 563                             &cmd->discard_root);
 564         if (r)
 565                 return r;
 566
 567         r = dm_tm_pre_commit(cmd->tm);
 568         if (r < 0)
 569                 return r;
 570
 571         r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
 572         if (r < 0)
 573                 return r;
 574
 575         r = superblock_lock(cmd, &sblock);
 576         if (r)
 577                 return r;
 578
 579         disk_super = dm_block_data(sblock);
 580
 581         if (mutator)
 582                 update_flags(disk_super, mutator);
 583
 584         disk_super->mapping_root = cpu_to_le64(cmd->root);
 585         disk_super->hint_root = cpu_to_le64(cmd->hint_root);
 586         disk_super->discard_root = cpu_to_le64(cmd->discard_root);
 587         disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
 588         disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
 589         disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
 590         strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
 591         disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
 592         disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
 593         disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
 594
 595         disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
 596         disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
 597         disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
 598         disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
 599
 600         r = dm_sm_copy_root(cmd->metadata_sm, &disk_super->metadata_space_map_root,
 601                             metadata_len);
 602         if (r < 0) {
 603                 dm_bm_unlock(sblock);
 604                 return r;
 605         }
 606
 607         return dm_tm_commit(cmd->tm, sblock);
 608 }
 609
 610 /*----------------------------------------------------------------*/
 611
 612 /*
 613  * The mappings are held in a dm-array that has 64-bit values stored in
 614  * little-endian format.  The index is the cblock, the high 48bits of the
 615  * value are the oblock and the low 16 bit the flags.
 616  */
 617 #define FLAGS_MASK ((1 << 16) - 1)
 618
 619 static __le64 pack_value(dm_oblock_t block, unsigned flags)
 620 {
 621         uint64_t value = from_oblock(block);
 622         value <<= 16;
 623         value = value | (flags & FLAGS_MASK);
 624         return cpu_to_le64(value);
 625 }
 626
 627 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
 628 {
 629         uint64_t value = le64_to_cpu(value_le);
 630         uint64_t b = value >> 16;
 631         *block = to_oblock(b);
 632         *flags = value & FLAGS_MASK;
 633 }
 634
 635 /*----------------------------------------------------------------*/
 636
 637 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
 638                                                  sector_t data_block_size,
 639                                                  bool may_format_device,
 640                                                  size_t policy_hint_size)
 641 {
 642         int r;
 643         struct dm_cache_metadata *cmd;
 644
 645         cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
 646         if (!cmd) {
 647                 DMERR("could not allocate metadata struct");
 648                 return NULL;
 649         }
 650
 651         init_rwsem(&cmd->root_lock);
 652         cmd->bdev = bdev;
 653         cmd->data_block_size = data_block_size;
 654         cmd->cache_blocks = 0;
 655         cmd->policy_hint_size = policy_hint_size;
 656         cmd->changed = true;
 657
 658         r = __create_persistent_data_objects(cmd, may_format_device);
 659         if (r) {
 660                 kfree(cmd);
 661                 return ERR_PTR(r);
 662         }
 663
 664         r = __begin_transaction_flags(cmd, clear_clean_shutdown);
 665         if (r < 0) {
 666                 dm_cache_metadata_close(cmd);
 667                 return ERR_PTR(r);
 668         }
 669
 670         return cmd;
 671 }
 672
 673 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
 674 {
 675         __destroy_persistent_data_objects(cmd);
 676         kfree(cmd);
 677 }
 678
 679 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
 680 {
 681         int r;
 682         __le64 null_mapping = pack_value(0, 0);
 683
 684         down_write(&cmd->root_lock);
 685         __dm_bless_for_disk(&null_mapping);
 686         r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
 687                             from_cblock(new_cache_size),
 688                             &null_mapping, &cmd->root);
 689         if (!r)
 690                 cmd->cache_blocks = new_cache_size;
 691         cmd->changed = true;
 692         up_write(&cmd->root_lock);
 693
 694         return r;
 695 }
 696
 697 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
 698                                    sector_t discard_block_size,
 699                                    dm_dblock_t new_nr_entries)
 700 {
 701         int r;
 702
 703         down_write(&cmd->root_lock);
 704         r = dm_bitset_resize(&cmd->discard_info,
 705                              cmd->discard_root,
 706                              from_dblock(cmd->discard_nr_blocks),
 707                              from_dblock(new_nr_entries),
 708                              false, &cmd->discard_root);
 709         if (!r) {
 710                 cmd->discard_block_size = discard_block_size;
 711                 cmd->discard_nr_blocks = new_nr_entries;
 712         }
 713
 714         cmd->changed = true;
 715         up_write(&cmd->root_lock);
 716
 717         return r;
 718 }
 719
 720 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
 721 {
 722         return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
 723                                  from_dblock(b), &cmd->discard_root);
 724 }
 725
 726 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
 727 {
 728         return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
 729                                    from_dblock(b), &cmd->discard_root);
 730 }
 731
 732 static int __is_discarded(struct dm_cache_metadata *cmd, dm_dblock_t b,
 733                           bool *is_discarded)
 734 {
 735         return dm_bitset_test_bit(&cmd->discard_info, cmd->discard_root,
 736                                   from_dblock(b), &cmd->discard_root,
 737                                   is_discarded);
 738 }
 739
 740 static int __discard(struct dm_cache_metadata *cmd,
 741                      dm_dblock_t dblock, bool discard)
 742 {
 743         int r;
 744
 745         r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
 746         if (r)
 747                 return r;
 748
 749         cmd->changed = true;
 750         return 0;
 751 }
 752
 753 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
 754                          dm_dblock_t dblock, bool discard)
 755 {
 756         int r;
 757
 758         down_write(&cmd->root_lock);
 759         r = __discard(cmd, dblock, discard);
 760         up_write(&cmd->root_lock);
 761
 762         return r;
 763 }
 764
 765 static int __load_discards(struct dm_cache_metadata *cmd,
 766                            load_discard_fn fn, void *context)
 767 {
 768         int r = 0;
 769         dm_block_t b;
 770         bool discard;
 771
 772         for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
 773                 dm_dblock_t dblock = to_dblock(b);
 774
 775                 if (cmd->clean_when_opened) {
 776                         r = __is_discarded(cmd, dblock, &discard);
 777                         if (r)
 778                                 return r;
 779                 } else
 780                         discard = false;
 781
 782                 r = fn(context, cmd->discard_block_size, dblock, discard);
 783                 if (r)
 784                         break;
 785         }
 786
 787         return r;
 788 }
 789
 790 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
 791                            load_discard_fn fn, void *context)
 792 {
 793         int r;
 794
 795         down_read(&cmd->root_lock);
 796         r = __load_discards(cmd, fn, context);
 797         up_read(&cmd->root_lock);
 798
 799         return r;
 800 }
 801
 802 dm_cblock_t dm_cache_size(struct dm_cache_metadata *cmd)
 803 {
 804         dm_cblock_t r;
 805
 806         down_read(&cmd->root_lock);
 807         r = cmd->cache_blocks;
 808         up_read(&cmd->root_lock);
 809
 810         return r;
 811 }
 812
 813 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
 814 {
 815         int r;
 816         __le64 value = pack_value(0, 0);
 817
 818         __dm_bless_for_disk(&value);
 819         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
 820                                &value, &cmd->root);
 821         if (r)
 822                 return r;
 823
 824         cmd->changed = true;
 825         return 0;
 826 }
 827
 828 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
 829 {
 830         int r;
 831
 832         down_write(&cmd->root_lock);
 833         r = __remove(cmd, cblock);
 834         up_write(&cmd->root_lock);
 835
 836         return r;
 837 }
 838
 839 static int __insert(struct dm_cache_metadata *cmd,
 840                     dm_cblock_t cblock, dm_oblock_t oblock)
 841 {
 842         int r;
 843         __le64 value = pack_value(oblock, M_VALID);
 844         __dm_bless_for_disk(&value);
 845
 846         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
 847                                &value, &cmd->root);
 848         if (r)
 849                 return r;
 850
 851         cmd->changed = true;
 852         return 0;
 853 }
 854
 855 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
 856                             dm_cblock_t cblock, dm_oblock_t oblock)
 857 {
 858         int r;
 859
 860         down_write(&cmd->root_lock);
 861         r = __insert(cmd, cblock, oblock);
 862         up_write(&cmd->root_lock);
 863
 864         return r;
 865 }
 866
 867 struct thunk {
 868         load_mapping_fn fn;
 869         void *context;
 870
 871         struct dm_cache_metadata *cmd;
 872         bool respect_dirty_flags;
 873         bool hints_valid;
 874 };
 875
 876 static bool policy_unchanged(struct dm_cache_metadata *cmd,
 877                              struct dm_cache_policy *policy)
 878 {
 879         const char *policy_name = dm_cache_policy_get_name(policy);
 880         const unsigned *policy_version = dm_cache_policy_get_version(policy);
 881         size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
 882
 883         /*
 884          * Ensure policy names match.
 885          */
 886         if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
 887                 return false;
 888
 889         /*
 890          * Ensure policy major versions match.
 891          */
 892         if (cmd->policy_version[0] != policy_version[0])
 893                 return false;
 894
 895         /*
 896          * Ensure policy hint sizes match.
 897          */
 898         if (cmd->policy_hint_size != policy_hint_size)
 899                 return false;
 900
 901         return true;
 902 }
 903
 904 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
 905 {
 906         return cmd->hint_root && cmd->policy_hint_size;
 907 }
 908
 909 static bool hints_array_available(struct dm_cache_metadata *cmd,
 910                                   struct dm_cache_policy *policy)
 911 {
 912         return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
 913                 hints_array_initialized(cmd);
 914 }
 915
 916 static int __load_mapping(void *context, uint64_t cblock, void *leaf)
 917 {
 918         int r = 0;
 919         bool dirty;
 920         __le64 value;
 921         __le32 hint_value = 0;
 922         dm_oblock_t oblock;
 923         unsigned flags;
 924         struct thunk *thunk = context;
 925         struct dm_cache_metadata *cmd = thunk->cmd;
 926
 927         memcpy(&value, leaf, sizeof(value));
 928         unpack_value(value, &oblock, &flags);
 929
 930         if (flags & M_VALID) {
 931                 if (thunk->hints_valid) {
 932                         r = dm_array_get_value(&cmd->hint_info, cmd->hint_root,
 933                                                cblock, &hint_value);
 934                         if (r && r != -ENODATA)
 935                                 return r;
 936                 }
 937
 938                 dirty = thunk->respect_dirty_flags ? (flags & M_DIRTY) : true;
 939                 r = thunk->fn(thunk->context, oblock, to_cblock(cblock),
 940                               dirty, le32_to_cpu(hint_value), thunk->hints_valid);
 941         }
 942
 943         return r;
 944 }
 945
 946 static int __load_mappings(struct dm_cache_metadata *cmd,
 947                            struct dm_cache_policy *policy,
 948                            load_mapping_fn fn, void *context)
 949 {
 950         struct thunk thunk;
 951
 952         thunk.fn = fn;
 953         thunk.context = context;
 954
 955         thunk.cmd = cmd;
 956         thunk.respect_dirty_flags = cmd->clean_when_opened;
 957         thunk.hints_valid = hints_array_available(cmd, policy);
 958
 959         return dm_array_walk(&cmd->info, cmd->root, __load_mapping, &thunk);
 960 }
 961
 962 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
 963                            struct dm_cache_policy *policy,
 964                            load_mapping_fn fn, void *context)
 965 {
 966         int r;
 967
 968         down_read(&cmd->root_lock);
 969         r = __load_mappings(cmd, policy, fn, context);
 970         up_read(&cmd->root_lock);
 971
 972         return r;
 973 }
 974
 975 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
 976 {
 977         int r = 0;
 978         __le64 value;
 979         dm_oblock_t oblock;
 980         unsigned flags;
 981
 982         memcpy(&value, leaf, sizeof(value));
 983         unpack_value(value, &oblock, &flags);
 984
 985         return r;
 986 }
 987
 988 static int __dump_mappings(struct dm_cache_metadata *cmd)
 989 {
 990         return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
 991 }
 992
 993 void dm_cache_dump(struct dm_cache_metadata *cmd)
 994 {
 995         down_read(&cmd->root_lock);
 996         __dump_mappings(cmd);
 997         up_read(&cmd->root_lock);
 998 }
 999
1000 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1001 {
1002         int r;
1003
1004         down_read(&cmd->root_lock);
1005         r = cmd->changed;
1006         up_read(&cmd->root_lock);
1007
1008         return r;
1009 }
1010
1011 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1012 {
1013         int r;
1014         unsigned flags;
1015         dm_oblock_t oblock;
1016         __le64 value;
1017
1018         r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1019         if (r)
1020                 return r;
1021
1022         unpack_value(value, &oblock, &flags);
1023
1024         if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1025                 /* nothing to be done */
1026                 return 0;
1027
1028         value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1029         __dm_bless_for_disk(&value);
1030
1031         r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1032                                &value, &cmd->root);
1033         if (r)
1034                 return r;
1035
1036         cmd->changed = true;
1037         return 0;
1038
1039 }
1040
1041 int dm_cache_set_dirty(struct dm_cache_metadata *cmd,
1042                        dm_cblock_t cblock, bool dirty)
1043 {
1044         int r;
1045
1046         down_write(&cmd->root_lock);
1047         r = __dirty(cmd, cblock, dirty);
1048         up_write(&cmd->root_lock);
1049
1050         return r;
1051 }
1052
1053 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1054                                  struct dm_cache_statistics *stats)
1055 {
1056         down_read(&cmd->root_lock);
1057         *stats = cmd->stats;
1058         up_read(&cmd->root_lock);
1059 }
1060
1061 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1062                                  struct dm_cache_statistics *stats)
1063 {
1064         down_write(&cmd->root_lock);
1065         cmd->stats = *stats;
1066         up_write(&cmd->root_lock);
1067 }
1068
1069 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1070 {
1071         int r;
1072         flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1073                                  clear_clean_shutdown);
1074
1075         down_write(&cmd->root_lock);
1076         r = __commit_transaction(cmd, mutator);
1077         if (r)
1078                 goto out;
1079
1080         r = __begin_transaction(cmd);
1081
1082 out:
1083         up_write(&cmd->root_lock);
1084         return r;
1085 }
1086
1087 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1088                                            dm_block_t *result)
1089 {
1090         int r = -EINVAL;
1091
1092         down_read(&cmd->root_lock);
1093         r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1094         up_read(&cmd->root_lock);
1095
1096         return r;
1097 }
1098
1099 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1100                                    dm_block_t *result)
1101 {
1102         int r = -EINVAL;
1103
1104         down_read(&cmd->root_lock);
1105         r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1106         up_read(&cmd->root_lock);
1107
1108         return r;
1109 }
1110
1111 /*----------------------------------------------------------------*/
1112
1113 static int begin_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1114 {
1115         int r;
1116         __le32 value;
1117         size_t hint_size;
1118         const char *policy_name = dm_cache_policy_get_name(policy);
1119         const unsigned *policy_version = dm_cache_policy_get_version(policy);
1120
1121         if (!policy_name[0] ||
1122             (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1123                 return -EINVAL;
1124
1125         if (!policy_unchanged(cmd, policy)) {
1126                 strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1127                 memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1128
1129                 hint_size = dm_cache_policy_get_hint_size(policy);
1130                 if (!hint_size)
1131                         return 0; /* short-circuit hints initialization */
1132                 cmd->policy_hint_size = hint_size;
1133
1134                 if (cmd->hint_root) {
1135                         r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1136                         if (r)
1137                                 return r;
1138                 }
1139
1140                 r = dm_array_empty(&cmd->hint_info, &cmd->hint_root);
1141                 if (r)
1142                         return r;
1143
1144                 value = cpu_to_le32(0);
1145                 __dm_bless_for_disk(&value);
1146                 r = dm_array_resize(&cmd->hint_info, cmd->hint_root, 0,
1147                                     from_cblock(cmd->cache_blocks),
1148                                     &value, &cmd->hint_root);
1149                 if (r)
1150                         return r;
1151         }
1152
1153         return 0;
1154 }
1155
1156 int dm_cache_begin_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1157 {
1158         int r;
1159
1160         down_write(&cmd->root_lock);
1161         r = begin_hints(cmd, policy);
1162         up_write(&cmd->root_lock);
1163
1164         return r;
1165 }
1166
1167 static int save_hint(struct dm_cache_metadata *cmd, dm_cblock_t cblock,
1168                      uint32_t hint)
1169 {
1170         int r;
1171         __le32 value = cpu_to_le32(hint);
1172         __dm_bless_for_disk(&value);
1173
1174         r = dm_array_set_value(&cmd->hint_info, cmd->hint_root,
1175                                from_cblock(cblock), &value, &cmd->hint_root);
1176         cmd->changed = true;
1177
1178         return r;
1179 }
1180
1181 int dm_cache_save_hint(struct dm_cache_metadata *cmd, dm_cblock_t cblock,
1182                        uint32_t hint)
1183 {
1184         int r;
1185
1186         if (!hints_array_initialized(cmd))
1187                 return 0;
1188
1189         down_write(&cmd->root_lock);
1190         r = save_hint(cmd, cblock, hint);
1191         up_write(&cmd->root_lock);
1192
1193         return r;
1194 }