drivers/md/dm-thin-metadata.c

   1 /*
   2  * Copyright (C) 2011-2012 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-thin-metadata.h"
   8 #include "persistent-data/dm-btree.h"
   9 #include "persistent-data/dm-space-map.h"
  10 #include "persistent-data/dm-space-map-disk.h"
  11 #include "persistent-data/dm-transaction-manager.h"
  12
  13 #include <linux/list.h>
  14 #include <linux/device-mapper.h>
  15 #include <linux/workqueue.h>
  16
  17 /*--------------------------------------------------------------------------
  18  * As far as the metadata goes, there is:
  19  *
  20  * - A superblock in block zero, taking up fewer than 512 bytes for
  21  *   atomic writes.
  22  *
  23  * - A space map managing the metadata blocks.
  24  *
  25  * - A space map managing the data blocks.
  26  *
  27  * - A btree mapping our internal thin dev ids onto struct disk_device_details.
  28  *
  29  * - A hierarchical btree, with 2 levels which effectively maps (thin
  30  *   dev id, virtual block) -> block_time.  Block time is a 64-bit
  31  *   field holding the time in the low 24 bits, and block in the top 48
  32  *   bits.
  33  *
  34  * BTrees consist solely of btree_nodes, that fill a block.  Some are
  35  * internal nodes, as such their values are a __le64 pointing to other
  36  * nodes.  Leaf nodes can store data of any reasonable size (ie. much
  37  * smaller than the block size).  The nodes consist of the header,
  38  * followed by an array of keys, followed by an array of values.  We have
  39  * to binary search on the keys so they're all held together to help the
  40  * cpu cache.
  41  *
  42  * Space maps have 2 btrees:
  43  *
  44  * - One maps a uint64_t onto a struct index_entry.  Which points to a
  45  *   bitmap block, and has some details about how many free entries there
  46  *   are etc.
  47  *
  48  * - The bitmap blocks have a header (for the checksum).  Then the rest
  49  *   of the block is pairs of bits.  With the meaning being:
  50  *
  51  *   0 - ref count is 0
  52  *   1 - ref count is 1
  53  *   2 - ref count is 2
  54  *   3 - ref count is higher than 2
  55  *
  56  * - If the count is higher than 2 then the ref count is entered in a
  57  *   second btree that directly maps the block_address to a uint32_t ref
  58  *   count.
  59  *
  60  * The space map metadata variant doesn't have a bitmaps btree.  Instead
  61  * it has one single blocks worth of index_entries.  This avoids
  62  * recursive issues with the bitmap btree needing to allocate space in
  63  * order to insert.  With a small data block size such as 64k the
  64  * metadata support data devices that are hundreds of terrabytes.
  65  *
  66  * The space maps allocate space linearly from front to back.  Space that
  67  * is freed in a transaction is never recycled within that transaction.
  68  * To try and avoid fragmenting _free_ space the allocator always goes
  69  * back and fills in gaps.
  70  *
  71  * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
  72  * from the block manager.
  73  *--------------------------------------------------------------------------*/
  74
  75 #define DM_MSG_PREFIX   "thin metadata"
  76
  77 #define THIN_SUPERBLOCK_MAGIC 27022010
  78 #define THIN_SUPERBLOCK_LOCATION 0
  79 #define THIN_VERSION 2
  80 #define SECTOR_TO_BLOCK_SHIFT 3
  81
  82 /*
  83  * For btree insert:
  84  *  3 for btree insert +
  85  *  2 for btree lookup used within space map
  86  * For btree remove:
  87  *  2 for shadow spine +
  88  *  4 for rebalance 3 child node
  89  */
  90 #define THIN_MAX_CONCURRENT_LOCKS 6
  91
  92 /* This should be plenty */
  93 #define SPACE_MAP_ROOT_SIZE 128
  94
  95 /*
  96  * Little endian on-disk superblock and device details.
  97  */
  98 struct thin_disk_superblock {
  99         __le32 csum;    /* Checksum of superblock except for this field. */
 100         __le32 flags;
 101         __le64 blocknr; /* This block number, dm_block_t. */
 102
 103         __u8 uuid[16];
 104         __le64 magic;
 105         __le32 version;
 106         __le32 time;
 107
 108         __le64 trans_id;
 109
 110         /*
 111          * Root held by userspace transactions.
 112          */
 113         __le64 held_root;
 114
 115         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 116         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 117
 118         /*
 119          * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 120          */
 121         __le64 data_mapping_root;
 122
 123         /*
 124          * Device detail root mapping dev_id -> device_details
 125          */
 126         __le64 device_details_root;
 127
 128         __le32 data_block_size;         /* In 512-byte sectors. */
 129
 130         __le32 metadata_block_size;     /* In 512-byte sectors. */
 131         __le64 metadata_nr_blocks;
 132
 133         __le32 compat_flags;
 134         __le32 compat_ro_flags;
 135         __le32 incompat_flags;
 136 } __packed;
 137
 138 struct disk_device_details {
 139         __le64 mapped_blocks;
 140         __le64 transaction_id;          /* When created. */
 141         __le32 creation_time;
 142         __le32 snapshotted_time;
 143 } __packed;
 144
 145 struct dm_pool_metadata {
 146         struct hlist_node hash;
 147
 148         struct block_device *bdev;
 149         struct dm_block_manager *bm;
 150         struct dm_space_map *metadata_sm;
 151         struct dm_space_map *data_sm;
 152         struct dm_transaction_manager *tm;
 153         struct dm_transaction_manager *nb_tm;
 154
 155         /*
 156          * Two-level btree.
 157          * First level holds thin_dev_t.
 158          * Second level holds mappings.
 159          */
 160         struct dm_btree_info info;
 161
 162         /*
 163          * Non-blocking version of the above.
 164          */
 165         struct dm_btree_info nb_info;
 166
 167         /*
 168          * Just the top level for deleting whole devices.
 169          */
 170         struct dm_btree_info tl_info;
 171
 172         /*
 173          * Just the bottom level for creating new devices.
 174          */
 175         struct dm_btree_info bl_info;
 176
 177         /*
 178          * Describes the device details btree.
 179          */
 180         struct dm_btree_info details_info;
 181
 182         struct rw_semaphore root_lock;
 183         uint32_t time;
 184         dm_block_t root;
 185         dm_block_t details_root;
 186         struct list_head thin_devices;
 187         uint64_t trans_id;
 188         unsigned long flags;
 189         sector_t data_block_size;
 190
 191         /*
 192          * Pre-commit callback.
 193          *
 194          * This allows the thin provisioning target to run a callback before
 195          * the metadata are committed.
 196          */
 197         dm_pool_pre_commit_fn pre_commit_fn;
 198         void *pre_commit_context;
 199
 200         /*
 201          * We reserve a section of the metadata for commit overhead.
 202          * All reported space does *not* include this.
 203          */
 204         dm_block_t metadata_reserve;
 205
 206         /*
 207          * Set if a transaction has to be aborted but the attempt to roll back
 208          * to the previous (good) transaction failed.  The only pool metadata
 209          * operation possible in this state is the closing of the device.
 210          */
 211         bool fail_io:1;
 212
 213         /*
 214          * Set once a thin-pool has been accessed through one of the interfaces
 215          * that imply the pool is in-service (e.g. thin devices created/deleted,
 216          * thin-pool message, metadata snapshots, etc).
 217          */
 218         bool in_service:1;
 219
 220         /*
 221          * Reading the space map roots can fail, so we read it into these
 222          * buffers before the superblock is locked and updated.
 223          */
 224         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 225         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 226 };
 227
 228 struct dm_thin_device {
 229         struct list_head list;
 230         struct dm_pool_metadata *pmd;
 231         dm_thin_id id;
 232
 233         int open_count;
 234         bool changed:1;
 235         bool aborted_with_changes:1;
 236         uint64_t mapped_blocks;
 237         uint64_t transaction_id;
 238         uint32_t creation_time;
 239         uint32_t snapshotted_time;
 240 };
 241
 242 /*----------------------------------------------------------------
 243  * superblock validator
 244  *--------------------------------------------------------------*/
 245
 246 #define SUPERBLOCK_CSUM_XOR 160774
 247
 248 static void sb_prepare_for_write(struct dm_block_validator *v,
 249                                  struct dm_block *b,
 250                                  size_t block_size)
 251 {
 252         struct thin_disk_superblock *disk_super = dm_block_data(b);
 253
 254         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 255         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 256                                                       block_size - sizeof(__le32),
 257                                                       SUPERBLOCK_CSUM_XOR));
 258 }
 259
 260 static int sb_check(struct dm_block_validator *v,
 261                     struct dm_block *b,
 262                     size_t block_size)
 263 {
 264         struct thin_disk_superblock *disk_super = dm_block_data(b);
 265         __le32 csum_le;
 266
 267         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 268                 DMERR("sb_check failed: blocknr %llu: "
 269                       "wanted %llu", le64_to_cpu(disk_super->blocknr),
 270                       (unsigned long long)dm_block_location(b));
 271                 return -ENOTBLK;
 272         }
 273
 274         if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 275                 DMERR("sb_check failed: magic %llu: "
 276                       "wanted %llu", le64_to_cpu(disk_super->magic),
 277                       (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 278                 return -EILSEQ;
 279         }
 280
 281         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 282                                              block_size - sizeof(__le32),
 283                                              SUPERBLOCK_CSUM_XOR));
 284         if (csum_le != disk_super->csum) {
 285                 DMERR("sb_check failed: csum %u: wanted %u",
 286                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 287                 return -EILSEQ;
 288         }
 289
 290         return 0;
 291 }
 292
 293 static struct dm_block_validator sb_validator = {
 294         .name = "superblock",
 295         .prepare_for_write = sb_prepare_for_write,
 296         .check = sb_check
 297 };
 298
 299 /*----------------------------------------------------------------
 300  * Methods for the btree value types
 301  *--------------------------------------------------------------*/
 302
 303 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 304 {
 305         return (b << 24) | t;
 306 }
 307
 308 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 309 {
 310         *b = v >> 24;
 311         *t = v & ((1 << 24) - 1);
 312 }
 313
 314 static void data_block_inc(void *context, const void *value_le)
 315 {
 316         struct dm_space_map *sm = context;
 317         __le64 v_le;
 318         uint64_t b;
 319         uint32_t t;
 320
 321         memcpy(&v_le, value_le, sizeof(v_le));
 322         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 323         dm_sm_inc_block(sm, b);
 324 }
 325
 326 static void data_block_dec(void *context, const void *value_le)
 327 {
 328         struct dm_space_map *sm = context;
 329         __le64 v_le;
 330         uint64_t b;
 331         uint32_t t;
 332
 333         memcpy(&v_le, value_le, sizeof(v_le));
 334         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 335         dm_sm_dec_block(sm, b);
 336 }
 337
 338 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 339 {
 340         __le64 v1_le, v2_le;
 341         uint64_t b1, b2;
 342         uint32_t t;
 343
 344         memcpy(&v1_le, value1_le, sizeof(v1_le));
 345         memcpy(&v2_le, value2_le, sizeof(v2_le));
 346         unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 347         unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 348
 349         return b1 == b2;
 350 }
 351
 352 static void subtree_inc(void *context, const void *value)
 353 {
 354         struct dm_btree_info *info = context;
 355         __le64 root_le;
 356         uint64_t root;
 357
 358         memcpy(&root_le, value, sizeof(root_le));
 359         root = le64_to_cpu(root_le);
 360         dm_tm_inc(info->tm, root);
 361 }
 362
 363 static void subtree_dec(void *context, const void *value)
 364 {
 365         struct dm_btree_info *info = context;
 366         __le64 root_le;
 367         uint64_t root;
 368
 369         memcpy(&root_le, value, sizeof(root_le));
 370         root = le64_to_cpu(root_le);
 371         if (dm_btree_del(info, root))
 372                 DMERR("btree delete failed");
 373 }
 374
 375 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 376 {
 377         __le64 v1_le, v2_le;
 378         memcpy(&v1_le, value1_le, sizeof(v1_le));
 379         memcpy(&v2_le, value2_le, sizeof(v2_le));
 380
 381         return v1_le == v2_le;
 382 }
 383
 384 /*----------------------------------------------------------------*/
 385
 386 /*
 387  * Variant that is used for in-core only changes or code that
 388  * shouldn't put the pool in service on its own (e.g. commit).
 389  */
 390 static inline void pmd_write_lock_in_core(struct dm_pool_metadata *pmd)
 391         __acquires(pmd->root_lock)
 392 {
 393         down_write(&pmd->root_lock);
 394 }
 395
 396 static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
 397 {
 398         pmd_write_lock_in_core(pmd);
 399         if (unlikely(!pmd->in_service))
 400                 pmd->in_service = true;
 401 }
 402
 403 static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
 404         __releases(pmd->root_lock)
 405 {
 406         up_write(&pmd->root_lock);
 407 }
 408
 409 /*----------------------------------------------------------------*/
 410
 411 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 412                                 struct dm_block **sblock)
 413 {
 414         return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 415                                      &sb_validator, sblock);
 416 }
 417
 418 static int superblock_lock(struct dm_pool_metadata *pmd,
 419                            struct dm_block **sblock)
 420 {
 421         return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 422                                 &sb_validator, sblock);
 423 }
 424
 425 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 426 {
 427         int r;
 428         unsigned i;
 429         struct dm_block *b;
 430         __le64 *data_le, zero = cpu_to_le64(0);
 431         unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 432
 433         /*
 434          * We can't use a validator here - it may be all zeroes.
 435          */
 436         r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 437         if (r)
 438                 return r;
 439
 440         data_le = dm_block_data(b);
 441         *result = 1;
 442         for (i = 0; i < block_size; i++) {
 443                 if (data_le[i] != zero) {
 444                         *result = 0;
 445                         break;
 446                 }
 447         }
 448
 449         dm_bm_unlock(b);
 450
 451         return 0;
 452 }
 453
 454 static void __setup_btree_details(struct dm_pool_metadata *pmd)
 455 {
 456         pmd->info.tm = pmd->tm;
 457         pmd->info.levels = 2;
 458         pmd->info.value_type.context = pmd->data_sm;
 459         pmd->info.value_type.size = sizeof(__le64);
 460         pmd->info.value_type.inc = data_block_inc;
 461         pmd->info.value_type.dec = data_block_dec;
 462         pmd->info.value_type.equal = data_block_equal;
 463
 464         memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 465         pmd->nb_info.tm = pmd->nb_tm;
 466
 467         pmd->tl_info.tm = pmd->tm;
 468         pmd->tl_info.levels = 1;
 469         pmd->tl_info.value_type.context = &pmd->bl_info;
 470         pmd->tl_info.value_type.size = sizeof(__le64);
 471         pmd->tl_info.value_type.inc = subtree_inc;
 472         pmd->tl_info.value_type.dec = subtree_dec;
 473         pmd->tl_info.value_type.equal = subtree_equal;
 474
 475         pmd->bl_info.tm = pmd->tm;
 476         pmd->bl_info.levels = 1;
 477         pmd->bl_info.value_type.context = pmd->data_sm;
 478         pmd->bl_info.value_type.size = sizeof(__le64);
 479         pmd->bl_info.value_type.inc = data_block_inc;
 480         pmd->bl_info.value_type.dec = data_block_dec;
 481         pmd->bl_info.value_type.equal = data_block_equal;
 482
 483         pmd->details_info.tm = pmd->tm;
 484         pmd->details_info.levels = 1;
 485         pmd->details_info.value_type.context = NULL;
 486         pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 487         pmd->details_info.value_type.inc = NULL;
 488         pmd->details_info.value_type.dec = NULL;
 489         pmd->details_info.value_type.equal = NULL;
 490 }
 491
 492 static int save_sm_roots(struct dm_pool_metadata *pmd)
 493 {
 494         int r;
 495         size_t len;
 496
 497         r = dm_sm_root_size(pmd->metadata_sm, &len);
 498         if (r < 0)
 499                 return r;
 500
 501         r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 502         if (r < 0)
 503                 return r;
 504
 505         r = dm_sm_root_size(pmd->data_sm, &len);
 506         if (r < 0)
 507                 return r;
 508
 509         return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 510 }
 511
 512 static void copy_sm_roots(struct dm_pool_metadata *pmd,
 513                           struct thin_disk_superblock *disk)
 514 {
 515         memcpy(&disk->metadata_space_map_root,
 516                &pmd->metadata_space_map_root,
 517                sizeof(pmd->metadata_space_map_root));
 518
 519         memcpy(&disk->data_space_map_root,
 520                &pmd->data_space_map_root,
 521                sizeof(pmd->data_space_map_root));
 522 }
 523
 524 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 525 {
 526         int r;
 527         struct dm_block *sblock;
 528         struct thin_disk_superblock *disk_super;
 529         sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 530
 531         if (bdev_size > THIN_METADATA_MAX_SECTORS)
 532                 bdev_size = THIN_METADATA_MAX_SECTORS;
 533
 534         r = dm_sm_commit(pmd->data_sm);
 535         if (r < 0)
 536                 return r;
 537
 538         r = dm_tm_pre_commit(pmd->tm);
 539         if (r < 0)
 540                 return r;
 541
 542         r = save_sm_roots(pmd);
 543         if (r < 0)
 544                 return r;
 545
 546         r = superblock_lock_zero(pmd, &sblock);
 547         if (r)
 548                 return r;
 549
 550         disk_super = dm_block_data(sblock);
 551         disk_super->flags = 0;
 552         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 553         disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 554         disk_super->version = cpu_to_le32(THIN_VERSION);
 555         disk_super->time = 0;
 556         disk_super->trans_id = 0;
 557         disk_super->held_root = 0;
 558
 559         copy_sm_roots(pmd, disk_super);
 560
 561         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 562         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 563         disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 564         disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 565         disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 566
 567         return dm_tm_commit(pmd->tm, sblock);
 568 }
 569
 570 static int __format_metadata(struct dm_pool_metadata *pmd)
 571 {
 572         int r;
 573
 574         r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 575                                  &pmd->tm, &pmd->metadata_sm);
 576         if (r < 0) {
 577                 DMERR("tm_create_with_sm failed");
 578                 return r;
 579         }
 580
 581         pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 582         if (IS_ERR(pmd->data_sm)) {
 583                 DMERR("sm_disk_create failed");
 584                 r = PTR_ERR(pmd->data_sm);
 585                 goto bad_cleanup_tm;
 586         }
 587
 588         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 589         if (!pmd->nb_tm) {
 590                 DMERR("could not create non-blocking clone tm");
 591                 r = -ENOMEM;
 592                 goto bad_cleanup_data_sm;
 593         }
 594
 595         __setup_btree_details(pmd);
 596
 597         r = dm_btree_empty(&pmd->info, &pmd->root);
 598         if (r < 0)
 599                 goto bad_cleanup_nb_tm;
 600
 601         r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 602         if (r < 0) {
 603                 DMERR("couldn't create devices root");
 604                 goto bad_cleanup_nb_tm;
 605         }
 606
 607         r = __write_initial_superblock(pmd);
 608         if (r)
 609                 goto bad_cleanup_nb_tm;
 610
 611         return 0;
 612
 613 bad_cleanup_nb_tm:
 614         dm_tm_destroy(pmd->nb_tm);
 615 bad_cleanup_data_sm:
 616         dm_sm_destroy(pmd->data_sm);
 617 bad_cleanup_tm:
 618         dm_tm_destroy(pmd->tm);
 619         dm_sm_destroy(pmd->metadata_sm);
 620
 621         return r;
 622 }
 623
 624 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 625                                      struct dm_pool_metadata *pmd)
 626 {
 627         uint32_t features;
 628
 629         features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 630         if (features) {
 631                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 632                       (unsigned long)features);
 633                 return -EINVAL;
 634         }
 635
 636         /*
 637          * Check for read-only metadata to skip the following RDWR checks.
 638          */
 639         if (get_disk_ro(pmd->bdev->bd_disk))
 640                 return 0;
 641
 642         features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 643         if (features) {
 644                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 645                       (unsigned long)features);
 646                 return -EINVAL;
 647         }
 648
 649         return 0;
 650 }
 651
 652 static int __open_metadata(struct dm_pool_metadata *pmd)
 653 {
 654         int r;
 655         struct dm_block *sblock;
 656         struct thin_disk_superblock *disk_super;
 657
 658         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 659                             &sb_validator, &sblock);
 660         if (r < 0) {
 661                 DMERR("couldn't read superblock");
 662                 return r;
 663         }
 664
 665         disk_super = dm_block_data(sblock);
 666
 667         /* Verify the data block size hasn't changed */
 668         if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 669                 DMERR("changing the data block size (from %u to %llu) is not supported",
 670                       le32_to_cpu(disk_super->data_block_size),
 671                       (unsigned long long)pmd->data_block_size);
 672                 r = -EINVAL;
 673                 goto bad_unlock_sblock;
 674         }
 675
 676         r = __check_incompat_features(disk_super, pmd);
 677         if (r < 0)
 678                 goto bad_unlock_sblock;
 679
 680         r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 681                                disk_super->metadata_space_map_root,
 682                                sizeof(disk_super->metadata_space_map_root),
 683                                &pmd->tm, &pmd->metadata_sm);
 684         if (r < 0) {
 685                 DMERR("tm_open_with_sm failed");
 686                 goto bad_unlock_sblock;
 687         }
 688
 689         pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 690                                        sizeof(disk_super->data_space_map_root));
 691         if (IS_ERR(pmd->data_sm)) {
 692                 DMERR("sm_disk_open failed");
 693                 r = PTR_ERR(pmd->data_sm);
 694                 goto bad_cleanup_tm;
 695         }
 696
 697         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 698         if (!pmd->nb_tm) {
 699                 DMERR("could not create non-blocking clone tm");
 700                 r = -ENOMEM;
 701                 goto bad_cleanup_data_sm;
 702         }
 703
 704         /*
 705          * For pool metadata opening process, root setting is redundant
 706          * because it will be set again in __begin_transaction(). But dm
 707          * pool aborting process really needs to get last transaction's
 708          * root to avoid accessing broken btree.
 709          */
 710         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 711         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 712
 713         __setup_btree_details(pmd);
 714         dm_bm_unlock(sblock);
 715
 716         return 0;
 717
 718 bad_cleanup_data_sm:
 719         dm_sm_destroy(pmd->data_sm);
 720 bad_cleanup_tm:
 721         dm_tm_destroy(pmd->tm);
 722         dm_sm_destroy(pmd->metadata_sm);
 723 bad_unlock_sblock:
 724         dm_bm_unlock(sblock);
 725
 726         return r;
 727 }
 728
 729 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 730 {
 731         int r, unformatted;
 732
 733         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 734         if (r)
 735                 return r;
 736
 737         if (unformatted)
 738                 return format_device ? __format_metadata(pmd) : -EPERM;
 739
 740         return __open_metadata(pmd);
 741 }
 742
 743 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 744 {
 745         int r;
 746
 747         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 748                                           THIN_MAX_CONCURRENT_LOCKS);
 749         if (IS_ERR(pmd->bm)) {
 750                 DMERR("could not create block manager");
 751                 r = PTR_ERR(pmd->bm);
 752                 pmd->bm = NULL;
 753                 return r;
 754         }
 755
 756         r = __open_or_format_metadata(pmd, format_device);
 757         if (r) {
 758                 dm_block_manager_destroy(pmd->bm);
 759                 pmd->bm = NULL;
 760         }
 761
 762         return r;
 763 }
 764
 765 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd,
 766                                               bool destroy_bm)
 767 {
 768         dm_sm_destroy(pmd->data_sm);
 769         dm_sm_destroy(pmd->metadata_sm);
 770         dm_tm_destroy(pmd->nb_tm);
 771         dm_tm_destroy(pmd->tm);
 772         if (destroy_bm)
 773                 dm_block_manager_destroy(pmd->bm);
 774 }
 775
 776 static int __begin_transaction(struct dm_pool_metadata *pmd)
 777 {
 778         int r;
 779         struct thin_disk_superblock *disk_super;
 780         struct dm_block *sblock;
 781
 782         /*
 783          * We re-read the superblock every time.  Shouldn't need to do this
 784          * really.
 785          */
 786         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 787                             &sb_validator, &sblock);
 788         if (r)
 789                 return r;
 790
 791         disk_super = dm_block_data(sblock);
 792         pmd->time = le32_to_cpu(disk_super->time);
 793         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 794         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 795         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 796         pmd->flags = le32_to_cpu(disk_super->flags);
 797         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 798
 799         dm_bm_unlock(sblock);
 800         return 0;
 801 }
 802
 803 static int __write_changed_details(struct dm_pool_metadata *pmd)
 804 {
 805         int r;
 806         struct dm_thin_device *td, *tmp;
 807         struct disk_device_details details;
 808         uint64_t key;
 809
 810         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 811                 if (!td->changed)
 812                         continue;
 813
 814                 key = td->id;
 815
 816                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 817                 details.transaction_id = cpu_to_le64(td->transaction_id);
 818                 details.creation_time = cpu_to_le32(td->creation_time);
 819                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 820                 __dm_bless_for_disk(&details);
 821
 822                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 823                                     &key, &details, &pmd->details_root);
 824                 if (r)
 825                         return r;
 826
 827                 if (td->open_count)
 828                         td->changed = 0;
 829                 else {
 830                         list_del(&td->list);
 831                         kfree(td);
 832                 }
 833         }
 834
 835         return 0;
 836 }
 837
 838 static int __commit_transaction(struct dm_pool_metadata *pmd)
 839 {
 840         int r;
 841         struct thin_disk_superblock *disk_super;
 842         struct dm_block *sblock;
 843
 844         /*
 845          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 846          */
 847         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 848         BUG_ON(!rwsem_is_locked(&pmd->root_lock));
 849
 850         if (unlikely(!pmd->in_service))
 851                 return 0;
 852
 853         if (pmd->pre_commit_fn) {
 854                 r = pmd->pre_commit_fn(pmd->pre_commit_context);
 855                 if (r < 0) {
 856                         DMERR("pre-commit callback failed");
 857                         return r;
 858                 }
 859         }
 860
 861         r = __write_changed_details(pmd);
 862         if (r < 0)
 863                 return r;
 864
 865         r = dm_sm_commit(pmd->data_sm);
 866         if (r < 0)
 867                 return r;
 868
 869         r = dm_tm_pre_commit(pmd->tm);
 870         if (r < 0)
 871                 return r;
 872
 873         r = save_sm_roots(pmd);
 874         if (r < 0)
 875                 return r;
 876
 877         r = superblock_lock(pmd, &sblock);
 878         if (r)
 879                 return r;
 880
 881         disk_super = dm_block_data(sblock);
 882         disk_super->time = cpu_to_le32(pmd->time);
 883         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 884         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 885         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 886         disk_super->flags = cpu_to_le32(pmd->flags);
 887
 888         copy_sm_roots(pmd, disk_super);
 889
 890         return dm_tm_commit(pmd->tm, sblock);
 891 }
 892
 893 static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 894 {
 895         int r;
 896         dm_block_t total;
 897         dm_block_t max_blocks = 4096; /* 16M */
 898
 899         r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 900         if (r) {
 901                 DMERR("could not get size of metadata device");
 902                 pmd->metadata_reserve = max_blocks;
 903         } else
 904                 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 905 }
 906
 907 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 908                                                sector_t data_block_size,
 909                                                bool format_device)
 910 {
 911         int r;
 912         struct dm_pool_metadata *pmd;
 913
 914         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 915         if (!pmd) {
 916                 DMERR("could not allocate metadata struct");
 917                 return ERR_PTR(-ENOMEM);
 918         }
 919
 920         init_rwsem(&pmd->root_lock);
 921         pmd->time = 0;
 922         INIT_LIST_HEAD(&pmd->thin_devices);
 923         pmd->fail_io = false;
 924         pmd->in_service = false;
 925         pmd->bdev = bdev;
 926         pmd->data_block_size = data_block_size;
 927         pmd->pre_commit_fn = NULL;
 928         pmd->pre_commit_context = NULL;
 929
 930         r = __create_persistent_data_objects(pmd, format_device);
 931         if (r) {
 932                 kfree(pmd);
 933                 return ERR_PTR(r);
 934         }
 935
 936         r = __begin_transaction(pmd);
 937         if (r < 0) {
 938                 if (dm_pool_metadata_close(pmd) < 0)
 939                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 940                 return ERR_PTR(r);
 941         }
 942
 943         __set_metadata_reserve(pmd);
 944
 945         return pmd;
 946 }
 947
 948 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 949 {
 950         int r;
 951         unsigned open_devices = 0;
 952         struct dm_thin_device *td, *tmp;
 953
 954         down_read(&pmd->root_lock);
 955         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 956                 if (td->open_count)
 957                         open_devices++;
 958                 else {
 959                         list_del(&td->list);
 960                         kfree(td);
 961                 }
 962         }
 963         up_read(&pmd->root_lock);
 964
 965         if (open_devices) {
 966                 DMERR("attempt to close pmd when %u device(s) are still open",
 967                        open_devices);
 968                 return -EBUSY;
 969         }
 970
 971         pmd_write_lock_in_core(pmd);
 972         if (!pmd->fail_io && !dm_bm_is_read_only(pmd->bm)) {
 973                 r = __commit_transaction(pmd);
 974                 if (r < 0)
 975                         DMWARN("%s: __commit_transaction() failed, error = %d",
 976                                __func__, r);
 977         }
 978         pmd_write_unlock(pmd);
 979         if (!pmd->fail_io)
 980                 __destroy_persistent_data_objects(pmd, true);
 981
 982         kfree(pmd);
 983         return 0;
 984 }
 985
 986 /*
 987  * __open_device: Returns @td corresponding to device with id @dev,
 988  * creating it if @create is set and incrementing @td->open_count.
 989  * On failure, @td is undefined.
 990  */
 991 static int __open_device(struct dm_pool_metadata *pmd,
 992                          dm_thin_id dev, int create,
 993                          struct dm_thin_device **td)
 994 {
 995         int r, changed = 0;
 996         struct dm_thin_device *td2;
 997         uint64_t key = dev;
 998         struct disk_device_details details_le;
 999
1000         /*
1001          * If the device is already open, return it.
1002          */
1003         list_for_each_entry(td2, &pmd->thin_devices, list)
1004                 if (td2->id == dev) {
1005                         /*
1006                          * May not create an already-open device.
1007                          */
1008                         if (create)
1009                                 return -EEXIST;
1010
1011                         td2->open_count++;
1012                         *td = td2;
1013                         return 0;
1014                 }
1015
1016         /*
1017          * Check the device exists.
1018          */
1019         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1020                             &key, &details_le);
1021         if (r) {
1022                 if (r != -ENODATA || !create)
1023                         return r;
1024
1025                 /*
1026                  * Create new device.
1027                  */
1028                 changed = 1;
1029                 details_le.mapped_blocks = 0;
1030                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
1031                 details_le.creation_time = cpu_to_le32(pmd->time);
1032                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
1033         }
1034
1035         *td = kmalloc(sizeof(**td), GFP_NOIO);
1036         if (!*td)
1037                 return -ENOMEM;
1038
1039         (*td)->pmd = pmd;
1040         (*td)->id = dev;
1041         (*td)->open_count = 1;
1042         (*td)->changed = changed;
1043         (*td)->aborted_with_changes = false;
1044         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1045         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1046         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
1047         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1048
1049         list_add(&(*td)->list, &pmd->thin_devices);
1050
1051         return 0;
1052 }
1053
1054 static void __close_device(struct dm_thin_device *td)
1055 {
1056         --td->open_count;
1057 }
1058
1059 static int __create_thin(struct dm_pool_metadata *pmd,
1060                          dm_thin_id dev)
1061 {
1062         int r;
1063         dm_block_t dev_root;
1064         uint64_t key = dev;
1065         struct disk_device_details details_le;
1066         struct dm_thin_device *td;
1067         __le64 value;
1068
1069         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1070                             &key, &details_le);
1071         if (!r)
1072                 return -EEXIST;
1073
1074         /*
1075          * Create an empty btree for the mappings.
1076          */
1077         r = dm_btree_empty(&pmd->bl_info, &dev_root);
1078         if (r)
1079                 return r;
1080
1081         /*
1082          * Insert it into the main mapping tree.
1083          */
1084         value = cpu_to_le64(dev_root);
1085         __dm_bless_for_disk(&value);
1086         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1087         if (r) {
1088                 dm_btree_del(&pmd->bl_info, dev_root);
1089                 return r;
1090         }
1091
1092         r = __open_device(pmd, dev, 1, &td);
1093         if (r) {
1094                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1095                 dm_btree_del(&pmd->bl_info, dev_root);
1096                 return r;
1097         }
1098         __close_device(td);
1099
1100         return r;
1101 }
1102
1103 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1104 {
1105         int r = -EINVAL;
1106
1107         pmd_write_lock(pmd);
1108         if (!pmd->fail_io)
1109                 r = __create_thin(pmd, dev);
1110         pmd_write_unlock(pmd);
1111
1112         return r;
1113 }
1114
1115 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1116                                   struct dm_thin_device *snap,
1117                                   dm_thin_id origin, uint32_t time)
1118 {
1119         int r;
1120         struct dm_thin_device *td;
1121
1122         r = __open_device(pmd, origin, 0, &td);
1123         if (r)
1124                 return r;
1125
1126         td->changed = 1;
1127         td->snapshotted_time = time;
1128
1129         snap->mapped_blocks = td->mapped_blocks;
1130         snap->snapshotted_time = time;
1131         __close_device(td);
1132
1133         return 0;
1134 }
1135
1136 static int __create_snap(struct dm_pool_metadata *pmd,
1137                          dm_thin_id dev, dm_thin_id origin)
1138 {
1139         int r;
1140         dm_block_t origin_root;
1141         uint64_t key = origin, dev_key = dev;
1142         struct dm_thin_device *td;
1143         struct disk_device_details details_le;
1144         __le64 value;
1145
1146         /* check this device is unused */
1147         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1148                             &dev_key, &details_le);
1149         if (!r)
1150                 return -EEXIST;
1151
1152         /* find the mapping tree for the origin */
1153         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1154         if (r)
1155                 return r;
1156         origin_root = le64_to_cpu(value);
1157
1158         /* clone the origin, an inc will do */
1159         dm_tm_inc(pmd->tm, origin_root);
1160
1161         /* insert into the main mapping tree */
1162         value = cpu_to_le64(origin_root);
1163         __dm_bless_for_disk(&value);
1164         key = dev;
1165         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1166         if (r) {
1167                 dm_tm_dec(pmd->tm, origin_root);
1168                 return r;
1169         }
1170
1171         pmd->time++;
1172
1173         r = __open_device(pmd, dev, 1, &td);
1174         if (r)
1175                 goto bad;
1176
1177         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1178         __close_device(td);
1179
1180         if (r)
1181                 goto bad;
1182
1183         return 0;
1184
1185 bad:
1186         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1187         dm_btree_remove(&pmd->details_info, pmd->details_root,
1188                         &key, &pmd->details_root);
1189         return r;
1190 }
1191
1192 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1193                                  dm_thin_id dev,
1194                                  dm_thin_id origin)
1195 {
1196         int r = -EINVAL;
1197
1198         pmd_write_lock(pmd);
1199         if (!pmd->fail_io)
1200                 r = __create_snap(pmd, dev, origin);
1201         pmd_write_unlock(pmd);
1202
1203         return r;
1204 }
1205
1206 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1207 {
1208         int r;
1209         uint64_t key = dev;
1210         struct dm_thin_device *td;
1211
1212         /* TODO: failure should mark the transaction invalid */
1213         r = __open_device(pmd, dev, 0, &td);
1214         if (r)
1215                 return r;
1216
1217         if (td->open_count > 1) {
1218                 __close_device(td);
1219                 return -EBUSY;
1220         }
1221
1222         list_del(&td->list);
1223         kfree(td);
1224         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1225                             &key, &pmd->details_root);
1226         if (r)
1227                 return r;
1228
1229         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1230         if (r)
1231                 return r;
1232
1233         return 0;
1234 }
1235
1236 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1237                                dm_thin_id dev)
1238 {
1239         int r = -EINVAL;
1240
1241         pmd_write_lock(pmd);
1242         if (!pmd->fail_io)
1243                 r = __delete_device(pmd, dev);
1244         pmd_write_unlock(pmd);
1245
1246         return r;
1247 }
1248
1249 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1250                                         uint64_t current_id,
1251                                         uint64_t new_id)
1252 {
1253         int r = -EINVAL;
1254
1255         pmd_write_lock(pmd);
1256
1257         if (pmd->fail_io)
1258                 goto out;
1259
1260         if (pmd->trans_id != current_id) {
1261                 DMERR("mismatched transaction id");
1262                 goto out;
1263         }
1264
1265         pmd->trans_id = new_id;
1266         r = 0;
1267
1268 out:
1269         pmd_write_unlock(pmd);
1270
1271         return r;
1272 }
1273
1274 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1275                                         uint64_t *result)
1276 {
1277         int r = -EINVAL;
1278
1279         down_read(&pmd->root_lock);
1280         if (!pmd->fail_io) {
1281                 *result = pmd->trans_id;
1282                 r = 0;
1283         }
1284         up_read(&pmd->root_lock);
1285
1286         return r;
1287 }
1288
1289 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1290 {
1291         int r, inc;
1292         struct thin_disk_superblock *disk_super;
1293         struct dm_block *copy, *sblock;
1294         dm_block_t held_root;
1295
1296         /*
1297          * We commit to ensure the btree roots which we increment in a
1298          * moment are up to date.
1299          */
1300         r = __commit_transaction(pmd);
1301         if (r < 0) {
1302                 DMWARN("%s: __commit_transaction() failed, error = %d",
1303                        __func__, r);
1304                 return r;
1305         }
1306
1307         /*
1308          * Copy the superblock.
1309          */
1310         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1311         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1312                                &sb_validator, &copy, &inc);
1313         if (r)
1314                 return r;
1315
1316         BUG_ON(!inc);
1317
1318         held_root = dm_block_location(copy);
1319         disk_super = dm_block_data(copy);
1320
1321         if (le64_to_cpu(disk_super->held_root)) {
1322                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1323
1324                 dm_tm_dec(pmd->tm, held_root);
1325                 dm_tm_unlock(pmd->tm, copy);
1326                 return -EBUSY;
1327         }
1328
1329         /*
1330          * Wipe the spacemap since we're not publishing this.
1331          */
1332         memset(&disk_super->data_space_map_root, 0,
1333                sizeof(disk_super->data_space_map_root));
1334         memset(&disk_super->metadata_space_map_root, 0,
1335                sizeof(disk_super->metadata_space_map_root));
1336
1337         /*
1338          * Increment the data structures that need to be preserved.
1339          */
1340         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1341         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1342         dm_tm_unlock(pmd->tm, copy);
1343
1344         /*
1345          * Write the held root into the superblock.
1346          */
1347         r = superblock_lock(pmd, &sblock);
1348         if (r) {
1349                 dm_tm_dec(pmd->tm, held_root);
1350                 return r;
1351         }
1352
1353         disk_super = dm_block_data(sblock);
1354         disk_super->held_root = cpu_to_le64(held_root);
1355         dm_bm_unlock(sblock);
1356         return 0;
1357 }
1358
1359 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1360 {
1361         int r = -EINVAL;
1362
1363         pmd_write_lock(pmd);
1364         if (!pmd->fail_io)
1365                 r = __reserve_metadata_snap(pmd);
1366         pmd_write_unlock(pmd);
1367
1368         return r;
1369 }
1370
1371 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1372 {
1373         int r;
1374         struct thin_disk_superblock *disk_super;
1375         struct dm_block *sblock, *copy;
1376         dm_block_t held_root;
1377
1378         r = superblock_lock(pmd, &sblock);
1379         if (r)
1380                 return r;
1381
1382         disk_super = dm_block_data(sblock);
1383         held_root = le64_to_cpu(disk_super->held_root);
1384         disk_super->held_root = cpu_to_le64(0);
1385
1386         dm_bm_unlock(sblock);
1387
1388         if (!held_root) {
1389                 DMWARN("No pool metadata snapshot found: nothing to release.");
1390                 return -EINVAL;
1391         }
1392
1393         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1394         if (r)
1395                 return r;
1396
1397         disk_super = dm_block_data(copy);
1398         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1399         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1400         dm_sm_dec_block(pmd->metadata_sm, held_root);
1401
1402         dm_tm_unlock(pmd->tm, copy);
1403
1404         return 0;
1405 }
1406
1407 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1408 {
1409         int r = -EINVAL;
1410
1411         pmd_write_lock(pmd);
1412         if (!pmd->fail_io)
1413                 r = __release_metadata_snap(pmd);
1414         pmd_write_unlock(pmd);
1415
1416         return r;
1417 }
1418
1419 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1420                                dm_block_t *result)
1421 {
1422         int r;
1423         struct thin_disk_superblock *disk_super;
1424         struct dm_block *sblock;
1425
1426         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1427                             &sb_validator, &sblock);
1428         if (r)
1429                 return r;
1430
1431         disk_super = dm_block_data(sblock);
1432         *result = le64_to_cpu(disk_super->held_root);
1433
1434         dm_bm_unlock(sblock);
1435
1436         return 0;
1437 }
1438
1439 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1440                               dm_block_t *result)
1441 {
1442         int r = -EINVAL;
1443
1444         down_read(&pmd->root_lock);
1445         if (!pmd->fail_io)
1446                 r = __get_metadata_snap(pmd, result);
1447         up_read(&pmd->root_lock);
1448
1449         return r;
1450 }
1451
1452 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1453                              struct dm_thin_device **td)
1454 {
1455         int r = -EINVAL;
1456
1457         pmd_write_lock_in_core(pmd);
1458         if (!pmd->fail_io)
1459                 r = __open_device(pmd, dev, 0, td);
1460         pmd_write_unlock(pmd);
1461
1462         return r;
1463 }
1464
1465 int dm_pool_close_thin_device(struct dm_thin_device *td)
1466 {
1467         pmd_write_lock_in_core(td->pmd);
1468         __close_device(td);
1469         pmd_write_unlock(td->pmd);
1470
1471         return 0;
1472 }
1473
1474 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1475 {
1476         return td->id;
1477 }
1478
1479 /*
1480  * Check whether @time (of block creation) is older than @td's last snapshot.
1481  * If so then the associated block is shared with the last snapshot device.
1482  * Any block on a device created *after* the device last got snapshotted is
1483  * necessarily not shared.
1484  */
1485 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1486 {
1487         return td->snapshotted_time > time;
1488 }
1489
1490 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1491                                  struct dm_thin_lookup_result *result)
1492 {
1493         uint64_t block_time = 0;
1494         dm_block_t exception_block;
1495         uint32_t exception_time;
1496
1497         block_time = le64_to_cpu(value);
1498         unpack_block_time(block_time, &exception_block, &exception_time);
1499         result->block = exception_block;
1500         result->shared = __snapshotted_since(td, exception_time);
1501 }
1502
1503 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1504                         int can_issue_io, struct dm_thin_lookup_result *result)
1505 {
1506         int r;
1507         __le64 value;
1508         struct dm_pool_metadata *pmd = td->pmd;
1509         dm_block_t keys[2] = { td->id, block };
1510         struct dm_btree_info *info;
1511
1512         if (can_issue_io) {
1513                 info = &pmd->info;
1514         } else
1515                 info = &pmd->nb_info;
1516
1517         r = dm_btree_lookup(info, pmd->root, keys, &value);
1518         if (!r)
1519                 unpack_lookup_result(td, value, result);
1520
1521         return r;
1522 }
1523
1524 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1525                        int can_issue_io, struct dm_thin_lookup_result *result)
1526 {
1527         int r;
1528         struct dm_pool_metadata *pmd = td->pmd;
1529
1530         down_read(&pmd->root_lock);
1531         if (pmd->fail_io) {
1532                 up_read(&pmd->root_lock);
1533                 return -EINVAL;
1534         }
1535
1536         r = __find_block(td, block, can_issue_io, result);
1537
1538         up_read(&pmd->root_lock);
1539         return r;
1540 }
1541
1542 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1543                                           dm_block_t *vblock,
1544                                           struct dm_thin_lookup_result *result)
1545 {
1546         int r;
1547         __le64 value;
1548         struct dm_pool_metadata *pmd = td->pmd;
1549         dm_block_t keys[2] = { td->id, block };
1550
1551         r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1552         if (!r)
1553                 unpack_lookup_result(td, value, result);
1554
1555         return r;
1556 }
1557
1558 static int __find_mapped_range(struct dm_thin_device *td,
1559                                dm_block_t begin, dm_block_t end,
1560                                dm_block_t *thin_begin, dm_block_t *thin_end,
1561                                dm_block_t *pool_begin, bool *maybe_shared)
1562 {
1563         int r;
1564         dm_block_t pool_end;
1565         struct dm_thin_lookup_result lookup;
1566
1567         if (end < begin)
1568                 return -ENODATA;
1569
1570         r = __find_next_mapped_block(td, begin, &begin, &lookup);
1571         if (r)
1572                 return r;
1573
1574         if (begin >= end)
1575                 return -ENODATA;
1576
1577         *thin_begin = begin;
1578         *pool_begin = lookup.block;
1579         *maybe_shared = lookup.shared;
1580
1581         begin++;
1582         pool_end = *pool_begin + 1;
1583         while (begin != end) {
1584                 r = __find_block(td, begin, true, &lookup);
1585                 if (r) {
1586                         if (r == -ENODATA)
1587                                 break;
1588                         else
1589                                 return r;
1590                 }
1591
1592                 if ((lookup.block != pool_end) ||
1593                     (lookup.shared != *maybe_shared))
1594                         break;
1595
1596                 pool_end++;
1597                 begin++;
1598         }
1599
1600         *thin_end = begin;
1601         return 0;
1602 }
1603
1604 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1605                               dm_block_t begin, dm_block_t end,
1606                               dm_block_t *thin_begin, dm_block_t *thin_end,
1607                               dm_block_t *pool_begin, bool *maybe_shared)
1608 {
1609         int r = -EINVAL;
1610         struct dm_pool_metadata *pmd = td->pmd;
1611
1612         down_read(&pmd->root_lock);
1613         if (!pmd->fail_io) {
1614                 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1615                                         pool_begin, maybe_shared);
1616         }
1617         up_read(&pmd->root_lock);
1618
1619         return r;
1620 }
1621
1622 static int __insert(struct dm_thin_device *td, dm_block_t block,
1623                     dm_block_t data_block)
1624 {
1625         int r, inserted;
1626         __le64 value;
1627         struct dm_pool_metadata *pmd = td->pmd;
1628         dm_block_t keys[2] = { td->id, block };
1629
1630         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1631         __dm_bless_for_disk(&value);
1632
1633         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1634                                    &pmd->root, &inserted);
1635         if (r)
1636                 return r;
1637
1638         td->changed = 1;
1639         if (inserted)
1640                 td->mapped_blocks++;
1641
1642         return 0;
1643 }
1644
1645 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1646                          dm_block_t data_block)
1647 {
1648         int r = -EINVAL;
1649
1650         pmd_write_lock(td->pmd);
1651         if (!td->pmd->fail_io)
1652                 r = __insert(td, block, data_block);
1653         pmd_write_unlock(td->pmd);
1654
1655         return r;
1656 }
1657
1658 static int __remove(struct dm_thin_device *td, dm_block_t block)
1659 {
1660         int r;
1661         struct dm_pool_metadata *pmd = td->pmd;
1662         dm_block_t keys[2] = { td->id, block };
1663
1664         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1665         if (r)
1666                 return r;
1667
1668         td->mapped_blocks--;
1669         td->changed = 1;
1670
1671         return 0;
1672 }
1673
1674 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1675 {
1676         int r;
1677         unsigned count, total_count = 0;
1678         struct dm_pool_metadata *pmd = td->pmd;
1679         dm_block_t keys[1] = { td->id };
1680         __le64 value;
1681         dm_block_t mapping_root;
1682
1683         /*
1684          * Find the mapping tree
1685          */
1686         r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1687         if (r)
1688                 return r;
1689
1690         /*
1691          * Remove from the mapping tree, taking care to inc the
1692          * ref count so it doesn't get deleted.
1693          */
1694         mapping_root = le64_to_cpu(value);
1695         dm_tm_inc(pmd->tm, mapping_root);
1696         r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1697         if (r)
1698                 return r;
1699
1700         /*
1701          * Remove leaves stops at the first unmapped entry, so we have to
1702          * loop round finding mapped ranges.
1703          */
1704         while (begin < end) {
1705                 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1706                 if (r == -ENODATA)
1707                         break;
1708
1709                 if (r)
1710                         return r;
1711
1712                 if (begin >= end)
1713                         break;
1714
1715                 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1716                 if (r)
1717                         return r;
1718
1719                 total_count += count;
1720         }
1721
1722         td->mapped_blocks -= total_count;
1723         td->changed = 1;
1724
1725         /*
1726          * Reinsert the mapping tree.
1727          */
1728         value = cpu_to_le64(mapping_root);
1729         __dm_bless_for_disk(&value);
1730         return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1731 }
1732
1733 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1734 {
1735         int r = -EINVAL;
1736
1737         pmd_write_lock(td->pmd);
1738         if (!td->pmd->fail_io)
1739                 r = __remove(td, block);
1740         pmd_write_unlock(td->pmd);
1741
1742         return r;
1743 }
1744
1745 int dm_thin_remove_range(struct dm_thin_device *td,
1746                          dm_block_t begin, dm_block_t end)
1747 {
1748         int r = -EINVAL;
1749
1750         pmd_write_lock(td->pmd);
1751         if (!td->pmd->fail_io)
1752                 r = __remove_range(td, begin, end);
1753         pmd_write_unlock(td->pmd);
1754
1755         return r;
1756 }
1757
1758 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1759 {
1760         int r;
1761         uint32_t ref_count;
1762
1763         down_read(&pmd->root_lock);
1764         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1765         if (!r)
1766                 *result = (ref_count > 1);
1767         up_read(&pmd->root_lock);
1768
1769         return r;
1770 }
1771
1772 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1773 {
1774         int r = 0;
1775
1776         pmd_write_lock(pmd);
1777         for (; b != e; b++) {
1778                 r = dm_sm_inc_block(pmd->data_sm, b);
1779                 if (r)
1780                         break;
1781         }
1782         pmd_write_unlock(pmd);
1783
1784         return r;
1785 }
1786
1787 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1788 {
1789         int r = 0;
1790
1791         pmd_write_lock(pmd);
1792         for (; b != e; b++) {
1793                 r = dm_sm_dec_block(pmd->data_sm, b);
1794                 if (r)
1795                         break;
1796         }
1797         pmd_write_unlock(pmd);
1798
1799         return r;
1800 }
1801
1802 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1803 {
1804         int r;
1805
1806         down_read(&td->pmd->root_lock);
1807         r = td->changed;
1808         up_read(&td->pmd->root_lock);
1809
1810         return r;
1811 }
1812
1813 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1814 {
1815         bool r = false;
1816         struct dm_thin_device *td, *tmp;
1817
1818         down_read(&pmd->root_lock);
1819         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1820                 if (td->changed) {
1821                         r = td->changed;
1822                         break;
1823                 }
1824         }
1825         up_read(&pmd->root_lock);
1826
1827         return r;
1828 }
1829
1830 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1831 {
1832         bool r;
1833
1834         down_read(&td->pmd->root_lock);
1835         r = td->aborted_with_changes;
1836         up_read(&td->pmd->root_lock);
1837
1838         return r;
1839 }
1840
1841 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1842 {
1843         int r = -EINVAL;
1844
1845         pmd_write_lock(pmd);
1846         if (!pmd->fail_io)
1847                 r = dm_sm_new_block(pmd->data_sm, result);
1848         pmd_write_unlock(pmd);
1849
1850         return r;
1851 }
1852
1853 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1854 {
1855         int r = -EINVAL;
1856
1857         /*
1858          * Care is taken to not have commit be what
1859          * triggers putting the thin-pool in-service.
1860          */
1861         pmd_write_lock_in_core(pmd);
1862         if (pmd->fail_io)
1863                 goto out;
1864
1865         r = __commit_transaction(pmd);
1866         if (r < 0)
1867                 goto out;
1868
1869         /*
1870          * Open the next transaction.
1871          */
1872         r = __begin_transaction(pmd);
1873 out:
1874         pmd_write_unlock(pmd);
1875         return r;
1876 }
1877
1878 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1879 {
1880         struct dm_thin_device *td;
1881
1882         list_for_each_entry(td, &pmd->thin_devices, list)
1883                 td->aborted_with_changes = td->changed;
1884 }
1885
1886 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1887 {
1888         int r = -EINVAL;
1889         struct dm_block_manager *old_bm = NULL, *new_bm = NULL;
1890
1891         /* fail_io is double-checked with pmd->root_lock held below */
1892         if (unlikely(pmd->fail_io))
1893                 return r;
1894
1895         /*
1896          * Replacement block manager (new_bm) is created and old_bm destroyed outside of
1897          * pmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of
1898          * shrinker associated with the block manager's bufio client vs pmd root_lock).
1899          * - must take shrinker_rwsem without holding pmd->root_lock
1900          */
1901         new_bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
1902                                          THIN_MAX_CONCURRENT_LOCKS);
1903
1904         pmd_write_lock(pmd);
1905         if (pmd->fail_io) {
1906                 pmd_write_unlock(pmd);
1907                 goto out;
1908         }
1909
1910         __set_abort_with_changes_flags(pmd);
1911         __destroy_persistent_data_objects(pmd, false);
1912         old_bm = pmd->bm;
1913         if (IS_ERR(new_bm)) {
1914                 DMERR("could not create block manager during abort");
1915                 pmd->bm = NULL;
1916                 r = PTR_ERR(new_bm);
1917                 goto out_unlock;
1918         }
1919
1920         pmd->bm = new_bm;
1921         r = __open_or_format_metadata(pmd, false);
1922         if (r) {
1923                 pmd->bm = NULL;
1924                 goto out_unlock;
1925         }
1926         new_bm = NULL;
1927 out_unlock:
1928         if (r)
1929                 pmd->fail_io = true;
1930         pmd_write_unlock(pmd);
1931         dm_block_manager_destroy(old_bm);
1932 out:
1933         if (new_bm && !IS_ERR(new_bm))
1934                 dm_block_manager_destroy(new_bm);
1935
1936         return r;
1937 }
1938
1939 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1940 {
1941         int r = -EINVAL;
1942
1943         down_read(&pmd->root_lock);
1944         if (!pmd->fail_io)
1945                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1946         up_read(&pmd->root_lock);
1947
1948         return r;
1949 }
1950
1951 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1952                                           dm_block_t *result)
1953 {
1954         int r = -EINVAL;
1955
1956         down_read(&pmd->root_lock);
1957         if (!pmd->fail_io)
1958                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1959
1960         if (!r) {
1961                 if (*result < pmd->metadata_reserve)
1962                         *result = 0;
1963                 else
1964                         *result -= pmd->metadata_reserve;
1965         }
1966         up_read(&pmd->root_lock);
1967
1968         return r;
1969 }
1970
1971 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1972                                   dm_block_t *result)
1973 {
1974         int r = -EINVAL;
1975
1976         down_read(&pmd->root_lock);
1977         if (!pmd->fail_io)
1978                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1979         up_read(&pmd->root_lock);
1980
1981         return r;
1982 }
1983
1984 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1985 {
1986         int r = -EINVAL;
1987
1988         down_read(&pmd->root_lock);
1989         if (!pmd->fail_io)
1990                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1991         up_read(&pmd->root_lock);
1992
1993         return r;
1994 }
1995
1996 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1997 {
1998         int r = -EINVAL;
1999         struct dm_pool_metadata *pmd = td->pmd;
2000
2001         down_read(&pmd->root_lock);
2002         if (!pmd->fail_io) {
2003                 *result = td->mapped_blocks;
2004                 r = 0;
2005         }
2006         up_read(&pmd->root_lock);
2007
2008         return r;
2009 }
2010
2011 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
2012 {
2013         int r;
2014         __le64 value_le;
2015         dm_block_t thin_root;
2016         struct dm_pool_metadata *pmd = td->pmd;
2017
2018         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
2019         if (r)
2020                 return r;
2021
2022         thin_root = le64_to_cpu(value_le);
2023
2024         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
2025 }
2026
2027 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
2028                                      dm_block_t *result)
2029 {
2030         int r = -EINVAL;
2031         struct dm_pool_metadata *pmd = td->pmd;
2032
2033         down_read(&pmd->root_lock);
2034         if (!pmd->fail_io)
2035                 r = __highest_block(td, result);
2036         up_read(&pmd->root_lock);
2037
2038         return r;
2039 }
2040
2041 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
2042 {
2043         int r;
2044         dm_block_t old_count;
2045
2046         r = dm_sm_get_nr_blocks(sm, &old_count);
2047         if (r)
2048                 return r;
2049
2050         if (new_count == old_count)
2051                 return 0;
2052
2053         if (new_count < old_count) {
2054                 DMERR("cannot reduce size of space map");
2055                 return -EINVAL;
2056         }
2057
2058         return dm_sm_extend(sm, new_count - old_count);
2059 }
2060
2061 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2062 {
2063         int r = -EINVAL;
2064
2065         pmd_write_lock(pmd);
2066         if (!pmd->fail_io)
2067                 r = __resize_space_map(pmd->data_sm, new_count);
2068         pmd_write_unlock(pmd);
2069
2070         return r;
2071 }
2072
2073 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2074 {
2075         int r = -EINVAL;
2076
2077         pmd_write_lock(pmd);
2078         if (!pmd->fail_io) {
2079                 r = __resize_space_map(pmd->metadata_sm, new_count);
2080                 if (!r)
2081                         __set_metadata_reserve(pmd);
2082         }
2083         pmd_write_unlock(pmd);
2084
2085         return r;
2086 }
2087
2088 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2089 {
2090         pmd_write_lock_in_core(pmd);
2091         dm_bm_set_read_only(pmd->bm);
2092         pmd_write_unlock(pmd);
2093 }
2094
2095 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2096 {
2097         pmd_write_lock_in_core(pmd);
2098         dm_bm_set_read_write(pmd->bm);
2099         pmd_write_unlock(pmd);
2100 }
2101
2102 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2103                                         dm_block_t threshold,
2104                                         dm_sm_threshold_fn fn,
2105                                         void *context)
2106 {
2107         int r = -EINVAL;
2108
2109         pmd_write_lock_in_core(pmd);
2110         if (!pmd->fail_io) {
2111                 r = dm_sm_register_threshold_callback(pmd->metadata_sm,
2112                                                       threshold, fn, context);
2113         }
2114         pmd_write_unlock(pmd);
2115
2116         return r;
2117 }
2118
2119 void dm_pool_register_pre_commit_callback(struct dm_pool_metadata *pmd,
2120                                           dm_pool_pre_commit_fn fn,
2121                                           void *context)
2122 {
2123         pmd_write_lock_in_core(pmd);
2124         pmd->pre_commit_fn = fn;
2125         pmd->pre_commit_context = context;
2126         pmd_write_unlock(pmd);
2127 }
2128
2129 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2130 {
2131         int r = -EINVAL;
2132         struct dm_block *sblock;
2133         struct thin_disk_superblock *disk_super;
2134
2135         pmd_write_lock(pmd);
2136         if (pmd->fail_io)
2137                 goto out;
2138
2139         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2140
2141         r = superblock_lock(pmd, &sblock);
2142         if (r) {
2143                 DMERR("couldn't lock superblock");
2144                 goto out;
2145         }
2146
2147         disk_super = dm_block_data(sblock);
2148         disk_super->flags = cpu_to_le32(pmd->flags);
2149
2150         dm_bm_unlock(sblock);
2151 out:
2152         pmd_write_unlock(pmd);
2153         return r;
2154 }
2155
2156 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2157 {
2158         bool needs_check;
2159
2160         down_read(&pmd->root_lock);
2161         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2162         up_read(&pmd->root_lock);
2163
2164         return needs_check;
2165 }
2166
2167 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2168 {
2169         down_read(&pmd->root_lock);
2170         if (!pmd->fail_io)
2171                 dm_tm_issue_prefetches(pmd->tm);
2172         up_read(&pmd->root_lock);
2173 }