kernel/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 THIN_METADATA_CACHE_SIZE 64
  81 #define SECTOR_TO_BLOCK_SHIFT 3
  82
  83 /*
  84  *  3 for btree insert +
  85  *  2 for btree lookup used within space map
  86  */
  87 #define THIN_MAX_CONCURRENT_LOCKS 5
  88
  89 /* This should be plenty */
  90 #define SPACE_MAP_ROOT_SIZE 128
  91
  92 /*
  93  * Little endian on-disk superblock and device details.
  94  */
  95 struct thin_disk_superblock {
  96         __le32 csum;    /* Checksum of superblock except for this field. */
  97         __le32 flags;
  98         __le64 blocknr; /* This block number, dm_block_t. */
  99
 100         __u8 uuid[16];
 101         __le64 magic;
 102         __le32 version;
 103         __le32 time;
 104
 105         __le64 trans_id;
 106
 107         /*
 108          * Root held by userspace transactions.
 109          */
 110         __le64 held_root;
 111
 112         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 113         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 114
 115         /*
 116          * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 117          */
 118         __le64 data_mapping_root;
 119
 120         /*
 121          * Device detail root mapping dev_id -> device_details
 122          */
 123         __le64 device_details_root;
 124
 125         __le32 data_block_size;         /* In 512-byte sectors. */
 126
 127         __le32 metadata_block_size;     /* In 512-byte sectors. */
 128         __le64 metadata_nr_blocks;
 129
 130         __le32 compat_flags;
 131         __le32 compat_ro_flags;
 132         __le32 incompat_flags;
 133 } __packed;
 134
 135 struct disk_device_details {
 136         __le64 mapped_blocks;
 137         __le64 transaction_id;          /* When created. */
 138         __le32 creation_time;
 139         __le32 snapshotted_time;
 140 } __packed;
 141
 142 struct dm_pool_metadata {
 143         struct hlist_node hash;
 144
 145         struct block_device *bdev;
 146         struct dm_block_manager *bm;
 147         struct dm_space_map *metadata_sm;
 148         struct dm_space_map *data_sm;
 149         struct dm_transaction_manager *tm;
 150         struct dm_transaction_manager *nb_tm;
 151
 152         /*
 153          * Two-level btree.
 154          * First level holds thin_dev_t.
 155          * Second level holds mappings.
 156          */
 157         struct dm_btree_info info;
 158
 159         /*
 160          * Non-blocking version of the above.
 161          */
 162         struct dm_btree_info nb_info;
 163
 164         /*
 165          * Just the top level for deleting whole devices.
 166          */
 167         struct dm_btree_info tl_info;
 168
 169         /*
 170          * Just the bottom level for creating new devices.
 171          */
 172         struct dm_btree_info bl_info;
 173
 174         /*
 175          * Describes the device details btree.
 176          */
 177         struct dm_btree_info details_info;
 178
 179         struct rw_semaphore root_lock;
 180         uint32_t time;
 181         dm_block_t root;
 182         dm_block_t details_root;
 183         struct list_head thin_devices;
 184         uint64_t trans_id;
 185         unsigned long flags;
 186         sector_t data_block_size;
 187         bool read_only:1;
 188
 189         /*
 190          * Set if a transaction has to be aborted but the attempt to roll back
 191          * to the previous (good) transaction failed.  The only pool metadata
 192          * operation possible in this state is the closing of the device.
 193          */
 194         bool fail_io:1;
 195
 196         /*
 197          * Reading the space map roots can fail, so we read it into these
 198          * buffers before the superblock is locked and updated.
 199          */
 200         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 201         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 202 };
 203
 204 struct dm_thin_device {
 205         struct list_head list;
 206         struct dm_pool_metadata *pmd;
 207         dm_thin_id id;
 208
 209         int open_count;
 210         bool changed:1;
 211         bool aborted_with_changes:1;
 212         uint64_t mapped_blocks;
 213         uint64_t transaction_id;
 214         uint32_t creation_time;
 215         uint32_t snapshotted_time;
 216 };
 217
 218 /*----------------------------------------------------------------
 219  * superblock validator
 220  *--------------------------------------------------------------*/
 221
 222 #define SUPERBLOCK_CSUM_XOR 160774
 223
 224 static void sb_prepare_for_write(struct dm_block_validator *v,
 225                                  struct dm_block *b,
 226                                  size_t block_size)
 227 {
 228         struct thin_disk_superblock *disk_super = dm_block_data(b);
 229
 230         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 231         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 232                                                       block_size - sizeof(__le32),
 233                                                       SUPERBLOCK_CSUM_XOR));
 234 }
 235
 236 static int sb_check(struct dm_block_validator *v,
 237                     struct dm_block *b,
 238                     size_t block_size)
 239 {
 240         struct thin_disk_superblock *disk_super = dm_block_data(b);
 241         __le32 csum_le;
 242
 243         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 244                 DMERR("sb_check failed: blocknr %llu: "
 245                       "wanted %llu", le64_to_cpu(disk_super->blocknr),
 246                       (unsigned long long)dm_block_location(b));
 247                 return -ENOTBLK;
 248         }
 249
 250         if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 251                 DMERR("sb_check failed: magic %llu: "
 252                       "wanted %llu", le64_to_cpu(disk_super->magic),
 253                       (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 254                 return -EILSEQ;
 255         }
 256
 257         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 258                                              block_size - sizeof(__le32),
 259                                              SUPERBLOCK_CSUM_XOR));
 260         if (csum_le != disk_super->csum) {
 261                 DMERR("sb_check failed: csum %u: wanted %u",
 262                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 263                 return -EILSEQ;
 264         }
 265
 266         return 0;
 267 }
 268
 269 static struct dm_block_validator sb_validator = {
 270         .name = "superblock",
 271         .prepare_for_write = sb_prepare_for_write,
 272         .check = sb_check
 273 };
 274
 275 /*----------------------------------------------------------------
 276  * Methods for the btree value types
 277  *--------------------------------------------------------------*/
 278
 279 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 280 {
 281         return (b << 24) | t;
 282 }
 283
 284 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 285 {
 286         *b = v >> 24;
 287         *t = v & ((1 << 24) - 1);
 288 }
 289
 290 static void data_block_inc(void *context, const void *value_le)
 291 {
 292         struct dm_space_map *sm = context;
 293         __le64 v_le;
 294         uint64_t b;
 295         uint32_t t;
 296
 297         memcpy(&v_le, value_le, sizeof(v_le));
 298         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 299         dm_sm_inc_block(sm, b);
 300 }
 301
 302 static void data_block_dec(void *context, const void *value_le)
 303 {
 304         struct dm_space_map *sm = context;
 305         __le64 v_le;
 306         uint64_t b;
 307         uint32_t t;
 308
 309         memcpy(&v_le, value_le, sizeof(v_le));
 310         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 311         dm_sm_dec_block(sm, b);
 312 }
 313
 314 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 315 {
 316         __le64 v1_le, v2_le;
 317         uint64_t b1, b2;
 318         uint32_t t;
 319
 320         memcpy(&v1_le, value1_le, sizeof(v1_le));
 321         memcpy(&v2_le, value2_le, sizeof(v2_le));
 322         unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 323         unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 324
 325         return b1 == b2;
 326 }
 327
 328 static void subtree_inc(void *context, const void *value)
 329 {
 330         struct dm_btree_info *info = context;
 331         __le64 root_le;
 332         uint64_t root;
 333
 334         memcpy(&root_le, value, sizeof(root_le));
 335         root = le64_to_cpu(root_le);
 336         dm_tm_inc(info->tm, root);
 337 }
 338
 339 static void subtree_dec(void *context, const void *value)
 340 {
 341         struct dm_btree_info *info = context;
 342         __le64 root_le;
 343         uint64_t root;
 344
 345         memcpy(&root_le, value, sizeof(root_le));
 346         root = le64_to_cpu(root_le);
 347         if (dm_btree_del(info, root))
 348                 DMERR("btree delete failed\n");
 349 }
 350
 351 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 352 {
 353         __le64 v1_le, v2_le;
 354         memcpy(&v1_le, value1_le, sizeof(v1_le));
 355         memcpy(&v2_le, value2_le, sizeof(v2_le));
 356
 357         return v1_le == v2_le;
 358 }
 359
 360 /*----------------------------------------------------------------*/
 361
 362 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 363                                 struct dm_block **sblock)
 364 {
 365         return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 366                                      &sb_validator, sblock);
 367 }
 368
 369 static int superblock_lock(struct dm_pool_metadata *pmd,
 370                            struct dm_block **sblock)
 371 {
 372         return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 373                                 &sb_validator, sblock);
 374 }
 375
 376 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 377 {
 378         int r;
 379         unsigned i;
 380         struct dm_block *b;
 381         __le64 *data_le, zero = cpu_to_le64(0);
 382         unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 383
 384         /*
 385          * We can't use a validator here - it may be all zeroes.
 386          */
 387         r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 388         if (r)
 389                 return r;
 390
 391         data_le = dm_block_data(b);
 392         *result = 1;
 393         for (i = 0; i < block_size; i++) {
 394                 if (data_le[i] != zero) {
 395                         *result = 0;
 396                         break;
 397                 }
 398         }
 399
 400         return dm_bm_unlock(b);
 401 }
 402
 403 static void __setup_btree_details(struct dm_pool_metadata *pmd)
 404 {
 405         pmd->info.tm = pmd->tm;
 406         pmd->info.levels = 2;
 407         pmd->info.value_type.context = pmd->data_sm;
 408         pmd->info.value_type.size = sizeof(__le64);
 409         pmd->info.value_type.inc = data_block_inc;
 410         pmd->info.value_type.dec = data_block_dec;
 411         pmd->info.value_type.equal = data_block_equal;
 412
 413         memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 414         pmd->nb_info.tm = pmd->nb_tm;
 415
 416         pmd->tl_info.tm = pmd->tm;
 417         pmd->tl_info.levels = 1;
 418         pmd->tl_info.value_type.context = &pmd->bl_info;
 419         pmd->tl_info.value_type.size = sizeof(__le64);
 420         pmd->tl_info.value_type.inc = subtree_inc;
 421         pmd->tl_info.value_type.dec = subtree_dec;
 422         pmd->tl_info.value_type.equal = subtree_equal;
 423
 424         pmd->bl_info.tm = pmd->tm;
 425         pmd->bl_info.levels = 1;
 426         pmd->bl_info.value_type.context = pmd->data_sm;
 427         pmd->bl_info.value_type.size = sizeof(__le64);
 428         pmd->bl_info.value_type.inc = data_block_inc;
 429         pmd->bl_info.value_type.dec = data_block_dec;
 430         pmd->bl_info.value_type.equal = data_block_equal;
 431
 432         pmd->details_info.tm = pmd->tm;
 433         pmd->details_info.levels = 1;
 434         pmd->details_info.value_type.context = NULL;
 435         pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 436         pmd->details_info.value_type.inc = NULL;
 437         pmd->details_info.value_type.dec = NULL;
 438         pmd->details_info.value_type.equal = NULL;
 439 }
 440
 441 static int save_sm_roots(struct dm_pool_metadata *pmd)
 442 {
 443         int r;
 444         size_t len;
 445
 446         r = dm_sm_root_size(pmd->metadata_sm, &len);
 447         if (r < 0)
 448                 return r;
 449
 450         r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 451         if (r < 0)
 452                 return r;
 453
 454         r = dm_sm_root_size(pmd->data_sm, &len);
 455         if (r < 0)
 456                 return r;
 457
 458         return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 459 }
 460
 461 static void copy_sm_roots(struct dm_pool_metadata *pmd,
 462                           struct thin_disk_superblock *disk)
 463 {
 464         memcpy(&disk->metadata_space_map_root,
 465                &pmd->metadata_space_map_root,
 466                sizeof(pmd->metadata_space_map_root));
 467
 468         memcpy(&disk->data_space_map_root,
 469                &pmd->data_space_map_root,
 470                sizeof(pmd->data_space_map_root));
 471 }
 472
 473 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 474 {
 475         int r;
 476         struct dm_block *sblock;
 477         struct thin_disk_superblock *disk_super;
 478         sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 479
 480         if (bdev_size > THIN_METADATA_MAX_SECTORS)
 481                 bdev_size = THIN_METADATA_MAX_SECTORS;
 482
 483         r = dm_sm_commit(pmd->data_sm);
 484         if (r < 0)
 485                 return r;
 486
 487         r = save_sm_roots(pmd);
 488         if (r < 0)
 489                 return r;
 490
 491         r = dm_tm_pre_commit(pmd->tm);
 492         if (r < 0)
 493                 return r;
 494
 495         r = superblock_lock_zero(pmd, &sblock);
 496         if (r)
 497                 return r;
 498
 499         disk_super = dm_block_data(sblock);
 500         disk_super->flags = 0;
 501         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 502         disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 503         disk_super->version = cpu_to_le32(THIN_VERSION);
 504         disk_super->time = 0;
 505         disk_super->trans_id = 0;
 506         disk_super->held_root = 0;
 507
 508         copy_sm_roots(pmd, disk_super);
 509
 510         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 511         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 512         disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 513         disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 514         disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 515
 516         return dm_tm_commit(pmd->tm, sblock);
 517 }
 518
 519 static int __format_metadata(struct dm_pool_metadata *pmd)
 520 {
 521         int r;
 522
 523         r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 524                                  &pmd->tm, &pmd->metadata_sm);
 525         if (r < 0) {
 526                 DMERR("tm_create_with_sm failed");
 527                 return r;
 528         }
 529
 530         pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 531         if (IS_ERR(pmd->data_sm)) {
 532                 DMERR("sm_disk_create failed");
 533                 r = PTR_ERR(pmd->data_sm);
 534                 goto bad_cleanup_tm;
 535         }
 536
 537         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 538         if (!pmd->nb_tm) {
 539                 DMERR("could not create non-blocking clone tm");
 540                 r = -ENOMEM;
 541                 goto bad_cleanup_data_sm;
 542         }
 543
 544         __setup_btree_details(pmd);
 545
 546         r = dm_btree_empty(&pmd->info, &pmd->root);
 547         if (r < 0)
 548                 goto bad_cleanup_nb_tm;
 549
 550         r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 551         if (r < 0) {
 552                 DMERR("couldn't create devices root");
 553                 goto bad_cleanup_nb_tm;
 554         }
 555
 556         r = __write_initial_superblock(pmd);
 557         if (r)
 558                 goto bad_cleanup_nb_tm;
 559
 560         return 0;
 561
 562 bad_cleanup_nb_tm:
 563         dm_tm_destroy(pmd->nb_tm);
 564 bad_cleanup_data_sm:
 565         dm_sm_destroy(pmd->data_sm);
 566 bad_cleanup_tm:
 567         dm_tm_destroy(pmd->tm);
 568         dm_sm_destroy(pmd->metadata_sm);
 569
 570         return r;
 571 }
 572
 573 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 574                                      struct dm_pool_metadata *pmd)
 575 {
 576         uint32_t features;
 577
 578         features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 579         if (features) {
 580                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 581                       (unsigned long)features);
 582                 return -EINVAL;
 583         }
 584
 585         /*
 586          * Check for read-only metadata to skip the following RDWR checks.
 587          */
 588         if (get_disk_ro(pmd->bdev->bd_disk))
 589                 return 0;
 590
 591         features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 592         if (features) {
 593                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 594                       (unsigned long)features);
 595                 return -EINVAL;
 596         }
 597
 598         return 0;
 599 }
 600
 601 static int __open_metadata(struct dm_pool_metadata *pmd)
 602 {
 603         int r;
 604         struct dm_block *sblock;
 605         struct thin_disk_superblock *disk_super;
 606
 607         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 608                             &sb_validator, &sblock);
 609         if (r < 0) {
 610                 DMERR("couldn't read superblock");
 611                 return r;
 612         }
 613
 614         disk_super = dm_block_data(sblock);
 615
 616         /* Verify the data block size hasn't changed */
 617         if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 618                 DMERR("changing the data block size (from %u to %llu) is not supported",
 619                       le32_to_cpu(disk_super->data_block_size),
 620                       (unsigned long long)pmd->data_block_size);
 621                 r = -EINVAL;
 622                 goto bad_unlock_sblock;
 623         }
 624
 625         r = __check_incompat_features(disk_super, pmd);
 626         if (r < 0)
 627                 goto bad_unlock_sblock;
 628
 629         r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 630                                disk_super->metadata_space_map_root,
 631                                sizeof(disk_super->metadata_space_map_root),
 632                                &pmd->tm, &pmd->metadata_sm);
 633         if (r < 0) {
 634                 DMERR("tm_open_with_sm failed");
 635                 goto bad_unlock_sblock;
 636         }
 637
 638         pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 639                                        sizeof(disk_super->data_space_map_root));
 640         if (IS_ERR(pmd->data_sm)) {
 641                 DMERR("sm_disk_open failed");
 642                 r = PTR_ERR(pmd->data_sm);
 643                 goto bad_cleanup_tm;
 644         }
 645
 646         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 647         if (!pmd->nb_tm) {
 648                 DMERR("could not create non-blocking clone tm");
 649                 r = -ENOMEM;
 650                 goto bad_cleanup_data_sm;
 651         }
 652
 653         __setup_btree_details(pmd);
 654         return dm_bm_unlock(sblock);
 655
 656 bad_cleanup_data_sm:
 657         dm_sm_destroy(pmd->data_sm);
 658 bad_cleanup_tm:
 659         dm_tm_destroy(pmd->tm);
 660         dm_sm_destroy(pmd->metadata_sm);
 661 bad_unlock_sblock:
 662         dm_bm_unlock(sblock);
 663
 664         return r;
 665 }
 666
 667 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 668 {
 669         int r, unformatted;
 670
 671         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 672         if (r)
 673                 return r;
 674
 675         if (unformatted)
 676                 return format_device ? __format_metadata(pmd) : -EPERM;
 677
 678         return __open_metadata(pmd);
 679 }
 680
 681 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 682 {
 683         int r;
 684
 685         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 686                                           THIN_METADATA_CACHE_SIZE,
 687                                           THIN_MAX_CONCURRENT_LOCKS);
 688         if (IS_ERR(pmd->bm)) {
 689                 DMERR("could not create block manager");
 690                 return PTR_ERR(pmd->bm);
 691         }
 692
 693         r = __open_or_format_metadata(pmd, format_device);
 694         if (r)
 695                 dm_block_manager_destroy(pmd->bm);
 696
 697         return r;
 698 }
 699
 700 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 701 {
 702         dm_sm_destroy(pmd->data_sm);
 703         dm_sm_destroy(pmd->metadata_sm);
 704         dm_tm_destroy(pmd->nb_tm);
 705         dm_tm_destroy(pmd->tm);
 706         dm_block_manager_destroy(pmd->bm);
 707 }
 708
 709 static int __begin_transaction(struct dm_pool_metadata *pmd)
 710 {
 711         int r;
 712         struct thin_disk_superblock *disk_super;
 713         struct dm_block *sblock;
 714
 715         /*
 716          * We re-read the superblock every time.  Shouldn't need to do this
 717          * really.
 718          */
 719         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 720                             &sb_validator, &sblock);
 721         if (r)
 722                 return r;
 723
 724         disk_super = dm_block_data(sblock);
 725         pmd->time = le32_to_cpu(disk_super->time);
 726         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 727         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 728         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 729         pmd->flags = le32_to_cpu(disk_super->flags);
 730         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 731
 732         dm_bm_unlock(sblock);
 733         return 0;
 734 }
 735
 736 static int __write_changed_details(struct dm_pool_metadata *pmd)
 737 {
 738         int r;
 739         struct dm_thin_device *td, *tmp;
 740         struct disk_device_details details;
 741         uint64_t key;
 742
 743         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 744                 if (!td->changed)
 745                         continue;
 746
 747                 key = td->id;
 748
 749                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 750                 details.transaction_id = cpu_to_le64(td->transaction_id);
 751                 details.creation_time = cpu_to_le32(td->creation_time);
 752                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 753                 __dm_bless_for_disk(&details);
 754
 755                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 756                                     &key, &details, &pmd->details_root);
 757                 if (r)
 758                         return r;
 759
 760                 if (td->open_count)
 761                         td->changed = 0;
 762                 else {
 763                         list_del(&td->list);
 764                         kfree(td);
 765                 }
 766         }
 767
 768         return 0;
 769 }
 770
 771 static int __commit_transaction(struct dm_pool_metadata *pmd)
 772 {
 773         int r;
 774         size_t metadata_len, data_len;
 775         struct thin_disk_superblock *disk_super;
 776         struct dm_block *sblock;
 777
 778         /*
 779          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 780          */
 781         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 782
 783         r = __write_changed_details(pmd);
 784         if (r < 0)
 785                 return r;
 786
 787         r = dm_sm_commit(pmd->data_sm);
 788         if (r < 0)
 789                 return r;
 790
 791         r = dm_tm_pre_commit(pmd->tm);
 792         if (r < 0)
 793                 return r;
 794
 795         r = dm_sm_root_size(pmd->metadata_sm, &metadata_len);
 796         if (r < 0)
 797                 return r;
 798
 799         r = dm_sm_root_size(pmd->data_sm, &data_len);
 800         if (r < 0)
 801                 return r;
 802
 803         r = save_sm_roots(pmd);
 804         if (r < 0)
 805                 return r;
 806
 807         r = superblock_lock(pmd, &sblock);
 808         if (r)
 809                 return r;
 810
 811         disk_super = dm_block_data(sblock);
 812         disk_super->time = cpu_to_le32(pmd->time);
 813         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 814         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 815         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 816         disk_super->flags = cpu_to_le32(pmd->flags);
 817
 818         copy_sm_roots(pmd, disk_super);
 819
 820         return dm_tm_commit(pmd->tm, sblock);
 821 }
 822
 823 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 824                                                sector_t data_block_size,
 825                                                bool format_device)
 826 {
 827         int r;
 828         struct dm_pool_metadata *pmd;
 829
 830         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 831         if (!pmd) {
 832                 DMERR("could not allocate metadata struct");
 833                 return ERR_PTR(-ENOMEM);
 834         }
 835
 836         init_rwsem(&pmd->root_lock);
 837         pmd->time = 0;
 838         INIT_LIST_HEAD(&pmd->thin_devices);
 839         pmd->read_only = false;
 840         pmd->fail_io = false;
 841         pmd->bdev = bdev;
 842         pmd->data_block_size = data_block_size;
 843
 844         r = __create_persistent_data_objects(pmd, format_device);
 845         if (r) {
 846                 kfree(pmd);
 847                 return ERR_PTR(r);
 848         }
 849
 850         r = __begin_transaction(pmd);
 851         if (r < 0) {
 852                 if (dm_pool_metadata_close(pmd) < 0)
 853                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 854                 return ERR_PTR(r);
 855         }
 856
 857         return pmd;
 858 }
 859
 860 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 861 {
 862         int r;
 863         unsigned open_devices = 0;
 864         struct dm_thin_device *td, *tmp;
 865
 866         down_read(&pmd->root_lock);
 867         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 868                 if (td->open_count)
 869                         open_devices++;
 870                 else {
 871                         list_del(&td->list);
 872                         kfree(td);
 873                 }
 874         }
 875         up_read(&pmd->root_lock);
 876
 877         if (open_devices) {
 878                 DMERR("attempt to close pmd when %u device(s) are still open",
 879                        open_devices);
 880                 return -EBUSY;
 881         }
 882
 883         if (!pmd->read_only && !pmd->fail_io) {
 884                 r = __commit_transaction(pmd);
 885                 if (r < 0)
 886                         DMWARN("%s: __commit_transaction() failed, error = %d",
 887                                __func__, r);
 888         }
 889
 890         if (!pmd->fail_io)
 891                 __destroy_persistent_data_objects(pmd);
 892
 893         kfree(pmd);
 894         return 0;
 895 }
 896
 897 /*
 898  * __open_device: Returns @td corresponding to device with id @dev,
 899  * creating it if @create is set and incrementing @td->open_count.
 900  * On failure, @td is undefined.
 901  */
 902 static int __open_device(struct dm_pool_metadata *pmd,
 903                          dm_thin_id dev, int create,
 904                          struct dm_thin_device **td)
 905 {
 906         int r, changed = 0;
 907         struct dm_thin_device *td2;
 908         uint64_t key = dev;
 909         struct disk_device_details details_le;
 910
 911         /*
 912          * If the device is already open, return it.
 913          */
 914         list_for_each_entry(td2, &pmd->thin_devices, list)
 915                 if (td2->id == dev) {
 916                         /*
 917                          * May not create an already-open device.
 918                          */
 919                         if (create)
 920                                 return -EEXIST;
 921
 922                         td2->open_count++;
 923                         *td = td2;
 924                         return 0;
 925                 }
 926
 927         /*
 928          * Check the device exists.
 929          */
 930         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 931                             &key, &details_le);
 932         if (r) {
 933                 if (r != -ENODATA || !create)
 934                         return r;
 935
 936                 /*
 937                  * Create new device.
 938                  */
 939                 changed = 1;
 940                 details_le.mapped_blocks = 0;
 941                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
 942                 details_le.creation_time = cpu_to_le32(pmd->time);
 943                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
 944         }
 945
 946         *td = kmalloc(sizeof(**td), GFP_NOIO);
 947         if (!*td)
 948                 return -ENOMEM;
 949
 950         (*td)->pmd = pmd;
 951         (*td)->id = dev;
 952         (*td)->open_count = 1;
 953         (*td)->changed = changed;
 954         (*td)->aborted_with_changes = false;
 955         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
 956         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
 957         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
 958         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
 959
 960         list_add(&(*td)->list, &pmd->thin_devices);
 961
 962         return 0;
 963 }
 964
 965 static void __close_device(struct dm_thin_device *td)
 966 {
 967         --td->open_count;
 968 }
 969
 970 static int __create_thin(struct dm_pool_metadata *pmd,
 971                          dm_thin_id dev)
 972 {
 973         int r;
 974         dm_block_t dev_root;
 975         uint64_t key = dev;
 976         struct disk_device_details details_le;
 977         struct dm_thin_device *td;
 978         __le64 value;
 979
 980         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 981                             &key, &details_le);
 982         if (!r)
 983                 return -EEXIST;
 984
 985         /*
 986          * Create an empty btree for the mappings.
 987          */
 988         r = dm_btree_empty(&pmd->bl_info, &dev_root);
 989         if (r)
 990                 return r;
 991
 992         /*
 993          * Insert it into the main mapping tree.
 994          */
 995         value = cpu_to_le64(dev_root);
 996         __dm_bless_for_disk(&value);
 997         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
 998         if (r) {
 999                 dm_btree_del(&pmd->bl_info, dev_root);
1000                 return r;
1001         }
1002
1003         r = __open_device(pmd, dev, 1, &td);
1004         if (r) {
1005                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1006                 dm_btree_del(&pmd->bl_info, dev_root);
1007                 return r;
1008         }
1009         __close_device(td);
1010
1011         return r;
1012 }
1013
1014 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1015 {
1016         int r = -EINVAL;
1017
1018         down_write(&pmd->root_lock);
1019         if (!pmd->fail_io)
1020                 r = __create_thin(pmd, dev);
1021         up_write(&pmd->root_lock);
1022
1023         return r;
1024 }
1025
1026 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1027                                   struct dm_thin_device *snap,
1028                                   dm_thin_id origin, uint32_t time)
1029 {
1030         int r;
1031         struct dm_thin_device *td;
1032
1033         r = __open_device(pmd, origin, 0, &td);
1034         if (r)
1035                 return r;
1036
1037         td->changed = 1;
1038         td->snapshotted_time = time;
1039
1040         snap->mapped_blocks = td->mapped_blocks;
1041         snap->snapshotted_time = time;
1042         __close_device(td);
1043
1044         return 0;
1045 }
1046
1047 static int __create_snap(struct dm_pool_metadata *pmd,
1048                          dm_thin_id dev, dm_thin_id origin)
1049 {
1050         int r;
1051         dm_block_t origin_root;
1052         uint64_t key = origin, dev_key = dev;
1053         struct dm_thin_device *td;
1054         struct disk_device_details details_le;
1055         __le64 value;
1056
1057         /* check this device is unused */
1058         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1059                             &dev_key, &details_le);
1060         if (!r)
1061                 return -EEXIST;
1062
1063         /* find the mapping tree for the origin */
1064         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1065         if (r)
1066                 return r;
1067         origin_root = le64_to_cpu(value);
1068
1069         /* clone the origin, an inc will do */
1070         dm_tm_inc(pmd->tm, origin_root);
1071
1072         /* insert into the main mapping tree */
1073         value = cpu_to_le64(origin_root);
1074         __dm_bless_for_disk(&value);
1075         key = dev;
1076         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1077         if (r) {
1078                 dm_tm_dec(pmd->tm, origin_root);
1079                 return r;
1080         }
1081
1082         pmd->time++;
1083
1084         r = __open_device(pmd, dev, 1, &td);
1085         if (r)
1086                 goto bad;
1087
1088         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1089         __close_device(td);
1090
1091         if (r)
1092                 goto bad;
1093
1094         return 0;
1095
1096 bad:
1097         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1098         dm_btree_remove(&pmd->details_info, pmd->details_root,
1099                         &key, &pmd->details_root);
1100         return r;
1101 }
1102
1103 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1104                                  dm_thin_id dev,
1105                                  dm_thin_id origin)
1106 {
1107         int r = -EINVAL;
1108
1109         down_write(&pmd->root_lock);
1110         if (!pmd->fail_io)
1111                 r = __create_snap(pmd, dev, origin);
1112         up_write(&pmd->root_lock);
1113
1114         return r;
1115 }
1116
1117 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1118 {
1119         int r;
1120         uint64_t key = dev;
1121         struct dm_thin_device *td;
1122
1123         /* TODO: failure should mark the transaction invalid */
1124         r = __open_device(pmd, dev, 0, &td);
1125         if (r)
1126                 return r;
1127
1128         if (td->open_count > 1) {
1129                 __close_device(td);
1130                 return -EBUSY;
1131         }
1132
1133         list_del(&td->list);
1134         kfree(td);
1135         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1136                             &key, &pmd->details_root);
1137         if (r)
1138                 return r;
1139
1140         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1141         if (r)
1142                 return r;
1143
1144         return 0;
1145 }
1146
1147 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1148                                dm_thin_id dev)
1149 {
1150         int r = -EINVAL;
1151
1152         down_write(&pmd->root_lock);
1153         if (!pmd->fail_io)
1154                 r = __delete_device(pmd, dev);
1155         up_write(&pmd->root_lock);
1156
1157         return r;
1158 }
1159
1160 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1161                                         uint64_t current_id,
1162                                         uint64_t new_id)
1163 {
1164         int r = -EINVAL;
1165
1166         down_write(&pmd->root_lock);
1167
1168         if (pmd->fail_io)
1169                 goto out;
1170
1171         if (pmd->trans_id != current_id) {
1172                 DMERR("mismatched transaction id");
1173                 goto out;
1174         }
1175
1176         pmd->trans_id = new_id;
1177         r = 0;
1178
1179 out:
1180         up_write(&pmd->root_lock);
1181
1182         return r;
1183 }
1184
1185 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1186                                         uint64_t *result)
1187 {
1188         int r = -EINVAL;
1189
1190         down_read(&pmd->root_lock);
1191         if (!pmd->fail_io) {
1192                 *result = pmd->trans_id;
1193                 r = 0;
1194         }
1195         up_read(&pmd->root_lock);
1196
1197         return r;
1198 }
1199
1200 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1201 {
1202         int r, inc;
1203         struct thin_disk_superblock *disk_super;
1204         struct dm_block *copy, *sblock;
1205         dm_block_t held_root;
1206
1207         /*
1208          * Copy the superblock.
1209          */
1210         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1211         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1212                                &sb_validator, &copy, &inc);
1213         if (r)
1214                 return r;
1215
1216         BUG_ON(!inc);
1217
1218         held_root = dm_block_location(copy);
1219         disk_super = dm_block_data(copy);
1220
1221         if (le64_to_cpu(disk_super->held_root)) {
1222                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1223
1224                 dm_tm_dec(pmd->tm, held_root);
1225                 dm_tm_unlock(pmd->tm, copy);
1226                 return -EBUSY;
1227         }
1228
1229         /*
1230          * Wipe the spacemap since we're not publishing this.
1231          */
1232         memset(&disk_super->data_space_map_root, 0,
1233                sizeof(disk_super->data_space_map_root));
1234         memset(&disk_super->metadata_space_map_root, 0,
1235                sizeof(disk_super->metadata_space_map_root));
1236
1237         /*
1238          * Increment the data structures that need to be preserved.
1239          */
1240         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1241         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1242         dm_tm_unlock(pmd->tm, copy);
1243
1244         /*
1245          * Write the held root into the superblock.
1246          */
1247         r = superblock_lock(pmd, &sblock);
1248         if (r) {
1249                 dm_tm_dec(pmd->tm, held_root);
1250                 return r;
1251         }
1252
1253         disk_super = dm_block_data(sblock);
1254         disk_super->held_root = cpu_to_le64(held_root);
1255         dm_bm_unlock(sblock);
1256         return 0;
1257 }
1258
1259 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1260 {
1261         int r = -EINVAL;
1262
1263         down_write(&pmd->root_lock);
1264         if (!pmd->fail_io)
1265                 r = __reserve_metadata_snap(pmd);
1266         up_write(&pmd->root_lock);
1267
1268         return r;
1269 }
1270
1271 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1272 {
1273         int r;
1274         struct thin_disk_superblock *disk_super;
1275         struct dm_block *sblock, *copy;
1276         dm_block_t held_root;
1277
1278         r = superblock_lock(pmd, &sblock);
1279         if (r)
1280                 return r;
1281
1282         disk_super = dm_block_data(sblock);
1283         held_root = le64_to_cpu(disk_super->held_root);
1284         disk_super->held_root = cpu_to_le64(0);
1285
1286         dm_bm_unlock(sblock);
1287
1288         if (!held_root) {
1289                 DMWARN("No pool metadata snapshot found: nothing to release.");
1290                 return -EINVAL;
1291         }
1292
1293         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1294         if (r)
1295                 return r;
1296
1297         disk_super = dm_block_data(copy);
1298         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1299         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1300         dm_sm_dec_block(pmd->metadata_sm, held_root);
1301
1302         return dm_tm_unlock(pmd->tm, copy);
1303 }
1304
1305 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1306 {
1307         int r = -EINVAL;
1308
1309         down_write(&pmd->root_lock);
1310         if (!pmd->fail_io)
1311                 r = __release_metadata_snap(pmd);
1312         up_write(&pmd->root_lock);
1313
1314         return r;
1315 }
1316
1317 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1318                                dm_block_t *result)
1319 {
1320         int r;
1321         struct thin_disk_superblock *disk_super;
1322         struct dm_block *sblock;
1323
1324         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1325                             &sb_validator, &sblock);
1326         if (r)
1327                 return r;
1328
1329         disk_super = dm_block_data(sblock);
1330         *result = le64_to_cpu(disk_super->held_root);
1331
1332         return dm_bm_unlock(sblock);
1333 }
1334
1335 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1336                               dm_block_t *result)
1337 {
1338         int r = -EINVAL;
1339
1340         down_read(&pmd->root_lock);
1341         if (!pmd->fail_io)
1342                 r = __get_metadata_snap(pmd, result);
1343         up_read(&pmd->root_lock);
1344
1345         return r;
1346 }
1347
1348 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1349                              struct dm_thin_device **td)
1350 {
1351         int r = -EINVAL;
1352
1353         down_write(&pmd->root_lock);
1354         if (!pmd->fail_io)
1355                 r = __open_device(pmd, dev, 0, td);
1356         up_write(&pmd->root_lock);
1357
1358         return r;
1359 }
1360
1361 int dm_pool_close_thin_device(struct dm_thin_device *td)
1362 {
1363         down_write(&td->pmd->root_lock);
1364         __close_device(td);
1365         up_write(&td->pmd->root_lock);
1366
1367         return 0;
1368 }
1369
1370 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1371 {
1372         return td->id;
1373 }
1374
1375 /*
1376  * Check whether @time (of block creation) is older than @td's last snapshot.
1377  * If so then the associated block is shared with the last snapshot device.
1378  * Any block on a device created *after* the device last got snapshotted is
1379  * necessarily not shared.
1380  */
1381 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1382 {
1383         return td->snapshotted_time > time;
1384 }
1385
1386 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1387                        int can_issue_io, struct dm_thin_lookup_result *result)
1388 {
1389         int r;
1390         __le64 value;
1391         struct dm_pool_metadata *pmd = td->pmd;
1392         dm_block_t keys[2] = { td->id, block };
1393         struct dm_btree_info *info;
1394
1395         if (pmd->fail_io)
1396                 return -EINVAL;
1397
1398         down_read(&pmd->root_lock);
1399
1400         if (can_issue_io) {
1401                 info = &pmd->info;
1402         } else
1403                 info = &pmd->nb_info;
1404
1405         r = dm_btree_lookup(info, pmd->root, keys, &value);
1406         if (!r) {
1407                 uint64_t block_time = 0;
1408                 dm_block_t exception_block;
1409                 uint32_t exception_time;
1410
1411                 block_time = le64_to_cpu(value);
1412                 unpack_block_time(block_time, &exception_block,
1413                                   &exception_time);
1414                 result->block = exception_block;
1415                 result->shared = __snapshotted_since(td, exception_time);
1416         }
1417
1418         up_read(&pmd->root_lock);
1419         return r;
1420 }
1421
1422 static int __insert(struct dm_thin_device *td, dm_block_t block,
1423                     dm_block_t data_block)
1424 {
1425         int r, inserted;
1426         __le64 value;
1427         struct dm_pool_metadata *pmd = td->pmd;
1428         dm_block_t keys[2] = { td->id, block };
1429
1430         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1431         __dm_bless_for_disk(&value);
1432
1433         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1434                                    &pmd->root, &inserted);
1435         if (r)
1436                 return r;
1437
1438         td->changed = 1;
1439         if (inserted)
1440                 td->mapped_blocks++;
1441
1442         return 0;
1443 }
1444
1445 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1446                          dm_block_t data_block)
1447 {
1448         int r = -EINVAL;
1449
1450         down_write(&td->pmd->root_lock);
1451         if (!td->pmd->fail_io)
1452                 r = __insert(td, block, data_block);
1453         up_write(&td->pmd->root_lock);
1454
1455         return r;
1456 }
1457
1458 static int __remove(struct dm_thin_device *td, dm_block_t block)
1459 {
1460         int r;
1461         struct dm_pool_metadata *pmd = td->pmd;
1462         dm_block_t keys[2] = { td->id, block };
1463
1464         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1465         if (r)
1466                 return r;
1467
1468         td->mapped_blocks--;
1469         td->changed = 1;
1470
1471         return 0;
1472 }
1473
1474 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1475 {
1476         int r = -EINVAL;
1477
1478         down_write(&td->pmd->root_lock);
1479         if (!td->pmd->fail_io)
1480                 r = __remove(td, block);
1481         up_write(&td->pmd->root_lock);
1482
1483         return r;
1484 }
1485
1486 int dm_pool_block_is_used(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1487 {
1488         int r;
1489         uint32_t ref_count;
1490
1491         down_read(&pmd->root_lock);
1492         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1493         if (!r)
1494                 *result = (ref_count != 0);
1495         up_read(&pmd->root_lock);
1496
1497         return r;
1498 }
1499
1500 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1501 {
1502         int r;
1503
1504         down_read(&td->pmd->root_lock);
1505         r = td->changed;
1506         up_read(&td->pmd->root_lock);
1507
1508         return r;
1509 }
1510
1511 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1512 {
1513         bool r = false;
1514         struct dm_thin_device *td, *tmp;
1515
1516         down_read(&pmd->root_lock);
1517         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1518                 if (td->changed) {
1519                         r = td->changed;
1520                         break;
1521                 }
1522         }
1523         up_read(&pmd->root_lock);
1524
1525         return r;
1526 }
1527
1528 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1529 {
1530         bool r;
1531
1532         down_read(&td->pmd->root_lock);
1533         r = td->aborted_with_changes;
1534         up_read(&td->pmd->root_lock);
1535
1536         return r;
1537 }
1538
1539 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1540 {
1541         int r = -EINVAL;
1542
1543         down_write(&pmd->root_lock);
1544         if (!pmd->fail_io)
1545                 r = dm_sm_new_block(pmd->data_sm, result);
1546         up_write(&pmd->root_lock);
1547
1548         return r;
1549 }
1550
1551 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1552 {
1553         int r = -EINVAL;
1554
1555         down_write(&pmd->root_lock);
1556         if (pmd->fail_io)
1557                 goto out;
1558
1559         r = __commit_transaction(pmd);
1560         if (r <= 0)
1561                 goto out;
1562
1563         /*
1564          * Open the next transaction.
1565          */
1566         r = __begin_transaction(pmd);
1567 out:
1568         up_write(&pmd->root_lock);
1569         return r;
1570 }
1571
1572 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1573 {
1574         struct dm_thin_device *td;
1575
1576         list_for_each_entry(td, &pmd->thin_devices, list)
1577                 td->aborted_with_changes = td->changed;
1578 }
1579
1580 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1581 {
1582         int r = -EINVAL;
1583
1584         down_write(&pmd->root_lock);
1585         if (pmd->fail_io)
1586                 goto out;
1587
1588         __set_abort_with_changes_flags(pmd);
1589         __destroy_persistent_data_objects(pmd);
1590         r = __create_persistent_data_objects(pmd, false);
1591         if (r)
1592                 pmd->fail_io = true;
1593
1594 out:
1595         up_write(&pmd->root_lock);
1596
1597         return r;
1598 }
1599
1600 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1601 {
1602         int r = -EINVAL;
1603
1604         down_read(&pmd->root_lock);
1605         if (!pmd->fail_io)
1606                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1607         up_read(&pmd->root_lock);
1608
1609         return r;
1610 }
1611
1612 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1613                                           dm_block_t *result)
1614 {
1615         int r = -EINVAL;
1616
1617         down_read(&pmd->root_lock);
1618         if (!pmd->fail_io)
1619                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1620         up_read(&pmd->root_lock);
1621
1622         return r;
1623 }
1624
1625 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1626                                   dm_block_t *result)
1627 {
1628         int r = -EINVAL;
1629
1630         down_read(&pmd->root_lock);
1631         if (!pmd->fail_io)
1632                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1633         up_read(&pmd->root_lock);
1634
1635         return r;
1636 }
1637
1638 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1639 {
1640         int r = -EINVAL;
1641
1642         down_read(&pmd->root_lock);
1643         if (!pmd->fail_io)
1644                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1645         up_read(&pmd->root_lock);
1646
1647         return r;
1648 }
1649
1650 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1651 {
1652         int r = -EINVAL;
1653         struct dm_pool_metadata *pmd = td->pmd;
1654
1655         down_read(&pmd->root_lock);
1656         if (!pmd->fail_io) {
1657                 *result = td->mapped_blocks;
1658                 r = 0;
1659         }
1660         up_read(&pmd->root_lock);
1661
1662         return r;
1663 }
1664
1665 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1666 {
1667         int r;
1668         __le64 value_le;
1669         dm_block_t thin_root;
1670         struct dm_pool_metadata *pmd = td->pmd;
1671
1672         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1673         if (r)
1674                 return r;
1675
1676         thin_root = le64_to_cpu(value_le);
1677
1678         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1679 }
1680
1681 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1682                                      dm_block_t *result)
1683 {
1684         int r = -EINVAL;
1685         struct dm_pool_metadata *pmd = td->pmd;
1686
1687         down_read(&pmd->root_lock);
1688         if (!pmd->fail_io)
1689                 r = __highest_block(td, result);
1690         up_read(&pmd->root_lock);
1691
1692         return r;
1693 }
1694
1695 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1696 {
1697         int r;
1698         dm_block_t old_count;
1699
1700         r = dm_sm_get_nr_blocks(sm, &old_count);
1701         if (r)
1702                 return r;
1703
1704         if (new_count == old_count)
1705                 return 0;
1706
1707         if (new_count < old_count) {
1708                 DMERR("cannot reduce size of space map");
1709                 return -EINVAL;
1710         }
1711
1712         return dm_sm_extend(sm, new_count - old_count);
1713 }
1714
1715 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1716 {
1717         int r = -EINVAL;
1718
1719         down_write(&pmd->root_lock);
1720         if (!pmd->fail_io)
1721                 r = __resize_space_map(pmd->data_sm, new_count);
1722         up_write(&pmd->root_lock);
1723
1724         return r;
1725 }
1726
1727 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1728 {
1729         int r = -EINVAL;
1730
1731         down_write(&pmd->root_lock);
1732         if (!pmd->fail_io)
1733                 r = __resize_space_map(pmd->metadata_sm, new_count);
1734         up_write(&pmd->root_lock);
1735
1736         return r;
1737 }
1738
1739 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1740 {
1741         down_write(&pmd->root_lock);
1742         pmd->read_only = true;
1743         dm_bm_set_read_only(pmd->bm);
1744         up_write(&pmd->root_lock);
1745 }
1746
1747 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
1748 {
1749         down_write(&pmd->root_lock);
1750         pmd->read_only = false;
1751         dm_bm_set_read_write(pmd->bm);
1752         up_write(&pmd->root_lock);
1753 }
1754
1755 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
1756                                         dm_block_t threshold,
1757                                         dm_sm_threshold_fn fn,
1758                                         void *context)
1759 {
1760         int r;
1761
1762         down_write(&pmd->root_lock);
1763         r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
1764         up_write(&pmd->root_lock);
1765
1766         return r;
1767 }
1768
1769 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
1770 {
1771         int r;
1772         struct dm_block *sblock;
1773         struct thin_disk_superblock *disk_super;
1774
1775         down_write(&pmd->root_lock);
1776         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
1777
1778         r = superblock_lock(pmd, &sblock);
1779         if (r) {
1780                 DMERR("couldn't read superblock");
1781                 goto out;
1782         }
1783
1784         disk_super = dm_block_data(sblock);
1785         disk_super->flags = cpu_to_le32(pmd->flags);
1786
1787         dm_bm_unlock(sblock);
1788 out:
1789         up_write(&pmd->root_lock);
1790         return r;
1791 }
1792
1793 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
1794 {
1795         bool needs_check;
1796
1797         down_read(&pmd->root_lock);
1798         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
1799         up_read(&pmd->root_lock);
1800
1801         return needs_check;
1802 }
1803
1804 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
1805 {
1806         dm_tm_issue_prefetches(pmd->tm);
1807 }