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          * Set if a transaction has to be aborted but the attempt to roll back
 193          * to the previous (good) transaction failed.  The only pool metadata
 194          * operation possible in this state is the closing of the device.
 195          */
 196         bool fail_io:1;
 197
 198         /*
 199          * Reading the space map roots can fail, so we read it into these
 200          * buffers before the superblock is locked and updated.
 201          */
 202         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 203         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 204 };
 205
 206 struct dm_thin_device {
 207         struct list_head list;
 208         struct dm_pool_metadata *pmd;
 209         dm_thin_id id;
 210
 211         int open_count;
 212         bool changed:1;
 213         bool aborted_with_changes:1;
 214         uint64_t mapped_blocks;
 215         uint64_t transaction_id;
 216         uint32_t creation_time;
 217         uint32_t snapshotted_time;
 218 };
 219
 220 /*----------------------------------------------------------------
 221  * superblock validator
 222  *--------------------------------------------------------------*/
 223
 224 #define SUPERBLOCK_CSUM_XOR 160774
 225
 226 static void sb_prepare_for_write(struct dm_block_validator *v,
 227                                  struct dm_block *b,
 228                                  size_t block_size)
 229 {
 230         struct thin_disk_superblock *disk_super = dm_block_data(b);
 231
 232         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 233         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 234                                                       block_size - sizeof(__le32),
 235                                                       SUPERBLOCK_CSUM_XOR));
 236 }
 237
 238 static int sb_check(struct dm_block_validator *v,
 239                     struct dm_block *b,
 240                     size_t block_size)
 241 {
 242         struct thin_disk_superblock *disk_super = dm_block_data(b);
 243         __le32 csum_le;
 244
 245         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 246                 DMERR("sb_check failed: blocknr %llu: "
 247                       "wanted %llu", le64_to_cpu(disk_super->blocknr),
 248                       (unsigned long long)dm_block_location(b));
 249                 return -ENOTBLK;
 250         }
 251
 252         if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 253                 DMERR("sb_check failed: magic %llu: "
 254                       "wanted %llu", le64_to_cpu(disk_super->magic),
 255                       (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 256                 return -EILSEQ;
 257         }
 258
 259         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 260                                              block_size - sizeof(__le32),
 261                                              SUPERBLOCK_CSUM_XOR));
 262         if (csum_le != disk_super->csum) {
 263                 DMERR("sb_check failed: csum %u: wanted %u",
 264                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 265                 return -EILSEQ;
 266         }
 267
 268         return 0;
 269 }
 270
 271 static struct dm_block_validator sb_validator = {
 272         .name = "superblock",
 273         .prepare_for_write = sb_prepare_for_write,
 274         .check = sb_check
 275 };
 276
 277 /*----------------------------------------------------------------
 278  * Methods for the btree value types
 279  *--------------------------------------------------------------*/
 280
 281 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 282 {
 283         return (b << 24) | t;
 284 }
 285
 286 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 287 {
 288         *b = v >> 24;
 289         *t = v & ((1 << 24) - 1);
 290 }
 291
 292 static void data_block_inc(void *context, const void *value_le)
 293 {
 294         struct dm_space_map *sm = context;
 295         __le64 v_le;
 296         uint64_t b;
 297         uint32_t t;
 298
 299         memcpy(&v_le, value_le, sizeof(v_le));
 300         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 301         dm_sm_inc_block(sm, b);
 302 }
 303
 304 static void data_block_dec(void *context, const void *value_le)
 305 {
 306         struct dm_space_map *sm = context;
 307         __le64 v_le;
 308         uint64_t b;
 309         uint32_t t;
 310
 311         memcpy(&v_le, value_le, sizeof(v_le));
 312         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 313         dm_sm_dec_block(sm, b);
 314 }
 315
 316 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 317 {
 318         __le64 v1_le, v2_le;
 319         uint64_t b1, b2;
 320         uint32_t t;
 321
 322         memcpy(&v1_le, value1_le, sizeof(v1_le));
 323         memcpy(&v2_le, value2_le, sizeof(v2_le));
 324         unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 325         unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 326
 327         return b1 == b2;
 328 }
 329
 330 static void subtree_inc(void *context, const void *value)
 331 {
 332         struct dm_btree_info *info = context;
 333         __le64 root_le;
 334         uint64_t root;
 335
 336         memcpy(&root_le, value, sizeof(root_le));
 337         root = le64_to_cpu(root_le);
 338         dm_tm_inc(info->tm, root);
 339 }
 340
 341 static void subtree_dec(void *context, const void *value)
 342 {
 343         struct dm_btree_info *info = context;
 344         __le64 root_le;
 345         uint64_t root;
 346
 347         memcpy(&root_le, value, sizeof(root_le));
 348         root = le64_to_cpu(root_le);
 349         if (dm_btree_del(info, root))
 350                 DMERR("btree delete failed");
 351 }
 352
 353 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 354 {
 355         __le64 v1_le, v2_le;
 356         memcpy(&v1_le, value1_le, sizeof(v1_le));
 357         memcpy(&v2_le, value2_le, sizeof(v2_le));
 358
 359         return v1_le == v2_le;
 360 }
 361
 362 /*----------------------------------------------------------------*/
 363
 364 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 365                                 struct dm_block **sblock)
 366 {
 367         return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 368                                      &sb_validator, sblock);
 369 }
 370
 371 static int superblock_lock(struct dm_pool_metadata *pmd,
 372                            struct dm_block **sblock)
 373 {
 374         return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 375                                 &sb_validator, sblock);
 376 }
 377
 378 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 379 {
 380         int r;
 381         unsigned i;
 382         struct dm_block *b;
 383         __le64 *data_le, zero = cpu_to_le64(0);
 384         unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 385
 386         /*
 387          * We can't use a validator here - it may be all zeroes.
 388          */
 389         r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 390         if (r)
 391                 return r;
 392
 393         data_le = dm_block_data(b);
 394         *result = 1;
 395         for (i = 0; i < block_size; i++) {
 396                 if (data_le[i] != zero) {
 397                         *result = 0;
 398                         break;
 399                 }
 400         }
 401
 402         dm_bm_unlock(b);
 403
 404         return 0;
 405 }
 406
 407 static void __setup_btree_details(struct dm_pool_metadata *pmd)
 408 {
 409         pmd->info.tm = pmd->tm;
 410         pmd->info.levels = 2;
 411         pmd->info.value_type.context = pmd->data_sm;
 412         pmd->info.value_type.size = sizeof(__le64);
 413         pmd->info.value_type.inc = data_block_inc;
 414         pmd->info.value_type.dec = data_block_dec;
 415         pmd->info.value_type.equal = data_block_equal;
 416
 417         memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 418         pmd->nb_info.tm = pmd->nb_tm;
 419
 420         pmd->tl_info.tm = pmd->tm;
 421         pmd->tl_info.levels = 1;
 422         pmd->tl_info.value_type.context = &pmd->bl_info;
 423         pmd->tl_info.value_type.size = sizeof(__le64);
 424         pmd->tl_info.value_type.inc = subtree_inc;
 425         pmd->tl_info.value_type.dec = subtree_dec;
 426         pmd->tl_info.value_type.equal = subtree_equal;
 427
 428         pmd->bl_info.tm = pmd->tm;
 429         pmd->bl_info.levels = 1;
 430         pmd->bl_info.value_type.context = pmd->data_sm;
 431         pmd->bl_info.value_type.size = sizeof(__le64);
 432         pmd->bl_info.value_type.inc = data_block_inc;
 433         pmd->bl_info.value_type.dec = data_block_dec;
 434         pmd->bl_info.value_type.equal = data_block_equal;
 435
 436         pmd->details_info.tm = pmd->tm;
 437         pmd->details_info.levels = 1;
 438         pmd->details_info.value_type.context = NULL;
 439         pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 440         pmd->details_info.value_type.inc = NULL;
 441         pmd->details_info.value_type.dec = NULL;
 442         pmd->details_info.value_type.equal = NULL;
 443 }
 444
 445 static int save_sm_roots(struct dm_pool_metadata *pmd)
 446 {
 447         int r;
 448         size_t len;
 449
 450         r = dm_sm_root_size(pmd->metadata_sm, &len);
 451         if (r < 0)
 452                 return r;
 453
 454         r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 455         if (r < 0)
 456                 return r;
 457
 458         r = dm_sm_root_size(pmd->data_sm, &len);
 459         if (r < 0)
 460                 return r;
 461
 462         return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 463 }
 464
 465 static void copy_sm_roots(struct dm_pool_metadata *pmd,
 466                           struct thin_disk_superblock *disk)
 467 {
 468         memcpy(&disk->metadata_space_map_root,
 469                &pmd->metadata_space_map_root,
 470                sizeof(pmd->metadata_space_map_root));
 471
 472         memcpy(&disk->data_space_map_root,
 473                &pmd->data_space_map_root,
 474                sizeof(pmd->data_space_map_root));
 475 }
 476
 477 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 478 {
 479         int r;
 480         struct dm_block *sblock;
 481         struct thin_disk_superblock *disk_super;
 482         sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 483
 484         if (bdev_size > THIN_METADATA_MAX_SECTORS)
 485                 bdev_size = THIN_METADATA_MAX_SECTORS;
 486
 487         r = dm_sm_commit(pmd->data_sm);
 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 = save_sm_roots(pmd);
 496         if (r < 0)
 497                 return r;
 498
 499         r = superblock_lock_zero(pmd, &sblock);
 500         if (r)
 501                 return r;
 502
 503         disk_super = dm_block_data(sblock);
 504         disk_super->flags = 0;
 505         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 506         disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 507         disk_super->version = cpu_to_le32(THIN_VERSION);
 508         disk_super->time = 0;
 509         disk_super->trans_id = 0;
 510         disk_super->held_root = 0;
 511
 512         copy_sm_roots(pmd, disk_super);
 513
 514         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 515         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 516         disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 517         disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 518         disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 519
 520         return dm_tm_commit(pmd->tm, sblock);
 521 }
 522
 523 static int __format_metadata(struct dm_pool_metadata *pmd)
 524 {
 525         int r;
 526
 527         r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 528                                  &pmd->tm, &pmd->metadata_sm);
 529         if (r < 0) {
 530                 DMERR("tm_create_with_sm failed");
 531                 return r;
 532         }
 533
 534         pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 535         if (IS_ERR(pmd->data_sm)) {
 536                 DMERR("sm_disk_create failed");
 537                 r = PTR_ERR(pmd->data_sm);
 538                 goto bad_cleanup_tm;
 539         }
 540
 541         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 542         if (!pmd->nb_tm) {
 543                 DMERR("could not create non-blocking clone tm");
 544                 r = -ENOMEM;
 545                 goto bad_cleanup_data_sm;
 546         }
 547
 548         __setup_btree_details(pmd);
 549
 550         r = dm_btree_empty(&pmd->info, &pmd->root);
 551         if (r < 0)
 552                 goto bad_cleanup_nb_tm;
 553
 554         r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 555         if (r < 0) {
 556                 DMERR("couldn't create devices root");
 557                 goto bad_cleanup_nb_tm;
 558         }
 559
 560         r = __write_initial_superblock(pmd);
 561         if (r)
 562                 goto bad_cleanup_nb_tm;
 563
 564         return 0;
 565
 566 bad_cleanup_nb_tm:
 567         dm_tm_destroy(pmd->nb_tm);
 568 bad_cleanup_data_sm:
 569         dm_sm_destroy(pmd->data_sm);
 570 bad_cleanup_tm:
 571         dm_tm_destroy(pmd->tm);
 572         dm_sm_destroy(pmd->metadata_sm);
 573
 574         return r;
 575 }
 576
 577 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 578                                      struct dm_pool_metadata *pmd)
 579 {
 580         uint32_t features;
 581
 582         features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 583         if (features) {
 584                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 585                       (unsigned long)features);
 586                 return -EINVAL;
 587         }
 588
 589         /*
 590          * Check for read-only metadata to skip the following RDWR checks.
 591          */
 592         if (get_disk_ro(pmd->bdev->bd_disk))
 593                 return 0;
 594
 595         features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 596         if (features) {
 597                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 598                       (unsigned long)features);
 599                 return -EINVAL;
 600         }
 601
 602         return 0;
 603 }
 604
 605 static int __open_metadata(struct dm_pool_metadata *pmd)
 606 {
 607         int r;
 608         struct dm_block *sblock;
 609         struct thin_disk_superblock *disk_super;
 610
 611         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 612                             &sb_validator, &sblock);
 613         if (r < 0) {
 614                 DMERR("couldn't read superblock");
 615                 return r;
 616         }
 617
 618         disk_super = dm_block_data(sblock);
 619
 620         /* Verify the data block size hasn't changed */
 621         if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 622                 DMERR("changing the data block size (from %u to %llu) is not supported",
 623                       le32_to_cpu(disk_super->data_block_size),
 624                       (unsigned long long)pmd->data_block_size);
 625                 r = -EINVAL;
 626                 goto bad_unlock_sblock;
 627         }
 628
 629         r = __check_incompat_features(disk_super, pmd);
 630         if (r < 0)
 631                 goto bad_unlock_sblock;
 632
 633         r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 634                                disk_super->metadata_space_map_root,
 635                                sizeof(disk_super->metadata_space_map_root),
 636                                &pmd->tm, &pmd->metadata_sm);
 637         if (r < 0) {
 638                 DMERR("tm_open_with_sm failed");
 639                 goto bad_unlock_sblock;
 640         }
 641
 642         pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 643                                        sizeof(disk_super->data_space_map_root));
 644         if (IS_ERR(pmd->data_sm)) {
 645                 DMERR("sm_disk_open failed");
 646                 r = PTR_ERR(pmd->data_sm);
 647                 goto bad_cleanup_tm;
 648         }
 649
 650         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 651         if (!pmd->nb_tm) {
 652                 DMERR("could not create non-blocking clone tm");
 653                 r = -ENOMEM;
 654                 goto bad_cleanup_data_sm;
 655         }
 656
 657         __setup_btree_details(pmd);
 658         dm_bm_unlock(sblock);
 659
 660         return 0;
 661
 662 bad_cleanup_data_sm:
 663         dm_sm_destroy(pmd->data_sm);
 664 bad_cleanup_tm:
 665         dm_tm_destroy(pmd->tm);
 666         dm_sm_destroy(pmd->metadata_sm);
 667 bad_unlock_sblock:
 668         dm_bm_unlock(sblock);
 669
 670         return r;
 671 }
 672
 673 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 674 {
 675         int r, unformatted;
 676
 677         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 678         if (r)
 679                 return r;
 680
 681         if (unformatted)
 682                 return format_device ? __format_metadata(pmd) : -EPERM;
 683
 684         return __open_metadata(pmd);
 685 }
 686
 687 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 688 {
 689         int r;
 690
 691         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 692                                           THIN_MAX_CONCURRENT_LOCKS);
 693         if (IS_ERR(pmd->bm)) {
 694                 DMERR("could not create block manager");
 695                 return PTR_ERR(pmd->bm);
 696         }
 697
 698         r = __open_or_format_metadata(pmd, format_device);
 699         if (r)
 700                 dm_block_manager_destroy(pmd->bm);
 701
 702         return r;
 703 }
 704
 705 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 706 {
 707         dm_sm_destroy(pmd->data_sm);
 708         dm_sm_destroy(pmd->metadata_sm);
 709         dm_tm_destroy(pmd->nb_tm);
 710         dm_tm_destroy(pmd->tm);
 711         dm_block_manager_destroy(pmd->bm);
 712 }
 713
 714 static int __begin_transaction(struct dm_pool_metadata *pmd)
 715 {
 716         int r;
 717         struct thin_disk_superblock *disk_super;
 718         struct dm_block *sblock;
 719
 720         /*
 721          * We re-read the superblock every time.  Shouldn't need to do this
 722          * really.
 723          */
 724         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 725                             &sb_validator, &sblock);
 726         if (r)
 727                 return r;
 728
 729         disk_super = dm_block_data(sblock);
 730         pmd->time = le32_to_cpu(disk_super->time);
 731         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 732         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 733         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 734         pmd->flags = le32_to_cpu(disk_super->flags);
 735         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 736
 737         dm_bm_unlock(sblock);
 738         return 0;
 739 }
 740
 741 static int __write_changed_details(struct dm_pool_metadata *pmd)
 742 {
 743         int r;
 744         struct dm_thin_device *td, *tmp;
 745         struct disk_device_details details;
 746         uint64_t key;
 747
 748         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 749                 if (!td->changed)
 750                         continue;
 751
 752                 key = td->id;
 753
 754                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 755                 details.transaction_id = cpu_to_le64(td->transaction_id);
 756                 details.creation_time = cpu_to_le32(td->creation_time);
 757                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 758                 __dm_bless_for_disk(&details);
 759
 760                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 761                                     &key, &details, &pmd->details_root);
 762                 if (r)
 763                         return r;
 764
 765                 if (td->open_count)
 766                         td->changed = 0;
 767                 else {
 768                         list_del(&td->list);
 769                         kfree(td);
 770                 }
 771         }
 772
 773         return 0;
 774 }
 775
 776 static int __commit_transaction(struct dm_pool_metadata *pmd)
 777 {
 778         int r;
 779         size_t metadata_len, data_len;
 780         struct thin_disk_superblock *disk_super;
 781         struct dm_block *sblock;
 782
 783         /*
 784          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 785          */
 786         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 787
 788         r = __write_changed_details(pmd);
 789         if (r < 0)
 790                 return r;
 791
 792         r = dm_sm_commit(pmd->data_sm);
 793         if (r < 0)
 794                 return r;
 795
 796         r = dm_tm_pre_commit(pmd->tm);
 797         if (r < 0)
 798                 return r;
 799
 800         r = dm_sm_root_size(pmd->metadata_sm, &metadata_len);
 801         if (r < 0)
 802                 return r;
 803
 804         r = dm_sm_root_size(pmd->data_sm, &data_len);
 805         if (r < 0)
 806                 return r;
 807
 808         r = save_sm_roots(pmd);
 809         if (r < 0)
 810                 return r;
 811
 812         r = superblock_lock(pmd, &sblock);
 813         if (r)
 814                 return r;
 815
 816         disk_super = dm_block_data(sblock);
 817         disk_super->time = cpu_to_le32(pmd->time);
 818         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 819         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 820         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 821         disk_super->flags = cpu_to_le32(pmd->flags);
 822
 823         copy_sm_roots(pmd, disk_super);
 824
 825         return dm_tm_commit(pmd->tm, sblock);
 826 }
 827
 828 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 829                                                sector_t data_block_size,
 830                                                bool format_device)
 831 {
 832         int r;
 833         struct dm_pool_metadata *pmd;
 834
 835         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 836         if (!pmd) {
 837                 DMERR("could not allocate metadata struct");
 838                 return ERR_PTR(-ENOMEM);
 839         }
 840
 841         init_rwsem(&pmd->root_lock);
 842         pmd->time = 0;
 843         INIT_LIST_HEAD(&pmd->thin_devices);
 844         pmd->fail_io = false;
 845         pmd->bdev = bdev;
 846         pmd->data_block_size = data_block_size;
 847
 848         r = __create_persistent_data_objects(pmd, format_device);
 849         if (r) {
 850                 kfree(pmd);
 851                 return ERR_PTR(r);
 852         }
 853
 854         r = __begin_transaction(pmd);
 855         if (r < 0) {
 856                 if (dm_pool_metadata_close(pmd) < 0)
 857                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 858                 return ERR_PTR(r);
 859         }
 860
 861         return pmd;
 862 }
 863
 864 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 865 {
 866         int r;
 867         unsigned open_devices = 0;
 868         struct dm_thin_device *td, *tmp;
 869
 870         down_read(&pmd->root_lock);
 871         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 872                 if (td->open_count)
 873                         open_devices++;
 874                 else {
 875                         list_del(&td->list);
 876                         kfree(td);
 877                 }
 878         }
 879         up_read(&pmd->root_lock);
 880
 881         if (open_devices) {
 882                 DMERR("attempt to close pmd when %u device(s) are still open",
 883                        open_devices);
 884                 return -EBUSY;
 885         }
 886
 887         if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
 888                 r = __commit_transaction(pmd);
 889                 if (r < 0)
 890                         DMWARN("%s: __commit_transaction() failed, error = %d",
 891                                __func__, r);
 892         }
 893
 894         if (!pmd->fail_io)
 895                 __destroy_persistent_data_objects(pmd);
 896
 897         kfree(pmd);
 898         return 0;
 899 }
 900
 901 /*
 902  * __open_device: Returns @td corresponding to device with id @dev,
 903  * creating it if @create is set and incrementing @td->open_count.
 904  * On failure, @td is undefined.
 905  */
 906 static int __open_device(struct dm_pool_metadata *pmd,
 907                          dm_thin_id dev, int create,
 908                          struct dm_thin_device **td)
 909 {
 910         int r, changed = 0;
 911         struct dm_thin_device *td2;
 912         uint64_t key = dev;
 913         struct disk_device_details details_le;
 914
 915         /*
 916          * If the device is already open, return it.
 917          */
 918         list_for_each_entry(td2, &pmd->thin_devices, list)
 919                 if (td2->id == dev) {
 920                         /*
 921                          * May not create an already-open device.
 922                          */
 923                         if (create)
 924                                 return -EEXIST;
 925
 926                         td2->open_count++;
 927                         *td = td2;
 928                         return 0;
 929                 }
 930
 931         /*
 932          * Check the device exists.
 933          */
 934         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 935                             &key, &details_le);
 936         if (r) {
 937                 if (r != -ENODATA || !create)
 938                         return r;
 939
 940                 /*
 941                  * Create new device.
 942                  */
 943                 changed = 1;
 944                 details_le.mapped_blocks = 0;
 945                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
 946                 details_le.creation_time = cpu_to_le32(pmd->time);
 947                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
 948         }
 949
 950         *td = kmalloc(sizeof(**td), GFP_NOIO);
 951         if (!*td)
 952                 return -ENOMEM;
 953
 954         (*td)->pmd = pmd;
 955         (*td)->id = dev;
 956         (*td)->open_count = 1;
 957         (*td)->changed = changed;
 958         (*td)->aborted_with_changes = false;
 959         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
 960         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
 961         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
 962         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
 963
 964         list_add(&(*td)->list, &pmd->thin_devices);
 965
 966         return 0;
 967 }
 968
 969 static void __close_device(struct dm_thin_device *td)
 970 {
 971         --td->open_count;
 972 }
 973
 974 static int __create_thin(struct dm_pool_metadata *pmd,
 975                          dm_thin_id dev)
 976 {
 977         int r;
 978         dm_block_t dev_root;
 979         uint64_t key = dev;
 980         struct disk_device_details details_le;
 981         struct dm_thin_device *td;
 982         __le64 value;
 983
 984         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 985                             &key, &details_le);
 986         if (!r)
 987                 return -EEXIST;
 988
 989         /*
 990          * Create an empty btree for the mappings.
 991          */
 992         r = dm_btree_empty(&pmd->bl_info, &dev_root);
 993         if (r)
 994                 return r;
 995
 996         /*
 997          * Insert it into the main mapping tree.
 998          */
 999         value = cpu_to_le64(dev_root);
1000         __dm_bless_for_disk(&value);
1001         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1002         if (r) {
1003                 dm_btree_del(&pmd->bl_info, dev_root);
1004                 return r;
1005         }
1006
1007         r = __open_device(pmd, dev, 1, &td);
1008         if (r) {
1009                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1010                 dm_btree_del(&pmd->bl_info, dev_root);
1011                 return r;
1012         }
1013         __close_device(td);
1014
1015         return r;
1016 }
1017
1018 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1019 {
1020         int r = -EINVAL;
1021
1022         down_write(&pmd->root_lock);
1023         if (!pmd->fail_io)
1024                 r = __create_thin(pmd, dev);
1025         up_write(&pmd->root_lock);
1026
1027         return r;
1028 }
1029
1030 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1031                                   struct dm_thin_device *snap,
1032                                   dm_thin_id origin, uint32_t time)
1033 {
1034         int r;
1035         struct dm_thin_device *td;
1036
1037         r = __open_device(pmd, origin, 0, &td);
1038         if (r)
1039                 return r;
1040
1041         td->changed = 1;
1042         td->snapshotted_time = time;
1043
1044         snap->mapped_blocks = td->mapped_blocks;
1045         snap->snapshotted_time = time;
1046         __close_device(td);
1047
1048         return 0;
1049 }
1050
1051 static int __create_snap(struct dm_pool_metadata *pmd,
1052                          dm_thin_id dev, dm_thin_id origin)
1053 {
1054         int r;
1055         dm_block_t origin_root;
1056         uint64_t key = origin, dev_key = dev;
1057         struct dm_thin_device *td;
1058         struct disk_device_details details_le;
1059         __le64 value;
1060
1061         /* check this device is unused */
1062         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1063                             &dev_key, &details_le);
1064         if (!r)
1065                 return -EEXIST;
1066
1067         /* find the mapping tree for the origin */
1068         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1069         if (r)
1070                 return r;
1071         origin_root = le64_to_cpu(value);
1072
1073         /* clone the origin, an inc will do */
1074         dm_tm_inc(pmd->tm, origin_root);
1075
1076         /* insert into the main mapping tree */
1077         value = cpu_to_le64(origin_root);
1078         __dm_bless_for_disk(&value);
1079         key = dev;
1080         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1081         if (r) {
1082                 dm_tm_dec(pmd->tm, origin_root);
1083                 return r;
1084         }
1085
1086         pmd->time++;
1087
1088         r = __open_device(pmd, dev, 1, &td);
1089         if (r)
1090                 goto bad;
1091
1092         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1093         __close_device(td);
1094
1095         if (r)
1096                 goto bad;
1097
1098         return 0;
1099
1100 bad:
1101         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1102         dm_btree_remove(&pmd->details_info, pmd->details_root,
1103                         &key, &pmd->details_root);
1104         return r;
1105 }
1106
1107 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1108                                  dm_thin_id dev,
1109                                  dm_thin_id origin)
1110 {
1111         int r = -EINVAL;
1112
1113         down_write(&pmd->root_lock);
1114         if (!pmd->fail_io)
1115                 r = __create_snap(pmd, dev, origin);
1116         up_write(&pmd->root_lock);
1117
1118         return r;
1119 }
1120
1121 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1122 {
1123         int r;
1124         uint64_t key = dev;
1125         struct dm_thin_device *td;
1126
1127         /* TODO: failure should mark the transaction invalid */
1128         r = __open_device(pmd, dev, 0, &td);
1129         if (r)
1130                 return r;
1131
1132         if (td->open_count > 1) {
1133                 __close_device(td);
1134                 return -EBUSY;
1135         }
1136
1137         list_del(&td->list);
1138         kfree(td);
1139         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1140                             &key, &pmd->details_root);
1141         if (r)
1142                 return r;
1143
1144         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1145         if (r)
1146                 return r;
1147
1148         return 0;
1149 }
1150
1151 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1152                                dm_thin_id dev)
1153 {
1154         int r = -EINVAL;
1155
1156         down_write(&pmd->root_lock);
1157         if (!pmd->fail_io)
1158                 r = __delete_device(pmd, dev);
1159         up_write(&pmd->root_lock);
1160
1161         return r;
1162 }
1163
1164 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1165                                         uint64_t current_id,
1166                                         uint64_t new_id)
1167 {
1168         int r = -EINVAL;
1169
1170         down_write(&pmd->root_lock);
1171
1172         if (pmd->fail_io)
1173                 goto out;
1174
1175         if (pmd->trans_id != current_id) {
1176                 DMERR("mismatched transaction id");
1177                 goto out;
1178         }
1179
1180         pmd->trans_id = new_id;
1181         r = 0;
1182
1183 out:
1184         up_write(&pmd->root_lock);
1185
1186         return r;
1187 }
1188
1189 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1190                                         uint64_t *result)
1191 {
1192         int r = -EINVAL;
1193
1194         down_read(&pmd->root_lock);
1195         if (!pmd->fail_io) {
1196                 *result = pmd->trans_id;
1197                 r = 0;
1198         }
1199         up_read(&pmd->root_lock);
1200
1201         return r;
1202 }
1203
1204 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1205 {
1206         int r, inc;
1207         struct thin_disk_superblock *disk_super;
1208         struct dm_block *copy, *sblock;
1209         dm_block_t held_root;
1210
1211         /*
1212          * We commit to ensure the btree roots which we increment in a
1213          * moment are up to date.
1214          */
1215         __commit_transaction(pmd);
1216
1217         /*
1218          * Copy the superblock.
1219          */
1220         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1221         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1222                                &sb_validator, &copy, &inc);
1223         if (r)
1224                 return r;
1225
1226         BUG_ON(!inc);
1227
1228         held_root = dm_block_location(copy);
1229         disk_super = dm_block_data(copy);
1230
1231         if (le64_to_cpu(disk_super->held_root)) {
1232                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1233
1234                 dm_tm_dec(pmd->tm, held_root);
1235                 dm_tm_unlock(pmd->tm, copy);
1236                 return -EBUSY;
1237         }
1238
1239         /*
1240          * Wipe the spacemap since we're not publishing this.
1241          */
1242         memset(&disk_super->data_space_map_root, 0,
1243                sizeof(disk_super->data_space_map_root));
1244         memset(&disk_super->metadata_space_map_root, 0,
1245                sizeof(disk_super->metadata_space_map_root));
1246
1247         /*
1248          * Increment the data structures that need to be preserved.
1249          */
1250         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1251         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1252         dm_tm_unlock(pmd->tm, copy);
1253
1254         /*
1255          * Write the held root into the superblock.
1256          */
1257         r = superblock_lock(pmd, &sblock);
1258         if (r) {
1259                 dm_tm_dec(pmd->tm, held_root);
1260                 return r;
1261         }
1262
1263         disk_super = dm_block_data(sblock);
1264         disk_super->held_root = cpu_to_le64(held_root);
1265         dm_bm_unlock(sblock);
1266         return 0;
1267 }
1268
1269 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1270 {
1271         int r = -EINVAL;
1272
1273         down_write(&pmd->root_lock);
1274         if (!pmd->fail_io)
1275                 r = __reserve_metadata_snap(pmd);
1276         up_write(&pmd->root_lock);
1277
1278         return r;
1279 }
1280
1281 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1282 {
1283         int r;
1284         struct thin_disk_superblock *disk_super;
1285         struct dm_block *sblock, *copy;
1286         dm_block_t held_root;
1287
1288         r = superblock_lock(pmd, &sblock);
1289         if (r)
1290                 return r;
1291
1292         disk_super = dm_block_data(sblock);
1293         held_root = le64_to_cpu(disk_super->held_root);
1294         disk_super->held_root = cpu_to_le64(0);
1295
1296         dm_bm_unlock(sblock);
1297
1298         if (!held_root) {
1299                 DMWARN("No pool metadata snapshot found: nothing to release.");
1300                 return -EINVAL;
1301         }
1302
1303         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1304         if (r)
1305                 return r;
1306
1307         disk_super = dm_block_data(copy);
1308         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1309         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1310         dm_sm_dec_block(pmd->metadata_sm, held_root);
1311
1312         dm_tm_unlock(pmd->tm, copy);
1313
1314         return 0;
1315 }
1316
1317 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1318 {
1319         int r = -EINVAL;
1320
1321         down_write(&pmd->root_lock);
1322         if (!pmd->fail_io)
1323                 r = __release_metadata_snap(pmd);
1324         up_write(&pmd->root_lock);
1325
1326         return r;
1327 }
1328
1329 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1330                                dm_block_t *result)
1331 {
1332         int r;
1333         struct thin_disk_superblock *disk_super;
1334         struct dm_block *sblock;
1335
1336         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1337                             &sb_validator, &sblock);
1338         if (r)
1339                 return r;
1340
1341         disk_super = dm_block_data(sblock);
1342         *result = le64_to_cpu(disk_super->held_root);
1343
1344         dm_bm_unlock(sblock);
1345
1346         return 0;
1347 }
1348
1349 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1350                               dm_block_t *result)
1351 {
1352         int r = -EINVAL;
1353
1354         down_read(&pmd->root_lock);
1355         if (!pmd->fail_io)
1356                 r = __get_metadata_snap(pmd, result);
1357         up_read(&pmd->root_lock);
1358
1359         return r;
1360 }
1361
1362 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1363                              struct dm_thin_device **td)
1364 {
1365         int r = -EINVAL;
1366
1367         down_write(&pmd->root_lock);
1368         if (!pmd->fail_io)
1369                 r = __open_device(pmd, dev, 0, td);
1370         up_write(&pmd->root_lock);
1371
1372         return r;
1373 }
1374
1375 int dm_pool_close_thin_device(struct dm_thin_device *td)
1376 {
1377         down_write(&td->pmd->root_lock);
1378         __close_device(td);
1379         up_write(&td->pmd->root_lock);
1380
1381         return 0;
1382 }
1383
1384 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1385 {
1386         return td->id;
1387 }
1388
1389 /*
1390  * Check whether @time (of block creation) is older than @td's last snapshot.
1391  * If so then the associated block is shared with the last snapshot device.
1392  * Any block on a device created *after* the device last got snapshotted is
1393  * necessarily not shared.
1394  */
1395 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1396 {
1397         return td->snapshotted_time > time;
1398 }
1399
1400 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1401                                  struct dm_thin_lookup_result *result)
1402 {
1403         uint64_t block_time = 0;
1404         dm_block_t exception_block;
1405         uint32_t exception_time;
1406
1407         block_time = le64_to_cpu(value);
1408         unpack_block_time(block_time, &exception_block, &exception_time);
1409         result->block = exception_block;
1410         result->shared = __snapshotted_since(td, exception_time);
1411 }
1412
1413 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1414                         int can_issue_io, struct dm_thin_lookup_result *result)
1415 {
1416         int r;
1417         __le64 value;
1418         struct dm_pool_metadata *pmd = td->pmd;
1419         dm_block_t keys[2] = { td->id, block };
1420         struct dm_btree_info *info;
1421
1422         if (can_issue_io) {
1423                 info = &pmd->info;
1424         } else
1425                 info = &pmd->nb_info;
1426
1427         r = dm_btree_lookup(info, pmd->root, keys, &value);
1428         if (!r)
1429                 unpack_lookup_result(td, value, result);
1430
1431         return r;
1432 }
1433
1434 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1435                        int can_issue_io, struct dm_thin_lookup_result *result)
1436 {
1437         int r;
1438         struct dm_pool_metadata *pmd = td->pmd;
1439
1440         down_read(&pmd->root_lock);
1441         if (pmd->fail_io) {
1442                 up_read(&pmd->root_lock);
1443                 return -EINVAL;
1444         }
1445
1446         r = __find_block(td, block, can_issue_io, result);
1447
1448         up_read(&pmd->root_lock);
1449         return r;
1450 }
1451
1452 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1453                                           dm_block_t *vblock,
1454                                           struct dm_thin_lookup_result *result)
1455 {
1456         int r;
1457         __le64 value;
1458         struct dm_pool_metadata *pmd = td->pmd;
1459         dm_block_t keys[2] = { td->id, block };
1460
1461         r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1462         if (!r)
1463                 unpack_lookup_result(td, value, result);
1464
1465         return r;
1466 }
1467
1468 static int __find_mapped_range(struct dm_thin_device *td,
1469                                dm_block_t begin, dm_block_t end,
1470                                dm_block_t *thin_begin, dm_block_t *thin_end,
1471                                dm_block_t *pool_begin, bool *maybe_shared)
1472 {
1473         int r;
1474         dm_block_t pool_end;
1475         struct dm_thin_lookup_result lookup;
1476
1477         if (end < begin)
1478                 return -ENODATA;
1479
1480         r = __find_next_mapped_block(td, begin, &begin, &lookup);
1481         if (r)
1482                 return r;
1483
1484         if (begin >= end)
1485                 return -ENODATA;
1486
1487         *thin_begin = begin;
1488         *pool_begin = lookup.block;
1489         *maybe_shared = lookup.shared;
1490
1491         begin++;
1492         pool_end = *pool_begin + 1;
1493         while (begin != end) {
1494                 r = __find_block(td, begin, true, &lookup);
1495                 if (r) {
1496                         if (r == -ENODATA)
1497                                 break;
1498                         else
1499                                 return r;
1500                 }
1501
1502                 if ((lookup.block != pool_end) ||
1503                     (lookup.shared != *maybe_shared))
1504                         break;
1505
1506                 pool_end++;
1507                 begin++;
1508         }
1509
1510         *thin_end = begin;
1511         return 0;
1512 }
1513
1514 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1515                               dm_block_t begin, dm_block_t end,
1516                               dm_block_t *thin_begin, dm_block_t *thin_end,
1517                               dm_block_t *pool_begin, bool *maybe_shared)
1518 {
1519         int r = -EINVAL;
1520         struct dm_pool_metadata *pmd = td->pmd;
1521
1522         down_read(&pmd->root_lock);
1523         if (!pmd->fail_io) {
1524                 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1525                                         pool_begin, maybe_shared);
1526         }
1527         up_read(&pmd->root_lock);
1528
1529         return r;
1530 }
1531
1532 static int __insert(struct dm_thin_device *td, dm_block_t block,
1533                     dm_block_t data_block)
1534 {
1535         int r, inserted;
1536         __le64 value;
1537         struct dm_pool_metadata *pmd = td->pmd;
1538         dm_block_t keys[2] = { td->id, block };
1539
1540         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1541         __dm_bless_for_disk(&value);
1542
1543         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1544                                    &pmd->root, &inserted);
1545         if (r)
1546                 return r;
1547
1548         td->changed = 1;
1549         if (inserted)
1550                 td->mapped_blocks++;
1551
1552         return 0;
1553 }
1554
1555 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1556                          dm_block_t data_block)
1557 {
1558         int r = -EINVAL;
1559
1560         down_write(&td->pmd->root_lock);
1561         if (!td->pmd->fail_io)
1562                 r = __insert(td, block, data_block);
1563         up_write(&td->pmd->root_lock);
1564
1565         return r;
1566 }
1567
1568 static int __remove(struct dm_thin_device *td, dm_block_t block)
1569 {
1570         int r;
1571         struct dm_pool_metadata *pmd = td->pmd;
1572         dm_block_t keys[2] = { td->id, block };
1573
1574         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1575         if (r)
1576                 return r;
1577
1578         td->mapped_blocks--;
1579         td->changed = 1;
1580
1581         return 0;
1582 }
1583
1584 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1585 {
1586         int r;
1587         unsigned count, total_count = 0;
1588         struct dm_pool_metadata *pmd = td->pmd;
1589         dm_block_t keys[1] = { td->id };
1590         __le64 value;
1591         dm_block_t mapping_root;
1592
1593         /*
1594          * Find the mapping tree
1595          */
1596         r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1597         if (r)
1598                 return r;
1599
1600         /*
1601          * Remove from the mapping tree, taking care to inc the
1602          * ref count so it doesn't get deleted.
1603          */
1604         mapping_root = le64_to_cpu(value);
1605         dm_tm_inc(pmd->tm, mapping_root);
1606         r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1607         if (r)
1608                 return r;
1609
1610         /*
1611          * Remove leaves stops at the first unmapped entry, so we have to
1612          * loop round finding mapped ranges.
1613          */
1614         while (begin < end) {
1615                 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1616                 if (r == -ENODATA)
1617                         break;
1618
1619                 if (r)
1620                         return r;
1621
1622                 if (begin >= end)
1623                         break;
1624
1625                 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1626                 if (r)
1627                         return r;
1628
1629                 total_count += count;
1630         }
1631
1632         td->mapped_blocks -= total_count;
1633         td->changed = 1;
1634
1635         /*
1636          * Reinsert the mapping tree.
1637          */
1638         value = cpu_to_le64(mapping_root);
1639         __dm_bless_for_disk(&value);
1640         return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1641 }
1642
1643 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1644 {
1645         int r = -EINVAL;
1646
1647         down_write(&td->pmd->root_lock);
1648         if (!td->pmd->fail_io)
1649                 r = __remove(td, block);
1650         up_write(&td->pmd->root_lock);
1651
1652         return r;
1653 }
1654
1655 int dm_thin_remove_range(struct dm_thin_device *td,
1656                          dm_block_t begin, dm_block_t end)
1657 {
1658         int r = -EINVAL;
1659
1660         down_write(&td->pmd->root_lock);
1661         if (!td->pmd->fail_io)
1662                 r = __remove_range(td, begin, end);
1663         up_write(&td->pmd->root_lock);
1664
1665         return r;
1666 }
1667
1668 int dm_pool_block_is_used(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1669 {
1670         int r;
1671         uint32_t ref_count;
1672
1673         down_read(&pmd->root_lock);
1674         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1675         if (!r)
1676                 *result = (ref_count != 0);
1677         up_read(&pmd->root_lock);
1678
1679         return r;
1680 }
1681
1682 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1683 {
1684         int r = 0;
1685
1686         down_write(&pmd->root_lock);
1687         for (; b != e; b++) {
1688                 r = dm_sm_inc_block(pmd->data_sm, b);
1689                 if (r)
1690                         break;
1691         }
1692         up_write(&pmd->root_lock);
1693
1694         return r;
1695 }
1696
1697 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1698 {
1699         int r = 0;
1700
1701         down_write(&pmd->root_lock);
1702         for (; b != e; b++) {
1703                 r = dm_sm_dec_block(pmd->data_sm, b);
1704                 if (r)
1705                         break;
1706         }
1707         up_write(&pmd->root_lock);
1708
1709         return r;
1710 }
1711
1712 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1713 {
1714         int r;
1715
1716         down_read(&td->pmd->root_lock);
1717         r = td->changed;
1718         up_read(&td->pmd->root_lock);
1719
1720         return r;
1721 }
1722
1723 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1724 {
1725         bool r = false;
1726         struct dm_thin_device *td, *tmp;
1727
1728         down_read(&pmd->root_lock);
1729         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1730                 if (td->changed) {
1731                         r = td->changed;
1732                         break;
1733                 }
1734         }
1735         up_read(&pmd->root_lock);
1736
1737         return r;
1738 }
1739
1740 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1741 {
1742         bool r;
1743
1744         down_read(&td->pmd->root_lock);
1745         r = td->aborted_with_changes;
1746         up_read(&td->pmd->root_lock);
1747
1748         return r;
1749 }
1750
1751 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1752 {
1753         int r = -EINVAL;
1754
1755         down_write(&pmd->root_lock);
1756         if (!pmd->fail_io)
1757                 r = dm_sm_new_block(pmd->data_sm, result);
1758         up_write(&pmd->root_lock);
1759
1760         return r;
1761 }
1762
1763 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1764 {
1765         int r = -EINVAL;
1766
1767         down_write(&pmd->root_lock);
1768         if (pmd->fail_io)
1769                 goto out;
1770
1771         r = __commit_transaction(pmd);
1772         if (r <= 0)
1773                 goto out;
1774
1775         /*
1776          * Open the next transaction.
1777          */
1778         r = __begin_transaction(pmd);
1779 out:
1780         up_write(&pmd->root_lock);
1781         return r;
1782 }
1783
1784 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1785 {
1786         struct dm_thin_device *td;
1787
1788         list_for_each_entry(td, &pmd->thin_devices, list)
1789                 td->aborted_with_changes = td->changed;
1790 }
1791
1792 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1793 {
1794         int r = -EINVAL;
1795
1796         down_write(&pmd->root_lock);
1797         if (pmd->fail_io)
1798                 goto out;
1799
1800         __set_abort_with_changes_flags(pmd);
1801         __destroy_persistent_data_objects(pmd);
1802         r = __create_persistent_data_objects(pmd, false);
1803         if (r)
1804                 pmd->fail_io = true;
1805
1806 out:
1807         up_write(&pmd->root_lock);
1808
1809         return r;
1810 }
1811
1812 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1813 {
1814         int r = -EINVAL;
1815
1816         down_read(&pmd->root_lock);
1817         if (!pmd->fail_io)
1818                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1819         up_read(&pmd->root_lock);
1820
1821         return r;
1822 }
1823
1824 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1825                                           dm_block_t *result)
1826 {
1827         int r = -EINVAL;
1828
1829         down_read(&pmd->root_lock);
1830         if (!pmd->fail_io)
1831                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1832         up_read(&pmd->root_lock);
1833
1834         return r;
1835 }
1836
1837 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1838                                   dm_block_t *result)
1839 {
1840         int r = -EINVAL;
1841
1842         down_read(&pmd->root_lock);
1843         if (!pmd->fail_io)
1844                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1845         up_read(&pmd->root_lock);
1846
1847         return r;
1848 }
1849
1850 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1851 {
1852         int r = -EINVAL;
1853
1854         down_read(&pmd->root_lock);
1855         if (!pmd->fail_io)
1856                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1857         up_read(&pmd->root_lock);
1858
1859         return r;
1860 }
1861
1862 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1863 {
1864         int r = -EINVAL;
1865         struct dm_pool_metadata *pmd = td->pmd;
1866
1867         down_read(&pmd->root_lock);
1868         if (!pmd->fail_io) {
1869                 *result = td->mapped_blocks;
1870                 r = 0;
1871         }
1872         up_read(&pmd->root_lock);
1873
1874         return r;
1875 }
1876
1877 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1878 {
1879         int r;
1880         __le64 value_le;
1881         dm_block_t thin_root;
1882         struct dm_pool_metadata *pmd = td->pmd;
1883
1884         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1885         if (r)
1886                 return r;
1887
1888         thin_root = le64_to_cpu(value_le);
1889
1890         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1891 }
1892
1893 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1894                                      dm_block_t *result)
1895 {
1896         int r = -EINVAL;
1897         struct dm_pool_metadata *pmd = td->pmd;
1898
1899         down_read(&pmd->root_lock);
1900         if (!pmd->fail_io)
1901                 r = __highest_block(td, result);
1902         up_read(&pmd->root_lock);
1903
1904         return r;
1905 }
1906
1907 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1908 {
1909         int r;
1910         dm_block_t old_count;
1911
1912         r = dm_sm_get_nr_blocks(sm, &old_count);
1913         if (r)
1914                 return r;
1915
1916         if (new_count == old_count)
1917                 return 0;
1918
1919         if (new_count < old_count) {
1920                 DMERR("cannot reduce size of space map");
1921                 return -EINVAL;
1922         }
1923
1924         return dm_sm_extend(sm, new_count - old_count);
1925 }
1926
1927 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1928 {
1929         int r = -EINVAL;
1930
1931         down_write(&pmd->root_lock);
1932         if (!pmd->fail_io)
1933                 r = __resize_space_map(pmd->data_sm, new_count);
1934         up_write(&pmd->root_lock);
1935
1936         return r;
1937 }
1938
1939 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1940 {
1941         int r = -EINVAL;
1942
1943         down_write(&pmd->root_lock);
1944         if (!pmd->fail_io)
1945                 r = __resize_space_map(pmd->metadata_sm, new_count);
1946         up_write(&pmd->root_lock);
1947
1948         return r;
1949 }
1950
1951 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1952 {
1953         down_write(&pmd->root_lock);
1954         dm_bm_set_read_only(pmd->bm);
1955         up_write(&pmd->root_lock);
1956 }
1957
1958 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
1959 {
1960         down_write(&pmd->root_lock);
1961         dm_bm_set_read_write(pmd->bm);
1962         up_write(&pmd->root_lock);
1963 }
1964
1965 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
1966                                         dm_block_t threshold,
1967                                         dm_sm_threshold_fn fn,
1968                                         void *context)
1969 {
1970         int r;
1971
1972         down_write(&pmd->root_lock);
1973         r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
1974         up_write(&pmd->root_lock);
1975
1976         return r;
1977 }
1978
1979 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
1980 {
1981         int r;
1982         struct dm_block *sblock;
1983         struct thin_disk_superblock *disk_super;
1984
1985         down_write(&pmd->root_lock);
1986         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
1987
1988         r = superblock_lock(pmd, &sblock);
1989         if (r) {
1990                 DMERR("couldn't read superblock");
1991                 goto out;
1992         }
1993
1994         disk_super = dm_block_data(sblock);
1995         disk_super->flags = cpu_to_le32(pmd->flags);
1996
1997         dm_bm_unlock(sblock);
1998 out:
1999         up_write(&pmd->root_lock);
2000         return r;
2001 }
2002
2003 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2004 {
2005         bool needs_check;
2006
2007         down_read(&pmd->root_lock);
2008         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2009         up_read(&pmd->root_lock);
2010
2011         return needs_check;
2012 }
2013
2014 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2015 {
2016         down_read(&pmd->root_lock);
2017         if (!pmd->fail_io)
2018                 dm_tm_issue_prefetches(pmd->tm);
2019         up_read(&pmd->root_lock);
2020 }