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