drivers/md/dm-thin-metadata.c

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