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         /*
 665          * For pool metadata opening process, root setting is redundant
 666          * because it will be set again in __begin_transaction(). But dm
 667          * pool aborting process really needs to get last transaction's
 668          * root to avoid accessing broken btree.
 669          */
 670         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 671         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 672
 673         __setup_btree_details(pmd);
 674         dm_bm_unlock(sblock);
 675
 676         return 0;
 677
 678 bad_cleanup_data_sm:
 679         dm_sm_destroy(pmd->data_sm);
 680 bad_cleanup_tm:
 681         dm_tm_destroy(pmd->tm);
 682         dm_sm_destroy(pmd->metadata_sm);
 683 bad_unlock_sblock:
 684         dm_bm_unlock(sblock);
 685
 686         return r;
 687 }
 688
 689 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 690 {
 691         int r, unformatted;
 692
 693         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 694         if (r)
 695                 return r;
 696
 697         if (unformatted)
 698                 return format_device ? __format_metadata(pmd) : -EPERM;
 699
 700         return __open_metadata(pmd);
 701 }
 702
 703 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 704 {
 705         int r;
 706
 707         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 708                                           THIN_METADATA_CACHE_SIZE,
 709                                           THIN_MAX_CONCURRENT_LOCKS);
 710         if (IS_ERR(pmd->bm)) {
 711                 DMERR("could not create block manager");
 712                 r = PTR_ERR(pmd->bm);
 713                 pmd->bm = NULL;
 714                 return r;
 715         }
 716
 717         r = __open_or_format_metadata(pmd, format_device);
 718         if (r) {
 719                 dm_block_manager_destroy(pmd->bm);
 720                 pmd->bm = NULL;
 721         }
 722
 723         return r;
 724 }
 725
 726 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 727 {
 728         dm_sm_destroy(pmd->data_sm);
 729         dm_sm_destroy(pmd->metadata_sm);
 730         dm_tm_destroy(pmd->nb_tm);
 731         dm_tm_destroy(pmd->tm);
 732         dm_block_manager_destroy(pmd->bm);
 733 }
 734
 735 static int __begin_transaction(struct dm_pool_metadata *pmd)
 736 {
 737         int r;
 738         struct thin_disk_superblock *disk_super;
 739         struct dm_block *sblock;
 740
 741         /*
 742          * We re-read the superblock every time.  Shouldn't need to do this
 743          * really.
 744          */
 745         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 746                             &sb_validator, &sblock);
 747         if (r)
 748                 return r;
 749
 750         disk_super = dm_block_data(sblock);
 751         pmd->time = le32_to_cpu(disk_super->time);
 752         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 753         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 754         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 755         pmd->flags = le32_to_cpu(disk_super->flags);
 756         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 757
 758         dm_bm_unlock(sblock);
 759         return 0;
 760 }
 761
 762 static int __write_changed_details(struct dm_pool_metadata *pmd)
 763 {
 764         int r;
 765         struct dm_thin_device *td, *tmp;
 766         struct disk_device_details details;
 767         uint64_t key;
 768
 769         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 770                 if (!td->changed)
 771                         continue;
 772
 773                 key = td->id;
 774
 775                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 776                 details.transaction_id = cpu_to_le64(td->transaction_id);
 777                 details.creation_time = cpu_to_le32(td->creation_time);
 778                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 779                 __dm_bless_for_disk(&details);
 780
 781                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 782                                     &key, &details, &pmd->details_root);
 783                 if (r)
 784                         return r;
 785
 786                 if (td->open_count)
 787                         td->changed = 0;
 788                 else {
 789                         list_del(&td->list);
 790                         kfree(td);
 791                 }
 792         }
 793
 794         return 0;
 795 }
 796
 797 static int __commit_transaction(struct dm_pool_metadata *pmd)
 798 {
 799         int r;
 800         size_t metadata_len, data_len;
 801         struct thin_disk_superblock *disk_super;
 802         struct dm_block *sblock;
 803
 804         /*
 805          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 806          */
 807         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 808
 809         r = __write_changed_details(pmd);
 810         if (r < 0)
 811                 return r;
 812
 813         r = dm_sm_commit(pmd->data_sm);
 814         if (r < 0)
 815                 return r;
 816
 817         r = dm_tm_pre_commit(pmd->tm);
 818         if (r < 0)
 819                 return r;
 820
 821         r = dm_sm_root_size(pmd->metadata_sm, &metadata_len);
 822         if (r < 0)
 823                 return r;
 824
 825         r = dm_sm_root_size(pmd->data_sm, &data_len);
 826         if (r < 0)
 827                 return r;
 828
 829         r = save_sm_roots(pmd);
 830         if (r < 0)
 831                 return r;
 832
 833         r = superblock_lock(pmd, &sblock);
 834         if (r)
 835                 return r;
 836
 837         disk_super = dm_block_data(sblock);
 838         disk_super->time = cpu_to_le32(pmd->time);
 839         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 840         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 841         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 842         disk_super->flags = cpu_to_le32(pmd->flags);
 843
 844         copy_sm_roots(pmd, disk_super);
 845
 846         return dm_tm_commit(pmd->tm, sblock);
 847 }
 848
 849 static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 850 {
 851         int r;
 852         dm_block_t total;
 853         dm_block_t max_blocks = 4096; /* 16M */
 854
 855         r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 856         if (r) {
 857                 DMERR("could not get size of metadata device");
 858                 pmd->metadata_reserve = max_blocks;
 859         } else
 860                 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 861 }
 862
 863 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 864                                                sector_t data_block_size,
 865                                                bool format_device)
 866 {
 867         int r;
 868         struct dm_pool_metadata *pmd;
 869
 870         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 871         if (!pmd) {
 872                 DMERR("could not allocate metadata struct");
 873                 return ERR_PTR(-ENOMEM);
 874         }
 875
 876         init_rwsem(&pmd->root_lock);
 877         pmd->time = 0;
 878         INIT_LIST_HEAD(&pmd->thin_devices);
 879         pmd->fail_io = false;
 880         pmd->bdev = bdev;
 881         pmd->data_block_size = data_block_size;
 882
 883         r = __create_persistent_data_objects(pmd, format_device);
 884         if (r) {
 885                 kfree(pmd);
 886                 return ERR_PTR(r);
 887         }
 888
 889         r = __begin_transaction(pmd);
 890         if (r < 0) {
 891                 if (dm_pool_metadata_close(pmd) < 0)
 892                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 893                 return ERR_PTR(r);
 894         }
 895
 896         __set_metadata_reserve(pmd);
 897
 898         return pmd;
 899 }
 900
 901 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 902 {
 903         int r;
 904         unsigned open_devices = 0;
 905         struct dm_thin_device *td, *tmp;
 906
 907         down_read(&pmd->root_lock);
 908         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 909                 if (td->open_count)
 910                         open_devices++;
 911                 else {
 912                         list_del(&td->list);
 913                         kfree(td);
 914                 }
 915         }
 916         up_read(&pmd->root_lock);
 917
 918         if (open_devices) {
 919                 DMERR("attempt to close pmd when %u device(s) are still open",
 920                        open_devices);
 921                 return -EBUSY;
 922         }
 923
 924         if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
 925                 r = __commit_transaction(pmd);
 926                 if (r < 0)
 927                         DMWARN("%s: __commit_transaction() failed, error = %d",
 928                                __func__, r);
 929         }
 930
 931         if (!pmd->fail_io)
 932                 __destroy_persistent_data_objects(pmd);
 933
 934         kfree(pmd);
 935         return 0;
 936 }
 937
 938 /*
 939  * __open_device: Returns @td corresponding to device with id @dev,
 940  * creating it if @create is set and incrementing @td->open_count.
 941  * On failure, @td is undefined.
 942  */
 943 static int __open_device(struct dm_pool_metadata *pmd,
 944                          dm_thin_id dev, int create,
 945                          struct dm_thin_device **td)
 946 {
 947         int r, changed = 0;
 948         struct dm_thin_device *td2;
 949         uint64_t key = dev;
 950         struct disk_device_details details_le;
 951
 952         /*
 953          * If the device is already open, return it.
 954          */
 955         list_for_each_entry(td2, &pmd->thin_devices, list)
 956                 if (td2->id == dev) {
 957                         /*
 958                          * May not create an already-open device.
 959                          */
 960                         if (create)
 961                                 return -EEXIST;
 962
 963                         td2->open_count++;
 964                         *td = td2;
 965                         return 0;
 966                 }
 967
 968         /*
 969          * Check the device exists.
 970          */
 971         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 972                             &key, &details_le);
 973         if (r) {
 974                 if (r != -ENODATA || !create)
 975                         return r;
 976
 977                 /*
 978                  * Create new device.
 979                  */
 980                 changed = 1;
 981                 details_le.mapped_blocks = 0;
 982                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
 983                 details_le.creation_time = cpu_to_le32(pmd->time);
 984                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
 985         }
 986
 987         *td = kmalloc(sizeof(**td), GFP_NOIO);
 988         if (!*td)
 989                 return -ENOMEM;
 990
 991         (*td)->pmd = pmd;
 992         (*td)->id = dev;
 993         (*td)->open_count = 1;
 994         (*td)->changed = changed;
 995         (*td)->aborted_with_changes = false;
 996         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
 997         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
 998         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
 999         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1000
1001         list_add(&(*td)->list, &pmd->thin_devices);
1002
1003         return 0;
1004 }
1005
1006 static void __close_device(struct dm_thin_device *td)
1007 {
1008         --td->open_count;
1009 }
1010
1011 static int __create_thin(struct dm_pool_metadata *pmd,
1012                          dm_thin_id dev)
1013 {
1014         int r;
1015         dm_block_t dev_root;
1016         uint64_t key = dev;
1017         struct disk_device_details details_le;
1018         struct dm_thin_device *td;
1019         __le64 value;
1020
1021         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1022                             &key, &details_le);
1023         if (!r)
1024                 return -EEXIST;
1025
1026         /*
1027          * Create an empty btree for the mappings.
1028          */
1029         r = dm_btree_empty(&pmd->bl_info, &dev_root);
1030         if (r)
1031                 return r;
1032
1033         /*
1034          * Insert it into the main mapping tree.
1035          */
1036         value = cpu_to_le64(dev_root);
1037         __dm_bless_for_disk(&value);
1038         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1039         if (r) {
1040                 dm_btree_del(&pmd->bl_info, dev_root);
1041                 return r;
1042         }
1043
1044         r = __open_device(pmd, dev, 1, &td);
1045         if (r) {
1046                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1047                 dm_btree_del(&pmd->bl_info, dev_root);
1048                 return r;
1049         }
1050         __close_device(td);
1051
1052         return r;
1053 }
1054
1055 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1056 {
1057         int r = -EINVAL;
1058
1059         down_write(&pmd->root_lock);
1060         if (!pmd->fail_io)
1061                 r = __create_thin(pmd, dev);
1062         up_write(&pmd->root_lock);
1063
1064         return r;
1065 }
1066
1067 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1068                                   struct dm_thin_device *snap,
1069                                   dm_thin_id origin, uint32_t time)
1070 {
1071         int r;
1072         struct dm_thin_device *td;
1073
1074         r = __open_device(pmd, origin, 0, &td);
1075         if (r)
1076                 return r;
1077
1078         td->changed = 1;
1079         td->snapshotted_time = time;
1080
1081         snap->mapped_blocks = td->mapped_blocks;
1082         snap->snapshotted_time = time;
1083         __close_device(td);
1084
1085         return 0;
1086 }
1087
1088 static int __create_snap(struct dm_pool_metadata *pmd,
1089                          dm_thin_id dev, dm_thin_id origin)
1090 {
1091         int r;
1092         dm_block_t origin_root;
1093         uint64_t key = origin, dev_key = dev;
1094         struct dm_thin_device *td;
1095         struct disk_device_details details_le;
1096         __le64 value;
1097
1098         /* check this device is unused */
1099         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1100                             &dev_key, &details_le);
1101         if (!r)
1102                 return -EEXIST;
1103
1104         /* find the mapping tree for the origin */
1105         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1106         if (r)
1107                 return r;
1108         origin_root = le64_to_cpu(value);
1109
1110         /* clone the origin, an inc will do */
1111         dm_tm_inc(pmd->tm, origin_root);
1112
1113         /* insert into the main mapping tree */
1114         value = cpu_to_le64(origin_root);
1115         __dm_bless_for_disk(&value);
1116         key = dev;
1117         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1118         if (r) {
1119                 dm_tm_dec(pmd->tm, origin_root);
1120                 return r;
1121         }
1122
1123         pmd->time++;
1124
1125         r = __open_device(pmd, dev, 1, &td);
1126         if (r)
1127                 goto bad;
1128
1129         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1130         __close_device(td);
1131
1132         if (r)
1133                 goto bad;
1134
1135         return 0;
1136
1137 bad:
1138         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1139         dm_btree_remove(&pmd->details_info, pmd->details_root,
1140                         &key, &pmd->details_root);
1141         return r;
1142 }
1143
1144 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1145                                  dm_thin_id dev,
1146                                  dm_thin_id origin)
1147 {
1148         int r = -EINVAL;
1149
1150         down_write(&pmd->root_lock);
1151         if (!pmd->fail_io)
1152                 r = __create_snap(pmd, dev, origin);
1153         up_write(&pmd->root_lock);
1154
1155         return r;
1156 }
1157
1158 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1159 {
1160         int r;
1161         uint64_t key = dev;
1162         struct dm_thin_device *td;
1163
1164         /* TODO: failure should mark the transaction invalid */
1165         r = __open_device(pmd, dev, 0, &td);
1166         if (r)
1167                 return r;
1168
1169         if (td->open_count > 1) {
1170                 __close_device(td);
1171                 return -EBUSY;
1172         }
1173
1174         list_del(&td->list);
1175         kfree(td);
1176         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1177                             &key, &pmd->details_root);
1178         if (r)
1179                 return r;
1180
1181         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1182         if (r)
1183                 return r;
1184
1185         return 0;
1186 }
1187
1188 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1189                                dm_thin_id dev)
1190 {
1191         int r = -EINVAL;
1192
1193         down_write(&pmd->root_lock);
1194         if (!pmd->fail_io)
1195                 r = __delete_device(pmd, dev);
1196         up_write(&pmd->root_lock);
1197
1198         return r;
1199 }
1200
1201 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1202                                         uint64_t current_id,
1203                                         uint64_t new_id)
1204 {
1205         int r = -EINVAL;
1206
1207         down_write(&pmd->root_lock);
1208
1209         if (pmd->fail_io)
1210                 goto out;
1211
1212         if (pmd->trans_id != current_id) {
1213                 DMERR("mismatched transaction id");
1214                 goto out;
1215         }
1216
1217         pmd->trans_id = new_id;
1218         r = 0;
1219
1220 out:
1221         up_write(&pmd->root_lock);
1222
1223         return r;
1224 }
1225
1226 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1227                                         uint64_t *result)
1228 {
1229         int r = -EINVAL;
1230
1231         down_read(&pmd->root_lock);
1232         if (!pmd->fail_io) {
1233                 *result = pmd->trans_id;
1234                 r = 0;
1235         }
1236         up_read(&pmd->root_lock);
1237
1238         return r;
1239 }
1240
1241 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1242 {
1243         int r, inc;
1244         struct thin_disk_superblock *disk_super;
1245         struct dm_block *copy, *sblock;
1246         dm_block_t held_root;
1247
1248         /*
1249          * We commit to ensure the btree roots which we increment in a
1250          * moment are up to date.
1251          */
1252         __commit_transaction(pmd);
1253
1254         /*
1255          * Copy the superblock.
1256          */
1257         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1258         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1259                                &sb_validator, &copy, &inc);
1260         if (r)
1261                 return r;
1262
1263         BUG_ON(!inc);
1264
1265         held_root = dm_block_location(copy);
1266         disk_super = dm_block_data(copy);
1267
1268         if (le64_to_cpu(disk_super->held_root)) {
1269                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1270
1271                 dm_tm_dec(pmd->tm, held_root);
1272                 dm_tm_unlock(pmd->tm, copy);
1273                 return -EBUSY;
1274         }
1275
1276         /*
1277          * Wipe the spacemap since we're not publishing this.
1278          */
1279         memset(&disk_super->data_space_map_root, 0,
1280                sizeof(disk_super->data_space_map_root));
1281         memset(&disk_super->metadata_space_map_root, 0,
1282                sizeof(disk_super->metadata_space_map_root));
1283
1284         /*
1285          * Increment the data structures that need to be preserved.
1286          */
1287         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1288         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1289         dm_tm_unlock(pmd->tm, copy);
1290
1291         /*
1292          * Write the held root into the superblock.
1293          */
1294         r = superblock_lock(pmd, &sblock);
1295         if (r) {
1296                 dm_tm_dec(pmd->tm, held_root);
1297                 return r;
1298         }
1299
1300         disk_super = dm_block_data(sblock);
1301         disk_super->held_root = cpu_to_le64(held_root);
1302         dm_bm_unlock(sblock);
1303         return 0;
1304 }
1305
1306 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1307 {
1308         int r = -EINVAL;
1309
1310         down_write(&pmd->root_lock);
1311         if (!pmd->fail_io)
1312                 r = __reserve_metadata_snap(pmd);
1313         up_write(&pmd->root_lock);
1314
1315         return r;
1316 }
1317
1318 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1319 {
1320         int r;
1321         struct thin_disk_superblock *disk_super;
1322         struct dm_block *sblock, *copy;
1323         dm_block_t held_root;
1324
1325         r = superblock_lock(pmd, &sblock);
1326         if (r)
1327                 return r;
1328
1329         disk_super = dm_block_data(sblock);
1330         held_root = le64_to_cpu(disk_super->held_root);
1331         disk_super->held_root = cpu_to_le64(0);
1332
1333         dm_bm_unlock(sblock);
1334
1335         if (!held_root) {
1336                 DMWARN("No pool metadata snapshot found: nothing to release.");
1337                 return -EINVAL;
1338         }
1339
1340         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1341         if (r)
1342                 return r;
1343
1344         disk_super = dm_block_data(copy);
1345         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1346         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1347         dm_sm_dec_block(pmd->metadata_sm, held_root);
1348
1349         dm_tm_unlock(pmd->tm, copy);
1350
1351         return 0;
1352 }
1353
1354 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1355 {
1356         int r = -EINVAL;
1357
1358         down_write(&pmd->root_lock);
1359         if (!pmd->fail_io)
1360                 r = __release_metadata_snap(pmd);
1361         up_write(&pmd->root_lock);
1362
1363         return r;
1364 }
1365
1366 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1367                                dm_block_t *result)
1368 {
1369         int r;
1370         struct thin_disk_superblock *disk_super;
1371         struct dm_block *sblock;
1372
1373         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1374                             &sb_validator, &sblock);
1375         if (r)
1376                 return r;
1377
1378         disk_super = dm_block_data(sblock);
1379         *result = le64_to_cpu(disk_super->held_root);
1380
1381         dm_bm_unlock(sblock);
1382
1383         return 0;
1384 }
1385
1386 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1387                               dm_block_t *result)
1388 {
1389         int r = -EINVAL;
1390
1391         down_read(&pmd->root_lock);
1392         if (!pmd->fail_io)
1393                 r = __get_metadata_snap(pmd, result);
1394         up_read(&pmd->root_lock);
1395
1396         return r;
1397 }
1398
1399 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1400                              struct dm_thin_device **td)
1401 {
1402         int r = -EINVAL;
1403
1404         down_write(&pmd->root_lock);
1405         if (!pmd->fail_io)
1406                 r = __open_device(pmd, dev, 0, td);
1407         up_write(&pmd->root_lock);
1408
1409         return r;
1410 }
1411
1412 int dm_pool_close_thin_device(struct dm_thin_device *td)
1413 {
1414         down_write(&td->pmd->root_lock);
1415         __close_device(td);
1416         up_write(&td->pmd->root_lock);
1417
1418         return 0;
1419 }
1420
1421 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1422 {
1423         return td->id;
1424 }
1425
1426 /*
1427  * Check whether @time (of block creation) is older than @td's last snapshot.
1428  * If so then the associated block is shared with the last snapshot device.
1429  * Any block on a device created *after* the device last got snapshotted is
1430  * necessarily not shared.
1431  */
1432 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1433 {
1434         return td->snapshotted_time > time;
1435 }
1436
1437 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1438                                  struct dm_thin_lookup_result *result)
1439 {
1440         uint64_t block_time = 0;
1441         dm_block_t exception_block;
1442         uint32_t exception_time;
1443
1444         block_time = le64_to_cpu(value);
1445         unpack_block_time(block_time, &exception_block, &exception_time);
1446         result->block = exception_block;
1447         result->shared = __snapshotted_since(td, exception_time);
1448 }
1449
1450 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1451                         int can_issue_io, struct dm_thin_lookup_result *result)
1452 {
1453         int r;
1454         __le64 value;
1455         struct dm_pool_metadata *pmd = td->pmd;
1456         dm_block_t keys[2] = { td->id, block };
1457         struct dm_btree_info *info;
1458
1459         if (can_issue_io) {
1460                 info = &pmd->info;
1461         } else
1462                 info = &pmd->nb_info;
1463
1464         r = dm_btree_lookup(info, pmd->root, keys, &value);
1465         if (!r)
1466                 unpack_lookup_result(td, value, result);
1467
1468         return r;
1469 }
1470
1471 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1472                        int can_issue_io, struct dm_thin_lookup_result *result)
1473 {
1474         int r;
1475         struct dm_pool_metadata *pmd = td->pmd;
1476
1477         down_read(&pmd->root_lock);
1478         if (pmd->fail_io) {
1479                 up_read(&pmd->root_lock);
1480                 return -EINVAL;
1481         }
1482
1483         r = __find_block(td, block, can_issue_io, result);
1484
1485         up_read(&pmd->root_lock);
1486         return r;
1487 }
1488
1489 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1490                                           dm_block_t *vblock,
1491                                           struct dm_thin_lookup_result *result)
1492 {
1493         int r;
1494         __le64 value;
1495         struct dm_pool_metadata *pmd = td->pmd;
1496         dm_block_t keys[2] = { td->id, block };
1497
1498         r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1499         if (!r)
1500                 unpack_lookup_result(td, value, result);
1501
1502         return r;
1503 }
1504
1505 static int __find_mapped_range(struct dm_thin_device *td,
1506                                dm_block_t begin, dm_block_t end,
1507                                dm_block_t *thin_begin, dm_block_t *thin_end,
1508                                dm_block_t *pool_begin, bool *maybe_shared)
1509 {
1510         int r;
1511         dm_block_t pool_end;
1512         struct dm_thin_lookup_result lookup;
1513
1514         if (end < begin)
1515                 return -ENODATA;
1516
1517         r = __find_next_mapped_block(td, begin, &begin, &lookup);
1518         if (r)
1519                 return r;
1520
1521         if (begin >= end)
1522                 return -ENODATA;
1523
1524         *thin_begin = begin;
1525         *pool_begin = lookup.block;
1526         *maybe_shared = lookup.shared;
1527
1528         begin++;
1529         pool_end = *pool_begin + 1;
1530         while (begin != end) {
1531                 r = __find_block(td, begin, true, &lookup);
1532                 if (r) {
1533                         if (r == -ENODATA)
1534                                 break;
1535                         else
1536                                 return r;
1537                 }
1538
1539                 if ((lookup.block != pool_end) ||
1540                     (lookup.shared != *maybe_shared))
1541                         break;
1542
1543                 pool_end++;
1544                 begin++;
1545         }
1546
1547         *thin_end = begin;
1548         return 0;
1549 }
1550
1551 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1552                               dm_block_t begin, dm_block_t end,
1553                               dm_block_t *thin_begin, dm_block_t *thin_end,
1554                               dm_block_t *pool_begin, bool *maybe_shared)
1555 {
1556         int r = -EINVAL;
1557         struct dm_pool_metadata *pmd = td->pmd;
1558
1559         down_read(&pmd->root_lock);
1560         if (!pmd->fail_io) {
1561                 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1562                                         pool_begin, maybe_shared);
1563         }
1564         up_read(&pmd->root_lock);
1565
1566         return r;
1567 }
1568
1569 static int __insert(struct dm_thin_device *td, dm_block_t block,
1570                     dm_block_t data_block)
1571 {
1572         int r, inserted;
1573         __le64 value;
1574         struct dm_pool_metadata *pmd = td->pmd;
1575         dm_block_t keys[2] = { td->id, block };
1576
1577         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1578         __dm_bless_for_disk(&value);
1579
1580         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1581                                    &pmd->root, &inserted);
1582         if (r)
1583                 return r;
1584
1585         td->changed = 1;
1586         if (inserted)
1587                 td->mapped_blocks++;
1588
1589         return 0;
1590 }
1591
1592 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1593                          dm_block_t data_block)
1594 {
1595         int r = -EINVAL;
1596
1597         down_write(&td->pmd->root_lock);
1598         if (!td->pmd->fail_io)
1599                 r = __insert(td, block, data_block);
1600         up_write(&td->pmd->root_lock);
1601
1602         return r;
1603 }
1604
1605 static int __remove(struct dm_thin_device *td, dm_block_t block)
1606 {
1607         int r;
1608         struct dm_pool_metadata *pmd = td->pmd;
1609         dm_block_t keys[2] = { td->id, block };
1610
1611         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1612         if (r)
1613                 return r;
1614
1615         td->mapped_blocks--;
1616         td->changed = 1;
1617
1618         return 0;
1619 }
1620
1621 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1622 {
1623         int r;
1624         unsigned count, total_count = 0;
1625         struct dm_pool_metadata *pmd = td->pmd;
1626         dm_block_t keys[1] = { td->id };
1627         __le64 value;
1628         dm_block_t mapping_root;
1629
1630         /*
1631          * Find the mapping tree
1632          */
1633         r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1634         if (r)
1635                 return r;
1636
1637         /*
1638          * Remove from the mapping tree, taking care to inc the
1639          * ref count so it doesn't get deleted.
1640          */
1641         mapping_root = le64_to_cpu(value);
1642         dm_tm_inc(pmd->tm, mapping_root);
1643         r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1644         if (r)
1645                 return r;
1646
1647         /*
1648          * Remove leaves stops at the first unmapped entry, so we have to
1649          * loop round finding mapped ranges.
1650          */
1651         while (begin < end) {
1652                 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1653                 if (r == -ENODATA)
1654                         break;
1655
1656                 if (r)
1657                         return r;
1658
1659                 if (begin >= end)
1660                         break;
1661
1662                 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1663                 if (r)
1664                         return r;
1665
1666                 total_count += count;
1667         }
1668
1669         td->mapped_blocks -= total_count;
1670         td->changed = 1;
1671
1672         /*
1673          * Reinsert the mapping tree.
1674          */
1675         value = cpu_to_le64(mapping_root);
1676         __dm_bless_for_disk(&value);
1677         return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1678 }
1679
1680 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1681 {
1682         int r = -EINVAL;
1683
1684         down_write(&td->pmd->root_lock);
1685         if (!td->pmd->fail_io)
1686                 r = __remove(td, block);
1687         up_write(&td->pmd->root_lock);
1688
1689         return r;
1690 }
1691
1692 int dm_thin_remove_range(struct dm_thin_device *td,
1693                          dm_block_t begin, dm_block_t end)
1694 {
1695         int r = -EINVAL;
1696
1697         down_write(&td->pmd->root_lock);
1698         if (!td->pmd->fail_io)
1699                 r = __remove_range(td, begin, end);
1700         up_write(&td->pmd->root_lock);
1701
1702         return r;
1703 }
1704
1705 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1706 {
1707         int r;
1708         uint32_t ref_count;
1709
1710         down_read(&pmd->root_lock);
1711         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1712         if (!r)
1713                 *result = (ref_count > 1);
1714         up_read(&pmd->root_lock);
1715
1716         return r;
1717 }
1718
1719 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1720 {
1721         int r = 0;
1722
1723         down_write(&pmd->root_lock);
1724         for (; b != e; b++) {
1725                 r = dm_sm_inc_block(pmd->data_sm, b);
1726                 if (r)
1727                         break;
1728         }
1729         up_write(&pmd->root_lock);
1730
1731         return r;
1732 }
1733
1734 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1735 {
1736         int r = 0;
1737
1738         down_write(&pmd->root_lock);
1739         for (; b != e; b++) {
1740                 r = dm_sm_dec_block(pmd->data_sm, b);
1741                 if (r)
1742                         break;
1743         }
1744         up_write(&pmd->root_lock);
1745
1746         return r;
1747 }
1748
1749 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1750 {
1751         int r;
1752
1753         down_read(&td->pmd->root_lock);
1754         r = td->changed;
1755         up_read(&td->pmd->root_lock);
1756
1757         return r;
1758 }
1759
1760 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1761 {
1762         bool r = false;
1763         struct dm_thin_device *td, *tmp;
1764
1765         down_read(&pmd->root_lock);
1766         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1767                 if (td->changed) {
1768                         r = td->changed;
1769                         break;
1770                 }
1771         }
1772         up_read(&pmd->root_lock);
1773
1774         return r;
1775 }
1776
1777 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1778 {
1779         bool r;
1780
1781         down_read(&td->pmd->root_lock);
1782         r = td->aborted_with_changes;
1783         up_read(&td->pmd->root_lock);
1784
1785         return r;
1786 }
1787
1788 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1789 {
1790         int r = -EINVAL;
1791
1792         down_write(&pmd->root_lock);
1793         if (!pmd->fail_io)
1794                 r = dm_sm_new_block(pmd->data_sm, result);
1795         up_write(&pmd->root_lock);
1796
1797         return r;
1798 }
1799
1800 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1801 {
1802         int r = -EINVAL;
1803
1804         down_write(&pmd->root_lock);
1805         if (pmd->fail_io)
1806                 goto out;
1807
1808         r = __commit_transaction(pmd);
1809         if (r <= 0)
1810                 goto out;
1811
1812         /*
1813          * Open the next transaction.
1814          */
1815         r = __begin_transaction(pmd);
1816 out:
1817         up_write(&pmd->root_lock);
1818         return r;
1819 }
1820
1821 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1822 {
1823         struct dm_thin_device *td;
1824
1825         list_for_each_entry(td, &pmd->thin_devices, list)
1826                 td->aborted_with_changes = td->changed;
1827 }
1828
1829 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1830 {
1831         int r = -EINVAL;
1832
1833         down_write(&pmd->root_lock);
1834         if (pmd->fail_io)
1835                 goto out;
1836
1837         __set_abort_with_changes_flags(pmd);
1838         __destroy_persistent_data_objects(pmd);
1839         r = __create_persistent_data_objects(pmd, false);
1840         if (r)
1841                 pmd->fail_io = true;
1842
1843 out:
1844         up_write(&pmd->root_lock);
1845
1846         return r;
1847 }
1848
1849 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1850 {
1851         int r = -EINVAL;
1852
1853         down_read(&pmd->root_lock);
1854         if (!pmd->fail_io)
1855                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1856         up_read(&pmd->root_lock);
1857
1858         return r;
1859 }
1860
1861 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1862                                           dm_block_t *result)
1863 {
1864         int r = -EINVAL;
1865
1866         down_read(&pmd->root_lock);
1867         if (!pmd->fail_io)
1868                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1869
1870         if (!r) {
1871                 if (*result < pmd->metadata_reserve)
1872                         *result = 0;
1873                 else
1874                         *result -= pmd->metadata_reserve;
1875         }
1876         up_read(&pmd->root_lock);
1877
1878         return r;
1879 }
1880
1881 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1882                                   dm_block_t *result)
1883 {
1884         int r = -EINVAL;
1885
1886         down_read(&pmd->root_lock);
1887         if (!pmd->fail_io)
1888                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1889         up_read(&pmd->root_lock);
1890
1891         return r;
1892 }
1893
1894 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1895 {
1896         int r = -EINVAL;
1897
1898         down_read(&pmd->root_lock);
1899         if (!pmd->fail_io)
1900                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1901         up_read(&pmd->root_lock);
1902
1903         return r;
1904 }
1905
1906 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1907 {
1908         int r = -EINVAL;
1909         struct dm_pool_metadata *pmd = td->pmd;
1910
1911         down_read(&pmd->root_lock);
1912         if (!pmd->fail_io) {
1913                 *result = td->mapped_blocks;
1914                 r = 0;
1915         }
1916         up_read(&pmd->root_lock);
1917
1918         return r;
1919 }
1920
1921 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1922 {
1923         int r;
1924         __le64 value_le;
1925         dm_block_t thin_root;
1926         struct dm_pool_metadata *pmd = td->pmd;
1927
1928         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1929         if (r)
1930                 return r;
1931
1932         thin_root = le64_to_cpu(value_le);
1933
1934         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1935 }
1936
1937 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1938                                      dm_block_t *result)
1939 {
1940         int r = -EINVAL;
1941         struct dm_pool_metadata *pmd = td->pmd;
1942
1943         down_read(&pmd->root_lock);
1944         if (!pmd->fail_io)
1945                 r = __highest_block(td, result);
1946         up_read(&pmd->root_lock);
1947
1948         return r;
1949 }
1950
1951 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1952 {
1953         int r;
1954         dm_block_t old_count;
1955
1956         r = dm_sm_get_nr_blocks(sm, &old_count);
1957         if (r)
1958                 return r;
1959
1960         if (new_count == old_count)
1961                 return 0;
1962
1963         if (new_count < old_count) {
1964                 DMERR("cannot reduce size of space map");
1965                 return -EINVAL;
1966         }
1967
1968         return dm_sm_extend(sm, new_count - old_count);
1969 }
1970
1971 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1972 {
1973         int r = -EINVAL;
1974
1975         down_write(&pmd->root_lock);
1976         if (!pmd->fail_io)
1977                 r = __resize_space_map(pmd->data_sm, new_count);
1978         up_write(&pmd->root_lock);
1979
1980         return r;
1981 }
1982
1983 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1984 {
1985         int r = -EINVAL;
1986
1987         down_write(&pmd->root_lock);
1988         if (!pmd->fail_io) {
1989                 r = __resize_space_map(pmd->metadata_sm, new_count);
1990                 if (!r)
1991                         __set_metadata_reserve(pmd);
1992         }
1993         up_write(&pmd->root_lock);
1994
1995         return r;
1996 }
1997
1998 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1999 {
2000         down_write(&pmd->root_lock);
2001         dm_bm_set_read_only(pmd->bm);
2002         up_write(&pmd->root_lock);
2003 }
2004
2005 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2006 {
2007         down_write(&pmd->root_lock);
2008         dm_bm_set_read_write(pmd->bm);
2009         up_write(&pmd->root_lock);
2010 }
2011
2012 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2013                                         dm_block_t threshold,
2014                                         dm_sm_threshold_fn fn,
2015                                         void *context)
2016 {
2017         int r;
2018
2019         down_write(&pmd->root_lock);
2020         r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
2021         up_write(&pmd->root_lock);
2022
2023         return r;
2024 }
2025
2026 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2027 {
2028         int r;
2029         struct dm_block *sblock;
2030         struct thin_disk_superblock *disk_super;
2031
2032         down_write(&pmd->root_lock);
2033         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2034
2035         r = superblock_lock(pmd, &sblock);
2036         if (r) {
2037                 DMERR("couldn't read superblock");
2038                 goto out;
2039         }
2040
2041         disk_super = dm_block_data(sblock);
2042         disk_super->flags = cpu_to_le32(pmd->flags);
2043
2044         dm_bm_unlock(sblock);
2045 out:
2046         up_write(&pmd->root_lock);
2047         return r;
2048 }
2049
2050 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2051 {
2052         bool needs_check;
2053
2054         down_read(&pmd->root_lock);
2055         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2056         up_read(&pmd->root_lock);
2057
2058         return needs_check;
2059 }
2060
2061 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2062 {
2063         down_read(&pmd->root_lock);
2064         if (!pmd->fail_io)
2065                 dm_tm_issue_prefetches(pmd->tm);
2066         up_read(&pmd->root_lock);
2067 }