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 40
  32  *   bits.
  33  *
  34  * BTrees consist solely of btree_nodes, that fill a block.  Some are
  35  * internal nodes, as such their values are a __le64 pointing to other
  36  * nodes.  Leaf nodes can store data of any reasonable size (ie. much
  37  * smaller than the block size).  The nodes consist of the header,
  38  * followed by an array of keys, followed by an array of values.  We have
  39  * to binary search on the keys so they're all held together to help the
  40  * cpu cache.
  41  *
  42  * Space maps have 2 btrees:
  43  *
  44  * - One maps a uint64_t onto a struct index_entry.  Which points to a
  45  *   bitmap block, and has some details about how many free entries there
  46  *   are etc.
  47  *
  48  * - The bitmap blocks have a header (for the checksum).  Then the rest
  49  *   of the block is pairs of bits.  With the meaning being:
  50  *
  51  *   0 - ref count is 0
  52  *   1 - ref count is 1
  53  *   2 - ref count is 2
  54  *   3 - ref count is higher than 2
  55  *
  56  * - If the count is higher than 2 then the ref count is entered in a
  57  *   second btree that directly maps the block_address to a uint32_t ref
  58  *   count.
  59  *
  60  * The space map metadata variant doesn't have a bitmaps btree.  Instead
  61  * it has one single blocks worth of index_entries.  This avoids
  62  * recursive issues with the bitmap btree needing to allocate space in
  63  * order to insert.  With a small data block size such as 64k the
  64  * metadata support data devices that are hundreds of terrabytes.
  65  *
  66  * The space maps allocate space linearly from front to back.  Space that
  67  * is freed in a transaction is never recycled within that transaction.
  68  * To try and avoid fragmenting _free_ space the allocator always goes
  69  * back and fills in gaps.
  70  *
  71  * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
  72  * from the block manager.
  73  *--------------------------------------------------------------------------*/
  74
  75 #define DM_MSG_PREFIX   "thin metadata"
  76
  77 #define THIN_SUPERBLOCK_MAGIC 27022010
  78 #define THIN_SUPERBLOCK_LOCATION 0
  79 #define THIN_VERSION 2
  80 #define SECTOR_TO_BLOCK_SHIFT 3
  81
  82 /*
  83  * For btree insert:
  84  *  3 for btree insert +
  85  *  2 for btree lookup used within space map
  86  * For btree remove:
  87  *  2 for shadow spine +
  88  *  4 for rebalance 3 child node
  89  */
  90 #define THIN_MAX_CONCURRENT_LOCKS 6
  91
  92 /* This should be plenty */
  93 #define SPACE_MAP_ROOT_SIZE 128
  94
  95 /*
  96  * Little endian on-disk superblock and device details.
  97  */
  98 struct thin_disk_superblock {
  99         __le32 csum;    /* Checksum of superblock except for this field. */
 100         __le32 flags;
 101         __le64 blocknr; /* This block number, dm_block_t. */
 102
 103         __u8 uuid[16];
 104         __le64 magic;
 105         __le32 version;
 106         __le32 time;
 107
 108         __le64 trans_id;
 109
 110         /*
 111          * Root held by userspace transactions.
 112          */
 113         __le64 held_root;
 114
 115         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 116         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 117
 118         /*
 119          * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 120          */
 121         __le64 data_mapping_root;
 122
 123         /*
 124          * Device detail root mapping dev_id -> device_details
 125          */
 126         __le64 device_details_root;
 127
 128         __le32 data_block_size;         /* In 512-byte sectors. */
 129
 130         __le32 metadata_block_size;     /* In 512-byte sectors. */
 131         __le64 metadata_nr_blocks;
 132
 133         __le32 compat_flags;
 134         __le32 compat_ro_flags;
 135         __le32 incompat_flags;
 136 } __packed;
 137
 138 struct disk_device_details {
 139         __le64 mapped_blocks;
 140         __le64 transaction_id;          /* When created. */
 141         __le32 creation_time;
 142         __le32 snapshotted_time;
 143 } __packed;
 144
 145 struct dm_pool_metadata {
 146         struct hlist_node hash;
 147
 148         struct block_device *bdev;
 149         struct dm_block_manager *bm;
 150         struct dm_space_map *metadata_sm;
 151         struct dm_space_map *data_sm;
 152         struct dm_transaction_manager *tm;
 153         struct dm_transaction_manager *nb_tm;
 154
 155         /*
 156          * Two-level btree.
 157          * First level holds thin_dev_t.
 158          * Second level holds mappings.
 159          */
 160         struct dm_btree_info info;
 161
 162         /*
 163          * Non-blocking version of the above.
 164          */
 165         struct dm_btree_info nb_info;
 166
 167         /*
 168          * Just the top level for deleting whole devices.
 169          */
 170         struct dm_btree_info tl_info;
 171
 172         /*
 173          * Just the bottom level for creating new devices.
 174          */
 175         struct dm_btree_info bl_info;
 176
 177         /*
 178          * Describes the device details btree.
 179          */
 180         struct dm_btree_info details_info;
 181
 182         struct rw_semaphore root_lock;
 183         uint32_t time;
 184         dm_block_t root;
 185         dm_block_t details_root;
 186         struct list_head thin_devices;
 187         uint64_t trans_id;
 188         unsigned long flags;
 189         sector_t data_block_size;
 190
 191         /*
 192          * Pre-commit callback.
 193          *
 194          * This allows the thin provisioning target to run a callback before
 195          * the metadata are committed.
 196          */
 197         dm_pool_pre_commit_fn pre_commit_fn;
 198         void *pre_commit_context;
 199
 200         /*
 201          * We reserve a section of the metadata for commit overhead.
 202          * All reported space does *not* include this.
 203          */
 204         dm_block_t metadata_reserve;
 205
 206         /*
 207          * Set if a transaction has to be aborted but the attempt to roll back
 208          * to the previous (good) transaction failed.  The only pool metadata
 209          * operation possible in this state is the closing of the device.
 210          */
 211         bool fail_io:1;
 212
 213         /*
 214          * Set once a thin-pool has been accessed through one of the interfaces
 215          * that imply the pool is in-service (e.g. thin devices created/deleted,
 216          * thin-pool message, metadata snapshots, etc).
 217          */
 218         bool in_service:1;
 219
 220         /*
 221          * Reading the space map roots can fail, so we read it into these
 222          * buffers before the superblock is locked and updated.
 223          */
 224         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 225         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 226 };
 227
 228 struct dm_thin_device {
 229         struct list_head list;
 230         struct dm_pool_metadata *pmd;
 231         dm_thin_id id;
 232
 233         int open_count;
 234         bool changed:1;
 235         bool aborted_with_changes:1;
 236         uint64_t mapped_blocks;
 237         uint64_t transaction_id;
 238         uint32_t creation_time;
 239         uint32_t snapshotted_time;
 240 };
 241
 242 /*----------------------------------------------------------------
 243  * superblock validator
 244  *--------------------------------------------------------------*/
 245
 246 #define SUPERBLOCK_CSUM_XOR 160774
 247
 248 static void sb_prepare_for_write(struct dm_block_validator *v,
 249                                  struct dm_block *b,
 250                                  size_t block_size)
 251 {
 252         struct thin_disk_superblock *disk_super = dm_block_data(b);
 253
 254         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 255         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 256                                                       block_size - sizeof(__le32),
 257                                                       SUPERBLOCK_CSUM_XOR));
 258 }
 259
 260 static int sb_check(struct dm_block_validator *v,
 261                     struct dm_block *b,
 262                     size_t block_size)
 263 {
 264         struct thin_disk_superblock *disk_super = dm_block_data(b);
 265         __le32 csum_le;
 266
 267         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 268                 DMERR("sb_check failed: blocknr %llu: "
 269                       "wanted %llu", le64_to_cpu(disk_super->blocknr),
 270                       (unsigned long long)dm_block_location(b));
 271                 return -ENOTBLK;
 272         }
 273
 274         if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 275                 DMERR("sb_check failed: magic %llu: "
 276                       "wanted %llu", le64_to_cpu(disk_super->magic),
 277                       (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 278                 return -EILSEQ;
 279         }
 280
 281         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 282                                              block_size - sizeof(__le32),
 283                                              SUPERBLOCK_CSUM_XOR));
 284         if (csum_le != disk_super->csum) {
 285                 DMERR("sb_check failed: csum %u: wanted %u",
 286                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 287                 return -EILSEQ;
 288         }
 289
 290         return 0;
 291 }
 292
 293 static struct dm_block_validator sb_validator = {
 294         .name = "superblock",
 295         .prepare_for_write = sb_prepare_for_write,
 296         .check = sb_check
 297 };
 298
 299 /*----------------------------------------------------------------
 300  * Methods for the btree value types
 301  *--------------------------------------------------------------*/
 302
 303 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 304 {
 305         return (b << 24) | t;
 306 }
 307
 308 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 309 {
 310         *b = v >> 24;
 311         *t = v & ((1 << 24) - 1);
 312 }
 313
 314 /*
 315  * It's more efficient to call dm_sm_{inc,dec}_blocks as few times as
 316  * possible.  'with_runs' reads contiguous runs of blocks, and calls the
 317  * given sm function.
 318  */
 319 typedef int (*run_fn)(struct dm_space_map *, dm_block_t, dm_block_t);
 320
 321 static void with_runs(struct dm_space_map *sm, const __le64 *value_le, unsigned count, run_fn fn)
 322 {
 323         uint64_t b, begin, end;
 324         uint32_t t;
 325         bool in_run = false;
 326         unsigned i;
 327
 328         for (i = 0; i < count; i++, value_le++) {
 329                 /* We know value_le is 8 byte aligned */
 330                 unpack_block_time(le64_to_cpu(*value_le), &b, &t);
 331
 332                 if (in_run) {
 333                         if (b == end) {
 334                                 end++;
 335                         } else {
 336                                 fn(sm, begin, end);
 337                                 begin = b;
 338                                 end = b + 1;
 339                         }
 340                 } else {
 341                         in_run = true;
 342                         begin = b;
 343                         end = b + 1;
 344                 }
 345         }
 346
 347         if (in_run)
 348                 fn(sm, begin, end);
 349 }
 350
 351 static void data_block_inc(void *context, const void *value_le, unsigned count)
 352 {
 353         with_runs((struct dm_space_map *) context,
 354                   (const __le64 *) value_le, count, dm_sm_inc_blocks);
 355 }
 356
 357 static void data_block_dec(void *context, const void *value_le, unsigned count)
 358 {
 359         with_runs((struct dm_space_map *) context,
 360                   (const __le64 *) value_le, count, dm_sm_dec_blocks);
 361 }
 362
 363 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 364 {
 365         __le64 v1_le, v2_le;
 366         uint64_t b1, b2;
 367         uint32_t t;
 368
 369         memcpy(&v1_le, value1_le, sizeof(v1_le));
 370         memcpy(&v2_le, value2_le, sizeof(v2_le));
 371         unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 372         unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 373
 374         return b1 == b2;
 375 }
 376
 377 static void subtree_inc(void *context, const void *value, unsigned count)
 378 {
 379         struct dm_btree_info *info = context;
 380         const __le64 *root_le = value;
 381         unsigned i;
 382
 383         for (i = 0; i < count; i++, root_le++)
 384                 dm_tm_inc(info->tm, le64_to_cpu(*root_le));
 385 }
 386
 387 static void subtree_dec(void *context, const void *value, unsigned count)
 388 {
 389         struct dm_btree_info *info = context;
 390         const __le64 *root_le = value;
 391         unsigned i;
 392
 393         for (i = 0; i < count; i++, root_le++)
 394                 if (dm_btree_del(info, le64_to_cpu(*root_le)))
 395                         DMERR("btree delete failed");
 396 }
 397
 398 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 399 {
 400         __le64 v1_le, v2_le;
 401         memcpy(&v1_le, value1_le, sizeof(v1_le));
 402         memcpy(&v2_le, value2_le, sizeof(v2_le));
 403
 404         return v1_le == v2_le;
 405 }
 406
 407 /*----------------------------------------------------------------*/
 408
 409 /*
 410  * Variant that is used for in-core only changes or code that
 411  * shouldn't put the pool in service on its own (e.g. commit).
 412  */
 413 static inline void pmd_write_lock_in_core(struct dm_pool_metadata *pmd)
 414         __acquires(pmd->root_lock)
 415 {
 416         down_write(&pmd->root_lock);
 417 }
 418
 419 static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
 420 {
 421         pmd_write_lock_in_core(pmd);
 422         if (unlikely(!pmd->in_service))
 423                 pmd->in_service = true;
 424 }
 425
 426 static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
 427         __releases(pmd->root_lock)
 428 {
 429         up_write(&pmd->root_lock);
 430 }
 431
 432 /*----------------------------------------------------------------*/
 433
 434 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 435                                 struct dm_block **sblock)
 436 {
 437         return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 438                                      &sb_validator, sblock);
 439 }
 440
 441 static int superblock_lock(struct dm_pool_metadata *pmd,
 442                            struct dm_block **sblock)
 443 {
 444         return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 445                                 &sb_validator, sblock);
 446 }
 447
 448 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 449 {
 450         int r;
 451         unsigned i;
 452         struct dm_block *b;
 453         __le64 *data_le, zero = cpu_to_le64(0);
 454         unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 455
 456         /*
 457          * We can't use a validator here - it may be all zeroes.
 458          */
 459         r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 460         if (r)
 461                 return r;
 462
 463         data_le = dm_block_data(b);
 464         *result = 1;
 465         for (i = 0; i < block_size; i++) {
 466                 if (data_le[i] != zero) {
 467                         *result = 0;
 468                         break;
 469                 }
 470         }
 471
 472         dm_bm_unlock(b);
 473
 474         return 0;
 475 }
 476
 477 static void __setup_btree_details(struct dm_pool_metadata *pmd)
 478 {
 479         pmd->info.tm = pmd->tm;
 480         pmd->info.levels = 2;
 481         pmd->info.value_type.context = pmd->data_sm;
 482         pmd->info.value_type.size = sizeof(__le64);
 483         pmd->info.value_type.inc = data_block_inc;
 484         pmd->info.value_type.dec = data_block_dec;
 485         pmd->info.value_type.equal = data_block_equal;
 486
 487         memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 488         pmd->nb_info.tm = pmd->nb_tm;
 489
 490         pmd->tl_info.tm = pmd->tm;
 491         pmd->tl_info.levels = 1;
 492         pmd->tl_info.value_type.context = &pmd->bl_info;
 493         pmd->tl_info.value_type.size = sizeof(__le64);
 494         pmd->tl_info.value_type.inc = subtree_inc;
 495         pmd->tl_info.value_type.dec = subtree_dec;
 496         pmd->tl_info.value_type.equal = subtree_equal;
 497
 498         pmd->bl_info.tm = pmd->tm;
 499         pmd->bl_info.levels = 1;
 500         pmd->bl_info.value_type.context = pmd->data_sm;
 501         pmd->bl_info.value_type.size = sizeof(__le64);
 502         pmd->bl_info.value_type.inc = data_block_inc;
 503         pmd->bl_info.value_type.dec = data_block_dec;
 504         pmd->bl_info.value_type.equal = data_block_equal;
 505
 506         pmd->details_info.tm = pmd->tm;
 507         pmd->details_info.levels = 1;
 508         pmd->details_info.value_type.context = NULL;
 509         pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 510         pmd->details_info.value_type.inc = NULL;
 511         pmd->details_info.value_type.dec = NULL;
 512         pmd->details_info.value_type.equal = NULL;
 513 }
 514
 515 static int save_sm_roots(struct dm_pool_metadata *pmd)
 516 {
 517         int r;
 518         size_t len;
 519
 520         r = dm_sm_root_size(pmd->metadata_sm, &len);
 521         if (r < 0)
 522                 return r;
 523
 524         r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 525         if (r < 0)
 526                 return r;
 527
 528         r = dm_sm_root_size(pmd->data_sm, &len);
 529         if (r < 0)
 530                 return r;
 531
 532         return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 533 }
 534
 535 static void copy_sm_roots(struct dm_pool_metadata *pmd,
 536                           struct thin_disk_superblock *disk)
 537 {
 538         memcpy(&disk->metadata_space_map_root,
 539                &pmd->metadata_space_map_root,
 540                sizeof(pmd->metadata_space_map_root));
 541
 542         memcpy(&disk->data_space_map_root,
 543                &pmd->data_space_map_root,
 544                sizeof(pmd->data_space_map_root));
 545 }
 546
 547 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 548 {
 549         int r;
 550         struct dm_block *sblock;
 551         struct thin_disk_superblock *disk_super;
 552         sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 553
 554         if (bdev_size > THIN_METADATA_MAX_SECTORS)
 555                 bdev_size = THIN_METADATA_MAX_SECTORS;
 556
 557         r = dm_sm_commit(pmd->data_sm);
 558         if (r < 0)
 559                 return r;
 560
 561         r = dm_tm_pre_commit(pmd->tm);
 562         if (r < 0)
 563                 return r;
 564
 565         r = save_sm_roots(pmd);
 566         if (r < 0)
 567                 return r;
 568
 569         r = superblock_lock_zero(pmd, &sblock);
 570         if (r)
 571                 return r;
 572
 573         disk_super = dm_block_data(sblock);
 574         disk_super->flags = 0;
 575         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 576         disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 577         disk_super->version = cpu_to_le32(THIN_VERSION);
 578         disk_super->time = 0;
 579         disk_super->trans_id = 0;
 580         disk_super->held_root = 0;
 581
 582         copy_sm_roots(pmd, disk_super);
 583
 584         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 585         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 586         disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 587         disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 588         disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 589
 590         return dm_tm_commit(pmd->tm, sblock);
 591 }
 592
 593 static int __format_metadata(struct dm_pool_metadata *pmd)
 594 {
 595         int r;
 596
 597         r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 598                                  &pmd->tm, &pmd->metadata_sm);
 599         if (r < 0) {
 600                 DMERR("tm_create_with_sm failed");
 601                 return r;
 602         }
 603
 604         pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 605         if (IS_ERR(pmd->data_sm)) {
 606                 DMERR("sm_disk_create failed");
 607                 r = PTR_ERR(pmd->data_sm);
 608                 goto bad_cleanup_tm;
 609         }
 610
 611         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 612         if (!pmd->nb_tm) {
 613                 DMERR("could not create non-blocking clone tm");
 614                 r = -ENOMEM;
 615                 goto bad_cleanup_data_sm;
 616         }
 617
 618         __setup_btree_details(pmd);
 619
 620         r = dm_btree_empty(&pmd->info, &pmd->root);
 621         if (r < 0)
 622                 goto bad_cleanup_nb_tm;
 623
 624         r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 625         if (r < 0) {
 626                 DMERR("couldn't create devices root");
 627                 goto bad_cleanup_nb_tm;
 628         }
 629
 630         r = __write_initial_superblock(pmd);
 631         if (r)
 632                 goto bad_cleanup_nb_tm;
 633
 634         return 0;
 635
 636 bad_cleanup_nb_tm:
 637         dm_tm_destroy(pmd->nb_tm);
 638 bad_cleanup_data_sm:
 639         dm_sm_destroy(pmd->data_sm);
 640 bad_cleanup_tm:
 641         dm_tm_destroy(pmd->tm);
 642         dm_sm_destroy(pmd->metadata_sm);
 643
 644         return r;
 645 }
 646
 647 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 648                                      struct dm_pool_metadata *pmd)
 649 {
 650         uint32_t features;
 651
 652         features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 653         if (features) {
 654                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 655                       (unsigned long)features);
 656                 return -EINVAL;
 657         }
 658
 659         /*
 660          * Check for read-only metadata to skip the following RDWR checks.
 661          */
 662         if (bdev_read_only(pmd->bdev))
 663                 return 0;
 664
 665         features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 666         if (features) {
 667                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 668                       (unsigned long)features);
 669                 return -EINVAL;
 670         }
 671
 672         return 0;
 673 }
 674
 675 static int __open_metadata(struct dm_pool_metadata *pmd)
 676 {
 677         int r;
 678         struct dm_block *sblock;
 679         struct thin_disk_superblock *disk_super;
 680
 681         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 682                             &sb_validator, &sblock);
 683         if (r < 0) {
 684                 DMERR("couldn't read superblock");
 685                 return r;
 686         }
 687
 688         disk_super = dm_block_data(sblock);
 689
 690         /* Verify the data block size hasn't changed */
 691         if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 692                 DMERR("changing the data block size (from %u to %llu) is not supported",
 693                       le32_to_cpu(disk_super->data_block_size),
 694                       (unsigned long long)pmd->data_block_size);
 695                 r = -EINVAL;
 696                 goto bad_unlock_sblock;
 697         }
 698
 699         r = __check_incompat_features(disk_super, pmd);
 700         if (r < 0)
 701                 goto bad_unlock_sblock;
 702
 703         r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 704                                disk_super->metadata_space_map_root,
 705                                sizeof(disk_super->metadata_space_map_root),
 706                                &pmd->tm, &pmd->metadata_sm);
 707         if (r < 0) {
 708                 DMERR("tm_open_with_sm failed");
 709                 goto bad_unlock_sblock;
 710         }
 711
 712         pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 713                                        sizeof(disk_super->data_space_map_root));
 714         if (IS_ERR(pmd->data_sm)) {
 715                 DMERR("sm_disk_open failed");
 716                 r = PTR_ERR(pmd->data_sm);
 717                 goto bad_cleanup_tm;
 718         }
 719
 720         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 721         if (!pmd->nb_tm) {
 722                 DMERR("could not create non-blocking clone tm");
 723                 r = -ENOMEM;
 724                 goto bad_cleanup_data_sm;
 725         }
 726
 727         __setup_btree_details(pmd);
 728         dm_bm_unlock(sblock);
 729
 730         return 0;
 731
 732 bad_cleanup_data_sm:
 733         dm_sm_destroy(pmd->data_sm);
 734 bad_cleanup_tm:
 735         dm_tm_destroy(pmd->tm);
 736         dm_sm_destroy(pmd->metadata_sm);
 737 bad_unlock_sblock:
 738         dm_bm_unlock(sblock);
 739
 740         return r;
 741 }
 742
 743 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 744 {
 745         int r, unformatted;
 746
 747         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 748         if (r)
 749                 return r;
 750
 751         if (unformatted)
 752                 return format_device ? __format_metadata(pmd) : -EPERM;
 753
 754         return __open_metadata(pmd);
 755 }
 756
 757 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 758 {
 759         int r;
 760
 761         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 762                                           THIN_MAX_CONCURRENT_LOCKS);
 763         if (IS_ERR(pmd->bm)) {
 764                 DMERR("could not create block manager");
 765                 r = PTR_ERR(pmd->bm);
 766                 pmd->bm = NULL;
 767                 return r;
 768         }
 769
 770         r = __open_or_format_metadata(pmd, format_device);
 771         if (r) {
 772                 dm_block_manager_destroy(pmd->bm);
 773                 pmd->bm = NULL;
 774         }
 775
 776         return r;
 777 }
 778
 779 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 780 {
 781         dm_sm_destroy(pmd->data_sm);
 782         dm_sm_destroy(pmd->metadata_sm);
 783         dm_tm_destroy(pmd->nb_tm);
 784         dm_tm_destroy(pmd->tm);
 785         dm_block_manager_destroy(pmd->bm);
 786 }
 787
 788 static int __begin_transaction(struct dm_pool_metadata *pmd)
 789 {
 790         int r;
 791         struct thin_disk_superblock *disk_super;
 792         struct dm_block *sblock;
 793
 794         /*
 795          * We re-read the superblock every time.  Shouldn't need to do this
 796          * really.
 797          */
 798         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 799                             &sb_validator, &sblock);
 800         if (r)
 801                 return r;
 802
 803         disk_super = dm_block_data(sblock);
 804         pmd->time = le32_to_cpu(disk_super->time);
 805         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 806         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 807         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 808         pmd->flags = le32_to_cpu(disk_super->flags);
 809         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 810
 811         dm_bm_unlock(sblock);
 812         return 0;
 813 }
 814
 815 static int __write_changed_details(struct dm_pool_metadata *pmd)
 816 {
 817         int r;
 818         struct dm_thin_device *td, *tmp;
 819         struct disk_device_details details;
 820         uint64_t key;
 821
 822         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 823                 if (!td->changed)
 824                         continue;
 825
 826                 key = td->id;
 827
 828                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 829                 details.transaction_id = cpu_to_le64(td->transaction_id);
 830                 details.creation_time = cpu_to_le32(td->creation_time);
 831                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 832                 __dm_bless_for_disk(&details);
 833
 834                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 835                                     &key, &details, &pmd->details_root);
 836                 if (r)
 837                         return r;
 838
 839                 if (td->open_count)
 840                         td->changed = false;
 841                 else {
 842                         list_del(&td->list);
 843                         kfree(td);
 844                 }
 845         }
 846
 847         return 0;
 848 }
 849
 850 static int __commit_transaction(struct dm_pool_metadata *pmd)
 851 {
 852         int r;
 853         struct thin_disk_superblock *disk_super;
 854         struct dm_block *sblock;
 855
 856         /*
 857          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 858          */
 859         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 860         BUG_ON(!rwsem_is_locked(&pmd->root_lock));
 861
 862         if (unlikely(!pmd->in_service))
 863                 return 0;
 864
 865         if (pmd->pre_commit_fn) {
 866                 r = pmd->pre_commit_fn(pmd->pre_commit_context);
 867                 if (r < 0) {
 868                         DMERR("pre-commit callback failed");
 869                         return r;
 870                 }
 871         }
 872
 873         r = __write_changed_details(pmd);
 874         if (r < 0)
 875                 return r;
 876
 877         r = dm_sm_commit(pmd->data_sm);
 878         if (r < 0)
 879                 return r;
 880
 881         r = dm_tm_pre_commit(pmd->tm);
 882         if (r < 0)
 883                 return r;
 884
 885         r = save_sm_roots(pmd);
 886         if (r < 0)
 887                 return r;
 888
 889         r = superblock_lock(pmd, &sblock);
 890         if (r)
 891                 return r;
 892
 893         disk_super = dm_block_data(sblock);
 894         disk_super->time = cpu_to_le32(pmd->time);
 895         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 896         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 897         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 898         disk_super->flags = cpu_to_le32(pmd->flags);
 899
 900         copy_sm_roots(pmd, disk_super);
 901
 902         return dm_tm_commit(pmd->tm, sblock);
 903 }
 904
 905 static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 906 {
 907         int r;
 908         dm_block_t total;
 909         dm_block_t max_blocks = 4096; /* 16M */
 910
 911         r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 912         if (r) {
 913                 DMERR("could not get size of metadata device");
 914                 pmd->metadata_reserve = max_blocks;
 915         } else
 916                 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 917 }
 918
 919 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 920                                                sector_t data_block_size,
 921                                                bool format_device)
 922 {
 923         int r;
 924         struct dm_pool_metadata *pmd;
 925
 926         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 927         if (!pmd) {
 928                 DMERR("could not allocate metadata struct");
 929                 return ERR_PTR(-ENOMEM);
 930         }
 931
 932         init_rwsem(&pmd->root_lock);
 933         pmd->time = 0;
 934         INIT_LIST_HEAD(&pmd->thin_devices);
 935         pmd->fail_io = false;
 936         pmd->in_service = false;
 937         pmd->bdev = bdev;
 938         pmd->data_block_size = data_block_size;
 939         pmd->pre_commit_fn = NULL;
 940         pmd->pre_commit_context = NULL;
 941
 942         r = __create_persistent_data_objects(pmd, format_device);
 943         if (r) {
 944                 kfree(pmd);
 945                 return ERR_PTR(r);
 946         }
 947
 948         r = __begin_transaction(pmd);
 949         if (r < 0) {
 950                 if (dm_pool_metadata_close(pmd) < 0)
 951                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 952                 return ERR_PTR(r);
 953         }
 954
 955         __set_metadata_reserve(pmd);
 956
 957         return pmd;
 958 }
 959
 960 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 961 {
 962         int r;
 963         unsigned open_devices = 0;
 964         struct dm_thin_device *td, *tmp;
 965
 966         down_read(&pmd->root_lock);
 967         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 968                 if (td->open_count)
 969                         open_devices++;
 970                 else {
 971                         list_del(&td->list);
 972                         kfree(td);
 973                 }
 974         }
 975         up_read(&pmd->root_lock);
 976
 977         if (open_devices) {
 978                 DMERR("attempt to close pmd when %u device(s) are still open",
 979                        open_devices);
 980                 return -EBUSY;
 981         }
 982
 983         pmd_write_lock_in_core(pmd);
 984         if (!pmd->fail_io && !dm_bm_is_read_only(pmd->bm)) {
 985                 r = __commit_transaction(pmd);
 986                 if (r < 0)
 987                         DMWARN("%s: __commit_transaction() failed, error = %d",
 988                                __func__, r);
 989         }
 990         pmd_write_unlock(pmd);
 991         if (!pmd->fail_io)
 992                 __destroy_persistent_data_objects(pmd);
 993
 994         kfree(pmd);
 995         return 0;
 996 }
 997
 998 /*
 999  * __open_device: Returns @td corresponding to device with id @dev,
1000  * creating it if @create is set and incrementing @td->open_count.
1001  * On failure, @td is undefined.
1002  */
1003 static int __open_device(struct dm_pool_metadata *pmd,
1004                          dm_thin_id dev, int create,
1005                          struct dm_thin_device **td)
1006 {
1007         int r, changed = 0;
1008         struct dm_thin_device *td2;
1009         uint64_t key = dev;
1010         struct disk_device_details details_le;
1011
1012         /*
1013          * If the device is already open, return it.
1014          */
1015         list_for_each_entry(td2, &pmd->thin_devices, list)
1016                 if (td2->id == dev) {
1017                         /*
1018                          * May not create an already-open device.
1019                          */
1020                         if (create)
1021                                 return -EEXIST;
1022
1023                         td2->open_count++;
1024                         *td = td2;
1025                         return 0;
1026                 }
1027
1028         /*
1029          * Check the device exists.
1030          */
1031         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1032                             &key, &details_le);
1033         if (r) {
1034                 if (r != -ENODATA || !create)
1035                         return r;
1036
1037                 /*
1038                  * Create new device.
1039                  */
1040                 changed = 1;
1041                 details_le.mapped_blocks = 0;
1042                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
1043                 details_le.creation_time = cpu_to_le32(pmd->time);
1044                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
1045         }
1046
1047         *td = kmalloc(sizeof(**td), GFP_NOIO);
1048         if (!*td)
1049                 return -ENOMEM;
1050
1051         (*td)->pmd = pmd;
1052         (*td)->id = dev;
1053         (*td)->open_count = 1;
1054         (*td)->changed = changed;
1055         (*td)->aborted_with_changes = false;
1056         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1057         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1058         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
1059         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1060
1061         list_add(&(*td)->list, &pmd->thin_devices);
1062
1063         return 0;
1064 }
1065
1066 static void __close_device(struct dm_thin_device *td)
1067 {
1068         --td->open_count;
1069 }
1070
1071 static int __create_thin(struct dm_pool_metadata *pmd,
1072                          dm_thin_id dev)
1073 {
1074         int r;
1075         dm_block_t dev_root;
1076         uint64_t key = dev;
1077         struct dm_thin_device *td;
1078         __le64 value;
1079
1080         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1081                             &key, NULL);
1082         if (!r)
1083                 return -EEXIST;
1084
1085         /*
1086          * Create an empty btree for the mappings.
1087          */
1088         r = dm_btree_empty(&pmd->bl_info, &dev_root);
1089         if (r)
1090                 return r;
1091
1092         /*
1093          * Insert it into the main mapping tree.
1094          */
1095         value = cpu_to_le64(dev_root);
1096         __dm_bless_for_disk(&value);
1097         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1098         if (r) {
1099                 dm_btree_del(&pmd->bl_info, dev_root);
1100                 return r;
1101         }
1102
1103         r = __open_device(pmd, dev, 1, &td);
1104         if (r) {
1105                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1106                 dm_btree_del(&pmd->bl_info, dev_root);
1107                 return r;
1108         }
1109         __close_device(td);
1110
1111         return r;
1112 }
1113
1114 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1115 {
1116         int r = -EINVAL;
1117
1118         pmd_write_lock(pmd);
1119         if (!pmd->fail_io)
1120                 r = __create_thin(pmd, dev);
1121         pmd_write_unlock(pmd);
1122
1123         return r;
1124 }
1125
1126 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1127                                   struct dm_thin_device *snap,
1128                                   dm_thin_id origin, uint32_t time)
1129 {
1130         int r;
1131         struct dm_thin_device *td;
1132
1133         r = __open_device(pmd, origin, 0, &td);
1134         if (r)
1135                 return r;
1136
1137         td->changed = true;
1138         td->snapshotted_time = time;
1139
1140         snap->mapped_blocks = td->mapped_blocks;
1141         snap->snapshotted_time = time;
1142         __close_device(td);
1143
1144         return 0;
1145 }
1146
1147 static int __create_snap(struct dm_pool_metadata *pmd,
1148                          dm_thin_id dev, dm_thin_id origin)
1149 {
1150         int r;
1151         dm_block_t origin_root;
1152         uint64_t key = origin, dev_key = dev;
1153         struct dm_thin_device *td;
1154         __le64 value;
1155
1156         /* check this device is unused */
1157         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1158                             &dev_key, NULL);
1159         if (!r)
1160                 return -EEXIST;
1161
1162         /* find the mapping tree for the origin */
1163         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1164         if (r)
1165                 return r;
1166         origin_root = le64_to_cpu(value);
1167
1168         /* clone the origin, an inc will do */
1169         dm_tm_inc(pmd->tm, origin_root);
1170
1171         /* insert into the main mapping tree */
1172         value = cpu_to_le64(origin_root);
1173         __dm_bless_for_disk(&value);
1174         key = dev;
1175         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1176         if (r) {
1177                 dm_tm_dec(pmd->tm, origin_root);
1178                 return r;
1179         }
1180
1181         pmd->time++;
1182
1183         r = __open_device(pmd, dev, 1, &td);
1184         if (r)
1185                 goto bad;
1186
1187         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1188         __close_device(td);
1189
1190         if (r)
1191                 goto bad;
1192
1193         return 0;
1194
1195 bad:
1196         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1197         dm_btree_remove(&pmd->details_info, pmd->details_root,
1198                         &key, &pmd->details_root);
1199         return r;
1200 }
1201
1202 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1203                                  dm_thin_id dev,
1204                                  dm_thin_id origin)
1205 {
1206         int r = -EINVAL;
1207
1208         pmd_write_lock(pmd);
1209         if (!pmd->fail_io)
1210                 r = __create_snap(pmd, dev, origin);
1211         pmd_write_unlock(pmd);
1212
1213         return r;
1214 }
1215
1216 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1217 {
1218         int r;
1219         uint64_t key = dev;
1220         struct dm_thin_device *td;
1221
1222         /* TODO: failure should mark the transaction invalid */
1223         r = __open_device(pmd, dev, 0, &td);
1224         if (r)
1225                 return r;
1226
1227         if (td->open_count > 1) {
1228                 __close_device(td);
1229                 return -EBUSY;
1230         }
1231
1232         list_del(&td->list);
1233         kfree(td);
1234         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1235                             &key, &pmd->details_root);
1236         if (r)
1237                 return r;
1238
1239         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1240         if (r)
1241                 return r;
1242
1243         return 0;
1244 }
1245
1246 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1247                                dm_thin_id dev)
1248 {
1249         int r = -EINVAL;
1250
1251         pmd_write_lock(pmd);
1252         if (!pmd->fail_io)
1253                 r = __delete_device(pmd, dev);
1254         pmd_write_unlock(pmd);
1255
1256         return r;
1257 }
1258
1259 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1260                                         uint64_t current_id,
1261                                         uint64_t new_id)
1262 {
1263         int r = -EINVAL;
1264
1265         pmd_write_lock(pmd);
1266
1267         if (pmd->fail_io)
1268                 goto out;
1269
1270         if (pmd->trans_id != current_id) {
1271                 DMERR("mismatched transaction id");
1272                 goto out;
1273         }
1274
1275         pmd->trans_id = new_id;
1276         r = 0;
1277
1278 out:
1279         pmd_write_unlock(pmd);
1280
1281         return r;
1282 }
1283
1284 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1285                                         uint64_t *result)
1286 {
1287         int r = -EINVAL;
1288
1289         down_read(&pmd->root_lock);
1290         if (!pmd->fail_io) {
1291                 *result = pmd->trans_id;
1292                 r = 0;
1293         }
1294         up_read(&pmd->root_lock);
1295
1296         return r;
1297 }
1298
1299 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1300 {
1301         int r, inc;
1302         struct thin_disk_superblock *disk_super;
1303         struct dm_block *copy, *sblock;
1304         dm_block_t held_root;
1305
1306         /*
1307          * We commit to ensure the btree roots which we increment in a
1308          * moment are up to date.
1309          */
1310         r = __commit_transaction(pmd);
1311         if (r < 0) {
1312                 DMWARN("%s: __commit_transaction() failed, error = %d",
1313                        __func__, r);
1314                 return r;
1315         }
1316
1317         /*
1318          * Copy the superblock.
1319          */
1320         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1321         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1322                                &sb_validator, &copy, &inc);
1323         if (r)
1324                 return r;
1325
1326         BUG_ON(!inc);
1327
1328         held_root = dm_block_location(copy);
1329         disk_super = dm_block_data(copy);
1330
1331         if (le64_to_cpu(disk_super->held_root)) {
1332                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1333
1334                 dm_tm_dec(pmd->tm, held_root);
1335                 dm_tm_unlock(pmd->tm, copy);
1336                 return -EBUSY;
1337         }
1338
1339         /*
1340          * Wipe the spacemap since we're not publishing this.
1341          */
1342         memset(&disk_super->data_space_map_root, 0,
1343                sizeof(disk_super->data_space_map_root));
1344         memset(&disk_super->metadata_space_map_root, 0,
1345                sizeof(disk_super->metadata_space_map_root));
1346
1347         /*
1348          * Increment the data structures that need to be preserved.
1349          */
1350         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1351         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1352         dm_tm_unlock(pmd->tm, copy);
1353
1354         /*
1355          * Write the held root into the superblock.
1356          */
1357         r = superblock_lock(pmd, &sblock);
1358         if (r) {
1359                 dm_tm_dec(pmd->tm, held_root);
1360                 return r;
1361         }
1362
1363         disk_super = dm_block_data(sblock);
1364         disk_super->held_root = cpu_to_le64(held_root);
1365         dm_bm_unlock(sblock);
1366         return 0;
1367 }
1368
1369 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1370 {
1371         int r = -EINVAL;
1372
1373         pmd_write_lock(pmd);
1374         if (!pmd->fail_io)
1375                 r = __reserve_metadata_snap(pmd);
1376         pmd_write_unlock(pmd);
1377
1378         return r;
1379 }
1380
1381 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1382 {
1383         int r;
1384         struct thin_disk_superblock *disk_super;
1385         struct dm_block *sblock, *copy;
1386         dm_block_t held_root;
1387
1388         r = superblock_lock(pmd, &sblock);
1389         if (r)
1390                 return r;
1391
1392         disk_super = dm_block_data(sblock);
1393         held_root = le64_to_cpu(disk_super->held_root);
1394         disk_super->held_root = cpu_to_le64(0);
1395
1396         dm_bm_unlock(sblock);
1397
1398         if (!held_root) {
1399                 DMWARN("No pool metadata snapshot found: nothing to release.");
1400                 return -EINVAL;
1401         }
1402
1403         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1404         if (r)
1405                 return r;
1406
1407         disk_super = dm_block_data(copy);
1408         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1409         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1410         dm_sm_dec_block(pmd->metadata_sm, held_root);
1411
1412         dm_tm_unlock(pmd->tm, copy);
1413
1414         return 0;
1415 }
1416
1417 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1418 {
1419         int r = -EINVAL;
1420
1421         pmd_write_lock(pmd);
1422         if (!pmd->fail_io)
1423                 r = __release_metadata_snap(pmd);
1424         pmd_write_unlock(pmd);
1425
1426         return r;
1427 }
1428
1429 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1430                                dm_block_t *result)
1431 {
1432         int r;
1433         struct thin_disk_superblock *disk_super;
1434         struct dm_block *sblock;
1435
1436         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1437                             &sb_validator, &sblock);
1438         if (r)
1439                 return r;
1440
1441         disk_super = dm_block_data(sblock);
1442         *result = le64_to_cpu(disk_super->held_root);
1443
1444         dm_bm_unlock(sblock);
1445
1446         return 0;
1447 }
1448
1449 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1450                               dm_block_t *result)
1451 {
1452         int r = -EINVAL;
1453
1454         down_read(&pmd->root_lock);
1455         if (!pmd->fail_io)
1456                 r = __get_metadata_snap(pmd, result);
1457         up_read(&pmd->root_lock);
1458
1459         return r;
1460 }
1461
1462 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1463                              struct dm_thin_device **td)
1464 {
1465         int r = -EINVAL;
1466
1467         pmd_write_lock_in_core(pmd);
1468         if (!pmd->fail_io)
1469                 r = __open_device(pmd, dev, 0, td);
1470         pmd_write_unlock(pmd);
1471
1472         return r;
1473 }
1474
1475 int dm_pool_close_thin_device(struct dm_thin_device *td)
1476 {
1477         pmd_write_lock_in_core(td->pmd);
1478         __close_device(td);
1479         pmd_write_unlock(td->pmd);
1480
1481         return 0;
1482 }
1483
1484 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1485 {
1486         return td->id;
1487 }
1488
1489 /*
1490  * Check whether @time (of block creation) is older than @td's last snapshot.
1491  * If so then the associated block is shared with the last snapshot device.
1492  * Any block on a device created *after* the device last got snapshotted is
1493  * necessarily not shared.
1494  */
1495 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1496 {
1497         return td->snapshotted_time > time;
1498 }
1499
1500 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1501                                  struct dm_thin_lookup_result *result)
1502 {
1503         uint64_t block_time = 0;
1504         dm_block_t exception_block;
1505         uint32_t exception_time;
1506
1507         block_time = le64_to_cpu(value);
1508         unpack_block_time(block_time, &exception_block, &exception_time);
1509         result->block = exception_block;
1510         result->shared = __snapshotted_since(td, exception_time);
1511 }
1512
1513 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1514                         int can_issue_io, struct dm_thin_lookup_result *result)
1515 {
1516         int r;
1517         __le64 value;
1518         struct dm_pool_metadata *pmd = td->pmd;
1519         dm_block_t keys[2] = { td->id, block };
1520         struct dm_btree_info *info;
1521
1522         if (can_issue_io) {
1523                 info = &pmd->info;
1524         } else
1525                 info = &pmd->nb_info;
1526
1527         r = dm_btree_lookup(info, pmd->root, keys, &value);
1528         if (!r)
1529                 unpack_lookup_result(td, value, result);
1530
1531         return r;
1532 }
1533
1534 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1535                        int can_issue_io, struct dm_thin_lookup_result *result)
1536 {
1537         int r;
1538         struct dm_pool_metadata *pmd = td->pmd;
1539
1540         down_read(&pmd->root_lock);
1541         if (pmd->fail_io) {
1542                 up_read(&pmd->root_lock);
1543                 return -EINVAL;
1544         }
1545
1546         r = __find_block(td, block, can_issue_io, result);
1547
1548         up_read(&pmd->root_lock);
1549         return r;
1550 }
1551
1552 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1553                                           dm_block_t *vblock,
1554                                           struct dm_thin_lookup_result *result)
1555 {
1556         int r;
1557         __le64 value;
1558         struct dm_pool_metadata *pmd = td->pmd;
1559         dm_block_t keys[2] = { td->id, block };
1560
1561         r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1562         if (!r)
1563                 unpack_lookup_result(td, value, result);
1564
1565         return r;
1566 }
1567
1568 static int __find_mapped_range(struct dm_thin_device *td,
1569                                dm_block_t begin, dm_block_t end,
1570                                dm_block_t *thin_begin, dm_block_t *thin_end,
1571                                dm_block_t *pool_begin, bool *maybe_shared)
1572 {
1573         int r;
1574         dm_block_t pool_end;
1575         struct dm_thin_lookup_result lookup;
1576
1577         if (end < begin)
1578                 return -ENODATA;
1579
1580         r = __find_next_mapped_block(td, begin, &begin, &lookup);
1581         if (r)
1582                 return r;
1583
1584         if (begin >= end)
1585                 return -ENODATA;
1586
1587         *thin_begin = begin;
1588         *pool_begin = lookup.block;
1589         *maybe_shared = lookup.shared;
1590
1591         begin++;
1592         pool_end = *pool_begin + 1;
1593         while (begin != end) {
1594                 r = __find_block(td, begin, true, &lookup);
1595                 if (r) {
1596                         if (r == -ENODATA)
1597                                 break;
1598                         else
1599                                 return r;
1600                 }
1601
1602                 if ((lookup.block != pool_end) ||
1603                     (lookup.shared != *maybe_shared))
1604                         break;
1605
1606                 pool_end++;
1607                 begin++;
1608         }
1609
1610         *thin_end = begin;
1611         return 0;
1612 }
1613
1614 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1615                               dm_block_t begin, dm_block_t end,
1616                               dm_block_t *thin_begin, dm_block_t *thin_end,
1617                               dm_block_t *pool_begin, bool *maybe_shared)
1618 {
1619         int r = -EINVAL;
1620         struct dm_pool_metadata *pmd = td->pmd;
1621
1622         down_read(&pmd->root_lock);
1623         if (!pmd->fail_io) {
1624                 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1625                                         pool_begin, maybe_shared);
1626         }
1627         up_read(&pmd->root_lock);
1628
1629         return r;
1630 }
1631
1632 static int __insert(struct dm_thin_device *td, dm_block_t block,
1633                     dm_block_t data_block)
1634 {
1635         int r, inserted;
1636         __le64 value;
1637         struct dm_pool_metadata *pmd = td->pmd;
1638         dm_block_t keys[2] = { td->id, block };
1639
1640         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1641         __dm_bless_for_disk(&value);
1642
1643         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1644                                    &pmd->root, &inserted);
1645         if (r)
1646                 return r;
1647
1648         td->changed = true;
1649         if (inserted)
1650                 td->mapped_blocks++;
1651
1652         return 0;
1653 }
1654
1655 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1656                          dm_block_t data_block)
1657 {
1658         int r = -EINVAL;
1659
1660         pmd_write_lock(td->pmd);
1661         if (!td->pmd->fail_io)
1662                 r = __insert(td, block, data_block);
1663         pmd_write_unlock(td->pmd);
1664
1665         return r;
1666 }
1667
1668 static int __remove(struct dm_thin_device *td, dm_block_t block)
1669 {
1670         int r;
1671         struct dm_pool_metadata *pmd = td->pmd;
1672         dm_block_t keys[2] = { td->id, block };
1673
1674         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1675         if (r)
1676                 return r;
1677
1678         td->mapped_blocks--;
1679         td->changed = true;
1680
1681         return 0;
1682 }
1683
1684 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1685 {
1686         int r;
1687         unsigned count, total_count = 0;
1688         struct dm_pool_metadata *pmd = td->pmd;
1689         dm_block_t keys[1] = { td->id };
1690         __le64 value;
1691         dm_block_t mapping_root;
1692
1693         /*
1694          * Find the mapping tree
1695          */
1696         r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1697         if (r)
1698                 return r;
1699
1700         /*
1701          * Remove from the mapping tree, taking care to inc the
1702          * ref count so it doesn't get deleted.
1703          */
1704         mapping_root = le64_to_cpu(value);
1705         dm_tm_inc(pmd->tm, mapping_root);
1706         r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1707         if (r)
1708                 return r;
1709
1710         /*
1711          * Remove leaves stops at the first unmapped entry, so we have to
1712          * loop round finding mapped ranges.
1713          */
1714         while (begin < end) {
1715                 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1716                 if (r == -ENODATA)
1717                         break;
1718
1719                 if (r)
1720                         return r;
1721
1722                 if (begin >= end)
1723                         break;
1724
1725                 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1726                 if (r)
1727                         return r;
1728
1729                 total_count += count;
1730         }
1731
1732         td->mapped_blocks -= total_count;
1733         td->changed = true;
1734
1735         /*
1736          * Reinsert the mapping tree.
1737          */
1738         value = cpu_to_le64(mapping_root);
1739         __dm_bless_for_disk(&value);
1740         return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1741 }
1742
1743 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1744 {
1745         int r = -EINVAL;
1746
1747         pmd_write_lock(td->pmd);
1748         if (!td->pmd->fail_io)
1749                 r = __remove(td, block);
1750         pmd_write_unlock(td->pmd);
1751
1752         return r;
1753 }
1754
1755 int dm_thin_remove_range(struct dm_thin_device *td,
1756                          dm_block_t begin, dm_block_t end)
1757 {
1758         int r = -EINVAL;
1759
1760         pmd_write_lock(td->pmd);
1761         if (!td->pmd->fail_io)
1762                 r = __remove_range(td, begin, end);
1763         pmd_write_unlock(td->pmd);
1764
1765         return r;
1766 }
1767
1768 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1769 {
1770         int r;
1771         uint32_t ref_count;
1772
1773         down_read(&pmd->root_lock);
1774         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1775         if (!r)
1776                 *result = (ref_count > 1);
1777         up_read(&pmd->root_lock);
1778
1779         return r;
1780 }
1781
1782 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1783 {
1784         int r = 0;
1785
1786         pmd_write_lock(pmd);
1787         r = dm_sm_inc_blocks(pmd->data_sm, b, e);
1788         pmd_write_unlock(pmd);
1789
1790         return r;
1791 }
1792
1793 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1794 {
1795         int r = 0;
1796
1797         pmd_write_lock(pmd);
1798         r = dm_sm_dec_blocks(pmd->data_sm, b, e);
1799         pmd_write_unlock(pmd);
1800
1801         return r;
1802 }
1803
1804 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1805 {
1806         int r;
1807
1808         down_read(&td->pmd->root_lock);
1809         r = td->changed;
1810         up_read(&td->pmd->root_lock);
1811
1812         return r;
1813 }
1814
1815 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1816 {
1817         bool r = false;
1818         struct dm_thin_device *td, *tmp;
1819
1820         down_read(&pmd->root_lock);
1821         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1822                 if (td->changed) {
1823                         r = td->changed;
1824                         break;
1825                 }
1826         }
1827         up_read(&pmd->root_lock);
1828
1829         return r;
1830 }
1831
1832 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1833 {
1834         bool r;
1835
1836         down_read(&td->pmd->root_lock);
1837         r = td->aborted_with_changes;
1838         up_read(&td->pmd->root_lock);
1839
1840         return r;
1841 }
1842
1843 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1844 {
1845         int r = -EINVAL;
1846
1847         pmd_write_lock(pmd);
1848         if (!pmd->fail_io)
1849                 r = dm_sm_new_block(pmd->data_sm, result);
1850         pmd_write_unlock(pmd);
1851
1852         return r;
1853 }
1854
1855 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1856 {
1857         int r = -EINVAL;
1858
1859         /*
1860          * Care is taken to not have commit be what
1861          * triggers putting the thin-pool in-service.
1862          */
1863         pmd_write_lock_in_core(pmd);
1864         if (pmd->fail_io)
1865                 goto out;
1866
1867         r = __commit_transaction(pmd);
1868         if (r < 0)
1869                 goto out;
1870
1871         /*
1872          * Open the next transaction.
1873          */
1874         r = __begin_transaction(pmd);
1875 out:
1876         pmd_write_unlock(pmd);
1877         return r;
1878 }
1879
1880 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1881 {
1882         struct dm_thin_device *td;
1883
1884         list_for_each_entry(td, &pmd->thin_devices, list)
1885                 td->aborted_with_changes = td->changed;
1886 }
1887
1888 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1889 {
1890         int r = -EINVAL;
1891
1892         pmd_write_lock(pmd);
1893         if (pmd->fail_io)
1894                 goto out;
1895
1896         __set_abort_with_changes_flags(pmd);
1897         __destroy_persistent_data_objects(pmd);
1898         r = __create_persistent_data_objects(pmd, false);
1899         if (r)
1900                 pmd->fail_io = true;
1901
1902 out:
1903         pmd_write_unlock(pmd);
1904
1905         return r;
1906 }
1907
1908 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1909 {
1910         int r = -EINVAL;
1911
1912         down_read(&pmd->root_lock);
1913         if (!pmd->fail_io)
1914                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1915         up_read(&pmd->root_lock);
1916
1917         return r;
1918 }
1919
1920 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1921                                           dm_block_t *result)
1922 {
1923         int r = -EINVAL;
1924
1925         down_read(&pmd->root_lock);
1926         if (!pmd->fail_io)
1927                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1928
1929         if (!r) {
1930                 if (*result < pmd->metadata_reserve)
1931                         *result = 0;
1932                 else
1933                         *result -= pmd->metadata_reserve;
1934         }
1935         up_read(&pmd->root_lock);
1936
1937         return r;
1938 }
1939
1940 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1941                                   dm_block_t *result)
1942 {
1943         int r = -EINVAL;
1944
1945         down_read(&pmd->root_lock);
1946         if (!pmd->fail_io)
1947                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1948         up_read(&pmd->root_lock);
1949
1950         return r;
1951 }
1952
1953 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1954 {
1955         int r = -EINVAL;
1956
1957         down_read(&pmd->root_lock);
1958         if (!pmd->fail_io)
1959                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1960         up_read(&pmd->root_lock);
1961
1962         return r;
1963 }
1964
1965 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1966 {
1967         int r = -EINVAL;
1968         struct dm_pool_metadata *pmd = td->pmd;
1969
1970         down_read(&pmd->root_lock);
1971         if (!pmd->fail_io) {
1972                 *result = td->mapped_blocks;
1973                 r = 0;
1974         }
1975         up_read(&pmd->root_lock);
1976
1977         return r;
1978 }
1979
1980 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1981 {
1982         int r;
1983         __le64 value_le;
1984         dm_block_t thin_root;
1985         struct dm_pool_metadata *pmd = td->pmd;
1986
1987         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1988         if (r)
1989                 return r;
1990
1991         thin_root = le64_to_cpu(value_le);
1992
1993         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1994 }
1995
1996 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1997                                      dm_block_t *result)
1998 {
1999         int r = -EINVAL;
2000         struct dm_pool_metadata *pmd = td->pmd;
2001
2002         down_read(&pmd->root_lock);
2003         if (!pmd->fail_io)
2004                 r = __highest_block(td, result);
2005         up_read(&pmd->root_lock);
2006
2007         return r;
2008 }
2009
2010 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
2011 {
2012         int r;
2013         dm_block_t old_count;
2014
2015         r = dm_sm_get_nr_blocks(sm, &old_count);
2016         if (r)
2017                 return r;
2018
2019         if (new_count == old_count)
2020                 return 0;
2021
2022         if (new_count < old_count) {
2023                 DMERR("cannot reduce size of space map");
2024                 return -EINVAL;
2025         }
2026
2027         return dm_sm_extend(sm, new_count - old_count);
2028 }
2029
2030 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2031 {
2032         int r = -EINVAL;
2033
2034         pmd_write_lock(pmd);
2035         if (!pmd->fail_io)
2036                 r = __resize_space_map(pmd->data_sm, new_count);
2037         pmd_write_unlock(pmd);
2038
2039         return r;
2040 }
2041
2042 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2043 {
2044         int r = -EINVAL;
2045
2046         pmd_write_lock(pmd);
2047         if (!pmd->fail_io) {
2048                 r = __resize_space_map(pmd->metadata_sm, new_count);
2049                 if (!r)
2050                         __set_metadata_reserve(pmd);
2051         }
2052         pmd_write_unlock(pmd);
2053
2054         return r;
2055 }
2056
2057 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2058 {
2059         pmd_write_lock_in_core(pmd);
2060         dm_bm_set_read_only(pmd->bm);
2061         pmd_write_unlock(pmd);
2062 }
2063
2064 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2065 {
2066         pmd_write_lock_in_core(pmd);
2067         dm_bm_set_read_write(pmd->bm);
2068         pmd_write_unlock(pmd);
2069 }
2070
2071 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2072                                         dm_block_t threshold,
2073                                         dm_sm_threshold_fn fn,
2074                                         void *context)
2075 {
2076         int r;
2077
2078         pmd_write_lock_in_core(pmd);
2079         r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
2080         pmd_write_unlock(pmd);
2081
2082         return r;
2083 }
2084
2085 void dm_pool_register_pre_commit_callback(struct dm_pool_metadata *pmd,
2086                                           dm_pool_pre_commit_fn fn,
2087                                           void *context)
2088 {
2089         pmd_write_lock_in_core(pmd);
2090         pmd->pre_commit_fn = fn;
2091         pmd->pre_commit_context = context;
2092         pmd_write_unlock(pmd);
2093 }
2094
2095 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2096 {
2097         int r = -EINVAL;
2098         struct dm_block *sblock;
2099         struct thin_disk_superblock *disk_super;
2100
2101         pmd_write_lock(pmd);
2102         if (pmd->fail_io)
2103                 goto out;
2104
2105         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2106
2107         r = superblock_lock(pmd, &sblock);
2108         if (r) {
2109                 DMERR("couldn't lock superblock");
2110                 goto out;
2111         }
2112
2113         disk_super = dm_block_data(sblock);
2114         disk_super->flags = cpu_to_le32(pmd->flags);
2115
2116         dm_bm_unlock(sblock);
2117 out:
2118         pmd_write_unlock(pmd);
2119         return r;
2120 }
2121
2122 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2123 {
2124         bool needs_check;
2125
2126         down_read(&pmd->root_lock);
2127         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2128         up_read(&pmd->root_lock);
2129
2130         return needs_check;
2131 }
2132
2133 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2134 {
2135         down_read(&pmd->root_lock);
2136         if (!pmd->fail_io)
2137                 dm_tm_issue_prefetches(pmd->tm);
2138         up_read(&pmd->root_lock);
2139 }