drivers/md/dm-thin-metadata.c

   1 /*
   2  * Copyright (C) 2011-2012 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-thin-metadata.h"
   8 #include "persistent-data/dm-btree.h"
   9 #include "persistent-data/dm-space-map.h"
  10 #include "persistent-data/dm-space-map-disk.h"
  11 #include "persistent-data/dm-transaction-manager.h"
  12
  13 #include <linux/list.h>
  14 #include <linux/device-mapper.h>
  15 #include <linux/workqueue.h>
  16
  17 /*--------------------------------------------------------------------------
  18  * As far as the metadata goes, there is:
  19  *
  20  * - A superblock in block zero, taking up fewer than 512 bytes for
  21  *   atomic writes.
  22  *
  23  * - A space map managing the metadata blocks.
  24  *
  25  * - A space map managing the data blocks.
  26  *
  27  * - A btree mapping our internal thin dev ids onto struct disk_device_details.
  28  *
  29  * - A hierarchical btree, with 2 levels which effectively maps (thin
  30  *   dev id, virtual block) -> block_time.  Block time is a 64-bit
  31  *   field holding the time in the low 24 bits, and block in the top 48
  32  *   bits.
  33  *
  34  * BTrees consist solely of btree_nodes, that fill a block.  Some are
  35  * internal nodes, as such their values are a __le64 pointing to other
  36  * nodes.  Leaf nodes can store data of any reasonable size (ie. much
  37  * smaller than the block size).  The nodes consist of the header,
  38  * followed by an array of keys, followed by an array of values.  We have
  39  * to binary search on the keys so they're all held together to help the
  40  * cpu cache.
  41  *
  42  * Space maps have 2 btrees:
  43  *
  44  * - One maps a uint64_t onto a struct index_entry.  Which points to a
  45  *   bitmap block, and has some details about how many free entries there
  46  *   are etc.
  47  *
  48  * - The bitmap blocks have a header (for the checksum).  Then the rest
  49  *   of the block is pairs of bits.  With the meaning being:
  50  *
  51  *   0 - ref count is 0
  52  *   1 - ref count is 1
  53  *   2 - ref count is 2
  54  *   3 - ref count is higher than 2
  55  *
  56  * - If the count is higher than 2 then the ref count is entered in a
  57  *   second btree that directly maps the block_address to a uint32_t ref
  58  *   count.
  59  *
  60  * The space map metadata variant doesn't have a bitmaps btree.  Instead
  61  * it has one single blocks worth of index_entries.  This avoids
  62  * recursive issues with the bitmap btree needing to allocate space in
  63  * order to insert.  With a small data block size such as 64k the
  64  * metadata support data devices that are hundreds of terrabytes.
  65  *
  66  * The space maps allocate space linearly from front to back.  Space that
  67  * is freed in a transaction is never recycled within that transaction.
  68  * To try and avoid fragmenting _free_ space the allocator always goes
  69  * back and fills in gaps.
  70  *
  71  * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
  72  * from the block manager.
  73  *--------------------------------------------------------------------------*/
  74
  75 #define DM_MSG_PREFIX   "thin metadata"
  76
  77 #define THIN_SUPERBLOCK_MAGIC 27022010
  78 #define THIN_SUPERBLOCK_LOCATION 0
  79 #define THIN_VERSION 2
  80 #define SECTOR_TO_BLOCK_SHIFT 3
  81
  82 /*
  83  * For btree insert:
  84  *  3 for btree insert +
  85  *  2 for btree lookup used within space map
  86  * For btree remove:
  87  *  2 for shadow spine +
  88  *  4 for rebalance 3 child node
  89  */
  90 #define THIN_MAX_CONCURRENT_LOCKS 6
  91
  92 /* This should be plenty */
  93 #define SPACE_MAP_ROOT_SIZE 128
  94
  95 /*
  96  * Little endian on-disk superblock and device details.
  97  */
  98 struct thin_disk_superblock {
  99         __le32 csum;    /* Checksum of superblock except for this field. */
 100         __le32 flags;
 101         __le64 blocknr; /* This block number, dm_block_t. */
 102
 103         __u8 uuid[16];
 104         __le64 magic;
 105         __le32 version;
 106         __le32 time;
 107
 108         __le64 trans_id;
 109
 110         /*
 111          * Root held by userspace transactions.
 112          */
 113         __le64 held_root;
 114
 115         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 116         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 117
 118         /*
 119          * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 120          */
 121         __le64 data_mapping_root;
 122
 123         /*
 124          * Device detail root mapping dev_id -> device_details
 125          */
 126         __le64 device_details_root;
 127
 128         __le32 data_block_size;         /* In 512-byte sectors. */
 129
 130         __le32 metadata_block_size;     /* In 512-byte sectors. */
 131         __le64 metadata_nr_blocks;
 132
 133         __le32 compat_flags;
 134         __le32 compat_ro_flags;
 135         __le32 incompat_flags;
 136 } __packed;
 137
 138 struct disk_device_details {
 139         __le64 mapped_blocks;
 140         __le64 transaction_id;          /* When created. */
 141         __le32 creation_time;
 142         __le32 snapshotted_time;
 143 } __packed;
 144
 145 struct dm_pool_metadata {
 146         struct hlist_node hash;
 147
 148         struct block_device *bdev;
 149         struct dm_block_manager *bm;
 150         struct dm_space_map *metadata_sm;
 151         struct dm_space_map *data_sm;
 152         struct dm_transaction_manager *tm;
 153         struct dm_transaction_manager *nb_tm;
 154
 155         /*
 156          * Two-level btree.
 157          * First level holds thin_dev_t.
 158          * Second level holds mappings.
 159          */
 160         struct dm_btree_info info;
 161
 162         /*
 163          * Non-blocking version of the above.
 164          */
 165         struct dm_btree_info nb_info;
 166
 167         /*
 168          * Just the top level for deleting whole devices.
 169          */
 170         struct dm_btree_info tl_info;
 171
 172         /*
 173          * Just the bottom level for creating new devices.
 174          */
 175         struct dm_btree_info bl_info;
 176
 177         /*
 178          * Describes the device details btree.
 179          */
 180         struct dm_btree_info details_info;
 181
 182         struct rw_semaphore root_lock;
 183         uint32_t time;
 184         dm_block_t root;
 185         dm_block_t details_root;
 186         struct list_head thin_devices;
 187         uint64_t trans_id;
 188         unsigned long flags;
 189         sector_t data_block_size;
 190
 191         /*
 192          * We reserve a section of the metadata for commit overhead.
 193          * All reported space does *not* include this.
 194          */
 195         dm_block_t metadata_reserve;
 196
 197         /*
 198          * Set if a transaction has to be aborted but the attempt to roll back
 199          * to the previous (good) transaction failed.  The only pool metadata
 200          * operation possible in this state is the closing of the device.
 201          */
 202         bool fail_io:1;
 203
 204         /*
 205          * Set once a thin-pool has been accessed through one of the interfaces
 206          * that imply the pool is in-service (e.g. thin devices created/deleted,
 207          * thin-pool message, metadata snapshots, etc).
 208          */
 209         bool in_service:1;
 210
 211         /*
 212          * Reading the space map roots can fail, so we read it into these
 213          * buffers before the superblock is locked and updated.
 214          */
 215         __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 216         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 217 };
 218
 219 struct dm_thin_device {
 220         struct list_head list;
 221         struct dm_pool_metadata *pmd;
 222         dm_thin_id id;
 223
 224         int open_count;
 225         bool changed:1;
 226         bool aborted_with_changes:1;
 227         uint64_t mapped_blocks;
 228         uint64_t transaction_id;
 229         uint32_t creation_time;
 230         uint32_t snapshotted_time;
 231 };
 232
 233 /*----------------------------------------------------------------
 234  * superblock validator
 235  *--------------------------------------------------------------*/
 236
 237 #define SUPERBLOCK_CSUM_XOR 160774
 238
 239 static void sb_prepare_for_write(struct dm_block_validator *v,
 240                                  struct dm_block *b,
 241                                  size_t block_size)
 242 {
 243         struct thin_disk_superblock *disk_super = dm_block_data(b);
 244
 245         disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 246         disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 247                                                       block_size - sizeof(__le32),
 248                                                       SUPERBLOCK_CSUM_XOR));
 249 }
 250
 251 static int sb_check(struct dm_block_validator *v,
 252                     struct dm_block *b,
 253                     size_t block_size)
 254 {
 255         struct thin_disk_superblock *disk_super = dm_block_data(b);
 256         __le32 csum_le;
 257
 258         if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 259                 DMERR("sb_check failed: blocknr %llu: "
 260                       "wanted %llu", le64_to_cpu(disk_super->blocknr),
 261                       (unsigned long long)dm_block_location(b));
 262                 return -ENOTBLK;
 263         }
 264
 265         if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 266                 DMERR("sb_check failed: magic %llu: "
 267                       "wanted %llu", le64_to_cpu(disk_super->magic),
 268                       (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 269                 return -EILSEQ;
 270         }
 271
 272         csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 273                                              block_size - sizeof(__le32),
 274                                              SUPERBLOCK_CSUM_XOR));
 275         if (csum_le != disk_super->csum) {
 276                 DMERR("sb_check failed: csum %u: wanted %u",
 277                       le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 278                 return -EILSEQ;
 279         }
 280
 281         return 0;
 282 }
 283
 284 static struct dm_block_validator sb_validator = {
 285         .name = "superblock",
 286         .prepare_for_write = sb_prepare_for_write,
 287         .check = sb_check
 288 };
 289
 290 /*----------------------------------------------------------------
 291  * Methods for the btree value types
 292  *--------------------------------------------------------------*/
 293
 294 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 295 {
 296         return (b << 24) | t;
 297 }
 298
 299 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 300 {
 301         *b = v >> 24;
 302         *t = v & ((1 << 24) - 1);
 303 }
 304
 305 static void data_block_inc(void *context, const void *value_le)
 306 {
 307         struct dm_space_map *sm = context;
 308         __le64 v_le;
 309         uint64_t b;
 310         uint32_t t;
 311
 312         memcpy(&v_le, value_le, sizeof(v_le));
 313         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 314         dm_sm_inc_block(sm, b);
 315 }
 316
 317 static void data_block_dec(void *context, const void *value_le)
 318 {
 319         struct dm_space_map *sm = context;
 320         __le64 v_le;
 321         uint64_t b;
 322         uint32_t t;
 323
 324         memcpy(&v_le, value_le, sizeof(v_le));
 325         unpack_block_time(le64_to_cpu(v_le), &b, &t);
 326         dm_sm_dec_block(sm, b);
 327 }
 328
 329 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 330 {
 331         __le64 v1_le, v2_le;
 332         uint64_t b1, b2;
 333         uint32_t t;
 334
 335         memcpy(&v1_le, value1_le, sizeof(v1_le));
 336         memcpy(&v2_le, value2_le, sizeof(v2_le));
 337         unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 338         unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 339
 340         return b1 == b2;
 341 }
 342
 343 static void subtree_inc(void *context, const void *value)
 344 {
 345         struct dm_btree_info *info = context;
 346         __le64 root_le;
 347         uint64_t root;
 348
 349         memcpy(&root_le, value, sizeof(root_le));
 350         root = le64_to_cpu(root_le);
 351         dm_tm_inc(info->tm, root);
 352 }
 353
 354 static void subtree_dec(void *context, const void *value)
 355 {
 356         struct dm_btree_info *info = context;
 357         __le64 root_le;
 358         uint64_t root;
 359
 360         memcpy(&root_le, value, sizeof(root_le));
 361         root = le64_to_cpu(root_le);
 362         if (dm_btree_del(info, root))
 363                 DMERR("btree delete failed");
 364 }
 365
 366 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 367 {
 368         __le64 v1_le, v2_le;
 369         memcpy(&v1_le, value1_le, sizeof(v1_le));
 370         memcpy(&v2_le, value2_le, sizeof(v2_le));
 371
 372         return v1_le == v2_le;
 373 }
 374
 375 /*----------------------------------------------------------------*/
 376
 377 /*
 378  * Variant that is used for in-core only changes or code that
 379  * shouldn't put the pool in service on its own (e.g. commit).
 380  */
 381 static inline void __pmd_write_lock(struct dm_pool_metadata *pmd)
 382         __acquires(pmd->root_lock)
 383 {
 384         down_write(&pmd->root_lock);
 385 }
 386 #define pmd_write_lock_in_core(pmd) __pmd_write_lock((pmd))
 387
 388 static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
 389 {
 390         __pmd_write_lock(pmd);
 391         if (unlikely(!pmd->in_service))
 392                 pmd->in_service = true;
 393 }
 394
 395 static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
 396         __releases(pmd->root_lock)
 397 {
 398         up_write(&pmd->root_lock);
 399 }
 400
 401 /*----------------------------------------------------------------*/
 402
 403 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 404                                 struct dm_block **sblock)
 405 {
 406         return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 407                                      &sb_validator, sblock);
 408 }
 409
 410 static int superblock_lock(struct dm_pool_metadata *pmd,
 411                            struct dm_block **sblock)
 412 {
 413         return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 414                                 &sb_validator, sblock);
 415 }
 416
 417 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 418 {
 419         int r;
 420         unsigned i;
 421         struct dm_block *b;
 422         __le64 *data_le, zero = cpu_to_le64(0);
 423         unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 424
 425         /*
 426          * We can't use a validator here - it may be all zeroes.
 427          */
 428         r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 429         if (r)
 430                 return r;
 431
 432         data_le = dm_block_data(b);
 433         *result = 1;
 434         for (i = 0; i < block_size; i++) {
 435                 if (data_le[i] != zero) {
 436                         *result = 0;
 437                         break;
 438                 }
 439         }
 440
 441         dm_bm_unlock(b);
 442
 443         return 0;
 444 }
 445
 446 static void __setup_btree_details(struct dm_pool_metadata *pmd)
 447 {
 448         pmd->info.tm = pmd->tm;
 449         pmd->info.levels = 2;
 450         pmd->info.value_type.context = pmd->data_sm;
 451         pmd->info.value_type.size = sizeof(__le64);
 452         pmd->info.value_type.inc = data_block_inc;
 453         pmd->info.value_type.dec = data_block_dec;
 454         pmd->info.value_type.equal = data_block_equal;
 455
 456         memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 457         pmd->nb_info.tm = pmd->nb_tm;
 458
 459         pmd->tl_info.tm = pmd->tm;
 460         pmd->tl_info.levels = 1;
 461         pmd->tl_info.value_type.context = &pmd->bl_info;
 462         pmd->tl_info.value_type.size = sizeof(__le64);
 463         pmd->tl_info.value_type.inc = subtree_inc;
 464         pmd->tl_info.value_type.dec = subtree_dec;
 465         pmd->tl_info.value_type.equal = subtree_equal;
 466
 467         pmd->bl_info.tm = pmd->tm;
 468         pmd->bl_info.levels = 1;
 469         pmd->bl_info.value_type.context = pmd->data_sm;
 470         pmd->bl_info.value_type.size = sizeof(__le64);
 471         pmd->bl_info.value_type.inc = data_block_inc;
 472         pmd->bl_info.value_type.dec = data_block_dec;
 473         pmd->bl_info.value_type.equal = data_block_equal;
 474
 475         pmd->details_info.tm = pmd->tm;
 476         pmd->details_info.levels = 1;
 477         pmd->details_info.value_type.context = NULL;
 478         pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 479         pmd->details_info.value_type.inc = NULL;
 480         pmd->details_info.value_type.dec = NULL;
 481         pmd->details_info.value_type.equal = NULL;
 482 }
 483
 484 static int save_sm_roots(struct dm_pool_metadata *pmd)
 485 {
 486         int r;
 487         size_t len;
 488
 489         r = dm_sm_root_size(pmd->metadata_sm, &len);
 490         if (r < 0)
 491                 return r;
 492
 493         r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 494         if (r < 0)
 495                 return r;
 496
 497         r = dm_sm_root_size(pmd->data_sm, &len);
 498         if (r < 0)
 499                 return r;
 500
 501         return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 502 }
 503
 504 static void copy_sm_roots(struct dm_pool_metadata *pmd,
 505                           struct thin_disk_superblock *disk)
 506 {
 507         memcpy(&disk->metadata_space_map_root,
 508                &pmd->metadata_space_map_root,
 509                sizeof(pmd->metadata_space_map_root));
 510
 511         memcpy(&disk->data_space_map_root,
 512                &pmd->data_space_map_root,
 513                sizeof(pmd->data_space_map_root));
 514 }
 515
 516 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 517 {
 518         int r;
 519         struct dm_block *sblock;
 520         struct thin_disk_superblock *disk_super;
 521         sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 522
 523         if (bdev_size > THIN_METADATA_MAX_SECTORS)
 524                 bdev_size = THIN_METADATA_MAX_SECTORS;
 525
 526         r = dm_sm_commit(pmd->data_sm);
 527         if (r < 0)
 528                 return r;
 529
 530         r = dm_tm_pre_commit(pmd->tm);
 531         if (r < 0)
 532                 return r;
 533
 534         r = save_sm_roots(pmd);
 535         if (r < 0)
 536                 return r;
 537
 538         r = superblock_lock_zero(pmd, &sblock);
 539         if (r)
 540                 return r;
 541
 542         disk_super = dm_block_data(sblock);
 543         disk_super->flags = 0;
 544         memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 545         disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 546         disk_super->version = cpu_to_le32(THIN_VERSION);
 547         disk_super->time = 0;
 548         disk_super->trans_id = 0;
 549         disk_super->held_root = 0;
 550
 551         copy_sm_roots(pmd, disk_super);
 552
 553         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 554         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 555         disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 556         disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 557         disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 558
 559         return dm_tm_commit(pmd->tm, sblock);
 560 }
 561
 562 static int __format_metadata(struct dm_pool_metadata *pmd)
 563 {
 564         int r;
 565
 566         r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 567                                  &pmd->tm, &pmd->metadata_sm);
 568         if (r < 0) {
 569                 DMERR("tm_create_with_sm failed");
 570                 return r;
 571         }
 572
 573         pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 574         if (IS_ERR(pmd->data_sm)) {
 575                 DMERR("sm_disk_create failed");
 576                 r = PTR_ERR(pmd->data_sm);
 577                 goto bad_cleanup_tm;
 578         }
 579
 580         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 581         if (!pmd->nb_tm) {
 582                 DMERR("could not create non-blocking clone tm");
 583                 r = -ENOMEM;
 584                 goto bad_cleanup_data_sm;
 585         }
 586
 587         __setup_btree_details(pmd);
 588
 589         r = dm_btree_empty(&pmd->info, &pmd->root);
 590         if (r < 0)
 591                 goto bad_cleanup_nb_tm;
 592
 593         r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 594         if (r < 0) {
 595                 DMERR("couldn't create devices root");
 596                 goto bad_cleanup_nb_tm;
 597         }
 598
 599         r = __write_initial_superblock(pmd);
 600         if (r)
 601                 goto bad_cleanup_nb_tm;
 602
 603         return 0;
 604
 605 bad_cleanup_nb_tm:
 606         dm_tm_destroy(pmd->nb_tm);
 607 bad_cleanup_data_sm:
 608         dm_sm_destroy(pmd->data_sm);
 609 bad_cleanup_tm:
 610         dm_tm_destroy(pmd->tm);
 611         dm_sm_destroy(pmd->metadata_sm);
 612
 613         return r;
 614 }
 615
 616 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 617                                      struct dm_pool_metadata *pmd)
 618 {
 619         uint32_t features;
 620
 621         features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 622         if (features) {
 623                 DMERR("could not access metadata due to unsupported optional features (%lx).",
 624                       (unsigned long)features);
 625                 return -EINVAL;
 626         }
 627
 628         /*
 629          * Check for read-only metadata to skip the following RDWR checks.
 630          */
 631         if (get_disk_ro(pmd->bdev->bd_disk))
 632                 return 0;
 633
 634         features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 635         if (features) {
 636                 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 637                       (unsigned long)features);
 638                 return -EINVAL;
 639         }
 640
 641         return 0;
 642 }
 643
 644 static int __open_metadata(struct dm_pool_metadata *pmd)
 645 {
 646         int r;
 647         struct dm_block *sblock;
 648         struct thin_disk_superblock *disk_super;
 649
 650         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 651                             &sb_validator, &sblock);
 652         if (r < 0) {
 653                 DMERR("couldn't read superblock");
 654                 return r;
 655         }
 656
 657         disk_super = dm_block_data(sblock);
 658
 659         /* Verify the data block size hasn't changed */
 660         if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 661                 DMERR("changing the data block size (from %u to %llu) is not supported",
 662                       le32_to_cpu(disk_super->data_block_size),
 663                       (unsigned long long)pmd->data_block_size);
 664                 r = -EINVAL;
 665                 goto bad_unlock_sblock;
 666         }
 667
 668         r = __check_incompat_features(disk_super, pmd);
 669         if (r < 0)
 670                 goto bad_unlock_sblock;
 671
 672         r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 673                                disk_super->metadata_space_map_root,
 674                                sizeof(disk_super->metadata_space_map_root),
 675                                &pmd->tm, &pmd->metadata_sm);
 676         if (r < 0) {
 677                 DMERR("tm_open_with_sm failed");
 678                 goto bad_unlock_sblock;
 679         }
 680
 681         pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 682                                        sizeof(disk_super->data_space_map_root));
 683         if (IS_ERR(pmd->data_sm)) {
 684                 DMERR("sm_disk_open failed");
 685                 r = PTR_ERR(pmd->data_sm);
 686                 goto bad_cleanup_tm;
 687         }
 688
 689         pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 690         if (!pmd->nb_tm) {
 691                 DMERR("could not create non-blocking clone tm");
 692                 r = -ENOMEM;
 693                 goto bad_cleanup_data_sm;
 694         }
 695
 696         __setup_btree_details(pmd);
 697         dm_bm_unlock(sblock);
 698
 699         return 0;
 700
 701 bad_cleanup_data_sm:
 702         dm_sm_destroy(pmd->data_sm);
 703 bad_cleanup_tm:
 704         dm_tm_destroy(pmd->tm);
 705         dm_sm_destroy(pmd->metadata_sm);
 706 bad_unlock_sblock:
 707         dm_bm_unlock(sblock);
 708
 709         return r;
 710 }
 711
 712 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 713 {
 714         int r, unformatted;
 715
 716         r = __superblock_all_zeroes(pmd->bm, &unformatted);
 717         if (r)
 718                 return r;
 719
 720         if (unformatted)
 721                 return format_device ? __format_metadata(pmd) : -EPERM;
 722
 723         return __open_metadata(pmd);
 724 }
 725
 726 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 727 {
 728         int r;
 729
 730         pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 731                                           THIN_MAX_CONCURRENT_LOCKS);
 732         if (IS_ERR(pmd->bm)) {
 733                 DMERR("could not create block manager");
 734                 return PTR_ERR(pmd->bm);
 735         }
 736
 737         r = __open_or_format_metadata(pmd, format_device);
 738         if (r)
 739                 dm_block_manager_destroy(pmd->bm);
 740
 741         return r;
 742 }
 743
 744 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 745 {
 746         dm_sm_destroy(pmd->data_sm);
 747         dm_sm_destroy(pmd->metadata_sm);
 748         dm_tm_destroy(pmd->nb_tm);
 749         dm_tm_destroy(pmd->tm);
 750         dm_block_manager_destroy(pmd->bm);
 751 }
 752
 753 static int __begin_transaction(struct dm_pool_metadata *pmd)
 754 {
 755         int r;
 756         struct thin_disk_superblock *disk_super;
 757         struct dm_block *sblock;
 758
 759         /*
 760          * We re-read the superblock every time.  Shouldn't need to do this
 761          * really.
 762          */
 763         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 764                             &sb_validator, &sblock);
 765         if (r)
 766                 return r;
 767
 768         disk_super = dm_block_data(sblock);
 769         pmd->time = le32_to_cpu(disk_super->time);
 770         pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 771         pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 772         pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 773         pmd->flags = le32_to_cpu(disk_super->flags);
 774         pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 775
 776         dm_bm_unlock(sblock);
 777         return 0;
 778 }
 779
 780 static int __write_changed_details(struct dm_pool_metadata *pmd)
 781 {
 782         int r;
 783         struct dm_thin_device *td, *tmp;
 784         struct disk_device_details details;
 785         uint64_t key;
 786
 787         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 788                 if (!td->changed)
 789                         continue;
 790
 791                 key = td->id;
 792
 793                 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 794                 details.transaction_id = cpu_to_le64(td->transaction_id);
 795                 details.creation_time = cpu_to_le32(td->creation_time);
 796                 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 797                 __dm_bless_for_disk(&details);
 798
 799                 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 800                                     &key, &details, &pmd->details_root);
 801                 if (r)
 802                         return r;
 803
 804                 if (td->open_count)
 805                         td->changed = 0;
 806                 else {
 807                         list_del(&td->list);
 808                         kfree(td);
 809                 }
 810         }
 811
 812         return 0;
 813 }
 814
 815 static int __commit_transaction(struct dm_pool_metadata *pmd)
 816 {
 817         int r;
 818         struct thin_disk_superblock *disk_super;
 819         struct dm_block *sblock;
 820
 821         /*
 822          * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 823          */
 824         BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 825
 826         if (unlikely(!pmd->in_service))
 827                 return 0;
 828
 829         r = __write_changed_details(pmd);
 830         if (r < 0)
 831                 return r;
 832
 833         r = dm_sm_commit(pmd->data_sm);
 834         if (r < 0)
 835                 return r;
 836
 837         r = dm_tm_pre_commit(pmd->tm);
 838         if (r < 0)
 839                 return r;
 840
 841         r = save_sm_roots(pmd);
 842         if (r < 0)
 843                 return r;
 844
 845         r = superblock_lock(pmd, &sblock);
 846         if (r)
 847                 return r;
 848
 849         disk_super = dm_block_data(sblock);
 850         disk_super->time = cpu_to_le32(pmd->time);
 851         disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 852         disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 853         disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 854         disk_super->flags = cpu_to_le32(pmd->flags);
 855
 856         copy_sm_roots(pmd, disk_super);
 857
 858         return dm_tm_commit(pmd->tm, sblock);
 859 }
 860
 861 static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 862 {
 863         int r;
 864         dm_block_t total;
 865         dm_block_t max_blocks = 4096; /* 16M */
 866
 867         r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 868         if (r) {
 869                 DMERR("could not get size of metadata device");
 870                 pmd->metadata_reserve = max_blocks;
 871         } else
 872                 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 873 }
 874
 875 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 876                                                sector_t data_block_size,
 877                                                bool format_device)
 878 {
 879         int r;
 880         struct dm_pool_metadata *pmd;
 881
 882         pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 883         if (!pmd) {
 884                 DMERR("could not allocate metadata struct");
 885                 return ERR_PTR(-ENOMEM);
 886         }
 887
 888         init_rwsem(&pmd->root_lock);
 889         pmd->time = 0;
 890         INIT_LIST_HEAD(&pmd->thin_devices);
 891         pmd->fail_io = false;
 892         pmd->in_service = false;
 893         pmd->bdev = bdev;
 894         pmd->data_block_size = data_block_size;
 895
 896         r = __create_persistent_data_objects(pmd, format_device);
 897         if (r) {
 898                 kfree(pmd);
 899                 return ERR_PTR(r);
 900         }
 901
 902         r = __begin_transaction(pmd);
 903         if (r < 0) {
 904                 if (dm_pool_metadata_close(pmd) < 0)
 905                         DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 906                 return ERR_PTR(r);
 907         }
 908
 909         __set_metadata_reserve(pmd);
 910
 911         return pmd;
 912 }
 913
 914 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 915 {
 916         int r;
 917         unsigned open_devices = 0;
 918         struct dm_thin_device *td, *tmp;
 919
 920         down_read(&pmd->root_lock);
 921         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 922                 if (td->open_count)
 923                         open_devices++;
 924                 else {
 925                         list_del(&td->list);
 926                         kfree(td);
 927                 }
 928         }
 929         up_read(&pmd->root_lock);
 930
 931         if (open_devices) {
 932                 DMERR("attempt to close pmd when %u device(s) are still open",
 933                        open_devices);
 934                 return -EBUSY;
 935         }
 936
 937         if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
 938                 r = __commit_transaction(pmd);
 939                 if (r < 0)
 940                         DMWARN("%s: __commit_transaction() failed, error = %d",
 941                                __func__, r);
 942         }
 943         if (!pmd->fail_io)
 944                 __destroy_persistent_data_objects(pmd);
 945
 946         kfree(pmd);
 947         return 0;
 948 }
 949
 950 /*
 951  * __open_device: Returns @td corresponding to device with id @dev,
 952  * creating it if @create is set and incrementing @td->open_count.
 953  * On failure, @td is undefined.
 954  */
 955 static int __open_device(struct dm_pool_metadata *pmd,
 956                          dm_thin_id dev, int create,
 957                          struct dm_thin_device **td)
 958 {
 959         int r, changed = 0;
 960         struct dm_thin_device *td2;
 961         uint64_t key = dev;
 962         struct disk_device_details details_le;
 963
 964         /*
 965          * If the device is already open, return it.
 966          */
 967         list_for_each_entry(td2, &pmd->thin_devices, list)
 968                 if (td2->id == dev) {
 969                         /*
 970                          * May not create an already-open device.
 971                          */
 972                         if (create)
 973                                 return -EEXIST;
 974
 975                         td2->open_count++;
 976                         *td = td2;
 977                         return 0;
 978                 }
 979
 980         /*
 981          * Check the device exists.
 982          */
 983         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 984                             &key, &details_le);
 985         if (r) {
 986                 if (r != -ENODATA || !create)
 987                         return r;
 988
 989                 /*
 990                  * Create new device.
 991                  */
 992                 changed = 1;
 993                 details_le.mapped_blocks = 0;
 994                 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
 995                 details_le.creation_time = cpu_to_le32(pmd->time);
 996                 details_le.snapshotted_time = cpu_to_le32(pmd->time);
 997         }
 998
 999         *td = kmalloc(sizeof(**td), GFP_NOIO);
1000         if (!*td)
1001                 return -ENOMEM;
1002
1003         (*td)->pmd = pmd;
1004         (*td)->id = dev;
1005         (*td)->open_count = 1;
1006         (*td)->changed = changed;
1007         (*td)->aborted_with_changes = false;
1008         (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1009         (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1010         (*td)->creation_time = le32_to_cpu(details_le.creation_time);
1011         (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1012
1013         list_add(&(*td)->list, &pmd->thin_devices);
1014
1015         return 0;
1016 }
1017
1018 static void __close_device(struct dm_thin_device *td)
1019 {
1020         --td->open_count;
1021 }
1022
1023 static int __create_thin(struct dm_pool_metadata *pmd,
1024                          dm_thin_id dev)
1025 {
1026         int r;
1027         dm_block_t dev_root;
1028         uint64_t key = dev;
1029         struct disk_device_details details_le;
1030         struct dm_thin_device *td;
1031         __le64 value;
1032
1033         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1034                             &key, &details_le);
1035         if (!r)
1036                 return -EEXIST;
1037
1038         /*
1039          * Create an empty btree for the mappings.
1040          */
1041         r = dm_btree_empty(&pmd->bl_info, &dev_root);
1042         if (r)
1043                 return r;
1044
1045         /*
1046          * Insert it into the main mapping tree.
1047          */
1048         value = cpu_to_le64(dev_root);
1049         __dm_bless_for_disk(&value);
1050         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1051         if (r) {
1052                 dm_btree_del(&pmd->bl_info, dev_root);
1053                 return r;
1054         }
1055
1056         r = __open_device(pmd, dev, 1, &td);
1057         if (r) {
1058                 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1059                 dm_btree_del(&pmd->bl_info, dev_root);
1060                 return r;
1061         }
1062         __close_device(td);
1063
1064         return r;
1065 }
1066
1067 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1068 {
1069         int r = -EINVAL;
1070
1071         pmd_write_lock(pmd);
1072         if (!pmd->fail_io)
1073                 r = __create_thin(pmd, dev);
1074         pmd_write_unlock(pmd);
1075
1076         return r;
1077 }
1078
1079 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1080                                   struct dm_thin_device *snap,
1081                                   dm_thin_id origin, uint32_t time)
1082 {
1083         int r;
1084         struct dm_thin_device *td;
1085
1086         r = __open_device(pmd, origin, 0, &td);
1087         if (r)
1088                 return r;
1089
1090         td->changed = 1;
1091         td->snapshotted_time = time;
1092
1093         snap->mapped_blocks = td->mapped_blocks;
1094         snap->snapshotted_time = time;
1095         __close_device(td);
1096
1097         return 0;
1098 }
1099
1100 static int __create_snap(struct dm_pool_metadata *pmd,
1101                          dm_thin_id dev, dm_thin_id origin)
1102 {
1103         int r;
1104         dm_block_t origin_root;
1105         uint64_t key = origin, dev_key = dev;
1106         struct dm_thin_device *td;
1107         struct disk_device_details details_le;
1108         __le64 value;
1109
1110         /* check this device is unused */
1111         r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1112                             &dev_key, &details_le);
1113         if (!r)
1114                 return -EEXIST;
1115
1116         /* find the mapping tree for the origin */
1117         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1118         if (r)
1119                 return r;
1120         origin_root = le64_to_cpu(value);
1121
1122         /* clone the origin, an inc will do */
1123         dm_tm_inc(pmd->tm, origin_root);
1124
1125         /* insert into the main mapping tree */
1126         value = cpu_to_le64(origin_root);
1127         __dm_bless_for_disk(&value);
1128         key = dev;
1129         r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1130         if (r) {
1131                 dm_tm_dec(pmd->tm, origin_root);
1132                 return r;
1133         }
1134
1135         pmd->time++;
1136
1137         r = __open_device(pmd, dev, 1, &td);
1138         if (r)
1139                 goto bad;
1140
1141         r = __set_snapshot_details(pmd, td, origin, pmd->time);
1142         __close_device(td);
1143
1144         if (r)
1145                 goto bad;
1146
1147         return 0;
1148
1149 bad:
1150         dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1151         dm_btree_remove(&pmd->details_info, pmd->details_root,
1152                         &key, &pmd->details_root);
1153         return r;
1154 }
1155
1156 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1157                                  dm_thin_id dev,
1158                                  dm_thin_id origin)
1159 {
1160         int r = -EINVAL;
1161
1162         pmd_write_lock(pmd);
1163         if (!pmd->fail_io)
1164                 r = __create_snap(pmd, dev, origin);
1165         pmd_write_unlock(pmd);
1166
1167         return r;
1168 }
1169
1170 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1171 {
1172         int r;
1173         uint64_t key = dev;
1174         struct dm_thin_device *td;
1175
1176         /* TODO: failure should mark the transaction invalid */
1177         r = __open_device(pmd, dev, 0, &td);
1178         if (r)
1179                 return r;
1180
1181         if (td->open_count > 1) {
1182                 __close_device(td);
1183                 return -EBUSY;
1184         }
1185
1186         list_del(&td->list);
1187         kfree(td);
1188         r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1189                             &key, &pmd->details_root);
1190         if (r)
1191                 return r;
1192
1193         r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1194         if (r)
1195                 return r;
1196
1197         return 0;
1198 }
1199
1200 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1201                                dm_thin_id dev)
1202 {
1203         int r = -EINVAL;
1204
1205         pmd_write_lock(pmd);
1206         if (!pmd->fail_io)
1207                 r = __delete_device(pmd, dev);
1208         pmd_write_unlock(pmd);
1209
1210         return r;
1211 }
1212
1213 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1214                                         uint64_t current_id,
1215                                         uint64_t new_id)
1216 {
1217         int r = -EINVAL;
1218
1219         pmd_write_lock(pmd);
1220
1221         if (pmd->fail_io)
1222                 goto out;
1223
1224         if (pmd->trans_id != current_id) {
1225                 DMERR("mismatched transaction id");
1226                 goto out;
1227         }
1228
1229         pmd->trans_id = new_id;
1230         r = 0;
1231
1232 out:
1233         pmd_write_unlock(pmd);
1234
1235         return r;
1236 }
1237
1238 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1239                                         uint64_t *result)
1240 {
1241         int r = -EINVAL;
1242
1243         down_read(&pmd->root_lock);
1244         if (!pmd->fail_io) {
1245                 *result = pmd->trans_id;
1246                 r = 0;
1247         }
1248         up_read(&pmd->root_lock);
1249
1250         return r;
1251 }
1252
1253 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1254 {
1255         int r, inc;
1256         struct thin_disk_superblock *disk_super;
1257         struct dm_block *copy, *sblock;
1258         dm_block_t held_root;
1259
1260         /*
1261          * We commit to ensure the btree roots which we increment in a
1262          * moment are up to date.
1263          */
1264         r = __commit_transaction(pmd);
1265         if (r < 0) {
1266                 DMWARN("%s: __commit_transaction() failed, error = %d",
1267                        __func__, r);
1268                 return r;
1269         }
1270
1271         /*
1272          * Copy the superblock.
1273          */
1274         dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1275         r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1276                                &sb_validator, &copy, &inc);
1277         if (r)
1278                 return r;
1279
1280         BUG_ON(!inc);
1281
1282         held_root = dm_block_location(copy);
1283         disk_super = dm_block_data(copy);
1284
1285         if (le64_to_cpu(disk_super->held_root)) {
1286                 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1287
1288                 dm_tm_dec(pmd->tm, held_root);
1289                 dm_tm_unlock(pmd->tm, copy);
1290                 return -EBUSY;
1291         }
1292
1293         /*
1294          * Wipe the spacemap since we're not publishing this.
1295          */
1296         memset(&disk_super->data_space_map_root, 0,
1297                sizeof(disk_super->data_space_map_root));
1298         memset(&disk_super->metadata_space_map_root, 0,
1299                sizeof(disk_super->metadata_space_map_root));
1300
1301         /*
1302          * Increment the data structures that need to be preserved.
1303          */
1304         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1305         dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1306         dm_tm_unlock(pmd->tm, copy);
1307
1308         /*
1309          * Write the held root into the superblock.
1310          */
1311         r = superblock_lock(pmd, &sblock);
1312         if (r) {
1313                 dm_tm_dec(pmd->tm, held_root);
1314                 return r;
1315         }
1316
1317         disk_super = dm_block_data(sblock);
1318         disk_super->held_root = cpu_to_le64(held_root);
1319         dm_bm_unlock(sblock);
1320         return 0;
1321 }
1322
1323 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1324 {
1325         int r = -EINVAL;
1326
1327         pmd_write_lock(pmd);
1328         if (!pmd->fail_io)
1329                 r = __reserve_metadata_snap(pmd);
1330         pmd_write_unlock(pmd);
1331
1332         return r;
1333 }
1334
1335 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1336 {
1337         int r;
1338         struct thin_disk_superblock *disk_super;
1339         struct dm_block *sblock, *copy;
1340         dm_block_t held_root;
1341
1342         r = superblock_lock(pmd, &sblock);
1343         if (r)
1344                 return r;
1345
1346         disk_super = dm_block_data(sblock);
1347         held_root = le64_to_cpu(disk_super->held_root);
1348         disk_super->held_root = cpu_to_le64(0);
1349
1350         dm_bm_unlock(sblock);
1351
1352         if (!held_root) {
1353                 DMWARN("No pool metadata snapshot found: nothing to release.");
1354                 return -EINVAL;
1355         }
1356
1357         r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1358         if (r)
1359                 return r;
1360
1361         disk_super = dm_block_data(copy);
1362         dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1363         dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1364         dm_sm_dec_block(pmd->metadata_sm, held_root);
1365
1366         dm_tm_unlock(pmd->tm, copy);
1367
1368         return 0;
1369 }
1370
1371 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1372 {
1373         int r = -EINVAL;
1374
1375         pmd_write_lock(pmd);
1376         if (!pmd->fail_io)
1377                 r = __release_metadata_snap(pmd);
1378         pmd_write_unlock(pmd);
1379
1380         return r;
1381 }
1382
1383 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1384                                dm_block_t *result)
1385 {
1386         int r;
1387         struct thin_disk_superblock *disk_super;
1388         struct dm_block *sblock;
1389
1390         r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1391                             &sb_validator, &sblock);
1392         if (r)
1393                 return r;
1394
1395         disk_super = dm_block_data(sblock);
1396         *result = le64_to_cpu(disk_super->held_root);
1397
1398         dm_bm_unlock(sblock);
1399
1400         return 0;
1401 }
1402
1403 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1404                               dm_block_t *result)
1405 {
1406         int r = -EINVAL;
1407
1408         down_read(&pmd->root_lock);
1409         if (!pmd->fail_io)
1410                 r = __get_metadata_snap(pmd, result);
1411         up_read(&pmd->root_lock);
1412
1413         return r;
1414 }
1415
1416 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1417                              struct dm_thin_device **td)
1418 {
1419         int r = -EINVAL;
1420
1421         pmd_write_lock_in_core(pmd);
1422         if (!pmd->fail_io)
1423                 r = __open_device(pmd, dev, 0, td);
1424         pmd_write_unlock(pmd);
1425
1426         return r;
1427 }
1428
1429 int dm_pool_close_thin_device(struct dm_thin_device *td)
1430 {
1431         pmd_write_lock_in_core(td->pmd);
1432         __close_device(td);
1433         pmd_write_unlock(td->pmd);
1434
1435         return 0;
1436 }
1437
1438 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1439 {
1440         return td->id;
1441 }
1442
1443 /*
1444  * Check whether @time (of block creation) is older than @td's last snapshot.
1445  * If so then the associated block is shared with the last snapshot device.
1446  * Any block on a device created *after* the device last got snapshotted is
1447  * necessarily not shared.
1448  */
1449 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1450 {
1451         return td->snapshotted_time > time;
1452 }
1453
1454 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1455                                  struct dm_thin_lookup_result *result)
1456 {
1457         uint64_t block_time = 0;
1458         dm_block_t exception_block;
1459         uint32_t exception_time;
1460
1461         block_time = le64_to_cpu(value);
1462         unpack_block_time(block_time, &exception_block, &exception_time);
1463         result->block = exception_block;
1464         result->shared = __snapshotted_since(td, exception_time);
1465 }
1466
1467 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1468                         int can_issue_io, struct dm_thin_lookup_result *result)
1469 {
1470         int r;
1471         __le64 value;
1472         struct dm_pool_metadata *pmd = td->pmd;
1473         dm_block_t keys[2] = { td->id, block };
1474         struct dm_btree_info *info;
1475
1476         if (can_issue_io) {
1477                 info = &pmd->info;
1478         } else
1479                 info = &pmd->nb_info;
1480
1481         r = dm_btree_lookup(info, pmd->root, keys, &value);
1482         if (!r)
1483                 unpack_lookup_result(td, value, result);
1484
1485         return r;
1486 }
1487
1488 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1489                        int can_issue_io, struct dm_thin_lookup_result *result)
1490 {
1491         int r;
1492         struct dm_pool_metadata *pmd = td->pmd;
1493
1494         down_read(&pmd->root_lock);
1495         if (pmd->fail_io) {
1496                 up_read(&pmd->root_lock);
1497                 return -EINVAL;
1498         }
1499
1500         r = __find_block(td, block, can_issue_io, result);
1501
1502         up_read(&pmd->root_lock);
1503         return r;
1504 }
1505
1506 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1507                                           dm_block_t *vblock,
1508                                           struct dm_thin_lookup_result *result)
1509 {
1510         int r;
1511         __le64 value;
1512         struct dm_pool_metadata *pmd = td->pmd;
1513         dm_block_t keys[2] = { td->id, block };
1514
1515         r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1516         if (!r)
1517                 unpack_lookup_result(td, value, result);
1518
1519         return r;
1520 }
1521
1522 static int __find_mapped_range(struct dm_thin_device *td,
1523                                dm_block_t begin, dm_block_t end,
1524                                dm_block_t *thin_begin, dm_block_t *thin_end,
1525                                dm_block_t *pool_begin, bool *maybe_shared)
1526 {
1527         int r;
1528         dm_block_t pool_end;
1529         struct dm_thin_lookup_result lookup;
1530
1531         if (end < begin)
1532                 return -ENODATA;
1533
1534         r = __find_next_mapped_block(td, begin, &begin, &lookup);
1535         if (r)
1536                 return r;
1537
1538         if (begin >= end)
1539                 return -ENODATA;
1540
1541         *thin_begin = begin;
1542         *pool_begin = lookup.block;
1543         *maybe_shared = lookup.shared;
1544
1545         begin++;
1546         pool_end = *pool_begin + 1;
1547         while (begin != end) {
1548                 r = __find_block(td, begin, true, &lookup);
1549                 if (r) {
1550                         if (r == -ENODATA)
1551                                 break;
1552                         else
1553                                 return r;
1554                 }
1555
1556                 if ((lookup.block != pool_end) ||
1557                     (lookup.shared != *maybe_shared))
1558                         break;
1559
1560                 pool_end++;
1561                 begin++;
1562         }
1563
1564         *thin_end = begin;
1565         return 0;
1566 }
1567
1568 int dm_thin_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 = -EINVAL;
1574         struct dm_pool_metadata *pmd = td->pmd;
1575
1576         down_read(&pmd->root_lock);
1577         if (!pmd->fail_io) {
1578                 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1579                                         pool_begin, maybe_shared);
1580         }
1581         up_read(&pmd->root_lock);
1582
1583         return r;
1584 }
1585
1586 static int __insert(struct dm_thin_device *td, dm_block_t block,
1587                     dm_block_t data_block)
1588 {
1589         int r, inserted;
1590         __le64 value;
1591         struct dm_pool_metadata *pmd = td->pmd;
1592         dm_block_t keys[2] = { td->id, block };
1593
1594         value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1595         __dm_bless_for_disk(&value);
1596
1597         r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1598                                    &pmd->root, &inserted);
1599         if (r)
1600                 return r;
1601
1602         td->changed = 1;
1603         if (inserted)
1604                 td->mapped_blocks++;
1605
1606         return 0;
1607 }
1608
1609 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1610                          dm_block_t data_block)
1611 {
1612         int r = -EINVAL;
1613
1614         pmd_write_lock(td->pmd);
1615         if (!td->pmd->fail_io)
1616                 r = __insert(td, block, data_block);
1617         pmd_write_unlock(td->pmd);
1618
1619         return r;
1620 }
1621
1622 static int __remove(struct dm_thin_device *td, dm_block_t block)
1623 {
1624         int r;
1625         struct dm_pool_metadata *pmd = td->pmd;
1626         dm_block_t keys[2] = { td->id, block };
1627
1628         r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1629         if (r)
1630                 return r;
1631
1632         td->mapped_blocks--;
1633         td->changed = 1;
1634
1635         return 0;
1636 }
1637
1638 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1639 {
1640         int r;
1641         unsigned count, total_count = 0;
1642         struct dm_pool_metadata *pmd = td->pmd;
1643         dm_block_t keys[1] = { td->id };
1644         __le64 value;
1645         dm_block_t mapping_root;
1646
1647         /*
1648          * Find the mapping tree
1649          */
1650         r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1651         if (r)
1652                 return r;
1653
1654         /*
1655          * Remove from the mapping tree, taking care to inc the
1656          * ref count so it doesn't get deleted.
1657          */
1658         mapping_root = le64_to_cpu(value);
1659         dm_tm_inc(pmd->tm, mapping_root);
1660         r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1661         if (r)
1662                 return r;
1663
1664         /*
1665          * Remove leaves stops at the first unmapped entry, so we have to
1666          * loop round finding mapped ranges.
1667          */
1668         while (begin < end) {
1669                 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1670                 if (r == -ENODATA)
1671                         break;
1672
1673                 if (r)
1674                         return r;
1675
1676                 if (begin >= end)
1677                         break;
1678
1679                 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1680                 if (r)
1681                         return r;
1682
1683                 total_count += count;
1684         }
1685
1686         td->mapped_blocks -= total_count;
1687         td->changed = 1;
1688
1689         /*
1690          * Reinsert the mapping tree.
1691          */
1692         value = cpu_to_le64(mapping_root);
1693         __dm_bless_for_disk(&value);
1694         return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1695 }
1696
1697 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1698 {
1699         int r = -EINVAL;
1700
1701         pmd_write_lock(td->pmd);
1702         if (!td->pmd->fail_io)
1703                 r = __remove(td, block);
1704         pmd_write_unlock(td->pmd);
1705
1706         return r;
1707 }
1708
1709 int dm_thin_remove_range(struct dm_thin_device *td,
1710                          dm_block_t begin, dm_block_t end)
1711 {
1712         int r = -EINVAL;
1713
1714         pmd_write_lock(td->pmd);
1715         if (!td->pmd->fail_io)
1716                 r = __remove_range(td, begin, end);
1717         pmd_write_unlock(td->pmd);
1718
1719         return r;
1720 }
1721
1722 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1723 {
1724         int r;
1725         uint32_t ref_count;
1726
1727         down_read(&pmd->root_lock);
1728         r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1729         if (!r)
1730                 *result = (ref_count > 1);
1731         up_read(&pmd->root_lock);
1732
1733         return r;
1734 }
1735
1736 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1737 {
1738         int r = 0;
1739
1740         pmd_write_lock(pmd);
1741         for (; b != e; b++) {
1742                 r = dm_sm_inc_block(pmd->data_sm, b);
1743                 if (r)
1744                         break;
1745         }
1746         pmd_write_unlock(pmd);
1747
1748         return r;
1749 }
1750
1751 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1752 {
1753         int r = 0;
1754
1755         pmd_write_lock(pmd);
1756         for (; b != e; b++) {
1757                 r = dm_sm_dec_block(pmd->data_sm, b);
1758                 if (r)
1759                         break;
1760         }
1761         pmd_write_unlock(pmd);
1762
1763         return r;
1764 }
1765
1766 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1767 {
1768         int r;
1769
1770         down_read(&td->pmd->root_lock);
1771         r = td->changed;
1772         up_read(&td->pmd->root_lock);
1773
1774         return r;
1775 }
1776
1777 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1778 {
1779         bool r = false;
1780         struct dm_thin_device *td, *tmp;
1781
1782         down_read(&pmd->root_lock);
1783         list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1784                 if (td->changed) {
1785                         r = td->changed;
1786                         break;
1787                 }
1788         }
1789         up_read(&pmd->root_lock);
1790
1791         return r;
1792 }
1793
1794 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1795 {
1796         bool r;
1797
1798         down_read(&td->pmd->root_lock);
1799         r = td->aborted_with_changes;
1800         up_read(&td->pmd->root_lock);
1801
1802         return r;
1803 }
1804
1805 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1806 {
1807         int r = -EINVAL;
1808
1809         pmd_write_lock(pmd);
1810         if (!pmd->fail_io)
1811                 r = dm_sm_new_block(pmd->data_sm, result);
1812         pmd_write_unlock(pmd);
1813
1814         return r;
1815 }
1816
1817 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1818 {
1819         int r = -EINVAL;
1820
1821         /*
1822          * Care is taken to not have commit be what
1823          * triggers putting the thin-pool in-service.
1824          */
1825         __pmd_write_lock(pmd);
1826         if (pmd->fail_io)
1827                 goto out;
1828
1829         r = __commit_transaction(pmd);
1830         if (r < 0)
1831                 goto out;
1832
1833         /*
1834          * Open the next transaction.
1835          */
1836         r = __begin_transaction(pmd);
1837 out:
1838         pmd_write_unlock(pmd);
1839         return r;
1840 }
1841
1842 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1843 {
1844         struct dm_thin_device *td;
1845
1846         list_for_each_entry(td, &pmd->thin_devices, list)
1847                 td->aborted_with_changes = td->changed;
1848 }
1849
1850 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1851 {
1852         int r = -EINVAL;
1853
1854         pmd_write_lock(pmd);
1855         if (pmd->fail_io)
1856                 goto out;
1857
1858         __set_abort_with_changes_flags(pmd);
1859         __destroy_persistent_data_objects(pmd);
1860         r = __create_persistent_data_objects(pmd, false);
1861         if (r)
1862                 pmd->fail_io = true;
1863
1864 out:
1865         pmd_write_unlock(pmd);
1866
1867         return r;
1868 }
1869
1870 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1871 {
1872         int r = -EINVAL;
1873
1874         down_read(&pmd->root_lock);
1875         if (!pmd->fail_io)
1876                 r = dm_sm_get_nr_free(pmd->data_sm, result);
1877         up_read(&pmd->root_lock);
1878
1879         return r;
1880 }
1881
1882 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1883                                           dm_block_t *result)
1884 {
1885         int r = -EINVAL;
1886
1887         down_read(&pmd->root_lock);
1888         if (!pmd->fail_io)
1889                 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1890
1891         if (!r) {
1892                 if (*result < pmd->metadata_reserve)
1893                         *result = 0;
1894                 else
1895                         *result -= pmd->metadata_reserve;
1896         }
1897         up_read(&pmd->root_lock);
1898
1899         return r;
1900 }
1901
1902 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1903                                   dm_block_t *result)
1904 {
1905         int r = -EINVAL;
1906
1907         down_read(&pmd->root_lock);
1908         if (!pmd->fail_io)
1909                 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1910         up_read(&pmd->root_lock);
1911
1912         return r;
1913 }
1914
1915 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1916 {
1917         int r = -EINVAL;
1918
1919         down_read(&pmd->root_lock);
1920         if (!pmd->fail_io)
1921                 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1922         up_read(&pmd->root_lock);
1923
1924         return r;
1925 }
1926
1927 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1928 {
1929         int r = -EINVAL;
1930         struct dm_pool_metadata *pmd = td->pmd;
1931
1932         down_read(&pmd->root_lock);
1933         if (!pmd->fail_io) {
1934                 *result = td->mapped_blocks;
1935                 r = 0;
1936         }
1937         up_read(&pmd->root_lock);
1938
1939         return r;
1940 }
1941
1942 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1943 {
1944         int r;
1945         __le64 value_le;
1946         dm_block_t thin_root;
1947         struct dm_pool_metadata *pmd = td->pmd;
1948
1949         r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1950         if (r)
1951                 return r;
1952
1953         thin_root = le64_to_cpu(value_le);
1954
1955         return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1956 }
1957
1958 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1959                                      dm_block_t *result)
1960 {
1961         int r = -EINVAL;
1962         struct dm_pool_metadata *pmd = td->pmd;
1963
1964         down_read(&pmd->root_lock);
1965         if (!pmd->fail_io)
1966                 r = __highest_block(td, result);
1967         up_read(&pmd->root_lock);
1968
1969         return r;
1970 }
1971
1972 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1973 {
1974         int r;
1975         dm_block_t old_count;
1976
1977         r = dm_sm_get_nr_blocks(sm, &old_count);
1978         if (r)
1979                 return r;
1980
1981         if (new_count == old_count)
1982                 return 0;
1983
1984         if (new_count < old_count) {
1985                 DMERR("cannot reduce size of space map");
1986                 return -EINVAL;
1987         }
1988
1989         return dm_sm_extend(sm, new_count - old_count);
1990 }
1991
1992 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1993 {
1994         int r = -EINVAL;
1995
1996         pmd_write_lock(pmd);
1997         if (!pmd->fail_io)
1998                 r = __resize_space_map(pmd->data_sm, new_count);
1999         pmd_write_unlock(pmd);
2000
2001         return r;
2002 }
2003
2004 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2005 {
2006         int r = -EINVAL;
2007
2008         pmd_write_lock(pmd);
2009         if (!pmd->fail_io) {
2010                 r = __resize_space_map(pmd->metadata_sm, new_count);
2011                 if (!r)
2012                         __set_metadata_reserve(pmd);
2013         }
2014         pmd_write_unlock(pmd);
2015
2016         return r;
2017 }
2018
2019 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2020 {
2021         pmd_write_lock_in_core(pmd);
2022         dm_bm_set_read_only(pmd->bm);
2023         pmd_write_unlock(pmd);
2024 }
2025
2026 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2027 {
2028         pmd_write_lock_in_core(pmd);
2029         dm_bm_set_read_write(pmd->bm);
2030         pmd_write_unlock(pmd);
2031 }
2032
2033 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2034                                         dm_block_t threshold,
2035                                         dm_sm_threshold_fn fn,
2036                                         void *context)
2037 {
2038         int r;
2039
2040         pmd_write_lock_in_core(pmd);
2041         r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
2042         pmd_write_unlock(pmd);
2043
2044         return r;
2045 }
2046
2047 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2048 {
2049         int r;
2050         struct dm_block *sblock;
2051         struct thin_disk_superblock *disk_super;
2052
2053         pmd_write_lock(pmd);
2054         pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2055
2056         r = superblock_lock(pmd, &sblock);
2057         if (r) {
2058                 DMERR("couldn't read superblock");
2059                 goto out;
2060         }
2061
2062         disk_super = dm_block_data(sblock);
2063         disk_super->flags = cpu_to_le32(pmd->flags);
2064
2065         dm_bm_unlock(sblock);
2066 out:
2067         pmd_write_unlock(pmd);
2068         return r;
2069 }
2070
2071 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2072 {
2073         bool needs_check;
2074
2075         down_read(&pmd->root_lock);
2076         needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2077         up_read(&pmd->root_lock);
2078
2079         return needs_check;
2080 }
2081
2082 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2083 {
2084         down_read(&pmd->root_lock);
2085         if (!pmd->fail_io)
2086                 dm_tm_issue_prefetches(pmd->tm);
2087         up_read(&pmd->root_lock);
2088 }