drivers/md/dm-thin-metadata.c

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