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