drivers/md/dm-thin-metadata.c

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