drivers/md/dm-era-target.c

   1 #include "dm.h"
   2 #include "persistent-data/dm-transaction-manager.h"
   3 #include "persistent-data/dm-bitset.h"
   4 #include "persistent-data/dm-space-map.h"
   5
   6 #include <linux/dm-io.h>
   7 #include <linux/dm-kcopyd.h>
   8 #include <linux/init.h>
   9 #include <linux/mempool.h>
  10 #include <linux/module.h>
  11 #include <linux/slab.h>
  12 #include <linux/vmalloc.h>
  13
  14 #define DM_MSG_PREFIX "era"
  15
  16 #define SUPERBLOCK_LOCATION 0
  17 #define SUPERBLOCK_MAGIC 2126579579
  18 #define SUPERBLOCK_CSUM_XOR 146538381
  19 #define MIN_ERA_VERSION 1
  20 #define MAX_ERA_VERSION 1
  21 #define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION
  22 #define MIN_BLOCK_SIZE 8
  23
  24 /*----------------------------------------------------------------
  25  * Writeset
  26  *--------------------------------------------------------------*/
  27 struct writeset_metadata {
  28         uint32_t nr_bits;
  29         dm_block_t root;
  30 };
  31
  32 struct writeset {
  33         struct writeset_metadata md;
  34
  35         /*
  36          * An in core copy of the bits to save constantly doing look ups on
  37          * disk.
  38          */
  39         unsigned long *bits;
  40 };
  41
  42 /*
  43  * This does not free off the on disk bitset as this will normally be done
  44  * after digesting into the era array.
  45  */
  46 static void writeset_free(struct writeset *ws)
  47 {
  48         vfree(ws->bits);
  49         ws->bits = NULL;
  50 }
  51
  52 static int setup_on_disk_bitset(struct dm_disk_bitset *info,
  53                                 unsigned nr_bits, dm_block_t *root)
  54 {
  55         int r;
  56
  57         r = dm_bitset_empty(info, root);
  58         if (r)
  59                 return r;
  60
  61         return dm_bitset_resize(info, *root, 0, nr_bits, false, root);
  62 }
  63
  64 static size_t bitset_size(unsigned nr_bits)
  65 {
  66         return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG);
  67 }
  68
  69 /*
  70  * Allocates memory for the in core bitset.
  71  */
  72 static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks)
  73 {
  74         ws->bits = vzalloc(bitset_size(nr_blocks));
  75         if (!ws->bits) {
  76                 DMERR("%s: couldn't allocate in memory bitset", __func__);
  77                 return -ENOMEM;
  78         }
  79
  80         return 0;
  81 }
  82
  83 /*
  84  * Wipes the in-core bitset, and creates a new on disk bitset.
  85  */
  86 static int writeset_init(struct dm_disk_bitset *info, struct writeset *ws,
  87                          dm_block_t nr_blocks)
  88 {
  89         int r;
  90
  91         memset(ws->bits, 0, bitset_size(nr_blocks));
  92
  93         ws->md.nr_bits = nr_blocks;
  94         r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root);
  95         if (r) {
  96                 DMERR("%s: setup_on_disk_bitset failed", __func__);
  97                 return r;
  98         }
  99
 100         return 0;
 101 }
 102
 103 static bool writeset_marked(struct writeset *ws, dm_block_t block)
 104 {
 105         return test_bit(block, ws->bits);
 106 }
 107
 108 static int writeset_marked_on_disk(struct dm_disk_bitset *info,
 109                                    struct writeset_metadata *m, dm_block_t block,
 110                                    bool *result)
 111 {
 112         dm_block_t old = m->root;
 113
 114         /*
 115          * The bitset was flushed when it was archived, so we know there'll
 116          * be no change to the root.
 117          */
 118         int r = dm_bitset_test_bit(info, m->root, block, &m->root, result);
 119         if (r) {
 120                 DMERR("%s: dm_bitset_test_bit failed", __func__);
 121                 return r;
 122         }
 123
 124         BUG_ON(m->root != old);
 125
 126         return r;
 127 }
 128
 129 /*
 130  * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was.
 131  */
 132 static int writeset_test_and_set(struct dm_disk_bitset *info,
 133                                  struct writeset *ws, uint32_t block)
 134 {
 135         int r;
 136
 137         if (!test_bit(block, ws->bits)) {
 138                 r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root);
 139                 if (r) {
 140                         /* FIXME: fail mode */
 141                         return r;
 142                 }
 143
 144                 return 0;
 145         }
 146
 147         return 1;
 148 }
 149
 150 /*----------------------------------------------------------------
 151  * On disk metadata layout
 152  *--------------------------------------------------------------*/
 153 #define SPACE_MAP_ROOT_SIZE 128
 154 #define UUID_LEN 16
 155
 156 struct writeset_disk {
 157         __le32 nr_bits;
 158         __le64 root;
 159 } __packed;
 160
 161 struct superblock_disk {
 162         __le32 csum;
 163         __le32 flags;
 164         __le64 blocknr;
 165
 166         __u8 uuid[UUID_LEN];
 167         __le64 magic;
 168         __le32 version;
 169
 170         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 171
 172         __le32 data_block_size;
 173         __le32 metadata_block_size;
 174         __le32 nr_blocks;
 175
 176         __le32 current_era;
 177         struct writeset_disk current_writeset;
 178
 179         /*
 180          * Only these two fields are valid within the metadata snapshot.
 181          */
 182         __le64 writeset_tree_root;
 183         __le64 era_array_root;
 184
 185         __le64 metadata_snap;
 186 } __packed;
 187
 188 /*----------------------------------------------------------------
 189  * Superblock validation
 190  *--------------------------------------------------------------*/
 191 static void sb_prepare_for_write(struct dm_block_validator *v,
 192                                  struct dm_block *b,
 193                                  size_t sb_block_size)
 194 {
 195         struct superblock_disk *disk = dm_block_data(b);
 196
 197         disk->blocknr = cpu_to_le64(dm_block_location(b));
 198         disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags,
 199                                                 sb_block_size - sizeof(__le32),
 200                                                 SUPERBLOCK_CSUM_XOR));
 201 }
 202
 203 static int check_metadata_version(struct superblock_disk *disk)
 204 {
 205         uint32_t metadata_version = le32_to_cpu(disk->version);
 206         if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) {
 207                 DMERR("Era metadata version %u found, but only versions between %u and %u supported.",
 208                       metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION);
 209                 return -EINVAL;
 210         }
 211
 212         return 0;
 213 }
 214
 215 static int sb_check(struct dm_block_validator *v,
 216                     struct dm_block *b,
 217                     size_t sb_block_size)
 218 {
 219         struct superblock_disk *disk = dm_block_data(b);
 220         __le32 csum_le;
 221
 222         if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) {
 223                 DMERR("sb_check failed: blocknr %llu: wanted %llu",
 224                       le64_to_cpu(disk->blocknr),
 225                       (unsigned long long)dm_block_location(b));
 226                 return -ENOTBLK;
 227         }
 228
 229         if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) {
 230                 DMERR("sb_check failed: magic %llu: wanted %llu",
 231                       le64_to_cpu(disk->magic),
 232                       (unsigned long long) SUPERBLOCK_MAGIC);
 233                 return -EILSEQ;
 234         }
 235
 236         csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags,
 237                                              sb_block_size - sizeof(__le32),
 238                                              SUPERBLOCK_CSUM_XOR));
 239         if (csum_le != disk->csum) {
 240                 DMERR("sb_check failed: csum %u: wanted %u",
 241                       le32_to_cpu(csum_le), le32_to_cpu(disk->csum));
 242                 return -EILSEQ;
 243         }
 244
 245         return check_metadata_version(disk);
 246 }
 247
 248 static struct dm_block_validator sb_validator = {
 249         .name = "superblock",
 250         .prepare_for_write = sb_prepare_for_write,
 251         .check = sb_check
 252 };
 253
 254 /*----------------------------------------------------------------
 255  * Low level metadata handling
 256  *--------------------------------------------------------------*/
 257 #define DM_ERA_METADATA_BLOCK_SIZE 4096
 258 #define ERA_MAX_CONCURRENT_LOCKS 5
 259
 260 struct era_metadata {
 261         struct block_device *bdev;
 262         struct dm_block_manager *bm;
 263         struct dm_space_map *sm;
 264         struct dm_transaction_manager *tm;
 265
 266         dm_block_t block_size;
 267         uint32_t nr_blocks;
 268
 269         uint32_t current_era;
 270
 271         /*
 272          * We preallocate 2 writesets.  When an era rolls over we
 273          * switch between them. This means the allocation is done at
 274          * preresume time, rather than on the io path.
 275          */
 276         struct writeset writesets[2];
 277         struct writeset *current_writeset;
 278
 279         dm_block_t writeset_tree_root;
 280         dm_block_t era_array_root;
 281
 282         struct dm_disk_bitset bitset_info;
 283         struct dm_btree_info writeset_tree_info;
 284         struct dm_array_info era_array_info;
 285
 286         dm_block_t metadata_snap;
 287
 288         /*
 289          * A flag that is set whenever a writeset has been archived.
 290          */
 291         bool archived_writesets;
 292
 293         /*
 294          * Reading the space map root can fail, so we read it into this
 295          * buffer before the superblock is locked and updated.
 296          */
 297         __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 298 };
 299
 300 static int superblock_read_lock(struct era_metadata *md,
 301                                 struct dm_block **sblock)
 302 {
 303         return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION,
 304                                &sb_validator, sblock);
 305 }
 306
 307 static int superblock_lock_zero(struct era_metadata *md,
 308                                 struct dm_block **sblock)
 309 {
 310         return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION,
 311                                      &sb_validator, sblock);
 312 }
 313
 314 static int superblock_lock(struct era_metadata *md,
 315                            struct dm_block **sblock)
 316 {
 317         return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION,
 318                                 &sb_validator, sblock);
 319 }
 320
 321 /* FIXME: duplication with cache and thin */
 322 static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
 323 {
 324         int r;
 325         unsigned i;
 326         struct dm_block *b;
 327         __le64 *data_le, zero = cpu_to_le64(0);
 328         unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
 329
 330         /*
 331          * We can't use a validator here - it may be all zeroes.
 332          */
 333         r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b);
 334         if (r)
 335                 return r;
 336
 337         data_le = dm_block_data(b);
 338         *result = true;
 339         for (i = 0; i < sb_block_size; i++) {
 340                 if (data_le[i] != zero) {
 341                         *result = false;
 342                         break;
 343                 }
 344         }
 345
 346         dm_bm_unlock(b);
 347
 348         return 0;
 349 }
 350
 351 /*----------------------------------------------------------------*/
 352
 353 static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk)
 354 {
 355         disk->nr_bits = cpu_to_le32(core->nr_bits);
 356         disk->root = cpu_to_le64(core->root);
 357 }
 358
 359 static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core)
 360 {
 361         core->nr_bits = le32_to_cpu(disk->nr_bits);
 362         core->root = le64_to_cpu(disk->root);
 363 }
 364
 365 static void ws_inc(void *context, const void *value)
 366 {
 367         struct era_metadata *md = context;
 368         struct writeset_disk ws_d;
 369         dm_block_t b;
 370
 371         memcpy(&ws_d, value, sizeof(ws_d));
 372         b = le64_to_cpu(ws_d.root);
 373
 374         dm_tm_inc(md->tm, b);
 375 }
 376
 377 static void ws_dec(void *context, const void *value)
 378 {
 379         struct era_metadata *md = context;
 380         struct writeset_disk ws_d;
 381         dm_block_t b;
 382
 383         memcpy(&ws_d, value, sizeof(ws_d));
 384         b = le64_to_cpu(ws_d.root);
 385
 386         dm_bitset_del(&md->bitset_info, b);
 387 }
 388
 389 static int ws_eq(void *context, const void *value1, const void *value2)
 390 {
 391         return !memcmp(value1, value2, sizeof(struct writeset_disk));
 392 }
 393
 394 /*----------------------------------------------------------------*/
 395
 396 static void setup_writeset_tree_info(struct era_metadata *md)
 397 {
 398         struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type;
 399         md->writeset_tree_info.tm = md->tm;
 400         md->writeset_tree_info.levels = 1;
 401         vt->context = md;
 402         vt->size = sizeof(struct writeset_disk);
 403         vt->inc = ws_inc;
 404         vt->dec = ws_dec;
 405         vt->equal = ws_eq;
 406 }
 407
 408 static void setup_era_array_info(struct era_metadata *md)
 409
 410 {
 411         struct dm_btree_value_type vt;
 412         vt.context = NULL;
 413         vt.size = sizeof(__le32);
 414         vt.inc = NULL;
 415         vt.dec = NULL;
 416         vt.equal = NULL;
 417
 418         dm_array_info_init(&md->era_array_info, md->tm, &vt);
 419 }
 420
 421 static void setup_infos(struct era_metadata *md)
 422 {
 423         dm_disk_bitset_init(md->tm, &md->bitset_info);
 424         setup_writeset_tree_info(md);
 425         setup_era_array_info(md);
 426 }
 427
 428 /*----------------------------------------------------------------*/
 429
 430 static int create_fresh_metadata(struct era_metadata *md)
 431 {
 432         int r;
 433
 434         r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION,
 435                                  &md->tm, &md->sm);
 436         if (r < 0) {
 437                 DMERR("dm_tm_create_with_sm failed");
 438                 return r;
 439         }
 440
 441         setup_infos(md);
 442
 443         r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root);
 444         if (r) {
 445                 DMERR("couldn't create new writeset tree");
 446                 goto bad;
 447         }
 448
 449         r = dm_array_empty(&md->era_array_info, &md->era_array_root);
 450         if (r) {
 451                 DMERR("couldn't create era array");
 452                 goto bad;
 453         }
 454
 455         return 0;
 456
 457 bad:
 458         dm_sm_destroy(md->sm);
 459         dm_tm_destroy(md->tm);
 460
 461         return r;
 462 }
 463
 464 static int save_sm_root(struct era_metadata *md)
 465 {
 466         int r;
 467         size_t metadata_len;
 468
 469         r = dm_sm_root_size(md->sm, &metadata_len);
 470         if (r < 0)
 471                 return r;
 472
 473         return dm_sm_copy_root(md->sm, &md->metadata_space_map_root,
 474                                metadata_len);
 475 }
 476
 477 static void copy_sm_root(struct era_metadata *md, struct superblock_disk *disk)
 478 {
 479         memcpy(&disk->metadata_space_map_root,
 480                &md->metadata_space_map_root,
 481                sizeof(md->metadata_space_map_root));
 482 }
 483
 484 /*
 485  * Writes a superblock, including the static fields that don't get updated
 486  * with every commit (possible optimisation here).  'md' should be fully
 487  * constructed when this is called.
 488  */
 489 static void prepare_superblock(struct era_metadata *md, struct superblock_disk *disk)
 490 {
 491         disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
 492         disk->flags = cpu_to_le32(0ul);
 493
 494         /* FIXME: can't keep blanking the uuid (uuid is currently unused though) */
 495         memset(disk->uuid, 0, sizeof(disk->uuid));
 496         disk->version = cpu_to_le32(MAX_ERA_VERSION);
 497
 498         copy_sm_root(md, disk);
 499
 500         disk->data_block_size = cpu_to_le32(md->block_size);
 501         disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
 502         disk->nr_blocks = cpu_to_le32(md->nr_blocks);
 503         disk->current_era = cpu_to_le32(md->current_era);
 504
 505         ws_pack(&md->current_writeset->md, &disk->current_writeset);
 506         disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root);
 507         disk->era_array_root = cpu_to_le64(md->era_array_root);
 508         disk->metadata_snap = cpu_to_le64(md->metadata_snap);
 509 }
 510
 511 static int write_superblock(struct era_metadata *md)
 512 {
 513         int r;
 514         struct dm_block *sblock;
 515         struct superblock_disk *disk;
 516
 517         r = save_sm_root(md);
 518         if (r) {
 519                 DMERR("%s: save_sm_root failed", __func__);
 520                 return r;
 521         }
 522
 523         r = superblock_lock_zero(md, &sblock);
 524         if (r)
 525                 return r;
 526
 527         disk = dm_block_data(sblock);
 528         prepare_superblock(md, disk);
 529
 530         return dm_tm_commit(md->tm, sblock);
 531 }
 532
 533 /*
 534  * Assumes block_size and the infos are set.
 535  */
 536 static int format_metadata(struct era_metadata *md)
 537 {
 538         int r;
 539
 540         r = create_fresh_metadata(md);
 541         if (r)
 542                 return r;
 543
 544         r = write_superblock(md);
 545         if (r) {
 546                 dm_sm_destroy(md->sm);
 547                 dm_tm_destroy(md->tm);
 548                 return r;
 549         }
 550
 551         return 0;
 552 }
 553
 554 static int open_metadata(struct era_metadata *md)
 555 {
 556         int r;
 557         struct dm_block *sblock;
 558         struct superblock_disk *disk;
 559
 560         r = superblock_read_lock(md, &sblock);
 561         if (r) {
 562                 DMERR("couldn't read_lock superblock");
 563                 return r;
 564         }
 565
 566         disk = dm_block_data(sblock);
 567
 568         /* Verify the data block size hasn't changed */
 569         if (le32_to_cpu(disk->data_block_size) != md->block_size) {
 570                 DMERR("changing the data block size (from %u to %llu) is not supported",
 571                       le32_to_cpu(disk->data_block_size), md->block_size);
 572                 r = -EINVAL;
 573                 goto bad;
 574         }
 575
 576         r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION,
 577                                disk->metadata_space_map_root,
 578                                sizeof(disk->metadata_space_map_root),
 579                                &md->tm, &md->sm);
 580         if (r) {
 581                 DMERR("dm_tm_open_with_sm failed");
 582                 goto bad;
 583         }
 584
 585         setup_infos(md);
 586
 587         md->nr_blocks = le32_to_cpu(disk->nr_blocks);
 588         md->current_era = le32_to_cpu(disk->current_era);
 589
 590         ws_unpack(&disk->current_writeset, &md->current_writeset->md);
 591         md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root);
 592         md->era_array_root = le64_to_cpu(disk->era_array_root);
 593         md->metadata_snap = le64_to_cpu(disk->metadata_snap);
 594         md->archived_writesets = true;
 595
 596         dm_bm_unlock(sblock);
 597
 598         return 0;
 599
 600 bad:
 601         dm_bm_unlock(sblock);
 602         return r;
 603 }
 604
 605 static int open_or_format_metadata(struct era_metadata *md,
 606                                    bool may_format)
 607 {
 608         int r;
 609         bool unformatted = false;
 610
 611         r = superblock_all_zeroes(md->bm, &unformatted);
 612         if (r)
 613                 return r;
 614
 615         if (unformatted)
 616                 return may_format ? format_metadata(md) : -EPERM;
 617
 618         return open_metadata(md);
 619 }
 620
 621 static int create_persistent_data_objects(struct era_metadata *md,
 622                                           bool may_format)
 623 {
 624         int r;
 625
 626         md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE,
 627                                          ERA_MAX_CONCURRENT_LOCKS);
 628         if (IS_ERR(md->bm)) {
 629                 DMERR("could not create block manager");
 630                 return PTR_ERR(md->bm);
 631         }
 632
 633         r = open_or_format_metadata(md, may_format);
 634         if (r)
 635                 dm_block_manager_destroy(md->bm);
 636
 637         return r;
 638 }
 639
 640 static void destroy_persistent_data_objects(struct era_metadata *md)
 641 {
 642         dm_sm_destroy(md->sm);
 643         dm_tm_destroy(md->tm);
 644         dm_block_manager_destroy(md->bm);
 645 }
 646
 647 /*
 648  * This waits until all era_map threads have picked up the new filter.
 649  */
 650 static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset)
 651 {
 652         rcu_assign_pointer(md->current_writeset, new_writeset);
 653         synchronize_rcu();
 654 }
 655
 656 /*----------------------------------------------------------------
 657  * Writesets get 'digested' into the main era array.
 658  *
 659  * We're using a coroutine here so the worker thread can do the digestion,
 660  * thus avoiding synchronisation of the metadata.  Digesting a whole
 661  * writeset in one go would cause too much latency.
 662  *--------------------------------------------------------------*/
 663 struct digest {
 664         uint32_t era;
 665         unsigned nr_bits, current_bit;
 666         struct writeset_metadata writeset;
 667         __le32 value;
 668         struct dm_disk_bitset info;
 669
 670         int (*step)(struct era_metadata *, struct digest *);
 671 };
 672
 673 static int metadata_digest_lookup_writeset(struct era_metadata *md,
 674                                            struct digest *d);
 675
 676 static int metadata_digest_remove_writeset(struct era_metadata *md,
 677                                            struct digest *d)
 678 {
 679         int r;
 680         uint64_t key = d->era;
 681
 682         r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root,
 683                             &key, &md->writeset_tree_root);
 684         if (r) {
 685                 DMERR("%s: dm_btree_remove failed", __func__);
 686                 return r;
 687         }
 688
 689         d->step = metadata_digest_lookup_writeset;
 690         return 0;
 691 }
 692
 693 #define INSERTS_PER_STEP 100
 694
 695 static int metadata_digest_transcribe_writeset(struct era_metadata *md,
 696                                                struct digest *d)
 697 {
 698         int r;
 699         bool marked;
 700         unsigned b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits);
 701
 702         for (b = d->current_bit; b < e; b++) {
 703                 r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked);
 704                 if (r) {
 705                         DMERR("%s: writeset_marked_on_disk failed", __func__);
 706                         return r;
 707                 }
 708
 709                 if (!marked)
 710                         continue;
 711
 712                 __dm_bless_for_disk(&d->value);
 713                 r = dm_array_set_value(&md->era_array_info, md->era_array_root,
 714                                        b, &d->value, &md->era_array_root);
 715                 if (r) {
 716                         DMERR("%s: dm_array_set_value failed", __func__);
 717                         return r;
 718                 }
 719         }
 720
 721         if (b == d->nr_bits)
 722                 d->step = metadata_digest_remove_writeset;
 723         else
 724                 d->current_bit = b;
 725
 726         return 0;
 727 }
 728
 729 static int metadata_digest_lookup_writeset(struct era_metadata *md,
 730                                            struct digest *d)
 731 {
 732         int r;
 733         uint64_t key;
 734         struct writeset_disk disk;
 735
 736         r = dm_btree_find_lowest_key(&md->writeset_tree_info,
 737                                      md->writeset_tree_root, &key);
 738         if (r < 0)
 739                 return r;
 740
 741         d->era = key;
 742
 743         r = dm_btree_lookup(&md->writeset_tree_info,
 744                             md->writeset_tree_root, &key, &disk);
 745         if (r) {
 746                 if (r == -ENODATA) {
 747                         d->step = NULL;
 748                         return 0;
 749                 }
 750
 751                 DMERR("%s: dm_btree_lookup failed", __func__);
 752                 return r;
 753         }
 754
 755         ws_unpack(&disk, &d->writeset);
 756         d->value = cpu_to_le32(key);
 757
 758         /*
 759          * We initialise another bitset info to avoid any caching side effects
 760          * with the previous one.
 761          */
 762         dm_disk_bitset_init(md->tm, &d->info);
 763
 764         d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks);
 765         d->current_bit = 0;
 766         d->step = metadata_digest_transcribe_writeset;
 767
 768         return 0;
 769 }
 770
 771 static int metadata_digest_start(struct era_metadata *md, struct digest *d)
 772 {
 773         if (d->step)
 774                 return 0;
 775
 776         memset(d, 0, sizeof(*d));
 777         d->step = metadata_digest_lookup_writeset;
 778
 779         return 0;
 780 }
 781
 782 /*----------------------------------------------------------------
 783  * High level metadata interface.  Target methods should use these, and not
 784  * the lower level ones.
 785  *--------------------------------------------------------------*/
 786 static struct era_metadata *metadata_open(struct block_device *bdev,
 787                                           sector_t block_size,
 788                                           bool may_format)
 789 {
 790         int r;
 791         struct era_metadata *md = kzalloc(sizeof(*md), GFP_KERNEL);
 792
 793         if (!md)
 794                 return NULL;
 795
 796         md->bdev = bdev;
 797         md->block_size = block_size;
 798
 799         md->writesets[0].md.root = INVALID_WRITESET_ROOT;
 800         md->writesets[1].md.root = INVALID_WRITESET_ROOT;
 801         md->current_writeset = &md->writesets[0];
 802
 803         r = create_persistent_data_objects(md, may_format);
 804         if (r) {
 805                 kfree(md);
 806                 return ERR_PTR(r);
 807         }
 808
 809         return md;
 810 }
 811
 812 static void metadata_close(struct era_metadata *md)
 813 {
 814         writeset_free(&md->writesets[0]);
 815         writeset_free(&md->writesets[1]);
 816         destroy_persistent_data_objects(md);
 817         kfree(md);
 818 }
 819
 820 static bool valid_nr_blocks(dm_block_t n)
 821 {
 822         /*
 823          * dm_bitset restricts us to 2^32.  test_bit & co. restrict us
 824          * further to 2^31 - 1
 825          */
 826         return n < (1ull << 31);
 827 }
 828
 829 static int metadata_resize(struct era_metadata *md, void *arg)
 830 {
 831         int r;
 832         dm_block_t *new_size = arg;
 833         __le32 value;
 834
 835         if (!valid_nr_blocks(*new_size)) {
 836                 DMERR("Invalid number of origin blocks %llu",
 837                       (unsigned long long) *new_size);
 838                 return -EINVAL;
 839         }
 840
 841         writeset_free(&md->writesets[0]);
 842         writeset_free(&md->writesets[1]);
 843
 844         r = writeset_alloc(&md->writesets[0], *new_size);
 845         if (r) {
 846                 DMERR("%s: writeset_alloc failed for writeset 0", __func__);
 847                 return r;
 848         }
 849
 850         r = writeset_alloc(&md->writesets[1], *new_size);
 851         if (r) {
 852                 DMERR("%s: writeset_alloc failed for writeset 1", __func__);
 853                 writeset_free(&md->writesets[0]);
 854                 return r;
 855         }
 856
 857         value = cpu_to_le32(0u);
 858         __dm_bless_for_disk(&value);
 859         r = dm_array_resize(&md->era_array_info, md->era_array_root,
 860                             md->nr_blocks, *new_size,
 861                             &value, &md->era_array_root);
 862         if (r) {
 863                 DMERR("%s: dm_array_resize failed", __func__);
 864                 writeset_free(&md->writesets[0]);
 865                 writeset_free(&md->writesets[1]);
 866                 return r;
 867         }
 868
 869         md->nr_blocks = *new_size;
 870         return 0;
 871 }
 872
 873 static int metadata_era_archive(struct era_metadata *md)
 874 {
 875         int r;
 876         uint64_t keys[1];
 877         struct writeset_disk value;
 878
 879         r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
 880                             &md->current_writeset->md.root);
 881         if (r) {
 882                 DMERR("%s: dm_bitset_flush failed", __func__);
 883                 return r;
 884         }
 885
 886         ws_pack(&md->current_writeset->md, &value);
 887
 888         keys[0] = md->current_era;
 889         __dm_bless_for_disk(&value);
 890         r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root,
 891                             keys, &value, &md->writeset_tree_root);
 892         if (r) {
 893                 DMERR("%s: couldn't insert writeset into btree", __func__);
 894                 /* FIXME: fail mode */
 895                 return r;
 896         }
 897
 898         md->current_writeset->md.root = INVALID_WRITESET_ROOT;
 899         md->archived_writesets = true;
 900
 901         return 0;
 902 }
 903
 904 static struct writeset *next_writeset(struct era_metadata *md)
 905 {
 906         return (md->current_writeset == &md->writesets[0]) ?
 907                 &md->writesets[1] : &md->writesets[0];
 908 }
 909
 910 static int metadata_new_era(struct era_metadata *md)
 911 {
 912         int r;
 913         struct writeset *new_writeset = next_writeset(md);
 914
 915         r = writeset_init(&md->bitset_info, new_writeset, md->nr_blocks);
 916         if (r) {
 917                 DMERR("%s: writeset_init failed", __func__);
 918                 return r;
 919         }
 920
 921         swap_writeset(md, new_writeset);
 922         md->current_era++;
 923
 924         return 0;
 925 }
 926
 927 static int metadata_era_rollover(struct era_metadata *md)
 928 {
 929         int r;
 930
 931         if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
 932                 r = metadata_era_archive(md);
 933                 if (r) {
 934                         DMERR("%s: metadata_archive_era failed", __func__);
 935                         /* FIXME: fail mode? */
 936                         return r;
 937                 }
 938         }
 939
 940         r = metadata_new_era(md);
 941         if (r) {
 942                 DMERR("%s: new era failed", __func__);
 943                 /* FIXME: fail mode */
 944                 return r;
 945         }
 946
 947         return 0;
 948 }
 949
 950 static bool metadata_current_marked(struct era_metadata *md, dm_block_t block)
 951 {
 952         bool r;
 953         struct writeset *ws;
 954
 955         rcu_read_lock();
 956         ws = rcu_dereference(md->current_writeset);
 957         r = writeset_marked(ws, block);
 958         rcu_read_unlock();
 959
 960         return r;
 961 }
 962
 963 static int metadata_commit(struct era_metadata *md)
 964 {
 965         int r;
 966         struct dm_block *sblock;
 967
 968         if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
 969                 r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
 970                                     &md->current_writeset->md.root);
 971                 if (r) {
 972                         DMERR("%s: bitset flush failed", __func__);
 973                         return r;
 974                 }
 975         }
 976
 977         r = dm_tm_pre_commit(md->tm);
 978         if (r) {
 979                 DMERR("%s: pre commit failed", __func__);
 980                 return r;
 981         }
 982
 983         r = save_sm_root(md);
 984         if (r) {
 985                 DMERR("%s: save_sm_root failed", __func__);
 986                 return r;
 987         }
 988
 989         r = superblock_lock(md, &sblock);
 990         if (r) {
 991                 DMERR("%s: superblock lock failed", __func__);
 992                 return r;
 993         }
 994
 995         prepare_superblock(md, dm_block_data(sblock));
 996
 997         return dm_tm_commit(md->tm, sblock);
 998 }
 999
1000 static int metadata_checkpoint(struct era_metadata *md)
1001 {
1002         /*
1003          * For now we just rollover, but later I want to put a check in to
1004          * avoid this if the filter is still pretty fresh.
1005          */
1006         return metadata_era_rollover(md);
1007 }
1008
1009 /*
1010  * Metadata snapshots allow userland to access era data.
1011  */
1012 static int metadata_take_snap(struct era_metadata *md)
1013 {
1014         int r, inc;
1015         struct dm_block *clone;
1016
1017         if (md->metadata_snap != SUPERBLOCK_LOCATION) {
1018                 DMERR("%s: metadata snapshot already exists", __func__);
1019                 return -EINVAL;
1020         }
1021
1022         r = metadata_era_rollover(md);
1023         if (r) {
1024                 DMERR("%s: era rollover failed", __func__);
1025                 return r;
1026         }
1027
1028         r = metadata_commit(md);
1029         if (r) {
1030                 DMERR("%s: pre commit failed", __func__);
1031                 return r;
1032         }
1033
1034         r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION);
1035         if (r) {
1036                 DMERR("%s: couldn't increment superblock", __func__);
1037                 return r;
1038         }
1039
1040         r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION,
1041                                &sb_validator, &clone, &inc);
1042         if (r) {
1043                 DMERR("%s: couldn't shadow superblock", __func__);
1044                 dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION);
1045                 return r;
1046         }
1047         BUG_ON(!inc);
1048
1049         r = dm_sm_inc_block(md->sm, md->writeset_tree_root);
1050         if (r) {
1051                 DMERR("%s: couldn't inc writeset tree root", __func__);
1052                 dm_tm_unlock(md->tm, clone);
1053                 return r;
1054         }
1055
1056         r = dm_sm_inc_block(md->sm, md->era_array_root);
1057         if (r) {
1058                 DMERR("%s: couldn't inc era tree root", __func__);
1059                 dm_sm_dec_block(md->sm, md->writeset_tree_root);
1060                 dm_tm_unlock(md->tm, clone);
1061                 return r;
1062         }
1063
1064         md->metadata_snap = dm_block_location(clone);
1065
1066         dm_tm_unlock(md->tm, clone);
1067
1068         return 0;
1069 }
1070
1071 static int metadata_drop_snap(struct era_metadata *md)
1072 {
1073         int r;
1074         dm_block_t location;
1075         struct dm_block *clone;
1076         struct superblock_disk *disk;
1077
1078         if (md->metadata_snap == SUPERBLOCK_LOCATION) {
1079                 DMERR("%s: no snap to drop", __func__);
1080                 return -EINVAL;
1081         }
1082
1083         r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone);
1084         if (r) {
1085                 DMERR("%s: couldn't read lock superblock clone", __func__);
1086                 return r;
1087         }
1088
1089         /*
1090          * Whatever happens now we'll commit with no record of the metadata
1091          * snap.
1092          */
1093         md->metadata_snap = SUPERBLOCK_LOCATION;
1094
1095         disk = dm_block_data(clone);
1096         r = dm_btree_del(&md->writeset_tree_info,
1097                          le64_to_cpu(disk->writeset_tree_root));
1098         if (r) {
1099                 DMERR("%s: error deleting writeset tree clone", __func__);
1100                 dm_tm_unlock(md->tm, clone);
1101                 return r;
1102         }
1103
1104         r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root));
1105         if (r) {
1106                 DMERR("%s: error deleting era array clone", __func__);
1107                 dm_tm_unlock(md->tm, clone);
1108                 return r;
1109         }
1110
1111         location = dm_block_location(clone);
1112         dm_tm_unlock(md->tm, clone);
1113
1114         return dm_sm_dec_block(md->sm, location);
1115 }
1116
1117 struct metadata_stats {
1118         dm_block_t used;
1119         dm_block_t total;
1120         dm_block_t snap;
1121         uint32_t era;
1122 };
1123
1124 static int metadata_get_stats(struct era_metadata *md, void *ptr)
1125 {
1126         int r;
1127         struct metadata_stats *s = ptr;
1128         dm_block_t nr_free, nr_total;
1129
1130         r = dm_sm_get_nr_free(md->sm, &nr_free);
1131         if (r) {
1132                 DMERR("dm_sm_get_nr_free returned %d", r);
1133                 return r;
1134         }
1135
1136         r = dm_sm_get_nr_blocks(md->sm, &nr_total);
1137         if (r) {
1138                 DMERR("dm_pool_get_metadata_dev_size returned %d", r);
1139                 return r;
1140         }
1141
1142         s->used = nr_total - nr_free;
1143         s->total = nr_total;
1144         s->snap = md->metadata_snap;
1145         s->era = md->current_era;
1146
1147         return 0;
1148 }
1149
1150 /*----------------------------------------------------------------*/
1151
1152 struct era {
1153         struct dm_target *ti;
1154         struct dm_target_callbacks callbacks;
1155
1156         struct dm_dev *metadata_dev;
1157         struct dm_dev *origin_dev;
1158
1159         dm_block_t nr_blocks;
1160         uint32_t sectors_per_block;
1161         int sectors_per_block_shift;
1162         struct era_metadata *md;
1163
1164         struct workqueue_struct *wq;
1165         struct work_struct worker;
1166
1167         spinlock_t deferred_lock;
1168         struct bio_list deferred_bios;
1169
1170         spinlock_t rpc_lock;
1171         struct list_head rpc_calls;
1172
1173         struct digest digest;
1174         atomic_t suspended;
1175 };
1176
1177 struct rpc {
1178         struct list_head list;
1179
1180         int (*fn0)(struct era_metadata *);
1181         int (*fn1)(struct era_metadata *, void *);
1182         void *arg;
1183         int result;
1184
1185         struct completion complete;
1186 };
1187
1188 /*----------------------------------------------------------------
1189  * Remapping.
1190  *---------------------------------------------------------------*/
1191 static bool block_size_is_power_of_two(struct era *era)
1192 {
1193         return era->sectors_per_block_shift >= 0;
1194 }
1195
1196 static dm_block_t get_block(struct era *era, struct bio *bio)
1197 {
1198         sector_t block_nr = bio->bi_iter.bi_sector;
1199
1200         if (!block_size_is_power_of_two(era))
1201                 (void) sector_div(block_nr, era->sectors_per_block);
1202         else
1203                 block_nr >>= era->sectors_per_block_shift;
1204
1205         return block_nr;
1206 }
1207
1208 static void remap_to_origin(struct era *era, struct bio *bio)
1209 {
1210         bio_set_dev(bio, era->origin_dev->bdev);
1211 }
1212
1213 /*----------------------------------------------------------------
1214  * Worker thread
1215  *--------------------------------------------------------------*/
1216 static void wake_worker(struct era *era)
1217 {
1218         if (!atomic_read(&era->suspended))
1219                 queue_work(era->wq, &era->worker);
1220 }
1221
1222 static void process_old_eras(struct era *era)
1223 {
1224         int r;
1225
1226         if (!era->digest.step)
1227                 return;
1228
1229         r = era->digest.step(era->md, &era->digest);
1230         if (r < 0) {
1231                 DMERR("%s: digest step failed, stopping digestion", __func__);
1232                 era->digest.step = NULL;
1233
1234         } else if (era->digest.step)
1235                 wake_worker(era);
1236 }
1237
1238 static void process_deferred_bios(struct era *era)
1239 {
1240         int r;
1241         struct bio_list deferred_bios, marked_bios;
1242         struct bio *bio;
1243         struct blk_plug plug;
1244         bool commit_needed = false;
1245         bool failed = false;
1246         struct writeset *ws = era->md->current_writeset;
1247
1248         bio_list_init(&deferred_bios);
1249         bio_list_init(&marked_bios);
1250
1251         spin_lock(&era->deferred_lock);
1252         bio_list_merge(&deferred_bios, &era->deferred_bios);
1253         bio_list_init(&era->deferred_bios);
1254         spin_unlock(&era->deferred_lock);
1255
1256         if (bio_list_empty(&deferred_bios))
1257                 return;
1258
1259         while ((bio = bio_list_pop(&deferred_bios))) {
1260                 r = writeset_test_and_set(&era->md->bitset_info, ws,
1261                                           get_block(era, bio));
1262                 if (r < 0) {
1263                         /*
1264                          * This is bad news, we need to rollback.
1265                          * FIXME: finish.
1266                          */
1267                         failed = true;
1268                 } else if (r == 0)
1269                         commit_needed = true;
1270
1271                 bio_list_add(&marked_bios, bio);
1272         }
1273
1274         if (commit_needed) {
1275                 r = metadata_commit(era->md);
1276                 if (r)
1277                         failed = true;
1278         }
1279
1280         if (failed)
1281                 while ((bio = bio_list_pop(&marked_bios)))
1282                         bio_io_error(bio);
1283         else {
1284                 blk_start_plug(&plug);
1285                 while ((bio = bio_list_pop(&marked_bios))) {
1286                         /*
1287                          * Only update the in-core writeset if the on-disk one
1288                          * was updated too.
1289                          */
1290                         if (commit_needed)
1291                                 set_bit(get_block(era, bio), ws->bits);
1292                         generic_make_request(bio);
1293                 }
1294                 blk_finish_plug(&plug);
1295         }
1296 }
1297
1298 static void process_rpc_calls(struct era *era)
1299 {
1300         int r;
1301         bool need_commit = false;
1302         struct list_head calls;
1303         struct rpc *rpc, *tmp;
1304
1305         INIT_LIST_HEAD(&calls);
1306         spin_lock(&era->rpc_lock);
1307         list_splice_init(&era->rpc_calls, &calls);
1308         spin_unlock(&era->rpc_lock);
1309
1310         list_for_each_entry_safe(rpc, tmp, &calls, list) {
1311                 rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg);
1312                 need_commit = true;
1313         }
1314
1315         if (need_commit) {
1316                 r = metadata_commit(era->md);
1317                 if (r)
1318                         list_for_each_entry_safe(rpc, tmp, &calls, list)
1319                                 rpc->result = r;
1320         }
1321
1322         list_for_each_entry_safe(rpc, tmp, &calls, list)
1323                 complete(&rpc->complete);
1324 }
1325
1326 static void kick_off_digest(struct era *era)
1327 {
1328         if (era->md->archived_writesets) {
1329                 era->md->archived_writesets = false;
1330                 metadata_digest_start(era->md, &era->digest);
1331         }
1332 }
1333
1334 static void do_work(struct work_struct *ws)
1335 {
1336         struct era *era = container_of(ws, struct era, worker);
1337
1338         kick_off_digest(era);
1339         process_old_eras(era);
1340         process_deferred_bios(era);
1341         process_rpc_calls(era);
1342 }
1343
1344 static void defer_bio(struct era *era, struct bio *bio)
1345 {
1346         spin_lock(&era->deferred_lock);
1347         bio_list_add(&era->deferred_bios, bio);
1348         spin_unlock(&era->deferred_lock);
1349
1350         wake_worker(era);
1351 }
1352
1353 /*
1354  * Make an rpc call to the worker to change the metadata.
1355  */
1356 static int perform_rpc(struct era *era, struct rpc *rpc)
1357 {
1358         rpc->result = 0;
1359         init_completion(&rpc->complete);
1360
1361         spin_lock(&era->rpc_lock);
1362         list_add(&rpc->list, &era->rpc_calls);
1363         spin_unlock(&era->rpc_lock);
1364
1365         wake_worker(era);
1366         wait_for_completion(&rpc->complete);
1367
1368         return rpc->result;
1369 }
1370
1371 static int in_worker0(struct era *era, int (*fn)(struct era_metadata *))
1372 {
1373         struct rpc rpc;
1374         rpc.fn0 = fn;
1375         rpc.fn1 = NULL;
1376
1377         return perform_rpc(era, &rpc);
1378 }
1379
1380 static int in_worker1(struct era *era,
1381                       int (*fn)(struct era_metadata *, void *), void *arg)
1382 {
1383         struct rpc rpc;
1384         rpc.fn0 = NULL;
1385         rpc.fn1 = fn;
1386         rpc.arg = arg;
1387
1388         return perform_rpc(era, &rpc);
1389 }
1390
1391 static void start_worker(struct era *era)
1392 {
1393         atomic_set(&era->suspended, 0);
1394 }
1395
1396 static void stop_worker(struct era *era)
1397 {
1398         atomic_set(&era->suspended, 1);
1399         drain_workqueue(era->wq);
1400 }
1401
1402 /*----------------------------------------------------------------
1403  * Target methods
1404  *--------------------------------------------------------------*/
1405 static int dev_is_congested(struct dm_dev *dev, int bdi_bits)
1406 {
1407         struct request_queue *q = bdev_get_queue(dev->bdev);
1408         return bdi_congested(q->backing_dev_info, bdi_bits);
1409 }
1410
1411 static int era_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1412 {
1413         struct era *era = container_of(cb, struct era, callbacks);
1414         return dev_is_congested(era->origin_dev, bdi_bits);
1415 }
1416
1417 static void era_destroy(struct era *era)
1418 {
1419         if (era->md)
1420                 metadata_close(era->md);
1421
1422         if (era->wq)
1423                 destroy_workqueue(era->wq);
1424
1425         if (era->origin_dev)
1426                 dm_put_device(era->ti, era->origin_dev);
1427
1428         if (era->metadata_dev)
1429                 dm_put_device(era->ti, era->metadata_dev);
1430
1431         kfree(era);
1432 }
1433
1434 static dm_block_t calc_nr_blocks(struct era *era)
1435 {
1436         return dm_sector_div_up(era->ti->len, era->sectors_per_block);
1437 }
1438
1439 static bool valid_block_size(dm_block_t block_size)
1440 {
1441         bool greater_than_zero = block_size > 0;
1442         bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0;
1443
1444         return greater_than_zero && multiple_of_min_block_size;
1445 }
1446
1447 /*
1448  * <metadata dev> <data dev> <data block size (sectors)>
1449  */
1450 static int era_ctr(struct dm_target *ti, unsigned argc, char **argv)
1451 {
1452         int r;
1453         char dummy;
1454         struct era *era;
1455         struct era_metadata *md;
1456
1457         if (argc != 3) {
1458                 ti->error = "Invalid argument count";
1459                 return -EINVAL;
1460         }
1461
1462         era = kzalloc(sizeof(*era), GFP_KERNEL);
1463         if (!era) {
1464                 ti->error = "Error allocating era structure";
1465                 return -ENOMEM;
1466         }
1467
1468         era->ti = ti;
1469
1470         r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &era->metadata_dev);
1471         if (r) {
1472                 ti->error = "Error opening metadata device";
1473                 era_destroy(era);
1474                 return -EINVAL;
1475         }
1476
1477         r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &era->origin_dev);
1478         if (r) {
1479                 ti->error = "Error opening data device";
1480                 era_destroy(era);
1481                 return -EINVAL;
1482         }
1483
1484         r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy);
1485         if (r != 1) {
1486                 ti->error = "Error parsing block size";
1487                 era_destroy(era);
1488                 return -EINVAL;
1489         }
1490
1491         r = dm_set_target_max_io_len(ti, era->sectors_per_block);
1492         if (r) {
1493                 ti->error = "could not set max io len";
1494                 era_destroy(era);
1495                 return -EINVAL;
1496         }
1497
1498         if (!valid_block_size(era->sectors_per_block)) {
1499                 ti->error = "Invalid block size";
1500                 era_destroy(era);
1501                 return -EINVAL;
1502         }
1503         if (era->sectors_per_block & (era->sectors_per_block - 1))
1504                 era->sectors_per_block_shift = -1;
1505         else
1506                 era->sectors_per_block_shift = __ffs(era->sectors_per_block);
1507
1508         md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true);
1509         if (IS_ERR(md)) {
1510                 ti->error = "Error reading metadata";
1511                 era_destroy(era);
1512                 return PTR_ERR(md);
1513         }
1514         era->md = md;
1515
1516         era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1517         if (!era->wq) {
1518                 ti->error = "could not create workqueue for metadata object";
1519                 era_destroy(era);
1520                 return -ENOMEM;
1521         }
1522         INIT_WORK(&era->worker, do_work);
1523
1524         spin_lock_init(&era->deferred_lock);
1525         bio_list_init(&era->deferred_bios);
1526
1527         spin_lock_init(&era->rpc_lock);
1528         INIT_LIST_HEAD(&era->rpc_calls);
1529
1530         ti->private = era;
1531         ti->num_flush_bios = 1;
1532         ti->flush_supported = true;
1533
1534         ti->num_discard_bios = 1;
1535         ti->discards_supported = true;
1536         era->callbacks.congested_fn = era_is_congested;
1537         dm_table_add_target_callbacks(ti->table, &era->callbacks);
1538
1539         return 0;
1540 }
1541
1542 static void era_dtr(struct dm_target *ti)
1543 {
1544         era_destroy(ti->private);
1545 }
1546
1547 static int era_map(struct dm_target *ti, struct bio *bio)
1548 {
1549         struct era *era = ti->private;
1550         dm_block_t block = get_block(era, bio);
1551
1552         /*
1553          * All bios get remapped to the origin device.  We do this now, but
1554          * it may not get issued until later.  Depending on whether the
1555          * block is marked in this era.
1556          */
1557         remap_to_origin(era, bio);
1558
1559         /*
1560          * REQ_PREFLUSH bios carry no data, so we're not interested in them.
1561          */
1562         if (!(bio->bi_opf & REQ_PREFLUSH) &&
1563             (bio_data_dir(bio) == WRITE) &&
1564             !metadata_current_marked(era->md, block)) {
1565                 defer_bio(era, bio);
1566                 return DM_MAPIO_SUBMITTED;
1567         }
1568
1569         return DM_MAPIO_REMAPPED;
1570 }
1571
1572 static void era_postsuspend(struct dm_target *ti)
1573 {
1574         int r;
1575         struct era *era = ti->private;
1576
1577         r = in_worker0(era, metadata_era_archive);
1578         if (r) {
1579                 DMERR("%s: couldn't archive current era", __func__);
1580                 /* FIXME: fail mode */
1581         }
1582
1583         stop_worker(era);
1584
1585         r = metadata_commit(era->md);
1586         if (r) {
1587                 DMERR("%s: metadata_commit failed", __func__);
1588                 /* FIXME: fail mode */
1589         }
1590 }
1591
1592 static int era_preresume(struct dm_target *ti)
1593 {
1594         int r;
1595         struct era *era = ti->private;
1596         dm_block_t new_size = calc_nr_blocks(era);
1597
1598         if (era->nr_blocks != new_size) {
1599                 r = metadata_resize(era->md, &new_size);
1600                 if (r) {
1601                         DMERR("%s: metadata_resize failed", __func__);
1602                         return r;
1603                 }
1604
1605                 r = metadata_commit(era->md);
1606                 if (r) {
1607                         DMERR("%s: metadata_commit failed", __func__);
1608                         return r;
1609                 }
1610
1611                 era->nr_blocks = new_size;
1612         }
1613
1614         start_worker(era);
1615
1616         r = in_worker0(era, metadata_era_rollover);
1617         if (r) {
1618                 DMERR("%s: metadata_era_rollover failed", __func__);
1619                 return r;
1620         }
1621
1622         return 0;
1623 }
1624
1625 /*
1626  * Status format:
1627  *
1628  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
1629  * <current era> <held metadata root | '-'>
1630  */
1631 static void era_status(struct dm_target *ti, status_type_t type,
1632                        unsigned status_flags, char *result, unsigned maxlen)
1633 {
1634         int r;
1635         struct era *era = ti->private;
1636         ssize_t sz = 0;
1637         struct metadata_stats stats;
1638         char buf[BDEVNAME_SIZE];
1639
1640         switch (type) {
1641         case STATUSTYPE_INFO:
1642                 r = in_worker1(era, metadata_get_stats, &stats);
1643                 if (r)
1644                         goto err;
1645
1646                 DMEMIT("%u %llu/%llu %u",
1647                        (unsigned) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
1648                        (unsigned long long) stats.used,
1649                        (unsigned long long) stats.total,
1650                        (unsigned) stats.era);
1651
1652                 if (stats.snap != SUPERBLOCK_LOCATION)
1653                         DMEMIT(" %llu", stats.snap);
1654                 else
1655                         DMEMIT(" -");
1656                 break;
1657
1658         case STATUSTYPE_TABLE:
1659                 format_dev_t(buf, era->metadata_dev->bdev->bd_dev);
1660                 DMEMIT("%s ", buf);
1661                 format_dev_t(buf, era->origin_dev->bdev->bd_dev);
1662                 DMEMIT("%s %u", buf, era->sectors_per_block);
1663                 break;
1664         }
1665
1666         return;
1667
1668 err:
1669         DMEMIT("Error");
1670 }
1671
1672 static int era_message(struct dm_target *ti, unsigned argc, char **argv)
1673 {
1674         struct era *era = ti->private;
1675
1676         if (argc != 1) {
1677                 DMERR("incorrect number of message arguments");
1678                 return -EINVAL;
1679         }
1680
1681         if (!strcasecmp(argv[0], "checkpoint"))
1682                 return in_worker0(era, metadata_checkpoint);
1683
1684         if (!strcasecmp(argv[0], "take_metadata_snap"))
1685                 return in_worker0(era, metadata_take_snap);
1686
1687         if (!strcasecmp(argv[0], "drop_metadata_snap"))
1688                 return in_worker0(era, metadata_drop_snap);
1689
1690         DMERR("unsupported message '%s'", argv[0]);
1691         return -EINVAL;
1692 }
1693
1694 static sector_t get_dev_size(struct dm_dev *dev)
1695 {
1696         return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1697 }
1698
1699 static int era_iterate_devices(struct dm_target *ti,
1700                                iterate_devices_callout_fn fn, void *data)
1701 {
1702         struct era *era = ti->private;
1703         return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data);
1704 }
1705
1706 static void era_io_hints(struct dm_target *ti, struct queue_limits *limits)
1707 {
1708         struct era *era = ti->private;
1709         uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
1710
1711         /*
1712          * If the system-determined stacked limits are compatible with the
1713          * era device's blocksize (io_opt is a factor) do not override them.
1714          */
1715         if (io_opt_sectors < era->sectors_per_block ||
1716             do_div(io_opt_sectors, era->sectors_per_block)) {
1717                 blk_limits_io_min(limits, 0);
1718                 blk_limits_io_opt(limits, era->sectors_per_block << SECTOR_SHIFT);
1719         }
1720 }
1721
1722 /*----------------------------------------------------------------*/
1723
1724 static struct target_type era_target = {
1725         .name = "era",
1726         .version = {1, 0, 0},
1727         .module = THIS_MODULE,
1728         .ctr = era_ctr,
1729         .dtr = era_dtr,
1730         .map = era_map,
1731         .postsuspend = era_postsuspend,
1732         .preresume = era_preresume,
1733         .status = era_status,
1734         .message = era_message,
1735         .iterate_devices = era_iterate_devices,
1736         .io_hints = era_io_hints
1737 };
1738
1739 static int __init dm_era_init(void)
1740 {
1741         int r;
1742
1743         r = dm_register_target(&era_target);
1744         if (r) {
1745                 DMERR("era target registration failed: %d", r);
1746                 return r;
1747         }
1748
1749         return 0;
1750 }
1751
1752 static void __exit dm_era_exit(void)
1753 {
1754         dm_unregister_target(&era_target);
1755 }
1756
1757 module_init(dm_era_init);
1758 module_exit(dm_era_exit);
1759
1760 MODULE_DESCRIPTION(DM_NAME " era target");
1761 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
1762 MODULE_LICENSE("GPL");