1 // SPDX-License-Identifier: GPL-2.0-only 2 #include "dm.h" 3 #include "persistent-data/dm-transaction-manager.h" 4 #include "persistent-data/dm-bitset.h" 5 #include "persistent-data/dm-space-map.h" 6 7 #include <linux/dm-io.h> 8 #include <linux/dm-kcopyd.h> 9 #include <linux/init.h> 10 #include <linux/mempool.h> 11 #include <linux/module.h> 12 #include <linux/slab.h> 13 #include <linux/vmalloc.h> 14 15 #define DM_MSG_PREFIX "era" 16 17 #define SUPERBLOCK_LOCATION 0 18 #define SUPERBLOCK_MAGIC 2126579579 19 #define SUPERBLOCK_CSUM_XOR 146538381 20 #define MIN_ERA_VERSION 1 21 #define MAX_ERA_VERSION 1 22 #define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION 23 #define MIN_BLOCK_SIZE 8 24 25 /* 26 *-------------------------------------------------------------- 27 * Writeset 28 *-------------------------------------------------------------- 29 */ 30 struct writeset_metadata { 31 uint32_t nr_bits; 32 dm_block_t root; 33 }; 34 35 struct writeset { 36 struct writeset_metadata md; 37 38 /* 39 * An in core copy of the bits to save constantly doing look ups on 40 * disk. 41 */ 42 unsigned long *bits; 43 }; 44 45 /* 46 * This does not free off the on disk bitset as this will normally be done 47 * after digesting into the era array. 48 */ 49 static void writeset_free(struct writeset *ws) 50 { 51 vfree(ws->bits); 52 ws->bits = NULL; 53 } 54 55 static int setup_on_disk_bitset(struct dm_disk_bitset *info, 56 unsigned int nr_bits, dm_block_t *root) 57 { 58 int r; 59 60 r = dm_bitset_empty(info, root); 61 if (r) 62 return r; 63 64 return dm_bitset_resize(info, *root, 0, nr_bits, false, root); 65 } 66 67 static size_t bitset_size(unsigned int nr_bits) 68 { 69 return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG); 70 } 71 72 /* 73 * Allocates memory for the in core bitset. 74 */ 75 static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks) 76 { 77 ws->bits = vzalloc(bitset_size(nr_blocks)); 78 if (!ws->bits) { 79 DMERR("%s: couldn't allocate in memory bitset", __func__); 80 return -ENOMEM; 81 } 82 83 return 0; 84 } 85 86 /* 87 * Wipes the in-core bitset, and creates a new on disk bitset. 88 */ 89 static int writeset_init(struct dm_disk_bitset *info, struct writeset *ws, 90 dm_block_t nr_blocks) 91 { 92 int r; 93 94 memset(ws->bits, 0, bitset_size(nr_blocks)); 95 96 ws->md.nr_bits = nr_blocks; 97 r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root); 98 if (r) { 99 DMERR("%s: setup_on_disk_bitset failed", __func__); 100 return r; 101 } 102 103 return 0; 104 } 105 106 static bool writeset_marked(struct writeset *ws, dm_block_t block) 107 { 108 return test_bit(block, ws->bits); 109 } 110 111 static int writeset_marked_on_disk(struct dm_disk_bitset *info, 112 struct writeset_metadata *m, dm_block_t block, 113 bool *result) 114 { 115 int r; 116 dm_block_t old = m->root; 117 118 /* 119 * The bitset was flushed when it was archived, so we know there'll 120 * be no change to the root. 121 */ 122 r = dm_bitset_test_bit(info, m->root, block, &m->root, result); 123 if (r) { 124 DMERR("%s: dm_bitset_test_bit failed", __func__); 125 return r; 126 } 127 128 BUG_ON(m->root != old); 129 130 return r; 131 } 132 133 /* 134 * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was. 135 */ 136 static int writeset_test_and_set(struct dm_disk_bitset *info, 137 struct writeset *ws, uint32_t block) 138 { 139 int r; 140 141 if (!test_bit(block, ws->bits)) { 142 r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root); 143 if (r) { 144 /* FIXME: fail mode */ 145 return r; 146 } 147 148 return 0; 149 } 150 151 return 1; 152 } 153 154 /* 155 *-------------------------------------------------------------- 156 * On disk metadata layout 157 *-------------------------------------------------------------- 158 */ 159 #define SPACE_MAP_ROOT_SIZE 128 160 #define UUID_LEN 16 161 162 struct writeset_disk { 163 __le32 nr_bits; 164 __le64 root; 165 } __packed; 166 167 struct superblock_disk { 168 __le32 csum; 169 __le32 flags; 170 __le64 blocknr; 171 172 __u8 uuid[UUID_LEN]; 173 __le64 magic; 174 __le32 version; 175 176 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE]; 177 178 __le32 data_block_size; 179 __le32 metadata_block_size; 180 __le32 nr_blocks; 181 182 __le32 current_era; 183 struct writeset_disk current_writeset; 184 185 /* 186 * Only these two fields are valid within the metadata snapshot. 187 */ 188 __le64 writeset_tree_root; 189 __le64 era_array_root; 190 191 __le64 metadata_snap; 192 } __packed; 193 194 /* 195 *-------------------------------------------------------------- 196 * Superblock validation 197 *-------------------------------------------------------------- 198 */ 199 static void sb_prepare_for_write(const struct dm_block_validator *v, 200 struct dm_block *b, 201 size_t sb_block_size) 202 { 203 struct superblock_disk *disk = dm_block_data(b); 204 205 disk->blocknr = cpu_to_le64(dm_block_location(b)); 206 disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags, 207 sb_block_size - sizeof(__le32), 208 SUPERBLOCK_CSUM_XOR)); 209 } 210 211 static int check_metadata_version(struct superblock_disk *disk) 212 { 213 uint32_t metadata_version = le32_to_cpu(disk->version); 214 215 if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) { 216 DMERR("Era metadata version %u found, but only versions between %u and %u supported.", 217 metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION); 218 return -EINVAL; 219 } 220 221 return 0; 222 } 223 224 static int sb_check(const struct dm_block_validator *v, 225 struct dm_block *b, 226 size_t sb_block_size) 227 { 228 struct superblock_disk *disk = dm_block_data(b); 229 __le32 csum_le; 230 231 if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) { 232 DMERR("%s failed: blocknr %llu: wanted %llu", 233 __func__, le64_to_cpu(disk->blocknr), 234 (unsigned long long)dm_block_location(b)); 235 return -ENOTBLK; 236 } 237 238 if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) { 239 DMERR("%s failed: magic %llu: wanted %llu", 240 __func__, le64_to_cpu(disk->magic), 241 (unsigned long long) SUPERBLOCK_MAGIC); 242 return -EILSEQ; 243 } 244 245 csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags, 246 sb_block_size - sizeof(__le32), 247 SUPERBLOCK_CSUM_XOR)); 248 if (csum_le != disk->csum) { 249 DMERR("%s failed: csum %u: wanted %u", 250 __func__, le32_to_cpu(csum_le), le32_to_cpu(disk->csum)); 251 return -EILSEQ; 252 } 253 254 return check_metadata_version(disk); 255 } 256 257 static const struct dm_block_validator sb_validator = { 258 .name = "superblock", 259 .prepare_for_write = sb_prepare_for_write, 260 .check = sb_check 261 }; 262 263 /* 264 *-------------------------------------------------------------- 265 * Low level metadata handling 266 *-------------------------------------------------------------- 267 */ 268 #define DM_ERA_METADATA_BLOCK_SIZE 4096 269 #define ERA_MAX_CONCURRENT_LOCKS 5 270 271 struct era_metadata { 272 struct block_device *bdev; 273 struct dm_block_manager *bm; 274 struct dm_space_map *sm; 275 struct dm_transaction_manager *tm; 276 277 dm_block_t block_size; 278 uint32_t nr_blocks; 279 280 uint32_t current_era; 281 282 /* 283 * We preallocate 2 writesets. When an era rolls over we 284 * switch between them. This means the allocation is done at 285 * preresume time, rather than on the io path. 286 */ 287 struct writeset writesets[2]; 288 struct writeset *current_writeset; 289 290 dm_block_t writeset_tree_root; 291 dm_block_t era_array_root; 292 293 struct dm_disk_bitset bitset_info; 294 struct dm_btree_info writeset_tree_info; 295 struct dm_array_info era_array_info; 296 297 dm_block_t metadata_snap; 298 299 /* 300 * A flag that is set whenever a writeset has been archived. 301 */ 302 bool archived_writesets; 303 304 /* 305 * Reading the space map root can fail, so we read it into this 306 * buffer before the superblock is locked and updated. 307 */ 308 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE]; 309 }; 310 311 static int superblock_read_lock(struct era_metadata *md, 312 struct dm_block **sblock) 313 { 314 return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION, 315 &sb_validator, sblock); 316 } 317 318 static int superblock_lock_zero(struct era_metadata *md, 319 struct dm_block **sblock) 320 { 321 return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION, 322 &sb_validator, sblock); 323 } 324 325 static int superblock_lock(struct era_metadata *md, 326 struct dm_block **sblock) 327 { 328 return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION, 329 &sb_validator, sblock); 330 } 331 332 /* FIXME: duplication with cache and thin */ 333 static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result) 334 { 335 int r; 336 unsigned int i; 337 struct dm_block *b; 338 __le64 *data_le, zero = cpu_to_le64(0); 339 unsigned int sb_block_size = dm_bm_block_size(bm) / sizeof(__le64); 340 341 /* 342 * We can't use a validator here - it may be all zeroes. 343 */ 344 r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b); 345 if (r) 346 return r; 347 348 data_le = dm_block_data(b); 349 *result = true; 350 for (i = 0; i < sb_block_size; i++) { 351 if (data_le[i] != zero) { 352 *result = false; 353 break; 354 } 355 } 356 357 dm_bm_unlock(b); 358 359 return 0; 360 } 361 362 /*----------------------------------------------------------------*/ 363 364 static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk) 365 { 366 disk->nr_bits = cpu_to_le32(core->nr_bits); 367 disk->root = cpu_to_le64(core->root); 368 } 369 370 static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core) 371 { 372 core->nr_bits = le32_to_cpu(disk->nr_bits); 373 core->root = le64_to_cpu(disk->root); 374 } 375 376 static void ws_inc(void *context, const void *value, unsigned int count) 377 { 378 struct era_metadata *md = context; 379 struct writeset_disk ws_d; 380 dm_block_t b; 381 unsigned int i; 382 383 for (i = 0; i < count; i++) { 384 memcpy(&ws_d, value + (i * sizeof(ws_d)), sizeof(ws_d)); 385 b = le64_to_cpu(ws_d.root); 386 dm_tm_inc(md->tm, b); 387 } 388 } 389 390 static void ws_dec(void *context, const void *value, unsigned int count) 391 { 392 struct era_metadata *md = context; 393 struct writeset_disk ws_d; 394 dm_block_t b; 395 unsigned int i; 396 397 for (i = 0; i < count; i++) { 398 memcpy(&ws_d, value + (i * sizeof(ws_d)), sizeof(ws_d)); 399 b = le64_to_cpu(ws_d.root); 400 dm_bitset_del(&md->bitset_info, b); 401 } 402 } 403 404 static int ws_eq(void *context, const void *value1, const void *value2) 405 { 406 return !memcmp(value1, value2, sizeof(struct writeset_disk)); 407 } 408 409 /*----------------------------------------------------------------*/ 410 411 static void setup_writeset_tree_info(struct era_metadata *md) 412 { 413 struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type; 414 415 md->writeset_tree_info.tm = md->tm; 416 md->writeset_tree_info.levels = 1; 417 vt->context = md; 418 vt->size = sizeof(struct writeset_disk); 419 vt->inc = ws_inc; 420 vt->dec = ws_dec; 421 vt->equal = ws_eq; 422 } 423 424 static void setup_era_array_info(struct era_metadata *md) 425 { 426 struct dm_btree_value_type vt; 427 428 vt.context = NULL; 429 vt.size = sizeof(__le32); 430 vt.inc = NULL; 431 vt.dec = NULL; 432 vt.equal = NULL; 433 434 dm_array_info_init(&md->era_array_info, md->tm, &vt); 435 } 436 437 static void setup_infos(struct era_metadata *md) 438 { 439 dm_disk_bitset_init(md->tm, &md->bitset_info); 440 setup_writeset_tree_info(md); 441 setup_era_array_info(md); 442 } 443 444 /*----------------------------------------------------------------*/ 445 446 static int create_fresh_metadata(struct era_metadata *md) 447 { 448 int r; 449 450 r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION, 451 &md->tm, &md->sm); 452 if (r < 0) { 453 DMERR("dm_tm_create_with_sm failed"); 454 return r; 455 } 456 457 setup_infos(md); 458 459 r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root); 460 if (r) { 461 DMERR("couldn't create new writeset tree"); 462 goto bad; 463 } 464 465 r = dm_array_empty(&md->era_array_info, &md->era_array_root); 466 if (r) { 467 DMERR("couldn't create era array"); 468 goto bad; 469 } 470 471 return 0; 472 473 bad: 474 dm_sm_destroy(md->sm); 475 dm_tm_destroy(md->tm); 476 477 return r; 478 } 479 480 static int save_sm_root(struct era_metadata *md) 481 { 482 int r; 483 size_t metadata_len; 484 485 r = dm_sm_root_size(md->sm, &metadata_len); 486 if (r < 0) 487 return r; 488 489 return dm_sm_copy_root(md->sm, &md->metadata_space_map_root, 490 metadata_len); 491 } 492 493 static void copy_sm_root(struct era_metadata *md, struct superblock_disk *disk) 494 { 495 memcpy(&disk->metadata_space_map_root, 496 &md->metadata_space_map_root, 497 sizeof(md->metadata_space_map_root)); 498 } 499 500 /* 501 * Writes a superblock, including the static fields that don't get updated 502 * with every commit (possible optimisation here). 'md' should be fully 503 * constructed when this is called. 504 */ 505 static void prepare_superblock(struct era_metadata *md, struct superblock_disk *disk) 506 { 507 disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC); 508 disk->flags = cpu_to_le32(0ul); 509 510 /* FIXME: can't keep blanking the uuid (uuid is currently unused though) */ 511 memset(disk->uuid, 0, sizeof(disk->uuid)); 512 disk->version = cpu_to_le32(MAX_ERA_VERSION); 513 514 copy_sm_root(md, disk); 515 516 disk->data_block_size = cpu_to_le32(md->block_size); 517 disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT); 518 disk->nr_blocks = cpu_to_le32(md->nr_blocks); 519 disk->current_era = cpu_to_le32(md->current_era); 520 521 ws_pack(&md->current_writeset->md, &disk->current_writeset); 522 disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root); 523 disk->era_array_root = cpu_to_le64(md->era_array_root); 524 disk->metadata_snap = cpu_to_le64(md->metadata_snap); 525 } 526 527 static int write_superblock(struct era_metadata *md) 528 { 529 int r; 530 struct dm_block *sblock; 531 struct superblock_disk *disk; 532 533 r = save_sm_root(md); 534 if (r) { 535 DMERR("%s: save_sm_root failed", __func__); 536 return r; 537 } 538 539 r = superblock_lock_zero(md, &sblock); 540 if (r) 541 return r; 542 543 disk = dm_block_data(sblock); 544 prepare_superblock(md, disk); 545 546 return dm_tm_commit(md->tm, sblock); 547 } 548 549 /* 550 * Assumes block_size and the infos are set. 551 */ 552 static int format_metadata(struct era_metadata *md) 553 { 554 int r; 555 556 r = create_fresh_metadata(md); 557 if (r) 558 return r; 559 560 r = write_superblock(md); 561 if (r) { 562 dm_sm_destroy(md->sm); 563 dm_tm_destroy(md->tm); 564 return r; 565 } 566 567 return 0; 568 } 569 570 static int open_metadata(struct era_metadata *md) 571 { 572 int r; 573 struct dm_block *sblock; 574 struct superblock_disk *disk; 575 576 r = superblock_read_lock(md, &sblock); 577 if (r) { 578 DMERR("couldn't read_lock superblock"); 579 return r; 580 } 581 582 disk = dm_block_data(sblock); 583 584 /* Verify the data block size hasn't changed */ 585 if (le32_to_cpu(disk->data_block_size) != md->block_size) { 586 DMERR("changing the data block size (from %u to %llu) is not supported", 587 le32_to_cpu(disk->data_block_size), md->block_size); 588 r = -EINVAL; 589 goto bad; 590 } 591 592 r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION, 593 disk->metadata_space_map_root, 594 sizeof(disk->metadata_space_map_root), 595 &md->tm, &md->sm); 596 if (r) { 597 DMERR("dm_tm_open_with_sm failed"); 598 goto bad; 599 } 600 601 setup_infos(md); 602 603 md->nr_blocks = le32_to_cpu(disk->nr_blocks); 604 md->current_era = le32_to_cpu(disk->current_era); 605 606 ws_unpack(&disk->current_writeset, &md->current_writeset->md); 607 md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root); 608 md->era_array_root = le64_to_cpu(disk->era_array_root); 609 md->metadata_snap = le64_to_cpu(disk->metadata_snap); 610 md->archived_writesets = true; 611 612 dm_bm_unlock(sblock); 613 614 return 0; 615 616 bad: 617 dm_bm_unlock(sblock); 618 return r; 619 } 620 621 static int open_or_format_metadata(struct era_metadata *md, 622 bool may_format) 623 { 624 int r; 625 bool unformatted = false; 626 627 r = superblock_all_zeroes(md->bm, &unformatted); 628 if (r) 629 return r; 630 631 if (unformatted) 632 return may_format ? format_metadata(md) : -EPERM; 633 634 return open_metadata(md); 635 } 636 637 static int create_persistent_data_objects(struct era_metadata *md, 638 bool may_format) 639 { 640 int r; 641 642 md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE, 643 ERA_MAX_CONCURRENT_LOCKS); 644 if (IS_ERR(md->bm)) { 645 DMERR("could not create block manager"); 646 return PTR_ERR(md->bm); 647 } 648 649 r = open_or_format_metadata(md, may_format); 650 if (r) 651 dm_block_manager_destroy(md->bm); 652 653 return r; 654 } 655 656 static void destroy_persistent_data_objects(struct era_metadata *md) 657 { 658 dm_sm_destroy(md->sm); 659 dm_tm_destroy(md->tm); 660 dm_block_manager_destroy(md->bm); 661 } 662 663 /* 664 * This waits until all era_map threads have picked up the new filter. 665 */ 666 static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset) 667 { 668 rcu_assign_pointer(md->current_writeset, new_writeset); 669 synchronize_rcu(); 670 } 671 672 /* 673 *------------------------------------------------------------------------ 674 * Writesets get 'digested' into the main era array. 675 * 676 * We're using a coroutine here so the worker thread can do the digestion, 677 * thus avoiding synchronisation of the metadata. Digesting a whole 678 * writeset in one go would cause too much latency. 679 *------------------------------------------------------------------------ 680 */ 681 struct digest { 682 uint32_t era; 683 unsigned int nr_bits, current_bit; 684 struct writeset_metadata writeset; 685 __le32 value; 686 struct dm_disk_bitset info; 687 688 int (*step)(struct era_metadata *md, struct digest *d); 689 }; 690 691 static int metadata_digest_lookup_writeset(struct era_metadata *md, 692 struct digest *d); 693 694 static int metadata_digest_remove_writeset(struct era_metadata *md, 695 struct digest *d) 696 { 697 int r; 698 uint64_t key = d->era; 699 700 r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root, 701 &key, &md->writeset_tree_root); 702 if (r) { 703 DMERR("%s: dm_btree_remove failed", __func__); 704 return r; 705 } 706 707 d->step = metadata_digest_lookup_writeset; 708 return 0; 709 } 710 711 #define INSERTS_PER_STEP 100 712 713 static int metadata_digest_transcribe_writeset(struct era_metadata *md, 714 struct digest *d) 715 { 716 int r; 717 bool marked; 718 unsigned int b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits); 719 720 for (b = d->current_bit; b < e; b++) { 721 r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked); 722 if (r) { 723 DMERR("%s: writeset_marked_on_disk failed", __func__); 724 return r; 725 } 726 727 if (!marked) 728 continue; 729 730 __dm_bless_for_disk(&d->value); 731 r = dm_array_set_value(&md->era_array_info, md->era_array_root, 732 b, &d->value, &md->era_array_root); 733 if (r) { 734 DMERR("%s: dm_array_set_value failed", __func__); 735 return r; 736 } 737 } 738 739 if (b == d->nr_bits) 740 d->step = metadata_digest_remove_writeset; 741 else 742 d->current_bit = b; 743 744 return 0; 745 } 746 747 static int metadata_digest_lookup_writeset(struct era_metadata *md, 748 struct digest *d) 749 { 750 int r; 751 uint64_t key; 752 struct writeset_disk disk; 753 754 r = dm_btree_find_lowest_key(&md->writeset_tree_info, 755 md->writeset_tree_root, &key); 756 if (r < 0) 757 return r; 758 759 d->era = key; 760 761 r = dm_btree_lookup(&md->writeset_tree_info, 762 md->writeset_tree_root, &key, &disk); 763 if (r) { 764 if (r == -ENODATA) { 765 d->step = NULL; 766 return 0; 767 } 768 769 DMERR("%s: dm_btree_lookup failed", __func__); 770 return r; 771 } 772 773 ws_unpack(&disk, &d->writeset); 774 d->value = cpu_to_le32(key); 775 776 /* 777 * We initialise another bitset info to avoid any caching side effects 778 * with the previous one. 779 */ 780 dm_disk_bitset_init(md->tm, &d->info); 781 782 d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks); 783 d->current_bit = 0; 784 d->step = metadata_digest_transcribe_writeset; 785 786 return 0; 787 } 788 789 static int metadata_digest_start(struct era_metadata *md, struct digest *d) 790 { 791 if (d->step) 792 return 0; 793 794 memset(d, 0, sizeof(*d)); 795 d->step = metadata_digest_lookup_writeset; 796 797 return 0; 798 } 799 800 /* 801 *----------------------------------------------------------------- 802 * High level metadata interface. Target methods should use these, 803 * and not the lower level ones. 804 *----------------------------------------------------------------- 805 */ 806 static struct era_metadata *metadata_open(struct block_device *bdev, 807 sector_t block_size, 808 bool may_format) 809 { 810 int r; 811 struct era_metadata *md = kzalloc_obj(*md); 812 813 if (!md) { 814 DMERR("could not allocate metadata struct"); 815 return ERR_PTR(-ENOMEM); 816 } 817 818 md->bdev = bdev; 819 md->block_size = block_size; 820 821 md->writesets[0].md.root = INVALID_WRITESET_ROOT; 822 md->writesets[1].md.root = INVALID_WRITESET_ROOT; 823 md->current_writeset = &md->writesets[0]; 824 825 r = create_persistent_data_objects(md, may_format); 826 if (r) { 827 kfree(md); 828 return ERR_PTR(r); 829 } 830 831 return md; 832 } 833 834 static void metadata_close(struct era_metadata *md) 835 { 836 writeset_free(&md->writesets[0]); 837 writeset_free(&md->writesets[1]); 838 destroy_persistent_data_objects(md); 839 kfree(md); 840 } 841 842 static bool valid_nr_blocks(dm_block_t n) 843 { 844 /* 845 * dm_bitset restricts us to 2^32. test_bit & co. restrict us 846 * further to 2^31 - 1 847 */ 848 return n < (1ull << 31); 849 } 850 851 static int metadata_resize(struct era_metadata *md, void *arg) 852 { 853 int r; 854 dm_block_t *new_size = arg; 855 __le32 value; 856 857 if (!valid_nr_blocks(*new_size)) { 858 DMERR("Invalid number of origin blocks %llu", 859 (unsigned long long) *new_size); 860 return -EINVAL; 861 } 862 863 writeset_free(&md->writesets[0]); 864 writeset_free(&md->writesets[1]); 865 866 r = writeset_alloc(&md->writesets[0], *new_size); 867 if (r) { 868 DMERR("%s: writeset_alloc failed for writeset 0", __func__); 869 return r; 870 } 871 872 r = writeset_alloc(&md->writesets[1], *new_size); 873 if (r) { 874 DMERR("%s: writeset_alloc failed for writeset 1", __func__); 875 writeset_free(&md->writesets[0]); 876 return r; 877 } 878 879 value = cpu_to_le32(0u); 880 __dm_bless_for_disk(&value); 881 r = dm_array_resize(&md->era_array_info, md->era_array_root, 882 md->nr_blocks, *new_size, 883 &value, &md->era_array_root); 884 if (r) { 885 DMERR("%s: dm_array_resize failed", __func__); 886 writeset_free(&md->writesets[0]); 887 writeset_free(&md->writesets[1]); 888 return r; 889 } 890 891 md->nr_blocks = *new_size; 892 return 0; 893 } 894 895 static int metadata_era_archive(struct era_metadata *md) 896 { 897 int r; 898 uint64_t keys[1]; 899 struct writeset_disk value; 900 901 r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root, 902 &md->current_writeset->md.root); 903 if (r) { 904 DMERR("%s: dm_bitset_flush failed", __func__); 905 return r; 906 } 907 908 ws_pack(&md->current_writeset->md, &value); 909 910 keys[0] = md->current_era; 911 __dm_bless_for_disk(&value); 912 r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root, 913 keys, &value, &md->writeset_tree_root); 914 if (r) { 915 DMERR("%s: couldn't insert writeset into btree", __func__); 916 /* FIXME: fail mode */ 917 return r; 918 } 919 920 md->current_writeset->md.root = INVALID_WRITESET_ROOT; 921 md->archived_writesets = true; 922 923 return 0; 924 } 925 926 static struct writeset *next_writeset(struct era_metadata *md) 927 { 928 return (md->current_writeset == &md->writesets[0]) ? 929 &md->writesets[1] : &md->writesets[0]; 930 } 931 932 static int metadata_new_era(struct era_metadata *md) 933 { 934 int r; 935 struct writeset *new_writeset = next_writeset(md); 936 937 r = writeset_init(&md->bitset_info, new_writeset, md->nr_blocks); 938 if (r) { 939 DMERR("%s: writeset_init failed", __func__); 940 return r; 941 } 942 943 swap_writeset(md, new_writeset); 944 md->current_era++; 945 946 return 0; 947 } 948 949 static int metadata_era_rollover(struct era_metadata *md) 950 { 951 int r; 952 953 if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) { 954 r = metadata_era_archive(md); 955 if (r) { 956 DMERR("%s: metadata_archive_era failed", __func__); 957 /* FIXME: fail mode? */ 958 return r; 959 } 960 } 961 962 r = metadata_new_era(md); 963 if (r) { 964 DMERR("%s: new era failed", __func__); 965 /* FIXME: fail mode */ 966 return r; 967 } 968 969 return 0; 970 } 971 972 static bool metadata_current_marked(struct era_metadata *md, dm_block_t block) 973 { 974 bool r; 975 struct writeset *ws; 976 977 rcu_read_lock(); 978 ws = rcu_dereference(md->current_writeset); 979 r = writeset_marked(ws, block); 980 rcu_read_unlock(); 981 982 return r; 983 } 984 985 static int metadata_commit(struct era_metadata *md) 986 { 987 int r; 988 struct dm_block *sblock; 989 990 if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) { 991 r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root, 992 &md->current_writeset->md.root); 993 if (r) { 994 DMERR("%s: bitset flush failed", __func__); 995 return r; 996 } 997 } 998 999 r = dm_tm_pre_commit(md->tm); 1000 if (r) { 1001 DMERR("%s: pre commit failed", __func__); 1002 return r; 1003 } 1004 1005 r = save_sm_root(md); 1006 if (r) { 1007 DMERR("%s: save_sm_root failed", __func__); 1008 return r; 1009 } 1010 1011 r = superblock_lock(md, &sblock); 1012 if (r) { 1013 DMERR("%s: superblock lock failed", __func__); 1014 return r; 1015 } 1016 1017 prepare_superblock(md, dm_block_data(sblock)); 1018 1019 return dm_tm_commit(md->tm, sblock); 1020 } 1021 1022 static int metadata_checkpoint(struct era_metadata *md) 1023 { 1024 /* 1025 * For now we just rollover, but later I want to put a check in to 1026 * avoid this if the filter is still pretty fresh. 1027 */ 1028 return metadata_era_rollover(md); 1029 } 1030 1031 /* 1032 * Metadata snapshots allow userland to access era data. 1033 */ 1034 static int metadata_take_snap(struct era_metadata *md) 1035 { 1036 int r, inc; 1037 dm_block_t location; 1038 struct dm_block *clone; 1039 1040 if (md->metadata_snap != SUPERBLOCK_LOCATION) { 1041 DMERR("%s: metadata snapshot already exists", __func__); 1042 return -EINVAL; 1043 } 1044 1045 r = metadata_era_rollover(md); 1046 if (r) { 1047 DMERR("%s: era rollover failed", __func__); 1048 return r; 1049 } 1050 1051 r = metadata_commit(md); 1052 if (r) { 1053 DMERR("%s: pre commit failed", __func__); 1054 return r; 1055 } 1056 1057 r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION); 1058 if (r) { 1059 DMERR("%s: couldn't increment superblock", __func__); 1060 return r; 1061 } 1062 1063 r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION, 1064 &sb_validator, &clone, &inc); 1065 if (r) { 1066 DMERR("%s: couldn't shadow superblock", __func__); 1067 dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION); 1068 return r; 1069 } 1070 BUG_ON(!inc); 1071 1072 r = dm_sm_inc_block(md->sm, md->writeset_tree_root); 1073 if (r) { 1074 DMERR("%s: couldn't inc writeset tree root", __func__); 1075 location = dm_block_location(clone); 1076 dm_tm_unlock(md->tm, clone); 1077 dm_sm_dec_block(md->sm, location); 1078 return r; 1079 } 1080 1081 r = dm_sm_inc_block(md->sm, md->era_array_root); 1082 if (r) { 1083 DMERR("%s: couldn't inc era tree root", __func__); 1084 dm_sm_dec_block(md->sm, md->writeset_tree_root); 1085 location = dm_block_location(clone); 1086 dm_tm_unlock(md->tm, clone); 1087 dm_sm_dec_block(md->sm, location); 1088 return r; 1089 } 1090 1091 md->metadata_snap = dm_block_location(clone); 1092 1093 dm_tm_unlock(md->tm, clone); 1094 1095 return 0; 1096 } 1097 1098 static int metadata_drop_snap(struct era_metadata *md) 1099 { 1100 int r; 1101 dm_block_t location; 1102 struct dm_block *clone; 1103 struct superblock_disk *disk; 1104 1105 if (md->metadata_snap == SUPERBLOCK_LOCATION) { 1106 DMERR("%s: no snap to drop", __func__); 1107 return -EINVAL; 1108 } 1109 1110 r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone); 1111 if (r) { 1112 DMERR("%s: couldn't read lock superblock clone", __func__); 1113 return r; 1114 } 1115 1116 /* 1117 * Whatever happens now we'll commit with no record of the metadata 1118 * snap. 1119 */ 1120 md->metadata_snap = SUPERBLOCK_LOCATION; 1121 1122 disk = dm_block_data(clone); 1123 r = dm_btree_del(&md->writeset_tree_info, 1124 le64_to_cpu(disk->writeset_tree_root)); 1125 if (r) { 1126 DMERR("%s: error deleting writeset tree clone", __func__); 1127 dm_tm_unlock(md->tm, clone); 1128 return r; 1129 } 1130 1131 r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root)); 1132 if (r) { 1133 DMERR("%s: error deleting era array clone", __func__); 1134 dm_tm_unlock(md->tm, clone); 1135 return r; 1136 } 1137 1138 location = dm_block_location(clone); 1139 dm_tm_unlock(md->tm, clone); 1140 1141 return dm_sm_dec_block(md->sm, location); 1142 } 1143 1144 struct metadata_stats { 1145 dm_block_t used; 1146 dm_block_t total; 1147 dm_block_t snap; 1148 uint32_t era; 1149 }; 1150 1151 static int metadata_get_stats(struct era_metadata *md, void *ptr) 1152 { 1153 int r; 1154 struct metadata_stats *s = ptr; 1155 dm_block_t nr_free, nr_total; 1156 1157 r = dm_sm_get_nr_free(md->sm, &nr_free); 1158 if (r) { 1159 DMERR("dm_sm_get_nr_free returned %d", r); 1160 return r; 1161 } 1162 1163 r = dm_sm_get_nr_blocks(md->sm, &nr_total); 1164 if (r) { 1165 DMERR("dm_pool_get_metadata_dev_size returned %d", r); 1166 return r; 1167 } 1168 1169 s->used = nr_total - nr_free; 1170 s->total = nr_total; 1171 s->snap = md->metadata_snap; 1172 s->era = md->current_era; 1173 1174 return 0; 1175 } 1176 1177 /*----------------------------------------------------------------*/ 1178 1179 struct era { 1180 struct dm_target *ti; 1181 1182 struct dm_dev *metadata_dev; 1183 struct dm_dev *origin_dev; 1184 1185 dm_block_t nr_blocks; 1186 uint32_t sectors_per_block; 1187 int sectors_per_block_shift; 1188 struct era_metadata *md; 1189 1190 struct workqueue_struct *wq; 1191 struct work_struct worker; 1192 1193 spinlock_t deferred_lock; 1194 struct bio_list deferred_bios; 1195 1196 spinlock_t rpc_lock; 1197 struct list_head rpc_calls; 1198 1199 struct digest digest; 1200 atomic_t suspended; 1201 }; 1202 1203 struct rpc { 1204 struct list_head list; 1205 1206 int (*fn0)(struct era_metadata *md); 1207 int (*fn1)(struct era_metadata *md, void *ref); 1208 void *arg; 1209 int result; 1210 1211 struct completion complete; 1212 }; 1213 1214 /* 1215 *--------------------------------------------------------------- 1216 * Remapping. 1217 *--------------------------------------------------------------- 1218 */ 1219 static bool block_size_is_power_of_two(struct era *era) 1220 { 1221 return era->sectors_per_block_shift >= 0; 1222 } 1223 1224 static dm_block_t get_block(struct era *era, struct bio *bio) 1225 { 1226 sector_t block_nr = bio->bi_iter.bi_sector; 1227 1228 if (!block_size_is_power_of_two(era)) 1229 (void) sector_div(block_nr, era->sectors_per_block); 1230 else 1231 block_nr >>= era->sectors_per_block_shift; 1232 1233 return block_nr; 1234 } 1235 1236 static void remap_to_origin(struct era *era, struct bio *bio) 1237 { 1238 bio_set_dev(bio, era->origin_dev->bdev); 1239 bio->bi_iter.bi_sector = dm_target_offset(era->ti, bio->bi_iter.bi_sector); 1240 } 1241 1242 /* 1243 *-------------------------------------------------------------- 1244 * Worker thread 1245 *-------------------------------------------------------------- 1246 */ 1247 static void wake_worker(struct era *era) 1248 { 1249 if (!atomic_read(&era->suspended)) 1250 queue_work(era->wq, &era->worker); 1251 } 1252 1253 static void process_old_eras(struct era *era) 1254 { 1255 int r; 1256 1257 if (!era->digest.step) 1258 return; 1259 1260 r = era->digest.step(era->md, &era->digest); 1261 if (r < 0) { 1262 DMERR("%s: digest step failed, stopping digestion", __func__); 1263 era->digest.step = NULL; 1264 1265 } else if (era->digest.step) 1266 wake_worker(era); 1267 } 1268 1269 static void process_deferred_bios(struct era *era) 1270 { 1271 int r; 1272 struct bio_list deferred_bios, marked_bios; 1273 struct bio *bio; 1274 struct blk_plug plug; 1275 bool commit_needed = false; 1276 bool failed = false; 1277 struct writeset *ws = era->md->current_writeset; 1278 1279 bio_list_init(&deferred_bios); 1280 bio_list_init(&marked_bios); 1281 1282 spin_lock(&era->deferred_lock); 1283 bio_list_merge_init(&deferred_bios, &era->deferred_bios); 1284 spin_unlock(&era->deferred_lock); 1285 1286 if (bio_list_empty(&deferred_bios)) 1287 return; 1288 1289 while ((bio = bio_list_pop(&deferred_bios))) { 1290 r = writeset_test_and_set(&era->md->bitset_info, ws, 1291 get_block(era, bio)); 1292 if (r < 0) { 1293 /* 1294 * This is bad news, we need to rollback. 1295 * FIXME: finish. 1296 */ 1297 failed = true; 1298 } else if (r == 0) 1299 commit_needed = true; 1300 1301 bio_list_add(&marked_bios, bio); 1302 } 1303 1304 if (commit_needed) { 1305 r = metadata_commit(era->md); 1306 if (r) 1307 failed = true; 1308 } 1309 1310 if (failed) 1311 while ((bio = bio_list_pop(&marked_bios))) 1312 bio_io_error(bio); 1313 else { 1314 blk_start_plug(&plug); 1315 while ((bio = bio_list_pop(&marked_bios))) { 1316 /* 1317 * Only update the in-core writeset if the on-disk one 1318 * was updated too. 1319 */ 1320 if (commit_needed) 1321 set_bit(get_block(era, bio), ws->bits); 1322 submit_bio_noacct(bio); 1323 } 1324 blk_finish_plug(&plug); 1325 } 1326 } 1327 1328 static void process_rpc_calls(struct era *era) 1329 { 1330 int r; 1331 bool need_commit = false; 1332 struct list_head calls; 1333 struct rpc *rpc, *tmp; 1334 1335 INIT_LIST_HEAD(&calls); 1336 spin_lock(&era->rpc_lock); 1337 list_splice_init(&era->rpc_calls, &calls); 1338 spin_unlock(&era->rpc_lock); 1339 1340 list_for_each_entry_safe(rpc, tmp, &calls, list) { 1341 rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg); 1342 need_commit = true; 1343 } 1344 1345 if (need_commit) { 1346 r = metadata_commit(era->md); 1347 if (r) 1348 list_for_each_entry_safe(rpc, tmp, &calls, list) 1349 rpc->result = r; 1350 } 1351 1352 list_for_each_entry_safe(rpc, tmp, &calls, list) 1353 complete(&rpc->complete); 1354 } 1355 1356 static void kick_off_digest(struct era *era) 1357 { 1358 if (era->md->archived_writesets) { 1359 era->md->archived_writesets = false; 1360 metadata_digest_start(era->md, &era->digest); 1361 } 1362 } 1363 1364 static void do_work(struct work_struct *ws) 1365 { 1366 struct era *era = container_of(ws, struct era, worker); 1367 1368 kick_off_digest(era); 1369 process_old_eras(era); 1370 process_deferred_bios(era); 1371 process_rpc_calls(era); 1372 } 1373 1374 static void defer_bio(struct era *era, struct bio *bio) 1375 { 1376 spin_lock(&era->deferred_lock); 1377 bio_list_add(&era->deferred_bios, bio); 1378 spin_unlock(&era->deferred_lock); 1379 1380 wake_worker(era); 1381 } 1382 1383 /* 1384 * Make an rpc call to the worker to change the metadata. 1385 */ 1386 static int perform_rpc(struct era *era, struct rpc *rpc) 1387 { 1388 rpc->result = 0; 1389 init_completion(&rpc->complete); 1390 1391 spin_lock(&era->rpc_lock); 1392 list_add(&rpc->list, &era->rpc_calls); 1393 spin_unlock(&era->rpc_lock); 1394 1395 wake_worker(era); 1396 wait_for_completion(&rpc->complete); 1397 1398 return rpc->result; 1399 } 1400 1401 static int in_worker0(struct era *era, int (*fn)(struct era_metadata *md)) 1402 { 1403 struct rpc rpc; 1404 1405 rpc.fn0 = fn; 1406 rpc.fn1 = NULL; 1407 1408 return perform_rpc(era, &rpc); 1409 } 1410 1411 static int in_worker1(struct era *era, 1412 int (*fn)(struct era_metadata *md, void *ref), void *arg) 1413 { 1414 struct rpc rpc; 1415 1416 rpc.fn0 = NULL; 1417 rpc.fn1 = fn; 1418 rpc.arg = arg; 1419 1420 return perform_rpc(era, &rpc); 1421 } 1422 1423 static void start_worker(struct era *era) 1424 { 1425 atomic_set(&era->suspended, 0); 1426 } 1427 1428 static void stop_worker(struct era *era) 1429 { 1430 atomic_set(&era->suspended, 1); 1431 drain_workqueue(era->wq); 1432 } 1433 1434 /* 1435 *-------------------------------------------------------------- 1436 * Target methods 1437 *-------------------------------------------------------------- 1438 */ 1439 static void era_destroy(struct era *era) 1440 { 1441 if (era->md) 1442 metadata_close(era->md); 1443 1444 if (era->wq) 1445 destroy_workqueue(era->wq); 1446 1447 if (era->origin_dev) 1448 dm_put_device(era->ti, era->origin_dev); 1449 1450 if (era->metadata_dev) 1451 dm_put_device(era->ti, era->metadata_dev); 1452 1453 kfree(era); 1454 } 1455 1456 static dm_block_t calc_nr_blocks(struct era *era) 1457 { 1458 return dm_sector_div_up(era->ti->len, era->sectors_per_block); 1459 } 1460 1461 static bool valid_block_size(dm_block_t block_size) 1462 { 1463 bool greater_than_zero = block_size > 0; 1464 bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0; 1465 1466 return greater_than_zero && multiple_of_min_block_size; 1467 } 1468 1469 /* 1470 * <metadata dev> <data dev> <data block size (sectors)> 1471 */ 1472 static int era_ctr(struct dm_target *ti, unsigned int argc, char **argv) 1473 { 1474 int r; 1475 char dummy; 1476 struct era *era; 1477 struct era_metadata *md; 1478 1479 if (argc != 3) { 1480 ti->error = "Invalid argument count"; 1481 return -EINVAL; 1482 } 1483 1484 era = kzalloc_obj(*era); 1485 if (!era) { 1486 ti->error = "Error allocating era structure"; 1487 return -ENOMEM; 1488 } 1489 1490 era->ti = ti; 1491 1492 r = dm_get_device(ti, argv[0], BLK_OPEN_READ | BLK_OPEN_WRITE, 1493 &era->metadata_dev); 1494 if (r) { 1495 ti->error = "Error opening metadata device"; 1496 era_destroy(era); 1497 return r; 1498 } 1499 1500 r = dm_get_device(ti, argv[1], BLK_OPEN_READ | BLK_OPEN_WRITE, 1501 &era->origin_dev); 1502 if (r) { 1503 ti->error = "Error opening data device"; 1504 era_destroy(era); 1505 return r; 1506 } 1507 1508 r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy); 1509 if (r != 1) { 1510 ti->error = "Error parsing block size"; 1511 era_destroy(era); 1512 return -EINVAL; 1513 } 1514 1515 r = dm_set_target_max_io_len(ti, era->sectors_per_block); 1516 if (r) { 1517 ti->error = "could not set max io len"; 1518 era_destroy(era); 1519 return r; 1520 } 1521 1522 if (!valid_block_size(era->sectors_per_block)) { 1523 ti->error = "Invalid block size"; 1524 era_destroy(era); 1525 return -EINVAL; 1526 } 1527 if (era->sectors_per_block & (era->sectors_per_block - 1)) 1528 era->sectors_per_block_shift = -1; 1529 else 1530 era->sectors_per_block_shift = __ffs(era->sectors_per_block); 1531 1532 md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true); 1533 if (IS_ERR(md)) { 1534 ti->error = "Error reading metadata"; 1535 era_destroy(era); 1536 return PTR_ERR(md); 1537 } 1538 era->md = md; 1539 1540 era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM); 1541 if (!era->wq) { 1542 ti->error = "could not create workqueue for metadata object"; 1543 era_destroy(era); 1544 return -ENOMEM; 1545 } 1546 INIT_WORK(&era->worker, do_work); 1547 1548 spin_lock_init(&era->deferred_lock); 1549 bio_list_init(&era->deferred_bios); 1550 1551 spin_lock_init(&era->rpc_lock); 1552 INIT_LIST_HEAD(&era->rpc_calls); 1553 1554 ti->private = era; 1555 ti->num_flush_bios = 1; 1556 ti->flush_supported = true; 1557 1558 ti->num_discard_bios = 1; 1559 1560 return 0; 1561 } 1562 1563 static void era_dtr(struct dm_target *ti) 1564 { 1565 era_destroy(ti->private); 1566 } 1567 1568 static int era_map(struct dm_target *ti, struct bio *bio) 1569 { 1570 struct era *era = ti->private; 1571 dm_block_t block; 1572 1573 /* 1574 * All bios get remapped to the origin device. We do this now, but 1575 * it may not get issued until later. Depending on whether the 1576 * block is marked in this era. 1577 */ 1578 remap_to_origin(era, bio); 1579 block = get_block(era, bio); 1580 1581 /* 1582 * REQ_PREFLUSH bios carry no data, so we're not interested in them. 1583 */ 1584 if (!(bio->bi_opf & REQ_PREFLUSH) && 1585 (bio_data_dir(bio) == WRITE) && 1586 !metadata_current_marked(era->md, block)) { 1587 defer_bio(era, bio); 1588 return DM_MAPIO_SUBMITTED; 1589 } 1590 1591 return DM_MAPIO_REMAPPED; 1592 } 1593 1594 static void era_postsuspend(struct dm_target *ti) 1595 { 1596 int r; 1597 struct era *era = ti->private; 1598 1599 r = in_worker0(era, metadata_era_archive); 1600 if (r) { 1601 DMERR("%s: couldn't archive current era", __func__); 1602 /* FIXME: fail mode */ 1603 } 1604 1605 stop_worker(era); 1606 1607 r = metadata_commit(era->md); 1608 if (r) { 1609 DMERR("%s: metadata_commit failed", __func__); 1610 /* FIXME: fail mode */ 1611 } 1612 } 1613 1614 static int era_preresume(struct dm_target *ti) 1615 { 1616 int r; 1617 struct era *era = ti->private; 1618 dm_block_t new_size = calc_nr_blocks(era); 1619 1620 if (era->nr_blocks != new_size) { 1621 r = metadata_resize(era->md, &new_size); 1622 if (r) { 1623 DMERR("%s: metadata_resize failed", __func__); 1624 return r; 1625 } 1626 1627 r = metadata_commit(era->md); 1628 if (r) { 1629 DMERR("%s: metadata_commit failed", __func__); 1630 return r; 1631 } 1632 1633 era->nr_blocks = new_size; 1634 } 1635 1636 start_worker(era); 1637 1638 r = in_worker0(era, metadata_era_rollover); 1639 if (r) { 1640 DMERR("%s: metadata_era_rollover failed", __func__); 1641 return r; 1642 } 1643 1644 return 0; 1645 } 1646 1647 /* 1648 * Status format: 1649 * 1650 * <metadata block size> <#used metadata blocks>/<#total metadata blocks> 1651 * <current era> <held metadata root | '-'> 1652 */ 1653 static void era_status(struct dm_target *ti, status_type_t type, 1654 unsigned int status_flags, char *result, unsigned int maxlen) 1655 { 1656 int r; 1657 struct era *era = ti->private; 1658 ssize_t sz = 0; 1659 struct metadata_stats stats; 1660 char buf[BDEVNAME_SIZE]; 1661 1662 switch (type) { 1663 case STATUSTYPE_INFO: 1664 r = in_worker1(era, metadata_get_stats, &stats); 1665 if (r) 1666 goto err; 1667 1668 DMEMIT("%u %llu/%llu %u", 1669 (unsigned int) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT), 1670 (unsigned long long) stats.used, 1671 (unsigned long long) stats.total, 1672 (unsigned int) stats.era); 1673 1674 if (stats.snap != SUPERBLOCK_LOCATION) 1675 DMEMIT(" %llu", stats.snap); 1676 else 1677 DMEMIT(" -"); 1678 break; 1679 1680 case STATUSTYPE_TABLE: 1681 format_dev_t(buf, era->metadata_dev->bdev->bd_dev); 1682 DMEMIT("%s ", buf); 1683 format_dev_t(buf, era->origin_dev->bdev->bd_dev); 1684 DMEMIT("%s %u", buf, era->sectors_per_block); 1685 break; 1686 1687 case STATUSTYPE_IMA: 1688 *result = '\0'; 1689 break; 1690 } 1691 1692 return; 1693 1694 err: 1695 DMEMIT("Error"); 1696 } 1697 1698 static int era_message(struct dm_target *ti, unsigned int argc, char **argv, 1699 char *result, unsigned int maxlen) 1700 { 1701 struct era *era = ti->private; 1702 1703 if (argc != 1) { 1704 DMERR("incorrect number of message arguments"); 1705 return -EINVAL; 1706 } 1707 1708 if (!strcasecmp(argv[0], "checkpoint")) 1709 return in_worker0(era, metadata_checkpoint); 1710 1711 if (!strcasecmp(argv[0], "take_metadata_snap")) 1712 return in_worker0(era, metadata_take_snap); 1713 1714 if (!strcasecmp(argv[0], "drop_metadata_snap")) 1715 return in_worker0(era, metadata_drop_snap); 1716 1717 DMERR("unsupported message '%s'", argv[0]); 1718 return -EINVAL; 1719 } 1720 1721 static sector_t get_dev_size(struct dm_dev *dev) 1722 { 1723 return bdev_nr_sectors(dev->bdev); 1724 } 1725 1726 static int era_iterate_devices(struct dm_target *ti, 1727 iterate_devices_callout_fn fn, void *data) 1728 { 1729 struct era *era = ti->private; 1730 1731 return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data); 1732 } 1733 1734 static void era_io_hints(struct dm_target *ti, struct queue_limits *limits) 1735 { 1736 struct era *era = ti->private; 1737 uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT; 1738 1739 /* 1740 * If the system-determined stacked limits are compatible with the 1741 * era device's blocksize (io_opt is a factor) do not override them. 1742 */ 1743 if (io_opt_sectors < era->sectors_per_block || 1744 do_div(io_opt_sectors, era->sectors_per_block)) { 1745 limits->io_min = 0; 1746 limits->io_opt = era->sectors_per_block << SECTOR_SHIFT; 1747 } 1748 } 1749 1750 /*----------------------------------------------------------------*/ 1751 1752 static struct target_type era_target = { 1753 .name = "era", 1754 .version = {1, 0, 0}, 1755 .module = THIS_MODULE, 1756 .ctr = era_ctr, 1757 .dtr = era_dtr, 1758 .map = era_map, 1759 .postsuspend = era_postsuspend, 1760 .preresume = era_preresume, 1761 .status = era_status, 1762 .message = era_message, 1763 .iterate_devices = era_iterate_devices, 1764 .io_hints = era_io_hints 1765 }; 1766 module_dm(era); 1767 1768 MODULE_DESCRIPTION(DM_NAME " era target"); 1769 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>"); 1770 MODULE_LICENSE("GPL"); 1771