1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2011-2012 Red Hat, Inc. 4 * 5 * This file is released under the GPL. 6 */ 7 8 #include "dm-thin-metadata.h" 9 #include "persistent-data/dm-btree.h" 10 #include "persistent-data/dm-space-map.h" 11 #include "persistent-data/dm-space-map-disk.h" 12 #include "persistent-data/dm-transaction-manager.h" 13 14 #include <linux/list.h> 15 #include <linux/device-mapper.h> 16 #include <linux/workqueue.h> 17 18 /* 19 *-------------------------------------------------------------------------- 20 * As far as the metadata goes, there is: 21 * 22 * - A superblock in block zero, taking up fewer than 512 bytes for 23 * atomic writes. 24 * 25 * - A space map managing the metadata blocks. 26 * 27 * - A space map managing the data blocks. 28 * 29 * - A btree mapping our internal thin dev ids onto struct disk_device_details. 30 * 31 * - A hierarchical btree, with 2 levels which effectively maps (thin 32 * dev id, virtual block) -> block_time. Block time is a 64-bit 33 * field holding the time in the low 24 bits, and block in the top 40 34 * bits. 35 * 36 * BTrees consist solely of btree_nodes, that fill a block. Some are 37 * internal nodes, as such their values are a __le64 pointing to other 38 * nodes. Leaf nodes can store data of any reasonable size (ie. much 39 * smaller than the block size). The nodes consist of the header, 40 * followed by an array of keys, followed by an array of values. We have 41 * to binary search on the keys so they're all held together to help the 42 * cpu cache. 43 * 44 * Space maps have 2 btrees: 45 * 46 * - One maps a uint64_t onto a struct index_entry. Which points to a 47 * bitmap block, and has some details about how many free entries there 48 * are etc. 49 * 50 * - The bitmap blocks have a header (for the checksum). Then the rest 51 * of the block is pairs of bits. With the meaning being: 52 * 53 * 0 - ref count is 0 54 * 1 - ref count is 1 55 * 2 - ref count is 2 56 * 3 - ref count is higher than 2 57 * 58 * - If the count is higher than 2 then the ref count is entered in a 59 * second btree that directly maps the block_address to a uint32_t ref 60 * count. 61 * 62 * The space map metadata variant doesn't have a bitmaps btree. Instead 63 * it has one single blocks worth of index_entries. This avoids 64 * recursive issues with the bitmap btree needing to allocate space in 65 * order to insert. With a small data block size such as 64k the 66 * metadata support data devices that are hundreds of terrabytes. 67 * 68 * The space maps allocate space linearly from front to back. Space that 69 * is freed in a transaction is never recycled within that transaction. 70 * To try and avoid fragmenting _free_ space the allocator always goes 71 * back and fills in gaps. 72 * 73 * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks 74 * from the block manager. 75 *-------------------------------------------------------------------------- 76 */ 77 78 #define DM_MSG_PREFIX "thin metadata" 79 80 #define THIN_SUPERBLOCK_MAGIC 27022010 81 #define THIN_SUPERBLOCK_LOCATION 0 82 #define THIN_VERSION 2 83 #define SECTOR_TO_BLOCK_SHIFT 3 84 85 /* 86 * For btree insert: 87 * 3 for btree insert + 88 * 2 for btree lookup used within space map 89 * For btree remove: 90 * 2 for shadow spine + 91 * 4 for rebalance 3 child node 92 */ 93 #define THIN_MAX_CONCURRENT_LOCKS 6 94 95 /* This should be plenty */ 96 #define SPACE_MAP_ROOT_SIZE 128 97 98 /* 99 * Little endian on-disk superblock and device details. 100 */ 101 struct thin_disk_superblock { 102 __le32 csum; /* Checksum of superblock except for this field. */ 103 __le32 flags; 104 __le64 blocknr; /* This block number, dm_block_t. */ 105 106 __u8 uuid[16]; 107 __le64 magic; 108 __le32 version; 109 __le32 time; 110 111 __le64 trans_id; 112 113 /* 114 * Root held by userspace transactions. 115 */ 116 __le64 held_root; 117 118 __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE]; 119 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE]; 120 121 /* 122 * 2-level btree mapping (dev_id, (dev block, time)) -> data block 123 */ 124 __le64 data_mapping_root; 125 126 /* 127 * Device detail root mapping dev_id -> device_details 128 */ 129 __le64 device_details_root; 130 131 __le32 data_block_size; /* In 512-byte sectors. */ 132 133 __le32 metadata_block_size; /* In 512-byte sectors. */ 134 __le64 metadata_nr_blocks; 135 136 __le32 compat_flags; 137 __le32 compat_ro_flags; 138 __le32 incompat_flags; 139 } __packed; 140 141 struct disk_device_details { 142 __le64 mapped_blocks; 143 __le64 transaction_id; /* When created. */ 144 __le32 creation_time; 145 __le32 snapshotted_time; 146 } __packed; 147 148 struct dm_pool_metadata { 149 struct hlist_node hash; 150 151 struct block_device *bdev; 152 struct dm_block_manager *bm; 153 struct dm_space_map *metadata_sm; 154 struct dm_space_map *data_sm; 155 struct dm_transaction_manager *tm; 156 struct dm_transaction_manager *nb_tm; 157 158 /* 159 * Two-level btree. 160 * First level holds thin_dev_t. 161 * Second level holds mappings. 162 */ 163 struct dm_btree_info info; 164 165 /* 166 * Non-blocking version of the above. 167 */ 168 struct dm_btree_info nb_info; 169 170 /* 171 * Just the top level for deleting whole devices. 172 */ 173 struct dm_btree_info tl_info; 174 175 /* 176 * Just the bottom level for creating new devices. 177 */ 178 struct dm_btree_info bl_info; 179 180 /* 181 * Describes the device details btree. 182 */ 183 struct dm_btree_info details_info; 184 185 struct rw_semaphore root_lock; 186 uint32_t time; 187 dm_block_t root; 188 dm_block_t details_root; 189 struct list_head thin_devices; 190 uint64_t trans_id; 191 unsigned long flags; 192 sector_t data_block_size; 193 194 /* 195 * Pre-commit callback. 196 * 197 * This allows the thin provisioning target to run a callback before 198 * the metadata are committed. 199 */ 200 dm_pool_pre_commit_fn pre_commit_fn; 201 void *pre_commit_context; 202 203 /* 204 * We reserve a section of the metadata for commit overhead. 205 * All reported space does *not* include this. 206 */ 207 dm_block_t metadata_reserve; 208 209 /* 210 * Set if a transaction has to be aborted but the attempt to roll back 211 * to the previous (good) transaction failed. The only pool metadata 212 * operation possible in this state is the closing of the device. 213 */ 214 bool fail_io:1; 215 216 /* 217 * Set once a thin-pool has been accessed through one of the interfaces 218 * that imply the pool is in-service (e.g. thin devices created/deleted, 219 * thin-pool message, metadata snapshots, etc). 220 */ 221 bool in_service:1; 222 223 /* 224 * Reading the space map roots can fail, so we read it into these 225 * buffers before the superblock is locked and updated. 226 */ 227 __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE]; 228 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE]; 229 }; 230 231 struct dm_thin_device { 232 struct list_head list; 233 struct dm_pool_metadata *pmd; 234 dm_thin_id id; 235 236 int open_count; 237 bool changed:1; 238 bool aborted_with_changes:1; 239 uint64_t mapped_blocks; 240 uint64_t transaction_id; 241 uint32_t creation_time; 242 uint32_t snapshotted_time; 243 }; 244 245 /* 246 *-------------------------------------------------------------- 247 * superblock validator 248 *-------------------------------------------------------------- 249 */ 250 #define SUPERBLOCK_CSUM_XOR 160774 251 252 static void sb_prepare_for_write(struct dm_block_validator *v, 253 struct dm_block *b, 254 size_t block_size) 255 { 256 struct thin_disk_superblock *disk_super = dm_block_data(b); 257 258 disk_super->blocknr = cpu_to_le64(dm_block_location(b)); 259 disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags, 260 block_size - sizeof(__le32), 261 SUPERBLOCK_CSUM_XOR)); 262 } 263 264 static int sb_check(struct dm_block_validator *v, 265 struct dm_block *b, 266 size_t block_size) 267 { 268 struct thin_disk_superblock *disk_super = dm_block_data(b); 269 __le32 csum_le; 270 271 if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) { 272 DMERR("%s failed: blocknr %llu: wanted %llu", 273 __func__, le64_to_cpu(disk_super->blocknr), 274 (unsigned long long)dm_block_location(b)); 275 return -ENOTBLK; 276 } 277 278 if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) { 279 DMERR("%s failed: magic %llu: wanted %llu", 280 __func__, le64_to_cpu(disk_super->magic), 281 (unsigned long long)THIN_SUPERBLOCK_MAGIC); 282 return -EILSEQ; 283 } 284 285 csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags, 286 block_size - sizeof(__le32), 287 SUPERBLOCK_CSUM_XOR)); 288 if (csum_le != disk_super->csum) { 289 DMERR("%s failed: csum %u: wanted %u", 290 __func__, le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum)); 291 return -EILSEQ; 292 } 293 294 return 0; 295 } 296 297 static struct dm_block_validator sb_validator = { 298 .name = "superblock", 299 .prepare_for_write = sb_prepare_for_write, 300 .check = sb_check 301 }; 302 303 /* 304 *-------------------------------------------------------------- 305 * Methods for the btree value types 306 *-------------------------------------------------------------- 307 */ 308 static uint64_t pack_block_time(dm_block_t b, uint32_t t) 309 { 310 return (b << 24) | t; 311 } 312 313 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t) 314 { 315 *b = v >> 24; 316 *t = v & ((1 << 24) - 1); 317 } 318 319 /* 320 * It's more efficient to call dm_sm_{inc,dec}_blocks as few times as 321 * possible. 'with_runs' reads contiguous runs of blocks, and calls the 322 * given sm function. 323 */ 324 typedef int (*run_fn)(struct dm_space_map *, dm_block_t, dm_block_t); 325 326 static void with_runs(struct dm_space_map *sm, const __le64 *value_le, unsigned int count, run_fn fn) 327 { 328 uint64_t b, begin, end; 329 uint32_t t; 330 bool in_run = false; 331 unsigned int i; 332 333 for (i = 0; i < count; i++, value_le++) { 334 /* We know value_le is 8 byte aligned */ 335 unpack_block_time(le64_to_cpu(*value_le), &b, &t); 336 337 if (in_run) { 338 if (b == end) { 339 end++; 340 } else { 341 fn(sm, begin, end); 342 begin = b; 343 end = b + 1; 344 } 345 } else { 346 in_run = true; 347 begin = b; 348 end = b + 1; 349 } 350 } 351 352 if (in_run) 353 fn(sm, begin, end); 354 } 355 356 static void data_block_inc(void *context, const void *value_le, unsigned int count) 357 { 358 with_runs((struct dm_space_map *) context, 359 (const __le64 *) value_le, count, dm_sm_inc_blocks); 360 } 361 362 static void data_block_dec(void *context, const void *value_le, unsigned int count) 363 { 364 with_runs((struct dm_space_map *) context, 365 (const __le64 *) value_le, count, dm_sm_dec_blocks); 366 } 367 368 static int data_block_equal(void *context, const void *value1_le, const void *value2_le) 369 { 370 __le64 v1_le, v2_le; 371 uint64_t b1, b2; 372 uint32_t t; 373 374 memcpy(&v1_le, value1_le, sizeof(v1_le)); 375 memcpy(&v2_le, value2_le, sizeof(v2_le)); 376 unpack_block_time(le64_to_cpu(v1_le), &b1, &t); 377 unpack_block_time(le64_to_cpu(v2_le), &b2, &t); 378 379 return b1 == b2; 380 } 381 382 static void subtree_inc(void *context, const void *value, unsigned int count) 383 { 384 struct dm_btree_info *info = context; 385 const __le64 *root_le = value; 386 unsigned int i; 387 388 for (i = 0; i < count; i++, root_le++) 389 dm_tm_inc(info->tm, le64_to_cpu(*root_le)); 390 } 391 392 static void subtree_dec(void *context, const void *value, unsigned int count) 393 { 394 struct dm_btree_info *info = context; 395 const __le64 *root_le = value; 396 unsigned int i; 397 398 for (i = 0; i < count; i++, root_le++) 399 if (dm_btree_del(info, le64_to_cpu(*root_le))) 400 DMERR("btree delete failed"); 401 } 402 403 static int subtree_equal(void *context, const void *value1_le, const void *value2_le) 404 { 405 __le64 v1_le, v2_le; 406 407 memcpy(&v1_le, value1_le, sizeof(v1_le)); 408 memcpy(&v2_le, value2_le, sizeof(v2_le)); 409 410 return v1_le == v2_le; 411 } 412 413 /*----------------------------------------------------------------*/ 414 415 /* 416 * Variant that is used for in-core only changes or code that 417 * shouldn't put the pool in service on its own (e.g. commit). 418 */ 419 static inline void pmd_write_lock_in_core(struct dm_pool_metadata *pmd) 420 __acquires(pmd->root_lock) 421 { 422 down_write(&pmd->root_lock); 423 } 424 425 static inline void pmd_write_lock(struct dm_pool_metadata *pmd) 426 { 427 pmd_write_lock_in_core(pmd); 428 if (unlikely(!pmd->in_service)) 429 pmd->in_service = true; 430 } 431 432 static inline void pmd_write_unlock(struct dm_pool_metadata *pmd) 433 __releases(pmd->root_lock) 434 { 435 up_write(&pmd->root_lock); 436 } 437 438 /*----------------------------------------------------------------*/ 439 440 static int superblock_lock_zero(struct dm_pool_metadata *pmd, 441 struct dm_block **sblock) 442 { 443 return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION, 444 &sb_validator, sblock); 445 } 446 447 static int superblock_lock(struct dm_pool_metadata *pmd, 448 struct dm_block **sblock) 449 { 450 return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION, 451 &sb_validator, sblock); 452 } 453 454 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result) 455 { 456 int r; 457 unsigned int i; 458 struct dm_block *b; 459 __le64 *data_le, zero = cpu_to_le64(0); 460 unsigned int block_size = dm_bm_block_size(bm) / sizeof(__le64); 461 462 /* 463 * We can't use a validator here - it may be all zeroes. 464 */ 465 r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b); 466 if (r) 467 return r; 468 469 data_le = dm_block_data(b); 470 *result = 1; 471 for (i = 0; i < block_size; i++) { 472 if (data_le[i] != zero) { 473 *result = 0; 474 break; 475 } 476 } 477 478 dm_bm_unlock(b); 479 480 return 0; 481 } 482 483 static void __setup_btree_details(struct dm_pool_metadata *pmd) 484 { 485 pmd->info.tm = pmd->tm; 486 pmd->info.levels = 2; 487 pmd->info.value_type.context = pmd->data_sm; 488 pmd->info.value_type.size = sizeof(__le64); 489 pmd->info.value_type.inc = data_block_inc; 490 pmd->info.value_type.dec = data_block_dec; 491 pmd->info.value_type.equal = data_block_equal; 492 493 memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info)); 494 pmd->nb_info.tm = pmd->nb_tm; 495 496 pmd->tl_info.tm = pmd->tm; 497 pmd->tl_info.levels = 1; 498 pmd->tl_info.value_type.context = &pmd->bl_info; 499 pmd->tl_info.value_type.size = sizeof(__le64); 500 pmd->tl_info.value_type.inc = subtree_inc; 501 pmd->tl_info.value_type.dec = subtree_dec; 502 pmd->tl_info.value_type.equal = subtree_equal; 503 504 pmd->bl_info.tm = pmd->tm; 505 pmd->bl_info.levels = 1; 506 pmd->bl_info.value_type.context = pmd->data_sm; 507 pmd->bl_info.value_type.size = sizeof(__le64); 508 pmd->bl_info.value_type.inc = data_block_inc; 509 pmd->bl_info.value_type.dec = data_block_dec; 510 pmd->bl_info.value_type.equal = data_block_equal; 511 512 pmd->details_info.tm = pmd->tm; 513 pmd->details_info.levels = 1; 514 pmd->details_info.value_type.context = NULL; 515 pmd->details_info.value_type.size = sizeof(struct disk_device_details); 516 pmd->details_info.value_type.inc = NULL; 517 pmd->details_info.value_type.dec = NULL; 518 pmd->details_info.value_type.equal = NULL; 519 } 520 521 static int save_sm_roots(struct dm_pool_metadata *pmd) 522 { 523 int r; 524 size_t len; 525 526 r = dm_sm_root_size(pmd->metadata_sm, &len); 527 if (r < 0) 528 return r; 529 530 r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len); 531 if (r < 0) 532 return r; 533 534 r = dm_sm_root_size(pmd->data_sm, &len); 535 if (r < 0) 536 return r; 537 538 return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len); 539 } 540 541 static void copy_sm_roots(struct dm_pool_metadata *pmd, 542 struct thin_disk_superblock *disk) 543 { 544 memcpy(&disk->metadata_space_map_root, 545 &pmd->metadata_space_map_root, 546 sizeof(pmd->metadata_space_map_root)); 547 548 memcpy(&disk->data_space_map_root, 549 &pmd->data_space_map_root, 550 sizeof(pmd->data_space_map_root)); 551 } 552 553 static int __write_initial_superblock(struct dm_pool_metadata *pmd) 554 { 555 int r; 556 struct dm_block *sblock; 557 struct thin_disk_superblock *disk_super; 558 sector_t bdev_size = bdev_nr_sectors(pmd->bdev); 559 560 if (bdev_size > THIN_METADATA_MAX_SECTORS) 561 bdev_size = THIN_METADATA_MAX_SECTORS; 562 563 r = dm_sm_commit(pmd->data_sm); 564 if (r < 0) 565 return r; 566 567 r = dm_tm_pre_commit(pmd->tm); 568 if (r < 0) 569 return r; 570 571 r = save_sm_roots(pmd); 572 if (r < 0) 573 return r; 574 575 r = superblock_lock_zero(pmd, &sblock); 576 if (r) 577 return r; 578 579 disk_super = dm_block_data(sblock); 580 disk_super->flags = 0; 581 memset(disk_super->uuid, 0, sizeof(disk_super->uuid)); 582 disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC); 583 disk_super->version = cpu_to_le32(THIN_VERSION); 584 disk_super->time = 0; 585 disk_super->trans_id = 0; 586 disk_super->held_root = 0; 587 588 copy_sm_roots(pmd, disk_super); 589 590 disk_super->data_mapping_root = cpu_to_le64(pmd->root); 591 disk_super->device_details_root = cpu_to_le64(pmd->details_root); 592 disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE); 593 disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT); 594 disk_super->data_block_size = cpu_to_le32(pmd->data_block_size); 595 596 return dm_tm_commit(pmd->tm, sblock); 597 } 598 599 static int __format_metadata(struct dm_pool_metadata *pmd) 600 { 601 int r; 602 603 r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION, 604 &pmd->tm, &pmd->metadata_sm); 605 if (r < 0) { 606 DMERR("tm_create_with_sm failed"); 607 return r; 608 } 609 610 pmd->data_sm = dm_sm_disk_create(pmd->tm, 0); 611 if (IS_ERR(pmd->data_sm)) { 612 DMERR("sm_disk_create failed"); 613 r = PTR_ERR(pmd->data_sm); 614 goto bad_cleanup_tm; 615 } 616 617 pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm); 618 if (!pmd->nb_tm) { 619 DMERR("could not create non-blocking clone tm"); 620 r = -ENOMEM; 621 goto bad_cleanup_data_sm; 622 } 623 624 __setup_btree_details(pmd); 625 626 r = dm_btree_empty(&pmd->info, &pmd->root); 627 if (r < 0) 628 goto bad_cleanup_nb_tm; 629 630 r = dm_btree_empty(&pmd->details_info, &pmd->details_root); 631 if (r < 0) { 632 DMERR("couldn't create devices root"); 633 goto bad_cleanup_nb_tm; 634 } 635 636 r = __write_initial_superblock(pmd); 637 if (r) 638 goto bad_cleanup_nb_tm; 639 640 return 0; 641 642 bad_cleanup_nb_tm: 643 dm_tm_destroy(pmd->nb_tm); 644 bad_cleanup_data_sm: 645 dm_sm_destroy(pmd->data_sm); 646 bad_cleanup_tm: 647 dm_tm_destroy(pmd->tm); 648 dm_sm_destroy(pmd->metadata_sm); 649 650 return r; 651 } 652 653 static int __check_incompat_features(struct thin_disk_superblock *disk_super, 654 struct dm_pool_metadata *pmd) 655 { 656 uint32_t features; 657 658 features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP; 659 if (features) { 660 DMERR("could not access metadata due to unsupported optional features (%lx).", 661 (unsigned long)features); 662 return -EINVAL; 663 } 664 665 /* 666 * Check for read-only metadata to skip the following RDWR checks. 667 */ 668 if (bdev_read_only(pmd->bdev)) 669 return 0; 670 671 features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP; 672 if (features) { 673 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).", 674 (unsigned long)features); 675 return -EINVAL; 676 } 677 678 return 0; 679 } 680 681 static int __open_metadata(struct dm_pool_metadata *pmd) 682 { 683 int r; 684 struct dm_block *sblock; 685 struct thin_disk_superblock *disk_super; 686 687 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION, 688 &sb_validator, &sblock); 689 if (r < 0) { 690 DMERR("couldn't read superblock"); 691 return r; 692 } 693 694 disk_super = dm_block_data(sblock); 695 696 /* Verify the data block size hasn't changed */ 697 if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) { 698 DMERR("changing the data block size (from %u to %llu) is not supported", 699 le32_to_cpu(disk_super->data_block_size), 700 (unsigned long long)pmd->data_block_size); 701 r = -EINVAL; 702 goto bad_unlock_sblock; 703 } 704 705 r = __check_incompat_features(disk_super, pmd); 706 if (r < 0) 707 goto bad_unlock_sblock; 708 709 r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION, 710 disk_super->metadata_space_map_root, 711 sizeof(disk_super->metadata_space_map_root), 712 &pmd->tm, &pmd->metadata_sm); 713 if (r < 0) { 714 DMERR("tm_open_with_sm failed"); 715 goto bad_unlock_sblock; 716 } 717 718 pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root, 719 sizeof(disk_super->data_space_map_root)); 720 if (IS_ERR(pmd->data_sm)) { 721 DMERR("sm_disk_open failed"); 722 r = PTR_ERR(pmd->data_sm); 723 goto bad_cleanup_tm; 724 } 725 726 pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm); 727 if (!pmd->nb_tm) { 728 DMERR("could not create non-blocking clone tm"); 729 r = -ENOMEM; 730 goto bad_cleanup_data_sm; 731 } 732 733 /* 734 * For pool metadata opening process, root setting is redundant 735 * because it will be set again in __begin_transaction(). But dm 736 * pool aborting process really needs to get last transaction's 737 * root to avoid accessing broken btree. 738 */ 739 pmd->root = le64_to_cpu(disk_super->data_mapping_root); 740 pmd->details_root = le64_to_cpu(disk_super->device_details_root); 741 742 __setup_btree_details(pmd); 743 dm_bm_unlock(sblock); 744 745 return 0; 746 747 bad_cleanup_data_sm: 748 dm_sm_destroy(pmd->data_sm); 749 bad_cleanup_tm: 750 dm_tm_destroy(pmd->tm); 751 dm_sm_destroy(pmd->metadata_sm); 752 bad_unlock_sblock: 753 dm_bm_unlock(sblock); 754 755 return r; 756 } 757 758 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device) 759 { 760 int r, unformatted; 761 762 r = __superblock_all_zeroes(pmd->bm, &unformatted); 763 if (r) 764 return r; 765 766 if (unformatted) 767 return format_device ? __format_metadata(pmd) : -EPERM; 768 769 return __open_metadata(pmd); 770 } 771 772 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device) 773 { 774 int r; 775 776 pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT, 777 THIN_MAX_CONCURRENT_LOCKS); 778 if (IS_ERR(pmd->bm)) { 779 DMERR("could not create block manager"); 780 r = PTR_ERR(pmd->bm); 781 pmd->bm = NULL; 782 return r; 783 } 784 785 r = __open_or_format_metadata(pmd, format_device); 786 if (r) { 787 dm_block_manager_destroy(pmd->bm); 788 pmd->bm = NULL; 789 } 790 791 return r; 792 } 793 794 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd, 795 bool destroy_bm) 796 { 797 dm_sm_destroy(pmd->data_sm); 798 dm_sm_destroy(pmd->metadata_sm); 799 dm_tm_destroy(pmd->nb_tm); 800 dm_tm_destroy(pmd->tm); 801 if (destroy_bm) 802 dm_block_manager_destroy(pmd->bm); 803 } 804 805 static int __begin_transaction(struct dm_pool_metadata *pmd) 806 { 807 int r; 808 struct thin_disk_superblock *disk_super; 809 struct dm_block *sblock; 810 811 /* 812 * We re-read the superblock every time. Shouldn't need to do this 813 * really. 814 */ 815 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION, 816 &sb_validator, &sblock); 817 if (r) 818 return r; 819 820 disk_super = dm_block_data(sblock); 821 pmd->time = le32_to_cpu(disk_super->time); 822 pmd->root = le64_to_cpu(disk_super->data_mapping_root); 823 pmd->details_root = le64_to_cpu(disk_super->device_details_root); 824 pmd->trans_id = le64_to_cpu(disk_super->trans_id); 825 pmd->flags = le32_to_cpu(disk_super->flags); 826 pmd->data_block_size = le32_to_cpu(disk_super->data_block_size); 827 828 dm_bm_unlock(sblock); 829 return 0; 830 } 831 832 static int __write_changed_details(struct dm_pool_metadata *pmd) 833 { 834 int r; 835 struct dm_thin_device *td, *tmp; 836 struct disk_device_details details; 837 uint64_t key; 838 839 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) { 840 if (!td->changed) 841 continue; 842 843 key = td->id; 844 845 details.mapped_blocks = cpu_to_le64(td->mapped_blocks); 846 details.transaction_id = cpu_to_le64(td->transaction_id); 847 details.creation_time = cpu_to_le32(td->creation_time); 848 details.snapshotted_time = cpu_to_le32(td->snapshotted_time); 849 __dm_bless_for_disk(&details); 850 851 r = dm_btree_insert(&pmd->details_info, pmd->details_root, 852 &key, &details, &pmd->details_root); 853 if (r) 854 return r; 855 856 if (td->open_count) 857 td->changed = false; 858 else { 859 list_del(&td->list); 860 kfree(td); 861 } 862 } 863 864 return 0; 865 } 866 867 static int __commit_transaction(struct dm_pool_metadata *pmd) 868 { 869 int r; 870 struct thin_disk_superblock *disk_super; 871 struct dm_block *sblock; 872 873 /* 874 * We need to know if the thin_disk_superblock exceeds a 512-byte sector. 875 */ 876 BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512); 877 BUG_ON(!rwsem_is_locked(&pmd->root_lock)); 878 879 if (unlikely(!pmd->in_service)) 880 return 0; 881 882 if (pmd->pre_commit_fn) { 883 r = pmd->pre_commit_fn(pmd->pre_commit_context); 884 if (r < 0) { 885 DMERR("pre-commit callback failed"); 886 return r; 887 } 888 } 889 890 r = __write_changed_details(pmd); 891 if (r < 0) 892 return r; 893 894 r = dm_sm_commit(pmd->data_sm); 895 if (r < 0) 896 return r; 897 898 r = dm_tm_pre_commit(pmd->tm); 899 if (r < 0) 900 return r; 901 902 r = save_sm_roots(pmd); 903 if (r < 0) 904 return r; 905 906 r = superblock_lock(pmd, &sblock); 907 if (r) 908 return r; 909 910 disk_super = dm_block_data(sblock); 911 disk_super->time = cpu_to_le32(pmd->time); 912 disk_super->data_mapping_root = cpu_to_le64(pmd->root); 913 disk_super->device_details_root = cpu_to_le64(pmd->details_root); 914 disk_super->trans_id = cpu_to_le64(pmd->trans_id); 915 disk_super->flags = cpu_to_le32(pmd->flags); 916 917 copy_sm_roots(pmd, disk_super); 918 919 return dm_tm_commit(pmd->tm, sblock); 920 } 921 922 static void __set_metadata_reserve(struct dm_pool_metadata *pmd) 923 { 924 int r; 925 dm_block_t total; 926 dm_block_t max_blocks = 4096; /* 16M */ 927 928 r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total); 929 if (r) { 930 DMERR("could not get size of metadata device"); 931 pmd->metadata_reserve = max_blocks; 932 } else 933 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10)); 934 } 935 936 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev, 937 sector_t data_block_size, 938 bool format_device) 939 { 940 int r; 941 struct dm_pool_metadata *pmd; 942 943 pmd = kmalloc(sizeof(*pmd), GFP_KERNEL); 944 if (!pmd) { 945 DMERR("could not allocate metadata struct"); 946 return ERR_PTR(-ENOMEM); 947 } 948 949 init_rwsem(&pmd->root_lock); 950 pmd->time = 0; 951 INIT_LIST_HEAD(&pmd->thin_devices); 952 pmd->fail_io = false; 953 pmd->in_service = false; 954 pmd->bdev = bdev; 955 pmd->data_block_size = data_block_size; 956 pmd->pre_commit_fn = NULL; 957 pmd->pre_commit_context = NULL; 958 959 r = __create_persistent_data_objects(pmd, format_device); 960 if (r) { 961 kfree(pmd); 962 return ERR_PTR(r); 963 } 964 965 r = __begin_transaction(pmd); 966 if (r < 0) { 967 if (dm_pool_metadata_close(pmd) < 0) 968 DMWARN("%s: dm_pool_metadata_close() failed.", __func__); 969 return ERR_PTR(r); 970 } 971 972 __set_metadata_reserve(pmd); 973 974 return pmd; 975 } 976 977 int dm_pool_metadata_close(struct dm_pool_metadata *pmd) 978 { 979 int r; 980 unsigned int open_devices = 0; 981 struct dm_thin_device *td, *tmp; 982 983 down_read(&pmd->root_lock); 984 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) { 985 if (td->open_count) 986 open_devices++; 987 else { 988 list_del(&td->list); 989 kfree(td); 990 } 991 } 992 up_read(&pmd->root_lock); 993 994 if (open_devices) { 995 DMERR("attempt to close pmd when %u device(s) are still open", 996 open_devices); 997 return -EBUSY; 998 } 999 1000 pmd_write_lock_in_core(pmd); 1001 if (!pmd->fail_io && !dm_bm_is_read_only(pmd->bm)) { 1002 r = __commit_transaction(pmd); 1003 if (r < 0) 1004 DMWARN("%s: __commit_transaction() failed, error = %d", 1005 __func__, r); 1006 } 1007 pmd_write_unlock(pmd); 1008 if (!pmd->fail_io) 1009 __destroy_persistent_data_objects(pmd, true); 1010 1011 kfree(pmd); 1012 return 0; 1013 } 1014 1015 /* 1016 * __open_device: Returns @td corresponding to device with id @dev, 1017 * creating it if @create is set and incrementing @td->open_count. 1018 * On failure, @td is undefined. 1019 */ 1020 static int __open_device(struct dm_pool_metadata *pmd, 1021 dm_thin_id dev, int create, 1022 struct dm_thin_device **td) 1023 { 1024 int r, changed = 0; 1025 struct dm_thin_device *td2; 1026 uint64_t key = dev; 1027 struct disk_device_details details_le; 1028 1029 /* 1030 * If the device is already open, return it. 1031 */ 1032 list_for_each_entry(td2, &pmd->thin_devices, list) 1033 if (td2->id == dev) { 1034 /* 1035 * May not create an already-open device. 1036 */ 1037 if (create) 1038 return -EEXIST; 1039 1040 td2->open_count++; 1041 *td = td2; 1042 return 0; 1043 } 1044 1045 /* 1046 * Check the device exists. 1047 */ 1048 r = dm_btree_lookup(&pmd->details_info, pmd->details_root, 1049 &key, &details_le); 1050 if (r) { 1051 if (r != -ENODATA || !create) 1052 return r; 1053 1054 /* 1055 * Create new device. 1056 */ 1057 changed = 1; 1058 details_le.mapped_blocks = 0; 1059 details_le.transaction_id = cpu_to_le64(pmd->trans_id); 1060 details_le.creation_time = cpu_to_le32(pmd->time); 1061 details_le.snapshotted_time = cpu_to_le32(pmd->time); 1062 } 1063 1064 *td = kmalloc(sizeof(**td), GFP_NOIO); 1065 if (!*td) 1066 return -ENOMEM; 1067 1068 (*td)->pmd = pmd; 1069 (*td)->id = dev; 1070 (*td)->open_count = 1; 1071 (*td)->changed = changed; 1072 (*td)->aborted_with_changes = false; 1073 (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks); 1074 (*td)->transaction_id = le64_to_cpu(details_le.transaction_id); 1075 (*td)->creation_time = le32_to_cpu(details_le.creation_time); 1076 (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time); 1077 1078 list_add(&(*td)->list, &pmd->thin_devices); 1079 1080 return 0; 1081 } 1082 1083 static void __close_device(struct dm_thin_device *td) 1084 { 1085 --td->open_count; 1086 } 1087 1088 static int __create_thin(struct dm_pool_metadata *pmd, 1089 dm_thin_id dev) 1090 { 1091 int r; 1092 dm_block_t dev_root; 1093 uint64_t key = dev; 1094 struct dm_thin_device *td; 1095 __le64 value; 1096 1097 r = dm_btree_lookup(&pmd->details_info, pmd->details_root, 1098 &key, NULL); 1099 if (!r) 1100 return -EEXIST; 1101 1102 /* 1103 * Create an empty btree for the mappings. 1104 */ 1105 r = dm_btree_empty(&pmd->bl_info, &dev_root); 1106 if (r) 1107 return r; 1108 1109 /* 1110 * Insert it into the main mapping tree. 1111 */ 1112 value = cpu_to_le64(dev_root); 1113 __dm_bless_for_disk(&value); 1114 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root); 1115 if (r) { 1116 dm_btree_del(&pmd->bl_info, dev_root); 1117 return r; 1118 } 1119 1120 r = __open_device(pmd, dev, 1, &td); 1121 if (r) { 1122 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root); 1123 dm_btree_del(&pmd->bl_info, dev_root); 1124 return r; 1125 } 1126 __close_device(td); 1127 1128 return r; 1129 } 1130 1131 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev) 1132 { 1133 int r = -EINVAL; 1134 1135 pmd_write_lock(pmd); 1136 if (!pmd->fail_io) 1137 r = __create_thin(pmd, dev); 1138 pmd_write_unlock(pmd); 1139 1140 return r; 1141 } 1142 1143 static int __set_snapshot_details(struct dm_pool_metadata *pmd, 1144 struct dm_thin_device *snap, 1145 dm_thin_id origin, uint32_t time) 1146 { 1147 int r; 1148 struct dm_thin_device *td; 1149 1150 r = __open_device(pmd, origin, 0, &td); 1151 if (r) 1152 return r; 1153 1154 td->changed = true; 1155 td->snapshotted_time = time; 1156 1157 snap->mapped_blocks = td->mapped_blocks; 1158 snap->snapshotted_time = time; 1159 __close_device(td); 1160 1161 return 0; 1162 } 1163 1164 static int __create_snap(struct dm_pool_metadata *pmd, 1165 dm_thin_id dev, dm_thin_id origin) 1166 { 1167 int r; 1168 dm_block_t origin_root; 1169 uint64_t key = origin, dev_key = dev; 1170 struct dm_thin_device *td; 1171 __le64 value; 1172 1173 /* check this device is unused */ 1174 r = dm_btree_lookup(&pmd->details_info, pmd->details_root, 1175 &dev_key, NULL); 1176 if (!r) 1177 return -EEXIST; 1178 1179 /* find the mapping tree for the origin */ 1180 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value); 1181 if (r) 1182 return r; 1183 origin_root = le64_to_cpu(value); 1184 1185 /* clone the origin, an inc will do */ 1186 dm_tm_inc(pmd->tm, origin_root); 1187 1188 /* insert into the main mapping tree */ 1189 value = cpu_to_le64(origin_root); 1190 __dm_bless_for_disk(&value); 1191 key = dev; 1192 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root); 1193 if (r) { 1194 dm_tm_dec(pmd->tm, origin_root); 1195 return r; 1196 } 1197 1198 pmd->time++; 1199 1200 r = __open_device(pmd, dev, 1, &td); 1201 if (r) 1202 goto bad; 1203 1204 r = __set_snapshot_details(pmd, td, origin, pmd->time); 1205 __close_device(td); 1206 1207 if (r) 1208 goto bad; 1209 1210 return 0; 1211 1212 bad: 1213 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root); 1214 dm_btree_remove(&pmd->details_info, pmd->details_root, 1215 &key, &pmd->details_root); 1216 return r; 1217 } 1218 1219 int dm_pool_create_snap(struct dm_pool_metadata *pmd, 1220 dm_thin_id dev, 1221 dm_thin_id origin) 1222 { 1223 int r = -EINVAL; 1224 1225 pmd_write_lock(pmd); 1226 if (!pmd->fail_io) 1227 r = __create_snap(pmd, dev, origin); 1228 pmd_write_unlock(pmd); 1229 1230 return r; 1231 } 1232 1233 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev) 1234 { 1235 int r; 1236 uint64_t key = dev; 1237 struct dm_thin_device *td; 1238 1239 /* TODO: failure should mark the transaction invalid */ 1240 r = __open_device(pmd, dev, 0, &td); 1241 if (r) 1242 return r; 1243 1244 if (td->open_count > 1) { 1245 __close_device(td); 1246 return -EBUSY; 1247 } 1248 1249 list_del(&td->list); 1250 kfree(td); 1251 r = dm_btree_remove(&pmd->details_info, pmd->details_root, 1252 &key, &pmd->details_root); 1253 if (r) 1254 return r; 1255 1256 r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root); 1257 if (r) 1258 return r; 1259 1260 return 0; 1261 } 1262 1263 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd, 1264 dm_thin_id dev) 1265 { 1266 int r = -EINVAL; 1267 1268 pmd_write_lock(pmd); 1269 if (!pmd->fail_io) 1270 r = __delete_device(pmd, dev); 1271 pmd_write_unlock(pmd); 1272 1273 return r; 1274 } 1275 1276 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd, 1277 uint64_t current_id, 1278 uint64_t new_id) 1279 { 1280 int r = -EINVAL; 1281 1282 pmd_write_lock(pmd); 1283 1284 if (pmd->fail_io) 1285 goto out; 1286 1287 if (pmd->trans_id != current_id) { 1288 DMERR("mismatched transaction id"); 1289 goto out; 1290 } 1291 1292 pmd->trans_id = new_id; 1293 r = 0; 1294 1295 out: 1296 pmd_write_unlock(pmd); 1297 1298 return r; 1299 } 1300 1301 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd, 1302 uint64_t *result) 1303 { 1304 int r = -EINVAL; 1305 1306 down_read(&pmd->root_lock); 1307 if (!pmd->fail_io) { 1308 *result = pmd->trans_id; 1309 r = 0; 1310 } 1311 up_read(&pmd->root_lock); 1312 1313 return r; 1314 } 1315 1316 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd) 1317 { 1318 int r, inc; 1319 struct thin_disk_superblock *disk_super; 1320 struct dm_block *copy, *sblock; 1321 dm_block_t held_root; 1322 1323 /* 1324 * We commit to ensure the btree roots which we increment in a 1325 * moment are up to date. 1326 */ 1327 r = __commit_transaction(pmd); 1328 if (r < 0) { 1329 DMWARN("%s: __commit_transaction() failed, error = %d", 1330 __func__, r); 1331 return r; 1332 } 1333 1334 /* 1335 * Copy the superblock. 1336 */ 1337 dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION); 1338 r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION, 1339 &sb_validator, ©, &inc); 1340 if (r) 1341 return r; 1342 1343 BUG_ON(!inc); 1344 1345 held_root = dm_block_location(copy); 1346 disk_super = dm_block_data(copy); 1347 1348 if (le64_to_cpu(disk_super->held_root)) { 1349 DMWARN("Pool metadata snapshot already exists: release this before taking another."); 1350 1351 dm_tm_dec(pmd->tm, held_root); 1352 dm_tm_unlock(pmd->tm, copy); 1353 return -EBUSY; 1354 } 1355 1356 /* 1357 * Wipe the spacemap since we're not publishing this. 1358 */ 1359 memset(&disk_super->data_space_map_root, 0, 1360 sizeof(disk_super->data_space_map_root)); 1361 memset(&disk_super->metadata_space_map_root, 0, 1362 sizeof(disk_super->metadata_space_map_root)); 1363 1364 /* 1365 * Increment the data structures that need to be preserved. 1366 */ 1367 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root)); 1368 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root)); 1369 dm_tm_unlock(pmd->tm, copy); 1370 1371 /* 1372 * Write the held root into the superblock. 1373 */ 1374 r = superblock_lock(pmd, &sblock); 1375 if (r) { 1376 dm_tm_dec(pmd->tm, held_root); 1377 return r; 1378 } 1379 1380 disk_super = dm_block_data(sblock); 1381 disk_super->held_root = cpu_to_le64(held_root); 1382 dm_bm_unlock(sblock); 1383 return 0; 1384 } 1385 1386 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd) 1387 { 1388 int r = -EINVAL; 1389 1390 pmd_write_lock(pmd); 1391 if (!pmd->fail_io) 1392 r = __reserve_metadata_snap(pmd); 1393 pmd_write_unlock(pmd); 1394 1395 return r; 1396 } 1397 1398 static int __release_metadata_snap(struct dm_pool_metadata *pmd) 1399 { 1400 int r; 1401 struct thin_disk_superblock *disk_super; 1402 struct dm_block *sblock, *copy; 1403 dm_block_t held_root; 1404 1405 r = superblock_lock(pmd, &sblock); 1406 if (r) 1407 return r; 1408 1409 disk_super = dm_block_data(sblock); 1410 held_root = le64_to_cpu(disk_super->held_root); 1411 disk_super->held_root = cpu_to_le64(0); 1412 1413 dm_bm_unlock(sblock); 1414 1415 if (!held_root) { 1416 DMWARN("No pool metadata snapshot found: nothing to release."); 1417 return -EINVAL; 1418 } 1419 1420 r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, ©); 1421 if (r) 1422 return r; 1423 1424 disk_super = dm_block_data(copy); 1425 dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root)); 1426 dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root)); 1427 dm_sm_dec_block(pmd->metadata_sm, held_root); 1428 1429 dm_tm_unlock(pmd->tm, copy); 1430 1431 return 0; 1432 } 1433 1434 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd) 1435 { 1436 int r = -EINVAL; 1437 1438 pmd_write_lock(pmd); 1439 if (!pmd->fail_io) 1440 r = __release_metadata_snap(pmd); 1441 pmd_write_unlock(pmd); 1442 1443 return r; 1444 } 1445 1446 static int __get_metadata_snap(struct dm_pool_metadata *pmd, 1447 dm_block_t *result) 1448 { 1449 int r; 1450 struct thin_disk_superblock *disk_super; 1451 struct dm_block *sblock; 1452 1453 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION, 1454 &sb_validator, &sblock); 1455 if (r) 1456 return r; 1457 1458 disk_super = dm_block_data(sblock); 1459 *result = le64_to_cpu(disk_super->held_root); 1460 1461 dm_bm_unlock(sblock); 1462 1463 return 0; 1464 } 1465 1466 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd, 1467 dm_block_t *result) 1468 { 1469 int r = -EINVAL; 1470 1471 down_read(&pmd->root_lock); 1472 if (!pmd->fail_io) 1473 r = __get_metadata_snap(pmd, result); 1474 up_read(&pmd->root_lock); 1475 1476 return r; 1477 } 1478 1479 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev, 1480 struct dm_thin_device **td) 1481 { 1482 int r = -EINVAL; 1483 1484 pmd_write_lock_in_core(pmd); 1485 if (!pmd->fail_io) 1486 r = __open_device(pmd, dev, 0, td); 1487 pmd_write_unlock(pmd); 1488 1489 return r; 1490 } 1491 1492 int dm_pool_close_thin_device(struct dm_thin_device *td) 1493 { 1494 pmd_write_lock_in_core(td->pmd); 1495 __close_device(td); 1496 pmd_write_unlock(td->pmd); 1497 1498 return 0; 1499 } 1500 1501 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td) 1502 { 1503 return td->id; 1504 } 1505 1506 /* 1507 * Check whether @time (of block creation) is older than @td's last snapshot. 1508 * If so then the associated block is shared with the last snapshot device. 1509 * Any block on a device created *after* the device last got snapshotted is 1510 * necessarily not shared. 1511 */ 1512 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time) 1513 { 1514 return td->snapshotted_time > time; 1515 } 1516 1517 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value, 1518 struct dm_thin_lookup_result *result) 1519 { 1520 uint64_t block_time = 0; 1521 dm_block_t exception_block; 1522 uint32_t exception_time; 1523 1524 block_time = le64_to_cpu(value); 1525 unpack_block_time(block_time, &exception_block, &exception_time); 1526 result->block = exception_block; 1527 result->shared = __snapshotted_since(td, exception_time); 1528 } 1529 1530 static int __find_block(struct dm_thin_device *td, dm_block_t block, 1531 int can_issue_io, struct dm_thin_lookup_result *result) 1532 { 1533 int r; 1534 __le64 value; 1535 struct dm_pool_metadata *pmd = td->pmd; 1536 dm_block_t keys[2] = { td->id, block }; 1537 struct dm_btree_info *info; 1538 1539 if (can_issue_io) 1540 info = &pmd->info; 1541 else 1542 info = &pmd->nb_info; 1543 1544 r = dm_btree_lookup(info, pmd->root, keys, &value); 1545 if (!r) 1546 unpack_lookup_result(td, value, result); 1547 1548 return r; 1549 } 1550 1551 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block, 1552 int can_issue_io, struct dm_thin_lookup_result *result) 1553 { 1554 int r; 1555 struct dm_pool_metadata *pmd = td->pmd; 1556 1557 down_read(&pmd->root_lock); 1558 if (pmd->fail_io) { 1559 up_read(&pmd->root_lock); 1560 return -EINVAL; 1561 } 1562 1563 r = __find_block(td, block, can_issue_io, result); 1564 1565 up_read(&pmd->root_lock); 1566 return r; 1567 } 1568 1569 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block, 1570 dm_block_t *vblock, 1571 struct dm_thin_lookup_result *result) 1572 { 1573 int r; 1574 __le64 value; 1575 struct dm_pool_metadata *pmd = td->pmd; 1576 dm_block_t keys[2] = { td->id, block }; 1577 1578 r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value); 1579 if (!r) 1580 unpack_lookup_result(td, value, result); 1581 1582 return r; 1583 } 1584 1585 static int __find_mapped_range(struct dm_thin_device *td, 1586 dm_block_t begin, dm_block_t end, 1587 dm_block_t *thin_begin, dm_block_t *thin_end, 1588 dm_block_t *pool_begin, bool *maybe_shared) 1589 { 1590 int r; 1591 dm_block_t pool_end; 1592 struct dm_thin_lookup_result lookup; 1593 1594 if (end < begin) 1595 return -ENODATA; 1596 1597 r = __find_next_mapped_block(td, begin, &begin, &lookup); 1598 if (r) 1599 return r; 1600 1601 if (begin >= end) 1602 return -ENODATA; 1603 1604 *thin_begin = begin; 1605 *pool_begin = lookup.block; 1606 *maybe_shared = lookup.shared; 1607 1608 begin++; 1609 pool_end = *pool_begin + 1; 1610 while (begin != end) { 1611 r = __find_block(td, begin, true, &lookup); 1612 if (r) { 1613 if (r == -ENODATA) 1614 break; 1615 1616 return r; 1617 } 1618 1619 if ((lookup.block != pool_end) || 1620 (lookup.shared != *maybe_shared)) 1621 break; 1622 1623 pool_end++; 1624 begin++; 1625 } 1626 1627 *thin_end = begin; 1628 return 0; 1629 } 1630 1631 int dm_thin_find_mapped_range(struct dm_thin_device *td, 1632 dm_block_t begin, dm_block_t end, 1633 dm_block_t *thin_begin, dm_block_t *thin_end, 1634 dm_block_t *pool_begin, bool *maybe_shared) 1635 { 1636 int r = -EINVAL; 1637 struct dm_pool_metadata *pmd = td->pmd; 1638 1639 down_read(&pmd->root_lock); 1640 if (!pmd->fail_io) { 1641 r = __find_mapped_range(td, begin, end, thin_begin, thin_end, 1642 pool_begin, maybe_shared); 1643 } 1644 up_read(&pmd->root_lock); 1645 1646 return r; 1647 } 1648 1649 static int __insert(struct dm_thin_device *td, dm_block_t block, 1650 dm_block_t data_block) 1651 { 1652 int r, inserted; 1653 __le64 value; 1654 struct dm_pool_metadata *pmd = td->pmd; 1655 dm_block_t keys[2] = { td->id, block }; 1656 1657 value = cpu_to_le64(pack_block_time(data_block, pmd->time)); 1658 __dm_bless_for_disk(&value); 1659 1660 r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value, 1661 &pmd->root, &inserted); 1662 if (r) 1663 return r; 1664 1665 td->changed = true; 1666 if (inserted) 1667 td->mapped_blocks++; 1668 1669 return 0; 1670 } 1671 1672 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block, 1673 dm_block_t data_block) 1674 { 1675 int r = -EINVAL; 1676 1677 pmd_write_lock(td->pmd); 1678 if (!td->pmd->fail_io) 1679 r = __insert(td, block, data_block); 1680 pmd_write_unlock(td->pmd); 1681 1682 return r; 1683 } 1684 1685 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end) 1686 { 1687 int r; 1688 unsigned int count, total_count = 0; 1689 struct dm_pool_metadata *pmd = td->pmd; 1690 dm_block_t keys[1] = { td->id }; 1691 __le64 value; 1692 dm_block_t mapping_root; 1693 1694 /* 1695 * Find the mapping tree 1696 */ 1697 r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value); 1698 if (r) 1699 return r; 1700 1701 /* 1702 * Remove from the mapping tree, taking care to inc the 1703 * ref count so it doesn't get deleted. 1704 */ 1705 mapping_root = le64_to_cpu(value); 1706 dm_tm_inc(pmd->tm, mapping_root); 1707 r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root); 1708 if (r) 1709 return r; 1710 1711 /* 1712 * Remove leaves stops at the first unmapped entry, so we have to 1713 * loop round finding mapped ranges. 1714 */ 1715 while (begin < end) { 1716 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value); 1717 if (r == -ENODATA) 1718 break; 1719 1720 if (r) 1721 return r; 1722 1723 if (begin >= end) 1724 break; 1725 1726 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count); 1727 if (r) 1728 return r; 1729 1730 total_count += count; 1731 } 1732 1733 td->mapped_blocks -= total_count; 1734 td->changed = true; 1735 1736 /* 1737 * Reinsert the mapping tree. 1738 */ 1739 value = cpu_to_le64(mapping_root); 1740 __dm_bless_for_disk(&value); 1741 return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root); 1742 } 1743 1744 int dm_thin_remove_range(struct dm_thin_device *td, 1745 dm_block_t begin, dm_block_t end) 1746 { 1747 int r = -EINVAL; 1748 1749 pmd_write_lock(td->pmd); 1750 if (!td->pmd->fail_io) 1751 r = __remove_range(td, begin, end); 1752 pmd_write_unlock(td->pmd); 1753 1754 return r; 1755 } 1756 1757 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result) 1758 { 1759 int r; 1760 uint32_t ref_count; 1761 1762 down_read(&pmd->root_lock); 1763 r = dm_sm_get_count(pmd->data_sm, b, &ref_count); 1764 if (!r) 1765 *result = (ref_count > 1); 1766 up_read(&pmd->root_lock); 1767 1768 return r; 1769 } 1770 1771 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e) 1772 { 1773 int r = 0; 1774 1775 pmd_write_lock(pmd); 1776 r = dm_sm_inc_blocks(pmd->data_sm, b, e); 1777 pmd_write_unlock(pmd); 1778 1779 return r; 1780 } 1781 1782 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e) 1783 { 1784 int r = 0; 1785 1786 pmd_write_lock(pmd); 1787 r = dm_sm_dec_blocks(pmd->data_sm, b, e); 1788 pmd_write_unlock(pmd); 1789 1790 return r; 1791 } 1792 1793 bool dm_thin_changed_this_transaction(struct dm_thin_device *td) 1794 { 1795 int r; 1796 1797 down_read(&td->pmd->root_lock); 1798 r = td->changed; 1799 up_read(&td->pmd->root_lock); 1800 1801 return r; 1802 } 1803 1804 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd) 1805 { 1806 bool r = false; 1807 struct dm_thin_device *td, *tmp; 1808 1809 down_read(&pmd->root_lock); 1810 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) { 1811 if (td->changed) { 1812 r = td->changed; 1813 break; 1814 } 1815 } 1816 up_read(&pmd->root_lock); 1817 1818 return r; 1819 } 1820 1821 bool dm_thin_aborted_changes(struct dm_thin_device *td) 1822 { 1823 bool r; 1824 1825 down_read(&td->pmd->root_lock); 1826 r = td->aborted_with_changes; 1827 up_read(&td->pmd->root_lock); 1828 1829 return r; 1830 } 1831 1832 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result) 1833 { 1834 int r = -EINVAL; 1835 1836 pmd_write_lock(pmd); 1837 if (!pmd->fail_io) 1838 r = dm_sm_new_block(pmd->data_sm, result); 1839 pmd_write_unlock(pmd); 1840 1841 return r; 1842 } 1843 1844 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd) 1845 { 1846 int r = -EINVAL; 1847 1848 /* 1849 * Care is taken to not have commit be what 1850 * triggers putting the thin-pool in-service. 1851 */ 1852 pmd_write_lock_in_core(pmd); 1853 if (pmd->fail_io) 1854 goto out; 1855 1856 r = __commit_transaction(pmd); 1857 if (r < 0) 1858 goto out; 1859 1860 /* 1861 * Open the next transaction. 1862 */ 1863 r = __begin_transaction(pmd); 1864 out: 1865 pmd_write_unlock(pmd); 1866 return r; 1867 } 1868 1869 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd) 1870 { 1871 struct dm_thin_device *td; 1872 1873 list_for_each_entry(td, &pmd->thin_devices, list) 1874 td->aborted_with_changes = td->changed; 1875 } 1876 1877 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd) 1878 { 1879 int r = -EINVAL; 1880 struct dm_block_manager *old_bm = NULL, *new_bm = NULL; 1881 1882 /* fail_io is double-checked with pmd->root_lock held below */ 1883 if (unlikely(pmd->fail_io)) 1884 return r; 1885 1886 /* 1887 * Replacement block manager (new_bm) is created and old_bm destroyed outside of 1888 * pmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of 1889 * shrinker associated with the block manager's bufio client vs pmd root_lock). 1890 * - must take shrinker_rwsem without holding pmd->root_lock 1891 */ 1892 new_bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT, 1893 THIN_MAX_CONCURRENT_LOCKS); 1894 1895 pmd_write_lock(pmd); 1896 if (pmd->fail_io) { 1897 pmd_write_unlock(pmd); 1898 goto out; 1899 } 1900 1901 __set_abort_with_changes_flags(pmd); 1902 __destroy_persistent_data_objects(pmd, false); 1903 old_bm = pmd->bm; 1904 if (IS_ERR(new_bm)) { 1905 DMERR("could not create block manager during abort"); 1906 pmd->bm = NULL; 1907 r = PTR_ERR(new_bm); 1908 goto out_unlock; 1909 } 1910 1911 pmd->bm = new_bm; 1912 r = __open_or_format_metadata(pmd, false); 1913 if (r) { 1914 pmd->bm = NULL; 1915 goto out_unlock; 1916 } 1917 new_bm = NULL; 1918 out_unlock: 1919 if (r) 1920 pmd->fail_io = true; 1921 pmd_write_unlock(pmd); 1922 dm_block_manager_destroy(old_bm); 1923 out: 1924 if (new_bm && !IS_ERR(new_bm)) 1925 dm_block_manager_destroy(new_bm); 1926 1927 return r; 1928 } 1929 1930 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result) 1931 { 1932 int r = -EINVAL; 1933 1934 down_read(&pmd->root_lock); 1935 if (!pmd->fail_io) 1936 r = dm_sm_get_nr_free(pmd->data_sm, result); 1937 up_read(&pmd->root_lock); 1938 1939 return r; 1940 } 1941 1942 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd, 1943 dm_block_t *result) 1944 { 1945 int r = -EINVAL; 1946 1947 down_read(&pmd->root_lock); 1948 if (!pmd->fail_io) 1949 r = dm_sm_get_nr_free(pmd->metadata_sm, result); 1950 1951 if (!r) { 1952 if (*result < pmd->metadata_reserve) 1953 *result = 0; 1954 else 1955 *result -= pmd->metadata_reserve; 1956 } 1957 up_read(&pmd->root_lock); 1958 1959 return r; 1960 } 1961 1962 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd, 1963 dm_block_t *result) 1964 { 1965 int r = -EINVAL; 1966 1967 down_read(&pmd->root_lock); 1968 if (!pmd->fail_io) 1969 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result); 1970 up_read(&pmd->root_lock); 1971 1972 return r; 1973 } 1974 1975 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result) 1976 { 1977 int r = -EINVAL; 1978 1979 down_read(&pmd->root_lock); 1980 if (!pmd->fail_io) 1981 r = dm_sm_get_nr_blocks(pmd->data_sm, result); 1982 up_read(&pmd->root_lock); 1983 1984 return r; 1985 } 1986 1987 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result) 1988 { 1989 int r = -EINVAL; 1990 struct dm_pool_metadata *pmd = td->pmd; 1991 1992 down_read(&pmd->root_lock); 1993 if (!pmd->fail_io) { 1994 *result = td->mapped_blocks; 1995 r = 0; 1996 } 1997 up_read(&pmd->root_lock); 1998 1999 return r; 2000 } 2001 2002 static int __highest_block(struct dm_thin_device *td, dm_block_t *result) 2003 { 2004 int r; 2005 __le64 value_le; 2006 dm_block_t thin_root; 2007 struct dm_pool_metadata *pmd = td->pmd; 2008 2009 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le); 2010 if (r) 2011 return r; 2012 2013 thin_root = le64_to_cpu(value_le); 2014 2015 return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result); 2016 } 2017 2018 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td, 2019 dm_block_t *result) 2020 { 2021 int r = -EINVAL; 2022 struct dm_pool_metadata *pmd = td->pmd; 2023 2024 down_read(&pmd->root_lock); 2025 if (!pmd->fail_io) 2026 r = __highest_block(td, result); 2027 up_read(&pmd->root_lock); 2028 2029 return r; 2030 } 2031 2032 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count) 2033 { 2034 int r; 2035 dm_block_t old_count; 2036 2037 r = dm_sm_get_nr_blocks(sm, &old_count); 2038 if (r) 2039 return r; 2040 2041 if (new_count == old_count) 2042 return 0; 2043 2044 if (new_count < old_count) { 2045 DMERR("cannot reduce size of space map"); 2046 return -EINVAL; 2047 } 2048 2049 return dm_sm_extend(sm, new_count - old_count); 2050 } 2051 2052 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count) 2053 { 2054 int r = -EINVAL; 2055 2056 pmd_write_lock(pmd); 2057 if (!pmd->fail_io) 2058 r = __resize_space_map(pmd->data_sm, new_count); 2059 pmd_write_unlock(pmd); 2060 2061 return r; 2062 } 2063 2064 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count) 2065 { 2066 int r = -EINVAL; 2067 2068 pmd_write_lock(pmd); 2069 if (!pmd->fail_io) { 2070 r = __resize_space_map(pmd->metadata_sm, new_count); 2071 if (!r) 2072 __set_metadata_reserve(pmd); 2073 } 2074 pmd_write_unlock(pmd); 2075 2076 return r; 2077 } 2078 2079 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd) 2080 { 2081 pmd_write_lock_in_core(pmd); 2082 dm_bm_set_read_only(pmd->bm); 2083 pmd_write_unlock(pmd); 2084 } 2085 2086 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd) 2087 { 2088 pmd_write_lock_in_core(pmd); 2089 dm_bm_set_read_write(pmd->bm); 2090 pmd_write_unlock(pmd); 2091 } 2092 2093 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd, 2094 dm_block_t threshold, 2095 dm_sm_threshold_fn fn, 2096 void *context) 2097 { 2098 int r = -EINVAL; 2099 2100 pmd_write_lock_in_core(pmd); 2101 if (!pmd->fail_io) { 2102 r = dm_sm_register_threshold_callback(pmd->metadata_sm, 2103 threshold, fn, context); 2104 } 2105 pmd_write_unlock(pmd); 2106 2107 return r; 2108 } 2109 2110 void dm_pool_register_pre_commit_callback(struct dm_pool_metadata *pmd, 2111 dm_pool_pre_commit_fn fn, 2112 void *context) 2113 { 2114 pmd_write_lock_in_core(pmd); 2115 pmd->pre_commit_fn = fn; 2116 pmd->pre_commit_context = context; 2117 pmd_write_unlock(pmd); 2118 } 2119 2120 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd) 2121 { 2122 int r = -EINVAL; 2123 struct dm_block *sblock; 2124 struct thin_disk_superblock *disk_super; 2125 2126 pmd_write_lock(pmd); 2127 if (pmd->fail_io) 2128 goto out; 2129 2130 pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG; 2131 2132 r = superblock_lock(pmd, &sblock); 2133 if (r) { 2134 DMERR("couldn't lock superblock"); 2135 goto out; 2136 } 2137 2138 disk_super = dm_block_data(sblock); 2139 disk_super->flags = cpu_to_le32(pmd->flags); 2140 2141 dm_bm_unlock(sblock); 2142 out: 2143 pmd_write_unlock(pmd); 2144 return r; 2145 } 2146 2147 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd) 2148 { 2149 bool needs_check; 2150 2151 down_read(&pmd->root_lock); 2152 needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG; 2153 up_read(&pmd->root_lock); 2154 2155 return needs_check; 2156 } 2157 2158 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd) 2159 { 2160 down_read(&pmd->root_lock); 2161 if (!pmd->fail_io) 2162 dm_tm_issue_prefetches(pmd->tm); 2163 up_read(&pmd->root_lock); 2164 } 2165