Linux Audio

Check our new training course

Loading...
v6.2
   1/*
   2 * Copyright (C) 2011-2012 Red Hat, Inc.
   3 *
   4 * This file is released under the GPL.
   5 */
   6
   7#include "dm-thin-metadata.h"
   8#include "persistent-data/dm-btree.h"
   9#include "persistent-data/dm-space-map.h"
  10#include "persistent-data/dm-space-map-disk.h"
  11#include "persistent-data/dm-transaction-manager.h"
  12
  13#include <linux/list.h>
  14#include <linux/device-mapper.h>
  15#include <linux/workqueue.h>
  16
  17/*--------------------------------------------------------------------------
  18 * As far as the metadata goes, there is:
  19 *
  20 * - A superblock in block zero, taking up fewer than 512 bytes for
  21 *   atomic writes.
  22 *
  23 * - A space map managing the metadata blocks.
  24 *
  25 * - A space map managing the data blocks.
  26 *
  27 * - A btree mapping our internal thin dev ids onto struct disk_device_details.
  28 *
  29 * - A hierarchical btree, with 2 levels which effectively maps (thin
  30 *   dev id, virtual block) -> block_time.  Block time is a 64-bit
  31 *   field holding the time in the low 24 bits, and block in the top 40
  32 *   bits.
  33 *
  34 * BTrees consist solely of btree_nodes, that fill a block.  Some are
  35 * internal nodes, as such their values are a __le64 pointing to other
  36 * nodes.  Leaf nodes can store data of any reasonable size (ie. much
  37 * smaller than the block size).  The nodes consist of the header,
  38 * followed by an array of keys, followed by an array of values.  We have
  39 * to binary search on the keys so they're all held together to help the
  40 * cpu cache.
  41 *
  42 * Space maps have 2 btrees:
  43 *
  44 * - One maps a uint64_t onto a struct index_entry.  Which points to a
  45 *   bitmap block, and has some details about how many free entries there
  46 *   are etc.
  47 *
  48 * - The bitmap blocks have a header (for the checksum).  Then the rest
  49 *   of the block is pairs of bits.  With the meaning being:
  50 *
  51 *   0 - ref count is 0
  52 *   1 - ref count is 1
  53 *   2 - ref count is 2
  54 *   3 - ref count is higher than 2
  55 *
  56 * - If the count is higher than 2 then the ref count is entered in a
  57 *   second btree that directly maps the block_address to a uint32_t ref
  58 *   count.
  59 *
  60 * The space map metadata variant doesn't have a bitmaps btree.  Instead
  61 * it has one single blocks worth of index_entries.  This avoids
  62 * recursive issues with the bitmap btree needing to allocate space in
  63 * order to insert.  With a small data block size such as 64k the
  64 * metadata support data devices that are hundreds of terrabytes.
  65 *
  66 * The space maps allocate space linearly from front to back.  Space that
  67 * is freed in a transaction is never recycled within that transaction.
  68 * To try and avoid fragmenting _free_ space the allocator always goes
  69 * back and fills in gaps.
  70 *
  71 * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
  72 * from the block manager.
  73 *--------------------------------------------------------------------------*/
  74
  75#define DM_MSG_PREFIX   "thin metadata"
  76
  77#define THIN_SUPERBLOCK_MAGIC 27022010
  78#define THIN_SUPERBLOCK_LOCATION 0
  79#define THIN_VERSION 2
  80#define SECTOR_TO_BLOCK_SHIFT 3
  81
  82/*
  83 * For btree insert:
  84 *  3 for btree insert +
  85 *  2 for btree lookup used within space map
  86 * For btree remove:
  87 *  2 for shadow spine +
  88 *  4 for rebalance 3 child node
  89 */
  90#define THIN_MAX_CONCURRENT_LOCKS 6
  91
  92/* This should be plenty */
  93#define SPACE_MAP_ROOT_SIZE 128
  94
  95/*
  96 * Little endian on-disk superblock and device details.
  97 */
  98struct thin_disk_superblock {
  99	__le32 csum;	/* Checksum of superblock except for this field. */
 100	__le32 flags;
 101	__le64 blocknr;	/* This block number, dm_block_t. */
 102
 103	__u8 uuid[16];
 104	__le64 magic;
 105	__le32 version;
 106	__le32 time;
 107
 108	__le64 trans_id;
 109
 110	/*
 111	 * Root held by userspace transactions.
 112	 */
 113	__le64 held_root;
 114
 115	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 116	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 117
 118	/*
 119	 * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 120	 */
 121	__le64 data_mapping_root;
 122
 123	/*
 124	 * Device detail root mapping dev_id -> device_details
 125	 */
 126	__le64 device_details_root;
 127
 128	__le32 data_block_size;		/* In 512-byte sectors. */
 129
 130	__le32 metadata_block_size;	/* In 512-byte sectors. */
 131	__le64 metadata_nr_blocks;
 132
 133	__le32 compat_flags;
 134	__le32 compat_ro_flags;
 135	__le32 incompat_flags;
 136} __packed;
 137
 138struct disk_device_details {
 139	__le64 mapped_blocks;
 140	__le64 transaction_id;		/* When created. */
 141	__le32 creation_time;
 142	__le32 snapshotted_time;
 143} __packed;
 144
 145struct dm_pool_metadata {
 146	struct hlist_node hash;
 147
 148	struct block_device *bdev;
 149	struct dm_block_manager *bm;
 150	struct dm_space_map *metadata_sm;
 151	struct dm_space_map *data_sm;
 152	struct dm_transaction_manager *tm;
 153	struct dm_transaction_manager *nb_tm;
 154
 155	/*
 156	 * Two-level btree.
 157	 * First level holds thin_dev_t.
 158	 * Second level holds mappings.
 159	 */
 160	struct dm_btree_info info;
 161
 162	/*
 163	 * Non-blocking version of the above.
 164	 */
 165	struct dm_btree_info nb_info;
 166
 167	/*
 168	 * Just the top level for deleting whole devices.
 169	 */
 170	struct dm_btree_info tl_info;
 171
 172	/*
 173	 * Just the bottom level for creating new devices.
 174	 */
 175	struct dm_btree_info bl_info;
 176
 177	/*
 178	 * Describes the device details btree.
 179	 */
 180	struct dm_btree_info details_info;
 181
 182	struct rw_semaphore root_lock;
 183	uint32_t time;
 184	dm_block_t root;
 185	dm_block_t details_root;
 186	struct list_head thin_devices;
 187	uint64_t trans_id;
 188	unsigned long flags;
 189	sector_t data_block_size;
 190
 191	/*
 192	 * Pre-commit callback.
 193	 *
 194	 * This allows the thin provisioning target to run a callback before
 195	 * the metadata are committed.
 196	 */
 197	dm_pool_pre_commit_fn pre_commit_fn;
 198	void *pre_commit_context;
 199
 200	/*
 201	 * We reserve a section of the metadata for commit overhead.
 202	 * All reported space does *not* include this.
 203	 */
 204	dm_block_t metadata_reserve;
 205
 206	/*
 207	 * Set if a transaction has to be aborted but the attempt to roll back
 208	 * to the previous (good) transaction failed.  The only pool metadata
 209	 * operation possible in this state is the closing of the device.
 210	 */
 211	bool fail_io:1;
 212
 213	/*
 214	 * Set once a thin-pool has been accessed through one of the interfaces
 215	 * that imply the pool is in-service (e.g. thin devices created/deleted,
 216	 * thin-pool message, metadata snapshots, etc).
 217	 */
 218	bool in_service:1;
 219
 220	/*
 221	 * Reading the space map roots can fail, so we read it into these
 222	 * buffers before the superblock is locked and updated.
 223	 */
 224	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 225	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 226};
 227
 228struct dm_thin_device {
 229	struct list_head list;
 230	struct dm_pool_metadata *pmd;
 231	dm_thin_id id;
 232
 233	int open_count;
 234	bool changed:1;
 235	bool aborted_with_changes:1;
 236	uint64_t mapped_blocks;
 237	uint64_t transaction_id;
 238	uint32_t creation_time;
 239	uint32_t snapshotted_time;
 240};
 241
 242/*----------------------------------------------------------------
 243 * superblock validator
 244 *--------------------------------------------------------------*/
 245
 246#define SUPERBLOCK_CSUM_XOR 160774
 247
 248static void sb_prepare_for_write(struct dm_block_validator *v,
 249				 struct dm_block *b,
 250				 size_t block_size)
 251{
 252	struct thin_disk_superblock *disk_super = dm_block_data(b);
 253
 254	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 255	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 256						      block_size - sizeof(__le32),
 257						      SUPERBLOCK_CSUM_XOR));
 258}
 259
 260static int sb_check(struct dm_block_validator *v,
 261		    struct dm_block *b,
 262		    size_t block_size)
 263{
 264	struct thin_disk_superblock *disk_super = dm_block_data(b);
 265	__le32 csum_le;
 266
 267	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 268		DMERR("sb_check failed: blocknr %llu: "
 269		      "wanted %llu", le64_to_cpu(disk_super->blocknr),
 270		      (unsigned long long)dm_block_location(b));
 271		return -ENOTBLK;
 272	}
 273
 274	if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 275		DMERR("sb_check failed: magic %llu: "
 276		      "wanted %llu", le64_to_cpu(disk_super->magic),
 277		      (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 278		return -EILSEQ;
 279	}
 280
 281	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 282					     block_size - sizeof(__le32),
 283					     SUPERBLOCK_CSUM_XOR));
 284	if (csum_le != disk_super->csum) {
 285		DMERR("sb_check failed: csum %u: wanted %u",
 286		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 287		return -EILSEQ;
 288	}
 289
 290	return 0;
 291}
 292
 293static struct dm_block_validator sb_validator = {
 294	.name = "superblock",
 295	.prepare_for_write = sb_prepare_for_write,
 296	.check = sb_check
 297};
 298
 299/*----------------------------------------------------------------
 300 * Methods for the btree value types
 301 *--------------------------------------------------------------*/
 302
 303static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 304{
 305	return (b << 24) | t;
 306}
 307
 308static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 309{
 310	*b = v >> 24;
 311	*t = v & ((1 << 24) - 1);
 312}
 313
 314/*
 315 * It's more efficient to call dm_sm_{inc,dec}_blocks as few times as
 316 * possible.  'with_runs' reads contiguous runs of blocks, and calls the
 317 * given sm function.
 318 */
 319typedef int (*run_fn)(struct dm_space_map *, dm_block_t, dm_block_t);
 320
 321static void with_runs(struct dm_space_map *sm, const __le64 *value_le, unsigned count, run_fn fn)
 322{
 323	uint64_t b, begin, end;
 
 
 324	uint32_t t;
 325	bool in_run = false;
 326	unsigned i;
 327
 328	for (i = 0; i < count; i++, value_le++) {
 329		/* We know value_le is 8 byte aligned */
 330		unpack_block_time(le64_to_cpu(*value_le), &b, &t);
 331
 332		if (in_run) {
 333			if (b == end) {
 334				end++;
 335			} else {
 336				fn(sm, begin, end);
 337				begin = b;
 338				end = b + 1;
 339			}
 340		} else {
 341			in_run = true;
 342			begin = b;
 343			end = b + 1;
 344		}
 345	}
 346
 347	if (in_run)
 348		fn(sm, begin, end);
 349}
 350
 351static void data_block_inc(void *context, const void *value_le, unsigned count)
 352{
 353	with_runs((struct dm_space_map *) context,
 354		  (const __le64 *) value_le, count, dm_sm_inc_blocks);
 355}
 
 356
 357static void data_block_dec(void *context, const void *value_le, unsigned count)
 358{
 359	with_runs((struct dm_space_map *) context,
 360		  (const __le64 *) value_le, count, dm_sm_dec_blocks);
 361}
 362
 363static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 364{
 365	__le64 v1_le, v2_le;
 366	uint64_t b1, b2;
 367	uint32_t t;
 368
 369	memcpy(&v1_le, value1_le, sizeof(v1_le));
 370	memcpy(&v2_le, value2_le, sizeof(v2_le));
 371	unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 372	unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 373
 374	return b1 == b2;
 375}
 376
 377static void subtree_inc(void *context, const void *value, unsigned count)
 378{
 379	struct dm_btree_info *info = context;
 380	const __le64 *root_le = value;
 381	unsigned i;
 382
 383	for (i = 0; i < count; i++, root_le++)
 384		dm_tm_inc(info->tm, le64_to_cpu(*root_le));
 
 385}
 386
 387static void subtree_dec(void *context, const void *value, unsigned count)
 388{
 389	struct dm_btree_info *info = context;
 390	const __le64 *root_le = value;
 391	unsigned i;
 392
 393	for (i = 0; i < count; i++, root_le++)
 394		if (dm_btree_del(info, le64_to_cpu(*root_le)))
 395			DMERR("btree delete failed");
 
 396}
 397
 398static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 399{
 400	__le64 v1_le, v2_le;
 401	memcpy(&v1_le, value1_le, sizeof(v1_le));
 402	memcpy(&v2_le, value2_le, sizeof(v2_le));
 403
 404	return v1_le == v2_le;
 405}
 406
 407/*----------------------------------------------------------------*/
 408
 409/*
 410 * Variant that is used for in-core only changes or code that
 411 * shouldn't put the pool in service on its own (e.g. commit).
 412 */
 413static inline void pmd_write_lock_in_core(struct dm_pool_metadata *pmd)
 414	__acquires(pmd->root_lock)
 415{
 416	down_write(&pmd->root_lock);
 417}
 
 418
 419static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
 420{
 421	pmd_write_lock_in_core(pmd);
 422	if (unlikely(!pmd->in_service))
 423		pmd->in_service = true;
 424}
 425
 426static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
 427	__releases(pmd->root_lock)
 428{
 429	up_write(&pmd->root_lock);
 430}
 431
 432/*----------------------------------------------------------------*/
 433
 434static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 435				struct dm_block **sblock)
 436{
 437	return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 438				     &sb_validator, sblock);
 439}
 440
 441static int superblock_lock(struct dm_pool_metadata *pmd,
 442			   struct dm_block **sblock)
 443{
 444	return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 445				&sb_validator, sblock);
 446}
 447
 448static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 449{
 450	int r;
 451	unsigned i;
 452	struct dm_block *b;
 453	__le64 *data_le, zero = cpu_to_le64(0);
 454	unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 455
 456	/*
 457	 * We can't use a validator here - it may be all zeroes.
 458	 */
 459	r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 460	if (r)
 461		return r;
 462
 463	data_le = dm_block_data(b);
 464	*result = 1;
 465	for (i = 0; i < block_size; i++) {
 466		if (data_le[i] != zero) {
 467			*result = 0;
 468			break;
 469		}
 470	}
 471
 472	dm_bm_unlock(b);
 473
 474	return 0;
 475}
 476
 477static void __setup_btree_details(struct dm_pool_metadata *pmd)
 478{
 479	pmd->info.tm = pmd->tm;
 480	pmd->info.levels = 2;
 481	pmd->info.value_type.context = pmd->data_sm;
 482	pmd->info.value_type.size = sizeof(__le64);
 483	pmd->info.value_type.inc = data_block_inc;
 484	pmd->info.value_type.dec = data_block_dec;
 485	pmd->info.value_type.equal = data_block_equal;
 486
 487	memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 488	pmd->nb_info.tm = pmd->nb_tm;
 489
 490	pmd->tl_info.tm = pmd->tm;
 491	pmd->tl_info.levels = 1;
 492	pmd->tl_info.value_type.context = &pmd->bl_info;
 493	pmd->tl_info.value_type.size = sizeof(__le64);
 494	pmd->tl_info.value_type.inc = subtree_inc;
 495	pmd->tl_info.value_type.dec = subtree_dec;
 496	pmd->tl_info.value_type.equal = subtree_equal;
 497
 498	pmd->bl_info.tm = pmd->tm;
 499	pmd->bl_info.levels = 1;
 500	pmd->bl_info.value_type.context = pmd->data_sm;
 501	pmd->bl_info.value_type.size = sizeof(__le64);
 502	pmd->bl_info.value_type.inc = data_block_inc;
 503	pmd->bl_info.value_type.dec = data_block_dec;
 504	pmd->bl_info.value_type.equal = data_block_equal;
 505
 506	pmd->details_info.tm = pmd->tm;
 507	pmd->details_info.levels = 1;
 508	pmd->details_info.value_type.context = NULL;
 509	pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 510	pmd->details_info.value_type.inc = NULL;
 511	pmd->details_info.value_type.dec = NULL;
 512	pmd->details_info.value_type.equal = NULL;
 513}
 514
 515static int save_sm_roots(struct dm_pool_metadata *pmd)
 516{
 517	int r;
 518	size_t len;
 519
 520	r = dm_sm_root_size(pmd->metadata_sm, &len);
 521	if (r < 0)
 522		return r;
 523
 524	r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 525	if (r < 0)
 526		return r;
 527
 528	r = dm_sm_root_size(pmd->data_sm, &len);
 529	if (r < 0)
 530		return r;
 531
 532	return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 533}
 534
 535static void copy_sm_roots(struct dm_pool_metadata *pmd,
 536			  struct thin_disk_superblock *disk)
 537{
 538	memcpy(&disk->metadata_space_map_root,
 539	       &pmd->metadata_space_map_root,
 540	       sizeof(pmd->metadata_space_map_root));
 541
 542	memcpy(&disk->data_space_map_root,
 543	       &pmd->data_space_map_root,
 544	       sizeof(pmd->data_space_map_root));
 545}
 546
 547static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 548{
 549	int r;
 550	struct dm_block *sblock;
 551	struct thin_disk_superblock *disk_super;
 552	sector_t bdev_size = bdev_nr_sectors(pmd->bdev);
 553
 554	if (bdev_size > THIN_METADATA_MAX_SECTORS)
 555		bdev_size = THIN_METADATA_MAX_SECTORS;
 556
 557	r = dm_sm_commit(pmd->data_sm);
 558	if (r < 0)
 559		return r;
 560
 561	r = dm_tm_pre_commit(pmd->tm);
 562	if (r < 0)
 563		return r;
 564
 565	r = save_sm_roots(pmd);
 566	if (r < 0)
 567		return r;
 568
 569	r = superblock_lock_zero(pmd, &sblock);
 570	if (r)
 571		return r;
 572
 573	disk_super = dm_block_data(sblock);
 574	disk_super->flags = 0;
 575	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 576	disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 577	disk_super->version = cpu_to_le32(THIN_VERSION);
 578	disk_super->time = 0;
 579	disk_super->trans_id = 0;
 580	disk_super->held_root = 0;
 581
 582	copy_sm_roots(pmd, disk_super);
 583
 584	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 585	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 586	disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 587	disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 588	disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 589
 590	return dm_tm_commit(pmd->tm, sblock);
 591}
 592
 593static int __format_metadata(struct dm_pool_metadata *pmd)
 594{
 595	int r;
 596
 597	r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 598				 &pmd->tm, &pmd->metadata_sm);
 599	if (r < 0) {
 600		DMERR("tm_create_with_sm failed");
 601		return r;
 602	}
 603
 604	pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 605	if (IS_ERR(pmd->data_sm)) {
 606		DMERR("sm_disk_create failed");
 607		r = PTR_ERR(pmd->data_sm);
 608		goto bad_cleanup_tm;
 609	}
 610
 611	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 612	if (!pmd->nb_tm) {
 613		DMERR("could not create non-blocking clone tm");
 614		r = -ENOMEM;
 615		goto bad_cleanup_data_sm;
 616	}
 617
 618	__setup_btree_details(pmd);
 619
 620	r = dm_btree_empty(&pmd->info, &pmd->root);
 621	if (r < 0)
 622		goto bad_cleanup_nb_tm;
 623
 624	r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 625	if (r < 0) {
 626		DMERR("couldn't create devices root");
 627		goto bad_cleanup_nb_tm;
 628	}
 629
 630	r = __write_initial_superblock(pmd);
 631	if (r)
 632		goto bad_cleanup_nb_tm;
 633
 634	return 0;
 635
 636bad_cleanup_nb_tm:
 637	dm_tm_destroy(pmd->nb_tm);
 638bad_cleanup_data_sm:
 639	dm_sm_destroy(pmd->data_sm);
 640bad_cleanup_tm:
 641	dm_tm_destroy(pmd->tm);
 642	dm_sm_destroy(pmd->metadata_sm);
 643
 644	return r;
 645}
 646
 647static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 648				     struct dm_pool_metadata *pmd)
 649{
 650	uint32_t features;
 651
 652	features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 653	if (features) {
 654		DMERR("could not access metadata due to unsupported optional features (%lx).",
 655		      (unsigned long)features);
 656		return -EINVAL;
 657	}
 658
 659	/*
 660	 * Check for read-only metadata to skip the following RDWR checks.
 661	 */
 662	if (bdev_read_only(pmd->bdev))
 663		return 0;
 664
 665	features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 666	if (features) {
 667		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 668		      (unsigned long)features);
 669		return -EINVAL;
 670	}
 671
 672	return 0;
 673}
 674
 675static int __open_metadata(struct dm_pool_metadata *pmd)
 676{
 677	int r;
 678	struct dm_block *sblock;
 679	struct thin_disk_superblock *disk_super;
 680
 681	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 682			    &sb_validator, &sblock);
 683	if (r < 0) {
 684		DMERR("couldn't read superblock");
 685		return r;
 686	}
 687
 688	disk_super = dm_block_data(sblock);
 689
 690	/* Verify the data block size hasn't changed */
 691	if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 692		DMERR("changing the data block size (from %u to %llu) is not supported",
 693		      le32_to_cpu(disk_super->data_block_size),
 694		      (unsigned long long)pmd->data_block_size);
 695		r = -EINVAL;
 696		goto bad_unlock_sblock;
 697	}
 698
 699	r = __check_incompat_features(disk_super, pmd);
 700	if (r < 0)
 701		goto bad_unlock_sblock;
 702
 703	r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 704			       disk_super->metadata_space_map_root,
 705			       sizeof(disk_super->metadata_space_map_root),
 706			       &pmd->tm, &pmd->metadata_sm);
 707	if (r < 0) {
 708		DMERR("tm_open_with_sm failed");
 709		goto bad_unlock_sblock;
 710	}
 711
 712	pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 713				       sizeof(disk_super->data_space_map_root));
 714	if (IS_ERR(pmd->data_sm)) {
 715		DMERR("sm_disk_open failed");
 716		r = PTR_ERR(pmd->data_sm);
 717		goto bad_cleanup_tm;
 718	}
 719
 720	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 721	if (!pmd->nb_tm) {
 722		DMERR("could not create non-blocking clone tm");
 723		r = -ENOMEM;
 724		goto bad_cleanup_data_sm;
 725	}
 726
 727	/*
 728	 * For pool metadata opening process, root setting is redundant
 729	 * because it will be set again in __begin_transaction(). But dm
 730	 * pool aborting process really needs to get last transaction's
 731	 * root to avoid accessing broken btree.
 732	 */
 733	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 734	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 735
 736	__setup_btree_details(pmd);
 737	dm_bm_unlock(sblock);
 738
 739	return 0;
 740
 741bad_cleanup_data_sm:
 742	dm_sm_destroy(pmd->data_sm);
 743bad_cleanup_tm:
 744	dm_tm_destroy(pmd->tm);
 745	dm_sm_destroy(pmd->metadata_sm);
 746bad_unlock_sblock:
 747	dm_bm_unlock(sblock);
 748
 749	return r;
 750}
 751
 752static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 753{
 754	int r, unformatted;
 755
 756	r = __superblock_all_zeroes(pmd->bm, &unformatted);
 757	if (r)
 758		return r;
 759
 760	if (unformatted)
 761		return format_device ? __format_metadata(pmd) : -EPERM;
 762
 763	return __open_metadata(pmd);
 764}
 765
 766static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 767{
 768	int r;
 769
 770	pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 771					  THIN_MAX_CONCURRENT_LOCKS);
 772	if (IS_ERR(pmd->bm)) {
 773		DMERR("could not create block manager");
 774		r = PTR_ERR(pmd->bm);
 775		pmd->bm = NULL;
 776		return r;
 777	}
 778
 779	r = __open_or_format_metadata(pmd, format_device);
 780	if (r) {
 781		dm_block_manager_destroy(pmd->bm);
 782		pmd->bm = NULL;
 783	}
 784
 785	return r;
 786}
 787
 788static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd,
 789					      bool destroy_bm)
 790{
 791	dm_sm_destroy(pmd->data_sm);
 792	dm_sm_destroy(pmd->metadata_sm);
 793	dm_tm_destroy(pmd->nb_tm);
 794	dm_tm_destroy(pmd->tm);
 795	if (destroy_bm)
 796		dm_block_manager_destroy(pmd->bm);
 797}
 798
 799static int __begin_transaction(struct dm_pool_metadata *pmd)
 800{
 801	int r;
 802	struct thin_disk_superblock *disk_super;
 803	struct dm_block *sblock;
 804
 805	/*
 806	 * We re-read the superblock every time.  Shouldn't need to do this
 807	 * really.
 808	 */
 809	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 810			    &sb_validator, &sblock);
 811	if (r)
 812		return r;
 813
 814	disk_super = dm_block_data(sblock);
 815	pmd->time = le32_to_cpu(disk_super->time);
 816	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 817	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 818	pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 819	pmd->flags = le32_to_cpu(disk_super->flags);
 820	pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 821
 822	dm_bm_unlock(sblock);
 823	return 0;
 824}
 825
 826static int __write_changed_details(struct dm_pool_metadata *pmd)
 827{
 828	int r;
 829	struct dm_thin_device *td, *tmp;
 830	struct disk_device_details details;
 831	uint64_t key;
 832
 833	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 834		if (!td->changed)
 835			continue;
 836
 837		key = td->id;
 838
 839		details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 840		details.transaction_id = cpu_to_le64(td->transaction_id);
 841		details.creation_time = cpu_to_le32(td->creation_time);
 842		details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 843		__dm_bless_for_disk(&details);
 844
 845		r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 846				    &key, &details, &pmd->details_root);
 847		if (r)
 848			return r;
 849
 850		if (td->open_count)
 851			td->changed = false;
 852		else {
 853			list_del(&td->list);
 854			kfree(td);
 855		}
 856	}
 857
 858	return 0;
 859}
 860
 861static int __commit_transaction(struct dm_pool_metadata *pmd)
 862{
 863	int r;
 864	struct thin_disk_superblock *disk_super;
 865	struct dm_block *sblock;
 866
 867	/*
 868	 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 869	 */
 870	BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 871	BUG_ON(!rwsem_is_locked(&pmd->root_lock));
 872
 873	if (unlikely(!pmd->in_service))
 874		return 0;
 875
 876	if (pmd->pre_commit_fn) {
 877		r = pmd->pre_commit_fn(pmd->pre_commit_context);
 878		if (r < 0) {
 879			DMERR("pre-commit callback failed");
 880			return r;
 881		}
 882	}
 883
 884	r = __write_changed_details(pmd);
 885	if (r < 0)
 886		return r;
 887
 888	r = dm_sm_commit(pmd->data_sm);
 889	if (r < 0)
 890		return r;
 891
 892	r = dm_tm_pre_commit(pmd->tm);
 893	if (r < 0)
 894		return r;
 895
 896	r = save_sm_roots(pmd);
 897	if (r < 0)
 898		return r;
 899
 900	r = superblock_lock(pmd, &sblock);
 901	if (r)
 902		return r;
 903
 904	disk_super = dm_block_data(sblock);
 905	disk_super->time = cpu_to_le32(pmd->time);
 906	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 907	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 908	disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 909	disk_super->flags = cpu_to_le32(pmd->flags);
 910
 911	copy_sm_roots(pmd, disk_super);
 912
 913	return dm_tm_commit(pmd->tm, sblock);
 914}
 915
 916static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 917{
 918	int r;
 919	dm_block_t total;
 920	dm_block_t max_blocks = 4096; /* 16M */
 921
 922	r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 923	if (r) {
 924		DMERR("could not get size of metadata device");
 925		pmd->metadata_reserve = max_blocks;
 926	} else
 927		pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 928}
 929
 930struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 931					       sector_t data_block_size,
 932					       bool format_device)
 933{
 934	int r;
 935	struct dm_pool_metadata *pmd;
 936
 937	pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 938	if (!pmd) {
 939		DMERR("could not allocate metadata struct");
 940		return ERR_PTR(-ENOMEM);
 941	}
 942
 943	init_rwsem(&pmd->root_lock);
 944	pmd->time = 0;
 945	INIT_LIST_HEAD(&pmd->thin_devices);
 946	pmd->fail_io = false;
 947	pmd->in_service = false;
 948	pmd->bdev = bdev;
 949	pmd->data_block_size = data_block_size;
 950	pmd->pre_commit_fn = NULL;
 951	pmd->pre_commit_context = NULL;
 952
 953	r = __create_persistent_data_objects(pmd, format_device);
 954	if (r) {
 955		kfree(pmd);
 956		return ERR_PTR(r);
 957	}
 958
 959	r = __begin_transaction(pmd);
 960	if (r < 0) {
 961		if (dm_pool_metadata_close(pmd) < 0)
 962			DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 963		return ERR_PTR(r);
 964	}
 965
 966	__set_metadata_reserve(pmd);
 967
 968	return pmd;
 969}
 970
 971int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 972{
 973	int r;
 974	unsigned open_devices = 0;
 975	struct dm_thin_device *td, *tmp;
 976
 977	down_read(&pmd->root_lock);
 978	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 979		if (td->open_count)
 980			open_devices++;
 981		else {
 982			list_del(&td->list);
 983			kfree(td);
 984		}
 985	}
 986	up_read(&pmd->root_lock);
 987
 988	if (open_devices) {
 989		DMERR("attempt to close pmd when %u device(s) are still open",
 990		       open_devices);
 991		return -EBUSY;
 992	}
 993
 994	pmd_write_lock_in_core(pmd);
 995	if (!pmd->fail_io && !dm_bm_is_read_only(pmd->bm)) {
 996		r = __commit_transaction(pmd);
 997		if (r < 0)
 998			DMWARN("%s: __commit_transaction() failed, error = %d",
 999			       __func__, r);
1000	}
1001	pmd_write_unlock(pmd);
1002	if (!pmd->fail_io)
1003		__destroy_persistent_data_objects(pmd, true);
1004
1005	kfree(pmd);
1006	return 0;
1007}
1008
1009/*
1010 * __open_device: Returns @td corresponding to device with id @dev,
1011 * creating it if @create is set and incrementing @td->open_count.
1012 * On failure, @td is undefined.
1013 */
1014static int __open_device(struct dm_pool_metadata *pmd,
1015			 dm_thin_id dev, int create,
1016			 struct dm_thin_device **td)
1017{
1018	int r, changed = 0;
1019	struct dm_thin_device *td2;
1020	uint64_t key = dev;
1021	struct disk_device_details details_le;
1022
1023	/*
1024	 * If the device is already open, return it.
1025	 */
1026	list_for_each_entry(td2, &pmd->thin_devices, list)
1027		if (td2->id == dev) {
1028			/*
1029			 * May not create an already-open device.
1030			 */
1031			if (create)
1032				return -EEXIST;
1033
1034			td2->open_count++;
1035			*td = td2;
1036			return 0;
1037		}
1038
1039	/*
1040	 * Check the device exists.
1041	 */
1042	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1043			    &key, &details_le);
1044	if (r) {
1045		if (r != -ENODATA || !create)
1046			return r;
1047
1048		/*
1049		 * Create new device.
1050		 */
1051		changed = 1;
1052		details_le.mapped_blocks = 0;
1053		details_le.transaction_id = cpu_to_le64(pmd->trans_id);
1054		details_le.creation_time = cpu_to_le32(pmd->time);
1055		details_le.snapshotted_time = cpu_to_le32(pmd->time);
1056	}
1057
1058	*td = kmalloc(sizeof(**td), GFP_NOIO);
1059	if (!*td)
1060		return -ENOMEM;
1061
1062	(*td)->pmd = pmd;
1063	(*td)->id = dev;
1064	(*td)->open_count = 1;
1065	(*td)->changed = changed;
1066	(*td)->aborted_with_changes = false;
1067	(*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1068	(*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1069	(*td)->creation_time = le32_to_cpu(details_le.creation_time);
1070	(*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1071
1072	list_add(&(*td)->list, &pmd->thin_devices);
1073
1074	return 0;
1075}
1076
1077static void __close_device(struct dm_thin_device *td)
1078{
1079	--td->open_count;
1080}
1081
1082static int __create_thin(struct dm_pool_metadata *pmd,
1083			 dm_thin_id dev)
1084{
1085	int r;
1086	dm_block_t dev_root;
1087	uint64_t key = dev;
 
1088	struct dm_thin_device *td;
1089	__le64 value;
1090
1091	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1092			    &key, NULL);
1093	if (!r)
1094		return -EEXIST;
1095
1096	/*
1097	 * Create an empty btree for the mappings.
1098	 */
1099	r = dm_btree_empty(&pmd->bl_info, &dev_root);
1100	if (r)
1101		return r;
1102
1103	/*
1104	 * Insert it into the main mapping tree.
1105	 */
1106	value = cpu_to_le64(dev_root);
1107	__dm_bless_for_disk(&value);
1108	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1109	if (r) {
1110		dm_btree_del(&pmd->bl_info, dev_root);
1111		return r;
1112	}
1113
1114	r = __open_device(pmd, dev, 1, &td);
1115	if (r) {
1116		dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1117		dm_btree_del(&pmd->bl_info, dev_root);
1118		return r;
1119	}
1120	__close_device(td);
1121
1122	return r;
1123}
1124
1125int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1126{
1127	int r = -EINVAL;
1128
1129	pmd_write_lock(pmd);
1130	if (!pmd->fail_io)
1131		r = __create_thin(pmd, dev);
1132	pmd_write_unlock(pmd);
1133
1134	return r;
1135}
1136
1137static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1138				  struct dm_thin_device *snap,
1139				  dm_thin_id origin, uint32_t time)
1140{
1141	int r;
1142	struct dm_thin_device *td;
1143
1144	r = __open_device(pmd, origin, 0, &td);
1145	if (r)
1146		return r;
1147
1148	td->changed = true;
1149	td->snapshotted_time = time;
1150
1151	snap->mapped_blocks = td->mapped_blocks;
1152	snap->snapshotted_time = time;
1153	__close_device(td);
1154
1155	return 0;
1156}
1157
1158static int __create_snap(struct dm_pool_metadata *pmd,
1159			 dm_thin_id dev, dm_thin_id origin)
1160{
1161	int r;
1162	dm_block_t origin_root;
1163	uint64_t key = origin, dev_key = dev;
1164	struct dm_thin_device *td;
 
1165	__le64 value;
1166
1167	/* check this device is unused */
1168	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1169			    &dev_key, NULL);
1170	if (!r)
1171		return -EEXIST;
1172
1173	/* find the mapping tree for the origin */
1174	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1175	if (r)
1176		return r;
1177	origin_root = le64_to_cpu(value);
1178
1179	/* clone the origin, an inc will do */
1180	dm_tm_inc(pmd->tm, origin_root);
1181
1182	/* insert into the main mapping tree */
1183	value = cpu_to_le64(origin_root);
1184	__dm_bless_for_disk(&value);
1185	key = dev;
1186	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1187	if (r) {
1188		dm_tm_dec(pmd->tm, origin_root);
1189		return r;
1190	}
1191
1192	pmd->time++;
1193
1194	r = __open_device(pmd, dev, 1, &td);
1195	if (r)
1196		goto bad;
1197
1198	r = __set_snapshot_details(pmd, td, origin, pmd->time);
1199	__close_device(td);
1200
1201	if (r)
1202		goto bad;
1203
1204	return 0;
1205
1206bad:
1207	dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1208	dm_btree_remove(&pmd->details_info, pmd->details_root,
1209			&key, &pmd->details_root);
1210	return r;
1211}
1212
1213int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1214				 dm_thin_id dev,
1215				 dm_thin_id origin)
1216{
1217	int r = -EINVAL;
1218
1219	pmd_write_lock(pmd);
1220	if (!pmd->fail_io)
1221		r = __create_snap(pmd, dev, origin);
1222	pmd_write_unlock(pmd);
1223
1224	return r;
1225}
1226
1227static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1228{
1229	int r;
1230	uint64_t key = dev;
1231	struct dm_thin_device *td;
1232
1233	/* TODO: failure should mark the transaction invalid */
1234	r = __open_device(pmd, dev, 0, &td);
1235	if (r)
1236		return r;
1237
1238	if (td->open_count > 1) {
1239		__close_device(td);
1240		return -EBUSY;
1241	}
1242
1243	list_del(&td->list);
1244	kfree(td);
1245	r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1246			    &key, &pmd->details_root);
1247	if (r)
1248		return r;
1249
1250	r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1251	if (r)
1252		return r;
1253
1254	return 0;
1255}
1256
1257int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1258			       dm_thin_id dev)
1259{
1260	int r = -EINVAL;
1261
1262	pmd_write_lock(pmd);
1263	if (!pmd->fail_io)
1264		r = __delete_device(pmd, dev);
1265	pmd_write_unlock(pmd);
1266
1267	return r;
1268}
1269
1270int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1271					uint64_t current_id,
1272					uint64_t new_id)
1273{
1274	int r = -EINVAL;
1275
1276	pmd_write_lock(pmd);
1277
1278	if (pmd->fail_io)
1279		goto out;
1280
1281	if (pmd->trans_id != current_id) {
1282		DMERR("mismatched transaction id");
1283		goto out;
1284	}
1285
1286	pmd->trans_id = new_id;
1287	r = 0;
1288
1289out:
1290	pmd_write_unlock(pmd);
1291
1292	return r;
1293}
1294
1295int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1296					uint64_t *result)
1297{
1298	int r = -EINVAL;
1299
1300	down_read(&pmd->root_lock);
1301	if (!pmd->fail_io) {
1302		*result = pmd->trans_id;
1303		r = 0;
1304	}
1305	up_read(&pmd->root_lock);
1306
1307	return r;
1308}
1309
1310static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1311{
1312	int r, inc;
1313	struct thin_disk_superblock *disk_super;
1314	struct dm_block *copy, *sblock;
1315	dm_block_t held_root;
1316
1317	/*
1318	 * We commit to ensure the btree roots which we increment in a
1319	 * moment are up to date.
1320	 */
1321	r = __commit_transaction(pmd);
1322	if (r < 0) {
1323		DMWARN("%s: __commit_transaction() failed, error = %d",
1324		       __func__, r);
1325		return r;
1326	}
1327
1328	/*
1329	 * Copy the superblock.
1330	 */
1331	dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1332	r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1333			       &sb_validator, &copy, &inc);
1334	if (r)
1335		return r;
1336
1337	BUG_ON(!inc);
1338
1339	held_root = dm_block_location(copy);
1340	disk_super = dm_block_data(copy);
1341
1342	if (le64_to_cpu(disk_super->held_root)) {
1343		DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1344
1345		dm_tm_dec(pmd->tm, held_root);
1346		dm_tm_unlock(pmd->tm, copy);
1347		return -EBUSY;
1348	}
1349
1350	/*
1351	 * Wipe the spacemap since we're not publishing this.
1352	 */
1353	memset(&disk_super->data_space_map_root, 0,
1354	       sizeof(disk_super->data_space_map_root));
1355	memset(&disk_super->metadata_space_map_root, 0,
1356	       sizeof(disk_super->metadata_space_map_root));
1357
1358	/*
1359	 * Increment the data structures that need to be preserved.
1360	 */
1361	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1362	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1363	dm_tm_unlock(pmd->tm, copy);
1364
1365	/*
1366	 * Write the held root into the superblock.
1367	 */
1368	r = superblock_lock(pmd, &sblock);
1369	if (r) {
1370		dm_tm_dec(pmd->tm, held_root);
1371		return r;
1372	}
1373
1374	disk_super = dm_block_data(sblock);
1375	disk_super->held_root = cpu_to_le64(held_root);
1376	dm_bm_unlock(sblock);
1377	return 0;
1378}
1379
1380int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1381{
1382	int r = -EINVAL;
1383
1384	pmd_write_lock(pmd);
1385	if (!pmd->fail_io)
1386		r = __reserve_metadata_snap(pmd);
1387	pmd_write_unlock(pmd);
1388
1389	return r;
1390}
1391
1392static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1393{
1394	int r;
1395	struct thin_disk_superblock *disk_super;
1396	struct dm_block *sblock, *copy;
1397	dm_block_t held_root;
1398
1399	r = superblock_lock(pmd, &sblock);
1400	if (r)
1401		return r;
1402
1403	disk_super = dm_block_data(sblock);
1404	held_root = le64_to_cpu(disk_super->held_root);
1405	disk_super->held_root = cpu_to_le64(0);
1406
1407	dm_bm_unlock(sblock);
1408
1409	if (!held_root) {
1410		DMWARN("No pool metadata snapshot found: nothing to release.");
1411		return -EINVAL;
1412	}
1413
1414	r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1415	if (r)
1416		return r;
1417
1418	disk_super = dm_block_data(copy);
1419	dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1420	dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1421	dm_sm_dec_block(pmd->metadata_sm, held_root);
1422
1423	dm_tm_unlock(pmd->tm, copy);
1424
1425	return 0;
1426}
1427
1428int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1429{
1430	int r = -EINVAL;
1431
1432	pmd_write_lock(pmd);
1433	if (!pmd->fail_io)
1434		r = __release_metadata_snap(pmd);
1435	pmd_write_unlock(pmd);
1436
1437	return r;
1438}
1439
1440static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1441			       dm_block_t *result)
1442{
1443	int r;
1444	struct thin_disk_superblock *disk_super;
1445	struct dm_block *sblock;
1446
1447	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1448			    &sb_validator, &sblock);
1449	if (r)
1450		return r;
1451
1452	disk_super = dm_block_data(sblock);
1453	*result = le64_to_cpu(disk_super->held_root);
1454
1455	dm_bm_unlock(sblock);
1456
1457	return 0;
1458}
1459
1460int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1461			      dm_block_t *result)
1462{
1463	int r = -EINVAL;
1464
1465	down_read(&pmd->root_lock);
1466	if (!pmd->fail_io)
1467		r = __get_metadata_snap(pmd, result);
1468	up_read(&pmd->root_lock);
1469
1470	return r;
1471}
1472
1473int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1474			     struct dm_thin_device **td)
1475{
1476	int r = -EINVAL;
1477
1478	pmd_write_lock_in_core(pmd);
1479	if (!pmd->fail_io)
1480		r = __open_device(pmd, dev, 0, td);
1481	pmd_write_unlock(pmd);
1482
1483	return r;
1484}
1485
1486int dm_pool_close_thin_device(struct dm_thin_device *td)
1487{
1488	pmd_write_lock_in_core(td->pmd);
1489	__close_device(td);
1490	pmd_write_unlock(td->pmd);
1491
1492	return 0;
1493}
1494
1495dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1496{
1497	return td->id;
1498}
1499
1500/*
1501 * Check whether @time (of block creation) is older than @td's last snapshot.
1502 * If so then the associated block is shared with the last snapshot device.
1503 * Any block on a device created *after* the device last got snapshotted is
1504 * necessarily not shared.
1505 */
1506static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1507{
1508	return td->snapshotted_time > time;
1509}
1510
1511static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1512				 struct dm_thin_lookup_result *result)
1513{
1514	uint64_t block_time = 0;
1515	dm_block_t exception_block;
1516	uint32_t exception_time;
1517
1518	block_time = le64_to_cpu(value);
1519	unpack_block_time(block_time, &exception_block, &exception_time);
1520	result->block = exception_block;
1521	result->shared = __snapshotted_since(td, exception_time);
1522}
1523
1524static int __find_block(struct dm_thin_device *td, dm_block_t block,
1525			int can_issue_io, struct dm_thin_lookup_result *result)
1526{
1527	int r;
1528	__le64 value;
1529	struct dm_pool_metadata *pmd = td->pmd;
1530	dm_block_t keys[2] = { td->id, block };
1531	struct dm_btree_info *info;
1532
1533	if (can_issue_io) {
1534		info = &pmd->info;
1535	} else
1536		info = &pmd->nb_info;
1537
1538	r = dm_btree_lookup(info, pmd->root, keys, &value);
1539	if (!r)
1540		unpack_lookup_result(td, value, result);
1541
1542	return r;
1543}
1544
1545int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1546		       int can_issue_io, struct dm_thin_lookup_result *result)
1547{
1548	int r;
1549	struct dm_pool_metadata *pmd = td->pmd;
1550
1551	down_read(&pmd->root_lock);
1552	if (pmd->fail_io) {
1553		up_read(&pmd->root_lock);
1554		return -EINVAL;
1555	}
1556
1557	r = __find_block(td, block, can_issue_io, result);
1558
1559	up_read(&pmd->root_lock);
1560	return r;
1561}
1562
1563static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1564					  dm_block_t *vblock,
1565					  struct dm_thin_lookup_result *result)
1566{
1567	int r;
1568	__le64 value;
1569	struct dm_pool_metadata *pmd = td->pmd;
1570	dm_block_t keys[2] = { td->id, block };
1571
1572	r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1573	if (!r)
1574		unpack_lookup_result(td, value, result);
1575
1576	return r;
1577}
1578
1579static int __find_mapped_range(struct dm_thin_device *td,
1580			       dm_block_t begin, dm_block_t end,
1581			       dm_block_t *thin_begin, dm_block_t *thin_end,
1582			       dm_block_t *pool_begin, bool *maybe_shared)
1583{
1584	int r;
1585	dm_block_t pool_end;
1586	struct dm_thin_lookup_result lookup;
1587
1588	if (end < begin)
1589		return -ENODATA;
1590
1591	r = __find_next_mapped_block(td, begin, &begin, &lookup);
1592	if (r)
1593		return r;
1594
1595	if (begin >= end)
1596		return -ENODATA;
1597
1598	*thin_begin = begin;
1599	*pool_begin = lookup.block;
1600	*maybe_shared = lookup.shared;
1601
1602	begin++;
1603	pool_end = *pool_begin + 1;
1604	while (begin != end) {
1605		r = __find_block(td, begin, true, &lookup);
1606		if (r) {
1607			if (r == -ENODATA)
1608				break;
1609			else
1610				return r;
1611		}
1612
1613		if ((lookup.block != pool_end) ||
1614		    (lookup.shared != *maybe_shared))
1615			break;
1616
1617		pool_end++;
1618		begin++;
1619	}
1620
1621	*thin_end = begin;
1622	return 0;
1623}
1624
1625int dm_thin_find_mapped_range(struct dm_thin_device *td,
1626			      dm_block_t begin, dm_block_t end,
1627			      dm_block_t *thin_begin, dm_block_t *thin_end,
1628			      dm_block_t *pool_begin, bool *maybe_shared)
1629{
1630	int r = -EINVAL;
1631	struct dm_pool_metadata *pmd = td->pmd;
1632
1633	down_read(&pmd->root_lock);
1634	if (!pmd->fail_io) {
1635		r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1636					pool_begin, maybe_shared);
1637	}
1638	up_read(&pmd->root_lock);
1639
1640	return r;
1641}
1642
1643static int __insert(struct dm_thin_device *td, dm_block_t block,
1644		    dm_block_t data_block)
1645{
1646	int r, inserted;
1647	__le64 value;
1648	struct dm_pool_metadata *pmd = td->pmd;
1649	dm_block_t keys[2] = { td->id, block };
1650
1651	value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1652	__dm_bless_for_disk(&value);
1653
1654	r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1655				   &pmd->root, &inserted);
1656	if (r)
1657		return r;
1658
1659	td->changed = true;
1660	if (inserted)
1661		td->mapped_blocks++;
1662
1663	return 0;
1664}
1665
1666int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1667			 dm_block_t data_block)
1668{
1669	int r = -EINVAL;
1670
1671	pmd_write_lock(td->pmd);
1672	if (!td->pmd->fail_io)
1673		r = __insert(td, block, data_block);
1674	pmd_write_unlock(td->pmd);
1675
1676	return r;
1677}
1678
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1679static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1680{
1681	int r;
1682	unsigned count, total_count = 0;
1683	struct dm_pool_metadata *pmd = td->pmd;
1684	dm_block_t keys[1] = { td->id };
1685	__le64 value;
1686	dm_block_t mapping_root;
1687
1688	/*
1689	 * Find the mapping tree
1690	 */
1691	r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1692	if (r)
1693		return r;
1694
1695	/*
1696	 * Remove from the mapping tree, taking care to inc the
1697	 * ref count so it doesn't get deleted.
1698	 */
1699	mapping_root = le64_to_cpu(value);
1700	dm_tm_inc(pmd->tm, mapping_root);
1701	r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1702	if (r)
1703		return r;
1704
1705	/*
1706	 * Remove leaves stops at the first unmapped entry, so we have to
1707	 * loop round finding mapped ranges.
1708	 */
1709	while (begin < end) {
1710		r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1711		if (r == -ENODATA)
1712			break;
1713
1714		if (r)
1715			return r;
1716
1717		if (begin >= end)
1718			break;
1719
1720		r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1721		if (r)
1722			return r;
1723
1724		total_count += count;
1725	}
1726
1727	td->mapped_blocks -= total_count;
1728	td->changed = true;
1729
1730	/*
1731	 * Reinsert the mapping tree.
1732	 */
1733	value = cpu_to_le64(mapping_root);
1734	__dm_bless_for_disk(&value);
1735	return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1736}
1737
 
 
 
 
 
 
 
 
 
 
 
 
1738int dm_thin_remove_range(struct dm_thin_device *td,
1739			 dm_block_t begin, dm_block_t end)
1740{
1741	int r = -EINVAL;
1742
1743	pmd_write_lock(td->pmd);
1744	if (!td->pmd->fail_io)
1745		r = __remove_range(td, begin, end);
1746	pmd_write_unlock(td->pmd);
1747
1748	return r;
1749}
1750
1751int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1752{
1753	int r;
1754	uint32_t ref_count;
1755
1756	down_read(&pmd->root_lock);
1757	r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1758	if (!r)
1759		*result = (ref_count > 1);
1760	up_read(&pmd->root_lock);
1761
1762	return r;
1763}
1764
1765int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1766{
1767	int r = 0;
1768
1769	pmd_write_lock(pmd);
1770	r = dm_sm_inc_blocks(pmd->data_sm, b, e);
 
 
 
 
1771	pmd_write_unlock(pmd);
1772
1773	return r;
1774}
1775
1776int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1777{
1778	int r = 0;
1779
1780	pmd_write_lock(pmd);
1781	r = dm_sm_dec_blocks(pmd->data_sm, b, e);
 
 
 
 
1782	pmd_write_unlock(pmd);
1783
1784	return r;
1785}
1786
1787bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1788{
1789	int r;
1790
1791	down_read(&td->pmd->root_lock);
1792	r = td->changed;
1793	up_read(&td->pmd->root_lock);
1794
1795	return r;
1796}
1797
1798bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1799{
1800	bool r = false;
1801	struct dm_thin_device *td, *tmp;
1802
1803	down_read(&pmd->root_lock);
1804	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1805		if (td->changed) {
1806			r = td->changed;
1807			break;
1808		}
1809	}
1810	up_read(&pmd->root_lock);
1811
1812	return r;
1813}
1814
1815bool dm_thin_aborted_changes(struct dm_thin_device *td)
1816{
1817	bool r;
1818
1819	down_read(&td->pmd->root_lock);
1820	r = td->aborted_with_changes;
1821	up_read(&td->pmd->root_lock);
1822
1823	return r;
1824}
1825
1826int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1827{
1828	int r = -EINVAL;
1829
1830	pmd_write_lock(pmd);
1831	if (!pmd->fail_io)
1832		r = dm_sm_new_block(pmd->data_sm, result);
1833	pmd_write_unlock(pmd);
1834
1835	return r;
1836}
1837
1838int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1839{
1840	int r = -EINVAL;
1841
1842	/*
1843	 * Care is taken to not have commit be what
1844	 * triggers putting the thin-pool in-service.
1845	 */
1846	pmd_write_lock_in_core(pmd);
1847	if (pmd->fail_io)
1848		goto out;
1849
1850	r = __commit_transaction(pmd);
1851	if (r < 0)
1852		goto out;
1853
1854	/*
1855	 * Open the next transaction.
1856	 */
1857	r = __begin_transaction(pmd);
1858out:
1859	pmd_write_unlock(pmd);
1860	return r;
1861}
1862
1863static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1864{
1865	struct dm_thin_device *td;
1866
1867	list_for_each_entry(td, &pmd->thin_devices, list)
1868		td->aborted_with_changes = td->changed;
1869}
1870
1871int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1872{
1873	int r = -EINVAL;
1874	struct dm_block_manager *old_bm = NULL, *new_bm = NULL;
1875
1876	/* fail_io is double-checked with pmd->root_lock held below */
1877	if (unlikely(pmd->fail_io))
1878		return r;
1879
1880	/*
1881	 * Replacement block manager (new_bm) is created and old_bm destroyed outside of
1882	 * pmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of
1883	 * shrinker associated with the block manager's bufio client vs pmd root_lock).
1884	 * - must take shrinker_rwsem without holding pmd->root_lock
1885	 */
1886	new_bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
1887					 THIN_MAX_CONCURRENT_LOCKS);
1888
1889	pmd_write_lock(pmd);
1890	if (pmd->fail_io) {
1891		pmd_write_unlock(pmd);
1892		goto out;
1893	}
1894
1895	__set_abort_with_changes_flags(pmd);
1896	__destroy_persistent_data_objects(pmd, false);
1897	old_bm = pmd->bm;
1898	if (IS_ERR(new_bm)) {
1899		DMERR("could not create block manager during abort");
1900		pmd->bm = NULL;
1901		r = PTR_ERR(new_bm);
1902		goto out_unlock;
1903	}
1904
1905	pmd->bm = new_bm;
1906	r = __open_or_format_metadata(pmd, false);
1907	if (r) {
1908		pmd->bm = NULL;
1909		goto out_unlock;
1910	}
1911	new_bm = NULL;
1912out_unlock:
1913	if (r)
1914		pmd->fail_io = true;
1915	pmd_write_unlock(pmd);
1916	dm_block_manager_destroy(old_bm);
1917out:
1918	if (new_bm && !IS_ERR(new_bm))
1919		dm_block_manager_destroy(new_bm);
1920
1921	return r;
1922}
1923
1924int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1925{
1926	int r = -EINVAL;
1927
1928	down_read(&pmd->root_lock);
1929	if (!pmd->fail_io)
1930		r = dm_sm_get_nr_free(pmd->data_sm, result);
1931	up_read(&pmd->root_lock);
1932
1933	return r;
1934}
1935
1936int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1937					  dm_block_t *result)
1938{
1939	int r = -EINVAL;
1940
1941	down_read(&pmd->root_lock);
1942	if (!pmd->fail_io)
1943		r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1944
1945	if (!r) {
1946		if (*result < pmd->metadata_reserve)
1947			*result = 0;
1948		else
1949			*result -= pmd->metadata_reserve;
1950	}
1951	up_read(&pmd->root_lock);
1952
1953	return r;
1954}
1955
1956int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1957				  dm_block_t *result)
1958{
1959	int r = -EINVAL;
1960
1961	down_read(&pmd->root_lock);
1962	if (!pmd->fail_io)
1963		r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1964	up_read(&pmd->root_lock);
1965
1966	return r;
1967}
1968
1969int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1970{
1971	int r = -EINVAL;
1972
1973	down_read(&pmd->root_lock);
1974	if (!pmd->fail_io)
1975		r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1976	up_read(&pmd->root_lock);
1977
1978	return r;
1979}
1980
1981int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1982{
1983	int r = -EINVAL;
1984	struct dm_pool_metadata *pmd = td->pmd;
1985
1986	down_read(&pmd->root_lock);
1987	if (!pmd->fail_io) {
1988		*result = td->mapped_blocks;
1989		r = 0;
1990	}
1991	up_read(&pmd->root_lock);
1992
1993	return r;
1994}
1995
1996static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1997{
1998	int r;
1999	__le64 value_le;
2000	dm_block_t thin_root;
2001	struct dm_pool_metadata *pmd = td->pmd;
2002
2003	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
2004	if (r)
2005		return r;
2006
2007	thin_root = le64_to_cpu(value_le);
2008
2009	return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
2010}
2011
2012int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
2013				     dm_block_t *result)
2014{
2015	int r = -EINVAL;
2016	struct dm_pool_metadata *pmd = td->pmd;
2017
2018	down_read(&pmd->root_lock);
2019	if (!pmd->fail_io)
2020		r = __highest_block(td, result);
2021	up_read(&pmd->root_lock);
2022
2023	return r;
2024}
2025
2026static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
2027{
2028	int r;
2029	dm_block_t old_count;
2030
2031	r = dm_sm_get_nr_blocks(sm, &old_count);
2032	if (r)
2033		return r;
2034
2035	if (new_count == old_count)
2036		return 0;
2037
2038	if (new_count < old_count) {
2039		DMERR("cannot reduce size of space map");
2040		return -EINVAL;
2041	}
2042
2043	return dm_sm_extend(sm, new_count - old_count);
2044}
2045
2046int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2047{
2048	int r = -EINVAL;
2049
2050	pmd_write_lock(pmd);
2051	if (!pmd->fail_io)
2052		r = __resize_space_map(pmd->data_sm, new_count);
2053	pmd_write_unlock(pmd);
2054
2055	return r;
2056}
2057
2058int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2059{
2060	int r = -EINVAL;
2061
2062	pmd_write_lock(pmd);
2063	if (!pmd->fail_io) {
2064		r = __resize_space_map(pmd->metadata_sm, new_count);
2065		if (!r)
2066			__set_metadata_reserve(pmd);
2067	}
2068	pmd_write_unlock(pmd);
2069
2070	return r;
2071}
2072
2073void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2074{
2075	pmd_write_lock_in_core(pmd);
2076	dm_bm_set_read_only(pmd->bm);
2077	pmd_write_unlock(pmd);
2078}
2079
2080void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2081{
2082	pmd_write_lock_in_core(pmd);
2083	dm_bm_set_read_write(pmd->bm);
2084	pmd_write_unlock(pmd);
2085}
2086
2087int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2088					dm_block_t threshold,
2089					dm_sm_threshold_fn fn,
2090					void *context)
2091{
2092	int r = -EINVAL;
2093
2094	pmd_write_lock_in_core(pmd);
2095	if (!pmd->fail_io) {
2096		r = dm_sm_register_threshold_callback(pmd->metadata_sm,
2097						      threshold, fn, context);
2098	}
2099	pmd_write_unlock(pmd);
2100
2101	return r;
2102}
2103
2104void dm_pool_register_pre_commit_callback(struct dm_pool_metadata *pmd,
2105					  dm_pool_pre_commit_fn fn,
2106					  void *context)
2107{
2108	pmd_write_lock_in_core(pmd);
2109	pmd->pre_commit_fn = fn;
2110	pmd->pre_commit_context = context;
2111	pmd_write_unlock(pmd);
2112}
2113
2114int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2115{
2116	int r = -EINVAL;
2117	struct dm_block *sblock;
2118	struct thin_disk_superblock *disk_super;
2119
2120	pmd_write_lock(pmd);
2121	if (pmd->fail_io)
2122		goto out;
2123
2124	pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2125
2126	r = superblock_lock(pmd, &sblock);
2127	if (r) {
2128		DMERR("couldn't lock superblock");
2129		goto out;
2130	}
2131
2132	disk_super = dm_block_data(sblock);
2133	disk_super->flags = cpu_to_le32(pmd->flags);
2134
2135	dm_bm_unlock(sblock);
2136out:
2137	pmd_write_unlock(pmd);
2138	return r;
2139}
2140
2141bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2142{
2143	bool needs_check;
2144
2145	down_read(&pmd->root_lock);
2146	needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2147	up_read(&pmd->root_lock);
2148
2149	return needs_check;
2150}
2151
2152void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2153{
2154	down_read(&pmd->root_lock);
2155	if (!pmd->fail_io)
2156		dm_tm_issue_prefetches(pmd->tm);
2157	up_read(&pmd->root_lock);
2158}
v5.4
   1/*
   2 * Copyright (C) 2011-2012 Red Hat, Inc.
   3 *
   4 * This file is released under the GPL.
   5 */
   6
   7#include "dm-thin-metadata.h"
   8#include "persistent-data/dm-btree.h"
   9#include "persistent-data/dm-space-map.h"
  10#include "persistent-data/dm-space-map-disk.h"
  11#include "persistent-data/dm-transaction-manager.h"
  12
  13#include <linux/list.h>
  14#include <linux/device-mapper.h>
  15#include <linux/workqueue.h>
  16
  17/*--------------------------------------------------------------------------
  18 * As far as the metadata goes, there is:
  19 *
  20 * - A superblock in block zero, taking up fewer than 512 bytes for
  21 *   atomic writes.
  22 *
  23 * - A space map managing the metadata blocks.
  24 *
  25 * - A space map managing the data blocks.
  26 *
  27 * - A btree mapping our internal thin dev ids onto struct disk_device_details.
  28 *
  29 * - A hierarchical btree, with 2 levels which effectively maps (thin
  30 *   dev id, virtual block) -> block_time.  Block time is a 64-bit
  31 *   field holding the time in the low 24 bits, and block in the top 48
  32 *   bits.
  33 *
  34 * BTrees consist solely of btree_nodes, that fill a block.  Some are
  35 * internal nodes, as such their values are a __le64 pointing to other
  36 * nodes.  Leaf nodes can store data of any reasonable size (ie. much
  37 * smaller than the block size).  The nodes consist of the header,
  38 * followed by an array of keys, followed by an array of values.  We have
  39 * to binary search on the keys so they're all held together to help the
  40 * cpu cache.
  41 *
  42 * Space maps have 2 btrees:
  43 *
  44 * - One maps a uint64_t onto a struct index_entry.  Which points to a
  45 *   bitmap block, and has some details about how many free entries there
  46 *   are etc.
  47 *
  48 * - The bitmap blocks have a header (for the checksum).  Then the rest
  49 *   of the block is pairs of bits.  With the meaning being:
  50 *
  51 *   0 - ref count is 0
  52 *   1 - ref count is 1
  53 *   2 - ref count is 2
  54 *   3 - ref count is higher than 2
  55 *
  56 * - If the count is higher than 2 then the ref count is entered in a
  57 *   second btree that directly maps the block_address to a uint32_t ref
  58 *   count.
  59 *
  60 * The space map metadata variant doesn't have a bitmaps btree.  Instead
  61 * it has one single blocks worth of index_entries.  This avoids
  62 * recursive issues with the bitmap btree needing to allocate space in
  63 * order to insert.  With a small data block size such as 64k the
  64 * metadata support data devices that are hundreds of terrabytes.
  65 *
  66 * The space maps allocate space linearly from front to back.  Space that
  67 * is freed in a transaction is never recycled within that transaction.
  68 * To try and avoid fragmenting _free_ space the allocator always goes
  69 * back and fills in gaps.
  70 *
  71 * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
  72 * from the block manager.
  73 *--------------------------------------------------------------------------*/
  74
  75#define DM_MSG_PREFIX   "thin metadata"
  76
  77#define THIN_SUPERBLOCK_MAGIC 27022010
  78#define THIN_SUPERBLOCK_LOCATION 0
  79#define THIN_VERSION 2
  80#define SECTOR_TO_BLOCK_SHIFT 3
  81
  82/*
  83 * For btree insert:
  84 *  3 for btree insert +
  85 *  2 for btree lookup used within space map
  86 * For btree remove:
  87 *  2 for shadow spine +
  88 *  4 for rebalance 3 child node
  89 */
  90#define THIN_MAX_CONCURRENT_LOCKS 6
  91
  92/* This should be plenty */
  93#define SPACE_MAP_ROOT_SIZE 128
  94
  95/*
  96 * Little endian on-disk superblock and device details.
  97 */
  98struct thin_disk_superblock {
  99	__le32 csum;	/* Checksum of superblock except for this field. */
 100	__le32 flags;
 101	__le64 blocknr;	/* This block number, dm_block_t. */
 102
 103	__u8 uuid[16];
 104	__le64 magic;
 105	__le32 version;
 106	__le32 time;
 107
 108	__le64 trans_id;
 109
 110	/*
 111	 * Root held by userspace transactions.
 112	 */
 113	__le64 held_root;
 114
 115	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 116	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 117
 118	/*
 119	 * 2-level btree mapping (dev_id, (dev block, time)) -> data block
 120	 */
 121	__le64 data_mapping_root;
 122
 123	/*
 124	 * Device detail root mapping dev_id -> device_details
 125	 */
 126	__le64 device_details_root;
 127
 128	__le32 data_block_size;		/* In 512-byte sectors. */
 129
 130	__le32 metadata_block_size;	/* In 512-byte sectors. */
 131	__le64 metadata_nr_blocks;
 132
 133	__le32 compat_flags;
 134	__le32 compat_ro_flags;
 135	__le32 incompat_flags;
 136} __packed;
 137
 138struct disk_device_details {
 139	__le64 mapped_blocks;
 140	__le64 transaction_id;		/* When created. */
 141	__le32 creation_time;
 142	__le32 snapshotted_time;
 143} __packed;
 144
 145struct dm_pool_metadata {
 146	struct hlist_node hash;
 147
 148	struct block_device *bdev;
 149	struct dm_block_manager *bm;
 150	struct dm_space_map *metadata_sm;
 151	struct dm_space_map *data_sm;
 152	struct dm_transaction_manager *tm;
 153	struct dm_transaction_manager *nb_tm;
 154
 155	/*
 156	 * Two-level btree.
 157	 * First level holds thin_dev_t.
 158	 * Second level holds mappings.
 159	 */
 160	struct dm_btree_info info;
 161
 162	/*
 163	 * Non-blocking version of the above.
 164	 */
 165	struct dm_btree_info nb_info;
 166
 167	/*
 168	 * Just the top level for deleting whole devices.
 169	 */
 170	struct dm_btree_info tl_info;
 171
 172	/*
 173	 * Just the bottom level for creating new devices.
 174	 */
 175	struct dm_btree_info bl_info;
 176
 177	/*
 178	 * Describes the device details btree.
 179	 */
 180	struct dm_btree_info details_info;
 181
 182	struct rw_semaphore root_lock;
 183	uint32_t time;
 184	dm_block_t root;
 185	dm_block_t details_root;
 186	struct list_head thin_devices;
 187	uint64_t trans_id;
 188	unsigned long flags;
 189	sector_t data_block_size;
 190
 191	/*
 
 
 
 
 
 
 
 
 
 192	 * We reserve a section of the metadata for commit overhead.
 193	 * All reported space does *not* include this.
 194	 */
 195	dm_block_t metadata_reserve;
 196
 197	/*
 198	 * Set if a transaction has to be aborted but the attempt to roll back
 199	 * to the previous (good) transaction failed.  The only pool metadata
 200	 * operation possible in this state is the closing of the device.
 201	 */
 202	bool fail_io:1;
 203
 204	/*
 205	 * Set once a thin-pool has been accessed through one of the interfaces
 206	 * that imply the pool is in-service (e.g. thin devices created/deleted,
 207	 * thin-pool message, metadata snapshots, etc).
 208	 */
 209	bool in_service:1;
 210
 211	/*
 212	 * Reading the space map roots can fail, so we read it into these
 213	 * buffers before the superblock is locked and updated.
 214	 */
 215	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
 216	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
 217};
 218
 219struct dm_thin_device {
 220	struct list_head list;
 221	struct dm_pool_metadata *pmd;
 222	dm_thin_id id;
 223
 224	int open_count;
 225	bool changed:1;
 226	bool aborted_with_changes:1;
 227	uint64_t mapped_blocks;
 228	uint64_t transaction_id;
 229	uint32_t creation_time;
 230	uint32_t snapshotted_time;
 231};
 232
 233/*----------------------------------------------------------------
 234 * superblock validator
 235 *--------------------------------------------------------------*/
 236
 237#define SUPERBLOCK_CSUM_XOR 160774
 238
 239static void sb_prepare_for_write(struct dm_block_validator *v,
 240				 struct dm_block *b,
 241				 size_t block_size)
 242{
 243	struct thin_disk_superblock *disk_super = dm_block_data(b);
 244
 245	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
 246	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 247						      block_size - sizeof(__le32),
 248						      SUPERBLOCK_CSUM_XOR));
 249}
 250
 251static int sb_check(struct dm_block_validator *v,
 252		    struct dm_block *b,
 253		    size_t block_size)
 254{
 255	struct thin_disk_superblock *disk_super = dm_block_data(b);
 256	__le32 csum_le;
 257
 258	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
 259		DMERR("sb_check failed: blocknr %llu: "
 260		      "wanted %llu", le64_to_cpu(disk_super->blocknr),
 261		      (unsigned long long)dm_block_location(b));
 262		return -ENOTBLK;
 263	}
 264
 265	if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
 266		DMERR("sb_check failed: magic %llu: "
 267		      "wanted %llu", le64_to_cpu(disk_super->magic),
 268		      (unsigned long long)THIN_SUPERBLOCK_MAGIC);
 269		return -EILSEQ;
 270	}
 271
 272	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
 273					     block_size - sizeof(__le32),
 274					     SUPERBLOCK_CSUM_XOR));
 275	if (csum_le != disk_super->csum) {
 276		DMERR("sb_check failed: csum %u: wanted %u",
 277		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
 278		return -EILSEQ;
 279	}
 280
 281	return 0;
 282}
 283
 284static struct dm_block_validator sb_validator = {
 285	.name = "superblock",
 286	.prepare_for_write = sb_prepare_for_write,
 287	.check = sb_check
 288};
 289
 290/*----------------------------------------------------------------
 291 * Methods for the btree value types
 292 *--------------------------------------------------------------*/
 293
 294static uint64_t pack_block_time(dm_block_t b, uint32_t t)
 295{
 296	return (b << 24) | t;
 297}
 298
 299static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
 300{
 301	*b = v >> 24;
 302	*t = v & ((1 << 24) - 1);
 303}
 304
 305static void data_block_inc(void *context, const void *value_le)
 
 
 
 
 
 
 
 306{
 307	struct dm_space_map *sm = context;
 308	__le64 v_le;
 309	uint64_t b;
 310	uint32_t t;
 
 
 311
 312	memcpy(&v_le, value_le, sizeof(v_le));
 313	unpack_block_time(le64_to_cpu(v_le), &b, &t);
 314	dm_sm_inc_block(sm, b);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 315}
 316
 317static void data_block_dec(void *context, const void *value_le)
 318{
 319	struct dm_space_map *sm = context;
 320	__le64 v_le;
 321	uint64_t b;
 322	uint32_t t;
 323
 324	memcpy(&v_le, value_le, sizeof(v_le));
 325	unpack_block_time(le64_to_cpu(v_le), &b, &t);
 326	dm_sm_dec_block(sm, b);
 
 327}
 328
 329static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
 330{
 331	__le64 v1_le, v2_le;
 332	uint64_t b1, b2;
 333	uint32_t t;
 334
 335	memcpy(&v1_le, value1_le, sizeof(v1_le));
 336	memcpy(&v2_le, value2_le, sizeof(v2_le));
 337	unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
 338	unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
 339
 340	return b1 == b2;
 341}
 342
 343static void subtree_inc(void *context, const void *value)
 344{
 345	struct dm_btree_info *info = context;
 346	__le64 root_le;
 347	uint64_t root;
 348
 349	memcpy(&root_le, value, sizeof(root_le));
 350	root = le64_to_cpu(root_le);
 351	dm_tm_inc(info->tm, root);
 352}
 353
 354static void subtree_dec(void *context, const void *value)
 355{
 356	struct dm_btree_info *info = context;
 357	__le64 root_le;
 358	uint64_t root;
 359
 360	memcpy(&root_le, value, sizeof(root_le));
 361	root = le64_to_cpu(root_le);
 362	if (dm_btree_del(info, root))
 363		DMERR("btree delete failed");
 364}
 365
 366static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
 367{
 368	__le64 v1_le, v2_le;
 369	memcpy(&v1_le, value1_le, sizeof(v1_le));
 370	memcpy(&v2_le, value2_le, sizeof(v2_le));
 371
 372	return v1_le == v2_le;
 373}
 374
 375/*----------------------------------------------------------------*/
 376
 377/*
 378 * Variant that is used for in-core only changes or code that
 379 * shouldn't put the pool in service on its own (e.g. commit).
 380 */
 381static inline void __pmd_write_lock(struct dm_pool_metadata *pmd)
 382	__acquires(pmd->root_lock)
 383{
 384	down_write(&pmd->root_lock);
 385}
 386#define pmd_write_lock_in_core(pmd) __pmd_write_lock((pmd))
 387
 388static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
 389{
 390	__pmd_write_lock(pmd);
 391	if (unlikely(!pmd->in_service))
 392		pmd->in_service = true;
 393}
 394
 395static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
 396	__releases(pmd->root_lock)
 397{
 398	up_write(&pmd->root_lock);
 399}
 400
 401/*----------------------------------------------------------------*/
 402
 403static int superblock_lock_zero(struct dm_pool_metadata *pmd,
 404				struct dm_block **sblock)
 405{
 406	return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 407				     &sb_validator, sblock);
 408}
 409
 410static int superblock_lock(struct dm_pool_metadata *pmd,
 411			   struct dm_block **sblock)
 412{
 413	return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 414				&sb_validator, sblock);
 415}
 416
 417static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
 418{
 419	int r;
 420	unsigned i;
 421	struct dm_block *b;
 422	__le64 *data_le, zero = cpu_to_le64(0);
 423	unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
 424
 425	/*
 426	 * We can't use a validator here - it may be all zeroes.
 427	 */
 428	r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
 429	if (r)
 430		return r;
 431
 432	data_le = dm_block_data(b);
 433	*result = 1;
 434	for (i = 0; i < block_size; i++) {
 435		if (data_le[i] != zero) {
 436			*result = 0;
 437			break;
 438		}
 439	}
 440
 441	dm_bm_unlock(b);
 442
 443	return 0;
 444}
 445
 446static void __setup_btree_details(struct dm_pool_metadata *pmd)
 447{
 448	pmd->info.tm = pmd->tm;
 449	pmd->info.levels = 2;
 450	pmd->info.value_type.context = pmd->data_sm;
 451	pmd->info.value_type.size = sizeof(__le64);
 452	pmd->info.value_type.inc = data_block_inc;
 453	pmd->info.value_type.dec = data_block_dec;
 454	pmd->info.value_type.equal = data_block_equal;
 455
 456	memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
 457	pmd->nb_info.tm = pmd->nb_tm;
 458
 459	pmd->tl_info.tm = pmd->tm;
 460	pmd->tl_info.levels = 1;
 461	pmd->tl_info.value_type.context = &pmd->bl_info;
 462	pmd->tl_info.value_type.size = sizeof(__le64);
 463	pmd->tl_info.value_type.inc = subtree_inc;
 464	pmd->tl_info.value_type.dec = subtree_dec;
 465	pmd->tl_info.value_type.equal = subtree_equal;
 466
 467	pmd->bl_info.tm = pmd->tm;
 468	pmd->bl_info.levels = 1;
 469	pmd->bl_info.value_type.context = pmd->data_sm;
 470	pmd->bl_info.value_type.size = sizeof(__le64);
 471	pmd->bl_info.value_type.inc = data_block_inc;
 472	pmd->bl_info.value_type.dec = data_block_dec;
 473	pmd->bl_info.value_type.equal = data_block_equal;
 474
 475	pmd->details_info.tm = pmd->tm;
 476	pmd->details_info.levels = 1;
 477	pmd->details_info.value_type.context = NULL;
 478	pmd->details_info.value_type.size = sizeof(struct disk_device_details);
 479	pmd->details_info.value_type.inc = NULL;
 480	pmd->details_info.value_type.dec = NULL;
 481	pmd->details_info.value_type.equal = NULL;
 482}
 483
 484static int save_sm_roots(struct dm_pool_metadata *pmd)
 485{
 486	int r;
 487	size_t len;
 488
 489	r = dm_sm_root_size(pmd->metadata_sm, &len);
 490	if (r < 0)
 491		return r;
 492
 493	r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
 494	if (r < 0)
 495		return r;
 496
 497	r = dm_sm_root_size(pmd->data_sm, &len);
 498	if (r < 0)
 499		return r;
 500
 501	return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
 502}
 503
 504static void copy_sm_roots(struct dm_pool_metadata *pmd,
 505			  struct thin_disk_superblock *disk)
 506{
 507	memcpy(&disk->metadata_space_map_root,
 508	       &pmd->metadata_space_map_root,
 509	       sizeof(pmd->metadata_space_map_root));
 510
 511	memcpy(&disk->data_space_map_root,
 512	       &pmd->data_space_map_root,
 513	       sizeof(pmd->data_space_map_root));
 514}
 515
 516static int __write_initial_superblock(struct dm_pool_metadata *pmd)
 517{
 518	int r;
 519	struct dm_block *sblock;
 520	struct thin_disk_superblock *disk_super;
 521	sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
 522
 523	if (bdev_size > THIN_METADATA_MAX_SECTORS)
 524		bdev_size = THIN_METADATA_MAX_SECTORS;
 525
 526	r = dm_sm_commit(pmd->data_sm);
 527	if (r < 0)
 528		return r;
 529
 530	r = dm_tm_pre_commit(pmd->tm);
 531	if (r < 0)
 532		return r;
 533
 534	r = save_sm_roots(pmd);
 535	if (r < 0)
 536		return r;
 537
 538	r = superblock_lock_zero(pmd, &sblock);
 539	if (r)
 540		return r;
 541
 542	disk_super = dm_block_data(sblock);
 543	disk_super->flags = 0;
 544	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
 545	disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
 546	disk_super->version = cpu_to_le32(THIN_VERSION);
 547	disk_super->time = 0;
 548	disk_super->trans_id = 0;
 549	disk_super->held_root = 0;
 550
 551	copy_sm_roots(pmd, disk_super);
 552
 553	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 554	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 555	disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
 556	disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
 557	disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
 558
 559	return dm_tm_commit(pmd->tm, sblock);
 560}
 561
 562static int __format_metadata(struct dm_pool_metadata *pmd)
 563{
 564	int r;
 565
 566	r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 567				 &pmd->tm, &pmd->metadata_sm);
 568	if (r < 0) {
 569		DMERR("tm_create_with_sm failed");
 570		return r;
 571	}
 572
 573	pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
 574	if (IS_ERR(pmd->data_sm)) {
 575		DMERR("sm_disk_create failed");
 576		r = PTR_ERR(pmd->data_sm);
 577		goto bad_cleanup_tm;
 578	}
 579
 580	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 581	if (!pmd->nb_tm) {
 582		DMERR("could not create non-blocking clone tm");
 583		r = -ENOMEM;
 584		goto bad_cleanup_data_sm;
 585	}
 586
 587	__setup_btree_details(pmd);
 588
 589	r = dm_btree_empty(&pmd->info, &pmd->root);
 590	if (r < 0)
 591		goto bad_cleanup_nb_tm;
 592
 593	r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
 594	if (r < 0) {
 595		DMERR("couldn't create devices root");
 596		goto bad_cleanup_nb_tm;
 597	}
 598
 599	r = __write_initial_superblock(pmd);
 600	if (r)
 601		goto bad_cleanup_nb_tm;
 602
 603	return 0;
 604
 605bad_cleanup_nb_tm:
 606	dm_tm_destroy(pmd->nb_tm);
 607bad_cleanup_data_sm:
 608	dm_sm_destroy(pmd->data_sm);
 609bad_cleanup_tm:
 610	dm_tm_destroy(pmd->tm);
 611	dm_sm_destroy(pmd->metadata_sm);
 612
 613	return r;
 614}
 615
 616static int __check_incompat_features(struct thin_disk_superblock *disk_super,
 617				     struct dm_pool_metadata *pmd)
 618{
 619	uint32_t features;
 620
 621	features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
 622	if (features) {
 623		DMERR("could not access metadata due to unsupported optional features (%lx).",
 624		      (unsigned long)features);
 625		return -EINVAL;
 626	}
 627
 628	/*
 629	 * Check for read-only metadata to skip the following RDWR checks.
 630	 */
 631	if (get_disk_ro(pmd->bdev->bd_disk))
 632		return 0;
 633
 634	features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
 635	if (features) {
 636		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
 637		      (unsigned long)features);
 638		return -EINVAL;
 639	}
 640
 641	return 0;
 642}
 643
 644static int __open_metadata(struct dm_pool_metadata *pmd)
 645{
 646	int r;
 647	struct dm_block *sblock;
 648	struct thin_disk_superblock *disk_super;
 649
 650	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 651			    &sb_validator, &sblock);
 652	if (r < 0) {
 653		DMERR("couldn't read superblock");
 654		return r;
 655	}
 656
 657	disk_super = dm_block_data(sblock);
 658
 659	/* Verify the data block size hasn't changed */
 660	if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
 661		DMERR("changing the data block size (from %u to %llu) is not supported",
 662		      le32_to_cpu(disk_super->data_block_size),
 663		      (unsigned long long)pmd->data_block_size);
 664		r = -EINVAL;
 665		goto bad_unlock_sblock;
 666	}
 667
 668	r = __check_incompat_features(disk_super, pmd);
 669	if (r < 0)
 670		goto bad_unlock_sblock;
 671
 672	r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 673			       disk_super->metadata_space_map_root,
 674			       sizeof(disk_super->metadata_space_map_root),
 675			       &pmd->tm, &pmd->metadata_sm);
 676	if (r < 0) {
 677		DMERR("tm_open_with_sm failed");
 678		goto bad_unlock_sblock;
 679	}
 680
 681	pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
 682				       sizeof(disk_super->data_space_map_root));
 683	if (IS_ERR(pmd->data_sm)) {
 684		DMERR("sm_disk_open failed");
 685		r = PTR_ERR(pmd->data_sm);
 686		goto bad_cleanup_tm;
 687	}
 688
 689	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
 690	if (!pmd->nb_tm) {
 691		DMERR("could not create non-blocking clone tm");
 692		r = -ENOMEM;
 693		goto bad_cleanup_data_sm;
 694	}
 695
 
 
 
 
 
 
 
 
 
 696	__setup_btree_details(pmd);
 697	dm_bm_unlock(sblock);
 698
 699	return 0;
 700
 701bad_cleanup_data_sm:
 702	dm_sm_destroy(pmd->data_sm);
 703bad_cleanup_tm:
 704	dm_tm_destroy(pmd->tm);
 705	dm_sm_destroy(pmd->metadata_sm);
 706bad_unlock_sblock:
 707	dm_bm_unlock(sblock);
 708
 709	return r;
 710}
 711
 712static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
 713{
 714	int r, unformatted;
 715
 716	r = __superblock_all_zeroes(pmd->bm, &unformatted);
 717	if (r)
 718		return r;
 719
 720	if (unformatted)
 721		return format_device ? __format_metadata(pmd) : -EPERM;
 722
 723	return __open_metadata(pmd);
 724}
 725
 726static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
 727{
 728	int r;
 729
 730	pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
 731					  THIN_MAX_CONCURRENT_LOCKS);
 732	if (IS_ERR(pmd->bm)) {
 733		DMERR("could not create block manager");
 734		return PTR_ERR(pmd->bm);
 
 
 735	}
 736
 737	r = __open_or_format_metadata(pmd, format_device);
 738	if (r)
 739		dm_block_manager_destroy(pmd->bm);
 
 
 740
 741	return r;
 742}
 743
 744static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
 
 745{
 746	dm_sm_destroy(pmd->data_sm);
 747	dm_sm_destroy(pmd->metadata_sm);
 748	dm_tm_destroy(pmd->nb_tm);
 749	dm_tm_destroy(pmd->tm);
 750	dm_block_manager_destroy(pmd->bm);
 
 751}
 752
 753static int __begin_transaction(struct dm_pool_metadata *pmd)
 754{
 755	int r;
 756	struct thin_disk_superblock *disk_super;
 757	struct dm_block *sblock;
 758
 759	/*
 760	 * We re-read the superblock every time.  Shouldn't need to do this
 761	 * really.
 762	 */
 763	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
 764			    &sb_validator, &sblock);
 765	if (r)
 766		return r;
 767
 768	disk_super = dm_block_data(sblock);
 769	pmd->time = le32_to_cpu(disk_super->time);
 770	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
 771	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
 772	pmd->trans_id = le64_to_cpu(disk_super->trans_id);
 773	pmd->flags = le32_to_cpu(disk_super->flags);
 774	pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
 775
 776	dm_bm_unlock(sblock);
 777	return 0;
 778}
 779
 780static int __write_changed_details(struct dm_pool_metadata *pmd)
 781{
 782	int r;
 783	struct dm_thin_device *td, *tmp;
 784	struct disk_device_details details;
 785	uint64_t key;
 786
 787	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 788		if (!td->changed)
 789			continue;
 790
 791		key = td->id;
 792
 793		details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
 794		details.transaction_id = cpu_to_le64(td->transaction_id);
 795		details.creation_time = cpu_to_le32(td->creation_time);
 796		details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
 797		__dm_bless_for_disk(&details);
 798
 799		r = dm_btree_insert(&pmd->details_info, pmd->details_root,
 800				    &key, &details, &pmd->details_root);
 801		if (r)
 802			return r;
 803
 804		if (td->open_count)
 805			td->changed = 0;
 806		else {
 807			list_del(&td->list);
 808			kfree(td);
 809		}
 810	}
 811
 812	return 0;
 813}
 814
 815static int __commit_transaction(struct dm_pool_metadata *pmd)
 816{
 817	int r;
 818	struct thin_disk_superblock *disk_super;
 819	struct dm_block *sblock;
 820
 821	/*
 822	 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
 823	 */
 824	BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
 
 825
 826	if (unlikely(!pmd->in_service))
 827		return 0;
 828
 
 
 
 
 
 
 
 
 829	r = __write_changed_details(pmd);
 830	if (r < 0)
 831		return r;
 832
 833	r = dm_sm_commit(pmd->data_sm);
 834	if (r < 0)
 835		return r;
 836
 837	r = dm_tm_pre_commit(pmd->tm);
 838	if (r < 0)
 839		return r;
 840
 841	r = save_sm_roots(pmd);
 842	if (r < 0)
 843		return r;
 844
 845	r = superblock_lock(pmd, &sblock);
 846	if (r)
 847		return r;
 848
 849	disk_super = dm_block_data(sblock);
 850	disk_super->time = cpu_to_le32(pmd->time);
 851	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
 852	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
 853	disk_super->trans_id = cpu_to_le64(pmd->trans_id);
 854	disk_super->flags = cpu_to_le32(pmd->flags);
 855
 856	copy_sm_roots(pmd, disk_super);
 857
 858	return dm_tm_commit(pmd->tm, sblock);
 859}
 860
 861static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
 862{
 863	int r;
 864	dm_block_t total;
 865	dm_block_t max_blocks = 4096; /* 16M */
 866
 867	r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
 868	if (r) {
 869		DMERR("could not get size of metadata device");
 870		pmd->metadata_reserve = max_blocks;
 871	} else
 872		pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
 873}
 874
 875struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
 876					       sector_t data_block_size,
 877					       bool format_device)
 878{
 879	int r;
 880	struct dm_pool_metadata *pmd;
 881
 882	pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
 883	if (!pmd) {
 884		DMERR("could not allocate metadata struct");
 885		return ERR_PTR(-ENOMEM);
 886	}
 887
 888	init_rwsem(&pmd->root_lock);
 889	pmd->time = 0;
 890	INIT_LIST_HEAD(&pmd->thin_devices);
 891	pmd->fail_io = false;
 892	pmd->in_service = false;
 893	pmd->bdev = bdev;
 894	pmd->data_block_size = data_block_size;
 
 
 895
 896	r = __create_persistent_data_objects(pmd, format_device);
 897	if (r) {
 898		kfree(pmd);
 899		return ERR_PTR(r);
 900	}
 901
 902	r = __begin_transaction(pmd);
 903	if (r < 0) {
 904		if (dm_pool_metadata_close(pmd) < 0)
 905			DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
 906		return ERR_PTR(r);
 907	}
 908
 909	__set_metadata_reserve(pmd);
 910
 911	return pmd;
 912}
 913
 914int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
 915{
 916	int r;
 917	unsigned open_devices = 0;
 918	struct dm_thin_device *td, *tmp;
 919
 920	down_read(&pmd->root_lock);
 921	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
 922		if (td->open_count)
 923			open_devices++;
 924		else {
 925			list_del(&td->list);
 926			kfree(td);
 927		}
 928	}
 929	up_read(&pmd->root_lock);
 930
 931	if (open_devices) {
 932		DMERR("attempt to close pmd when %u device(s) are still open",
 933		       open_devices);
 934		return -EBUSY;
 935	}
 936
 937	if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
 
 938		r = __commit_transaction(pmd);
 939		if (r < 0)
 940			DMWARN("%s: __commit_transaction() failed, error = %d",
 941			       __func__, r);
 942	}
 
 943	if (!pmd->fail_io)
 944		__destroy_persistent_data_objects(pmd);
 945
 946	kfree(pmd);
 947	return 0;
 948}
 949
 950/*
 951 * __open_device: Returns @td corresponding to device with id @dev,
 952 * creating it if @create is set and incrementing @td->open_count.
 953 * On failure, @td is undefined.
 954 */
 955static int __open_device(struct dm_pool_metadata *pmd,
 956			 dm_thin_id dev, int create,
 957			 struct dm_thin_device **td)
 958{
 959	int r, changed = 0;
 960	struct dm_thin_device *td2;
 961	uint64_t key = dev;
 962	struct disk_device_details details_le;
 963
 964	/*
 965	 * If the device is already open, return it.
 966	 */
 967	list_for_each_entry(td2, &pmd->thin_devices, list)
 968		if (td2->id == dev) {
 969			/*
 970			 * May not create an already-open device.
 971			 */
 972			if (create)
 973				return -EEXIST;
 974
 975			td2->open_count++;
 976			*td = td2;
 977			return 0;
 978		}
 979
 980	/*
 981	 * Check the device exists.
 982	 */
 983	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
 984			    &key, &details_le);
 985	if (r) {
 986		if (r != -ENODATA || !create)
 987			return r;
 988
 989		/*
 990		 * Create new device.
 991		 */
 992		changed = 1;
 993		details_le.mapped_blocks = 0;
 994		details_le.transaction_id = cpu_to_le64(pmd->trans_id);
 995		details_le.creation_time = cpu_to_le32(pmd->time);
 996		details_le.snapshotted_time = cpu_to_le32(pmd->time);
 997	}
 998
 999	*td = kmalloc(sizeof(**td), GFP_NOIO);
1000	if (!*td)
1001		return -ENOMEM;
1002
1003	(*td)->pmd = pmd;
1004	(*td)->id = dev;
1005	(*td)->open_count = 1;
1006	(*td)->changed = changed;
1007	(*td)->aborted_with_changes = false;
1008	(*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1009	(*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1010	(*td)->creation_time = le32_to_cpu(details_le.creation_time);
1011	(*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1012
1013	list_add(&(*td)->list, &pmd->thin_devices);
1014
1015	return 0;
1016}
1017
1018static void __close_device(struct dm_thin_device *td)
1019{
1020	--td->open_count;
1021}
1022
1023static int __create_thin(struct dm_pool_metadata *pmd,
1024			 dm_thin_id dev)
1025{
1026	int r;
1027	dm_block_t dev_root;
1028	uint64_t key = dev;
1029	struct disk_device_details details_le;
1030	struct dm_thin_device *td;
1031	__le64 value;
1032
1033	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1034			    &key, &details_le);
1035	if (!r)
1036		return -EEXIST;
1037
1038	/*
1039	 * Create an empty btree for the mappings.
1040	 */
1041	r = dm_btree_empty(&pmd->bl_info, &dev_root);
1042	if (r)
1043		return r;
1044
1045	/*
1046	 * Insert it into the main mapping tree.
1047	 */
1048	value = cpu_to_le64(dev_root);
1049	__dm_bless_for_disk(&value);
1050	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1051	if (r) {
1052		dm_btree_del(&pmd->bl_info, dev_root);
1053		return r;
1054	}
1055
1056	r = __open_device(pmd, dev, 1, &td);
1057	if (r) {
1058		dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1059		dm_btree_del(&pmd->bl_info, dev_root);
1060		return r;
1061	}
1062	__close_device(td);
1063
1064	return r;
1065}
1066
1067int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1068{
1069	int r = -EINVAL;
1070
1071	pmd_write_lock(pmd);
1072	if (!pmd->fail_io)
1073		r = __create_thin(pmd, dev);
1074	pmd_write_unlock(pmd);
1075
1076	return r;
1077}
1078
1079static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1080				  struct dm_thin_device *snap,
1081				  dm_thin_id origin, uint32_t time)
1082{
1083	int r;
1084	struct dm_thin_device *td;
1085
1086	r = __open_device(pmd, origin, 0, &td);
1087	if (r)
1088		return r;
1089
1090	td->changed = 1;
1091	td->snapshotted_time = time;
1092
1093	snap->mapped_blocks = td->mapped_blocks;
1094	snap->snapshotted_time = time;
1095	__close_device(td);
1096
1097	return 0;
1098}
1099
1100static int __create_snap(struct dm_pool_metadata *pmd,
1101			 dm_thin_id dev, dm_thin_id origin)
1102{
1103	int r;
1104	dm_block_t origin_root;
1105	uint64_t key = origin, dev_key = dev;
1106	struct dm_thin_device *td;
1107	struct disk_device_details details_le;
1108	__le64 value;
1109
1110	/* check this device is unused */
1111	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1112			    &dev_key, &details_le);
1113	if (!r)
1114		return -EEXIST;
1115
1116	/* find the mapping tree for the origin */
1117	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1118	if (r)
1119		return r;
1120	origin_root = le64_to_cpu(value);
1121
1122	/* clone the origin, an inc will do */
1123	dm_tm_inc(pmd->tm, origin_root);
1124
1125	/* insert into the main mapping tree */
1126	value = cpu_to_le64(origin_root);
1127	__dm_bless_for_disk(&value);
1128	key = dev;
1129	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1130	if (r) {
1131		dm_tm_dec(pmd->tm, origin_root);
1132		return r;
1133	}
1134
1135	pmd->time++;
1136
1137	r = __open_device(pmd, dev, 1, &td);
1138	if (r)
1139		goto bad;
1140
1141	r = __set_snapshot_details(pmd, td, origin, pmd->time);
1142	__close_device(td);
1143
1144	if (r)
1145		goto bad;
1146
1147	return 0;
1148
1149bad:
1150	dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1151	dm_btree_remove(&pmd->details_info, pmd->details_root,
1152			&key, &pmd->details_root);
1153	return r;
1154}
1155
1156int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1157				 dm_thin_id dev,
1158				 dm_thin_id origin)
1159{
1160	int r = -EINVAL;
1161
1162	pmd_write_lock(pmd);
1163	if (!pmd->fail_io)
1164		r = __create_snap(pmd, dev, origin);
1165	pmd_write_unlock(pmd);
1166
1167	return r;
1168}
1169
1170static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1171{
1172	int r;
1173	uint64_t key = dev;
1174	struct dm_thin_device *td;
1175
1176	/* TODO: failure should mark the transaction invalid */
1177	r = __open_device(pmd, dev, 0, &td);
1178	if (r)
1179		return r;
1180
1181	if (td->open_count > 1) {
1182		__close_device(td);
1183		return -EBUSY;
1184	}
1185
1186	list_del(&td->list);
1187	kfree(td);
1188	r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1189			    &key, &pmd->details_root);
1190	if (r)
1191		return r;
1192
1193	r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1194	if (r)
1195		return r;
1196
1197	return 0;
1198}
1199
1200int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1201			       dm_thin_id dev)
1202{
1203	int r = -EINVAL;
1204
1205	pmd_write_lock(pmd);
1206	if (!pmd->fail_io)
1207		r = __delete_device(pmd, dev);
1208	pmd_write_unlock(pmd);
1209
1210	return r;
1211}
1212
1213int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1214					uint64_t current_id,
1215					uint64_t new_id)
1216{
1217	int r = -EINVAL;
1218
1219	pmd_write_lock(pmd);
1220
1221	if (pmd->fail_io)
1222		goto out;
1223
1224	if (pmd->trans_id != current_id) {
1225		DMERR("mismatched transaction id");
1226		goto out;
1227	}
1228
1229	pmd->trans_id = new_id;
1230	r = 0;
1231
1232out:
1233	pmd_write_unlock(pmd);
1234
1235	return r;
1236}
1237
1238int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1239					uint64_t *result)
1240{
1241	int r = -EINVAL;
1242
1243	down_read(&pmd->root_lock);
1244	if (!pmd->fail_io) {
1245		*result = pmd->trans_id;
1246		r = 0;
1247	}
1248	up_read(&pmd->root_lock);
1249
1250	return r;
1251}
1252
1253static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1254{
1255	int r, inc;
1256	struct thin_disk_superblock *disk_super;
1257	struct dm_block *copy, *sblock;
1258	dm_block_t held_root;
1259
1260	/*
1261	 * We commit to ensure the btree roots which we increment in a
1262	 * moment are up to date.
1263	 */
1264	r = __commit_transaction(pmd);
1265	if (r < 0) {
1266		DMWARN("%s: __commit_transaction() failed, error = %d",
1267		       __func__, r);
1268		return r;
1269	}
1270
1271	/*
1272	 * Copy the superblock.
1273	 */
1274	dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1275	r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1276			       &sb_validator, &copy, &inc);
1277	if (r)
1278		return r;
1279
1280	BUG_ON(!inc);
1281
1282	held_root = dm_block_location(copy);
1283	disk_super = dm_block_data(copy);
1284
1285	if (le64_to_cpu(disk_super->held_root)) {
1286		DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1287
1288		dm_tm_dec(pmd->tm, held_root);
1289		dm_tm_unlock(pmd->tm, copy);
1290		return -EBUSY;
1291	}
1292
1293	/*
1294	 * Wipe the spacemap since we're not publishing this.
1295	 */
1296	memset(&disk_super->data_space_map_root, 0,
1297	       sizeof(disk_super->data_space_map_root));
1298	memset(&disk_super->metadata_space_map_root, 0,
1299	       sizeof(disk_super->metadata_space_map_root));
1300
1301	/*
1302	 * Increment the data structures that need to be preserved.
1303	 */
1304	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1305	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1306	dm_tm_unlock(pmd->tm, copy);
1307
1308	/*
1309	 * Write the held root into the superblock.
1310	 */
1311	r = superblock_lock(pmd, &sblock);
1312	if (r) {
1313		dm_tm_dec(pmd->tm, held_root);
1314		return r;
1315	}
1316
1317	disk_super = dm_block_data(sblock);
1318	disk_super->held_root = cpu_to_le64(held_root);
1319	dm_bm_unlock(sblock);
1320	return 0;
1321}
1322
1323int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1324{
1325	int r = -EINVAL;
1326
1327	pmd_write_lock(pmd);
1328	if (!pmd->fail_io)
1329		r = __reserve_metadata_snap(pmd);
1330	pmd_write_unlock(pmd);
1331
1332	return r;
1333}
1334
1335static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1336{
1337	int r;
1338	struct thin_disk_superblock *disk_super;
1339	struct dm_block *sblock, *copy;
1340	dm_block_t held_root;
1341
1342	r = superblock_lock(pmd, &sblock);
1343	if (r)
1344		return r;
1345
1346	disk_super = dm_block_data(sblock);
1347	held_root = le64_to_cpu(disk_super->held_root);
1348	disk_super->held_root = cpu_to_le64(0);
1349
1350	dm_bm_unlock(sblock);
1351
1352	if (!held_root) {
1353		DMWARN("No pool metadata snapshot found: nothing to release.");
1354		return -EINVAL;
1355	}
1356
1357	r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1358	if (r)
1359		return r;
1360
1361	disk_super = dm_block_data(copy);
1362	dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1363	dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1364	dm_sm_dec_block(pmd->metadata_sm, held_root);
1365
1366	dm_tm_unlock(pmd->tm, copy);
1367
1368	return 0;
1369}
1370
1371int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1372{
1373	int r = -EINVAL;
1374
1375	pmd_write_lock(pmd);
1376	if (!pmd->fail_io)
1377		r = __release_metadata_snap(pmd);
1378	pmd_write_unlock(pmd);
1379
1380	return r;
1381}
1382
1383static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1384			       dm_block_t *result)
1385{
1386	int r;
1387	struct thin_disk_superblock *disk_super;
1388	struct dm_block *sblock;
1389
1390	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1391			    &sb_validator, &sblock);
1392	if (r)
1393		return r;
1394
1395	disk_super = dm_block_data(sblock);
1396	*result = le64_to_cpu(disk_super->held_root);
1397
1398	dm_bm_unlock(sblock);
1399
1400	return 0;
1401}
1402
1403int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1404			      dm_block_t *result)
1405{
1406	int r = -EINVAL;
1407
1408	down_read(&pmd->root_lock);
1409	if (!pmd->fail_io)
1410		r = __get_metadata_snap(pmd, result);
1411	up_read(&pmd->root_lock);
1412
1413	return r;
1414}
1415
1416int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1417			     struct dm_thin_device **td)
1418{
1419	int r = -EINVAL;
1420
1421	pmd_write_lock_in_core(pmd);
1422	if (!pmd->fail_io)
1423		r = __open_device(pmd, dev, 0, td);
1424	pmd_write_unlock(pmd);
1425
1426	return r;
1427}
1428
1429int dm_pool_close_thin_device(struct dm_thin_device *td)
1430{
1431	pmd_write_lock_in_core(td->pmd);
1432	__close_device(td);
1433	pmd_write_unlock(td->pmd);
1434
1435	return 0;
1436}
1437
1438dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1439{
1440	return td->id;
1441}
1442
1443/*
1444 * Check whether @time (of block creation) is older than @td's last snapshot.
1445 * If so then the associated block is shared with the last snapshot device.
1446 * Any block on a device created *after* the device last got snapshotted is
1447 * necessarily not shared.
1448 */
1449static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1450{
1451	return td->snapshotted_time > time;
1452}
1453
1454static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1455				 struct dm_thin_lookup_result *result)
1456{
1457	uint64_t block_time = 0;
1458	dm_block_t exception_block;
1459	uint32_t exception_time;
1460
1461	block_time = le64_to_cpu(value);
1462	unpack_block_time(block_time, &exception_block, &exception_time);
1463	result->block = exception_block;
1464	result->shared = __snapshotted_since(td, exception_time);
1465}
1466
1467static int __find_block(struct dm_thin_device *td, dm_block_t block,
1468			int can_issue_io, struct dm_thin_lookup_result *result)
1469{
1470	int r;
1471	__le64 value;
1472	struct dm_pool_metadata *pmd = td->pmd;
1473	dm_block_t keys[2] = { td->id, block };
1474	struct dm_btree_info *info;
1475
1476	if (can_issue_io) {
1477		info = &pmd->info;
1478	} else
1479		info = &pmd->nb_info;
1480
1481	r = dm_btree_lookup(info, pmd->root, keys, &value);
1482	if (!r)
1483		unpack_lookup_result(td, value, result);
1484
1485	return r;
1486}
1487
1488int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1489		       int can_issue_io, struct dm_thin_lookup_result *result)
1490{
1491	int r;
1492	struct dm_pool_metadata *pmd = td->pmd;
1493
1494	down_read(&pmd->root_lock);
1495	if (pmd->fail_io) {
1496		up_read(&pmd->root_lock);
1497		return -EINVAL;
1498	}
1499
1500	r = __find_block(td, block, can_issue_io, result);
1501
1502	up_read(&pmd->root_lock);
1503	return r;
1504}
1505
1506static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1507					  dm_block_t *vblock,
1508					  struct dm_thin_lookup_result *result)
1509{
1510	int r;
1511	__le64 value;
1512	struct dm_pool_metadata *pmd = td->pmd;
1513	dm_block_t keys[2] = { td->id, block };
1514
1515	r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1516	if (!r)
1517		unpack_lookup_result(td, value, result);
1518
1519	return r;
1520}
1521
1522static int __find_mapped_range(struct dm_thin_device *td,
1523			       dm_block_t begin, dm_block_t end,
1524			       dm_block_t *thin_begin, dm_block_t *thin_end,
1525			       dm_block_t *pool_begin, bool *maybe_shared)
1526{
1527	int r;
1528	dm_block_t pool_end;
1529	struct dm_thin_lookup_result lookup;
1530
1531	if (end < begin)
1532		return -ENODATA;
1533
1534	r = __find_next_mapped_block(td, begin, &begin, &lookup);
1535	if (r)
1536		return r;
1537
1538	if (begin >= end)
1539		return -ENODATA;
1540
1541	*thin_begin = begin;
1542	*pool_begin = lookup.block;
1543	*maybe_shared = lookup.shared;
1544
1545	begin++;
1546	pool_end = *pool_begin + 1;
1547	while (begin != end) {
1548		r = __find_block(td, begin, true, &lookup);
1549		if (r) {
1550			if (r == -ENODATA)
1551				break;
1552			else
1553				return r;
1554		}
1555
1556		if ((lookup.block != pool_end) ||
1557		    (lookup.shared != *maybe_shared))
1558			break;
1559
1560		pool_end++;
1561		begin++;
1562	}
1563
1564	*thin_end = begin;
1565	return 0;
1566}
1567
1568int dm_thin_find_mapped_range(struct dm_thin_device *td,
1569			      dm_block_t begin, dm_block_t end,
1570			      dm_block_t *thin_begin, dm_block_t *thin_end,
1571			      dm_block_t *pool_begin, bool *maybe_shared)
1572{
1573	int r = -EINVAL;
1574	struct dm_pool_metadata *pmd = td->pmd;
1575
1576	down_read(&pmd->root_lock);
1577	if (!pmd->fail_io) {
1578		r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1579					pool_begin, maybe_shared);
1580	}
1581	up_read(&pmd->root_lock);
1582
1583	return r;
1584}
1585
1586static int __insert(struct dm_thin_device *td, dm_block_t block,
1587		    dm_block_t data_block)
1588{
1589	int r, inserted;
1590	__le64 value;
1591	struct dm_pool_metadata *pmd = td->pmd;
1592	dm_block_t keys[2] = { td->id, block };
1593
1594	value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1595	__dm_bless_for_disk(&value);
1596
1597	r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1598				   &pmd->root, &inserted);
1599	if (r)
1600		return r;
1601
1602	td->changed = 1;
1603	if (inserted)
1604		td->mapped_blocks++;
1605
1606	return 0;
1607}
1608
1609int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1610			 dm_block_t data_block)
1611{
1612	int r = -EINVAL;
1613
1614	pmd_write_lock(td->pmd);
1615	if (!td->pmd->fail_io)
1616		r = __insert(td, block, data_block);
1617	pmd_write_unlock(td->pmd);
1618
1619	return r;
1620}
1621
1622static int __remove(struct dm_thin_device *td, dm_block_t block)
1623{
1624	int r;
1625	struct dm_pool_metadata *pmd = td->pmd;
1626	dm_block_t keys[2] = { td->id, block };
1627
1628	r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1629	if (r)
1630		return r;
1631
1632	td->mapped_blocks--;
1633	td->changed = 1;
1634
1635	return 0;
1636}
1637
1638static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1639{
1640	int r;
1641	unsigned count, total_count = 0;
1642	struct dm_pool_metadata *pmd = td->pmd;
1643	dm_block_t keys[1] = { td->id };
1644	__le64 value;
1645	dm_block_t mapping_root;
1646
1647	/*
1648	 * Find the mapping tree
1649	 */
1650	r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1651	if (r)
1652		return r;
1653
1654	/*
1655	 * Remove from the mapping tree, taking care to inc the
1656	 * ref count so it doesn't get deleted.
1657	 */
1658	mapping_root = le64_to_cpu(value);
1659	dm_tm_inc(pmd->tm, mapping_root);
1660	r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1661	if (r)
1662		return r;
1663
1664	/*
1665	 * Remove leaves stops at the first unmapped entry, so we have to
1666	 * loop round finding mapped ranges.
1667	 */
1668	while (begin < end) {
1669		r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1670		if (r == -ENODATA)
1671			break;
1672
1673		if (r)
1674			return r;
1675
1676		if (begin >= end)
1677			break;
1678
1679		r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1680		if (r)
1681			return r;
1682
1683		total_count += count;
1684	}
1685
1686	td->mapped_blocks -= total_count;
1687	td->changed = 1;
1688
1689	/*
1690	 * Reinsert the mapping tree.
1691	 */
1692	value = cpu_to_le64(mapping_root);
1693	__dm_bless_for_disk(&value);
1694	return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1695}
1696
1697int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1698{
1699	int r = -EINVAL;
1700
1701	pmd_write_lock(td->pmd);
1702	if (!td->pmd->fail_io)
1703		r = __remove(td, block);
1704	pmd_write_unlock(td->pmd);
1705
1706	return r;
1707}
1708
1709int dm_thin_remove_range(struct dm_thin_device *td,
1710			 dm_block_t begin, dm_block_t end)
1711{
1712	int r = -EINVAL;
1713
1714	pmd_write_lock(td->pmd);
1715	if (!td->pmd->fail_io)
1716		r = __remove_range(td, begin, end);
1717	pmd_write_unlock(td->pmd);
1718
1719	return r;
1720}
1721
1722int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1723{
1724	int r;
1725	uint32_t ref_count;
1726
1727	down_read(&pmd->root_lock);
1728	r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1729	if (!r)
1730		*result = (ref_count > 1);
1731	up_read(&pmd->root_lock);
1732
1733	return r;
1734}
1735
1736int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1737{
1738	int r = 0;
1739
1740	pmd_write_lock(pmd);
1741	for (; b != e; b++) {
1742		r = dm_sm_inc_block(pmd->data_sm, b);
1743		if (r)
1744			break;
1745	}
1746	pmd_write_unlock(pmd);
1747
1748	return r;
1749}
1750
1751int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1752{
1753	int r = 0;
1754
1755	pmd_write_lock(pmd);
1756	for (; b != e; b++) {
1757		r = dm_sm_dec_block(pmd->data_sm, b);
1758		if (r)
1759			break;
1760	}
1761	pmd_write_unlock(pmd);
1762
1763	return r;
1764}
1765
1766bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1767{
1768	int r;
1769
1770	down_read(&td->pmd->root_lock);
1771	r = td->changed;
1772	up_read(&td->pmd->root_lock);
1773
1774	return r;
1775}
1776
1777bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1778{
1779	bool r = false;
1780	struct dm_thin_device *td, *tmp;
1781
1782	down_read(&pmd->root_lock);
1783	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1784		if (td->changed) {
1785			r = td->changed;
1786			break;
1787		}
1788	}
1789	up_read(&pmd->root_lock);
1790
1791	return r;
1792}
1793
1794bool dm_thin_aborted_changes(struct dm_thin_device *td)
1795{
1796	bool r;
1797
1798	down_read(&td->pmd->root_lock);
1799	r = td->aborted_with_changes;
1800	up_read(&td->pmd->root_lock);
1801
1802	return r;
1803}
1804
1805int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1806{
1807	int r = -EINVAL;
1808
1809	pmd_write_lock(pmd);
1810	if (!pmd->fail_io)
1811		r = dm_sm_new_block(pmd->data_sm, result);
1812	pmd_write_unlock(pmd);
1813
1814	return r;
1815}
1816
1817int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1818{
1819	int r = -EINVAL;
1820
1821	/*
1822	 * Care is taken to not have commit be what
1823	 * triggers putting the thin-pool in-service.
1824	 */
1825	__pmd_write_lock(pmd);
1826	if (pmd->fail_io)
1827		goto out;
1828
1829	r = __commit_transaction(pmd);
1830	if (r < 0)
1831		goto out;
1832
1833	/*
1834	 * Open the next transaction.
1835	 */
1836	r = __begin_transaction(pmd);
1837out:
1838	pmd_write_unlock(pmd);
1839	return r;
1840}
1841
1842static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1843{
1844	struct dm_thin_device *td;
1845
1846	list_for_each_entry(td, &pmd->thin_devices, list)
1847		td->aborted_with_changes = td->changed;
1848}
1849
1850int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1851{
1852	int r = -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1853
1854	pmd_write_lock(pmd);
1855	if (pmd->fail_io)
 
1856		goto out;
 
1857
1858	__set_abort_with_changes_flags(pmd);
1859	__destroy_persistent_data_objects(pmd);
1860	r = __create_persistent_data_objects(pmd, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1861	if (r)
1862		pmd->fail_io = true;
1863
 
1864out:
1865	pmd_write_unlock(pmd);
 
1866
1867	return r;
1868}
1869
1870int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1871{
1872	int r = -EINVAL;
1873
1874	down_read(&pmd->root_lock);
1875	if (!pmd->fail_io)
1876		r = dm_sm_get_nr_free(pmd->data_sm, result);
1877	up_read(&pmd->root_lock);
1878
1879	return r;
1880}
1881
1882int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1883					  dm_block_t *result)
1884{
1885	int r = -EINVAL;
1886
1887	down_read(&pmd->root_lock);
1888	if (!pmd->fail_io)
1889		r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1890
1891	if (!r) {
1892		if (*result < pmd->metadata_reserve)
1893			*result = 0;
1894		else
1895			*result -= pmd->metadata_reserve;
1896	}
1897	up_read(&pmd->root_lock);
1898
1899	return r;
1900}
1901
1902int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1903				  dm_block_t *result)
1904{
1905	int r = -EINVAL;
1906
1907	down_read(&pmd->root_lock);
1908	if (!pmd->fail_io)
1909		r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1910	up_read(&pmd->root_lock);
1911
1912	return r;
1913}
1914
1915int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1916{
1917	int r = -EINVAL;
1918
1919	down_read(&pmd->root_lock);
1920	if (!pmd->fail_io)
1921		r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1922	up_read(&pmd->root_lock);
1923
1924	return r;
1925}
1926
1927int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1928{
1929	int r = -EINVAL;
1930	struct dm_pool_metadata *pmd = td->pmd;
1931
1932	down_read(&pmd->root_lock);
1933	if (!pmd->fail_io) {
1934		*result = td->mapped_blocks;
1935		r = 0;
1936	}
1937	up_read(&pmd->root_lock);
1938
1939	return r;
1940}
1941
1942static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1943{
1944	int r;
1945	__le64 value_le;
1946	dm_block_t thin_root;
1947	struct dm_pool_metadata *pmd = td->pmd;
1948
1949	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1950	if (r)
1951		return r;
1952
1953	thin_root = le64_to_cpu(value_le);
1954
1955	return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1956}
1957
1958int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1959				     dm_block_t *result)
1960{
1961	int r = -EINVAL;
1962	struct dm_pool_metadata *pmd = td->pmd;
1963
1964	down_read(&pmd->root_lock);
1965	if (!pmd->fail_io)
1966		r = __highest_block(td, result);
1967	up_read(&pmd->root_lock);
1968
1969	return r;
1970}
1971
1972static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1973{
1974	int r;
1975	dm_block_t old_count;
1976
1977	r = dm_sm_get_nr_blocks(sm, &old_count);
1978	if (r)
1979		return r;
1980
1981	if (new_count == old_count)
1982		return 0;
1983
1984	if (new_count < old_count) {
1985		DMERR("cannot reduce size of space map");
1986		return -EINVAL;
1987	}
1988
1989	return dm_sm_extend(sm, new_count - old_count);
1990}
1991
1992int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1993{
1994	int r = -EINVAL;
1995
1996	pmd_write_lock(pmd);
1997	if (!pmd->fail_io)
1998		r = __resize_space_map(pmd->data_sm, new_count);
1999	pmd_write_unlock(pmd);
2000
2001	return r;
2002}
2003
2004int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2005{
2006	int r = -EINVAL;
2007
2008	pmd_write_lock(pmd);
2009	if (!pmd->fail_io) {
2010		r = __resize_space_map(pmd->metadata_sm, new_count);
2011		if (!r)
2012			__set_metadata_reserve(pmd);
2013	}
2014	pmd_write_unlock(pmd);
2015
2016	return r;
2017}
2018
2019void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2020{
2021	pmd_write_lock_in_core(pmd);
2022	dm_bm_set_read_only(pmd->bm);
2023	pmd_write_unlock(pmd);
2024}
2025
2026void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2027{
2028	pmd_write_lock_in_core(pmd);
2029	dm_bm_set_read_write(pmd->bm);
2030	pmd_write_unlock(pmd);
2031}
2032
2033int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2034					dm_block_t threshold,
2035					dm_sm_threshold_fn fn,
2036					void *context)
2037{
2038	int r;
2039
2040	pmd_write_lock_in_core(pmd);
2041	r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
 
 
 
2042	pmd_write_unlock(pmd);
2043
2044	return r;
 
 
 
 
 
 
 
 
 
 
2045}
2046
2047int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2048{
2049	int r = -EINVAL;
2050	struct dm_block *sblock;
2051	struct thin_disk_superblock *disk_super;
2052
2053	pmd_write_lock(pmd);
2054	if (pmd->fail_io)
2055		goto out;
2056
2057	pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2058
2059	r = superblock_lock(pmd, &sblock);
2060	if (r) {
2061		DMERR("couldn't lock superblock");
2062		goto out;
2063	}
2064
2065	disk_super = dm_block_data(sblock);
2066	disk_super->flags = cpu_to_le32(pmd->flags);
2067
2068	dm_bm_unlock(sblock);
2069out:
2070	pmd_write_unlock(pmd);
2071	return r;
2072}
2073
2074bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2075{
2076	bool needs_check;
2077
2078	down_read(&pmd->root_lock);
2079	needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2080	up_read(&pmd->root_lock);
2081
2082	return needs_check;
2083}
2084
2085void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2086{
2087	down_read(&pmd->root_lock);
2088	if (!pmd->fail_io)
2089		dm_tm_issue_prefetches(pmd->tm);
2090	up_read(&pmd->root_lock);
2091}