Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   4 */
   5
   6#include <linux/blkdev.h>
   7#include <linux/ratelimit.h>
   8#include <linux/sched/mm.h>
   9#include <crypto/hash.h>
  10#include "ctree.h"
  11#include "discard.h"
  12#include "volumes.h"
  13#include "disk-io.h"
  14#include "ordered-data.h"
  15#include "transaction.h"
  16#include "backref.h"
  17#include "extent_io.h"
  18#include "dev-replace.h"
  19#include "check-integrity.h"
 
  20#include "raid56.h"
  21#include "block-group.h"
  22#include "zoned.h"
  23#include "fs.h"
  24#include "accessors.h"
  25#include "file-item.h"
  26#include "scrub.h"
  27
  28/*
  29 * This is only the first step towards a full-features scrub. It reads all
  30 * extent and super block and verifies the checksums. In case a bad checksum
  31 * is found or the extent cannot be read, good data will be written back if
  32 * any can be found.
  33 *
  34 * Future enhancements:
  35 *  - In case an unrepairable extent is encountered, track which files are
  36 *    affected and report them
  37 *  - track and record media errors, throw out bad devices
  38 *  - add a mode to also read unallocated space
  39 */
  40
  41struct scrub_block;
  42struct scrub_ctx;
  43
  44/*
  45 * The following three values only influence the performance.
  46 *
  47 * The last one configures the number of parallel and outstanding I/O
  48 * operations. The first one configures an upper limit for the number
  49 * of (dynamically allocated) pages that are added to a bio.
  50 */
  51#define SCRUB_SECTORS_PER_BIO	32	/* 128KiB per bio for 4KiB pages */
  52#define SCRUB_BIOS_PER_SCTX	64	/* 8MiB per device in flight for 4KiB pages */
 
  53
  54/*
  55 * The following value times PAGE_SIZE needs to be large enough to match the
  56 * largest node/leaf/sector size that shall be supported.
 
  57 */
  58#define SCRUB_MAX_SECTORS_PER_BLOCK	(BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
  59
  60#define SCRUB_MAX_PAGES			(DIV_ROUND_UP(BTRFS_MAX_METADATA_BLOCKSIZE, PAGE_SIZE))
  61
  62/*
  63 * Maximum number of mirrors that can be available for all profiles counting
  64 * the target device of dev-replace as one. During an active device replace
  65 * procedure, the target device of the copy operation is a mirror for the
  66 * filesystem data as well that can be used to read data in order to repair
  67 * read errors on other disks.
  68 *
  69 * Current value is derived from RAID1C4 with 4 copies.
  70 */
  71#define BTRFS_MAX_MIRRORS (4 + 1)
  72
  73struct scrub_recover {
  74	refcount_t		refs;
  75	struct btrfs_io_context	*bioc;
  76	u64			map_length;
  77};
  78
  79struct scrub_sector {
  80	struct scrub_block	*sblock;
 
 
  81	struct list_head	list;
  82	u64			flags;  /* extent flags */
  83	u64			generation;
  84	/* Offset in bytes to @sblock. */
  85	u32			offset;
 
  86	atomic_t		refs;
  87	unsigned int		have_csum:1;
  88	unsigned int		io_error:1;
 
 
 
  89	u8			csum[BTRFS_CSUM_SIZE];
  90
  91	struct scrub_recover	*recover;
  92};
  93
  94struct scrub_bio {
  95	int			index;
  96	struct scrub_ctx	*sctx;
  97	struct btrfs_device	*dev;
  98	struct bio		*bio;
  99	blk_status_t		status;
 100	u64			logical;
 101	u64			physical;
 102	struct scrub_sector	*sectors[SCRUB_SECTORS_PER_BIO];
 103	int			sector_count;
 
 
 
 
 104	int			next_free;
 105	struct work_struct	work;
 106};
 107
 108struct scrub_block {
 109	/*
 110	 * Each page will have its page::private used to record the logical
 111	 * bytenr.
 112	 */
 113	struct page		*pages[SCRUB_MAX_PAGES];
 114	struct scrub_sector	*sectors[SCRUB_MAX_SECTORS_PER_BLOCK];
 115	struct btrfs_device	*dev;
 116	/* Logical bytenr of the sblock */
 117	u64			logical;
 118	u64			physical;
 119	u64			physical_for_dev_replace;
 120	/* Length of sblock in bytes */
 121	u32			len;
 122	int			sector_count;
 123	int			mirror_num;
 124
 125	atomic_t		outstanding_sectors;
 126	refcount_t		refs; /* free mem on transition to zero */
 127	struct scrub_ctx	*sctx;
 128	struct scrub_parity	*sparity;
 129	struct {
 130		unsigned int	header_error:1;
 131		unsigned int	checksum_error:1;
 132		unsigned int	no_io_error_seen:1;
 133		unsigned int	generation_error:1; /* also sets header_error */
 134
 135		/* The following is for the data used to check parity */
 136		/* It is for the data with checksum */
 137		unsigned int	data_corrected:1;
 138	};
 139	struct work_struct	work;
 140};
 141
 142/* Used for the chunks with parity stripe such RAID5/6 */
 143struct scrub_parity {
 144	struct scrub_ctx	*sctx;
 145
 146	struct btrfs_device	*scrub_dev;
 147
 148	u64			logic_start;
 149
 150	u64			logic_end;
 151
 152	int			nsectors;
 153
 154	u32			stripe_len;
 155
 156	refcount_t		refs;
 157
 158	struct list_head	sectors_list;
 159
 160	/* Work of parity check and repair */
 161	struct work_struct	work;
 162
 163	/* Mark the parity blocks which have data */
 164	unsigned long		dbitmap;
 165
 166	/*
 167	 * Mark the parity blocks which have data, but errors happen when
 168	 * read data or check data
 169	 */
 170	unsigned long		ebitmap;
 
 
 171};
 172
 173struct scrub_ctx {
 174	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
 175	struct btrfs_fs_info	*fs_info;
 176	int			first_free;
 177	int			curr;
 178	atomic_t		bios_in_flight;
 179	atomic_t		workers_pending;
 180	spinlock_t		list_lock;
 181	wait_queue_head_t	list_wait;
 
 182	struct list_head	csum_list;
 183	atomic_t		cancel_req;
 184	int			readonly;
 185	int			sectors_per_bio;
 186
 187	/* State of IO submission throttling affecting the associated device */
 188	ktime_t			throttle_deadline;
 189	u64			throttle_sent;
 190
 191	int			is_dev_replace;
 192	u64			write_pointer;
 193
 194	struct scrub_bio        *wr_curr_bio;
 195	struct mutex            wr_lock;
 
 196	struct btrfs_device     *wr_tgtdev;
 197	bool                    flush_all_writes;
 198
 199	/*
 200	 * statistics
 201	 */
 202	struct btrfs_scrub_progress stat;
 203	spinlock_t		stat_lock;
 204
 205	/*
 206	 * Use a ref counter to avoid use-after-free issues. Scrub workers
 207	 * decrement bios_in_flight and workers_pending and then do a wakeup
 208	 * on the list_wait wait queue. We must ensure the main scrub task
 209	 * doesn't free the scrub context before or while the workers are
 210	 * doing the wakeup() call.
 211	 */
 212	refcount_t              refs;
 213};
 214
 215struct scrub_warning {
 216	struct btrfs_path	*path;
 217	u64			extent_item_size;
 218	const char		*errstr;
 219	u64			physical;
 220	u64			logical;
 221	struct btrfs_device	*dev;
 222};
 223
 224struct full_stripe_lock {
 225	struct rb_node node;
 226	u64 logical;
 227	u64 refs;
 228	struct mutex mutex;
 229};
 230
 231#ifndef CONFIG_64BIT
 232/* This structure is for archtectures whose (void *) is smaller than u64 */
 233struct scrub_page_private {
 234	u64 logical;
 235};
 236#endif
 237
 238static int attach_scrub_page_private(struct page *page, u64 logical)
 239{
 240#ifdef CONFIG_64BIT
 241	attach_page_private(page, (void *)logical);
 242	return 0;
 243#else
 244	struct scrub_page_private *spp;
 245
 246	spp = kmalloc(sizeof(*spp), GFP_KERNEL);
 247	if (!spp)
 248		return -ENOMEM;
 249	spp->logical = logical;
 250	attach_page_private(page, (void *)spp);
 251	return 0;
 252#endif
 253}
 254
 255static void detach_scrub_page_private(struct page *page)
 256{
 257#ifdef CONFIG_64BIT
 258	detach_page_private(page);
 259	return;
 260#else
 261	struct scrub_page_private *spp;
 262
 263	spp = detach_page_private(page);
 264	kfree(spp);
 265	return;
 266#endif
 267}
 268
 269static struct scrub_block *alloc_scrub_block(struct scrub_ctx *sctx,
 270					     struct btrfs_device *dev,
 271					     u64 logical, u64 physical,
 272					     u64 physical_for_dev_replace,
 273					     int mirror_num)
 274{
 275	struct scrub_block *sblock;
 276
 277	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
 278	if (!sblock)
 279		return NULL;
 280	refcount_set(&sblock->refs, 1);
 281	sblock->sctx = sctx;
 282	sblock->logical = logical;
 283	sblock->physical = physical;
 284	sblock->physical_for_dev_replace = physical_for_dev_replace;
 285	sblock->dev = dev;
 286	sblock->mirror_num = mirror_num;
 287	sblock->no_io_error_seen = 1;
 288	/*
 289	 * Scrub_block::pages will be allocated at alloc_scrub_sector() when
 290	 * the corresponding page is not allocated.
 291	 */
 292	return sblock;
 293}
 294
 295/*
 296 * Allocate a new scrub sector and attach it to @sblock.
 297 *
 298 * Will also allocate new pages for @sblock if needed.
 299 */
 300static struct scrub_sector *alloc_scrub_sector(struct scrub_block *sblock,
 301					       u64 logical)
 302{
 303	const pgoff_t page_index = (logical - sblock->logical) >> PAGE_SHIFT;
 304	struct scrub_sector *ssector;
 305
 306	/* We must never have scrub_block exceed U32_MAX in size. */
 307	ASSERT(logical - sblock->logical < U32_MAX);
 308
 309	ssector = kzalloc(sizeof(*ssector), GFP_KERNEL);
 310	if (!ssector)
 311		return NULL;
 312
 313	/* Allocate a new page if the slot is not allocated */
 314	if (!sblock->pages[page_index]) {
 315		int ret;
 316
 317		sblock->pages[page_index] = alloc_page(GFP_KERNEL);
 318		if (!sblock->pages[page_index]) {
 319			kfree(ssector);
 320			return NULL;
 321		}
 322		ret = attach_scrub_page_private(sblock->pages[page_index],
 323				sblock->logical + (page_index << PAGE_SHIFT));
 324		if (ret < 0) {
 325			kfree(ssector);
 326			__free_page(sblock->pages[page_index]);
 327			sblock->pages[page_index] = NULL;
 328			return NULL;
 329		}
 330	}
 331
 332	atomic_set(&ssector->refs, 1);
 333	ssector->sblock = sblock;
 334	/* The sector to be added should not be used */
 335	ASSERT(sblock->sectors[sblock->sector_count] == NULL);
 336	ssector->offset = logical - sblock->logical;
 337
 338	/* The sector count must be smaller than the limit */
 339	ASSERT(sblock->sector_count < SCRUB_MAX_SECTORS_PER_BLOCK);
 340
 341	sblock->sectors[sblock->sector_count] = ssector;
 342	sblock->sector_count++;
 343	sblock->len += sblock->sctx->fs_info->sectorsize;
 344
 345	return ssector;
 346}
 347
 348static struct page *scrub_sector_get_page(struct scrub_sector *ssector)
 349{
 350	struct scrub_block *sblock = ssector->sblock;
 351	pgoff_t index;
 352	/*
 353	 * When calling this function, ssector must be alreaday attached to the
 354	 * parent sblock.
 355	 */
 356	ASSERT(sblock);
 357
 358	/* The range should be inside the sblock range */
 359	ASSERT(ssector->offset < sblock->len);
 360
 361	index = ssector->offset >> PAGE_SHIFT;
 362	ASSERT(index < SCRUB_MAX_PAGES);
 363	ASSERT(sblock->pages[index]);
 364	ASSERT(PagePrivate(sblock->pages[index]));
 365	return sblock->pages[index];
 366}
 367
 368static unsigned int scrub_sector_get_page_offset(struct scrub_sector *ssector)
 369{
 370	struct scrub_block *sblock = ssector->sblock;
 371
 372	/*
 373	 * When calling this function, ssector must be already attached to the
 374	 * parent sblock.
 375	 */
 376	ASSERT(sblock);
 377
 378	/* The range should be inside the sblock range */
 379	ASSERT(ssector->offset < sblock->len);
 380
 381	return offset_in_page(ssector->offset);
 382}
 383
 384static char *scrub_sector_get_kaddr(struct scrub_sector *ssector)
 385{
 386	return page_address(scrub_sector_get_page(ssector)) +
 387	       scrub_sector_get_page_offset(ssector);
 388}
 389
 390static int bio_add_scrub_sector(struct bio *bio, struct scrub_sector *ssector,
 391				unsigned int len)
 392{
 393	return bio_add_page(bio, scrub_sector_get_page(ssector), len,
 394			    scrub_sector_get_page_offset(ssector));
 395}
 396
 397static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 398				     struct scrub_block *sblocks_for_recheck[]);
 399static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 400				struct scrub_block *sblock,
 401				int retry_failed_mirror);
 402static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 403static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 404					     struct scrub_block *sblock_good);
 405static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
 406					    struct scrub_block *sblock_good,
 407					    int sector_num, int force_write);
 408static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 409static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock,
 410					     int sector_num);
 411static int scrub_checksum_data(struct scrub_block *sblock);
 412static int scrub_checksum_tree_block(struct scrub_block *sblock);
 413static int scrub_checksum_super(struct scrub_block *sblock);
 
 414static void scrub_block_put(struct scrub_block *sblock);
 415static void scrub_sector_get(struct scrub_sector *sector);
 416static void scrub_sector_put(struct scrub_sector *sector);
 417static void scrub_parity_get(struct scrub_parity *sparity);
 418static void scrub_parity_put(struct scrub_parity *sparity);
 419static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
 420			 u64 physical, struct btrfs_device *dev, u64 flags,
 421			 u64 gen, int mirror_num, u8 *csum,
 422			 u64 physical_for_dev_replace);
 
 
 423static void scrub_bio_end_io(struct bio *bio);
 424static void scrub_bio_end_io_worker(struct work_struct *work);
 425static void scrub_block_complete(struct scrub_block *sblock);
 426static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
 427				 u64 extent_logical, u32 extent_len,
 428				 u64 *extent_physical,
 429				 struct btrfs_device **extent_dev,
 430				 int *extent_mirror_num);
 431static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
 432				      struct scrub_sector *sector);
 433static void scrub_wr_submit(struct scrub_ctx *sctx);
 434static void scrub_wr_bio_end_io(struct bio *bio);
 435static void scrub_wr_bio_end_io_worker(struct work_struct *work);
 
 
 436static void scrub_put_ctx(struct scrub_ctx *sctx);
 437
 438static inline int scrub_is_page_on_raid56(struct scrub_sector *sector)
 439{
 440	return sector->recover &&
 441	       (sector->recover->bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 442}
 443
 444static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 445{
 446	refcount_inc(&sctx->refs);
 447	atomic_inc(&sctx->bios_in_flight);
 448}
 449
 450static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 451{
 452	atomic_dec(&sctx->bios_in_flight);
 453	wake_up(&sctx->list_wait);
 454	scrub_put_ctx(sctx);
 455}
 456
 457static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 458{
 459	while (atomic_read(&fs_info->scrub_pause_req)) {
 460		mutex_unlock(&fs_info->scrub_lock);
 461		wait_event(fs_info->scrub_pause_wait,
 462		   atomic_read(&fs_info->scrub_pause_req) == 0);
 463		mutex_lock(&fs_info->scrub_lock);
 464	}
 465}
 466
 467static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 468{
 469	atomic_inc(&fs_info->scrubs_paused);
 470	wake_up(&fs_info->scrub_pause_wait);
 471}
 472
 473static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 474{
 475	mutex_lock(&fs_info->scrub_lock);
 476	__scrub_blocked_if_needed(fs_info);
 477	atomic_dec(&fs_info->scrubs_paused);
 478	mutex_unlock(&fs_info->scrub_lock);
 479
 480	wake_up(&fs_info->scrub_pause_wait);
 481}
 482
 483static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 484{
 485	scrub_pause_on(fs_info);
 486	scrub_pause_off(fs_info);
 487}
 488
 489/*
 490 * Insert new full stripe lock into full stripe locks tree
 491 *
 492 * Return pointer to existing or newly inserted full_stripe_lock structure if
 493 * everything works well.
 494 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 495 *
 496 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 497 * function
 498 */
 499static struct full_stripe_lock *insert_full_stripe_lock(
 500		struct btrfs_full_stripe_locks_tree *locks_root,
 501		u64 fstripe_logical)
 502{
 503	struct rb_node **p;
 504	struct rb_node *parent = NULL;
 505	struct full_stripe_lock *entry;
 506	struct full_stripe_lock *ret;
 507
 508	lockdep_assert_held(&locks_root->lock);
 509
 510	p = &locks_root->root.rb_node;
 511	while (*p) {
 512		parent = *p;
 513		entry = rb_entry(parent, struct full_stripe_lock, node);
 514		if (fstripe_logical < entry->logical) {
 515			p = &(*p)->rb_left;
 516		} else if (fstripe_logical > entry->logical) {
 517			p = &(*p)->rb_right;
 518		} else {
 519			entry->refs++;
 520			return entry;
 521		}
 522	}
 523
 524	/*
 525	 * Insert new lock.
 526	 */
 527	ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 528	if (!ret)
 529		return ERR_PTR(-ENOMEM);
 530	ret->logical = fstripe_logical;
 531	ret->refs = 1;
 532	mutex_init(&ret->mutex);
 533
 534	rb_link_node(&ret->node, parent, p);
 535	rb_insert_color(&ret->node, &locks_root->root);
 536	return ret;
 537}
 538
 539/*
 540 * Search for a full stripe lock of a block group
 541 *
 542 * Return pointer to existing full stripe lock if found
 543 * Return NULL if not found
 544 */
 545static struct full_stripe_lock *search_full_stripe_lock(
 546		struct btrfs_full_stripe_locks_tree *locks_root,
 547		u64 fstripe_logical)
 548{
 549	struct rb_node *node;
 550	struct full_stripe_lock *entry;
 551
 552	lockdep_assert_held(&locks_root->lock);
 553
 554	node = locks_root->root.rb_node;
 555	while (node) {
 556		entry = rb_entry(node, struct full_stripe_lock, node);
 557		if (fstripe_logical < entry->logical)
 558			node = node->rb_left;
 559		else if (fstripe_logical > entry->logical)
 560			node = node->rb_right;
 561		else
 562			return entry;
 563	}
 564	return NULL;
 565}
 566
 567/*
 568 * Helper to get full stripe logical from a normal bytenr.
 569 *
 570 * Caller must ensure @cache is a RAID56 block group.
 571 */
 572static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
 573{
 574	u64 ret;
 575
 576	/*
 577	 * Due to chunk item size limit, full stripe length should not be
 578	 * larger than U32_MAX. Just a sanity check here.
 579	 */
 580	WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 581
 582	/*
 583	 * round_down() can only handle power of 2, while RAID56 full
 584	 * stripe length can be 64KiB * n, so we need to manually round down.
 585	 */
 586	ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 587			cache->full_stripe_len + cache->start;
 588	return ret;
 589}
 590
 591/*
 592 * Lock a full stripe to avoid concurrency of recovery and read
 593 *
 594 * It's only used for profiles with parities (RAID5/6), for other profiles it
 595 * does nothing.
 596 *
 597 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 598 * So caller must call unlock_full_stripe() at the same context.
 599 *
 600 * Return <0 if encounters error.
 601 */
 602static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 603			    bool *locked_ret)
 604{
 605	struct btrfs_block_group *bg_cache;
 606	struct btrfs_full_stripe_locks_tree *locks_root;
 607	struct full_stripe_lock *existing;
 608	u64 fstripe_start;
 609	int ret = 0;
 610
 611	*locked_ret = false;
 612	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 613	if (!bg_cache) {
 614		ASSERT(0);
 615		return -ENOENT;
 616	}
 617
 618	/* Profiles not based on parity don't need full stripe lock */
 619	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 620		goto out;
 621	locks_root = &bg_cache->full_stripe_locks_root;
 622
 623	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 624
 625	/* Now insert the full stripe lock */
 626	mutex_lock(&locks_root->lock);
 627	existing = insert_full_stripe_lock(locks_root, fstripe_start);
 628	mutex_unlock(&locks_root->lock);
 629	if (IS_ERR(existing)) {
 630		ret = PTR_ERR(existing);
 631		goto out;
 632	}
 633	mutex_lock(&existing->mutex);
 634	*locked_ret = true;
 635out:
 636	btrfs_put_block_group(bg_cache);
 637	return ret;
 638}
 639
 640/*
 641 * Unlock a full stripe.
 642 *
 643 * NOTE: Caller must ensure it's the same context calling corresponding
 644 * lock_full_stripe().
 645 *
 646 * Return 0 if we unlock full stripe without problem.
 647 * Return <0 for error
 648 */
 649static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 650			      bool locked)
 651{
 652	struct btrfs_block_group *bg_cache;
 653	struct btrfs_full_stripe_locks_tree *locks_root;
 654	struct full_stripe_lock *fstripe_lock;
 655	u64 fstripe_start;
 656	bool freeit = false;
 657	int ret = 0;
 658
 659	/* If we didn't acquire full stripe lock, no need to continue */
 660	if (!locked)
 661		return 0;
 662
 663	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 664	if (!bg_cache) {
 665		ASSERT(0);
 666		return -ENOENT;
 667	}
 668	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 669		goto out;
 670
 671	locks_root = &bg_cache->full_stripe_locks_root;
 672	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 673
 674	mutex_lock(&locks_root->lock);
 675	fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 676	/* Unpaired unlock_full_stripe() detected */
 677	if (!fstripe_lock) {
 678		WARN_ON(1);
 679		ret = -ENOENT;
 680		mutex_unlock(&locks_root->lock);
 681		goto out;
 682	}
 683
 684	if (fstripe_lock->refs == 0) {
 685		WARN_ON(1);
 686		btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 687			fstripe_lock->logical);
 688	} else {
 689		fstripe_lock->refs--;
 690	}
 691
 692	if (fstripe_lock->refs == 0) {
 693		rb_erase(&fstripe_lock->node, &locks_root->root);
 694		freeit = true;
 695	}
 696	mutex_unlock(&locks_root->lock);
 697
 698	mutex_unlock(&fstripe_lock->mutex);
 699	if (freeit)
 700		kfree(fstripe_lock);
 701out:
 702	btrfs_put_block_group(bg_cache);
 703	return ret;
 704}
 705
 706static void scrub_free_csums(struct scrub_ctx *sctx)
 707{
 708	while (!list_empty(&sctx->csum_list)) {
 709		struct btrfs_ordered_sum *sum;
 710		sum = list_first_entry(&sctx->csum_list,
 711				       struct btrfs_ordered_sum, list);
 712		list_del(&sum->list);
 713		kfree(sum);
 714	}
 715}
 716
 717static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 718{
 719	int i;
 720
 721	if (!sctx)
 722		return;
 723
 724	/* this can happen when scrub is cancelled */
 725	if (sctx->curr != -1) {
 726		struct scrub_bio *sbio = sctx->bios[sctx->curr];
 727
 728		for (i = 0; i < sbio->sector_count; i++)
 729			scrub_block_put(sbio->sectors[i]->sblock);
 
 
 730		bio_put(sbio->bio);
 731	}
 732
 733	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 734		struct scrub_bio *sbio = sctx->bios[i];
 735
 736		if (!sbio)
 737			break;
 738		kfree(sbio);
 739	}
 740
 741	kfree(sctx->wr_curr_bio);
 742	scrub_free_csums(sctx);
 743	kfree(sctx);
 744}
 745
 746static void scrub_put_ctx(struct scrub_ctx *sctx)
 747{
 748	if (refcount_dec_and_test(&sctx->refs))
 749		scrub_free_ctx(sctx);
 750}
 751
 752static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 753		struct btrfs_fs_info *fs_info, int is_dev_replace)
 754{
 755	struct scrub_ctx *sctx;
 756	int		i;
 757
 758	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 759	if (!sctx)
 760		goto nomem;
 761	refcount_set(&sctx->refs, 1);
 762	sctx->is_dev_replace = is_dev_replace;
 763	sctx->sectors_per_bio = SCRUB_SECTORS_PER_BIO;
 764	sctx->curr = -1;
 765	sctx->fs_info = fs_info;
 766	INIT_LIST_HEAD(&sctx->csum_list);
 767	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 768		struct scrub_bio *sbio;
 769
 770		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 771		if (!sbio)
 772			goto nomem;
 773		sctx->bios[i] = sbio;
 774
 775		sbio->index = i;
 776		sbio->sctx = sctx;
 777		sbio->sector_count = 0;
 778		INIT_WORK(&sbio->work, scrub_bio_end_io_worker);
 
 779
 780		if (i != SCRUB_BIOS_PER_SCTX - 1)
 781			sctx->bios[i]->next_free = i + 1;
 782		else
 783			sctx->bios[i]->next_free = -1;
 784	}
 785	sctx->first_free = 0;
 786	atomic_set(&sctx->bios_in_flight, 0);
 787	atomic_set(&sctx->workers_pending, 0);
 788	atomic_set(&sctx->cancel_req, 0);
 
 789
 790	spin_lock_init(&sctx->list_lock);
 791	spin_lock_init(&sctx->stat_lock);
 792	init_waitqueue_head(&sctx->list_wait);
 793	sctx->throttle_deadline = 0;
 794
 795	WARN_ON(sctx->wr_curr_bio != NULL);
 796	mutex_init(&sctx->wr_lock);
 797	sctx->wr_curr_bio = NULL;
 798	if (is_dev_replace) {
 799		WARN_ON(!fs_info->dev_replace.tgtdev);
 
 800		sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 801		sctx->flush_all_writes = false;
 802	}
 803
 804	return sctx;
 805
 806nomem:
 807	scrub_free_ctx(sctx);
 808	return ERR_PTR(-ENOMEM);
 809}
 810
 811static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
 812				     u64 root, void *warn_ctx)
 813{
 
 814	u32 nlink;
 815	int ret;
 816	int i;
 817	unsigned nofs_flag;
 818	struct extent_buffer *eb;
 819	struct btrfs_inode_item *inode_item;
 820	struct scrub_warning *swarn = warn_ctx;
 821	struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 822	struct inode_fs_paths *ipath = NULL;
 823	struct btrfs_root *local_root;
 824	struct btrfs_key key;
 825
 826	local_root = btrfs_get_fs_root(fs_info, root, true);
 827	if (IS_ERR(local_root)) {
 828		ret = PTR_ERR(local_root);
 829		goto err;
 830	}
 831
 832	/*
 833	 * this makes the path point to (inum INODE_ITEM ioff)
 834	 */
 835	key.objectid = inum;
 836	key.type = BTRFS_INODE_ITEM_KEY;
 837	key.offset = 0;
 838
 839	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 840	if (ret) {
 841		btrfs_put_root(local_root);
 842		btrfs_release_path(swarn->path);
 843		goto err;
 844	}
 845
 846	eb = swarn->path->nodes[0];
 847	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 848					struct btrfs_inode_item);
 
 849	nlink = btrfs_inode_nlink(eb, inode_item);
 850	btrfs_release_path(swarn->path);
 851
 852	/*
 853	 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 854	 * uses GFP_NOFS in this context, so we keep it consistent but it does
 855	 * not seem to be strictly necessary.
 856	 */
 857	nofs_flag = memalloc_nofs_save();
 858	ipath = init_ipath(4096, local_root, swarn->path);
 859	memalloc_nofs_restore(nofs_flag);
 860	if (IS_ERR(ipath)) {
 861		btrfs_put_root(local_root);
 862		ret = PTR_ERR(ipath);
 863		ipath = NULL;
 864		goto err;
 865	}
 866	ret = paths_from_inode(inum, ipath);
 867
 868	if (ret < 0)
 869		goto err;
 870
 871	/*
 872	 * we deliberately ignore the bit ipath might have been too small to
 873	 * hold all of the paths here
 874	 */
 875	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 876		btrfs_warn_in_rcu(fs_info,
 877"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
 878				  swarn->errstr, swarn->logical,
 879				  btrfs_dev_name(swarn->dev),
 880				  swarn->physical,
 881				  root, inum, offset,
 882				  fs_info->sectorsize, nlink,
 883				  (char *)(unsigned long)ipath->fspath->val[i]);
 884
 885	btrfs_put_root(local_root);
 886	free_ipath(ipath);
 887	return 0;
 888
 889err:
 890	btrfs_warn_in_rcu(fs_info,
 891			  "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 892			  swarn->errstr, swarn->logical,
 893			  btrfs_dev_name(swarn->dev),
 894			  swarn->physical,
 895			  root, inum, offset, ret);
 896
 897	free_ipath(ipath);
 898	return 0;
 899}
 900
 901static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 902{
 903	struct btrfs_device *dev;
 904	struct btrfs_fs_info *fs_info;
 905	struct btrfs_path *path;
 906	struct btrfs_key found_key;
 907	struct extent_buffer *eb;
 908	struct btrfs_extent_item *ei;
 909	struct scrub_warning swarn;
 910	unsigned long ptr = 0;
 
 911	u64 flags = 0;
 912	u64 ref_root;
 913	u32 item_size;
 914	u8 ref_level = 0;
 915	int ret;
 916
 917	WARN_ON(sblock->sector_count < 1);
 918	dev = sblock->dev;
 919	fs_info = sblock->sctx->fs_info;
 920
 921	/* Super block error, no need to search extent tree. */
 922	if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 923		btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
 924			errstr, btrfs_dev_name(dev), sblock->physical);
 925		return;
 926	}
 927	path = btrfs_alloc_path();
 928	if (!path)
 929		return;
 930
 931	swarn.physical = sblock->physical;
 932	swarn.logical = sblock->logical;
 933	swarn.errstr = errstr;
 934	swarn.dev = NULL;
 935
 936	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 937				  &flags);
 938	if (ret < 0)
 939		goto out;
 940
 
 941	swarn.extent_item_size = found_key.offset;
 942
 943	eb = path->nodes[0];
 944	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 945	item_size = btrfs_item_size(eb, path->slots[0]);
 946
 947	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 948		do {
 949			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 950						      item_size, &ref_root,
 951						      &ref_level);
 952			btrfs_warn_in_rcu(fs_info,
 953"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 954				errstr, swarn.logical,
 955				btrfs_dev_name(dev),
 956				swarn.physical,
 957				ref_level ? "node" : "leaf",
 958				ret < 0 ? -1 : ref_level,
 959				ret < 0 ? -1 : ref_root);
 960		} while (ret != 1);
 961		btrfs_release_path(path);
 962	} else {
 963		struct btrfs_backref_walk_ctx ctx = { 0 };
 964
 965		btrfs_release_path(path);
 966
 967		ctx.bytenr = found_key.objectid;
 968		ctx.extent_item_pos = swarn.logical - found_key.objectid;
 969		ctx.fs_info = fs_info;
 970
 971		swarn.path = path;
 972		swarn.dev = dev;
 973
 974		iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
 
 975	}
 976
 977out:
 978	btrfs_free_path(path);
 979}
 980
 981static inline void scrub_get_recover(struct scrub_recover *recover)
 982{
 983	refcount_inc(&recover->refs);
 984}
 985
 986static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 987				     struct scrub_recover *recover)
 988{
 989	if (refcount_dec_and_test(&recover->refs)) {
 990		btrfs_bio_counter_dec(fs_info);
 991		btrfs_put_bioc(recover->bioc);
 992		kfree(recover);
 993	}
 994}
 995
 996/*
 997 * scrub_handle_errored_block gets called when either verification of the
 998 * sectors failed or the bio failed to read, e.g. with EIO. In the latter
 999 * case, this function handles all sectors in the bio, even though only one
1000 * may be bad.
1001 * The goal of this function is to repair the errored block by using the
1002 * contents of one of the mirrors.
1003 */
1004static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1005{
1006	struct scrub_ctx *sctx = sblock_to_check->sctx;
1007	struct btrfs_device *dev = sblock_to_check->dev;
1008	struct btrfs_fs_info *fs_info;
1009	u64 logical;
1010	unsigned int failed_mirror_index;
1011	unsigned int is_metadata;
1012	unsigned int have_csum;
1013	/* One scrub_block for each mirror */
1014	struct scrub_block *sblocks_for_recheck[BTRFS_MAX_MIRRORS] = { 0 };
1015	struct scrub_block *sblock_bad;
1016	int ret;
1017	int mirror_index;
1018	int sector_num;
1019	int success;
1020	bool full_stripe_locked;
1021	unsigned int nofs_flag;
1022	static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
1023				      DEFAULT_RATELIMIT_BURST);
1024
1025	BUG_ON(sblock_to_check->sector_count < 1);
1026	fs_info = sctx->fs_info;
1027	if (sblock_to_check->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1028		/*
1029		 * If we find an error in a super block, we just report it.
1030		 * They will get written with the next transaction commit
1031		 * anyway
1032		 */
1033		scrub_print_warning("super block error", sblock_to_check);
1034		spin_lock(&sctx->stat_lock);
1035		++sctx->stat.super_errors;
1036		spin_unlock(&sctx->stat_lock);
1037		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_CORRUPTION_ERRS);
1038		return 0;
1039	}
1040	logical = sblock_to_check->logical;
1041	ASSERT(sblock_to_check->mirror_num);
1042	failed_mirror_index = sblock_to_check->mirror_num - 1;
1043	is_metadata = !(sblock_to_check->sectors[0]->flags &
1044			BTRFS_EXTENT_FLAG_DATA);
1045	have_csum = sblock_to_check->sectors[0]->have_csum;
1046
1047	if (!sctx->is_dev_replace && btrfs_repair_one_zone(fs_info, logical))
1048		return 0;
1049
1050	/*
1051	 * We must use GFP_NOFS because the scrub task might be waiting for a
1052	 * worker task executing this function and in turn a transaction commit
1053	 * might be waiting the scrub task to pause (which needs to wait for all
1054	 * the worker tasks to complete before pausing).
1055	 * We do allocations in the workers through insert_full_stripe_lock()
1056	 * and scrub_add_sector_to_wr_bio(), which happens down the call chain of
1057	 * this function.
1058	 */
1059	nofs_flag = memalloc_nofs_save();
1060	/*
1061	 * For RAID5/6, race can happen for a different device scrub thread.
1062	 * For data corruption, Parity and Data threads will both try
1063	 * to recovery the data.
1064	 * Race can lead to doubly added csum error, or even unrecoverable
1065	 * error.
1066	 */
1067	ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1068	if (ret < 0) {
1069		memalloc_nofs_restore(nofs_flag);
1070		spin_lock(&sctx->stat_lock);
1071		if (ret == -ENOMEM)
1072			sctx->stat.malloc_errors++;
1073		sctx->stat.read_errors++;
1074		sctx->stat.uncorrectable_errors++;
1075		spin_unlock(&sctx->stat_lock);
1076		return ret;
1077	}
1078
1079	/*
1080	 * read all mirrors one after the other. This includes to
1081	 * re-read the extent or metadata block that failed (that was
1082	 * the cause that this fixup code is called) another time,
1083	 * sector by sector this time in order to know which sectors
1084	 * caused I/O errors and which ones are good (for all mirrors).
1085	 * It is the goal to handle the situation when more than one
1086	 * mirror contains I/O errors, but the errors do not
1087	 * overlap, i.e. the data can be repaired by selecting the
1088	 * sectors from those mirrors without I/O error on the
1089	 * particular sectors. One example (with blocks >= 2 * sectorsize)
1090	 * would be that mirror #1 has an I/O error on the first sector,
1091	 * the second sector is good, and mirror #2 has an I/O error on
1092	 * the second sector, but the first sector is good.
1093	 * Then the first sector of the first mirror can be repaired by
1094	 * taking the first sector of the second mirror, and the
1095	 * second sector of the second mirror can be repaired by
1096	 * copying the contents of the 2nd sector of the 1st mirror.
1097	 * One more note: if the sectors of one mirror contain I/O
1098	 * errors, the checksum cannot be verified. In order to get
1099	 * the best data for repairing, the first attempt is to find
1100	 * a mirror without I/O errors and with a validated checksum.
1101	 * Only if this is not possible, the sectors are picked from
1102	 * mirrors with I/O errors without considering the checksum.
1103	 * If the latter is the case, at the end, the checksum of the
1104	 * repaired area is verified in order to correctly maintain
1105	 * the statistics.
1106	 */
1107	for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1108		/*
1109		 * Note: the two members refs and outstanding_sectors are not
1110		 * used in the blocks that are used for the recheck procedure.
1111		 *
1112		 * But alloc_scrub_block() will initialize sblock::ref anyway,
1113		 * so we can use scrub_block_put() to clean them up.
1114		 *
1115		 * And here we don't setup the physical/dev for the sblock yet,
1116		 * they will be correctly initialized in scrub_setup_recheck_block().
1117		 */
1118		sblocks_for_recheck[mirror_index] = alloc_scrub_block(sctx, NULL,
1119							logical, 0, 0, mirror_index);
1120		if (!sblocks_for_recheck[mirror_index]) {
1121			spin_lock(&sctx->stat_lock);
1122			sctx->stat.malloc_errors++;
1123			sctx->stat.read_errors++;
1124			sctx->stat.uncorrectable_errors++;
1125			spin_unlock(&sctx->stat_lock);
1126			btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1127			goto out;
1128		}
1129	}
1130
1131	/* Setup the context, map the logical blocks and alloc the sectors */
1132	ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1133	if (ret) {
1134		spin_lock(&sctx->stat_lock);
1135		sctx->stat.read_errors++;
1136		sctx->stat.uncorrectable_errors++;
1137		spin_unlock(&sctx->stat_lock);
1138		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1139		goto out;
1140	}
1141	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1142	sblock_bad = sblocks_for_recheck[failed_mirror_index];
1143
1144	/* build and submit the bios for the failed mirror, check checksums */
1145	scrub_recheck_block(fs_info, sblock_bad, 1);
1146
1147	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1148	    sblock_bad->no_io_error_seen) {
1149		/*
1150		 * The error disappeared after reading sector by sector, or
1151		 * the area was part of a huge bio and other parts of the
1152		 * bio caused I/O errors, or the block layer merged several
1153		 * read requests into one and the error is caused by a
1154		 * different bio (usually one of the two latter cases is
1155		 * the cause)
1156		 */
1157		spin_lock(&sctx->stat_lock);
1158		sctx->stat.unverified_errors++;
1159		sblock_to_check->data_corrected = 1;
1160		spin_unlock(&sctx->stat_lock);
1161
1162		if (sctx->is_dev_replace)
1163			scrub_write_block_to_dev_replace(sblock_bad);
1164		goto out;
1165	}
1166
1167	if (!sblock_bad->no_io_error_seen) {
1168		spin_lock(&sctx->stat_lock);
1169		sctx->stat.read_errors++;
1170		spin_unlock(&sctx->stat_lock);
1171		if (__ratelimit(&rs))
1172			scrub_print_warning("i/o error", sblock_to_check);
1173		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1174	} else if (sblock_bad->checksum_error) {
1175		spin_lock(&sctx->stat_lock);
1176		sctx->stat.csum_errors++;
1177		spin_unlock(&sctx->stat_lock);
1178		if (__ratelimit(&rs))
1179			scrub_print_warning("checksum error", sblock_to_check);
1180		btrfs_dev_stat_inc_and_print(dev,
1181					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
1182	} else if (sblock_bad->header_error) {
1183		spin_lock(&sctx->stat_lock);
1184		sctx->stat.verify_errors++;
1185		spin_unlock(&sctx->stat_lock);
1186		if (__ratelimit(&rs))
1187			scrub_print_warning("checksum/header error",
1188					    sblock_to_check);
1189		if (sblock_bad->generation_error)
1190			btrfs_dev_stat_inc_and_print(dev,
1191				BTRFS_DEV_STAT_GENERATION_ERRS);
1192		else
1193			btrfs_dev_stat_inc_and_print(dev,
1194				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1195	}
1196
1197	if (sctx->readonly) {
1198		ASSERT(!sctx->is_dev_replace);
1199		goto out;
1200	}
1201
1202	/*
1203	 * now build and submit the bios for the other mirrors, check
1204	 * checksums.
1205	 * First try to pick the mirror which is completely without I/O
1206	 * errors and also does not have a checksum error.
1207	 * If one is found, and if a checksum is present, the full block
1208	 * that is known to contain an error is rewritten. Afterwards
1209	 * the block is known to be corrected.
1210	 * If a mirror is found which is completely correct, and no
1211	 * checksum is present, only those sectors are rewritten that had
1212	 * an I/O error in the block to be repaired, since it cannot be
1213	 * determined, which copy of the other sectors is better (and it
1214	 * could happen otherwise that a correct sector would be
1215	 * overwritten by a bad one).
1216	 */
1217	for (mirror_index = 0; ;mirror_index++) {
1218		struct scrub_block *sblock_other;
1219
1220		if (mirror_index == failed_mirror_index)
1221			continue;
1222
1223		/* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1224		if (!scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1225			if (mirror_index >= BTRFS_MAX_MIRRORS)
1226				break;
1227			if (!sblocks_for_recheck[mirror_index]->sector_count)
1228				break;
1229
1230			sblock_other = sblocks_for_recheck[mirror_index];
1231		} else {
1232			struct scrub_recover *r = sblock_bad->sectors[0]->recover;
1233			int max_allowed = r->bioc->num_stripes - r->bioc->num_tgtdevs;
 
1234
1235			if (mirror_index >= max_allowed)
1236				break;
1237			if (!sblocks_for_recheck[1]->sector_count)
1238				break;
1239
1240			ASSERT(failed_mirror_index == 0);
1241			sblock_other = sblocks_for_recheck[1];
1242			sblock_other->mirror_num = 1 + mirror_index;
1243		}
1244
1245		/* build and submit the bios, check checksums */
1246		scrub_recheck_block(fs_info, sblock_other, 0);
1247
1248		if (!sblock_other->header_error &&
1249		    !sblock_other->checksum_error &&
1250		    sblock_other->no_io_error_seen) {
1251			if (sctx->is_dev_replace) {
1252				scrub_write_block_to_dev_replace(sblock_other);
1253				goto corrected_error;
1254			} else {
1255				ret = scrub_repair_block_from_good_copy(
1256						sblock_bad, sblock_other);
1257				if (!ret)
1258					goto corrected_error;
1259			}
1260		}
1261	}
1262
1263	if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1264		goto did_not_correct_error;
1265
1266	/*
1267	 * In case of I/O errors in the area that is supposed to be
1268	 * repaired, continue by picking good copies of those sectors.
1269	 * Select the good sectors from mirrors to rewrite bad sectors from
1270	 * the area to fix. Afterwards verify the checksum of the block
1271	 * that is supposed to be repaired. This verification step is
1272	 * only done for the purpose of statistic counting and for the
1273	 * final scrub report, whether errors remain.
1274	 * A perfect algorithm could make use of the checksum and try
1275	 * all possible combinations of sectors from the different mirrors
1276	 * until the checksum verification succeeds. For example, when
1277	 * the 2nd sector of mirror #1 faces I/O errors, and the 2nd sector
1278	 * of mirror #2 is readable but the final checksum test fails,
1279	 * then the 2nd sector of mirror #3 could be tried, whether now
1280	 * the final checksum succeeds. But this would be a rare
1281	 * exception and is therefore not implemented. At least it is
1282	 * avoided that the good copy is overwritten.
1283	 * A more useful improvement would be to pick the sectors
1284	 * without I/O error based on sector sizes (512 bytes on legacy
1285	 * disks) instead of on sectorsize. Then maybe 512 byte of one
1286	 * mirror could be repaired by taking 512 byte of a different
1287	 * mirror, even if other 512 byte sectors in the same sectorsize
1288	 * area are unreadable.
1289	 */
1290	success = 1;
1291	for (sector_num = 0; sector_num < sblock_bad->sector_count;
1292	     sector_num++) {
1293		struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1294		struct scrub_block *sblock_other = NULL;
1295
1296		/* Skip no-io-error sectors in scrub */
1297		if (!sector_bad->io_error && !sctx->is_dev_replace)
1298			continue;
1299
1300		if (scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1301			/*
1302			 * In case of dev replace, if raid56 rebuild process
1303			 * didn't work out correct data, then copy the content
1304			 * in sblock_bad to make sure target device is identical
1305			 * to source device, instead of writing garbage data in
1306			 * sblock_for_recheck array to target device.
1307			 */
1308			sblock_other = NULL;
1309		} else if (sector_bad->io_error) {
1310			/* Try to find no-io-error sector in mirrors */
1311			for (mirror_index = 0;
1312			     mirror_index < BTRFS_MAX_MIRRORS &&
1313			     sblocks_for_recheck[mirror_index]->sector_count > 0;
1314			     mirror_index++) {
1315				if (!sblocks_for_recheck[mirror_index]->
1316				    sectors[sector_num]->io_error) {
1317					sblock_other = sblocks_for_recheck[mirror_index];
 
1318					break;
1319				}
1320			}
1321			if (!sblock_other)
1322				success = 0;
1323		}
1324
1325		if (sctx->is_dev_replace) {
1326			/*
1327			 * Did not find a mirror to fetch the sector from.
1328			 * scrub_write_sector_to_dev_replace() handles this
1329			 * case (sector->io_error), by filling the block with
1330			 * zeros before submitting the write request
 
1331			 */
1332			if (!sblock_other)
1333				sblock_other = sblock_bad;
1334
1335			if (scrub_write_sector_to_dev_replace(sblock_other,
1336							      sector_num) != 0) {
1337				atomic64_inc(
1338					&fs_info->dev_replace.num_write_errors);
1339				success = 0;
1340			}
1341		} else if (sblock_other) {
1342			ret = scrub_repair_sector_from_good_copy(sblock_bad,
1343								 sblock_other,
1344								 sector_num, 0);
1345			if (0 == ret)
1346				sector_bad->io_error = 0;
1347			else
1348				success = 0;
1349		}
1350	}
1351
1352	if (success && !sctx->is_dev_replace) {
1353		if (is_metadata || have_csum) {
1354			/*
1355			 * need to verify the checksum now that all
1356			 * sectors on disk are repaired (the write
1357			 * request for data to be repaired is on its way).
1358			 * Just be lazy and use scrub_recheck_block()
1359			 * which re-reads the data before the checksum
1360			 * is verified, but most likely the data comes out
1361			 * of the page cache.
1362			 */
1363			scrub_recheck_block(fs_info, sblock_bad, 1);
1364			if (!sblock_bad->header_error &&
1365			    !sblock_bad->checksum_error &&
1366			    sblock_bad->no_io_error_seen)
1367				goto corrected_error;
1368			else
1369				goto did_not_correct_error;
1370		} else {
1371corrected_error:
1372			spin_lock(&sctx->stat_lock);
1373			sctx->stat.corrected_errors++;
1374			sblock_to_check->data_corrected = 1;
1375			spin_unlock(&sctx->stat_lock);
1376			btrfs_err_rl_in_rcu(fs_info,
1377				"fixed up error at logical %llu on dev %s",
1378				logical, btrfs_dev_name(dev));
1379		}
1380	} else {
1381did_not_correct_error:
1382		spin_lock(&sctx->stat_lock);
1383		sctx->stat.uncorrectable_errors++;
1384		spin_unlock(&sctx->stat_lock);
1385		btrfs_err_rl_in_rcu(fs_info,
1386			"unable to fixup (regular) error at logical %llu on dev %s",
1387			logical, btrfs_dev_name(dev));
1388	}
1389
1390out:
1391	for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1392		struct scrub_block *sblock = sblocks_for_recheck[mirror_index];
1393		struct scrub_recover *recover;
1394		int sector_index;
1395
1396		/* Not allocated, continue checking the next mirror */
1397		if (!sblock)
1398			continue;
1399
1400		for (sector_index = 0; sector_index < sblock->sector_count;
1401		     sector_index++) {
1402			/*
1403			 * Here we just cleanup the recover, each sector will be
1404			 * properly cleaned up by later scrub_block_put()
1405			 */
1406			recover = sblock->sectors[sector_index]->recover;
1407			if (recover) {
1408				scrub_put_recover(fs_info, recover);
1409				sblock->sectors[sector_index]->recover = NULL;
1410			}
1411		}
1412		scrub_block_put(sblock);
1413	}
1414
1415	ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1416	memalloc_nofs_restore(nofs_flag);
1417	if (ret < 0)
1418		return ret;
1419	return 0;
1420}
1421
1422static inline int scrub_nr_raid_mirrors(struct btrfs_io_context *bioc)
1423{
1424	if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID5)
1425		return 2;
1426	else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID6)
1427		return 3;
1428	else
1429		return (int)bioc->num_stripes;
1430}
1431
1432static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1433						 u64 *raid_map,
 
1434						 int nstripes, int mirror,
1435						 int *stripe_index,
1436						 u64 *stripe_offset)
1437{
1438	int i;
1439
1440	if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1441		/* RAID5/6 */
1442		for (i = 0; i < nstripes; i++) {
1443			if (raid_map[i] == RAID6_Q_STRIPE ||
1444			    raid_map[i] == RAID5_P_STRIPE)
1445				continue;
1446
1447			if (logical >= raid_map[i] &&
1448			    logical < raid_map[i] + BTRFS_STRIPE_LEN)
1449				break;
1450		}
1451
1452		*stripe_index = i;
1453		*stripe_offset = logical - raid_map[i];
1454	} else {
1455		/* The other RAID type */
1456		*stripe_index = mirror;
1457		*stripe_offset = 0;
1458	}
1459}
1460
1461static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1462				     struct scrub_block *sblocks_for_recheck[])
1463{
1464	struct scrub_ctx *sctx = original_sblock->sctx;
1465	struct btrfs_fs_info *fs_info = sctx->fs_info;
1466	u64 logical = original_sblock->logical;
1467	u64 length = original_sblock->sector_count << fs_info->sectorsize_bits;
1468	u64 generation = original_sblock->sectors[0]->generation;
1469	u64 flags = original_sblock->sectors[0]->flags;
1470	u64 have_csum = original_sblock->sectors[0]->have_csum;
1471	struct scrub_recover *recover;
1472	struct btrfs_io_context *bioc;
1473	u64 sublen;
1474	u64 mapped_length;
1475	u64 stripe_offset;
1476	int stripe_index;
1477	int sector_index = 0;
1478	int mirror_index;
1479	int nmirrors;
1480	int ret;
1481
 
 
 
 
 
 
1482	while (length > 0) {
1483		sublen = min_t(u64, length, fs_info->sectorsize);
1484		mapped_length = sublen;
1485		bioc = NULL;
1486
1487		/*
1488		 * With a length of sectorsize, each returned stripe represents
1489		 * one mirror
1490		 */
1491		btrfs_bio_counter_inc_blocked(fs_info);
1492		ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1493				       logical, &mapped_length, &bioc);
1494		if (ret || !bioc || mapped_length < sublen) {
1495			btrfs_put_bioc(bioc);
1496			btrfs_bio_counter_dec(fs_info);
1497			return -EIO;
1498		}
1499
1500		recover = kzalloc(sizeof(struct scrub_recover), GFP_KERNEL);
1501		if (!recover) {
1502			btrfs_put_bioc(bioc);
1503			btrfs_bio_counter_dec(fs_info);
1504			return -ENOMEM;
1505		}
1506
1507		refcount_set(&recover->refs, 1);
1508		recover->bioc = bioc;
1509		recover->map_length = mapped_length;
1510
1511		ASSERT(sector_index < SCRUB_MAX_SECTORS_PER_BLOCK);
1512
1513		nmirrors = min(scrub_nr_raid_mirrors(bioc), BTRFS_MAX_MIRRORS);
1514
1515		for (mirror_index = 0; mirror_index < nmirrors;
1516		     mirror_index++) {
1517			struct scrub_block *sblock;
1518			struct scrub_sector *sector;
1519
1520			sblock = sblocks_for_recheck[mirror_index];
1521			sblock->sctx = sctx;
1522
1523			sector = alloc_scrub_sector(sblock, logical);
1524			if (!sector) {
 
1525				spin_lock(&sctx->stat_lock);
1526				sctx->stat.malloc_errors++;
1527				spin_unlock(&sctx->stat_lock);
1528				scrub_put_recover(fs_info, recover);
1529				return -ENOMEM;
1530			}
1531			sector->flags = flags;
1532			sector->generation = generation;
1533			sector->have_csum = have_csum;
 
 
 
 
1534			if (have_csum)
1535				memcpy(sector->csum,
1536				       original_sblock->sectors[0]->csum,
1537				       sctx->fs_info->csum_size);
1538
1539			scrub_stripe_index_and_offset(logical,
1540						      bioc->map_type,
1541						      bioc->raid_map,
1542						      bioc->num_stripes -
1543						      bioc->num_tgtdevs,
 
1544						      mirror_index,
1545						      &stripe_index,
1546						      &stripe_offset);
1547			/*
1548			 * We're at the first sector, also populate @sblock
1549			 * physical and dev.
1550			 */
1551			if (sector_index == 0) {
1552				sblock->physical =
1553					bioc->stripes[stripe_index].physical +
1554					stripe_offset;
1555				sblock->dev = bioc->stripes[stripe_index].dev;
1556				sblock->physical_for_dev_replace =
1557					original_sblock->physical_for_dev_replace;
1558			}
 
 
1559
1560			BUG_ON(sector_index >= original_sblock->sector_count);
1561			scrub_get_recover(recover);
1562			sector->recover = recover;
1563		}
1564		scrub_put_recover(fs_info, recover);
1565		length -= sublen;
1566		logical += sublen;
1567		sector_index++;
1568	}
1569
1570	return 0;
1571}
1572
1573static void scrub_bio_wait_endio(struct bio *bio)
1574{
1575	complete(bio->bi_private);
1576}
1577
1578static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1579					struct bio *bio,
1580					struct scrub_sector *sector)
1581{
1582	DECLARE_COMPLETION_ONSTACK(done);
 
 
1583
1584	bio->bi_iter.bi_sector = (sector->offset + sector->sblock->logical) >>
1585				 SECTOR_SHIFT;
1586	bio->bi_private = &done;
1587	bio->bi_end_io = scrub_bio_wait_endio;
1588	raid56_parity_recover(bio, sector->recover->bioc, sector->sblock->mirror_num);
 
 
 
 
 
 
1589
1590	wait_for_completion_io(&done);
1591	return blk_status_to_errno(bio->bi_status);
1592}
1593
1594static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1595					  struct scrub_block *sblock)
1596{
1597	struct scrub_sector *first_sector = sblock->sectors[0];
1598	struct bio *bio;
1599	int i;
1600
1601	/* All sectors in sblock belong to the same stripe on the same device. */
1602	ASSERT(sblock->dev);
1603	if (!sblock->dev->bdev)
1604		goto out;
1605
1606	bio = bio_alloc(sblock->dev->bdev, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
 
1607
1608	for (i = 0; i < sblock->sector_count; i++) {
1609		struct scrub_sector *sector = sblock->sectors[i];
1610
1611		bio_add_scrub_sector(bio, sector, fs_info->sectorsize);
 
1612	}
1613
1614	if (scrub_submit_raid56_bio_wait(fs_info, bio, first_sector)) {
1615		bio_put(bio);
1616		goto out;
1617	}
1618
1619	bio_put(bio);
1620
1621	scrub_recheck_block_checksum(sblock);
1622
1623	return;
1624out:
1625	for (i = 0; i < sblock->sector_count; i++)
1626		sblock->sectors[i]->io_error = 1;
1627
1628	sblock->no_io_error_seen = 0;
1629}
1630
1631/*
1632 * This function will check the on disk data for checksum errors, header errors
1633 * and read I/O errors. If any I/O errors happen, the exact sectors which are
1634 * errored are marked as being bad. The goal is to enable scrub to take those
1635 * sectors that are not errored from all the mirrors so that the sectors that
1636 * are errored in the just handled mirror can be repaired.
1637 */
1638static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1639				struct scrub_block *sblock,
1640				int retry_failed_mirror)
1641{
1642	int i;
1643
1644	sblock->no_io_error_seen = 1;
1645
1646	/* short cut for raid56 */
1647	if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->sectors[0]))
1648		return scrub_recheck_block_on_raid56(fs_info, sblock);
1649
1650	for (i = 0; i < sblock->sector_count; i++) {
1651		struct scrub_sector *sector = sblock->sectors[i];
1652		struct bio bio;
1653		struct bio_vec bvec;
1654
1655		if (sblock->dev->bdev == NULL) {
1656			sector->io_error = 1;
1657			sblock->no_io_error_seen = 0;
1658			continue;
1659		}
1660
1661		bio_init(&bio, sblock->dev->bdev, &bvec, 1, REQ_OP_READ);
1662		bio_add_scrub_sector(&bio, sector, fs_info->sectorsize);
1663		bio.bi_iter.bi_sector = (sblock->physical + sector->offset) >>
1664					SECTOR_SHIFT;
1665
1666		btrfsic_check_bio(&bio);
1667		if (submit_bio_wait(&bio)) {
1668			sector->io_error = 1;
 
 
1669			sblock->no_io_error_seen = 0;
1670		}
1671
1672		bio_uninit(&bio);
1673	}
1674
1675	if (sblock->no_io_error_seen)
1676		scrub_recheck_block_checksum(sblock);
1677}
1678
1679static inline int scrub_check_fsid(u8 fsid[], struct scrub_sector *sector)
 
1680{
1681	struct btrfs_fs_devices *fs_devices = sector->sblock->dev->fs_devices;
1682	int ret;
1683
1684	ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1685	return !ret;
1686}
1687
1688static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1689{
1690	sblock->header_error = 0;
1691	sblock->checksum_error = 0;
1692	sblock->generation_error = 0;
1693
1694	if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1695		scrub_checksum_data(sblock);
1696	else
1697		scrub_checksum_tree_block(sblock);
1698}
1699
1700static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1701					     struct scrub_block *sblock_good)
1702{
1703	int i;
1704	int ret = 0;
1705
1706	for (i = 0; i < sblock_bad->sector_count; i++) {
1707		int ret_sub;
1708
1709		ret_sub = scrub_repair_sector_from_good_copy(sblock_bad,
1710							     sblock_good, i, 1);
 
1711		if (ret_sub)
1712			ret = ret_sub;
1713	}
1714
1715	return ret;
1716}
1717
1718static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
1719					      struct scrub_block *sblock_good,
1720					      int sector_num, int force_write)
1721{
1722	struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1723	struct scrub_sector *sector_good = sblock_good->sectors[sector_num];
1724	struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1725	const u32 sectorsize = fs_info->sectorsize;
1726
 
 
1727	if (force_write || sblock_bad->header_error ||
1728	    sblock_bad->checksum_error || sector_bad->io_error) {
1729		struct bio bio;
1730		struct bio_vec bvec;
1731		int ret;
1732
1733		if (!sblock_bad->dev->bdev) {
1734			btrfs_warn_rl(fs_info,
1735				"scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1736			return -EIO;
1737		}
1738
1739		bio_init(&bio, sblock_bad->dev->bdev, &bvec, 1, REQ_OP_WRITE);
1740		bio.bi_iter.bi_sector = (sblock_bad->physical +
1741					 sector_bad->offset) >> SECTOR_SHIFT;
1742		ret = bio_add_scrub_sector(&bio, sector_good, sectorsize);
1743
1744		btrfsic_check_bio(&bio);
1745		ret = submit_bio_wait(&bio);
1746		bio_uninit(&bio);
 
 
1747
1748		if (ret) {
1749			btrfs_dev_stat_inc_and_print(sblock_bad->dev,
1750				BTRFS_DEV_STAT_WRITE_ERRS);
1751			atomic64_inc(&fs_info->dev_replace.num_write_errors);
 
1752			return -EIO;
1753		}
 
1754	}
1755
1756	return 0;
1757}
1758
1759static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1760{
1761	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1762	int i;
1763
1764	/*
1765	 * This block is used for the check of the parity on the source device,
1766	 * so the data needn't be written into the destination device.
1767	 */
1768	if (sblock->sparity)
1769		return;
1770
1771	for (i = 0; i < sblock->sector_count; i++) {
1772		int ret;
1773
1774		ret = scrub_write_sector_to_dev_replace(sblock, i);
1775		if (ret)
1776			atomic64_inc(&fs_info->dev_replace.num_write_errors);
1777	}
1778}
1779
1780static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock, int sector_num)
 
1781{
1782	const u32 sectorsize = sblock->sctx->fs_info->sectorsize;
1783	struct scrub_sector *sector = sblock->sectors[sector_num];
1784
1785	if (sector->io_error)
1786		memset(scrub_sector_get_kaddr(sector), 0, sectorsize);
 
1787
1788	return scrub_add_sector_to_wr_bio(sblock->sctx, sector);
1789}
1790
1791static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
 
1792{
1793	int ret = 0;
1794	u64 length;
1795
1796	if (!btrfs_is_zoned(sctx->fs_info))
1797		return 0;
1798
1799	if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1800		return 0;
1801
1802	if (sctx->write_pointer < physical) {
1803		length = physical - sctx->write_pointer;
1804
1805		ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1806						sctx->write_pointer, length);
1807		if (!ret)
1808			sctx->write_pointer = physical;
1809	}
1810	return ret;
1811}
1812
1813static void scrub_block_get(struct scrub_block *sblock)
1814{
1815	refcount_inc(&sblock->refs);
1816}
1817
1818static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
1819				      struct scrub_sector *sector)
1820{
1821	struct scrub_block *sblock = sector->sblock;
1822	struct scrub_bio *sbio;
1823	int ret;
1824	const u32 sectorsize = sctx->fs_info->sectorsize;
1825
1826	mutex_lock(&sctx->wr_lock);
1827again:
1828	if (!sctx->wr_curr_bio) {
1829		sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1830					      GFP_KERNEL);
1831		if (!sctx->wr_curr_bio) {
1832			mutex_unlock(&sctx->wr_lock);
1833			return -ENOMEM;
1834		}
1835		sctx->wr_curr_bio->sctx = sctx;
1836		sctx->wr_curr_bio->sector_count = 0;
1837	}
1838	sbio = sctx->wr_curr_bio;
1839	if (sbio->sector_count == 0) {
1840		ret = fill_writer_pointer_gap(sctx, sector->offset +
1841					      sblock->physical_for_dev_replace);
1842		if (ret) {
1843			mutex_unlock(&sctx->wr_lock);
1844			return ret;
1845		}
1846
1847		sbio->physical = sblock->physical_for_dev_replace + sector->offset;
1848		sbio->logical = sblock->logical + sector->offset;
1849		sbio->dev = sctx->wr_tgtdev;
1850		if (!sbio->bio) {
1851			sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
1852					      REQ_OP_WRITE, GFP_NOFS);
1853		}
1854		sbio->bio->bi_private = sbio;
1855		sbio->bio->bi_end_io = scrub_wr_bio_end_io;
1856		sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
 
 
 
 
1857		sbio->status = 0;
1858	} else if (sbio->physical + sbio->sector_count * sectorsize !=
1859		   sblock->physical_for_dev_replace + sector->offset ||
1860		   sbio->logical + sbio->sector_count * sectorsize !=
1861		   sblock->logical + sector->offset) {
1862		scrub_wr_submit(sctx);
1863		goto again;
1864	}
1865
1866	ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
1867	if (ret != sectorsize) {
1868		if (sbio->sector_count < 1) {
1869			bio_put(sbio->bio);
1870			sbio->bio = NULL;
1871			mutex_unlock(&sctx->wr_lock);
1872			return -EIO;
1873		}
1874		scrub_wr_submit(sctx);
1875		goto again;
1876	}
1877
1878	sbio->sectors[sbio->sector_count] = sector;
1879	scrub_sector_get(sector);
1880	/*
1881	 * Since ssector no longer holds a page, but uses sblock::pages, we
1882	 * have to ensure the sblock had not been freed before our write bio
1883	 * finished.
1884	 */
1885	scrub_block_get(sector->sblock);
1886
1887	sbio->sector_count++;
1888	if (sbio->sector_count == sctx->sectors_per_bio)
1889		scrub_wr_submit(sctx);
1890	mutex_unlock(&sctx->wr_lock);
1891
1892	return 0;
1893}
1894
1895static void scrub_wr_submit(struct scrub_ctx *sctx)
1896{
1897	struct scrub_bio *sbio;
1898
1899	if (!sctx->wr_curr_bio)
1900		return;
1901
1902	sbio = sctx->wr_curr_bio;
1903	sctx->wr_curr_bio = NULL;
 
1904	scrub_pending_bio_inc(sctx);
1905	/* process all writes in a single worker thread. Then the block layer
1906	 * orders the requests before sending them to the driver which
1907	 * doubled the write performance on spinning disks when measured
1908	 * with Linux 3.5 */
1909	btrfsic_check_bio(sbio->bio);
1910	submit_bio(sbio->bio);
1911
1912	if (btrfs_is_zoned(sctx->fs_info))
1913		sctx->write_pointer = sbio->physical + sbio->sector_count *
1914			sctx->fs_info->sectorsize;
1915}
1916
1917static void scrub_wr_bio_end_io(struct bio *bio)
1918{
1919	struct scrub_bio *sbio = bio->bi_private;
1920	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1921
1922	sbio->status = bio->bi_status;
1923	sbio->bio = bio;
1924
1925	INIT_WORK(&sbio->work, scrub_wr_bio_end_io_worker);
1926	queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1927}
1928
1929static void scrub_wr_bio_end_io_worker(struct work_struct *work)
1930{
1931	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1932	struct scrub_ctx *sctx = sbio->sctx;
1933	int i;
1934
1935	ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
1936	if (sbio->status) {
1937		struct btrfs_dev_replace *dev_replace =
1938			&sbio->sctx->fs_info->dev_replace;
1939
1940		for (i = 0; i < sbio->sector_count; i++) {
1941			struct scrub_sector *sector = sbio->sectors[i];
1942
1943			sector->io_error = 1;
1944			atomic64_inc(&dev_replace->num_write_errors);
1945		}
1946	}
1947
1948	/*
1949	 * In scrub_add_sector_to_wr_bio() we grab extra ref for sblock, now in
1950	 * endio we should put the sblock.
1951	 */
1952	for (i = 0; i < sbio->sector_count; i++) {
1953		scrub_block_put(sbio->sectors[i]->sblock);
1954		scrub_sector_put(sbio->sectors[i]);
1955	}
1956
1957	bio_put(sbio->bio);
1958	kfree(sbio);
1959	scrub_pending_bio_dec(sctx);
1960}
1961
1962static int scrub_checksum(struct scrub_block *sblock)
1963{
1964	u64 flags;
1965	int ret;
1966
1967	/*
1968	 * No need to initialize these stats currently,
1969	 * because this function only use return value
1970	 * instead of these stats value.
1971	 *
1972	 * Todo:
1973	 * always use stats
1974	 */
1975	sblock->header_error = 0;
1976	sblock->generation_error = 0;
1977	sblock->checksum_error = 0;
1978
1979	WARN_ON(sblock->sector_count < 1);
1980	flags = sblock->sectors[0]->flags;
1981	ret = 0;
1982	if (flags & BTRFS_EXTENT_FLAG_DATA)
1983		ret = scrub_checksum_data(sblock);
1984	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1985		ret = scrub_checksum_tree_block(sblock);
1986	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1987		ret = scrub_checksum_super(sblock);
1988	else
1989		WARN_ON(1);
1990	if (ret)
1991		scrub_handle_errored_block(sblock);
1992
1993	return ret;
1994}
1995
1996static int scrub_checksum_data(struct scrub_block *sblock)
1997{
1998	struct scrub_ctx *sctx = sblock->sctx;
1999	struct btrfs_fs_info *fs_info = sctx->fs_info;
2000	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2001	u8 csum[BTRFS_CSUM_SIZE];
2002	struct scrub_sector *sector;
2003	char *kaddr;
2004
2005	BUG_ON(sblock->sector_count < 1);
2006	sector = sblock->sectors[0];
2007	if (!sector->have_csum)
2008		return 0;
2009
2010	kaddr = scrub_sector_get_kaddr(sector);
2011
2012	shash->tfm = fs_info->csum_shash;
2013	crypto_shash_init(shash);
 
2014
2015	crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
2016
2017	if (memcmp(csum, sector->csum, fs_info->csum_size))
2018		sblock->checksum_error = 1;
 
2019	return sblock->checksum_error;
2020}
2021
2022static int scrub_checksum_tree_block(struct scrub_block *sblock)
2023{
2024	struct scrub_ctx *sctx = sblock->sctx;
2025	struct btrfs_header *h;
2026	struct btrfs_fs_info *fs_info = sctx->fs_info;
2027	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2028	u8 calculated_csum[BTRFS_CSUM_SIZE];
2029	u8 on_disk_csum[BTRFS_CSUM_SIZE];
2030	/*
2031	 * This is done in sectorsize steps even for metadata as there's a
2032	 * constraint for nodesize to be aligned to sectorsize. This will need
2033	 * to change so we don't misuse data and metadata units like that.
2034	 */
2035	const u32 sectorsize = sctx->fs_info->sectorsize;
2036	const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
2037	int i;
2038	struct scrub_sector *sector;
2039	char *kaddr;
2040
2041	BUG_ON(sblock->sector_count < 1);
2042
2043	/* Each member in sectors is just one sector */
2044	ASSERT(sblock->sector_count == num_sectors);
2045
2046	sector = sblock->sectors[0];
2047	kaddr = scrub_sector_get_kaddr(sector);
2048	h = (struct btrfs_header *)kaddr;
2049	memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
2050
2051	/*
2052	 * we don't use the getter functions here, as we
2053	 * a) don't have an extent buffer and
2054	 * b) the page is already kmapped
2055	 */
2056	if (sblock->logical != btrfs_stack_header_bytenr(h))
2057		sblock->header_error = 1;
2058
2059	if (sector->generation != btrfs_stack_header_generation(h)) {
2060		sblock->header_error = 1;
2061		sblock->generation_error = 1;
2062	}
2063
2064	if (!scrub_check_fsid(h->fsid, sector))
2065		sblock->header_error = 1;
2066
2067	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2068		   BTRFS_UUID_SIZE))
2069		sblock->header_error = 1;
2070
2071	shash->tfm = fs_info->csum_shash;
2072	crypto_shash_init(shash);
2073	crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
2074			    sectorsize - BTRFS_CSUM_SIZE);
2075
2076	for (i = 1; i < num_sectors; i++) {
2077		kaddr = scrub_sector_get_kaddr(sblock->sectors[i]);
2078		crypto_shash_update(shash, kaddr, sectorsize);
2079	}
2080
2081	crypto_shash_final(shash, calculated_csum);
2082	if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size))
2083		sblock->checksum_error = 1;
2084
2085	return sblock->header_error || sblock->checksum_error;
2086}
2087
2088static int scrub_checksum_super(struct scrub_block *sblock)
2089{
2090	struct btrfs_super_block *s;
2091	struct scrub_ctx *sctx = sblock->sctx;
2092	struct btrfs_fs_info *fs_info = sctx->fs_info;
2093	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2094	u8 calculated_csum[BTRFS_CSUM_SIZE];
2095	struct scrub_sector *sector;
2096	char *kaddr;
2097	int fail_gen = 0;
2098	int fail_cor = 0;
2099
2100	BUG_ON(sblock->sector_count < 1);
2101	sector = sblock->sectors[0];
2102	kaddr = scrub_sector_get_kaddr(sector);
2103	s = (struct btrfs_super_block *)kaddr;
2104
2105	if (sblock->logical != btrfs_super_bytenr(s))
2106		++fail_cor;
2107
2108	if (sector->generation != btrfs_super_generation(s))
2109		++fail_gen;
2110
2111	if (!scrub_check_fsid(s->fsid, sector))
2112		++fail_cor;
2113
2114	shash->tfm = fs_info->csum_shash;
2115	crypto_shash_init(shash);
2116	crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
2117			BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
2118
2119	if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
2120		++fail_cor;
2121
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2122	return fail_cor + fail_gen;
2123}
2124
 
 
 
 
 
2125static void scrub_block_put(struct scrub_block *sblock)
2126{
2127	if (refcount_dec_and_test(&sblock->refs)) {
2128		int i;
2129
2130		if (sblock->sparity)
2131			scrub_parity_put(sblock->sparity);
2132
2133		for (i = 0; i < sblock->sector_count; i++)
2134			scrub_sector_put(sblock->sectors[i]);
2135		for (i = 0; i < DIV_ROUND_UP(sblock->len, PAGE_SIZE); i++) {
2136			if (sblock->pages[i]) {
2137				detach_scrub_page_private(sblock->pages[i]);
2138				__free_page(sblock->pages[i]);
2139			}
2140		}
2141		kfree(sblock);
2142	}
2143}
2144
2145static void scrub_sector_get(struct scrub_sector *sector)
2146{
2147	atomic_inc(&sector->refs);
2148}
2149
2150static void scrub_sector_put(struct scrub_sector *sector)
2151{
2152	if (atomic_dec_and_test(&sector->refs))
2153		kfree(sector);
2154}
2155
2156/*
2157 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
2158 * second.  Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
2159 */
2160static void scrub_throttle(struct scrub_ctx *sctx)
2161{
2162	const int time_slice = 1000;
2163	struct scrub_bio *sbio;
2164	struct btrfs_device *device;
2165	s64 delta;
2166	ktime_t now;
2167	u32 div;
2168	u64 bwlimit;
2169
2170	sbio = sctx->bios[sctx->curr];
2171	device = sbio->dev;
2172	bwlimit = READ_ONCE(device->scrub_speed_max);
2173	if (bwlimit == 0)
2174		return;
2175
2176	/*
2177	 * Slice is divided into intervals when the IO is submitted, adjust by
2178	 * bwlimit and maximum of 64 intervals.
2179	 */
2180	div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
2181	div = min_t(u32, 64, div);
2182
2183	/* Start new epoch, set deadline */
2184	now = ktime_get();
2185	if (sctx->throttle_deadline == 0) {
2186		sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
2187		sctx->throttle_sent = 0;
2188	}
2189
2190	/* Still in the time to send? */
2191	if (ktime_before(now, sctx->throttle_deadline)) {
2192		/* If current bio is within the limit, send it */
2193		sctx->throttle_sent += sbio->bio->bi_iter.bi_size;
2194		if (sctx->throttle_sent <= div_u64(bwlimit, div))
2195			return;
2196
2197		/* We're over the limit, sleep until the rest of the slice */
2198		delta = ktime_ms_delta(sctx->throttle_deadline, now);
2199	} else {
2200		/* New request after deadline, start new epoch */
2201		delta = 0;
2202	}
2203
2204	if (delta) {
2205		long timeout;
2206
2207		timeout = div_u64(delta * HZ, 1000);
2208		schedule_timeout_interruptible(timeout);
2209	}
2210
2211	/* Next call will start the deadline period */
2212	sctx->throttle_deadline = 0;
2213}
2214
2215static void scrub_submit(struct scrub_ctx *sctx)
2216{
2217	struct scrub_bio *sbio;
2218
2219	if (sctx->curr == -1)
2220		return;
2221
2222	scrub_throttle(sctx);
2223
2224	sbio = sctx->bios[sctx->curr];
2225	sctx->curr = -1;
2226	scrub_pending_bio_inc(sctx);
2227	btrfsic_check_bio(sbio->bio);
2228	submit_bio(sbio->bio);
2229}
2230
2231static int scrub_add_sector_to_rd_bio(struct scrub_ctx *sctx,
2232				      struct scrub_sector *sector)
2233{
2234	struct scrub_block *sblock = sector->sblock;
2235	struct scrub_bio *sbio;
2236	const u32 sectorsize = sctx->fs_info->sectorsize;
2237	int ret;
2238
2239again:
2240	/*
2241	 * grab a fresh bio or wait for one to become available
2242	 */
2243	while (sctx->curr == -1) {
2244		spin_lock(&sctx->list_lock);
2245		sctx->curr = sctx->first_free;
2246		if (sctx->curr != -1) {
2247			sctx->first_free = sctx->bios[sctx->curr]->next_free;
2248			sctx->bios[sctx->curr]->next_free = -1;
2249			sctx->bios[sctx->curr]->sector_count = 0;
2250			spin_unlock(&sctx->list_lock);
2251		} else {
2252			spin_unlock(&sctx->list_lock);
2253			wait_event(sctx->list_wait, sctx->first_free != -1);
2254		}
2255	}
2256	sbio = sctx->bios[sctx->curr];
2257	if (sbio->sector_count == 0) {
2258		sbio->physical = sblock->physical + sector->offset;
2259		sbio->logical = sblock->logical + sector->offset;
2260		sbio->dev = sblock->dev;
2261		if (!sbio->bio) {
2262			sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
2263					      REQ_OP_READ, GFP_NOFS);
2264		}
2265		sbio->bio->bi_private = sbio;
2266		sbio->bio->bi_end_io = scrub_bio_end_io;
2267		sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
 
 
 
 
 
 
2268		sbio->status = 0;
2269	} else if (sbio->physical + sbio->sector_count * sectorsize !=
2270		   sblock->physical + sector->offset ||
2271		   sbio->logical + sbio->sector_count * sectorsize !=
2272		   sblock->logical + sector->offset ||
2273		   sbio->dev != sblock->dev) {
2274		scrub_submit(sctx);
2275		goto again;
2276	}
2277
2278	sbio->sectors[sbio->sector_count] = sector;
2279	ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
2280	if (ret != sectorsize) {
2281		if (sbio->sector_count < 1) {
2282			bio_put(sbio->bio);
2283			sbio->bio = NULL;
2284			return -EIO;
2285		}
2286		scrub_submit(sctx);
2287		goto again;
2288	}
2289
2290	scrub_block_get(sblock); /* one for the page added to the bio */
2291	atomic_inc(&sblock->outstanding_sectors);
2292	sbio->sector_count++;
2293	if (sbio->sector_count == sctx->sectors_per_bio)
2294		scrub_submit(sctx);
2295
2296	return 0;
2297}
2298
2299static void scrub_missing_raid56_end_io(struct bio *bio)
2300{
2301	struct scrub_block *sblock = bio->bi_private;
2302	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2303
2304	btrfs_bio_counter_dec(fs_info);
2305	if (bio->bi_status)
2306		sblock->no_io_error_seen = 0;
2307
2308	bio_put(bio);
2309
2310	queue_work(fs_info->scrub_workers, &sblock->work);
2311}
2312
2313static void scrub_missing_raid56_worker(struct work_struct *work)
2314{
2315	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2316	struct scrub_ctx *sctx = sblock->sctx;
2317	struct btrfs_fs_info *fs_info = sctx->fs_info;
2318	u64 logical;
2319	struct btrfs_device *dev;
2320
2321	logical = sblock->logical;
2322	dev = sblock->dev;
2323
2324	if (sblock->no_io_error_seen)
2325		scrub_recheck_block_checksum(sblock);
2326
2327	if (!sblock->no_io_error_seen) {
2328		spin_lock(&sctx->stat_lock);
2329		sctx->stat.read_errors++;
2330		spin_unlock(&sctx->stat_lock);
2331		btrfs_err_rl_in_rcu(fs_info,
2332			"IO error rebuilding logical %llu for dev %s",
2333			logical, btrfs_dev_name(dev));
2334	} else if (sblock->header_error || sblock->checksum_error) {
2335		spin_lock(&sctx->stat_lock);
2336		sctx->stat.uncorrectable_errors++;
2337		spin_unlock(&sctx->stat_lock);
2338		btrfs_err_rl_in_rcu(fs_info,
2339			"failed to rebuild valid logical %llu for dev %s",
2340			logical, btrfs_dev_name(dev));
2341	} else {
2342		scrub_write_block_to_dev_replace(sblock);
2343	}
2344
2345	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2346		mutex_lock(&sctx->wr_lock);
2347		scrub_wr_submit(sctx);
2348		mutex_unlock(&sctx->wr_lock);
2349	}
2350
2351	scrub_block_put(sblock);
2352	scrub_pending_bio_dec(sctx);
2353}
2354
2355static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2356{
2357	struct scrub_ctx *sctx = sblock->sctx;
2358	struct btrfs_fs_info *fs_info = sctx->fs_info;
2359	u64 length = sblock->sector_count << fs_info->sectorsize_bits;
2360	u64 logical = sblock->logical;
2361	struct btrfs_io_context *bioc = NULL;
2362	struct bio *bio;
2363	struct btrfs_raid_bio *rbio;
2364	int ret;
2365	int i;
2366
2367	btrfs_bio_counter_inc_blocked(fs_info);
2368	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2369			       &length, &bioc);
2370	if (ret || !bioc || !bioc->raid_map)
2371		goto bioc_out;
2372
2373	if (WARN_ON(!sctx->is_dev_replace ||
2374		    !(bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2375		/*
2376		 * We shouldn't be scrubbing a missing device. Even for dev
2377		 * replace, we should only get here for RAID 5/6. We either
2378		 * managed to mount something with no mirrors remaining or
2379		 * there's a bug in scrub_find_good_copy()/btrfs_map_block().
2380		 */
2381		goto bioc_out;
2382	}
2383
2384	bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2385	bio->bi_iter.bi_sector = logical >> 9;
2386	bio->bi_private = sblock;
2387	bio->bi_end_io = scrub_missing_raid56_end_io;
2388
2389	rbio = raid56_alloc_missing_rbio(bio, bioc);
2390	if (!rbio)
2391		goto rbio_out;
2392
2393	for (i = 0; i < sblock->sector_count; i++) {
2394		struct scrub_sector *sector = sblock->sectors[i];
2395
2396		raid56_add_scrub_pages(rbio, scrub_sector_get_page(sector),
2397				       scrub_sector_get_page_offset(sector),
2398				       sector->offset + sector->sblock->logical);
2399	}
2400
2401	INIT_WORK(&sblock->work, scrub_missing_raid56_worker);
2402	scrub_block_get(sblock);
2403	scrub_pending_bio_inc(sctx);
2404	raid56_submit_missing_rbio(rbio);
2405	btrfs_put_bioc(bioc);
2406	return;
2407
2408rbio_out:
2409	bio_put(bio);
2410bioc_out:
2411	btrfs_bio_counter_dec(fs_info);
2412	btrfs_put_bioc(bioc);
2413	spin_lock(&sctx->stat_lock);
2414	sctx->stat.malloc_errors++;
2415	spin_unlock(&sctx->stat_lock);
2416}
2417
2418static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
2419		       u64 physical, struct btrfs_device *dev, u64 flags,
2420		       u64 gen, int mirror_num, u8 *csum,
2421		       u64 physical_for_dev_replace)
2422{
2423	struct scrub_block *sblock;
2424	const u32 sectorsize = sctx->fs_info->sectorsize;
2425	int index;
2426
2427	sblock = alloc_scrub_block(sctx, dev, logical, physical,
2428				   physical_for_dev_replace, mirror_num);
2429	if (!sblock) {
2430		spin_lock(&sctx->stat_lock);
2431		sctx->stat.malloc_errors++;
2432		spin_unlock(&sctx->stat_lock);
2433		return -ENOMEM;
2434	}
2435
 
 
 
 
 
 
2436	for (index = 0; len > 0; index++) {
2437		struct scrub_sector *sector;
2438		/*
2439		 * Here we will allocate one page for one sector to scrub.
2440		 * This is fine if PAGE_SIZE == sectorsize, but will cost
2441		 * more memory for PAGE_SIZE > sectorsize case.
2442		 */
2443		u32 l = min(sectorsize, len);
2444
2445		sector = alloc_scrub_sector(sblock, logical);
2446		if (!sector) {
 
2447			spin_lock(&sctx->stat_lock);
2448			sctx->stat.malloc_errors++;
2449			spin_unlock(&sctx->stat_lock);
2450			scrub_block_put(sblock);
2451			return -ENOMEM;
2452		}
2453		sector->flags = flags;
2454		sector->generation = gen;
 
 
 
 
 
 
 
 
 
2455		if (csum) {
2456			sector->have_csum = 1;
2457			memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2458		} else {
2459			sector->have_csum = 0;
2460		}
 
 
 
 
2461		len -= l;
2462		logical += l;
2463		physical += l;
2464		physical_for_dev_replace += l;
2465	}
2466
2467	WARN_ON(sblock->sector_count == 0);
2468	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2469		/*
2470		 * This case should only be hit for RAID 5/6 device replace. See
2471		 * the comment in scrub_missing_raid56_pages() for details.
2472		 */
2473		scrub_missing_raid56_pages(sblock);
2474	} else {
2475		for (index = 0; index < sblock->sector_count; index++) {
2476			struct scrub_sector *sector = sblock->sectors[index];
2477			int ret;
2478
2479			ret = scrub_add_sector_to_rd_bio(sctx, sector);
2480			if (ret) {
2481				scrub_block_put(sblock);
2482				return ret;
2483			}
2484		}
2485
2486		if (flags & BTRFS_EXTENT_FLAG_SUPER)
2487			scrub_submit(sctx);
2488	}
2489
2490	/* last one frees, either here or in bio completion for last page */
2491	scrub_block_put(sblock);
2492	return 0;
2493}
2494
2495static void scrub_bio_end_io(struct bio *bio)
2496{
2497	struct scrub_bio *sbio = bio->bi_private;
2498	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2499
2500	sbio->status = bio->bi_status;
2501	sbio->bio = bio;
2502
2503	queue_work(fs_info->scrub_workers, &sbio->work);
2504}
2505
2506static void scrub_bio_end_io_worker(struct work_struct *work)
2507{
2508	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2509	struct scrub_ctx *sctx = sbio->sctx;
2510	int i;
2511
2512	ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
2513	if (sbio->status) {
2514		for (i = 0; i < sbio->sector_count; i++) {
2515			struct scrub_sector *sector = sbio->sectors[i];
2516
2517			sector->io_error = 1;
2518			sector->sblock->no_io_error_seen = 0;
2519		}
2520	}
2521
2522	/* Now complete the scrub_block items that have all pages completed */
2523	for (i = 0; i < sbio->sector_count; i++) {
2524		struct scrub_sector *sector = sbio->sectors[i];
2525		struct scrub_block *sblock = sector->sblock;
2526
2527		if (atomic_dec_and_test(&sblock->outstanding_sectors))
2528			scrub_block_complete(sblock);
2529		scrub_block_put(sblock);
2530	}
2531
2532	bio_put(sbio->bio);
2533	sbio->bio = NULL;
2534	spin_lock(&sctx->list_lock);
2535	sbio->next_free = sctx->first_free;
2536	sctx->first_free = sbio->index;
2537	spin_unlock(&sctx->list_lock);
2538
2539	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2540		mutex_lock(&sctx->wr_lock);
2541		scrub_wr_submit(sctx);
2542		mutex_unlock(&sctx->wr_lock);
2543	}
2544
2545	scrub_pending_bio_dec(sctx);
2546}
2547
2548static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2549				       unsigned long *bitmap,
2550				       u64 start, u32 len)
2551{
2552	u64 offset;
 
2553	u32 nsectors;
2554	u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2555
2556	if (len >= sparity->stripe_len) {
2557		bitmap_set(bitmap, 0, sparity->nsectors);
2558		return;
2559	}
2560
2561	start -= sparity->logic_start;
2562	start = div64_u64_rem(start, sparity->stripe_len, &offset);
2563	offset = offset >> sectorsize_bits;
2564	nsectors = len >> sectorsize_bits;
 
 
 
2565
2566	if (offset + nsectors <= sparity->nsectors) {
2567		bitmap_set(bitmap, offset, nsectors);
2568		return;
2569	}
2570
2571	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2572	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2573}
2574
2575static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2576						   u64 start, u32 len)
2577{
2578	__scrub_mark_bitmap(sparity, &sparity->ebitmap, start, len);
2579}
2580
2581static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2582						  u64 start, u32 len)
2583{
2584	__scrub_mark_bitmap(sparity, &sparity->dbitmap, start, len);
2585}
2586
2587static void scrub_block_complete(struct scrub_block *sblock)
2588{
2589	int corrupted = 0;
2590
2591	if (!sblock->no_io_error_seen) {
2592		corrupted = 1;
2593		scrub_handle_errored_block(sblock);
2594	} else {
2595		/*
2596		 * if has checksum error, write via repair mechanism in
2597		 * dev replace case, otherwise write here in dev replace
2598		 * case.
2599		 */
2600		corrupted = scrub_checksum(sblock);
2601		if (!corrupted && sblock->sctx->is_dev_replace)
2602			scrub_write_block_to_dev_replace(sblock);
2603	}
2604
2605	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2606		u64 start = sblock->logical;
2607		u64 end = sblock->logical +
2608			  sblock->sectors[sblock->sector_count - 1]->offset +
2609			  sblock->sctx->fs_info->sectorsize;
2610
2611		ASSERT(end - start <= U32_MAX);
2612		scrub_parity_mark_sectors_error(sblock->sparity,
2613						start, end - start);
2614	}
2615}
2616
2617static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2618{
2619	sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2620	list_del(&sum->list);
2621	kfree(sum);
2622}
2623
2624/*
2625 * Find the desired csum for range [logical, logical + sectorsize), and store
2626 * the csum into @csum.
2627 *
2628 * The search source is sctx->csum_list, which is a pre-populated list
2629 * storing bytenr ordered csum ranges.  We're responsible to cleanup any range
2630 * that is before @logical.
2631 *
2632 * Return 0 if there is no csum for the range.
2633 * Return 1 if there is csum for the range and copied to @csum.
2634 */
2635static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2636{
2637	bool found = false;
 
 
2638
2639	while (!list_empty(&sctx->csum_list)) {
2640		struct btrfs_ordered_sum *sum = NULL;
2641		unsigned long index;
2642		unsigned long num_sectors;
2643
2644		sum = list_first_entry(&sctx->csum_list,
2645				       struct btrfs_ordered_sum, list);
2646		/* The current csum range is beyond our range, no csum found */
2647		if (sum->bytenr > logical)
 
 
2648			break;
2649
2650		/*
2651		 * The current sum is before our bytenr, since scrub is always
2652		 * done in bytenr order, the csum will never be used anymore,
2653		 * clean it up so that later calls won't bother with the range,
2654		 * and continue search the next range.
2655		 */
2656		if (sum->bytenr + sum->len <= logical) {
2657			drop_csum_range(sctx, sum);
2658			continue;
2659		}
2660
2661		/* Now the csum range covers our bytenr, copy the csum */
2662		found = true;
2663		index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2664		num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2665
2666		memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2667		       sctx->fs_info->csum_size);
2668
2669		/* Cleanup the range if we're at the end of the csum range */
2670		if (index == num_sectors - 1)
2671			drop_csum_range(sctx, sum);
2672		break;
2673	}
2674	if (!found)
2675		return 0;
 
 
 
 
 
 
 
 
 
 
2676	return 1;
2677}
2678
2679/* scrub extent tries to collect up to 64 kB for each bio */
2680static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2681			u64 logical, u32 len,
2682			u64 physical, struct btrfs_device *dev, u64 flags,
2683			u64 gen, int mirror_num)
2684{
2685	struct btrfs_device *src_dev = dev;
2686	u64 src_physical = physical;
2687	int src_mirror = mirror_num;
2688	int ret;
2689	u8 csum[BTRFS_CSUM_SIZE];
2690	u32 blocksize;
2691
2692	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2693		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2694			blocksize = map->stripe_len;
2695		else
2696			blocksize = sctx->fs_info->sectorsize;
2697		spin_lock(&sctx->stat_lock);
2698		sctx->stat.data_extents_scrubbed++;
2699		sctx->stat.data_bytes_scrubbed += len;
2700		spin_unlock(&sctx->stat_lock);
2701	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2702		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2703			blocksize = map->stripe_len;
2704		else
2705			blocksize = sctx->fs_info->nodesize;
2706		spin_lock(&sctx->stat_lock);
2707		sctx->stat.tree_extents_scrubbed++;
2708		sctx->stat.tree_bytes_scrubbed += len;
2709		spin_unlock(&sctx->stat_lock);
2710	} else {
2711		blocksize = sctx->fs_info->sectorsize;
2712		WARN_ON(1);
2713	}
2714
2715	/*
2716	 * For dev-replace case, we can have @dev being a missing device.
2717	 * Regular scrub will avoid its execution on missing device at all,
2718	 * as that would trigger tons of read error.
2719	 *
2720	 * Reading from missing device will cause read error counts to
2721	 * increase unnecessarily.
2722	 * So here we change the read source to a good mirror.
2723	 */
2724	if (sctx->is_dev_replace && !dev->bdev)
2725		scrub_find_good_copy(sctx->fs_info, logical, len, &src_physical,
2726				     &src_dev, &src_mirror);
2727	while (len) {
2728		u32 l = min(len, blocksize);
2729		int have_csum = 0;
2730
2731		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2732			/* push csums to sbio */
2733			have_csum = scrub_find_csum(sctx, logical, csum);
2734			if (have_csum == 0)
2735				++sctx->stat.no_csum;
2736		}
2737		ret = scrub_sectors(sctx, logical, l, src_physical, src_dev,
2738				    flags, gen, src_mirror,
2739				    have_csum ? csum : NULL, physical);
2740		if (ret)
2741			return ret;
2742		len -= l;
2743		logical += l;
2744		physical += l;
2745		src_physical += l;
2746	}
2747	return 0;
2748}
2749
2750static int scrub_sectors_for_parity(struct scrub_parity *sparity,
2751				  u64 logical, u32 len,
2752				  u64 physical, struct btrfs_device *dev,
2753				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2754{
2755	struct scrub_ctx *sctx = sparity->sctx;
2756	struct scrub_block *sblock;
2757	const u32 sectorsize = sctx->fs_info->sectorsize;
2758	int index;
2759
2760	ASSERT(IS_ALIGNED(len, sectorsize));
2761
2762	sblock = alloc_scrub_block(sctx, dev, logical, physical, physical, mirror_num);
2763	if (!sblock) {
2764		spin_lock(&sctx->stat_lock);
2765		sctx->stat.malloc_errors++;
2766		spin_unlock(&sctx->stat_lock);
2767		return -ENOMEM;
2768	}
2769
 
 
 
 
 
2770	sblock->sparity = sparity;
2771	scrub_parity_get(sparity);
2772
2773	for (index = 0; len > 0; index++) {
2774		struct scrub_sector *sector;
 
2775
2776		sector = alloc_scrub_sector(sblock, logical);
2777		if (!sector) {
 
2778			spin_lock(&sctx->stat_lock);
2779			sctx->stat.malloc_errors++;
2780			spin_unlock(&sctx->stat_lock);
2781			scrub_block_put(sblock);
2782			return -ENOMEM;
2783		}
2784		sblock->sectors[index] = sector;
 
 
 
2785		/* For scrub parity */
2786		scrub_sector_get(sector);
2787		list_add_tail(&sector->list, &sparity->sectors_list);
2788		sector->flags = flags;
2789		sector->generation = gen;
 
 
 
 
 
2790		if (csum) {
2791			sector->have_csum = 1;
2792			memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2793		} else {
2794			sector->have_csum = 0;
2795		}
2796
2797		/* Iterate over the stripe range in sectorsize steps */
2798		len -= sectorsize;
2799		logical += sectorsize;
2800		physical += sectorsize;
 
 
2801	}
2802
2803	WARN_ON(sblock->sector_count == 0);
2804	for (index = 0; index < sblock->sector_count; index++) {
2805		struct scrub_sector *sector = sblock->sectors[index];
2806		int ret;
2807
2808		ret = scrub_add_sector_to_rd_bio(sctx, sector);
2809		if (ret) {
2810			scrub_block_put(sblock);
2811			return ret;
2812		}
2813	}
2814
2815	/* Last one frees, either here or in bio completion for last sector */
2816	scrub_block_put(sblock);
2817	return 0;
2818}
2819
2820static int scrub_extent_for_parity(struct scrub_parity *sparity,
2821				   u64 logical, u32 len,
2822				   u64 physical, struct btrfs_device *dev,
2823				   u64 flags, u64 gen, int mirror_num)
2824{
2825	struct scrub_ctx *sctx = sparity->sctx;
2826	int ret;
2827	u8 csum[BTRFS_CSUM_SIZE];
2828	u32 blocksize;
2829
2830	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2831		scrub_parity_mark_sectors_error(sparity, logical, len);
2832		return 0;
2833	}
2834
2835	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2836		blocksize = sparity->stripe_len;
2837	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2838		blocksize = sparity->stripe_len;
2839	} else {
2840		blocksize = sctx->fs_info->sectorsize;
2841		WARN_ON(1);
2842	}
2843
2844	while (len) {
2845		u32 l = min(len, blocksize);
2846		int have_csum = 0;
2847
2848		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2849			/* push csums to sbio */
2850			have_csum = scrub_find_csum(sctx, logical, csum);
2851			if (have_csum == 0)
2852				goto skip;
2853		}
2854		ret = scrub_sectors_for_parity(sparity, logical, l, physical, dev,
2855					     flags, gen, mirror_num,
2856					     have_csum ? csum : NULL);
2857		if (ret)
2858			return ret;
2859skip:
2860		len -= l;
2861		logical += l;
2862		physical += l;
2863	}
2864	return 0;
2865}
2866
2867/*
2868 * Given a physical address, this will calculate it's
2869 * logical offset. if this is a parity stripe, it will return
2870 * the most left data stripe's logical offset.
2871 *
2872 * return 0 if it is a data stripe, 1 means parity stripe.
2873 */
2874static int get_raid56_logic_offset(u64 physical, int num,
2875				   struct map_lookup *map, u64 *offset,
2876				   u64 *stripe_start)
2877{
2878	int i;
2879	int j = 0;
2880	u64 stripe_nr;
2881	u64 last_offset;
2882	u32 stripe_index;
2883	u32 rot;
2884	const int data_stripes = nr_data_stripes(map);
2885
2886	last_offset = (physical - map->stripes[num].physical) * data_stripes;
2887	if (stripe_start)
2888		*stripe_start = last_offset;
2889
2890	*offset = last_offset;
2891	for (i = 0; i < data_stripes; i++) {
2892		*offset = last_offset + i * map->stripe_len;
2893
2894		stripe_nr = div64_u64(*offset, map->stripe_len);
2895		stripe_nr = div_u64(stripe_nr, data_stripes);
2896
2897		/* Work out the disk rotation on this stripe-set */
2898		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2899		/* calculate which stripe this data locates */
2900		rot += i;
2901		stripe_index = rot % map->num_stripes;
2902		if (stripe_index == num)
2903			return 0;
2904		if (stripe_index < num)
2905			j++;
2906	}
2907	*offset = last_offset + j * map->stripe_len;
2908	return 1;
2909}
2910
2911static void scrub_free_parity(struct scrub_parity *sparity)
2912{
2913	struct scrub_ctx *sctx = sparity->sctx;
2914	struct scrub_sector *curr, *next;
2915	int nbits;
2916
2917	nbits = bitmap_weight(&sparity->ebitmap, sparity->nsectors);
2918	if (nbits) {
2919		spin_lock(&sctx->stat_lock);
2920		sctx->stat.read_errors += nbits;
2921		sctx->stat.uncorrectable_errors += nbits;
2922		spin_unlock(&sctx->stat_lock);
2923	}
2924
2925	list_for_each_entry_safe(curr, next, &sparity->sectors_list, list) {
2926		list_del_init(&curr->list);
2927		scrub_sector_put(curr);
2928	}
2929
2930	kfree(sparity);
2931}
2932
2933static void scrub_parity_bio_endio_worker(struct work_struct *work)
2934{
2935	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2936						    work);
2937	struct scrub_ctx *sctx = sparity->sctx;
2938
2939	btrfs_bio_counter_dec(sctx->fs_info);
2940	scrub_free_parity(sparity);
2941	scrub_pending_bio_dec(sctx);
2942}
2943
2944static void scrub_parity_bio_endio(struct bio *bio)
2945{
2946	struct scrub_parity *sparity = bio->bi_private;
2947	struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2948
2949	if (bio->bi_status)
2950		bitmap_or(&sparity->ebitmap, &sparity->ebitmap,
2951			  &sparity->dbitmap, sparity->nsectors);
2952
2953	bio_put(bio);
2954
2955	INIT_WORK(&sparity->work, scrub_parity_bio_endio_worker);
2956	queue_work(fs_info->scrub_parity_workers, &sparity->work);
 
2957}
2958
2959static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2960{
2961	struct scrub_ctx *sctx = sparity->sctx;
2962	struct btrfs_fs_info *fs_info = sctx->fs_info;
2963	struct bio *bio;
2964	struct btrfs_raid_bio *rbio;
2965	struct btrfs_io_context *bioc = NULL;
2966	u64 length;
2967	int ret;
2968
2969	if (!bitmap_andnot(&sparity->dbitmap, &sparity->dbitmap,
2970			   &sparity->ebitmap, sparity->nsectors))
2971		goto out;
2972
2973	length = sparity->logic_end - sparity->logic_start;
2974
2975	btrfs_bio_counter_inc_blocked(fs_info);
2976	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2977			       &length, &bioc);
2978	if (ret || !bioc || !bioc->raid_map)
2979		goto bioc_out;
2980
2981	bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2982	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2983	bio->bi_private = sparity;
2984	bio->bi_end_io = scrub_parity_bio_endio;
2985
2986	rbio = raid56_parity_alloc_scrub_rbio(bio, bioc,
2987					      sparity->scrub_dev,
2988					      &sparity->dbitmap,
2989					      sparity->nsectors);
2990	btrfs_put_bioc(bioc);
2991	if (!rbio)
2992		goto rbio_out;
2993
2994	scrub_pending_bio_inc(sctx);
2995	raid56_parity_submit_scrub_rbio(rbio);
2996	return;
2997
2998rbio_out:
2999	bio_put(bio);
3000bioc_out:
3001	btrfs_bio_counter_dec(fs_info);
3002	bitmap_or(&sparity->ebitmap, &sparity->ebitmap, &sparity->dbitmap,
 
3003		  sparity->nsectors);
3004	spin_lock(&sctx->stat_lock);
3005	sctx->stat.malloc_errors++;
3006	spin_unlock(&sctx->stat_lock);
3007out:
3008	scrub_free_parity(sparity);
3009}
3010
 
 
 
 
 
3011static void scrub_parity_get(struct scrub_parity *sparity)
3012{
3013	refcount_inc(&sparity->refs);
3014}
3015
3016static void scrub_parity_put(struct scrub_parity *sparity)
3017{
3018	if (!refcount_dec_and_test(&sparity->refs))
3019		return;
3020
3021	scrub_parity_check_and_repair(sparity);
3022}
3023
3024/*
3025 * Return 0 if the extent item range covers any byte of the range.
3026 * Return <0 if the extent item is before @search_start.
3027 * Return >0 if the extent item is after @start_start + @search_len.
3028 */
3029static int compare_extent_item_range(struct btrfs_path *path,
3030				     u64 search_start, u64 search_len)
3031{
3032	struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
3033	u64 len;
3034	struct btrfs_key key;
3035
3036	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3037	ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
3038	       key.type == BTRFS_METADATA_ITEM_KEY);
3039	if (key.type == BTRFS_METADATA_ITEM_KEY)
3040		len = fs_info->nodesize;
3041	else
3042		len = key.offset;
3043
3044	if (key.objectid + len <= search_start)
3045		return -1;
3046	if (key.objectid >= search_start + search_len)
3047		return 1;
3048	return 0;
3049}
3050
3051/*
3052 * Locate one extent item which covers any byte in range
3053 * [@search_start, @search_start + @search_length)
3054 *
3055 * If the path is not initialized, we will initialize the search by doing
3056 * a btrfs_search_slot().
3057 * If the path is already initialized, we will use the path as the initial
3058 * slot, to avoid duplicated btrfs_search_slot() calls.
3059 *
3060 * NOTE: If an extent item starts before @search_start, we will still
3061 * return the extent item. This is for data extent crossing stripe boundary.
3062 *
3063 * Return 0 if we found such extent item, and @path will point to the extent item.
3064 * Return >0 if no such extent item can be found, and @path will be released.
3065 * Return <0 if hit fatal error, and @path will be released.
3066 */
3067static int find_first_extent_item(struct btrfs_root *extent_root,
3068				  struct btrfs_path *path,
3069				  u64 search_start, u64 search_len)
3070{
3071	struct btrfs_fs_info *fs_info = extent_root->fs_info;
3072	struct btrfs_key key;
3073	int ret;
3074
3075	/* Continue using the existing path */
3076	if (path->nodes[0])
3077		goto search_forward;
3078
3079	if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3080		key.type = BTRFS_METADATA_ITEM_KEY;
3081	else
3082		key.type = BTRFS_EXTENT_ITEM_KEY;
3083	key.objectid = search_start;
3084	key.offset = (u64)-1;
3085
3086	ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3087	if (ret < 0)
3088		return ret;
3089
3090	ASSERT(ret > 0);
3091	/*
3092	 * Here we intentionally pass 0 as @min_objectid, as there could be
3093	 * an extent item starting before @search_start.
3094	 */
3095	ret = btrfs_previous_extent_item(extent_root, path, 0);
3096	if (ret < 0)
3097		return ret;
3098	/*
3099	 * No matter whether we have found an extent item, the next loop will
3100	 * properly do every check on the key.
3101	 */
3102search_forward:
3103	while (true) {
3104		btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3105		if (key.objectid >= search_start + search_len)
3106			break;
3107		if (key.type != BTRFS_METADATA_ITEM_KEY &&
3108		    key.type != BTRFS_EXTENT_ITEM_KEY)
3109			goto next;
3110
3111		ret = compare_extent_item_range(path, search_start, search_len);
3112		if (ret == 0)
3113			return ret;
3114		if (ret > 0)
3115			break;
3116next:
3117		path->slots[0]++;
3118		if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
3119			ret = btrfs_next_leaf(extent_root, path);
3120			if (ret) {
3121				/* Either no more item or fatal error */
3122				btrfs_release_path(path);
3123				return ret;
3124			}
3125		}
3126	}
3127	btrfs_release_path(path);
3128	return 1;
3129}
3130
3131static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
3132			    u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
3133{
3134	struct btrfs_key key;
3135	struct btrfs_extent_item *ei;
3136
3137	btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3138	ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
3139	       key.type == BTRFS_EXTENT_ITEM_KEY);
3140	*extent_start_ret = key.objectid;
3141	if (key.type == BTRFS_METADATA_ITEM_KEY)
3142		*size_ret = path->nodes[0]->fs_info->nodesize;
3143	else
3144		*size_ret = key.offset;
3145	ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
3146	*flags_ret = btrfs_extent_flags(path->nodes[0], ei);
3147	*generation_ret = btrfs_extent_generation(path->nodes[0], ei);
3148}
3149
3150static bool does_range_cross_boundary(u64 extent_start, u64 extent_len,
3151				      u64 boundary_start, u64 boudary_len)
3152{
3153	return (extent_start < boundary_start &&
3154		extent_start + extent_len > boundary_start) ||
3155	       (extent_start < boundary_start + boudary_len &&
3156		extent_start + extent_len > boundary_start + boudary_len);
3157}
3158
3159static int scrub_raid56_data_stripe_for_parity(struct scrub_ctx *sctx,
3160					       struct scrub_parity *sparity,
3161					       struct map_lookup *map,
3162					       struct btrfs_device *sdev,
3163					       struct btrfs_path *path,
3164					       u64 logical)
3165{
3166	struct btrfs_fs_info *fs_info = sctx->fs_info;
3167	struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
3168	struct btrfs_root *csum_root = btrfs_csum_root(fs_info, logical);
3169	u64 cur_logical = logical;
3170	int ret;
3171
3172	ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3173
3174	/* Path must not be populated */
3175	ASSERT(!path->nodes[0]);
3176
3177	while (cur_logical < logical + map->stripe_len) {
3178		struct btrfs_io_context *bioc = NULL;
3179		struct btrfs_device *extent_dev;
3180		u64 extent_start;
3181		u64 extent_size;
3182		u64 mapped_length;
3183		u64 extent_flags;
3184		u64 extent_gen;
3185		u64 extent_physical;
3186		u64 extent_mirror_num;
3187
3188		ret = find_first_extent_item(extent_root, path, cur_logical,
3189					     logical + map->stripe_len - cur_logical);
3190		/* No more extent item in this data stripe */
3191		if (ret > 0) {
3192			ret = 0;
3193			break;
3194		}
3195		if (ret < 0)
3196			break;
3197		get_extent_info(path, &extent_start, &extent_size, &extent_flags,
3198				&extent_gen);
3199
3200		/* Metadata should not cross stripe boundaries */
3201		if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3202		    does_range_cross_boundary(extent_start, extent_size,
3203					      logical, map->stripe_len)) {
3204			btrfs_err(fs_info,
3205	"scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3206				  extent_start, logical);
3207			spin_lock(&sctx->stat_lock);
3208			sctx->stat.uncorrectable_errors++;
3209			spin_unlock(&sctx->stat_lock);
3210			cur_logical += extent_size;
3211			continue;
3212		}
3213
3214		/* Skip hole range which doesn't have any extent */
3215		cur_logical = max(extent_start, cur_logical);
3216
3217		/* Truncate the range inside this data stripe */
3218		extent_size = min(extent_start + extent_size,
3219				  logical + map->stripe_len) - cur_logical;
3220		extent_start = cur_logical;
3221		ASSERT(extent_size <= U32_MAX);
3222
3223		scrub_parity_mark_sectors_data(sparity, extent_start, extent_size);
3224
3225		mapped_length = extent_size;
3226		ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_start,
3227				      &mapped_length, &bioc, 0);
3228		if (!ret && (!bioc || mapped_length < extent_size))
3229			ret = -EIO;
3230		if (ret) {
3231			btrfs_put_bioc(bioc);
3232			scrub_parity_mark_sectors_error(sparity, extent_start,
3233							extent_size);
3234			break;
3235		}
3236		extent_physical = bioc->stripes[0].physical;
3237		extent_mirror_num = bioc->mirror_num;
3238		extent_dev = bioc->stripes[0].dev;
3239		btrfs_put_bioc(bioc);
3240
3241		ret = btrfs_lookup_csums_list(csum_root, extent_start,
3242					      extent_start + extent_size - 1,
3243					      &sctx->csum_list, 1, false);
3244		if (ret) {
3245			scrub_parity_mark_sectors_error(sparity, extent_start,
3246							extent_size);
3247			break;
3248		}
3249
3250		ret = scrub_extent_for_parity(sparity, extent_start,
3251					      extent_size, extent_physical,
3252					      extent_dev, extent_flags,
3253					      extent_gen, extent_mirror_num);
3254		scrub_free_csums(sctx);
3255
3256		if (ret) {
3257			scrub_parity_mark_sectors_error(sparity, extent_start,
3258							extent_size);
3259			break;
3260		}
3261
3262		cond_resched();
3263		cur_logical += extent_size;
3264	}
3265	btrfs_release_path(path);
3266	return ret;
3267}
3268
3269static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3270						  struct map_lookup *map,
3271						  struct btrfs_device *sdev,
 
3272						  u64 logic_start,
3273						  u64 logic_end)
3274{
3275	struct btrfs_fs_info *fs_info = sctx->fs_info;
3276	struct btrfs_path *path;
3277	u64 cur_logical;
 
 
 
3278	int ret;
 
 
 
 
 
 
 
 
 
3279	struct scrub_parity *sparity;
3280	int nsectors;
 
 
 
3281
3282	path = btrfs_alloc_path();
3283	if (!path) {
3284		spin_lock(&sctx->stat_lock);
3285		sctx->stat.malloc_errors++;
3286		spin_unlock(&sctx->stat_lock);
3287		return -ENOMEM;
3288	}
3289	path->search_commit_root = 1;
3290	path->skip_locking = 1;
3291
3292	ASSERT(map->stripe_len <= U32_MAX);
3293	nsectors = map->stripe_len >> fs_info->sectorsize_bits;
3294	ASSERT(nsectors <= BITS_PER_LONG);
3295	sparity = kzalloc(sizeof(struct scrub_parity), GFP_NOFS);
3296	if (!sparity) {
3297		spin_lock(&sctx->stat_lock);
3298		sctx->stat.malloc_errors++;
3299		spin_unlock(&sctx->stat_lock);
3300		btrfs_free_path(path);
3301		return -ENOMEM;
3302	}
3303
3304	ASSERT(map->stripe_len <= U32_MAX);
3305	sparity->stripe_len = map->stripe_len;
3306	sparity->nsectors = nsectors;
3307	sparity->sctx = sctx;
3308	sparity->scrub_dev = sdev;
3309	sparity->logic_start = logic_start;
3310	sparity->logic_end = logic_end;
3311	refcount_set(&sparity->refs, 1);
3312	INIT_LIST_HEAD(&sparity->sectors_list);
 
 
3313
3314	ret = 0;
3315	for (cur_logical = logic_start; cur_logical < logic_end;
3316	     cur_logical += map->stripe_len) {
3317		ret = scrub_raid56_data_stripe_for_parity(sctx, sparity, map,
3318							  sdev, path, cur_logical);
3319		if (ret < 0)
3320			break;
3321	}
3322
3323	scrub_parity_put(sparity);
3324	scrub_submit(sctx);
3325	mutex_lock(&sctx->wr_lock);
3326	scrub_wr_submit(sctx);
3327	mutex_unlock(&sctx->wr_lock);
3328
3329	btrfs_free_path(path);
3330	return ret < 0 ? ret : 0;
3331}
3332
3333static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3334{
3335	if (!btrfs_is_zoned(sctx->fs_info))
3336		return;
 
 
 
 
 
 
 
 
3337
3338	sctx->flush_all_writes = true;
3339	scrub_submit(sctx);
3340	mutex_lock(&sctx->wr_lock);
3341	scrub_wr_submit(sctx);
3342	mutex_unlock(&sctx->wr_lock);
 
 
 
 
 
 
 
3343
3344	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3345}
 
 
3346
3347static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3348					u64 physical, u64 physical_end)
3349{
3350	struct btrfs_fs_info *fs_info = sctx->fs_info;
3351	int ret = 0;
3352
3353	if (!btrfs_is_zoned(fs_info))
3354		return 0;
 
 
3355
3356	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
 
3357
3358	mutex_lock(&sctx->wr_lock);
3359	if (sctx->write_pointer < physical_end) {
3360		ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3361						    physical,
3362						    sctx->write_pointer);
3363		if (ret)
3364			btrfs_err(fs_info,
3365				  "zoned: failed to recover write pointer");
3366	}
3367	mutex_unlock(&sctx->wr_lock);
3368	btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3369
3370	return ret;
3371}
3372
3373/*
3374 * Scrub one range which can only has simple mirror based profile.
3375 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
3376 *  RAID0/RAID10).
3377 *
3378 * Since we may need to handle a subset of block group, we need @logical_start
3379 * and @logical_length parameter.
3380 */
3381static int scrub_simple_mirror(struct scrub_ctx *sctx,
3382			       struct btrfs_root *extent_root,
3383			       struct btrfs_root *csum_root,
3384			       struct btrfs_block_group *bg,
3385			       struct map_lookup *map,
3386			       u64 logical_start, u64 logical_length,
3387			       struct btrfs_device *device,
3388			       u64 physical, int mirror_num)
3389{
3390	struct btrfs_fs_info *fs_info = sctx->fs_info;
3391	const u64 logical_end = logical_start + logical_length;
3392	/* An artificial limit, inherit from old scrub behavior */
3393	const u32 max_length = SZ_64K;
3394	struct btrfs_path path = { 0 };
3395	u64 cur_logical = logical_start;
3396	int ret;
3397
3398	/* The range must be inside the bg */
3399	ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
 
 
3400
3401	path.search_commit_root = 1;
3402	path.skip_locking = 1;
3403	/* Go through each extent items inside the logical range */
3404	while (cur_logical < logical_end) {
3405		u64 extent_start;
3406		u64 extent_len;
3407		u64 extent_flags;
3408		u64 extent_gen;
3409		u64 scrub_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3410
3411		/* Canceled? */
3412		if (atomic_read(&fs_info->scrub_cancel_req) ||
3413		    atomic_read(&sctx->cancel_req)) {
3414			ret = -ECANCELED;
3415			break;
3416		}
3417		/* Paused? */
3418		if (atomic_read(&fs_info->scrub_pause_req)) {
3419			/* Push queued extents */
3420			sctx->flush_all_writes = true;
3421			scrub_submit(sctx);
3422			mutex_lock(&sctx->wr_lock);
3423			scrub_wr_submit(sctx);
3424			mutex_unlock(&sctx->wr_lock);
3425			wait_event(sctx->list_wait,
3426				   atomic_read(&sctx->bios_in_flight) == 0);
3427			sctx->flush_all_writes = false;
3428			scrub_blocked_if_needed(fs_info);
3429		}
3430		/* Block group removed? */
3431		spin_lock(&bg->lock);
3432		if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
3433			spin_unlock(&bg->lock);
3434			ret = 0;
3435			break;
3436		}
3437		spin_unlock(&bg->lock);
3438
3439		ret = find_first_extent_item(extent_root, &path, cur_logical,
3440					     logical_end - cur_logical);
3441		if (ret > 0) {
3442			/* No more extent, just update the accounting */
3443			sctx->stat.last_physical = physical + logical_length;
3444			ret = 0;
3445			break;
3446		}
3447		if (ret < 0)
3448			break;
3449		get_extent_info(&path, &extent_start, &extent_len,
3450				&extent_flags, &extent_gen);
3451		/* Skip hole range which doesn't have any extent */
3452		cur_logical = max(extent_start, cur_logical);
3453
3454		/*
3455		 * Scrub len has three limits:
3456		 * - Extent size limit
3457		 * - Scrub range limit
3458		 *   This is especially imporatant for RAID0/RAID10 to reuse
3459		 *   this function
3460		 * - Max scrub size limit
3461		 */
3462		scrub_len = min(min(extent_start + extent_len,
3463				    logical_end), cur_logical + max_length) -
3464			    cur_logical;
3465
3466		if (extent_flags & BTRFS_EXTENT_FLAG_DATA) {
3467			ret = btrfs_lookup_csums_list(csum_root, cur_logical,
3468					cur_logical + scrub_len - 1,
3469					&sctx->csum_list, 1, false);
3470			if (ret)
3471				break;
3472		}
3473		if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3474		    does_range_cross_boundary(extent_start, extent_len,
3475					      logical_start, logical_length)) {
3476			btrfs_err(fs_info,
3477"scrub: tree block %llu spanning boundaries, ignored. boundary=[%llu, %llu)",
3478				  extent_start, logical_start, logical_end);
3479			spin_lock(&sctx->stat_lock);
3480			sctx->stat.uncorrectable_errors++;
3481			spin_unlock(&sctx->stat_lock);
3482			cur_logical += scrub_len;
3483			continue;
3484		}
3485		ret = scrub_extent(sctx, map, cur_logical, scrub_len,
3486				   cur_logical - logical_start + physical,
3487				   device, extent_flags, extent_gen,
3488				   mirror_num);
3489		scrub_free_csums(sctx);
3490		if (ret)
3491			break;
3492		if (sctx->is_dev_replace)
3493			sync_replace_for_zoned(sctx);
3494		cur_logical += scrub_len;
3495		/* Don't hold CPU for too long time */
3496		cond_resched();
3497	}
3498	btrfs_release_path(&path);
3499	return ret;
3500}
3501
3502/* Calculate the full stripe length for simple stripe based profiles */
3503static u64 simple_stripe_full_stripe_len(const struct map_lookup *map)
3504{
3505	ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3506			    BTRFS_BLOCK_GROUP_RAID10));
3507
3508	return map->num_stripes / map->sub_stripes * map->stripe_len;
3509}
 
 
3510
3511/* Get the logical bytenr for the stripe */
3512static u64 simple_stripe_get_logical(struct map_lookup *map,
3513				     struct btrfs_block_group *bg,
3514				     int stripe_index)
3515{
3516	ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3517			    BTRFS_BLOCK_GROUP_RAID10));
3518	ASSERT(stripe_index < map->num_stripes);
3519
3520	/*
3521	 * (stripe_index / sub_stripes) gives how many data stripes we need to
3522	 * skip.
3523	 */
3524	return (stripe_index / map->sub_stripes) * map->stripe_len + bg->start;
3525}
3526
3527/* Get the mirror number for the stripe */
3528static int simple_stripe_mirror_num(struct map_lookup *map, int stripe_index)
3529{
3530	ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3531			    BTRFS_BLOCK_GROUP_RAID10));
3532	ASSERT(stripe_index < map->num_stripes);
3533
3534	/* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
3535	return stripe_index % map->sub_stripes + 1;
3536}
3537
3538static int scrub_simple_stripe(struct scrub_ctx *sctx,
3539			       struct btrfs_root *extent_root,
3540			       struct btrfs_root *csum_root,
3541			       struct btrfs_block_group *bg,
3542			       struct map_lookup *map,
3543			       struct btrfs_device *device,
3544			       int stripe_index)
3545{
3546	const u64 logical_increment = simple_stripe_full_stripe_len(map);
3547	const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
3548	const u64 orig_physical = map->stripes[stripe_index].physical;
3549	const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
3550	u64 cur_logical = orig_logical;
3551	u64 cur_physical = orig_physical;
3552	int ret = 0;
3553
3554	while (cur_logical < bg->start + bg->length) {
3555		/*
3556		 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
3557		 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
3558		 * this stripe.
3559		 */
3560		ret = scrub_simple_mirror(sctx, extent_root, csum_root, bg, map,
3561					  cur_logical, map->stripe_len, device,
3562					  cur_physical, mirror_num);
3563		if (ret)
3564			return ret;
3565		/* Skip to next stripe which belongs to the target device */
3566		cur_logical += logical_increment;
3567		/* For physical offset, we just go to next stripe */
3568		cur_physical += map->stripe_len;
3569	}
3570	return ret;
 
 
 
 
 
 
 
 
 
 
 
3571}
3572
3573static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3574					   struct btrfs_block_group *bg,
3575					   struct extent_map *em,
3576					   struct btrfs_device *scrub_dev,
3577					   int stripe_index)
 
3578{
3579	struct btrfs_path *path;
3580	struct btrfs_fs_info *fs_info = sctx->fs_info;
3581	struct btrfs_root *root;
3582	struct btrfs_root *csum_root;
 
3583	struct blk_plug plug;
3584	struct map_lookup *map = em->map_lookup;
3585	const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3586	const u64 chunk_logical = bg->start;
3587	int ret;
3588	u64 physical = map->stripes[stripe_index].physical;
3589	const u64 dev_stripe_len = btrfs_calc_stripe_length(em);
3590	const u64 physical_end = physical + dev_stripe_len;
 
3591	u64 logical;
3592	u64 logic_end;
3593	/* The logical increment after finishing one stripe */
3594	u64 increment;
3595	/* Offset inside the chunk */
 
 
 
 
 
3596	u64 offset;
 
 
 
3597	u64 stripe_logical;
3598	u64 stripe_end;
 
 
3599	int stop_loop = 0;
3600
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3601	path = btrfs_alloc_path();
3602	if (!path)
3603		return -ENOMEM;
3604
 
 
 
 
 
 
3605	/*
3606	 * work on commit root. The related disk blocks are static as
3607	 * long as COW is applied. This means, it is save to rewrite
3608	 * them to repair disk errors without any race conditions
3609	 */
3610	path->search_commit_root = 1;
3611	path->skip_locking = 1;
3612	path->reada = READA_FORWARD;
3613
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3614	wait_event(sctx->list_wait,
3615		   atomic_read(&sctx->bios_in_flight) == 0);
3616	scrub_blocked_if_needed(fs_info);
3617
3618	root = btrfs_extent_root(fs_info, bg->start);
3619	csum_root = btrfs_csum_root(fs_info, bg->start);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3620
3621	/*
3622	 * collect all data csums for the stripe to avoid seeking during
3623	 * the scrub. This might currently (crc32) end up to be about 1MB
3624	 */
3625	blk_start_plug(&plug);
3626
3627	if (sctx->is_dev_replace &&
3628	    btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3629		mutex_lock(&sctx->wr_lock);
3630		sctx->write_pointer = physical;
3631		mutex_unlock(&sctx->wr_lock);
3632		sctx->flush_all_writes = true;
3633	}
3634
3635	/*
3636	 * There used to be a big double loop to handle all profiles using the
3637	 * same routine, which grows larger and more gross over time.
3638	 *
3639	 * So here we handle each profile differently, so simpler profiles
3640	 * have simpler scrubbing function.
3641	 */
3642	if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
3643			 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
3644		/*
3645		 * Above check rules out all complex profile, the remaining
3646		 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
3647		 * mirrored duplication without stripe.
3648		 *
3649		 * Only @physical and @mirror_num needs to calculated using
3650		 * @stripe_index.
3651		 */
3652		ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3653				bg->start, bg->length, scrub_dev,
3654				map->stripes[stripe_index].physical,
3655				stripe_index + 1);
3656		offset = 0;
3657		goto out;
3658	}
3659	if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
3660		ret = scrub_simple_stripe(sctx, root, csum_root, bg, map,
3661					  scrub_dev, stripe_index);
3662		offset = map->stripe_len * (stripe_index / map->sub_stripes);
3663		goto out;
3664	}
 
 
 
 
 
 
 
3665
3666	/* Only RAID56 goes through the old code */
3667	ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3668	ret = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3669
3670	/* Calculate the logical end of the stripe */
3671	get_raid56_logic_offset(physical_end, stripe_index,
3672				map, &logic_end, NULL);
3673	logic_end += chunk_logical;
3674
3675	/* Initialize @offset in case we need to go to out: label */
3676	get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
3677	increment = map->stripe_len * nr_data_stripes(map);
3678
3679	/*
3680	 * Due to the rotation, for RAID56 it's better to iterate each stripe
3681	 * using their physical offset.
3682	 */
3683	while (physical < physical_end) {
3684		ret = get_raid56_logic_offset(physical, stripe_index, map,
3685					      &logical, &stripe_logical);
3686		logical += chunk_logical;
3687		if (ret) {
3688			/* it is parity strip */
3689			stripe_logical += chunk_logical;
3690			stripe_end = stripe_logical + increment;
3691			ret = scrub_raid56_parity(sctx, map, scrub_dev,
3692						  stripe_logical,
3693						  stripe_end);
3694			if (ret)
3695				goto out;
3696			goto next;
 
 
 
 
 
 
 
 
3697		}
3698
3699		/*
3700		 * Now we're at a data stripe, scrub each extents in the range.
3701		 *
3702		 * At this stage, if we ignore the repair part, inside each data
3703		 * stripe it is no different than SINGLE profile.
3704		 * We can reuse scrub_simple_mirror() here, as the repair part
3705		 * is still based on @mirror_num.
3706		 */
3707		ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3708					  logical, map->stripe_len,
3709					  scrub_dev, physical, 1);
3710		if (ret < 0)
3711			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3712next:
 
 
 
 
3713		logical += increment;
3714		physical += map->stripe_len;
3715		spin_lock(&sctx->stat_lock);
3716		if (stop_loop)
3717			sctx->stat.last_physical =
3718				map->stripes[stripe_index].physical + dev_stripe_len;
3719		else
3720			sctx->stat.last_physical = physical;
3721		spin_unlock(&sctx->stat_lock);
3722		if (stop_loop)
3723			break;
3724	}
3725out:
3726	/* push queued extents */
3727	scrub_submit(sctx);
3728	mutex_lock(&sctx->wr_lock);
3729	scrub_wr_submit(sctx);
3730	mutex_unlock(&sctx->wr_lock);
3731
3732	blk_finish_plug(&plug);
3733	btrfs_free_path(path);
3734
3735	if (sctx->is_dev_replace && ret >= 0) {
3736		int ret2;
3737
3738		ret2 = sync_write_pointer_for_zoned(sctx,
3739				chunk_logical + offset,
3740				map->stripes[stripe_index].physical,
3741				physical_end);
3742		if (ret2)
3743			ret = ret2;
3744	}
3745
3746	return ret < 0 ? ret : 0;
3747}
3748
3749static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3750					  struct btrfs_block_group *bg,
3751					  struct btrfs_device *scrub_dev,
 
3752					  u64 dev_offset,
3753					  u64 dev_extent_len)
3754{
3755	struct btrfs_fs_info *fs_info = sctx->fs_info;
3756	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3757	struct map_lookup *map;
3758	struct extent_map *em;
3759	int i;
3760	int ret = 0;
3761
3762	read_lock(&map_tree->lock);
3763	em = lookup_extent_mapping(map_tree, bg->start, bg->length);
3764	read_unlock(&map_tree->lock);
3765
3766	if (!em) {
3767		/*
3768		 * Might have been an unused block group deleted by the cleaner
3769		 * kthread or relocation.
3770		 */
3771		spin_lock(&bg->lock);
3772		if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
3773			ret = -EINVAL;
3774		spin_unlock(&bg->lock);
3775
3776		return ret;
3777	}
3778	if (em->start != bg->start)
 
 
3779		goto out;
3780	if (em->len < dev_extent_len)
 
3781		goto out;
3782
3783	map = em->map_lookup;
3784	for (i = 0; i < map->num_stripes; ++i) {
3785		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3786		    map->stripes[i].physical == dev_offset) {
3787			ret = scrub_stripe(sctx, bg, em, scrub_dev, i);
 
3788			if (ret)
3789				goto out;
3790		}
3791	}
3792out:
3793	free_extent_map(em);
3794
3795	return ret;
3796}
3797
3798static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3799					  struct btrfs_block_group *cache)
3800{
3801	struct btrfs_fs_info *fs_info = cache->fs_info;
3802	struct btrfs_trans_handle *trans;
3803
3804	if (!btrfs_is_zoned(fs_info))
3805		return 0;
3806
3807	btrfs_wait_block_group_reservations(cache);
3808	btrfs_wait_nocow_writers(cache);
3809	btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3810
3811	trans = btrfs_join_transaction(root);
3812	if (IS_ERR(trans))
3813		return PTR_ERR(trans);
3814	return btrfs_commit_transaction(trans);
3815}
3816
3817static noinline_for_stack
3818int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3819			   struct btrfs_device *scrub_dev, u64 start, u64 end)
3820{
3821	struct btrfs_dev_extent *dev_extent = NULL;
3822	struct btrfs_path *path;
3823	struct btrfs_fs_info *fs_info = sctx->fs_info;
3824	struct btrfs_root *root = fs_info->dev_root;
 
3825	u64 chunk_offset;
3826	int ret = 0;
3827	int ro_set;
3828	int slot;
3829	struct extent_buffer *l;
3830	struct btrfs_key key;
3831	struct btrfs_key found_key;
3832	struct btrfs_block_group *cache;
3833	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3834
3835	path = btrfs_alloc_path();
3836	if (!path)
3837		return -ENOMEM;
3838
3839	path->reada = READA_FORWARD;
3840	path->search_commit_root = 1;
3841	path->skip_locking = 1;
3842
3843	key.objectid = scrub_dev->devid;
3844	key.offset = 0ull;
3845	key.type = BTRFS_DEV_EXTENT_KEY;
3846
3847	while (1) {
3848		u64 dev_extent_len;
3849
3850		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3851		if (ret < 0)
3852			break;
3853		if (ret > 0) {
3854			if (path->slots[0] >=
3855			    btrfs_header_nritems(path->nodes[0])) {
3856				ret = btrfs_next_leaf(root, path);
3857				if (ret < 0)
3858					break;
3859				if (ret > 0) {
3860					ret = 0;
3861					break;
3862				}
3863			} else {
3864				ret = 0;
3865			}
3866		}
3867
3868		l = path->nodes[0];
3869		slot = path->slots[0];
3870
3871		btrfs_item_key_to_cpu(l, &found_key, slot);
3872
3873		if (found_key.objectid != scrub_dev->devid)
3874			break;
3875
3876		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3877			break;
3878
3879		if (found_key.offset >= end)
3880			break;
3881
3882		if (found_key.offset < key.offset)
3883			break;
3884
3885		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3886		dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
3887
3888		if (found_key.offset + dev_extent_len <= start)
3889			goto skip;
3890
3891		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3892
3893		/*
3894		 * get a reference on the corresponding block group to prevent
3895		 * the chunk from going away while we scrub it
3896		 */
3897		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3898
3899		/* some chunks are removed but not committed to disk yet,
3900		 * continue scrubbing */
3901		if (!cache)
3902			goto skip;
3903
3904		ASSERT(cache->start <= chunk_offset);
3905		/*
3906		 * We are using the commit root to search for device extents, so
3907		 * that means we could have found a device extent item from a
3908		 * block group that was deleted in the current transaction. The
3909		 * logical start offset of the deleted block group, stored at
3910		 * @chunk_offset, might be part of the logical address range of
3911		 * a new block group (which uses different physical extents).
3912		 * In this case btrfs_lookup_block_group() has returned the new
3913		 * block group, and its start address is less than @chunk_offset.
3914		 *
3915		 * We skip such new block groups, because it's pointless to
3916		 * process them, as we won't find their extents because we search
3917		 * for them using the commit root of the extent tree. For a device
3918		 * replace it's also fine to skip it, we won't miss copying them
3919		 * to the target device because we have the write duplication
3920		 * setup through the regular write path (by btrfs_map_block()),
3921		 * and we have committed a transaction when we started the device
3922		 * replace, right after setting up the device replace state.
3923		 */
3924		if (cache->start < chunk_offset) {
3925			btrfs_put_block_group(cache);
3926			goto skip;
3927		}
3928
3929		if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3930			if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
3931				btrfs_put_block_group(cache);
3932				goto skip;
3933			}
3934		}
3935
3936		/*
3937		 * Make sure that while we are scrubbing the corresponding block
3938		 * group doesn't get its logical address and its device extents
3939		 * reused for another block group, which can possibly be of a
3940		 * different type and different profile. We do this to prevent
3941		 * false error detections and crashes due to bogus attempts to
3942		 * repair extents.
3943		 */
3944		spin_lock(&cache->lock);
3945		if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
3946			spin_unlock(&cache->lock);
3947			btrfs_put_block_group(cache);
3948			goto skip;
3949		}
3950		btrfs_freeze_block_group(cache);
3951		spin_unlock(&cache->lock);
3952
3953		/*
3954		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3955		 * to avoid deadlock caused by:
3956		 * btrfs_inc_block_group_ro()
3957		 * -> btrfs_wait_for_commit()
3958		 * -> btrfs_commit_transaction()
3959		 * -> btrfs_scrub_pause()
3960		 */
3961		scrub_pause_on(fs_info);
3962
3963		/*
3964		 * Don't do chunk preallocation for scrub.
3965		 *
3966		 * This is especially important for SYSTEM bgs, or we can hit
3967		 * -EFBIG from btrfs_finish_chunk_alloc() like:
3968		 * 1. The only SYSTEM bg is marked RO.
3969		 *    Since SYSTEM bg is small, that's pretty common.
3970		 * 2. New SYSTEM bg will be allocated
3971		 *    Due to regular version will allocate new chunk.
3972		 * 3. New SYSTEM bg is empty and will get cleaned up
3973		 *    Before cleanup really happens, it's marked RO again.
3974		 * 4. Empty SYSTEM bg get scrubbed
3975		 *    We go back to 2.
3976		 *
3977		 * This can easily boost the amount of SYSTEM chunks if cleaner
3978		 * thread can't be triggered fast enough, and use up all space
3979		 * of btrfs_super_block::sys_chunk_array
3980		 *
3981		 * While for dev replace, we need to try our best to mark block
3982		 * group RO, to prevent race between:
3983		 * - Write duplication
3984		 *   Contains latest data
3985		 * - Scrub copy
3986		 *   Contains data from commit tree
3987		 *
3988		 * If target block group is not marked RO, nocow writes can
3989		 * be overwritten by scrub copy, causing data corruption.
3990		 * So for dev-replace, it's not allowed to continue if a block
3991		 * group is not RO.
3992		 */
3993		ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3994		if (!ret && sctx->is_dev_replace) {
3995			ret = finish_extent_writes_for_zoned(root, cache);
3996			if (ret) {
3997				btrfs_dec_block_group_ro(cache);
3998				scrub_pause_off(fs_info);
3999				btrfs_put_block_group(cache);
4000				break;
4001			}
4002		}
4003
4004		if (ret == 0) {
4005			ro_set = 1;
4006		} else if (ret == -ENOSPC && !sctx->is_dev_replace) {
4007			/*
4008			 * btrfs_inc_block_group_ro return -ENOSPC when it
4009			 * failed in creating new chunk for metadata.
4010			 * It is not a problem for scrub, because
4011			 * metadata are always cowed, and our scrub paused
4012			 * commit_transactions.
4013			 */
4014			ro_set = 0;
4015		} else if (ret == -ETXTBSY) {
4016			btrfs_warn(fs_info,
4017		   "skipping scrub of block group %llu due to active swapfile",
4018				   cache->start);
4019			scrub_pause_off(fs_info);
4020			ret = 0;
4021			goto skip_unfreeze;
4022		} else {
4023			btrfs_warn(fs_info,
4024				   "failed setting block group ro: %d", ret);
4025			btrfs_unfreeze_block_group(cache);
4026			btrfs_put_block_group(cache);
4027			scrub_pause_off(fs_info);
4028			break;
4029		}
4030
4031		/*
4032		 * Now the target block is marked RO, wait for nocow writes to
4033		 * finish before dev-replace.
4034		 * COW is fine, as COW never overwrites extents in commit tree.
4035		 */
4036		if (sctx->is_dev_replace) {
4037			btrfs_wait_nocow_writers(cache);
4038			btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
4039					cache->length);
4040		}
4041
4042		scrub_pause_off(fs_info);
4043		down_write(&dev_replace->rwsem);
4044		dev_replace->cursor_right = found_key.offset + dev_extent_len;
4045		dev_replace->cursor_left = found_key.offset;
4046		dev_replace->item_needs_writeback = 1;
4047		up_write(&dev_replace->rwsem);
4048
4049		ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
4050				  dev_extent_len);
4051
4052		/*
4053		 * flush, submit all pending read and write bios, afterwards
4054		 * wait for them.
4055		 * Note that in the dev replace case, a read request causes
4056		 * write requests that are submitted in the read completion
4057		 * worker. Therefore in the current situation, it is required
4058		 * that all write requests are flushed, so that all read and
4059		 * write requests are really completed when bios_in_flight
4060		 * changes to 0.
4061		 */
4062		sctx->flush_all_writes = true;
4063		scrub_submit(sctx);
4064		mutex_lock(&sctx->wr_lock);
4065		scrub_wr_submit(sctx);
4066		mutex_unlock(&sctx->wr_lock);
4067
4068		wait_event(sctx->list_wait,
4069			   atomic_read(&sctx->bios_in_flight) == 0);
4070
4071		scrub_pause_on(fs_info);
4072
4073		/*
4074		 * must be called before we decrease @scrub_paused.
4075		 * make sure we don't block transaction commit while
4076		 * we are waiting pending workers finished.
4077		 */
4078		wait_event(sctx->list_wait,
4079			   atomic_read(&sctx->workers_pending) == 0);
4080		sctx->flush_all_writes = false;
4081
4082		scrub_pause_off(fs_info);
4083
4084		if (sctx->is_dev_replace &&
4085		    !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
4086						      cache, found_key.offset))
4087			ro_set = 0;
4088
4089		down_write(&dev_replace->rwsem);
4090		dev_replace->cursor_left = dev_replace->cursor_right;
4091		dev_replace->item_needs_writeback = 1;
4092		up_write(&dev_replace->rwsem);
4093
4094		if (ro_set)
4095			btrfs_dec_block_group_ro(cache);
4096
4097		/*
4098		 * We might have prevented the cleaner kthread from deleting
4099		 * this block group if it was already unused because we raced
4100		 * and set it to RO mode first. So add it back to the unused
4101		 * list, otherwise it might not ever be deleted unless a manual
4102		 * balance is triggered or it becomes used and unused again.
4103		 */
4104		spin_lock(&cache->lock);
4105		if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
4106		    !cache->ro && cache->reserved == 0 && cache->used == 0) {
4107			spin_unlock(&cache->lock);
4108			if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
4109				btrfs_discard_queue_work(&fs_info->discard_ctl,
4110							 cache);
4111			else
4112				btrfs_mark_bg_unused(cache);
4113		} else {
4114			spin_unlock(&cache->lock);
4115		}
4116skip_unfreeze:
4117		btrfs_unfreeze_block_group(cache);
4118		btrfs_put_block_group(cache);
4119		if (ret)
4120			break;
4121		if (sctx->is_dev_replace &&
4122		    atomic64_read(&dev_replace->num_write_errors) > 0) {
4123			ret = -EIO;
4124			break;
4125		}
4126		if (sctx->stat.malloc_errors > 0) {
4127			ret = -ENOMEM;
4128			break;
4129		}
4130skip:
4131		key.offset = found_key.offset + dev_extent_len;
4132		btrfs_release_path(path);
4133	}
4134
4135	btrfs_free_path(path);
4136
4137	return ret;
4138}
4139
4140static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4141					   struct btrfs_device *scrub_dev)
4142{
4143	int	i;
4144	u64	bytenr;
4145	u64	gen;
4146	int	ret;
4147	struct btrfs_fs_info *fs_info = sctx->fs_info;
4148
4149	if (BTRFS_FS_ERROR(fs_info))
4150		return -EROFS;
4151
4152	/* Seed devices of a new filesystem has their own generation. */
4153	if (scrub_dev->fs_devices != fs_info->fs_devices)
4154		gen = scrub_dev->generation;
4155	else
4156		gen = fs_info->last_trans_committed;
4157
4158	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4159		bytenr = btrfs_sb_offset(i);
4160		if (bytenr + BTRFS_SUPER_INFO_SIZE >
4161		    scrub_dev->commit_total_bytes)
4162			break;
4163		if (!btrfs_check_super_location(scrub_dev, bytenr))
4164			continue;
4165
4166		ret = scrub_sectors(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4167				    scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4168				    NULL, bytenr);
4169		if (ret)
4170			return ret;
4171	}
4172	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4173
4174	return 0;
4175}
4176
4177static void scrub_workers_put(struct btrfs_fs_info *fs_info)
4178{
4179	if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
4180					&fs_info->scrub_lock)) {
4181		struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
4182		struct workqueue_struct *scrub_wr_comp =
4183						fs_info->scrub_wr_completion_workers;
4184		struct workqueue_struct *scrub_parity =
4185						fs_info->scrub_parity_workers;
 
 
4186
4187		fs_info->scrub_workers = NULL;
4188		fs_info->scrub_wr_completion_workers = NULL;
4189		fs_info->scrub_parity_workers = NULL;
4190		mutex_unlock(&fs_info->scrub_lock);
4191
4192		if (scrub_workers)
4193			destroy_workqueue(scrub_workers);
4194		if (scrub_wr_comp)
4195			destroy_workqueue(scrub_wr_comp);
4196		if (scrub_parity)
4197			destroy_workqueue(scrub_parity);
4198	}
4199}
4200
4201/*
4202 * get a reference count on fs_info->scrub_workers. start worker if necessary
4203 */
4204static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4205						int is_dev_replace)
4206{
4207	struct workqueue_struct *scrub_workers = NULL;
4208	struct workqueue_struct *scrub_wr_comp = NULL;
4209	struct workqueue_struct *scrub_parity = NULL;
4210	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4211	int max_active = fs_info->thread_pool_size;
4212	int ret = -ENOMEM;
4213
4214	if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
4215		return 0;
4216
4217	scrub_workers = alloc_workqueue("btrfs-scrub", flags,
4218					is_dev_replace ? 1 : max_active);
4219	if (!scrub_workers)
4220		goto fail_scrub_workers;
4221
4222	scrub_wr_comp = alloc_workqueue("btrfs-scrubwrc", flags, max_active);
 
4223	if (!scrub_wr_comp)
4224		goto fail_scrub_wr_completion_workers;
4225
4226	scrub_parity = alloc_workqueue("btrfs-scrubparity", flags, max_active);
 
4227	if (!scrub_parity)
4228		goto fail_scrub_parity_workers;
4229
4230	mutex_lock(&fs_info->scrub_lock);
4231	if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
4232		ASSERT(fs_info->scrub_workers == NULL &&
4233		       fs_info->scrub_wr_completion_workers == NULL &&
4234		       fs_info->scrub_parity_workers == NULL);
4235		fs_info->scrub_workers = scrub_workers;
4236		fs_info->scrub_wr_completion_workers = scrub_wr_comp;
4237		fs_info->scrub_parity_workers = scrub_parity;
4238		refcount_set(&fs_info->scrub_workers_refcnt, 1);
4239		mutex_unlock(&fs_info->scrub_lock);
4240		return 0;
4241	}
4242	/* Other thread raced in and created the workers for us */
4243	refcount_inc(&fs_info->scrub_workers_refcnt);
4244	mutex_unlock(&fs_info->scrub_lock);
4245
4246	ret = 0;
4247	destroy_workqueue(scrub_parity);
4248fail_scrub_parity_workers:
4249	destroy_workqueue(scrub_wr_comp);
4250fail_scrub_wr_completion_workers:
4251	destroy_workqueue(scrub_workers);
4252fail_scrub_workers:
4253	return ret;
4254}
4255
4256int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4257		    u64 end, struct btrfs_scrub_progress *progress,
4258		    int readonly, int is_dev_replace)
4259{
4260	struct btrfs_dev_lookup_args args = { .devid = devid };
4261	struct scrub_ctx *sctx;
4262	int ret;
4263	struct btrfs_device *dev;
4264	unsigned int nofs_flag;
4265	bool need_commit = false;
4266
4267	if (btrfs_fs_closing(fs_info))
4268		return -EAGAIN;
4269
4270	/* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
4271	ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
4272
4273	/*
4274	 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
4275	 * value (max nodesize / min sectorsize), thus nodesize should always
4276	 * be fine.
4277	 */
4278	ASSERT(fs_info->nodesize <=
4279	       SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4280
4281	/* Allocate outside of device_list_mutex */
4282	sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4283	if (IS_ERR(sctx))
4284		return PTR_ERR(sctx);
4285
4286	ret = scrub_workers_get(fs_info, is_dev_replace);
4287	if (ret)
4288		goto out_free_ctx;
4289
4290	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4291	dev = btrfs_find_device(fs_info->fs_devices, &args);
4292	if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4293		     !is_dev_replace)) {
4294		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4295		ret = -ENODEV;
4296		goto out;
4297	}
4298
4299	if (!is_dev_replace && !readonly &&
4300	    !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4301		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4302		btrfs_err_in_rcu(fs_info,
4303			"scrub on devid %llu: filesystem on %s is not writable",
4304				 devid, btrfs_dev_name(dev));
4305		ret = -EROFS;
4306		goto out;
4307	}
4308
4309	mutex_lock(&fs_info->scrub_lock);
4310	if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4311	    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4312		mutex_unlock(&fs_info->scrub_lock);
4313		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4314		ret = -EIO;
4315		goto out;
4316	}
4317
4318	down_read(&fs_info->dev_replace.rwsem);
4319	if (dev->scrub_ctx ||
4320	    (!is_dev_replace &&
4321	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4322		up_read(&fs_info->dev_replace.rwsem);
4323		mutex_unlock(&fs_info->scrub_lock);
4324		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4325		ret = -EINPROGRESS;
4326		goto out;
4327	}
4328	up_read(&fs_info->dev_replace.rwsem);
4329
4330	sctx->readonly = readonly;
4331	dev->scrub_ctx = sctx;
4332	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4333
4334	/*
4335	 * checking @scrub_pause_req here, we can avoid
4336	 * race between committing transaction and scrubbing.
4337	 */
4338	__scrub_blocked_if_needed(fs_info);
4339	atomic_inc(&fs_info->scrubs_running);
4340	mutex_unlock(&fs_info->scrub_lock);
4341
4342	/*
4343	 * In order to avoid deadlock with reclaim when there is a transaction
4344	 * trying to pause scrub, make sure we use GFP_NOFS for all the
4345	 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
4346	 * invoked by our callees. The pausing request is done when the
4347	 * transaction commit starts, and it blocks the transaction until scrub
4348	 * is paused (done at specific points at scrub_stripe() or right above
4349	 * before incrementing fs_info->scrubs_running).
4350	 */
4351	nofs_flag = memalloc_nofs_save();
4352	if (!is_dev_replace) {
4353		u64 old_super_errors;
4354
4355		spin_lock(&sctx->stat_lock);
4356		old_super_errors = sctx->stat.super_errors;
4357		spin_unlock(&sctx->stat_lock);
4358
4359		btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4360		/*
4361		 * by holding device list mutex, we can
4362		 * kick off writing super in log tree sync.
4363		 */
4364		mutex_lock(&fs_info->fs_devices->device_list_mutex);
4365		ret = scrub_supers(sctx, dev);
4366		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4367
4368		spin_lock(&sctx->stat_lock);
4369		/*
4370		 * Super block errors found, but we can not commit transaction
4371		 * at current context, since btrfs_commit_transaction() needs
4372		 * to pause the current running scrub (hold by ourselves).
4373		 */
4374		if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
4375			need_commit = true;
4376		spin_unlock(&sctx->stat_lock);
4377	}
4378
4379	if (!ret)
4380		ret = scrub_enumerate_chunks(sctx, dev, start, end);
4381	memalloc_nofs_restore(nofs_flag);
4382
4383	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4384	atomic_dec(&fs_info->scrubs_running);
4385	wake_up(&fs_info->scrub_pause_wait);
4386
4387	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4388
4389	if (progress)
4390		memcpy(progress, &sctx->stat, sizeof(*progress));
4391
4392	if (!is_dev_replace)
4393		btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4394			ret ? "not finished" : "finished", devid, ret);
4395
4396	mutex_lock(&fs_info->scrub_lock);
4397	dev->scrub_ctx = NULL;
4398	mutex_unlock(&fs_info->scrub_lock);
4399
4400	scrub_workers_put(fs_info);
4401	scrub_put_ctx(sctx);
4402
4403	/*
4404	 * We found some super block errors before, now try to force a
4405	 * transaction commit, as scrub has finished.
4406	 */
4407	if (need_commit) {
4408		struct btrfs_trans_handle *trans;
4409
4410		trans = btrfs_start_transaction(fs_info->tree_root, 0);
4411		if (IS_ERR(trans)) {
4412			ret = PTR_ERR(trans);
4413			btrfs_err(fs_info,
4414	"scrub: failed to start transaction to fix super block errors: %d", ret);
4415			return ret;
4416		}
4417		ret = btrfs_commit_transaction(trans);
4418		if (ret < 0)
4419			btrfs_err(fs_info,
4420	"scrub: failed to commit transaction to fix super block errors: %d", ret);
4421	}
4422	return ret;
4423out:
4424	scrub_workers_put(fs_info);
4425out_free_ctx:
4426	scrub_free_ctx(sctx);
4427
4428	return ret;
4429}
4430
4431void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4432{
4433	mutex_lock(&fs_info->scrub_lock);
4434	atomic_inc(&fs_info->scrub_pause_req);
4435	while (atomic_read(&fs_info->scrubs_paused) !=
4436	       atomic_read(&fs_info->scrubs_running)) {
4437		mutex_unlock(&fs_info->scrub_lock);
4438		wait_event(fs_info->scrub_pause_wait,
4439			   atomic_read(&fs_info->scrubs_paused) ==
4440			   atomic_read(&fs_info->scrubs_running));
4441		mutex_lock(&fs_info->scrub_lock);
4442	}
4443	mutex_unlock(&fs_info->scrub_lock);
4444}
4445
4446void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4447{
4448	atomic_dec(&fs_info->scrub_pause_req);
4449	wake_up(&fs_info->scrub_pause_wait);
4450}
4451
4452int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4453{
4454	mutex_lock(&fs_info->scrub_lock);
4455	if (!atomic_read(&fs_info->scrubs_running)) {
4456		mutex_unlock(&fs_info->scrub_lock);
4457		return -ENOTCONN;
4458	}
4459
4460	atomic_inc(&fs_info->scrub_cancel_req);
4461	while (atomic_read(&fs_info->scrubs_running)) {
4462		mutex_unlock(&fs_info->scrub_lock);
4463		wait_event(fs_info->scrub_pause_wait,
4464			   atomic_read(&fs_info->scrubs_running) == 0);
4465		mutex_lock(&fs_info->scrub_lock);
4466	}
4467	atomic_dec(&fs_info->scrub_cancel_req);
4468	mutex_unlock(&fs_info->scrub_lock);
4469
4470	return 0;
4471}
4472
4473int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4474{
4475	struct btrfs_fs_info *fs_info = dev->fs_info;
4476	struct scrub_ctx *sctx;
4477
4478	mutex_lock(&fs_info->scrub_lock);
4479	sctx = dev->scrub_ctx;
4480	if (!sctx) {
4481		mutex_unlock(&fs_info->scrub_lock);
4482		return -ENOTCONN;
4483	}
4484	atomic_inc(&sctx->cancel_req);
4485	while (dev->scrub_ctx) {
4486		mutex_unlock(&fs_info->scrub_lock);
4487		wait_event(fs_info->scrub_pause_wait,
4488			   dev->scrub_ctx == NULL);
4489		mutex_lock(&fs_info->scrub_lock);
4490	}
4491	mutex_unlock(&fs_info->scrub_lock);
4492
4493	return 0;
4494}
4495
4496int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4497			 struct btrfs_scrub_progress *progress)
4498{
4499	struct btrfs_dev_lookup_args args = { .devid = devid };
4500	struct btrfs_device *dev;
4501	struct scrub_ctx *sctx = NULL;
4502
4503	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4504	dev = btrfs_find_device(fs_info->fs_devices, &args);
4505	if (dev)
4506		sctx = dev->scrub_ctx;
4507	if (sctx)
4508		memcpy(progress, &sctx->stat, sizeof(*progress));
4509	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4510
4511	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4512}
4513
4514static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
4515				 u64 extent_logical, u32 extent_len,
4516				 u64 *extent_physical,
4517				 struct btrfs_device **extent_dev,
4518				 int *extent_mirror_num)
4519{
4520	u64 mapped_length;
4521	struct btrfs_io_context *bioc = NULL;
4522	int ret;
4523
4524	mapped_length = extent_len;
4525	ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4526			      &mapped_length, &bioc, 0);
4527	if (ret || !bioc || mapped_length < extent_len ||
4528	    !bioc->stripes[0].dev->bdev) {
4529		btrfs_put_bioc(bioc);
4530		return;
4531	}
4532
4533	*extent_physical = bioc->stripes[0].physical;
4534	*extent_mirror_num = bioc->mirror_num;
4535	*extent_dev = bioc->stripes[0].dev;
4536	btrfs_put_bioc(bioc);
4537}
v5.9
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   4 */
   5
   6#include <linux/blkdev.h>
   7#include <linux/ratelimit.h>
   8#include <linux/sched/mm.h>
   9#include <crypto/hash.h>
  10#include "ctree.h"
  11#include "discard.h"
  12#include "volumes.h"
  13#include "disk-io.h"
  14#include "ordered-data.h"
  15#include "transaction.h"
  16#include "backref.h"
  17#include "extent_io.h"
  18#include "dev-replace.h"
  19#include "check-integrity.h"
  20#include "rcu-string.h"
  21#include "raid56.h"
  22#include "block-group.h"
 
 
 
 
 
  23
  24/*
  25 * This is only the first step towards a full-features scrub. It reads all
  26 * extent and super block and verifies the checksums. In case a bad checksum
  27 * is found or the extent cannot be read, good data will be written back if
  28 * any can be found.
  29 *
  30 * Future enhancements:
  31 *  - In case an unrepairable extent is encountered, track which files are
  32 *    affected and report them
  33 *  - track and record media errors, throw out bad devices
  34 *  - add a mode to also read unallocated space
  35 */
  36
  37struct scrub_block;
  38struct scrub_ctx;
  39
  40/*
  41 * the following three values only influence the performance.
 
  42 * The last one configures the number of parallel and outstanding I/O
  43 * operations. The first two values configure an upper limit for the number
  44 * of (dynamically allocated) pages that are added to a bio.
  45 */
  46#define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
  47#define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
  48#define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
  49
  50/*
  51 * the following value times PAGE_SIZE needs to be large enough to match the
  52 * largest node/leaf/sector size that shall be supported.
  53 * Values larger than BTRFS_STRIPE_LEN are not supported.
  54 */
  55#define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
 
 
 
 
 
 
 
 
 
 
 
 
 
  56
  57struct scrub_recover {
  58	refcount_t		refs;
  59	struct btrfs_bio	*bbio;
  60	u64			map_length;
  61};
  62
  63struct scrub_page {
  64	struct scrub_block	*sblock;
  65	struct page		*page;
  66	struct btrfs_device	*dev;
  67	struct list_head	list;
  68	u64			flags;  /* extent flags */
  69	u64			generation;
  70	u64			logical;
  71	u64			physical;
  72	u64			physical_for_dev_replace;
  73	atomic_t		refs;
  74	struct {
  75		unsigned int	mirror_num:8;
  76		unsigned int	have_csum:1;
  77		unsigned int	io_error:1;
  78	};
  79	u8			csum[BTRFS_CSUM_SIZE];
  80
  81	struct scrub_recover	*recover;
  82};
  83
  84struct scrub_bio {
  85	int			index;
  86	struct scrub_ctx	*sctx;
  87	struct btrfs_device	*dev;
  88	struct bio		*bio;
  89	blk_status_t		status;
  90	u64			logical;
  91	u64			physical;
  92#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  93	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
  94#else
  95	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
  96#endif
  97	int			page_count;
  98	int			next_free;
  99	struct btrfs_work	work;
 100};
 101
 102struct scrub_block {
 103	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 104	int			page_count;
 105	atomic_t		outstanding_pages;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 106	refcount_t		refs; /* free mem on transition to zero */
 107	struct scrub_ctx	*sctx;
 108	struct scrub_parity	*sparity;
 109	struct {
 110		unsigned int	header_error:1;
 111		unsigned int	checksum_error:1;
 112		unsigned int	no_io_error_seen:1;
 113		unsigned int	generation_error:1; /* also sets header_error */
 114
 115		/* The following is for the data used to check parity */
 116		/* It is for the data with checksum */
 117		unsigned int	data_corrected:1;
 118	};
 119	struct btrfs_work	work;
 120};
 121
 122/* Used for the chunks with parity stripe such RAID5/6 */
 123struct scrub_parity {
 124	struct scrub_ctx	*sctx;
 125
 126	struct btrfs_device	*scrub_dev;
 127
 128	u64			logic_start;
 129
 130	u64			logic_end;
 131
 132	int			nsectors;
 133
 134	u64			stripe_len;
 135
 136	refcount_t		refs;
 137
 138	struct list_head	spages;
 139
 140	/* Work of parity check and repair */
 141	struct btrfs_work	work;
 142
 143	/* Mark the parity blocks which have data */
 144	unsigned long		*dbitmap;
 145
 146	/*
 147	 * Mark the parity blocks which have data, but errors happen when
 148	 * read data or check data
 149	 */
 150	unsigned long		*ebitmap;
 151
 152	unsigned long		bitmap[];
 153};
 154
 155struct scrub_ctx {
 156	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
 157	struct btrfs_fs_info	*fs_info;
 158	int			first_free;
 159	int			curr;
 160	atomic_t		bios_in_flight;
 161	atomic_t		workers_pending;
 162	spinlock_t		list_lock;
 163	wait_queue_head_t	list_wait;
 164	u16			csum_size;
 165	struct list_head	csum_list;
 166	atomic_t		cancel_req;
 167	int			readonly;
 168	int			pages_per_rd_bio;
 
 
 
 
 169
 170	int			is_dev_replace;
 
 171
 172	struct scrub_bio        *wr_curr_bio;
 173	struct mutex            wr_lock;
 174	int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 175	struct btrfs_device     *wr_tgtdev;
 176	bool                    flush_all_writes;
 177
 178	/*
 179	 * statistics
 180	 */
 181	struct btrfs_scrub_progress stat;
 182	spinlock_t		stat_lock;
 183
 184	/*
 185	 * Use a ref counter to avoid use-after-free issues. Scrub workers
 186	 * decrement bios_in_flight and workers_pending and then do a wakeup
 187	 * on the list_wait wait queue. We must ensure the main scrub task
 188	 * doesn't free the scrub context before or while the workers are
 189	 * doing the wakeup() call.
 190	 */
 191	refcount_t              refs;
 192};
 193
 194struct scrub_warning {
 195	struct btrfs_path	*path;
 196	u64			extent_item_size;
 197	const char		*errstr;
 198	u64			physical;
 199	u64			logical;
 200	struct btrfs_device	*dev;
 201};
 202
 203struct full_stripe_lock {
 204	struct rb_node node;
 205	u64 logical;
 206	u64 refs;
 207	struct mutex mutex;
 208};
 209
 210static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 211static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 212static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 213static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 214				     struct scrub_block *sblocks_for_recheck);
 215static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 216				struct scrub_block *sblock,
 217				int retry_failed_mirror);
 218static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 219static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 220					     struct scrub_block *sblock_good);
 221static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 222					    struct scrub_block *sblock_good,
 223					    int page_num, int force_write);
 224static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 225static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 226					   int page_num);
 227static int scrub_checksum_data(struct scrub_block *sblock);
 228static int scrub_checksum_tree_block(struct scrub_block *sblock);
 229static int scrub_checksum_super(struct scrub_block *sblock);
 230static void scrub_block_get(struct scrub_block *sblock);
 231static void scrub_block_put(struct scrub_block *sblock);
 232static void scrub_page_get(struct scrub_page *spage);
 233static void scrub_page_put(struct scrub_page *spage);
 234static void scrub_parity_get(struct scrub_parity *sparity);
 235static void scrub_parity_put(struct scrub_parity *sparity);
 236static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 237				    struct scrub_page *spage);
 238static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 239		       u64 physical, struct btrfs_device *dev, u64 flags,
 240		       u64 gen, int mirror_num, u8 *csum, int force,
 241		       u64 physical_for_dev_replace);
 242static void scrub_bio_end_io(struct bio *bio);
 243static void scrub_bio_end_io_worker(struct btrfs_work *work);
 244static void scrub_block_complete(struct scrub_block *sblock);
 245static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 246			       u64 extent_logical, u64 extent_len,
 247			       u64 *extent_physical,
 248			       struct btrfs_device **extent_dev,
 249			       int *extent_mirror_num);
 250static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 251				    struct scrub_page *spage);
 252static void scrub_wr_submit(struct scrub_ctx *sctx);
 253static void scrub_wr_bio_end_io(struct bio *bio);
 254static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 255static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 256static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 257static void scrub_put_ctx(struct scrub_ctx *sctx);
 258
 259static inline int scrub_is_page_on_raid56(struct scrub_page *page)
 260{
 261	return page->recover &&
 262	       (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 263}
 264
 265static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 266{
 267	refcount_inc(&sctx->refs);
 268	atomic_inc(&sctx->bios_in_flight);
 269}
 270
 271static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 272{
 273	atomic_dec(&sctx->bios_in_flight);
 274	wake_up(&sctx->list_wait);
 275	scrub_put_ctx(sctx);
 276}
 277
 278static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 279{
 280	while (atomic_read(&fs_info->scrub_pause_req)) {
 281		mutex_unlock(&fs_info->scrub_lock);
 282		wait_event(fs_info->scrub_pause_wait,
 283		   atomic_read(&fs_info->scrub_pause_req) == 0);
 284		mutex_lock(&fs_info->scrub_lock);
 285	}
 286}
 287
 288static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 289{
 290	atomic_inc(&fs_info->scrubs_paused);
 291	wake_up(&fs_info->scrub_pause_wait);
 292}
 293
 294static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 295{
 296	mutex_lock(&fs_info->scrub_lock);
 297	__scrub_blocked_if_needed(fs_info);
 298	atomic_dec(&fs_info->scrubs_paused);
 299	mutex_unlock(&fs_info->scrub_lock);
 300
 301	wake_up(&fs_info->scrub_pause_wait);
 302}
 303
 304static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 305{
 306	scrub_pause_on(fs_info);
 307	scrub_pause_off(fs_info);
 308}
 309
 310/*
 311 * Insert new full stripe lock into full stripe locks tree
 312 *
 313 * Return pointer to existing or newly inserted full_stripe_lock structure if
 314 * everything works well.
 315 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 316 *
 317 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 318 * function
 319 */
 320static struct full_stripe_lock *insert_full_stripe_lock(
 321		struct btrfs_full_stripe_locks_tree *locks_root,
 322		u64 fstripe_logical)
 323{
 324	struct rb_node **p;
 325	struct rb_node *parent = NULL;
 326	struct full_stripe_lock *entry;
 327	struct full_stripe_lock *ret;
 328
 329	lockdep_assert_held(&locks_root->lock);
 330
 331	p = &locks_root->root.rb_node;
 332	while (*p) {
 333		parent = *p;
 334		entry = rb_entry(parent, struct full_stripe_lock, node);
 335		if (fstripe_logical < entry->logical) {
 336			p = &(*p)->rb_left;
 337		} else if (fstripe_logical > entry->logical) {
 338			p = &(*p)->rb_right;
 339		} else {
 340			entry->refs++;
 341			return entry;
 342		}
 343	}
 344
 345	/*
 346	 * Insert new lock.
 347	 */
 348	ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 349	if (!ret)
 350		return ERR_PTR(-ENOMEM);
 351	ret->logical = fstripe_logical;
 352	ret->refs = 1;
 353	mutex_init(&ret->mutex);
 354
 355	rb_link_node(&ret->node, parent, p);
 356	rb_insert_color(&ret->node, &locks_root->root);
 357	return ret;
 358}
 359
 360/*
 361 * Search for a full stripe lock of a block group
 362 *
 363 * Return pointer to existing full stripe lock if found
 364 * Return NULL if not found
 365 */
 366static struct full_stripe_lock *search_full_stripe_lock(
 367		struct btrfs_full_stripe_locks_tree *locks_root,
 368		u64 fstripe_logical)
 369{
 370	struct rb_node *node;
 371	struct full_stripe_lock *entry;
 372
 373	lockdep_assert_held(&locks_root->lock);
 374
 375	node = locks_root->root.rb_node;
 376	while (node) {
 377		entry = rb_entry(node, struct full_stripe_lock, node);
 378		if (fstripe_logical < entry->logical)
 379			node = node->rb_left;
 380		else if (fstripe_logical > entry->logical)
 381			node = node->rb_right;
 382		else
 383			return entry;
 384	}
 385	return NULL;
 386}
 387
 388/*
 389 * Helper to get full stripe logical from a normal bytenr.
 390 *
 391 * Caller must ensure @cache is a RAID56 block group.
 392 */
 393static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
 394{
 395	u64 ret;
 396
 397	/*
 398	 * Due to chunk item size limit, full stripe length should not be
 399	 * larger than U32_MAX. Just a sanity check here.
 400	 */
 401	WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 402
 403	/*
 404	 * round_down() can only handle power of 2, while RAID56 full
 405	 * stripe length can be 64KiB * n, so we need to manually round down.
 406	 */
 407	ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 408			cache->full_stripe_len + cache->start;
 409	return ret;
 410}
 411
 412/*
 413 * Lock a full stripe to avoid concurrency of recovery and read
 414 *
 415 * It's only used for profiles with parities (RAID5/6), for other profiles it
 416 * does nothing.
 417 *
 418 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 419 * So caller must call unlock_full_stripe() at the same context.
 420 *
 421 * Return <0 if encounters error.
 422 */
 423static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 424			    bool *locked_ret)
 425{
 426	struct btrfs_block_group *bg_cache;
 427	struct btrfs_full_stripe_locks_tree *locks_root;
 428	struct full_stripe_lock *existing;
 429	u64 fstripe_start;
 430	int ret = 0;
 431
 432	*locked_ret = false;
 433	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 434	if (!bg_cache) {
 435		ASSERT(0);
 436		return -ENOENT;
 437	}
 438
 439	/* Profiles not based on parity don't need full stripe lock */
 440	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 441		goto out;
 442	locks_root = &bg_cache->full_stripe_locks_root;
 443
 444	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 445
 446	/* Now insert the full stripe lock */
 447	mutex_lock(&locks_root->lock);
 448	existing = insert_full_stripe_lock(locks_root, fstripe_start);
 449	mutex_unlock(&locks_root->lock);
 450	if (IS_ERR(existing)) {
 451		ret = PTR_ERR(existing);
 452		goto out;
 453	}
 454	mutex_lock(&existing->mutex);
 455	*locked_ret = true;
 456out:
 457	btrfs_put_block_group(bg_cache);
 458	return ret;
 459}
 460
 461/*
 462 * Unlock a full stripe.
 463 *
 464 * NOTE: Caller must ensure it's the same context calling corresponding
 465 * lock_full_stripe().
 466 *
 467 * Return 0 if we unlock full stripe without problem.
 468 * Return <0 for error
 469 */
 470static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 471			      bool locked)
 472{
 473	struct btrfs_block_group *bg_cache;
 474	struct btrfs_full_stripe_locks_tree *locks_root;
 475	struct full_stripe_lock *fstripe_lock;
 476	u64 fstripe_start;
 477	bool freeit = false;
 478	int ret = 0;
 479
 480	/* If we didn't acquire full stripe lock, no need to continue */
 481	if (!locked)
 482		return 0;
 483
 484	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 485	if (!bg_cache) {
 486		ASSERT(0);
 487		return -ENOENT;
 488	}
 489	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 490		goto out;
 491
 492	locks_root = &bg_cache->full_stripe_locks_root;
 493	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 494
 495	mutex_lock(&locks_root->lock);
 496	fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 497	/* Unpaired unlock_full_stripe() detected */
 498	if (!fstripe_lock) {
 499		WARN_ON(1);
 500		ret = -ENOENT;
 501		mutex_unlock(&locks_root->lock);
 502		goto out;
 503	}
 504
 505	if (fstripe_lock->refs == 0) {
 506		WARN_ON(1);
 507		btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 508			fstripe_lock->logical);
 509	} else {
 510		fstripe_lock->refs--;
 511	}
 512
 513	if (fstripe_lock->refs == 0) {
 514		rb_erase(&fstripe_lock->node, &locks_root->root);
 515		freeit = true;
 516	}
 517	mutex_unlock(&locks_root->lock);
 518
 519	mutex_unlock(&fstripe_lock->mutex);
 520	if (freeit)
 521		kfree(fstripe_lock);
 522out:
 523	btrfs_put_block_group(bg_cache);
 524	return ret;
 525}
 526
 527static void scrub_free_csums(struct scrub_ctx *sctx)
 528{
 529	while (!list_empty(&sctx->csum_list)) {
 530		struct btrfs_ordered_sum *sum;
 531		sum = list_first_entry(&sctx->csum_list,
 532				       struct btrfs_ordered_sum, list);
 533		list_del(&sum->list);
 534		kfree(sum);
 535	}
 536}
 537
 538static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 539{
 540	int i;
 541
 542	if (!sctx)
 543		return;
 544
 545	/* this can happen when scrub is cancelled */
 546	if (sctx->curr != -1) {
 547		struct scrub_bio *sbio = sctx->bios[sctx->curr];
 548
 549		for (i = 0; i < sbio->page_count; i++) {
 550			WARN_ON(!sbio->pagev[i]->page);
 551			scrub_block_put(sbio->pagev[i]->sblock);
 552		}
 553		bio_put(sbio->bio);
 554	}
 555
 556	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 557		struct scrub_bio *sbio = sctx->bios[i];
 558
 559		if (!sbio)
 560			break;
 561		kfree(sbio);
 562	}
 563
 564	kfree(sctx->wr_curr_bio);
 565	scrub_free_csums(sctx);
 566	kfree(sctx);
 567}
 568
 569static void scrub_put_ctx(struct scrub_ctx *sctx)
 570{
 571	if (refcount_dec_and_test(&sctx->refs))
 572		scrub_free_ctx(sctx);
 573}
 574
 575static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 576		struct btrfs_fs_info *fs_info, int is_dev_replace)
 577{
 578	struct scrub_ctx *sctx;
 579	int		i;
 580
 581	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 582	if (!sctx)
 583		goto nomem;
 584	refcount_set(&sctx->refs, 1);
 585	sctx->is_dev_replace = is_dev_replace;
 586	sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 587	sctx->curr = -1;
 588	sctx->fs_info = fs_info;
 589	INIT_LIST_HEAD(&sctx->csum_list);
 590	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 591		struct scrub_bio *sbio;
 592
 593		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 594		if (!sbio)
 595			goto nomem;
 596		sctx->bios[i] = sbio;
 597
 598		sbio->index = i;
 599		sbio->sctx = sctx;
 600		sbio->page_count = 0;
 601		btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
 602				NULL);
 603
 604		if (i != SCRUB_BIOS_PER_SCTX - 1)
 605			sctx->bios[i]->next_free = i + 1;
 606		else
 607			sctx->bios[i]->next_free = -1;
 608	}
 609	sctx->first_free = 0;
 610	atomic_set(&sctx->bios_in_flight, 0);
 611	atomic_set(&sctx->workers_pending, 0);
 612	atomic_set(&sctx->cancel_req, 0);
 613	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 614
 615	spin_lock_init(&sctx->list_lock);
 616	spin_lock_init(&sctx->stat_lock);
 617	init_waitqueue_head(&sctx->list_wait);
 
 618
 619	WARN_ON(sctx->wr_curr_bio != NULL);
 620	mutex_init(&sctx->wr_lock);
 621	sctx->wr_curr_bio = NULL;
 622	if (is_dev_replace) {
 623		WARN_ON(!fs_info->dev_replace.tgtdev);
 624		sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 625		sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 626		sctx->flush_all_writes = false;
 627	}
 628
 629	return sctx;
 630
 631nomem:
 632	scrub_free_ctx(sctx);
 633	return ERR_PTR(-ENOMEM);
 634}
 635
 636static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 637				     void *warn_ctx)
 638{
 639	u64 isize;
 640	u32 nlink;
 641	int ret;
 642	int i;
 643	unsigned nofs_flag;
 644	struct extent_buffer *eb;
 645	struct btrfs_inode_item *inode_item;
 646	struct scrub_warning *swarn = warn_ctx;
 647	struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 648	struct inode_fs_paths *ipath = NULL;
 649	struct btrfs_root *local_root;
 650	struct btrfs_key key;
 651
 652	local_root = btrfs_get_fs_root(fs_info, root, true);
 653	if (IS_ERR(local_root)) {
 654		ret = PTR_ERR(local_root);
 655		goto err;
 656	}
 657
 658	/*
 659	 * this makes the path point to (inum INODE_ITEM ioff)
 660	 */
 661	key.objectid = inum;
 662	key.type = BTRFS_INODE_ITEM_KEY;
 663	key.offset = 0;
 664
 665	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 666	if (ret) {
 667		btrfs_put_root(local_root);
 668		btrfs_release_path(swarn->path);
 669		goto err;
 670	}
 671
 672	eb = swarn->path->nodes[0];
 673	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 674					struct btrfs_inode_item);
 675	isize = btrfs_inode_size(eb, inode_item);
 676	nlink = btrfs_inode_nlink(eb, inode_item);
 677	btrfs_release_path(swarn->path);
 678
 679	/*
 680	 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 681	 * uses GFP_NOFS in this context, so we keep it consistent but it does
 682	 * not seem to be strictly necessary.
 683	 */
 684	nofs_flag = memalloc_nofs_save();
 685	ipath = init_ipath(4096, local_root, swarn->path);
 686	memalloc_nofs_restore(nofs_flag);
 687	if (IS_ERR(ipath)) {
 688		btrfs_put_root(local_root);
 689		ret = PTR_ERR(ipath);
 690		ipath = NULL;
 691		goto err;
 692	}
 693	ret = paths_from_inode(inum, ipath);
 694
 695	if (ret < 0)
 696		goto err;
 697
 698	/*
 699	 * we deliberately ignore the bit ipath might have been too small to
 700	 * hold all of the paths here
 701	 */
 702	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 703		btrfs_warn_in_rcu(fs_info,
 704"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 705				  swarn->errstr, swarn->logical,
 706				  rcu_str_deref(swarn->dev->name),
 707				  swarn->physical,
 708				  root, inum, offset,
 709				  min(isize - offset, (u64)PAGE_SIZE), nlink,
 710				  (char *)(unsigned long)ipath->fspath->val[i]);
 711
 712	btrfs_put_root(local_root);
 713	free_ipath(ipath);
 714	return 0;
 715
 716err:
 717	btrfs_warn_in_rcu(fs_info,
 718			  "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 719			  swarn->errstr, swarn->logical,
 720			  rcu_str_deref(swarn->dev->name),
 721			  swarn->physical,
 722			  root, inum, offset, ret);
 723
 724	free_ipath(ipath);
 725	return 0;
 726}
 727
 728static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 729{
 730	struct btrfs_device *dev;
 731	struct btrfs_fs_info *fs_info;
 732	struct btrfs_path *path;
 733	struct btrfs_key found_key;
 734	struct extent_buffer *eb;
 735	struct btrfs_extent_item *ei;
 736	struct scrub_warning swarn;
 737	unsigned long ptr = 0;
 738	u64 extent_item_pos;
 739	u64 flags = 0;
 740	u64 ref_root;
 741	u32 item_size;
 742	u8 ref_level = 0;
 743	int ret;
 744
 745	WARN_ON(sblock->page_count < 1);
 746	dev = sblock->pagev[0]->dev;
 747	fs_info = sblock->sctx->fs_info;
 748
 
 
 
 
 
 
 749	path = btrfs_alloc_path();
 750	if (!path)
 751		return;
 752
 753	swarn.physical = sblock->pagev[0]->physical;
 754	swarn.logical = sblock->pagev[0]->logical;
 755	swarn.errstr = errstr;
 756	swarn.dev = NULL;
 757
 758	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 759				  &flags);
 760	if (ret < 0)
 761		goto out;
 762
 763	extent_item_pos = swarn.logical - found_key.objectid;
 764	swarn.extent_item_size = found_key.offset;
 765
 766	eb = path->nodes[0];
 767	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 768	item_size = btrfs_item_size_nr(eb, path->slots[0]);
 769
 770	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 771		do {
 772			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 773						      item_size, &ref_root,
 774						      &ref_level);
 775			btrfs_warn_in_rcu(fs_info,
 776"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 777				errstr, swarn.logical,
 778				rcu_str_deref(dev->name),
 779				swarn.physical,
 780				ref_level ? "node" : "leaf",
 781				ret < 0 ? -1 : ref_level,
 782				ret < 0 ? -1 : ref_root);
 783		} while (ret != 1);
 784		btrfs_release_path(path);
 785	} else {
 
 
 786		btrfs_release_path(path);
 
 
 
 
 
 787		swarn.path = path;
 788		swarn.dev = dev;
 789		iterate_extent_inodes(fs_info, found_key.objectid,
 790					extent_item_pos, 1,
 791					scrub_print_warning_inode, &swarn, false);
 792	}
 793
 794out:
 795	btrfs_free_path(path);
 796}
 797
 798static inline void scrub_get_recover(struct scrub_recover *recover)
 799{
 800	refcount_inc(&recover->refs);
 801}
 802
 803static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 804				     struct scrub_recover *recover)
 805{
 806	if (refcount_dec_and_test(&recover->refs)) {
 807		btrfs_bio_counter_dec(fs_info);
 808		btrfs_put_bbio(recover->bbio);
 809		kfree(recover);
 810	}
 811}
 812
 813/*
 814 * scrub_handle_errored_block gets called when either verification of the
 815 * pages failed or the bio failed to read, e.g. with EIO. In the latter
 816 * case, this function handles all pages in the bio, even though only one
 817 * may be bad.
 818 * The goal of this function is to repair the errored block by using the
 819 * contents of one of the mirrors.
 820 */
 821static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 822{
 823	struct scrub_ctx *sctx = sblock_to_check->sctx;
 824	struct btrfs_device *dev;
 825	struct btrfs_fs_info *fs_info;
 826	u64 logical;
 827	unsigned int failed_mirror_index;
 828	unsigned int is_metadata;
 829	unsigned int have_csum;
 830	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 
 831	struct scrub_block *sblock_bad;
 832	int ret;
 833	int mirror_index;
 834	int page_num;
 835	int success;
 836	bool full_stripe_locked;
 837	unsigned int nofs_flag;
 838	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 839				      DEFAULT_RATELIMIT_BURST);
 840
 841	BUG_ON(sblock_to_check->page_count < 1);
 842	fs_info = sctx->fs_info;
 843	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 844		/*
 845		 * if we find an error in a super block, we just report it.
 846		 * They will get written with the next transaction commit
 847		 * anyway
 848		 */
 
 849		spin_lock(&sctx->stat_lock);
 850		++sctx->stat.super_errors;
 851		spin_unlock(&sctx->stat_lock);
 
 852		return 0;
 853	}
 854	logical = sblock_to_check->pagev[0]->logical;
 855	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 856	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 857	is_metadata = !(sblock_to_check->pagev[0]->flags &
 858			BTRFS_EXTENT_FLAG_DATA);
 859	have_csum = sblock_to_check->pagev[0]->have_csum;
 860	dev = sblock_to_check->pagev[0]->dev;
 
 
 861
 862	/*
 863	 * We must use GFP_NOFS because the scrub task might be waiting for a
 864	 * worker task executing this function and in turn a transaction commit
 865	 * might be waiting the scrub task to pause (which needs to wait for all
 866	 * the worker tasks to complete before pausing).
 867	 * We do allocations in the workers through insert_full_stripe_lock()
 868	 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 869	 * this function.
 870	 */
 871	nofs_flag = memalloc_nofs_save();
 872	/*
 873	 * For RAID5/6, race can happen for a different device scrub thread.
 874	 * For data corruption, Parity and Data threads will both try
 875	 * to recovery the data.
 876	 * Race can lead to doubly added csum error, or even unrecoverable
 877	 * error.
 878	 */
 879	ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 880	if (ret < 0) {
 881		memalloc_nofs_restore(nofs_flag);
 882		spin_lock(&sctx->stat_lock);
 883		if (ret == -ENOMEM)
 884			sctx->stat.malloc_errors++;
 885		sctx->stat.read_errors++;
 886		sctx->stat.uncorrectable_errors++;
 887		spin_unlock(&sctx->stat_lock);
 888		return ret;
 889	}
 890
 891	/*
 892	 * read all mirrors one after the other. This includes to
 893	 * re-read the extent or metadata block that failed (that was
 894	 * the cause that this fixup code is called) another time,
 895	 * page by page this time in order to know which pages
 896	 * caused I/O errors and which ones are good (for all mirrors).
 897	 * It is the goal to handle the situation when more than one
 898	 * mirror contains I/O errors, but the errors do not
 899	 * overlap, i.e. the data can be repaired by selecting the
 900	 * pages from those mirrors without I/O error on the
 901	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 902	 * would be that mirror #1 has an I/O error on the first page,
 903	 * the second page is good, and mirror #2 has an I/O error on
 904	 * the second page, but the first page is good.
 905	 * Then the first page of the first mirror can be repaired by
 906	 * taking the first page of the second mirror, and the
 907	 * second page of the second mirror can be repaired by
 908	 * copying the contents of the 2nd page of the 1st mirror.
 909	 * One more note: if the pages of one mirror contain I/O
 910	 * errors, the checksum cannot be verified. In order to get
 911	 * the best data for repairing, the first attempt is to find
 912	 * a mirror without I/O errors and with a validated checksum.
 913	 * Only if this is not possible, the pages are picked from
 914	 * mirrors with I/O errors without considering the checksum.
 915	 * If the latter is the case, at the end, the checksum of the
 916	 * repaired area is verified in order to correctly maintain
 917	 * the statistics.
 918	 */
 919
 920	sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 921				      sizeof(*sblocks_for_recheck), GFP_KERNEL);
 922	if (!sblocks_for_recheck) {
 923		spin_lock(&sctx->stat_lock);
 924		sctx->stat.malloc_errors++;
 925		sctx->stat.read_errors++;
 926		sctx->stat.uncorrectable_errors++;
 927		spin_unlock(&sctx->stat_lock);
 928		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 929		goto out;
 
 
 
 
 
 
 
 
 
 
 
 930	}
 931
 932	/* setup the context, map the logical blocks and alloc the pages */
 933	ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 934	if (ret) {
 935		spin_lock(&sctx->stat_lock);
 936		sctx->stat.read_errors++;
 937		sctx->stat.uncorrectable_errors++;
 938		spin_unlock(&sctx->stat_lock);
 939		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 940		goto out;
 941	}
 942	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 943	sblock_bad = sblocks_for_recheck + failed_mirror_index;
 944
 945	/* build and submit the bios for the failed mirror, check checksums */
 946	scrub_recheck_block(fs_info, sblock_bad, 1);
 947
 948	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 949	    sblock_bad->no_io_error_seen) {
 950		/*
 951		 * the error disappeared after reading page by page, or
 952		 * the area was part of a huge bio and other parts of the
 953		 * bio caused I/O errors, or the block layer merged several
 954		 * read requests into one and the error is caused by a
 955		 * different bio (usually one of the two latter cases is
 956		 * the cause)
 957		 */
 958		spin_lock(&sctx->stat_lock);
 959		sctx->stat.unverified_errors++;
 960		sblock_to_check->data_corrected = 1;
 961		spin_unlock(&sctx->stat_lock);
 962
 963		if (sctx->is_dev_replace)
 964			scrub_write_block_to_dev_replace(sblock_bad);
 965		goto out;
 966	}
 967
 968	if (!sblock_bad->no_io_error_seen) {
 969		spin_lock(&sctx->stat_lock);
 970		sctx->stat.read_errors++;
 971		spin_unlock(&sctx->stat_lock);
 972		if (__ratelimit(&_rs))
 973			scrub_print_warning("i/o error", sblock_to_check);
 974		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 975	} else if (sblock_bad->checksum_error) {
 976		spin_lock(&sctx->stat_lock);
 977		sctx->stat.csum_errors++;
 978		spin_unlock(&sctx->stat_lock);
 979		if (__ratelimit(&_rs))
 980			scrub_print_warning("checksum error", sblock_to_check);
 981		btrfs_dev_stat_inc_and_print(dev,
 982					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
 983	} else if (sblock_bad->header_error) {
 984		spin_lock(&sctx->stat_lock);
 985		sctx->stat.verify_errors++;
 986		spin_unlock(&sctx->stat_lock);
 987		if (__ratelimit(&_rs))
 988			scrub_print_warning("checksum/header error",
 989					    sblock_to_check);
 990		if (sblock_bad->generation_error)
 991			btrfs_dev_stat_inc_and_print(dev,
 992				BTRFS_DEV_STAT_GENERATION_ERRS);
 993		else
 994			btrfs_dev_stat_inc_and_print(dev,
 995				BTRFS_DEV_STAT_CORRUPTION_ERRS);
 996	}
 997
 998	if (sctx->readonly) {
 999		ASSERT(!sctx->is_dev_replace);
1000		goto out;
1001	}
1002
1003	/*
1004	 * now build and submit the bios for the other mirrors, check
1005	 * checksums.
1006	 * First try to pick the mirror which is completely without I/O
1007	 * errors and also does not have a checksum error.
1008	 * If one is found, and if a checksum is present, the full block
1009	 * that is known to contain an error is rewritten. Afterwards
1010	 * the block is known to be corrected.
1011	 * If a mirror is found which is completely correct, and no
1012	 * checksum is present, only those pages are rewritten that had
1013	 * an I/O error in the block to be repaired, since it cannot be
1014	 * determined, which copy of the other pages is better (and it
1015	 * could happen otherwise that a correct page would be
1016	 * overwritten by a bad one).
1017	 */
1018	for (mirror_index = 0; ;mirror_index++) {
1019		struct scrub_block *sblock_other;
1020
1021		if (mirror_index == failed_mirror_index)
1022			continue;
1023
1024		/* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1025		if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1026			if (mirror_index >= BTRFS_MAX_MIRRORS)
1027				break;
1028			if (!sblocks_for_recheck[mirror_index].page_count)
1029				break;
1030
1031			sblock_other = sblocks_for_recheck + mirror_index;
1032		} else {
1033			struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1034			int max_allowed = r->bbio->num_stripes -
1035						r->bbio->num_tgtdevs;
1036
1037			if (mirror_index >= max_allowed)
1038				break;
1039			if (!sblocks_for_recheck[1].page_count)
1040				break;
1041
1042			ASSERT(failed_mirror_index == 0);
1043			sblock_other = sblocks_for_recheck + 1;
1044			sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1045		}
1046
1047		/* build and submit the bios, check checksums */
1048		scrub_recheck_block(fs_info, sblock_other, 0);
1049
1050		if (!sblock_other->header_error &&
1051		    !sblock_other->checksum_error &&
1052		    sblock_other->no_io_error_seen) {
1053			if (sctx->is_dev_replace) {
1054				scrub_write_block_to_dev_replace(sblock_other);
1055				goto corrected_error;
1056			} else {
1057				ret = scrub_repair_block_from_good_copy(
1058						sblock_bad, sblock_other);
1059				if (!ret)
1060					goto corrected_error;
1061			}
1062		}
1063	}
1064
1065	if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1066		goto did_not_correct_error;
1067
1068	/*
1069	 * In case of I/O errors in the area that is supposed to be
1070	 * repaired, continue by picking good copies of those pages.
1071	 * Select the good pages from mirrors to rewrite bad pages from
1072	 * the area to fix. Afterwards verify the checksum of the block
1073	 * that is supposed to be repaired. This verification step is
1074	 * only done for the purpose of statistic counting and for the
1075	 * final scrub report, whether errors remain.
1076	 * A perfect algorithm could make use of the checksum and try
1077	 * all possible combinations of pages from the different mirrors
1078	 * until the checksum verification succeeds. For example, when
1079	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1080	 * of mirror #2 is readable but the final checksum test fails,
1081	 * then the 2nd page of mirror #3 could be tried, whether now
1082	 * the final checksum succeeds. But this would be a rare
1083	 * exception and is therefore not implemented. At least it is
1084	 * avoided that the good copy is overwritten.
1085	 * A more useful improvement would be to pick the sectors
1086	 * without I/O error based on sector sizes (512 bytes on legacy
1087	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1088	 * mirror could be repaired by taking 512 byte of a different
1089	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1090	 * area are unreadable.
1091	 */
1092	success = 1;
1093	for (page_num = 0; page_num < sblock_bad->page_count;
1094	     page_num++) {
1095		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1096		struct scrub_block *sblock_other = NULL;
1097
1098		/* skip no-io-error page in scrub */
1099		if (!page_bad->io_error && !sctx->is_dev_replace)
1100			continue;
1101
1102		if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1103			/*
1104			 * In case of dev replace, if raid56 rebuild process
1105			 * didn't work out correct data, then copy the content
1106			 * in sblock_bad to make sure target device is identical
1107			 * to source device, instead of writing garbage data in
1108			 * sblock_for_recheck array to target device.
1109			 */
1110			sblock_other = NULL;
1111		} else if (page_bad->io_error) {
1112			/* try to find no-io-error page in mirrors */
1113			for (mirror_index = 0;
1114			     mirror_index < BTRFS_MAX_MIRRORS &&
1115			     sblocks_for_recheck[mirror_index].page_count > 0;
1116			     mirror_index++) {
1117				if (!sblocks_for_recheck[mirror_index].
1118				    pagev[page_num]->io_error) {
1119					sblock_other = sblocks_for_recheck +
1120						       mirror_index;
1121					break;
1122				}
1123			}
1124			if (!sblock_other)
1125				success = 0;
1126		}
1127
1128		if (sctx->is_dev_replace) {
1129			/*
1130			 * did not find a mirror to fetch the page
1131			 * from. scrub_write_page_to_dev_replace()
1132			 * handles this case (page->io_error), by
1133			 * filling the block with zeros before
1134			 * submitting the write request
1135			 */
1136			if (!sblock_other)
1137				sblock_other = sblock_bad;
1138
1139			if (scrub_write_page_to_dev_replace(sblock_other,
1140							    page_num) != 0) {
1141				atomic64_inc(
1142					&fs_info->dev_replace.num_write_errors);
1143				success = 0;
1144			}
1145		} else if (sblock_other) {
1146			ret = scrub_repair_page_from_good_copy(sblock_bad,
1147							       sblock_other,
1148							       page_num, 0);
1149			if (0 == ret)
1150				page_bad->io_error = 0;
1151			else
1152				success = 0;
1153		}
1154	}
1155
1156	if (success && !sctx->is_dev_replace) {
1157		if (is_metadata || have_csum) {
1158			/*
1159			 * need to verify the checksum now that all
1160			 * sectors on disk are repaired (the write
1161			 * request for data to be repaired is on its way).
1162			 * Just be lazy and use scrub_recheck_block()
1163			 * which re-reads the data before the checksum
1164			 * is verified, but most likely the data comes out
1165			 * of the page cache.
1166			 */
1167			scrub_recheck_block(fs_info, sblock_bad, 1);
1168			if (!sblock_bad->header_error &&
1169			    !sblock_bad->checksum_error &&
1170			    sblock_bad->no_io_error_seen)
1171				goto corrected_error;
1172			else
1173				goto did_not_correct_error;
1174		} else {
1175corrected_error:
1176			spin_lock(&sctx->stat_lock);
1177			sctx->stat.corrected_errors++;
1178			sblock_to_check->data_corrected = 1;
1179			spin_unlock(&sctx->stat_lock);
1180			btrfs_err_rl_in_rcu(fs_info,
1181				"fixed up error at logical %llu on dev %s",
1182				logical, rcu_str_deref(dev->name));
1183		}
1184	} else {
1185did_not_correct_error:
1186		spin_lock(&sctx->stat_lock);
1187		sctx->stat.uncorrectable_errors++;
1188		spin_unlock(&sctx->stat_lock);
1189		btrfs_err_rl_in_rcu(fs_info,
1190			"unable to fixup (regular) error at logical %llu on dev %s",
1191			logical, rcu_str_deref(dev->name));
1192	}
1193
1194out:
1195	if (sblocks_for_recheck) {
1196		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1197		     mirror_index++) {
1198			struct scrub_block *sblock = sblocks_for_recheck +
1199						     mirror_index;
1200			struct scrub_recover *recover;
1201			int page_index;
1202
1203			for (page_index = 0; page_index < sblock->page_count;
1204			     page_index++) {
1205				sblock->pagev[page_index]->sblock = NULL;
1206				recover = sblock->pagev[page_index]->recover;
1207				if (recover) {
1208					scrub_put_recover(fs_info, recover);
1209					sblock->pagev[page_index]->recover =
1210									NULL;
1211				}
1212				scrub_page_put(sblock->pagev[page_index]);
 
1213			}
1214		}
1215		kfree(sblocks_for_recheck);
1216	}
1217
1218	ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1219	memalloc_nofs_restore(nofs_flag);
1220	if (ret < 0)
1221		return ret;
1222	return 0;
1223}
1224
1225static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1226{
1227	if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1228		return 2;
1229	else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1230		return 3;
1231	else
1232		return (int)bbio->num_stripes;
1233}
1234
1235static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1236						 u64 *raid_map,
1237						 u64 mapped_length,
1238						 int nstripes, int mirror,
1239						 int *stripe_index,
1240						 u64 *stripe_offset)
1241{
1242	int i;
1243
1244	if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1245		/* RAID5/6 */
1246		for (i = 0; i < nstripes; i++) {
1247			if (raid_map[i] == RAID6_Q_STRIPE ||
1248			    raid_map[i] == RAID5_P_STRIPE)
1249				continue;
1250
1251			if (logical >= raid_map[i] &&
1252			    logical < raid_map[i] + mapped_length)
1253				break;
1254		}
1255
1256		*stripe_index = i;
1257		*stripe_offset = logical - raid_map[i];
1258	} else {
1259		/* The other RAID type */
1260		*stripe_index = mirror;
1261		*stripe_offset = 0;
1262	}
1263}
1264
1265static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1266				     struct scrub_block *sblocks_for_recheck)
1267{
1268	struct scrub_ctx *sctx = original_sblock->sctx;
1269	struct btrfs_fs_info *fs_info = sctx->fs_info;
1270	u64 length = original_sblock->page_count * PAGE_SIZE;
1271	u64 logical = original_sblock->pagev[0]->logical;
1272	u64 generation = original_sblock->pagev[0]->generation;
1273	u64 flags = original_sblock->pagev[0]->flags;
1274	u64 have_csum = original_sblock->pagev[0]->have_csum;
1275	struct scrub_recover *recover;
1276	struct btrfs_bio *bbio;
1277	u64 sublen;
1278	u64 mapped_length;
1279	u64 stripe_offset;
1280	int stripe_index;
1281	int page_index = 0;
1282	int mirror_index;
1283	int nmirrors;
1284	int ret;
1285
1286	/*
1287	 * note: the two members refs and outstanding_pages
1288	 * are not used (and not set) in the blocks that are used for
1289	 * the recheck procedure
1290	 */
1291
1292	while (length > 0) {
1293		sublen = min_t(u64, length, PAGE_SIZE);
1294		mapped_length = sublen;
1295		bbio = NULL;
1296
1297		/*
1298		 * with a length of PAGE_SIZE, each returned stripe
1299		 * represents one mirror
1300		 */
1301		btrfs_bio_counter_inc_blocked(fs_info);
1302		ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1303				logical, &mapped_length, &bbio);
1304		if (ret || !bbio || mapped_length < sublen) {
1305			btrfs_put_bbio(bbio);
1306			btrfs_bio_counter_dec(fs_info);
1307			return -EIO;
1308		}
1309
1310		recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1311		if (!recover) {
1312			btrfs_put_bbio(bbio);
1313			btrfs_bio_counter_dec(fs_info);
1314			return -ENOMEM;
1315		}
1316
1317		refcount_set(&recover->refs, 1);
1318		recover->bbio = bbio;
1319		recover->map_length = mapped_length;
1320
1321		BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1322
1323		nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1324
1325		for (mirror_index = 0; mirror_index < nmirrors;
1326		     mirror_index++) {
1327			struct scrub_block *sblock;
1328			struct scrub_page *page;
1329
1330			sblock = sblocks_for_recheck + mirror_index;
1331			sblock->sctx = sctx;
1332
1333			page = kzalloc(sizeof(*page), GFP_NOFS);
1334			if (!page) {
1335leave_nomem:
1336				spin_lock(&sctx->stat_lock);
1337				sctx->stat.malloc_errors++;
1338				spin_unlock(&sctx->stat_lock);
1339				scrub_put_recover(fs_info, recover);
1340				return -ENOMEM;
1341			}
1342			scrub_page_get(page);
1343			sblock->pagev[page_index] = page;
1344			page->sblock = sblock;
1345			page->flags = flags;
1346			page->generation = generation;
1347			page->logical = logical;
1348			page->have_csum = have_csum;
1349			if (have_csum)
1350				memcpy(page->csum,
1351				       original_sblock->pagev[0]->csum,
1352				       sctx->csum_size);
1353
1354			scrub_stripe_index_and_offset(logical,
1355						      bbio->map_type,
1356						      bbio->raid_map,
1357						      mapped_length,
1358						      bbio->num_stripes -
1359						      bbio->num_tgtdevs,
1360						      mirror_index,
1361						      &stripe_index,
1362						      &stripe_offset);
1363			page->physical = bbio->stripes[stripe_index].physical +
1364					 stripe_offset;
1365			page->dev = bbio->stripes[stripe_index].dev;
1366
1367			BUG_ON(page_index >= original_sblock->page_count);
1368			page->physical_for_dev_replace =
1369				original_sblock->pagev[page_index]->
1370				physical_for_dev_replace;
1371			/* for missing devices, dev->bdev is NULL */
1372			page->mirror_num = mirror_index + 1;
1373			sblock->page_count++;
1374			page->page = alloc_page(GFP_NOFS);
1375			if (!page->page)
1376				goto leave_nomem;
1377
 
1378			scrub_get_recover(recover);
1379			page->recover = recover;
1380		}
1381		scrub_put_recover(fs_info, recover);
1382		length -= sublen;
1383		logical += sublen;
1384		page_index++;
1385	}
1386
1387	return 0;
1388}
1389
1390static void scrub_bio_wait_endio(struct bio *bio)
1391{
1392	complete(bio->bi_private);
1393}
1394
1395static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1396					struct bio *bio,
1397					struct scrub_page *page)
1398{
1399	DECLARE_COMPLETION_ONSTACK(done);
1400	int ret;
1401	int mirror_num;
1402
1403	bio->bi_iter.bi_sector = page->logical >> 9;
 
1404	bio->bi_private = &done;
1405	bio->bi_end_io = scrub_bio_wait_endio;
1406
1407	mirror_num = page->sblock->pagev[0]->mirror_num;
1408	ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1409				    page->recover->map_length,
1410				    mirror_num, 0);
1411	if (ret)
1412		return ret;
1413
1414	wait_for_completion_io(&done);
1415	return blk_status_to_errno(bio->bi_status);
1416}
1417
1418static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1419					  struct scrub_block *sblock)
1420{
1421	struct scrub_page *first_page = sblock->pagev[0];
1422	struct bio *bio;
1423	int page_num;
1424
1425	/* All pages in sblock belong to the same stripe on the same device. */
1426	ASSERT(first_page->dev);
1427	if (!first_page->dev->bdev)
1428		goto out;
1429
1430	bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1431	bio_set_dev(bio, first_page->dev->bdev);
1432
1433	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1434		struct scrub_page *page = sblock->pagev[page_num];
1435
1436		WARN_ON(!page->page);
1437		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1438	}
1439
1440	if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1441		bio_put(bio);
1442		goto out;
1443	}
1444
1445	bio_put(bio);
1446
1447	scrub_recheck_block_checksum(sblock);
1448
1449	return;
1450out:
1451	for (page_num = 0; page_num < sblock->page_count; page_num++)
1452		sblock->pagev[page_num]->io_error = 1;
1453
1454	sblock->no_io_error_seen = 0;
1455}
1456
1457/*
1458 * this function will check the on disk data for checksum errors, header
1459 * errors and read I/O errors. If any I/O errors happen, the exact pages
1460 * which are errored are marked as being bad. The goal is to enable scrub
1461 * to take those pages that are not errored from all the mirrors so that
1462 * the pages that are errored in the just handled mirror can be repaired.
1463 */
1464static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1465				struct scrub_block *sblock,
1466				int retry_failed_mirror)
1467{
1468	int page_num;
1469
1470	sblock->no_io_error_seen = 1;
1471
1472	/* short cut for raid56 */
1473	if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1474		return scrub_recheck_block_on_raid56(fs_info, sblock);
1475
1476	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1477		struct bio *bio;
1478		struct scrub_page *page = sblock->pagev[page_num];
 
1479
1480		if (page->dev->bdev == NULL) {
1481			page->io_error = 1;
1482			sblock->no_io_error_seen = 0;
1483			continue;
1484		}
1485
1486		WARN_ON(!page->page);
1487		bio = btrfs_io_bio_alloc(1);
1488		bio_set_dev(bio, page->dev->bdev);
1489
1490		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1491		bio->bi_iter.bi_sector = page->physical >> 9;
1492		bio->bi_opf = REQ_OP_READ;
1493
1494		if (btrfsic_submit_bio_wait(bio)) {
1495			page->io_error = 1;
1496			sblock->no_io_error_seen = 0;
1497		}
1498
1499		bio_put(bio);
1500	}
1501
1502	if (sblock->no_io_error_seen)
1503		scrub_recheck_block_checksum(sblock);
1504}
1505
1506static inline int scrub_check_fsid(u8 fsid[],
1507				   struct scrub_page *spage)
1508{
1509	struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1510	int ret;
1511
1512	ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1513	return !ret;
1514}
1515
1516static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1517{
1518	sblock->header_error = 0;
1519	sblock->checksum_error = 0;
1520	sblock->generation_error = 0;
1521
1522	if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1523		scrub_checksum_data(sblock);
1524	else
1525		scrub_checksum_tree_block(sblock);
1526}
1527
1528static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1529					     struct scrub_block *sblock_good)
1530{
1531	int page_num;
1532	int ret = 0;
1533
1534	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1535		int ret_sub;
1536
1537		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1538							   sblock_good,
1539							   page_num, 1);
1540		if (ret_sub)
1541			ret = ret_sub;
1542	}
1543
1544	return ret;
1545}
1546
1547static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1548					    struct scrub_block *sblock_good,
1549					    int page_num, int force_write)
1550{
1551	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1552	struct scrub_page *page_good = sblock_good->pagev[page_num];
1553	struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
 
1554
1555	BUG_ON(page_bad->page == NULL);
1556	BUG_ON(page_good->page == NULL);
1557	if (force_write || sblock_bad->header_error ||
1558	    sblock_bad->checksum_error || page_bad->io_error) {
1559		struct bio *bio;
 
1560		int ret;
1561
1562		if (!page_bad->dev->bdev) {
1563			btrfs_warn_rl(fs_info,
1564				"scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1565			return -EIO;
1566		}
1567
1568		bio = btrfs_io_bio_alloc(1);
1569		bio_set_dev(bio, page_bad->dev->bdev);
1570		bio->bi_iter.bi_sector = page_bad->physical >> 9;
1571		bio->bi_opf = REQ_OP_WRITE;
1572
1573		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1574		if (PAGE_SIZE != ret) {
1575			bio_put(bio);
1576			return -EIO;
1577		}
1578
1579		if (btrfsic_submit_bio_wait(bio)) {
1580			btrfs_dev_stat_inc_and_print(page_bad->dev,
1581				BTRFS_DEV_STAT_WRITE_ERRS);
1582			atomic64_inc(&fs_info->dev_replace.num_write_errors);
1583			bio_put(bio);
1584			return -EIO;
1585		}
1586		bio_put(bio);
1587	}
1588
1589	return 0;
1590}
1591
1592static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1593{
1594	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1595	int page_num;
1596
1597	/*
1598	 * This block is used for the check of the parity on the source device,
1599	 * so the data needn't be written into the destination device.
1600	 */
1601	if (sblock->sparity)
1602		return;
1603
1604	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1605		int ret;
1606
1607		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1608		if (ret)
1609			atomic64_inc(&fs_info->dev_replace.num_write_errors);
1610	}
1611}
1612
1613static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1614					   int page_num)
1615{
1616	struct scrub_page *spage = sblock->pagev[page_num];
 
1617
1618	BUG_ON(spage->page == NULL);
1619	if (spage->io_error)
1620		clear_page(page_address(spage->page));
1621
1622	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1623}
1624
1625static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1626				    struct scrub_page *spage)
1627{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1628	struct scrub_bio *sbio;
1629	int ret;
 
1630
1631	mutex_lock(&sctx->wr_lock);
1632again:
1633	if (!sctx->wr_curr_bio) {
1634		sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1635					      GFP_KERNEL);
1636		if (!sctx->wr_curr_bio) {
1637			mutex_unlock(&sctx->wr_lock);
1638			return -ENOMEM;
1639		}
1640		sctx->wr_curr_bio->sctx = sctx;
1641		sctx->wr_curr_bio->page_count = 0;
1642	}
1643	sbio = sctx->wr_curr_bio;
1644	if (sbio->page_count == 0) {
1645		struct bio *bio;
 
 
 
 
 
1646
1647		sbio->physical = spage->physical_for_dev_replace;
1648		sbio->logical = spage->logical;
1649		sbio->dev = sctx->wr_tgtdev;
1650		bio = sbio->bio;
1651		if (!bio) {
1652			bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1653			sbio->bio = bio;
1654		}
1655
1656		bio->bi_private = sbio;
1657		bio->bi_end_io = scrub_wr_bio_end_io;
1658		bio_set_dev(bio, sbio->dev->bdev);
1659		bio->bi_iter.bi_sector = sbio->physical >> 9;
1660		bio->bi_opf = REQ_OP_WRITE;
1661		sbio->status = 0;
1662	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1663		   spage->physical_for_dev_replace ||
1664		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1665		   spage->logical) {
1666		scrub_wr_submit(sctx);
1667		goto again;
1668	}
1669
1670	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1671	if (ret != PAGE_SIZE) {
1672		if (sbio->page_count < 1) {
1673			bio_put(sbio->bio);
1674			sbio->bio = NULL;
1675			mutex_unlock(&sctx->wr_lock);
1676			return -EIO;
1677		}
1678		scrub_wr_submit(sctx);
1679		goto again;
1680	}
1681
1682	sbio->pagev[sbio->page_count] = spage;
1683	scrub_page_get(spage);
1684	sbio->page_count++;
1685	if (sbio->page_count == sctx->pages_per_wr_bio)
 
 
 
 
 
 
 
1686		scrub_wr_submit(sctx);
1687	mutex_unlock(&sctx->wr_lock);
1688
1689	return 0;
1690}
1691
1692static void scrub_wr_submit(struct scrub_ctx *sctx)
1693{
1694	struct scrub_bio *sbio;
1695
1696	if (!sctx->wr_curr_bio)
1697		return;
1698
1699	sbio = sctx->wr_curr_bio;
1700	sctx->wr_curr_bio = NULL;
1701	WARN_ON(!sbio->bio->bi_disk);
1702	scrub_pending_bio_inc(sctx);
1703	/* process all writes in a single worker thread. Then the block layer
1704	 * orders the requests before sending them to the driver which
1705	 * doubled the write performance on spinning disks when measured
1706	 * with Linux 3.5 */
1707	btrfsic_submit_bio(sbio->bio);
 
 
 
 
 
1708}
1709
1710static void scrub_wr_bio_end_io(struct bio *bio)
1711{
1712	struct scrub_bio *sbio = bio->bi_private;
1713	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1714
1715	sbio->status = bio->bi_status;
1716	sbio->bio = bio;
1717
1718	btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1719	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1720}
1721
1722static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1723{
1724	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1725	struct scrub_ctx *sctx = sbio->sctx;
1726	int i;
1727
1728	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1729	if (sbio->status) {
1730		struct btrfs_dev_replace *dev_replace =
1731			&sbio->sctx->fs_info->dev_replace;
1732
1733		for (i = 0; i < sbio->page_count; i++) {
1734			struct scrub_page *spage = sbio->pagev[i];
1735
1736			spage->io_error = 1;
1737			atomic64_inc(&dev_replace->num_write_errors);
1738		}
1739	}
1740
1741	for (i = 0; i < sbio->page_count; i++)
1742		scrub_page_put(sbio->pagev[i]);
 
 
 
 
 
 
1743
1744	bio_put(sbio->bio);
1745	kfree(sbio);
1746	scrub_pending_bio_dec(sctx);
1747}
1748
1749static int scrub_checksum(struct scrub_block *sblock)
1750{
1751	u64 flags;
1752	int ret;
1753
1754	/*
1755	 * No need to initialize these stats currently,
1756	 * because this function only use return value
1757	 * instead of these stats value.
1758	 *
1759	 * Todo:
1760	 * always use stats
1761	 */
1762	sblock->header_error = 0;
1763	sblock->generation_error = 0;
1764	sblock->checksum_error = 0;
1765
1766	WARN_ON(sblock->page_count < 1);
1767	flags = sblock->pagev[0]->flags;
1768	ret = 0;
1769	if (flags & BTRFS_EXTENT_FLAG_DATA)
1770		ret = scrub_checksum_data(sblock);
1771	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1772		ret = scrub_checksum_tree_block(sblock);
1773	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1774		(void)scrub_checksum_super(sblock);
1775	else
1776		WARN_ON(1);
1777	if (ret)
1778		scrub_handle_errored_block(sblock);
1779
1780	return ret;
1781}
1782
1783static int scrub_checksum_data(struct scrub_block *sblock)
1784{
1785	struct scrub_ctx *sctx = sblock->sctx;
1786	struct btrfs_fs_info *fs_info = sctx->fs_info;
1787	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1788	u8 csum[BTRFS_CSUM_SIZE];
1789	struct scrub_page *spage;
1790	char *kaddr;
1791
1792	BUG_ON(sblock->page_count < 1);
1793	spage = sblock->pagev[0];
1794	if (!spage->have_csum)
1795		return 0;
1796
1797	kaddr = page_address(spage->page);
1798
1799	shash->tfm = fs_info->csum_shash;
1800	crypto_shash_init(shash);
1801	crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum);
1802
1803	if (memcmp(csum, spage->csum, sctx->csum_size))
 
 
1804		sblock->checksum_error = 1;
1805
1806	return sblock->checksum_error;
1807}
1808
1809static int scrub_checksum_tree_block(struct scrub_block *sblock)
1810{
1811	struct scrub_ctx *sctx = sblock->sctx;
1812	struct btrfs_header *h;
1813	struct btrfs_fs_info *fs_info = sctx->fs_info;
1814	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1815	u8 calculated_csum[BTRFS_CSUM_SIZE];
1816	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1817	const int num_pages = sctx->fs_info->nodesize >> PAGE_SHIFT;
 
 
 
 
 
 
1818	int i;
1819	struct scrub_page *spage;
1820	char *kaddr;
1821
1822	BUG_ON(sblock->page_count < 1);
1823	spage = sblock->pagev[0];
1824	kaddr = page_address(spage->page);
 
 
 
 
1825	h = (struct btrfs_header *)kaddr;
1826	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1827
1828	/*
1829	 * we don't use the getter functions here, as we
1830	 * a) don't have an extent buffer and
1831	 * b) the page is already kmapped
1832	 */
1833	if (spage->logical != btrfs_stack_header_bytenr(h))
1834		sblock->header_error = 1;
1835
1836	if (spage->generation != btrfs_stack_header_generation(h)) {
1837		sblock->header_error = 1;
1838		sblock->generation_error = 1;
1839	}
1840
1841	if (!scrub_check_fsid(h->fsid, spage))
1842		sblock->header_error = 1;
1843
1844	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1845		   BTRFS_UUID_SIZE))
1846		sblock->header_error = 1;
1847
1848	shash->tfm = fs_info->csum_shash;
1849	crypto_shash_init(shash);
1850	crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
1851			    PAGE_SIZE - BTRFS_CSUM_SIZE);
1852
1853	for (i = 1; i < num_pages; i++) {
1854		kaddr = page_address(sblock->pagev[i]->page);
1855		crypto_shash_update(shash, kaddr, PAGE_SIZE);
1856	}
1857
1858	crypto_shash_final(shash, calculated_csum);
1859	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1860		sblock->checksum_error = 1;
1861
1862	return sblock->header_error || sblock->checksum_error;
1863}
1864
1865static int scrub_checksum_super(struct scrub_block *sblock)
1866{
1867	struct btrfs_super_block *s;
1868	struct scrub_ctx *sctx = sblock->sctx;
1869	struct btrfs_fs_info *fs_info = sctx->fs_info;
1870	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1871	u8 calculated_csum[BTRFS_CSUM_SIZE];
1872	struct scrub_page *spage;
1873	char *kaddr;
1874	int fail_gen = 0;
1875	int fail_cor = 0;
1876
1877	BUG_ON(sblock->page_count < 1);
1878	spage = sblock->pagev[0];
1879	kaddr = page_address(spage->page);
1880	s = (struct btrfs_super_block *)kaddr;
1881
1882	if (spage->logical != btrfs_super_bytenr(s))
1883		++fail_cor;
1884
1885	if (spage->generation != btrfs_super_generation(s))
1886		++fail_gen;
1887
1888	if (!scrub_check_fsid(s->fsid, spage))
1889		++fail_cor;
1890
1891	shash->tfm = fs_info->csum_shash;
1892	crypto_shash_init(shash);
1893	crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
1894			BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
1895
1896	if (memcmp(calculated_csum, s->csum, sctx->csum_size))
1897		++fail_cor;
1898
1899	if (fail_cor + fail_gen) {
1900		/*
1901		 * if we find an error in a super block, we just report it.
1902		 * They will get written with the next transaction commit
1903		 * anyway
1904		 */
1905		spin_lock(&sctx->stat_lock);
1906		++sctx->stat.super_errors;
1907		spin_unlock(&sctx->stat_lock);
1908		if (fail_cor)
1909			btrfs_dev_stat_inc_and_print(spage->dev,
1910				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1911		else
1912			btrfs_dev_stat_inc_and_print(spage->dev,
1913				BTRFS_DEV_STAT_GENERATION_ERRS);
1914	}
1915
1916	return fail_cor + fail_gen;
1917}
1918
1919static void scrub_block_get(struct scrub_block *sblock)
1920{
1921	refcount_inc(&sblock->refs);
1922}
1923
1924static void scrub_block_put(struct scrub_block *sblock)
1925{
1926	if (refcount_dec_and_test(&sblock->refs)) {
1927		int i;
1928
1929		if (sblock->sparity)
1930			scrub_parity_put(sblock->sparity);
1931
1932		for (i = 0; i < sblock->page_count; i++)
1933			scrub_page_put(sblock->pagev[i]);
 
 
 
 
 
 
1934		kfree(sblock);
1935	}
1936}
1937
1938static void scrub_page_get(struct scrub_page *spage)
1939{
1940	atomic_inc(&spage->refs);
1941}
1942
1943static void scrub_page_put(struct scrub_page *spage)
1944{
1945	if (atomic_dec_and_test(&spage->refs)) {
1946		if (spage->page)
1947			__free_page(spage->page);
1948		kfree(spage);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1949	}
 
 
 
 
 
 
 
 
 
 
1950}
1951
1952static void scrub_submit(struct scrub_ctx *sctx)
1953{
1954	struct scrub_bio *sbio;
1955
1956	if (sctx->curr == -1)
1957		return;
1958
 
 
1959	sbio = sctx->bios[sctx->curr];
1960	sctx->curr = -1;
1961	scrub_pending_bio_inc(sctx);
1962	btrfsic_submit_bio(sbio->bio);
 
1963}
1964
1965static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1966				    struct scrub_page *spage)
1967{
1968	struct scrub_block *sblock = spage->sblock;
1969	struct scrub_bio *sbio;
 
1970	int ret;
1971
1972again:
1973	/*
1974	 * grab a fresh bio or wait for one to become available
1975	 */
1976	while (sctx->curr == -1) {
1977		spin_lock(&sctx->list_lock);
1978		sctx->curr = sctx->first_free;
1979		if (sctx->curr != -1) {
1980			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1981			sctx->bios[sctx->curr]->next_free = -1;
1982			sctx->bios[sctx->curr]->page_count = 0;
1983			spin_unlock(&sctx->list_lock);
1984		} else {
1985			spin_unlock(&sctx->list_lock);
1986			wait_event(sctx->list_wait, sctx->first_free != -1);
1987		}
1988	}
1989	sbio = sctx->bios[sctx->curr];
1990	if (sbio->page_count == 0) {
1991		struct bio *bio;
1992
1993		sbio->physical = spage->physical;
1994		sbio->logical = spage->logical;
1995		sbio->dev = spage->dev;
1996		bio = sbio->bio;
1997		if (!bio) {
1998			bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
1999			sbio->bio = bio;
2000		}
2001
2002		bio->bi_private = sbio;
2003		bio->bi_end_io = scrub_bio_end_io;
2004		bio_set_dev(bio, sbio->dev->bdev);
2005		bio->bi_iter.bi_sector = sbio->physical >> 9;
2006		bio->bi_opf = REQ_OP_READ;
2007		sbio->status = 0;
2008	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2009		   spage->physical ||
2010		   sbio->logical + sbio->page_count * PAGE_SIZE !=
2011		   spage->logical ||
2012		   sbio->dev != spage->dev) {
2013		scrub_submit(sctx);
2014		goto again;
2015	}
2016
2017	sbio->pagev[sbio->page_count] = spage;
2018	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2019	if (ret != PAGE_SIZE) {
2020		if (sbio->page_count < 1) {
2021			bio_put(sbio->bio);
2022			sbio->bio = NULL;
2023			return -EIO;
2024		}
2025		scrub_submit(sctx);
2026		goto again;
2027	}
2028
2029	scrub_block_get(sblock); /* one for the page added to the bio */
2030	atomic_inc(&sblock->outstanding_pages);
2031	sbio->page_count++;
2032	if (sbio->page_count == sctx->pages_per_rd_bio)
2033		scrub_submit(sctx);
2034
2035	return 0;
2036}
2037
2038static void scrub_missing_raid56_end_io(struct bio *bio)
2039{
2040	struct scrub_block *sblock = bio->bi_private;
2041	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2042
 
2043	if (bio->bi_status)
2044		sblock->no_io_error_seen = 0;
2045
2046	bio_put(bio);
2047
2048	btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2049}
2050
2051static void scrub_missing_raid56_worker(struct btrfs_work *work)
2052{
2053	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2054	struct scrub_ctx *sctx = sblock->sctx;
2055	struct btrfs_fs_info *fs_info = sctx->fs_info;
2056	u64 logical;
2057	struct btrfs_device *dev;
2058
2059	logical = sblock->pagev[0]->logical;
2060	dev = sblock->pagev[0]->dev;
2061
2062	if (sblock->no_io_error_seen)
2063		scrub_recheck_block_checksum(sblock);
2064
2065	if (!sblock->no_io_error_seen) {
2066		spin_lock(&sctx->stat_lock);
2067		sctx->stat.read_errors++;
2068		spin_unlock(&sctx->stat_lock);
2069		btrfs_err_rl_in_rcu(fs_info,
2070			"IO error rebuilding logical %llu for dev %s",
2071			logical, rcu_str_deref(dev->name));
2072	} else if (sblock->header_error || sblock->checksum_error) {
2073		spin_lock(&sctx->stat_lock);
2074		sctx->stat.uncorrectable_errors++;
2075		spin_unlock(&sctx->stat_lock);
2076		btrfs_err_rl_in_rcu(fs_info,
2077			"failed to rebuild valid logical %llu for dev %s",
2078			logical, rcu_str_deref(dev->name));
2079	} else {
2080		scrub_write_block_to_dev_replace(sblock);
2081	}
2082
2083	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2084		mutex_lock(&sctx->wr_lock);
2085		scrub_wr_submit(sctx);
2086		mutex_unlock(&sctx->wr_lock);
2087	}
2088
2089	scrub_block_put(sblock);
2090	scrub_pending_bio_dec(sctx);
2091}
2092
2093static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2094{
2095	struct scrub_ctx *sctx = sblock->sctx;
2096	struct btrfs_fs_info *fs_info = sctx->fs_info;
2097	u64 length = sblock->page_count * PAGE_SIZE;
2098	u64 logical = sblock->pagev[0]->logical;
2099	struct btrfs_bio *bbio = NULL;
2100	struct bio *bio;
2101	struct btrfs_raid_bio *rbio;
2102	int ret;
2103	int i;
2104
2105	btrfs_bio_counter_inc_blocked(fs_info);
2106	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2107			&length, &bbio);
2108	if (ret || !bbio || !bbio->raid_map)
2109		goto bbio_out;
2110
2111	if (WARN_ON(!sctx->is_dev_replace ||
2112		    !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2113		/*
2114		 * We shouldn't be scrubbing a missing device. Even for dev
2115		 * replace, we should only get here for RAID 5/6. We either
2116		 * managed to mount something with no mirrors remaining or
2117		 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2118		 */
2119		goto bbio_out;
2120	}
2121
2122	bio = btrfs_io_bio_alloc(0);
2123	bio->bi_iter.bi_sector = logical >> 9;
2124	bio->bi_private = sblock;
2125	bio->bi_end_io = scrub_missing_raid56_end_io;
2126
2127	rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2128	if (!rbio)
2129		goto rbio_out;
2130
2131	for (i = 0; i < sblock->page_count; i++) {
2132		struct scrub_page *spage = sblock->pagev[i];
2133
2134		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
 
 
2135	}
2136
2137	btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2138	scrub_block_get(sblock);
2139	scrub_pending_bio_inc(sctx);
2140	raid56_submit_missing_rbio(rbio);
 
2141	return;
2142
2143rbio_out:
2144	bio_put(bio);
2145bbio_out:
2146	btrfs_bio_counter_dec(fs_info);
2147	btrfs_put_bbio(bbio);
2148	spin_lock(&sctx->stat_lock);
2149	sctx->stat.malloc_errors++;
2150	spin_unlock(&sctx->stat_lock);
2151}
2152
2153static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2154		       u64 physical, struct btrfs_device *dev, u64 flags,
2155		       u64 gen, int mirror_num, u8 *csum, int force,
2156		       u64 physical_for_dev_replace)
2157{
2158	struct scrub_block *sblock;
 
2159	int index;
2160
2161	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
 
2162	if (!sblock) {
2163		spin_lock(&sctx->stat_lock);
2164		sctx->stat.malloc_errors++;
2165		spin_unlock(&sctx->stat_lock);
2166		return -ENOMEM;
2167	}
2168
2169	/* one ref inside this function, plus one for each page added to
2170	 * a bio later on */
2171	refcount_set(&sblock->refs, 1);
2172	sblock->sctx = sctx;
2173	sblock->no_io_error_seen = 1;
2174
2175	for (index = 0; len > 0; index++) {
2176		struct scrub_page *spage;
2177		u64 l = min_t(u64, len, PAGE_SIZE);
 
 
 
 
 
2178
2179		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2180		if (!spage) {
2181leave_nomem:
2182			spin_lock(&sctx->stat_lock);
2183			sctx->stat.malloc_errors++;
2184			spin_unlock(&sctx->stat_lock);
2185			scrub_block_put(sblock);
2186			return -ENOMEM;
2187		}
2188		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2189		scrub_page_get(spage);
2190		sblock->pagev[index] = spage;
2191		spage->sblock = sblock;
2192		spage->dev = dev;
2193		spage->flags = flags;
2194		spage->generation = gen;
2195		spage->logical = logical;
2196		spage->physical = physical;
2197		spage->physical_for_dev_replace = physical_for_dev_replace;
2198		spage->mirror_num = mirror_num;
2199		if (csum) {
2200			spage->have_csum = 1;
2201			memcpy(spage->csum, csum, sctx->csum_size);
2202		} else {
2203			spage->have_csum = 0;
2204		}
2205		sblock->page_count++;
2206		spage->page = alloc_page(GFP_KERNEL);
2207		if (!spage->page)
2208			goto leave_nomem;
2209		len -= l;
2210		logical += l;
2211		physical += l;
2212		physical_for_dev_replace += l;
2213	}
2214
2215	WARN_ON(sblock->page_count == 0);
2216	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2217		/*
2218		 * This case should only be hit for RAID 5/6 device replace. See
2219		 * the comment in scrub_missing_raid56_pages() for details.
2220		 */
2221		scrub_missing_raid56_pages(sblock);
2222	} else {
2223		for (index = 0; index < sblock->page_count; index++) {
2224			struct scrub_page *spage = sblock->pagev[index];
2225			int ret;
2226
2227			ret = scrub_add_page_to_rd_bio(sctx, spage);
2228			if (ret) {
2229				scrub_block_put(sblock);
2230				return ret;
2231			}
2232		}
2233
2234		if (force)
2235			scrub_submit(sctx);
2236	}
2237
2238	/* last one frees, either here or in bio completion for last page */
2239	scrub_block_put(sblock);
2240	return 0;
2241}
2242
2243static void scrub_bio_end_io(struct bio *bio)
2244{
2245	struct scrub_bio *sbio = bio->bi_private;
2246	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2247
2248	sbio->status = bio->bi_status;
2249	sbio->bio = bio;
2250
2251	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2252}
2253
2254static void scrub_bio_end_io_worker(struct btrfs_work *work)
2255{
2256	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2257	struct scrub_ctx *sctx = sbio->sctx;
2258	int i;
2259
2260	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2261	if (sbio->status) {
2262		for (i = 0; i < sbio->page_count; i++) {
2263			struct scrub_page *spage = sbio->pagev[i];
2264
2265			spage->io_error = 1;
2266			spage->sblock->no_io_error_seen = 0;
2267		}
2268	}
2269
2270	/* now complete the scrub_block items that have all pages completed */
2271	for (i = 0; i < sbio->page_count; i++) {
2272		struct scrub_page *spage = sbio->pagev[i];
2273		struct scrub_block *sblock = spage->sblock;
2274
2275		if (atomic_dec_and_test(&sblock->outstanding_pages))
2276			scrub_block_complete(sblock);
2277		scrub_block_put(sblock);
2278	}
2279
2280	bio_put(sbio->bio);
2281	sbio->bio = NULL;
2282	spin_lock(&sctx->list_lock);
2283	sbio->next_free = sctx->first_free;
2284	sctx->first_free = sbio->index;
2285	spin_unlock(&sctx->list_lock);
2286
2287	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2288		mutex_lock(&sctx->wr_lock);
2289		scrub_wr_submit(sctx);
2290		mutex_unlock(&sctx->wr_lock);
2291	}
2292
2293	scrub_pending_bio_dec(sctx);
2294}
2295
2296static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2297				       unsigned long *bitmap,
2298				       u64 start, u64 len)
2299{
2300	u64 offset;
2301	u64 nsectors64;
2302	u32 nsectors;
2303	int sectorsize = sparity->sctx->fs_info->sectorsize;
2304
2305	if (len >= sparity->stripe_len) {
2306		bitmap_set(bitmap, 0, sparity->nsectors);
2307		return;
2308	}
2309
2310	start -= sparity->logic_start;
2311	start = div64_u64_rem(start, sparity->stripe_len, &offset);
2312	offset = div_u64(offset, sectorsize);
2313	nsectors64 = div_u64(len, sectorsize);
2314
2315	ASSERT(nsectors64 < UINT_MAX);
2316	nsectors = (u32)nsectors64;
2317
2318	if (offset + nsectors <= sparity->nsectors) {
2319		bitmap_set(bitmap, offset, nsectors);
2320		return;
2321	}
2322
2323	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2324	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2325}
2326
2327static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2328						   u64 start, u64 len)
2329{
2330	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2331}
2332
2333static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2334						  u64 start, u64 len)
2335{
2336	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2337}
2338
2339static void scrub_block_complete(struct scrub_block *sblock)
2340{
2341	int corrupted = 0;
2342
2343	if (!sblock->no_io_error_seen) {
2344		corrupted = 1;
2345		scrub_handle_errored_block(sblock);
2346	} else {
2347		/*
2348		 * if has checksum error, write via repair mechanism in
2349		 * dev replace case, otherwise write here in dev replace
2350		 * case.
2351		 */
2352		corrupted = scrub_checksum(sblock);
2353		if (!corrupted && sblock->sctx->is_dev_replace)
2354			scrub_write_block_to_dev_replace(sblock);
2355	}
2356
2357	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2358		u64 start = sblock->pagev[0]->logical;
2359		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2360			  PAGE_SIZE;
 
2361
 
2362		scrub_parity_mark_sectors_error(sblock->sparity,
2363						start, end - start);
2364	}
2365}
2366
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2367static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2368{
2369	struct btrfs_ordered_sum *sum = NULL;
2370	unsigned long index;
2371	unsigned long num_sectors;
2372
2373	while (!list_empty(&sctx->csum_list)) {
 
 
 
 
2374		sum = list_first_entry(&sctx->csum_list,
2375				       struct btrfs_ordered_sum, list);
 
2376		if (sum->bytenr > logical)
2377			return 0;
2378		if (sum->bytenr + sum->len > logical)
2379			break;
2380
2381		++sctx->stat.csum_discards;
2382		list_del(&sum->list);
2383		kfree(sum);
2384		sum = NULL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2385	}
2386	if (!sum)
2387		return 0;
2388
2389	index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2390	ASSERT(index < UINT_MAX);
2391
2392	num_sectors = sum->len / sctx->fs_info->sectorsize;
2393	memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2394	if (index == num_sectors - 1) {
2395		list_del(&sum->list);
2396		kfree(sum);
2397	}
2398	return 1;
2399}
2400
2401/* scrub extent tries to collect up to 64 kB for each bio */
2402static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2403			u64 logical, u64 len,
2404			u64 physical, struct btrfs_device *dev, u64 flags,
2405			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2406{
 
 
 
2407	int ret;
2408	u8 csum[BTRFS_CSUM_SIZE];
2409	u32 blocksize;
2410
2411	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2412		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2413			blocksize = map->stripe_len;
2414		else
2415			blocksize = sctx->fs_info->sectorsize;
2416		spin_lock(&sctx->stat_lock);
2417		sctx->stat.data_extents_scrubbed++;
2418		sctx->stat.data_bytes_scrubbed += len;
2419		spin_unlock(&sctx->stat_lock);
2420	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2421		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2422			blocksize = map->stripe_len;
2423		else
2424			blocksize = sctx->fs_info->nodesize;
2425		spin_lock(&sctx->stat_lock);
2426		sctx->stat.tree_extents_scrubbed++;
2427		sctx->stat.tree_bytes_scrubbed += len;
2428		spin_unlock(&sctx->stat_lock);
2429	} else {
2430		blocksize = sctx->fs_info->sectorsize;
2431		WARN_ON(1);
2432	}
2433
 
 
 
 
 
 
 
 
 
 
 
 
2434	while (len) {
2435		u64 l = min_t(u64, len, blocksize);
2436		int have_csum = 0;
2437
2438		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2439			/* push csums to sbio */
2440			have_csum = scrub_find_csum(sctx, logical, csum);
2441			if (have_csum == 0)
2442				++sctx->stat.no_csum;
2443		}
2444		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2445				  mirror_num, have_csum ? csum : NULL, 0,
2446				  physical_for_dev_replace);
2447		if (ret)
2448			return ret;
2449		len -= l;
2450		logical += l;
2451		physical += l;
2452		physical_for_dev_replace += l;
2453	}
2454	return 0;
2455}
2456
2457static int scrub_pages_for_parity(struct scrub_parity *sparity,
2458				  u64 logical, u64 len,
2459				  u64 physical, struct btrfs_device *dev,
2460				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2461{
2462	struct scrub_ctx *sctx = sparity->sctx;
2463	struct scrub_block *sblock;
 
2464	int index;
2465
2466	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
 
 
2467	if (!sblock) {
2468		spin_lock(&sctx->stat_lock);
2469		sctx->stat.malloc_errors++;
2470		spin_unlock(&sctx->stat_lock);
2471		return -ENOMEM;
2472	}
2473
2474	/* one ref inside this function, plus one for each page added to
2475	 * a bio later on */
2476	refcount_set(&sblock->refs, 1);
2477	sblock->sctx = sctx;
2478	sblock->no_io_error_seen = 1;
2479	sblock->sparity = sparity;
2480	scrub_parity_get(sparity);
2481
2482	for (index = 0; len > 0; index++) {
2483		struct scrub_page *spage;
2484		u64 l = min_t(u64, len, PAGE_SIZE);
2485
2486		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2487		if (!spage) {
2488leave_nomem:
2489			spin_lock(&sctx->stat_lock);
2490			sctx->stat.malloc_errors++;
2491			spin_unlock(&sctx->stat_lock);
2492			scrub_block_put(sblock);
2493			return -ENOMEM;
2494		}
2495		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2496		/* For scrub block */
2497		scrub_page_get(spage);
2498		sblock->pagev[index] = spage;
2499		/* For scrub parity */
2500		scrub_page_get(spage);
2501		list_add_tail(&spage->list, &sparity->spages);
2502		spage->sblock = sblock;
2503		spage->dev = dev;
2504		spage->flags = flags;
2505		spage->generation = gen;
2506		spage->logical = logical;
2507		spage->physical = physical;
2508		spage->mirror_num = mirror_num;
2509		if (csum) {
2510			spage->have_csum = 1;
2511			memcpy(spage->csum, csum, sctx->csum_size);
2512		} else {
2513			spage->have_csum = 0;
2514		}
2515		sblock->page_count++;
2516		spage->page = alloc_page(GFP_KERNEL);
2517		if (!spage->page)
2518			goto leave_nomem;
2519		len -= l;
2520		logical += l;
2521		physical += l;
2522	}
2523
2524	WARN_ON(sblock->page_count == 0);
2525	for (index = 0; index < sblock->page_count; index++) {
2526		struct scrub_page *spage = sblock->pagev[index];
2527		int ret;
2528
2529		ret = scrub_add_page_to_rd_bio(sctx, spage);
2530		if (ret) {
2531			scrub_block_put(sblock);
2532			return ret;
2533		}
2534	}
2535
2536	/* last one frees, either here or in bio completion for last page */
2537	scrub_block_put(sblock);
2538	return 0;
2539}
2540
2541static int scrub_extent_for_parity(struct scrub_parity *sparity,
2542				   u64 logical, u64 len,
2543				   u64 physical, struct btrfs_device *dev,
2544				   u64 flags, u64 gen, int mirror_num)
2545{
2546	struct scrub_ctx *sctx = sparity->sctx;
2547	int ret;
2548	u8 csum[BTRFS_CSUM_SIZE];
2549	u32 blocksize;
2550
2551	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2552		scrub_parity_mark_sectors_error(sparity, logical, len);
2553		return 0;
2554	}
2555
2556	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2557		blocksize = sparity->stripe_len;
2558	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2559		blocksize = sparity->stripe_len;
2560	} else {
2561		blocksize = sctx->fs_info->sectorsize;
2562		WARN_ON(1);
2563	}
2564
2565	while (len) {
2566		u64 l = min_t(u64, len, blocksize);
2567		int have_csum = 0;
2568
2569		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2570			/* push csums to sbio */
2571			have_csum = scrub_find_csum(sctx, logical, csum);
2572			if (have_csum == 0)
2573				goto skip;
2574		}
2575		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2576					     flags, gen, mirror_num,
2577					     have_csum ? csum : NULL);
2578		if (ret)
2579			return ret;
2580skip:
2581		len -= l;
2582		logical += l;
2583		physical += l;
2584	}
2585	return 0;
2586}
2587
2588/*
2589 * Given a physical address, this will calculate it's
2590 * logical offset. if this is a parity stripe, it will return
2591 * the most left data stripe's logical offset.
2592 *
2593 * return 0 if it is a data stripe, 1 means parity stripe.
2594 */
2595static int get_raid56_logic_offset(u64 physical, int num,
2596				   struct map_lookup *map, u64 *offset,
2597				   u64 *stripe_start)
2598{
2599	int i;
2600	int j = 0;
2601	u64 stripe_nr;
2602	u64 last_offset;
2603	u32 stripe_index;
2604	u32 rot;
2605	const int data_stripes = nr_data_stripes(map);
2606
2607	last_offset = (physical - map->stripes[num].physical) * data_stripes;
2608	if (stripe_start)
2609		*stripe_start = last_offset;
2610
2611	*offset = last_offset;
2612	for (i = 0; i < data_stripes; i++) {
2613		*offset = last_offset + i * map->stripe_len;
2614
2615		stripe_nr = div64_u64(*offset, map->stripe_len);
2616		stripe_nr = div_u64(stripe_nr, data_stripes);
2617
2618		/* Work out the disk rotation on this stripe-set */
2619		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2620		/* calculate which stripe this data locates */
2621		rot += i;
2622		stripe_index = rot % map->num_stripes;
2623		if (stripe_index == num)
2624			return 0;
2625		if (stripe_index < num)
2626			j++;
2627	}
2628	*offset = last_offset + j * map->stripe_len;
2629	return 1;
2630}
2631
2632static void scrub_free_parity(struct scrub_parity *sparity)
2633{
2634	struct scrub_ctx *sctx = sparity->sctx;
2635	struct scrub_page *curr, *next;
2636	int nbits;
2637
2638	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2639	if (nbits) {
2640		spin_lock(&sctx->stat_lock);
2641		sctx->stat.read_errors += nbits;
2642		sctx->stat.uncorrectable_errors += nbits;
2643		spin_unlock(&sctx->stat_lock);
2644	}
2645
2646	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2647		list_del_init(&curr->list);
2648		scrub_page_put(curr);
2649	}
2650
2651	kfree(sparity);
2652}
2653
2654static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2655{
2656	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2657						    work);
2658	struct scrub_ctx *sctx = sparity->sctx;
2659
 
2660	scrub_free_parity(sparity);
2661	scrub_pending_bio_dec(sctx);
2662}
2663
2664static void scrub_parity_bio_endio(struct bio *bio)
2665{
2666	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2667	struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2668
2669	if (bio->bi_status)
2670		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2671			  sparity->nsectors);
2672
2673	bio_put(bio);
2674
2675	btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2676			NULL);
2677	btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2678}
2679
2680static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2681{
2682	struct scrub_ctx *sctx = sparity->sctx;
2683	struct btrfs_fs_info *fs_info = sctx->fs_info;
2684	struct bio *bio;
2685	struct btrfs_raid_bio *rbio;
2686	struct btrfs_bio *bbio = NULL;
2687	u64 length;
2688	int ret;
2689
2690	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2691			   sparity->nsectors))
2692		goto out;
2693
2694	length = sparity->logic_end - sparity->logic_start;
2695
2696	btrfs_bio_counter_inc_blocked(fs_info);
2697	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2698			       &length, &bbio);
2699	if (ret || !bbio || !bbio->raid_map)
2700		goto bbio_out;
2701
2702	bio = btrfs_io_bio_alloc(0);
2703	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2704	bio->bi_private = sparity;
2705	bio->bi_end_io = scrub_parity_bio_endio;
2706
2707	rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2708					      length, sparity->scrub_dev,
2709					      sparity->dbitmap,
2710					      sparity->nsectors);
 
2711	if (!rbio)
2712		goto rbio_out;
2713
2714	scrub_pending_bio_inc(sctx);
2715	raid56_parity_submit_scrub_rbio(rbio);
2716	return;
2717
2718rbio_out:
2719	bio_put(bio);
2720bbio_out:
2721	btrfs_bio_counter_dec(fs_info);
2722	btrfs_put_bbio(bbio);
2723	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2724		  sparity->nsectors);
2725	spin_lock(&sctx->stat_lock);
2726	sctx->stat.malloc_errors++;
2727	spin_unlock(&sctx->stat_lock);
2728out:
2729	scrub_free_parity(sparity);
2730}
2731
2732static inline int scrub_calc_parity_bitmap_len(int nsectors)
2733{
2734	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2735}
2736
2737static void scrub_parity_get(struct scrub_parity *sparity)
2738{
2739	refcount_inc(&sparity->refs);
2740}
2741
2742static void scrub_parity_put(struct scrub_parity *sparity)
2743{
2744	if (!refcount_dec_and_test(&sparity->refs))
2745		return;
2746
2747	scrub_parity_check_and_repair(sparity);
2748}
2749
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2750static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2751						  struct map_lookup *map,
2752						  struct btrfs_device *sdev,
2753						  struct btrfs_path *path,
2754						  u64 logic_start,
2755						  u64 logic_end)
2756{
2757	struct btrfs_fs_info *fs_info = sctx->fs_info;
2758	struct btrfs_root *root = fs_info->extent_root;
2759	struct btrfs_root *csum_root = fs_info->csum_root;
2760	struct btrfs_extent_item *extent;
2761	struct btrfs_bio *bbio = NULL;
2762	u64 flags;
2763	int ret;
2764	int slot;
2765	struct extent_buffer *l;
2766	struct btrfs_key key;
2767	u64 generation;
2768	u64 extent_logical;
2769	u64 extent_physical;
2770	u64 extent_len;
2771	u64 mapped_length;
2772	struct btrfs_device *extent_dev;
2773	struct scrub_parity *sparity;
2774	int nsectors;
2775	int bitmap_len;
2776	int extent_mirror_num;
2777	int stop_loop = 0;
2778
2779	nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2780	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2781	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2782			  GFP_NOFS);
 
 
 
 
 
 
 
 
 
 
2783	if (!sparity) {
2784		spin_lock(&sctx->stat_lock);
2785		sctx->stat.malloc_errors++;
2786		spin_unlock(&sctx->stat_lock);
 
2787		return -ENOMEM;
2788	}
2789
 
2790	sparity->stripe_len = map->stripe_len;
2791	sparity->nsectors = nsectors;
2792	sparity->sctx = sctx;
2793	sparity->scrub_dev = sdev;
2794	sparity->logic_start = logic_start;
2795	sparity->logic_end = logic_end;
2796	refcount_set(&sparity->refs, 1);
2797	INIT_LIST_HEAD(&sparity->spages);
2798	sparity->dbitmap = sparity->bitmap;
2799	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2800
2801	ret = 0;
2802	while (logic_start < logic_end) {
2803		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2804			key.type = BTRFS_METADATA_ITEM_KEY;
2805		else
2806			key.type = BTRFS_EXTENT_ITEM_KEY;
2807		key.objectid = logic_start;
2808		key.offset = (u64)-1;
 
 
 
 
 
 
2809
2810		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2811		if (ret < 0)
2812			goto out;
2813
2814		if (ret > 0) {
2815			ret = btrfs_previous_extent_item(root, path, 0);
2816			if (ret < 0)
2817				goto out;
2818			if (ret > 0) {
2819				btrfs_release_path(path);
2820				ret = btrfs_search_slot(NULL, root, &key,
2821							path, 0, 0);
2822				if (ret < 0)
2823					goto out;
2824			}
2825		}
2826
2827		stop_loop = 0;
2828		while (1) {
2829			u64 bytes;
2830
2831			l = path->nodes[0];
2832			slot = path->slots[0];
2833			if (slot >= btrfs_header_nritems(l)) {
2834				ret = btrfs_next_leaf(root, path);
2835				if (ret == 0)
2836					continue;
2837				if (ret < 0)
2838					goto out;
2839
2840				stop_loop = 1;
2841				break;
2842			}
2843			btrfs_item_key_to_cpu(l, &key, slot);
2844
2845			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2846			    key.type != BTRFS_METADATA_ITEM_KEY)
2847				goto next;
 
 
2848
2849			if (key.type == BTRFS_METADATA_ITEM_KEY)
2850				bytes = fs_info->nodesize;
2851			else
2852				bytes = key.offset;
2853
2854			if (key.objectid + bytes <= logic_start)
2855				goto next;
2856
2857			if (key.objectid >= logic_end) {
2858				stop_loop = 1;
2859				break;
2860			}
 
 
 
 
 
 
 
2861
2862			while (key.objectid >= logic_start + map->stripe_len)
2863				logic_start += map->stripe_len;
2864
2865			extent = btrfs_item_ptr(l, slot,
2866						struct btrfs_extent_item);
2867			flags = btrfs_extent_flags(l, extent);
2868			generation = btrfs_extent_generation(l, extent);
2869
2870			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2871			    (key.objectid < logic_start ||
2872			     key.objectid + bytes >
2873			     logic_start + map->stripe_len)) {
2874				btrfs_err(fs_info,
2875					  "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2876					  key.objectid, logic_start);
2877				spin_lock(&sctx->stat_lock);
2878				sctx->stat.uncorrectable_errors++;
2879				spin_unlock(&sctx->stat_lock);
2880				goto next;
2881			}
2882again:
2883			extent_logical = key.objectid;
2884			extent_len = bytes;
 
 
 
 
2885
2886			if (extent_logical < logic_start) {
2887				extent_len -= logic_start - extent_logical;
2888				extent_logical = logic_start;
2889			}
2890
2891			if (extent_logical + extent_len >
2892			    logic_start + map->stripe_len)
2893				extent_len = logic_start + map->stripe_len -
2894					     extent_logical;
2895
2896			scrub_parity_mark_sectors_data(sparity, extent_logical,
2897						       extent_len);
2898
2899			mapped_length = extent_len;
2900			bbio = NULL;
2901			ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2902					extent_logical, &mapped_length, &bbio,
2903					0);
2904			if (!ret) {
2905				if (!bbio || mapped_length < extent_len)
2906					ret = -EIO;
2907			}
2908			if (ret) {
2909				btrfs_put_bbio(bbio);
2910				goto out;
2911			}
2912			extent_physical = bbio->stripes[0].physical;
2913			extent_mirror_num = bbio->mirror_num;
2914			extent_dev = bbio->stripes[0].dev;
2915			btrfs_put_bbio(bbio);
2916
2917			ret = btrfs_lookup_csums_range(csum_root,
2918						extent_logical,
2919						extent_logical + extent_len - 1,
2920						&sctx->csum_list, 1);
2921			if (ret)
2922				goto out;
2923
2924			ret = scrub_extent_for_parity(sparity, extent_logical,
2925						      extent_len,
2926						      extent_physical,
2927						      extent_dev, flags,
2928						      generation,
2929						      extent_mirror_num);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2930
2931			scrub_free_csums(sctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
2932
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2933			if (ret)
2934				goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2935
2936			if (extent_logical + extent_len <
2937			    key.objectid + bytes) {
2938				logic_start += map->stripe_len;
 
 
2939
2940				if (logic_start >= logic_end) {
2941					stop_loop = 1;
2942					break;
2943				}
2944
2945				if (logic_start < key.objectid + bytes) {
2946					cond_resched();
2947					goto again;
2948				}
2949			}
2950next:
2951			path->slots[0]++;
2952		}
2953
2954		btrfs_release_path(path);
 
 
 
 
 
2955
2956		if (stop_loop)
2957			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2958
2959		logic_start += map->stripe_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2960	}
2961out:
2962	if (ret < 0)
2963		scrub_parity_mark_sectors_error(sparity, logic_start,
2964						logic_end - logic_start);
2965	scrub_parity_put(sparity);
2966	scrub_submit(sctx);
2967	mutex_lock(&sctx->wr_lock);
2968	scrub_wr_submit(sctx);
2969	mutex_unlock(&sctx->wr_lock);
2970
2971	btrfs_release_path(path);
2972	return ret < 0 ? ret : 0;
2973}
2974
2975static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2976					   struct map_lookup *map,
 
2977					   struct btrfs_device *scrub_dev,
2978					   int num, u64 base, u64 length,
2979					   struct btrfs_block_group *cache)
2980{
2981	struct btrfs_path *path, *ppath;
2982	struct btrfs_fs_info *fs_info = sctx->fs_info;
2983	struct btrfs_root *root = fs_info->extent_root;
2984	struct btrfs_root *csum_root = fs_info->csum_root;
2985	struct btrfs_extent_item *extent;
2986	struct blk_plug plug;
2987	u64 flags;
 
 
2988	int ret;
2989	int slot;
2990	u64 nstripes;
2991	struct extent_buffer *l;
2992	u64 physical;
2993	u64 logical;
2994	u64 logic_end;
2995	u64 physical_end;
2996	u64 generation;
2997	int mirror_num;
2998	struct reada_control *reada1;
2999	struct reada_control *reada2;
3000	struct btrfs_key key;
3001	struct btrfs_key key_end;
3002	u64 increment = map->stripe_len;
3003	u64 offset;
3004	u64 extent_logical;
3005	u64 extent_physical;
3006	u64 extent_len;
3007	u64 stripe_logical;
3008	u64 stripe_end;
3009	struct btrfs_device *extent_dev;
3010	int extent_mirror_num;
3011	int stop_loop = 0;
3012
3013	physical = map->stripes[num].physical;
3014	offset = 0;
3015	nstripes = div64_u64(length, map->stripe_len);
3016	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3017		offset = map->stripe_len * num;
3018		increment = map->stripe_len * map->num_stripes;
3019		mirror_num = 1;
3020	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3021		int factor = map->num_stripes / map->sub_stripes;
3022		offset = map->stripe_len * (num / map->sub_stripes);
3023		increment = map->stripe_len * factor;
3024		mirror_num = num % map->sub_stripes + 1;
3025	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3026		increment = map->stripe_len;
3027		mirror_num = num % map->num_stripes + 1;
3028	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3029		increment = map->stripe_len;
3030		mirror_num = num % map->num_stripes + 1;
3031	} else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3032		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3033		increment = map->stripe_len * nr_data_stripes(map);
3034		mirror_num = 1;
3035	} else {
3036		increment = map->stripe_len;
3037		mirror_num = 1;
3038	}
3039
3040	path = btrfs_alloc_path();
3041	if (!path)
3042		return -ENOMEM;
3043
3044	ppath = btrfs_alloc_path();
3045	if (!ppath) {
3046		btrfs_free_path(path);
3047		return -ENOMEM;
3048	}
3049
3050	/*
3051	 * work on commit root. The related disk blocks are static as
3052	 * long as COW is applied. This means, it is save to rewrite
3053	 * them to repair disk errors without any race conditions
3054	 */
3055	path->search_commit_root = 1;
3056	path->skip_locking = 1;
 
3057
3058	ppath->search_commit_root = 1;
3059	ppath->skip_locking = 1;
3060	/*
3061	 * trigger the readahead for extent tree csum tree and wait for
3062	 * completion. During readahead, the scrub is officially paused
3063	 * to not hold off transaction commits
3064	 */
3065	logical = base + offset;
3066	physical_end = physical + nstripes * map->stripe_len;
3067	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3068		get_raid56_logic_offset(physical_end, num,
3069					map, &logic_end, NULL);
3070		logic_end += base;
3071	} else {
3072		logic_end = logical + increment * nstripes;
3073	}
3074	wait_event(sctx->list_wait,
3075		   atomic_read(&sctx->bios_in_flight) == 0);
3076	scrub_blocked_if_needed(fs_info);
3077
3078	/* FIXME it might be better to start readahead at commit root */
3079	key.objectid = logical;
3080	key.type = BTRFS_EXTENT_ITEM_KEY;
3081	key.offset = (u64)0;
3082	key_end.objectid = logic_end;
3083	key_end.type = BTRFS_METADATA_ITEM_KEY;
3084	key_end.offset = (u64)-1;
3085	reada1 = btrfs_reada_add(root, &key, &key_end);
3086
3087	key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3088	key.type = BTRFS_EXTENT_CSUM_KEY;
3089	key.offset = logical;
3090	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3091	key_end.type = BTRFS_EXTENT_CSUM_KEY;
3092	key_end.offset = logic_end;
3093	reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3094
3095	if (!IS_ERR(reada1))
3096		btrfs_reada_wait(reada1);
3097	if (!IS_ERR(reada2))
3098		btrfs_reada_wait(reada2);
3099
3100
3101	/*
3102	 * collect all data csums for the stripe to avoid seeking during
3103	 * the scrub. This might currently (crc32) end up to be about 1MB
3104	 */
3105	blk_start_plug(&plug);
3106
 
 
 
 
 
 
 
 
3107	/*
3108	 * now find all extents for each stripe and scrub them
 
 
 
 
3109	 */
3110	ret = 0;
3111	while (physical < physical_end) {
3112		/*
3113		 * canceled?
 
 
 
 
 
3114		 */
3115		if (atomic_read(&fs_info->scrub_cancel_req) ||
3116		    atomic_read(&sctx->cancel_req)) {
3117			ret = -ECANCELED;
3118			goto out;
3119		}
3120		/*
3121		 * check to see if we have to pause
3122		 */
3123		if (atomic_read(&fs_info->scrub_pause_req)) {
3124			/* push queued extents */
3125			sctx->flush_all_writes = true;
3126			scrub_submit(sctx);
3127			mutex_lock(&sctx->wr_lock);
3128			scrub_wr_submit(sctx);
3129			mutex_unlock(&sctx->wr_lock);
3130			wait_event(sctx->list_wait,
3131				   atomic_read(&sctx->bios_in_flight) == 0);
3132			sctx->flush_all_writes = false;
3133			scrub_blocked_if_needed(fs_info);
3134		}
3135
3136		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3137			ret = get_raid56_logic_offset(physical, num, map,
3138						      &logical,
3139						      &stripe_logical);
3140			logical += base;
3141			if (ret) {
3142				/* it is parity strip */
3143				stripe_logical += base;
3144				stripe_end = stripe_logical + increment;
3145				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3146							  ppath, stripe_logical,
3147							  stripe_end);
3148				if (ret)
3149					goto out;
3150				goto skip;
3151			}
3152		}
3153
3154		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3155			key.type = BTRFS_METADATA_ITEM_KEY;
3156		else
3157			key.type = BTRFS_EXTENT_ITEM_KEY;
3158		key.objectid = logical;
3159		key.offset = (u64)-1;
 
 
3160
3161		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3162		if (ret < 0)
3163			goto out;
3164
3165		if (ret > 0) {
3166			ret = btrfs_previous_extent_item(root, path, 0);
3167			if (ret < 0)
 
 
 
 
 
 
 
 
 
3168				goto out;
3169			if (ret > 0) {
3170				/* there's no smaller item, so stick with the
3171				 * larger one */
3172				btrfs_release_path(path);
3173				ret = btrfs_search_slot(NULL, root, &key,
3174							path, 0, 0);
3175				if (ret < 0)
3176					goto out;
3177			}
3178		}
3179
3180		stop_loop = 0;
3181		while (1) {
3182			u64 bytes;
3183
3184			l = path->nodes[0];
3185			slot = path->slots[0];
3186			if (slot >= btrfs_header_nritems(l)) {
3187				ret = btrfs_next_leaf(root, path);
3188				if (ret == 0)
3189					continue;
3190				if (ret < 0)
3191					goto out;
3192
3193				stop_loop = 1;
3194				break;
3195			}
3196			btrfs_item_key_to_cpu(l, &key, slot);
3197
3198			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3199			    key.type != BTRFS_METADATA_ITEM_KEY)
3200				goto next;
3201
3202			if (key.type == BTRFS_METADATA_ITEM_KEY)
3203				bytes = fs_info->nodesize;
3204			else
3205				bytes = key.offset;
3206
3207			if (key.objectid + bytes <= logical)
3208				goto next;
3209
3210			if (key.objectid >= logical + map->stripe_len) {
3211				/* out of this device extent */
3212				if (key.objectid >= logic_end)
3213					stop_loop = 1;
3214				break;
3215			}
3216
3217			/*
3218			 * If our block group was removed in the meanwhile, just
3219			 * stop scrubbing since there is no point in continuing.
3220			 * Continuing would prevent reusing its device extents
3221			 * for new block groups for a long time.
3222			 */
3223			spin_lock(&cache->lock);
3224			if (cache->removed) {
3225				spin_unlock(&cache->lock);
3226				ret = 0;
3227				goto out;
3228			}
3229			spin_unlock(&cache->lock);
3230
3231			extent = btrfs_item_ptr(l, slot,
3232						struct btrfs_extent_item);
3233			flags = btrfs_extent_flags(l, extent);
3234			generation = btrfs_extent_generation(l, extent);
3235
3236			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3237			    (key.objectid < logical ||
3238			     key.objectid + bytes >
3239			     logical + map->stripe_len)) {
3240				btrfs_err(fs_info,
3241					   "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3242				       key.objectid, logical);
3243				spin_lock(&sctx->stat_lock);
3244				sctx->stat.uncorrectable_errors++;
3245				spin_unlock(&sctx->stat_lock);
3246				goto next;
3247			}
3248
3249again:
3250			extent_logical = key.objectid;
3251			extent_len = bytes;
3252
3253			/*
3254			 * trim extent to this stripe
3255			 */
3256			if (extent_logical < logical) {
3257				extent_len -= logical - extent_logical;
3258				extent_logical = logical;
3259			}
3260			if (extent_logical + extent_len >
3261			    logical + map->stripe_len) {
3262				extent_len = logical + map->stripe_len -
3263					     extent_logical;
3264			}
3265
3266			extent_physical = extent_logical - logical + physical;
3267			extent_dev = scrub_dev;
3268			extent_mirror_num = mirror_num;
3269			if (sctx->is_dev_replace)
3270				scrub_remap_extent(fs_info, extent_logical,
3271						   extent_len, &extent_physical,
3272						   &extent_dev,
3273						   &extent_mirror_num);
3274
3275			if (flags & BTRFS_EXTENT_FLAG_DATA) {
3276				ret = btrfs_lookup_csums_range(csum_root,
3277						extent_logical,
3278						extent_logical + extent_len - 1,
3279						&sctx->csum_list, 1);
3280				if (ret)
3281					goto out;
3282			}
3283
3284			ret = scrub_extent(sctx, map, extent_logical, extent_len,
3285					   extent_physical, extent_dev, flags,
3286					   generation, extent_mirror_num,
3287					   extent_logical - logical + physical);
3288
3289			scrub_free_csums(sctx);
3290
3291			if (ret)
3292				goto out;
3293
3294			if (extent_logical + extent_len <
3295			    key.objectid + bytes) {
3296				if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3297					/*
3298					 * loop until we find next data stripe
3299					 * or we have finished all stripes.
3300					 */
3301loop:
3302					physical += map->stripe_len;
3303					ret = get_raid56_logic_offset(physical,
3304							num, map, &logical,
3305							&stripe_logical);
3306					logical += base;
3307
3308					if (ret && physical < physical_end) {
3309						stripe_logical += base;
3310						stripe_end = stripe_logical +
3311								increment;
3312						ret = scrub_raid56_parity(sctx,
3313							map, scrub_dev, ppath,
3314							stripe_logical,
3315							stripe_end);
3316						if (ret)
3317							goto out;
3318						goto loop;
3319					}
3320				} else {
3321					physical += map->stripe_len;
3322					logical += increment;
3323				}
3324				if (logical < key.objectid + bytes) {
3325					cond_resched();
3326					goto again;
3327				}
3328
3329				if (physical >= physical_end) {
3330					stop_loop = 1;
3331					break;
3332				}
3333			}
3334next:
3335			path->slots[0]++;
3336		}
3337		btrfs_release_path(path);
3338skip:
3339		logical += increment;
3340		physical += map->stripe_len;
3341		spin_lock(&sctx->stat_lock);
3342		if (stop_loop)
3343			sctx->stat.last_physical = map->stripes[num].physical +
3344						   length;
3345		else
3346			sctx->stat.last_physical = physical;
3347		spin_unlock(&sctx->stat_lock);
3348		if (stop_loop)
3349			break;
3350	}
3351out:
3352	/* push queued extents */
3353	scrub_submit(sctx);
3354	mutex_lock(&sctx->wr_lock);
3355	scrub_wr_submit(sctx);
3356	mutex_unlock(&sctx->wr_lock);
3357
3358	blk_finish_plug(&plug);
3359	btrfs_free_path(path);
3360	btrfs_free_path(ppath);
 
 
 
 
 
 
 
 
 
 
 
3361	return ret < 0 ? ret : 0;
3362}
3363
3364static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
 
3365					  struct btrfs_device *scrub_dev,
3366					  u64 chunk_offset, u64 length,
3367					  u64 dev_offset,
3368					  struct btrfs_block_group *cache)
3369{
3370	struct btrfs_fs_info *fs_info = sctx->fs_info;
3371	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3372	struct map_lookup *map;
3373	struct extent_map *em;
3374	int i;
3375	int ret = 0;
3376
3377	read_lock(&map_tree->lock);
3378	em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3379	read_unlock(&map_tree->lock);
3380
3381	if (!em) {
3382		/*
3383		 * Might have been an unused block group deleted by the cleaner
3384		 * kthread or relocation.
3385		 */
3386		spin_lock(&cache->lock);
3387		if (!cache->removed)
3388			ret = -EINVAL;
3389		spin_unlock(&cache->lock);
3390
3391		return ret;
3392	}
3393
3394	map = em->map_lookup;
3395	if (em->start != chunk_offset)
3396		goto out;
3397
3398	if (em->len < length)
3399		goto out;
3400
 
3401	for (i = 0; i < map->num_stripes; ++i) {
3402		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3403		    map->stripes[i].physical == dev_offset) {
3404			ret = scrub_stripe(sctx, map, scrub_dev, i,
3405					   chunk_offset, length, cache);
3406			if (ret)
3407				goto out;
3408		}
3409	}
3410out:
3411	free_extent_map(em);
3412
3413	return ret;
3414}
3415
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3416static noinline_for_stack
3417int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3418			   struct btrfs_device *scrub_dev, u64 start, u64 end)
3419{
3420	struct btrfs_dev_extent *dev_extent = NULL;
3421	struct btrfs_path *path;
3422	struct btrfs_fs_info *fs_info = sctx->fs_info;
3423	struct btrfs_root *root = fs_info->dev_root;
3424	u64 length;
3425	u64 chunk_offset;
3426	int ret = 0;
3427	int ro_set;
3428	int slot;
3429	struct extent_buffer *l;
3430	struct btrfs_key key;
3431	struct btrfs_key found_key;
3432	struct btrfs_block_group *cache;
3433	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3434
3435	path = btrfs_alloc_path();
3436	if (!path)
3437		return -ENOMEM;
3438
3439	path->reada = READA_FORWARD;
3440	path->search_commit_root = 1;
3441	path->skip_locking = 1;
3442
3443	key.objectid = scrub_dev->devid;
3444	key.offset = 0ull;
3445	key.type = BTRFS_DEV_EXTENT_KEY;
3446
3447	while (1) {
 
 
3448		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3449		if (ret < 0)
3450			break;
3451		if (ret > 0) {
3452			if (path->slots[0] >=
3453			    btrfs_header_nritems(path->nodes[0])) {
3454				ret = btrfs_next_leaf(root, path);
3455				if (ret < 0)
3456					break;
3457				if (ret > 0) {
3458					ret = 0;
3459					break;
3460				}
3461			} else {
3462				ret = 0;
3463			}
3464		}
3465
3466		l = path->nodes[0];
3467		slot = path->slots[0];
3468
3469		btrfs_item_key_to_cpu(l, &found_key, slot);
3470
3471		if (found_key.objectid != scrub_dev->devid)
3472			break;
3473
3474		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3475			break;
3476
3477		if (found_key.offset >= end)
3478			break;
3479
3480		if (found_key.offset < key.offset)
3481			break;
3482
3483		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3484		length = btrfs_dev_extent_length(l, dev_extent);
3485
3486		if (found_key.offset + length <= start)
3487			goto skip;
3488
3489		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3490
3491		/*
3492		 * get a reference on the corresponding block group to prevent
3493		 * the chunk from going away while we scrub it
3494		 */
3495		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3496
3497		/* some chunks are removed but not committed to disk yet,
3498		 * continue scrubbing */
3499		if (!cache)
3500			goto skip;
3501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3502		/*
3503		 * Make sure that while we are scrubbing the corresponding block
3504		 * group doesn't get its logical address and its device extents
3505		 * reused for another block group, which can possibly be of a
3506		 * different type and different profile. We do this to prevent
3507		 * false error detections and crashes due to bogus attempts to
3508		 * repair extents.
3509		 */
3510		spin_lock(&cache->lock);
3511		if (cache->removed) {
3512			spin_unlock(&cache->lock);
3513			btrfs_put_block_group(cache);
3514			goto skip;
3515		}
3516		btrfs_freeze_block_group(cache);
3517		spin_unlock(&cache->lock);
3518
3519		/*
3520		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3521		 * to avoid deadlock caused by:
3522		 * btrfs_inc_block_group_ro()
3523		 * -> btrfs_wait_for_commit()
3524		 * -> btrfs_commit_transaction()
3525		 * -> btrfs_scrub_pause()
3526		 */
3527		scrub_pause_on(fs_info);
3528
3529		/*
3530		 * Don't do chunk preallocation for scrub.
3531		 *
3532		 * This is especially important for SYSTEM bgs, or we can hit
3533		 * -EFBIG from btrfs_finish_chunk_alloc() like:
3534		 * 1. The only SYSTEM bg is marked RO.
3535		 *    Since SYSTEM bg is small, that's pretty common.
3536		 * 2. New SYSTEM bg will be allocated
3537		 *    Due to regular version will allocate new chunk.
3538		 * 3. New SYSTEM bg is empty and will get cleaned up
3539		 *    Before cleanup really happens, it's marked RO again.
3540		 * 4. Empty SYSTEM bg get scrubbed
3541		 *    We go back to 2.
3542		 *
3543		 * This can easily boost the amount of SYSTEM chunks if cleaner
3544		 * thread can't be triggered fast enough, and use up all space
3545		 * of btrfs_super_block::sys_chunk_array
3546		 *
3547		 * While for dev replace, we need to try our best to mark block
3548		 * group RO, to prevent race between:
3549		 * - Write duplication
3550		 *   Contains latest data
3551		 * - Scrub copy
3552		 *   Contains data from commit tree
3553		 *
3554		 * If target block group is not marked RO, nocow writes can
3555		 * be overwritten by scrub copy, causing data corruption.
3556		 * So for dev-replace, it's not allowed to continue if a block
3557		 * group is not RO.
3558		 */
3559		ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
 
 
 
 
 
 
 
 
 
 
3560		if (ret == 0) {
3561			ro_set = 1;
3562		} else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3563			/*
3564			 * btrfs_inc_block_group_ro return -ENOSPC when it
3565			 * failed in creating new chunk for metadata.
3566			 * It is not a problem for scrub, because
3567			 * metadata are always cowed, and our scrub paused
3568			 * commit_transactions.
3569			 */
3570			ro_set = 0;
 
 
 
 
 
 
 
3571		} else {
3572			btrfs_warn(fs_info,
3573				   "failed setting block group ro: %d", ret);
3574			btrfs_unfreeze_block_group(cache);
3575			btrfs_put_block_group(cache);
3576			scrub_pause_off(fs_info);
3577			break;
3578		}
3579
3580		/*
3581		 * Now the target block is marked RO, wait for nocow writes to
3582		 * finish before dev-replace.
3583		 * COW is fine, as COW never overwrites extents in commit tree.
3584		 */
3585		if (sctx->is_dev_replace) {
3586			btrfs_wait_nocow_writers(cache);
3587			btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3588					cache->length);
3589		}
3590
3591		scrub_pause_off(fs_info);
3592		down_write(&dev_replace->rwsem);
3593		dev_replace->cursor_right = found_key.offset + length;
3594		dev_replace->cursor_left = found_key.offset;
3595		dev_replace->item_needs_writeback = 1;
3596		up_write(&dev_replace->rwsem);
3597
3598		ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3599				  found_key.offset, cache);
3600
3601		/*
3602		 * flush, submit all pending read and write bios, afterwards
3603		 * wait for them.
3604		 * Note that in the dev replace case, a read request causes
3605		 * write requests that are submitted in the read completion
3606		 * worker. Therefore in the current situation, it is required
3607		 * that all write requests are flushed, so that all read and
3608		 * write requests are really completed when bios_in_flight
3609		 * changes to 0.
3610		 */
3611		sctx->flush_all_writes = true;
3612		scrub_submit(sctx);
3613		mutex_lock(&sctx->wr_lock);
3614		scrub_wr_submit(sctx);
3615		mutex_unlock(&sctx->wr_lock);
3616
3617		wait_event(sctx->list_wait,
3618			   atomic_read(&sctx->bios_in_flight) == 0);
3619
3620		scrub_pause_on(fs_info);
3621
3622		/*
3623		 * must be called before we decrease @scrub_paused.
3624		 * make sure we don't block transaction commit while
3625		 * we are waiting pending workers finished.
3626		 */
3627		wait_event(sctx->list_wait,
3628			   atomic_read(&sctx->workers_pending) == 0);
3629		sctx->flush_all_writes = false;
3630
3631		scrub_pause_off(fs_info);
3632
 
 
 
 
 
3633		down_write(&dev_replace->rwsem);
3634		dev_replace->cursor_left = dev_replace->cursor_right;
3635		dev_replace->item_needs_writeback = 1;
3636		up_write(&dev_replace->rwsem);
3637
3638		if (ro_set)
3639			btrfs_dec_block_group_ro(cache);
3640
3641		/*
3642		 * We might have prevented the cleaner kthread from deleting
3643		 * this block group if it was already unused because we raced
3644		 * and set it to RO mode first. So add it back to the unused
3645		 * list, otherwise it might not ever be deleted unless a manual
3646		 * balance is triggered or it becomes used and unused again.
3647		 */
3648		spin_lock(&cache->lock);
3649		if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3650		    cache->used == 0) {
3651			spin_unlock(&cache->lock);
3652			if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3653				btrfs_discard_queue_work(&fs_info->discard_ctl,
3654							 cache);
3655			else
3656				btrfs_mark_bg_unused(cache);
3657		} else {
3658			spin_unlock(&cache->lock);
3659		}
3660
3661		btrfs_unfreeze_block_group(cache);
3662		btrfs_put_block_group(cache);
3663		if (ret)
3664			break;
3665		if (sctx->is_dev_replace &&
3666		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3667			ret = -EIO;
3668			break;
3669		}
3670		if (sctx->stat.malloc_errors > 0) {
3671			ret = -ENOMEM;
3672			break;
3673		}
3674skip:
3675		key.offset = found_key.offset + length;
3676		btrfs_release_path(path);
3677	}
3678
3679	btrfs_free_path(path);
3680
3681	return ret;
3682}
3683
3684static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3685					   struct btrfs_device *scrub_dev)
3686{
3687	int	i;
3688	u64	bytenr;
3689	u64	gen;
3690	int	ret;
3691	struct btrfs_fs_info *fs_info = sctx->fs_info;
3692
3693	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3694		return -EROFS;
3695
3696	/* Seed devices of a new filesystem has their own generation. */
3697	if (scrub_dev->fs_devices != fs_info->fs_devices)
3698		gen = scrub_dev->generation;
3699	else
3700		gen = fs_info->last_trans_committed;
3701
3702	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3703		bytenr = btrfs_sb_offset(i);
3704		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3705		    scrub_dev->commit_total_bytes)
3706			break;
 
 
3707
3708		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3709				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3710				  NULL, 1, bytenr);
3711		if (ret)
3712			return ret;
3713	}
3714	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3715
3716	return 0;
3717}
3718
3719static void scrub_workers_put(struct btrfs_fs_info *fs_info)
3720{
3721	if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
3722					&fs_info->scrub_lock)) {
3723		struct btrfs_workqueue *scrub_workers = NULL;
3724		struct btrfs_workqueue *scrub_wr_comp = NULL;
3725		struct btrfs_workqueue *scrub_parity = NULL;
3726
3727		scrub_workers = fs_info->scrub_workers;
3728		scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3729		scrub_parity = fs_info->scrub_parity_workers;
3730
3731		fs_info->scrub_workers = NULL;
3732		fs_info->scrub_wr_completion_workers = NULL;
3733		fs_info->scrub_parity_workers = NULL;
3734		mutex_unlock(&fs_info->scrub_lock);
3735
3736		btrfs_destroy_workqueue(scrub_workers);
3737		btrfs_destroy_workqueue(scrub_wr_comp);
3738		btrfs_destroy_workqueue(scrub_parity);
 
 
 
3739	}
3740}
3741
3742/*
3743 * get a reference count on fs_info->scrub_workers. start worker if necessary
3744 */
3745static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3746						int is_dev_replace)
3747{
3748	struct btrfs_workqueue *scrub_workers = NULL;
3749	struct btrfs_workqueue *scrub_wr_comp = NULL;
3750	struct btrfs_workqueue *scrub_parity = NULL;
3751	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3752	int max_active = fs_info->thread_pool_size;
3753	int ret = -ENOMEM;
3754
3755	if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
3756		return 0;
3757
3758	scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub", flags,
3759					      is_dev_replace ? 1 : max_active, 4);
3760	if (!scrub_workers)
3761		goto fail_scrub_workers;
3762
3763	scrub_wr_comp = btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3764					      max_active, 2);
3765	if (!scrub_wr_comp)
3766		goto fail_scrub_wr_completion_workers;
3767
3768	scrub_parity = btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3769					     max_active, 2);
3770	if (!scrub_parity)
3771		goto fail_scrub_parity_workers;
3772
3773	mutex_lock(&fs_info->scrub_lock);
3774	if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3775		ASSERT(fs_info->scrub_workers == NULL &&
3776		       fs_info->scrub_wr_completion_workers == NULL &&
3777		       fs_info->scrub_parity_workers == NULL);
3778		fs_info->scrub_workers = scrub_workers;
3779		fs_info->scrub_wr_completion_workers = scrub_wr_comp;
3780		fs_info->scrub_parity_workers = scrub_parity;
3781		refcount_set(&fs_info->scrub_workers_refcnt, 1);
3782		mutex_unlock(&fs_info->scrub_lock);
3783		return 0;
3784	}
3785	/* Other thread raced in and created the workers for us */
3786	refcount_inc(&fs_info->scrub_workers_refcnt);
3787	mutex_unlock(&fs_info->scrub_lock);
3788
3789	ret = 0;
3790	btrfs_destroy_workqueue(scrub_parity);
3791fail_scrub_parity_workers:
3792	btrfs_destroy_workqueue(scrub_wr_comp);
3793fail_scrub_wr_completion_workers:
3794	btrfs_destroy_workqueue(scrub_workers);
3795fail_scrub_workers:
3796	return ret;
3797}
3798
3799int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3800		    u64 end, struct btrfs_scrub_progress *progress,
3801		    int readonly, int is_dev_replace)
3802{
 
3803	struct scrub_ctx *sctx;
3804	int ret;
3805	struct btrfs_device *dev;
3806	unsigned int nofs_flag;
 
3807
3808	if (btrfs_fs_closing(fs_info))
3809		return -EAGAIN;
3810
3811	if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3812		/*
3813		 * in this case scrub is unable to calculate the checksum
3814		 * the way scrub is implemented. Do not handle this
3815		 * situation at all because it won't ever happen.
3816		 */
3817		btrfs_err(fs_info,
3818			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3819		       fs_info->nodesize,
3820		       BTRFS_STRIPE_LEN);
3821		return -EINVAL;
3822	}
3823
3824	if (fs_info->sectorsize != PAGE_SIZE) {
3825		/* not supported for data w/o checksums */
3826		btrfs_err_rl(fs_info,
3827			   "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3828		       fs_info->sectorsize, PAGE_SIZE);
3829		return -EINVAL;
3830	}
3831
3832	if (fs_info->nodesize >
3833	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3834	    fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3835		/*
3836		 * would exhaust the array bounds of pagev member in
3837		 * struct scrub_block
3838		 */
3839		btrfs_err(fs_info,
3840			  "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3841		       fs_info->nodesize,
3842		       SCRUB_MAX_PAGES_PER_BLOCK,
3843		       fs_info->sectorsize,
3844		       SCRUB_MAX_PAGES_PER_BLOCK);
3845		return -EINVAL;
3846	}
3847
3848	/* Allocate outside of device_list_mutex */
3849	sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3850	if (IS_ERR(sctx))
3851		return PTR_ERR(sctx);
3852
3853	ret = scrub_workers_get(fs_info, is_dev_replace);
3854	if (ret)
3855		goto out_free_ctx;
3856
3857	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3858	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3859	if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3860		     !is_dev_replace)) {
3861		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3862		ret = -ENODEV;
3863		goto out;
3864	}
3865
3866	if (!is_dev_replace && !readonly &&
3867	    !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3868		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3869		btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3870				rcu_str_deref(dev->name));
 
3871		ret = -EROFS;
3872		goto out;
3873	}
3874
3875	mutex_lock(&fs_info->scrub_lock);
3876	if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3877	    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3878		mutex_unlock(&fs_info->scrub_lock);
3879		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3880		ret = -EIO;
3881		goto out;
3882	}
3883
3884	down_read(&fs_info->dev_replace.rwsem);
3885	if (dev->scrub_ctx ||
3886	    (!is_dev_replace &&
3887	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3888		up_read(&fs_info->dev_replace.rwsem);
3889		mutex_unlock(&fs_info->scrub_lock);
3890		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3891		ret = -EINPROGRESS;
3892		goto out;
3893	}
3894	up_read(&fs_info->dev_replace.rwsem);
3895
3896	sctx->readonly = readonly;
3897	dev->scrub_ctx = sctx;
3898	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3899
3900	/*
3901	 * checking @scrub_pause_req here, we can avoid
3902	 * race between committing transaction and scrubbing.
3903	 */
3904	__scrub_blocked_if_needed(fs_info);
3905	atomic_inc(&fs_info->scrubs_running);
3906	mutex_unlock(&fs_info->scrub_lock);
3907
3908	/*
3909	 * In order to avoid deadlock with reclaim when there is a transaction
3910	 * trying to pause scrub, make sure we use GFP_NOFS for all the
3911	 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3912	 * invoked by our callees. The pausing request is done when the
3913	 * transaction commit starts, and it blocks the transaction until scrub
3914	 * is paused (done at specific points at scrub_stripe() or right above
3915	 * before incrementing fs_info->scrubs_running).
3916	 */
3917	nofs_flag = memalloc_nofs_save();
3918	if (!is_dev_replace) {
 
 
 
 
 
 
3919		btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3920		/*
3921		 * by holding device list mutex, we can
3922		 * kick off writing super in log tree sync.
3923		 */
3924		mutex_lock(&fs_info->fs_devices->device_list_mutex);
3925		ret = scrub_supers(sctx, dev);
3926		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 
 
 
 
 
 
 
 
 
 
3927	}
3928
3929	if (!ret)
3930		ret = scrub_enumerate_chunks(sctx, dev, start, end);
3931	memalloc_nofs_restore(nofs_flag);
3932
3933	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3934	atomic_dec(&fs_info->scrubs_running);
3935	wake_up(&fs_info->scrub_pause_wait);
3936
3937	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3938
3939	if (progress)
3940		memcpy(progress, &sctx->stat, sizeof(*progress));
3941
3942	if (!is_dev_replace)
3943		btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3944			ret ? "not finished" : "finished", devid, ret);
3945
3946	mutex_lock(&fs_info->scrub_lock);
3947	dev->scrub_ctx = NULL;
3948	mutex_unlock(&fs_info->scrub_lock);
3949
3950	scrub_workers_put(fs_info);
3951	scrub_put_ctx(sctx);
3952
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3953	return ret;
3954out:
3955	scrub_workers_put(fs_info);
3956out_free_ctx:
3957	scrub_free_ctx(sctx);
3958
3959	return ret;
3960}
3961
3962void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3963{
3964	mutex_lock(&fs_info->scrub_lock);
3965	atomic_inc(&fs_info->scrub_pause_req);
3966	while (atomic_read(&fs_info->scrubs_paused) !=
3967	       atomic_read(&fs_info->scrubs_running)) {
3968		mutex_unlock(&fs_info->scrub_lock);
3969		wait_event(fs_info->scrub_pause_wait,
3970			   atomic_read(&fs_info->scrubs_paused) ==
3971			   atomic_read(&fs_info->scrubs_running));
3972		mutex_lock(&fs_info->scrub_lock);
3973	}
3974	mutex_unlock(&fs_info->scrub_lock);
3975}
3976
3977void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3978{
3979	atomic_dec(&fs_info->scrub_pause_req);
3980	wake_up(&fs_info->scrub_pause_wait);
3981}
3982
3983int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3984{
3985	mutex_lock(&fs_info->scrub_lock);
3986	if (!atomic_read(&fs_info->scrubs_running)) {
3987		mutex_unlock(&fs_info->scrub_lock);
3988		return -ENOTCONN;
3989	}
3990
3991	atomic_inc(&fs_info->scrub_cancel_req);
3992	while (atomic_read(&fs_info->scrubs_running)) {
3993		mutex_unlock(&fs_info->scrub_lock);
3994		wait_event(fs_info->scrub_pause_wait,
3995			   atomic_read(&fs_info->scrubs_running) == 0);
3996		mutex_lock(&fs_info->scrub_lock);
3997	}
3998	atomic_dec(&fs_info->scrub_cancel_req);
3999	mutex_unlock(&fs_info->scrub_lock);
4000
4001	return 0;
4002}
4003
4004int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4005{
4006	struct btrfs_fs_info *fs_info = dev->fs_info;
4007	struct scrub_ctx *sctx;
4008
4009	mutex_lock(&fs_info->scrub_lock);
4010	sctx = dev->scrub_ctx;
4011	if (!sctx) {
4012		mutex_unlock(&fs_info->scrub_lock);
4013		return -ENOTCONN;
4014	}
4015	atomic_inc(&sctx->cancel_req);
4016	while (dev->scrub_ctx) {
4017		mutex_unlock(&fs_info->scrub_lock);
4018		wait_event(fs_info->scrub_pause_wait,
4019			   dev->scrub_ctx == NULL);
4020		mutex_lock(&fs_info->scrub_lock);
4021	}
4022	mutex_unlock(&fs_info->scrub_lock);
4023
4024	return 0;
4025}
4026
4027int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4028			 struct btrfs_scrub_progress *progress)
4029{
 
4030	struct btrfs_device *dev;
4031	struct scrub_ctx *sctx = NULL;
4032
4033	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4034	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4035	if (dev)
4036		sctx = dev->scrub_ctx;
4037	if (sctx)
4038		memcpy(progress, &sctx->stat, sizeof(*progress));
4039	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4040
4041	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4042}
4043
4044static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4045			       u64 extent_logical, u64 extent_len,
4046			       u64 *extent_physical,
4047			       struct btrfs_device **extent_dev,
4048			       int *extent_mirror_num)
4049{
4050	u64 mapped_length;
4051	struct btrfs_bio *bbio = NULL;
4052	int ret;
4053
4054	mapped_length = extent_len;
4055	ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4056			      &mapped_length, &bbio, 0);
4057	if (ret || !bbio || mapped_length < extent_len ||
4058	    !bbio->stripes[0].dev->bdev) {
4059		btrfs_put_bbio(bbio);
4060		return;
4061	}
4062
4063	*extent_physical = bbio->stripes[0].physical;
4064	*extent_mirror_num = bbio->mirror_num;
4065	*extent_dev = bbio->stripes[0].dev;
4066	btrfs_put_bbio(bbio);
4067}