Linux Audio

Check our new training course

Embedded Linux training

Mar 31-Apr 8, 2025
Register
Loading...
v4.6
 
   1/*
   2 * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/blkdev.h>
  20#include <linux/ratelimit.h>
 
 
  21#include "ctree.h"
 
  22#include "volumes.h"
  23#include "disk-io.h"
  24#include "ordered-data.h"
  25#include "transaction.h"
  26#include "backref.h"
  27#include "extent_io.h"
  28#include "dev-replace.h"
  29#include "check-integrity.h"
  30#include "rcu-string.h"
  31#include "raid56.h"
 
 
  32
  33/*
  34 * This is only the first step towards a full-features scrub. It reads all
  35 * extent and super block and verifies the checksums. In case a bad checksum
  36 * is found or the extent cannot be read, good data will be written back if
  37 * any can be found.
  38 *
  39 * Future enhancements:
  40 *  - In case an unrepairable extent is encountered, track which files are
  41 *    affected and report them
  42 *  - track and record media errors, throw out bad devices
  43 *  - add a mode to also read unallocated space
  44 */
  45
  46struct scrub_block;
  47struct scrub_ctx;
  48
  49/*
  50 * the following three values only influence the performance.
  51 * The last one configures the number of parallel and outstanding I/O
  52 * operations. The first two values configure an upper limit for the number
  53 * of (dynamically allocated) pages that are added to a bio.
  54 */
  55#define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
  56#define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
  57#define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
  58
  59/*
  60 * the following value times PAGE_SIZE needs to be large enough to match the
  61 * largest node/leaf/sector size that shall be supported.
  62 * Values larger than BTRFS_STRIPE_LEN are not supported.
  63 */
  64#define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
  65
  66struct scrub_recover {
  67	atomic_t		refs;
  68	struct btrfs_bio	*bbio;
  69	u64			map_length;
  70};
  71
  72struct scrub_page {
  73	struct scrub_block	*sblock;
  74	struct page		*page;
  75	struct btrfs_device	*dev;
  76	struct list_head	list;
  77	u64			flags;  /* extent flags */
  78	u64			generation;
  79	u64			logical;
  80	u64			physical;
  81	u64			physical_for_dev_replace;
  82	atomic_t		refs;
  83	struct {
  84		unsigned int	mirror_num:8;
  85		unsigned int	have_csum:1;
  86		unsigned int	io_error:1;
  87	};
  88	u8			csum[BTRFS_CSUM_SIZE];
  89
  90	struct scrub_recover	*recover;
  91};
  92
  93struct scrub_bio {
  94	int			index;
  95	struct scrub_ctx	*sctx;
  96	struct btrfs_device	*dev;
  97	struct bio		*bio;
  98	int			err;
  99	u64			logical;
 100	u64			physical;
 101#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
 102	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
 103#else
 104	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
 105#endif
 106	int			page_count;
 107	int			next_free;
 108	struct btrfs_work	work;
 109};
 110
 111struct scrub_block {
 112	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 113	int			page_count;
 114	atomic_t		outstanding_pages;
 115	atomic_t		refs; /* free mem on transition to zero */
 116	struct scrub_ctx	*sctx;
 117	struct scrub_parity	*sparity;
 118	struct {
 119		unsigned int	header_error:1;
 120		unsigned int	checksum_error:1;
 121		unsigned int	no_io_error_seen:1;
 122		unsigned int	generation_error:1; /* also sets header_error */
 123
 124		/* The following is for the data used to check parity */
 125		/* It is for the data with checksum */
 126		unsigned int	data_corrected:1;
 127	};
 128	struct btrfs_work	work;
 129};
 130
 131/* Used for the chunks with parity stripe such RAID5/6 */
 132struct scrub_parity {
 133	struct scrub_ctx	*sctx;
 134
 135	struct btrfs_device	*scrub_dev;
 136
 137	u64			logic_start;
 138
 139	u64			logic_end;
 140
 141	int			nsectors;
 142
 143	int			stripe_len;
 144
 145	atomic_t		refs;
 146
 147	struct list_head	spages;
 148
 149	/* Work of parity check and repair */
 150	struct btrfs_work	work;
 151
 152	/* Mark the parity blocks which have data */
 153	unsigned long		*dbitmap;
 154
 155	/*
 156	 * Mark the parity blocks which have data, but errors happen when
 157	 * read data or check data
 158	 */
 159	unsigned long		*ebitmap;
 160
 161	unsigned long		bitmap[0];
 162};
 163
 164struct scrub_wr_ctx {
 165	struct scrub_bio *wr_curr_bio;
 166	struct btrfs_device *tgtdev;
 167	int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 168	atomic_t flush_all_writes;
 169	struct mutex wr_lock;
 170};
 171
 172struct scrub_ctx {
 173	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
 174	struct btrfs_root	*dev_root;
 175	int			first_free;
 176	int			curr;
 177	atomic_t		bios_in_flight;
 178	atomic_t		workers_pending;
 179	spinlock_t		list_lock;
 180	wait_queue_head_t	list_wait;
 181	u16			csum_size;
 182	struct list_head	csum_list;
 183	atomic_t		cancel_req;
 184	int			readonly;
 185	int			pages_per_rd_bio;
 186	u32			sectorsize;
 187	u32			nodesize;
 
 
 188
 189	int			is_dev_replace;
 190	struct scrub_wr_ctx	wr_ctx;
 
 
 
 
 
 
 191
 192	/*
 193	 * statistics
 194	 */
 195	struct btrfs_scrub_progress stat;
 196	spinlock_t		stat_lock;
 197
 198	/*
 199	 * Use a ref counter to avoid use-after-free issues. Scrub workers
 200	 * decrement bios_in_flight and workers_pending and then do a wakeup
 201	 * on the list_wait wait queue. We must ensure the main scrub task
 202	 * doesn't free the scrub context before or while the workers are
 203	 * doing the wakeup() call.
 204	 */
 205	atomic_t                refs;
 206};
 207
 208struct scrub_fixup_nodatasum {
 209	struct scrub_ctx	*sctx;
 210	struct btrfs_device	*dev;
 211	u64			logical;
 212	struct btrfs_root	*root;
 213	struct btrfs_work	work;
 214	int			mirror_num;
 215};
 216
 217struct scrub_nocow_inode {
 218	u64			inum;
 219	u64			offset;
 220	u64			root;
 221	struct list_head	list;
 222};
 223
 224struct scrub_copy_nocow_ctx {
 225	struct scrub_ctx	*sctx;
 226	u64			logical;
 227	u64			len;
 228	int			mirror_num;
 229	u64			physical_for_dev_replace;
 230	struct list_head	inodes;
 231	struct btrfs_work	work;
 232};
 233
 234struct scrub_warning {
 235	struct btrfs_path	*path;
 236	u64			extent_item_size;
 237	const char		*errstr;
 238	sector_t		sector;
 239	u64			logical;
 240	struct btrfs_device	*dev;
 241};
 242
 243static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 244static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 245static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 246static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 247static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 
 
 248static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 249				     struct scrub_block *sblocks_for_recheck);
 250static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 251				struct scrub_block *sblock,
 252				int retry_failed_mirror);
 253static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 254static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 255					     struct scrub_block *sblock_good);
 256static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 257					    struct scrub_block *sblock_good,
 258					    int page_num, int force_write);
 259static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 260static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 261					   int page_num);
 262static int scrub_checksum_data(struct scrub_block *sblock);
 263static int scrub_checksum_tree_block(struct scrub_block *sblock);
 264static int scrub_checksum_super(struct scrub_block *sblock);
 265static void scrub_block_get(struct scrub_block *sblock);
 266static void scrub_block_put(struct scrub_block *sblock);
 267static void scrub_page_get(struct scrub_page *spage);
 268static void scrub_page_put(struct scrub_page *spage);
 269static void scrub_parity_get(struct scrub_parity *sparity);
 270static void scrub_parity_put(struct scrub_parity *sparity);
 271static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 272				    struct scrub_page *spage);
 273static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 274		       u64 physical, struct btrfs_device *dev, u64 flags,
 275		       u64 gen, int mirror_num, u8 *csum, int force,
 276		       u64 physical_for_dev_replace);
 277static void scrub_bio_end_io(struct bio *bio);
 278static void scrub_bio_end_io_worker(struct btrfs_work *work);
 279static void scrub_block_complete(struct scrub_block *sblock);
 280static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 281			       u64 extent_logical, u64 extent_len,
 282			       u64 *extent_physical,
 283			       struct btrfs_device **extent_dev,
 284			       int *extent_mirror_num);
 285static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
 286			      struct scrub_wr_ctx *wr_ctx,
 287			      struct btrfs_fs_info *fs_info,
 288			      struct btrfs_device *dev,
 289			      int is_dev_replace);
 290static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
 291static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 292				    struct scrub_page *spage);
 293static void scrub_wr_submit(struct scrub_ctx *sctx);
 294static void scrub_wr_bio_end_io(struct bio *bio);
 295static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 296static int write_page_nocow(struct scrub_ctx *sctx,
 297			    u64 physical_for_dev_replace, struct page *page);
 298static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 299				      struct scrub_copy_nocow_ctx *ctx);
 300static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 301			    int mirror_num, u64 physical_for_dev_replace);
 302static void copy_nocow_pages_worker(struct btrfs_work *work);
 303static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 304static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 305static void scrub_put_ctx(struct scrub_ctx *sctx);
 306
 
 
 
 
 
 307
 308static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 309{
 310	atomic_inc(&sctx->refs);
 311	atomic_inc(&sctx->bios_in_flight);
 312}
 313
 314static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 315{
 316	atomic_dec(&sctx->bios_in_flight);
 317	wake_up(&sctx->list_wait);
 318	scrub_put_ctx(sctx);
 319}
 320
 321static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 322{
 323	while (atomic_read(&fs_info->scrub_pause_req)) {
 324		mutex_unlock(&fs_info->scrub_lock);
 325		wait_event(fs_info->scrub_pause_wait,
 326		   atomic_read(&fs_info->scrub_pause_req) == 0);
 327		mutex_lock(&fs_info->scrub_lock);
 328	}
 329}
 330
 331static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 332{
 333	atomic_inc(&fs_info->scrubs_paused);
 334	wake_up(&fs_info->scrub_pause_wait);
 335}
 336
 337static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 338{
 339	mutex_lock(&fs_info->scrub_lock);
 340	__scrub_blocked_if_needed(fs_info);
 341	atomic_dec(&fs_info->scrubs_paused);
 342	mutex_unlock(&fs_info->scrub_lock);
 343
 344	wake_up(&fs_info->scrub_pause_wait);
 345}
 346
 347static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 348{
 349	scrub_pause_on(fs_info);
 350	scrub_pause_off(fs_info);
 351}
 352
 353/*
 354 * used for workers that require transaction commits (i.e., for the
 355 * NOCOW case)
 
 
 
 
 
 
 356 */
 357static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 358{
 359	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 360
 361	atomic_inc(&sctx->refs);
 362	/*
 363	 * increment scrubs_running to prevent cancel requests from
 364	 * completing as long as a worker is running. we must also
 365	 * increment scrubs_paused to prevent deadlocking on pause
 366	 * requests used for transactions commits (as the worker uses a
 367	 * transaction context). it is safe to regard the worker
 368	 * as paused for all matters practical. effectively, we only
 369	 * avoid cancellation requests from completing.
 370	 */
 371	mutex_lock(&fs_info->scrub_lock);
 372	atomic_inc(&fs_info->scrubs_running);
 373	atomic_inc(&fs_info->scrubs_paused);
 374	mutex_unlock(&fs_info->scrub_lock);
 375
 376	/*
 377	 * check if @scrubs_running=@scrubs_paused condition
 378	 * inside wait_event() is not an atomic operation.
 379	 * which means we may inc/dec @scrub_running/paused
 380	 * at any time. Let's wake up @scrub_pause_wait as
 381	 * much as we can to let commit transaction blocked less.
 382	 */
 383	wake_up(&fs_info->scrub_pause_wait);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 384
 385	atomic_inc(&sctx->workers_pending);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 386}
 387
 388/* used for workers that require transaction commits */
 389static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 
 
 
 
 
 
 
 
 
 390{
 391	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 
 
 
 
 
 392
 393	/*
 394	 * see scrub_pending_trans_workers_inc() why we're pretending
 395	 * to be paused in the scrub counters
 396	 */
 397	mutex_lock(&fs_info->scrub_lock);
 398	atomic_dec(&fs_info->scrubs_running);
 399	atomic_dec(&fs_info->scrubs_paused);
 400	mutex_unlock(&fs_info->scrub_lock);
 401	atomic_dec(&sctx->workers_pending);
 402	wake_up(&fs_info->scrub_pause_wait);
 403	wake_up(&sctx->list_wait);
 404	scrub_put_ctx(sctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 405}
 406
 407static void scrub_free_csums(struct scrub_ctx *sctx)
 408{
 409	while (!list_empty(&sctx->csum_list)) {
 410		struct btrfs_ordered_sum *sum;
 411		sum = list_first_entry(&sctx->csum_list,
 412				       struct btrfs_ordered_sum, list);
 413		list_del(&sum->list);
 414		kfree(sum);
 415	}
 416}
 417
 418static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 419{
 420	int i;
 421
 422	if (!sctx)
 423		return;
 424
 425	scrub_free_wr_ctx(&sctx->wr_ctx);
 426
 427	/* this can happen when scrub is cancelled */
 428	if (sctx->curr != -1) {
 429		struct scrub_bio *sbio = sctx->bios[sctx->curr];
 430
 431		for (i = 0; i < sbio->page_count; i++) {
 432			WARN_ON(!sbio->pagev[i]->page);
 433			scrub_block_put(sbio->pagev[i]->sblock);
 434		}
 435		bio_put(sbio->bio);
 436	}
 437
 438	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 439		struct scrub_bio *sbio = sctx->bios[i];
 440
 441		if (!sbio)
 442			break;
 443		kfree(sbio);
 444	}
 445
 
 446	scrub_free_csums(sctx);
 447	kfree(sctx);
 448}
 449
 450static void scrub_put_ctx(struct scrub_ctx *sctx)
 451{
 452	if (atomic_dec_and_test(&sctx->refs))
 453		scrub_free_ctx(sctx);
 454}
 455
 456static noinline_for_stack
 457struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 458{
 459	struct scrub_ctx *sctx;
 460	int		i;
 461	struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 462	int ret;
 463
 464	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 465	if (!sctx)
 466		goto nomem;
 467	atomic_set(&sctx->refs, 1);
 468	sctx->is_dev_replace = is_dev_replace;
 469	sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 470	sctx->curr = -1;
 471	sctx->dev_root = dev->dev_root;
 
 472	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 473		struct scrub_bio *sbio;
 474
 475		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 476		if (!sbio)
 477			goto nomem;
 478		sctx->bios[i] = sbio;
 479
 480		sbio->index = i;
 481		sbio->sctx = sctx;
 482		sbio->page_count = 0;
 483		btrfs_init_work(&sbio->work, btrfs_scrub_helper,
 484				scrub_bio_end_io_worker, NULL, NULL);
 485
 486		if (i != SCRUB_BIOS_PER_SCTX - 1)
 487			sctx->bios[i]->next_free = i + 1;
 488		else
 489			sctx->bios[i]->next_free = -1;
 490	}
 491	sctx->first_free = 0;
 492	sctx->nodesize = dev->dev_root->nodesize;
 493	sctx->sectorsize = dev->dev_root->sectorsize;
 494	atomic_set(&sctx->bios_in_flight, 0);
 495	atomic_set(&sctx->workers_pending, 0);
 496	atomic_set(&sctx->cancel_req, 0);
 497	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 498	INIT_LIST_HEAD(&sctx->csum_list);
 499
 500	spin_lock_init(&sctx->list_lock);
 501	spin_lock_init(&sctx->stat_lock);
 502	init_waitqueue_head(&sctx->list_wait);
 
 503
 504	ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 505				 fs_info->dev_replace.tgtdev, is_dev_replace);
 506	if (ret) {
 507		scrub_free_ctx(sctx);
 508		return ERR_PTR(ret);
 
 
 
 509	}
 
 510	return sctx;
 511
 512nomem:
 513	scrub_free_ctx(sctx);
 514	return ERR_PTR(-ENOMEM);
 515}
 516
 517static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 518				     void *warn_ctx)
 519{
 520	u64 isize;
 521	u32 nlink;
 522	int ret;
 523	int i;
 
 524	struct extent_buffer *eb;
 525	struct btrfs_inode_item *inode_item;
 526	struct scrub_warning *swarn = warn_ctx;
 527	struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 528	struct inode_fs_paths *ipath = NULL;
 529	struct btrfs_root *local_root;
 530	struct btrfs_key root_key;
 531	struct btrfs_key key;
 532
 533	root_key.objectid = root;
 534	root_key.type = BTRFS_ROOT_ITEM_KEY;
 535	root_key.offset = (u64)-1;
 536	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 537	if (IS_ERR(local_root)) {
 538		ret = PTR_ERR(local_root);
 539		goto err;
 540	}
 541
 542	/*
 543	 * this makes the path point to (inum INODE_ITEM ioff)
 544	 */
 545	key.objectid = inum;
 546	key.type = BTRFS_INODE_ITEM_KEY;
 547	key.offset = 0;
 548
 549	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 550	if (ret) {
 
 551		btrfs_release_path(swarn->path);
 552		goto err;
 553	}
 554
 555	eb = swarn->path->nodes[0];
 556	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 557					struct btrfs_inode_item);
 558	isize = btrfs_inode_size(eb, inode_item);
 559	nlink = btrfs_inode_nlink(eb, inode_item);
 560	btrfs_release_path(swarn->path);
 561
 
 
 
 
 
 
 562	ipath = init_ipath(4096, local_root, swarn->path);
 
 563	if (IS_ERR(ipath)) {
 
 564		ret = PTR_ERR(ipath);
 565		ipath = NULL;
 566		goto err;
 567	}
 568	ret = paths_from_inode(inum, ipath);
 569
 570	if (ret < 0)
 571		goto err;
 572
 573	/*
 574	 * we deliberately ignore the bit ipath might have been too small to
 575	 * hold all of the paths here
 576	 */
 577	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 578		btrfs_warn_in_rcu(fs_info, "%s at logical %llu on dev "
 579			"%s, sector %llu, root %llu, inode %llu, offset %llu, "
 580			"length %llu, links %u (path: %s)", swarn->errstr,
 581			swarn->logical, rcu_str_deref(swarn->dev->name),
 582			(unsigned long long)swarn->sector, root, inum, offset,
 583			min(isize - offset, (u64)PAGE_SIZE), nlink,
 584			(char *)(unsigned long)ipath->fspath->val[i]);
 
 585
 
 586	free_ipath(ipath);
 587	return 0;
 588
 589err:
 590	btrfs_warn_in_rcu(fs_info, "%s at logical %llu on dev "
 591		"%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 592		"resolving failed with ret=%d", swarn->errstr,
 593		swarn->logical, rcu_str_deref(swarn->dev->name),
 594		(unsigned long long)swarn->sector, root, inum, offset, ret);
 
 595
 596	free_ipath(ipath);
 597	return 0;
 598}
 599
 600static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 601{
 602	struct btrfs_device *dev;
 603	struct btrfs_fs_info *fs_info;
 604	struct btrfs_path *path;
 605	struct btrfs_key found_key;
 606	struct extent_buffer *eb;
 607	struct btrfs_extent_item *ei;
 608	struct scrub_warning swarn;
 609	unsigned long ptr = 0;
 610	u64 extent_item_pos;
 611	u64 flags = 0;
 612	u64 ref_root;
 613	u32 item_size;
 614	u8 ref_level = 0;
 615	int ret;
 616
 617	WARN_ON(sblock->page_count < 1);
 618	dev = sblock->pagev[0]->dev;
 619	fs_info = sblock->sctx->dev_root->fs_info;
 620
 621	path = btrfs_alloc_path();
 622	if (!path)
 623		return;
 624
 625	swarn.sector = (sblock->pagev[0]->physical) >> 9;
 626	swarn.logical = sblock->pagev[0]->logical;
 627	swarn.errstr = errstr;
 628	swarn.dev = NULL;
 629
 630	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 631				  &flags);
 632	if (ret < 0)
 633		goto out;
 634
 635	extent_item_pos = swarn.logical - found_key.objectid;
 636	swarn.extent_item_size = found_key.offset;
 637
 638	eb = path->nodes[0];
 639	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 640	item_size = btrfs_item_size_nr(eb, path->slots[0]);
 641
 642	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 643		do {
 644			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 645						      item_size, &ref_root,
 646						      &ref_level);
 647			btrfs_warn_in_rcu(fs_info,
 648				"%s at logical %llu on dev %s, "
 649				"sector %llu: metadata %s (level %d) in tree "
 650				"%llu", errstr, swarn.logical,
 651				rcu_str_deref(dev->name),
 652				(unsigned long long)swarn.sector,
 653				ref_level ? "node" : "leaf",
 654				ret < 0 ? -1 : ref_level,
 655				ret < 0 ? -1 : ref_root);
 656		} while (ret != 1);
 657		btrfs_release_path(path);
 658	} else {
 659		btrfs_release_path(path);
 660		swarn.path = path;
 661		swarn.dev = dev;
 662		iterate_extent_inodes(fs_info, found_key.objectid,
 663					extent_item_pos, 1,
 664					scrub_print_warning_inode, &swarn);
 665	}
 666
 667out:
 668	btrfs_free_path(path);
 669}
 670
 671static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 672{
 673	struct page *page = NULL;
 674	unsigned long index;
 675	struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 676	int ret;
 677	int corrected = 0;
 678	struct btrfs_key key;
 679	struct inode *inode = NULL;
 680	struct btrfs_fs_info *fs_info;
 681	u64 end = offset + PAGE_SIZE - 1;
 682	struct btrfs_root *local_root;
 683	int srcu_index;
 684
 685	key.objectid = root;
 686	key.type = BTRFS_ROOT_ITEM_KEY;
 687	key.offset = (u64)-1;
 688
 689	fs_info = fixup->root->fs_info;
 690	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 691
 692	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 693	if (IS_ERR(local_root)) {
 694		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 695		return PTR_ERR(local_root);
 696	}
 697
 698	key.type = BTRFS_INODE_ITEM_KEY;
 699	key.objectid = inum;
 700	key.offset = 0;
 701	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 702	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 703	if (IS_ERR(inode))
 704		return PTR_ERR(inode);
 705
 706	index = offset >> PAGE_SHIFT;
 707
 708	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 709	if (!page) {
 710		ret = -ENOMEM;
 711		goto out;
 712	}
 713
 714	if (PageUptodate(page)) {
 715		if (PageDirty(page)) {
 716			/*
 717			 * we need to write the data to the defect sector. the
 718			 * data that was in that sector is not in memory,
 719			 * because the page was modified. we must not write the
 720			 * modified page to that sector.
 721			 *
 722			 * TODO: what could be done here: wait for the delalloc
 723			 *       runner to write out that page (might involve
 724			 *       COW) and see whether the sector is still
 725			 *       referenced afterwards.
 726			 *
 727			 * For the meantime, we'll treat this error
 728			 * incorrectable, although there is a chance that a
 729			 * later scrub will find the bad sector again and that
 730			 * there's no dirty page in memory, then.
 731			 */
 732			ret = -EIO;
 733			goto out;
 734		}
 735		ret = repair_io_failure(inode, offset, PAGE_SIZE,
 736					fixup->logical, page,
 737					offset - page_offset(page),
 738					fixup->mirror_num);
 739		unlock_page(page);
 740		corrected = !ret;
 741	} else {
 742		/*
 743		 * we need to get good data first. the general readpage path
 744		 * will call repair_io_failure for us, we just have to make
 745		 * sure we read the bad mirror.
 746		 */
 747		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 748					EXTENT_DAMAGED, GFP_NOFS);
 749		if (ret) {
 750			/* set_extent_bits should give proper error */
 751			WARN_ON(ret > 0);
 752			if (ret > 0)
 753				ret = -EFAULT;
 754			goto out;
 755		}
 756
 757		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 758						btrfs_get_extent,
 759						fixup->mirror_num);
 760		wait_on_page_locked(page);
 761
 762		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 763						end, EXTENT_DAMAGED, 0, NULL);
 764		if (!corrected)
 765			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 766						EXTENT_DAMAGED, GFP_NOFS);
 767	}
 768
 769out:
 770	if (page)
 771		put_page(page);
 772
 773	iput(inode);
 774
 775	if (ret < 0)
 776		return ret;
 777
 778	if (ret == 0 && corrected) {
 779		/*
 780		 * we only need to call readpage for one of the inodes belonging
 781		 * to this extent. so make iterate_extent_inodes stop
 782		 */
 783		return 1;
 784	}
 785
 786	return -EIO;
 787}
 788
 789static void scrub_fixup_nodatasum(struct btrfs_work *work)
 790{
 791	int ret;
 792	struct scrub_fixup_nodatasum *fixup;
 793	struct scrub_ctx *sctx;
 794	struct btrfs_trans_handle *trans = NULL;
 795	struct btrfs_path *path;
 796	int uncorrectable = 0;
 797
 798	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 799	sctx = fixup->sctx;
 800
 801	path = btrfs_alloc_path();
 802	if (!path) {
 803		spin_lock(&sctx->stat_lock);
 804		++sctx->stat.malloc_errors;
 805		spin_unlock(&sctx->stat_lock);
 806		uncorrectable = 1;
 807		goto out;
 808	}
 809
 810	trans = btrfs_join_transaction(fixup->root);
 811	if (IS_ERR(trans)) {
 812		uncorrectable = 1;
 813		goto out;
 814	}
 815
 816	/*
 817	 * the idea is to trigger a regular read through the standard path. we
 818	 * read a page from the (failed) logical address by specifying the
 819	 * corresponding copynum of the failed sector. thus, that readpage is
 820	 * expected to fail.
 821	 * that is the point where on-the-fly error correction will kick in
 822	 * (once it's finished) and rewrite the failed sector if a good copy
 823	 * can be found.
 824	 */
 825	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 826						path, scrub_fixup_readpage,
 827						fixup);
 828	if (ret < 0) {
 829		uncorrectable = 1;
 830		goto out;
 831	}
 832	WARN_ON(ret != 1);
 833
 834	spin_lock(&sctx->stat_lock);
 835	++sctx->stat.corrected_errors;
 836	spin_unlock(&sctx->stat_lock);
 837
 838out:
 839	if (trans && !IS_ERR(trans))
 840		btrfs_end_transaction(trans, fixup->root);
 841	if (uncorrectable) {
 842		spin_lock(&sctx->stat_lock);
 843		++sctx->stat.uncorrectable_errors;
 844		spin_unlock(&sctx->stat_lock);
 845		btrfs_dev_replace_stats_inc(
 846			&sctx->dev_root->fs_info->dev_replace.
 847			num_uncorrectable_read_errors);
 848		btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
 849		    "unable to fixup (nodatasum) error at logical %llu on dev %s",
 850			fixup->logical, rcu_str_deref(fixup->dev->name));
 851	}
 852
 853	btrfs_free_path(path);
 854	kfree(fixup);
 855
 856	scrub_pending_trans_workers_dec(sctx);
 857}
 858
 859static inline void scrub_get_recover(struct scrub_recover *recover)
 860{
 861	atomic_inc(&recover->refs);
 862}
 863
 864static inline void scrub_put_recover(struct scrub_recover *recover)
 
 865{
 866	if (atomic_dec_and_test(&recover->refs)) {
 
 867		btrfs_put_bbio(recover->bbio);
 868		kfree(recover);
 869	}
 870}
 871
 872/*
 873 * scrub_handle_errored_block gets called when either verification of the
 874 * pages failed or the bio failed to read, e.g. with EIO. In the latter
 875 * case, this function handles all pages in the bio, even though only one
 876 * may be bad.
 877 * The goal of this function is to repair the errored block by using the
 878 * contents of one of the mirrors.
 879 */
 880static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 881{
 882	struct scrub_ctx *sctx = sblock_to_check->sctx;
 883	struct btrfs_device *dev;
 884	struct btrfs_fs_info *fs_info;
 885	u64 length;
 886	u64 logical;
 887	unsigned int failed_mirror_index;
 888	unsigned int is_metadata;
 889	unsigned int have_csum;
 890	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 891	struct scrub_block *sblock_bad;
 892	int ret;
 893	int mirror_index;
 894	int page_num;
 895	int success;
 896	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 
 
 897				      DEFAULT_RATELIMIT_BURST);
 898
 899	BUG_ON(sblock_to_check->page_count < 1);
 900	fs_info = sctx->dev_root->fs_info;
 901	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 902		/*
 903		 * if we find an error in a super block, we just report it.
 904		 * They will get written with the next transaction commit
 905		 * anyway
 906		 */
 907		spin_lock(&sctx->stat_lock);
 908		++sctx->stat.super_errors;
 909		spin_unlock(&sctx->stat_lock);
 910		return 0;
 911	}
 912	length = sblock_to_check->page_count * PAGE_SIZE;
 913	logical = sblock_to_check->pagev[0]->logical;
 914	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 915	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 916	is_metadata = !(sblock_to_check->pagev[0]->flags &
 917			BTRFS_EXTENT_FLAG_DATA);
 918	have_csum = sblock_to_check->pagev[0]->have_csum;
 919	dev = sblock_to_check->pagev[0]->dev;
 920
 921	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 922		sblocks_for_recheck = NULL;
 923		goto nodatasum_case;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 924	}
 925
 926	/*
 927	 * read all mirrors one after the other. This includes to
 928	 * re-read the extent or metadata block that failed (that was
 929	 * the cause that this fixup code is called) another time,
 930	 * page by page this time in order to know which pages
 931	 * caused I/O errors and which ones are good (for all mirrors).
 932	 * It is the goal to handle the situation when more than one
 933	 * mirror contains I/O errors, but the errors do not
 934	 * overlap, i.e. the data can be repaired by selecting the
 935	 * pages from those mirrors without I/O error on the
 936	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 937	 * would be that mirror #1 has an I/O error on the first page,
 938	 * the second page is good, and mirror #2 has an I/O error on
 939	 * the second page, but the first page is good.
 940	 * Then the first page of the first mirror can be repaired by
 941	 * taking the first page of the second mirror, and the
 942	 * second page of the second mirror can be repaired by
 943	 * copying the contents of the 2nd page of the 1st mirror.
 944	 * One more note: if the pages of one mirror contain I/O
 945	 * errors, the checksum cannot be verified. In order to get
 946	 * the best data for repairing, the first attempt is to find
 947	 * a mirror without I/O errors and with a validated checksum.
 948	 * Only if this is not possible, the pages are picked from
 949	 * mirrors with I/O errors without considering the checksum.
 950	 * If the latter is the case, at the end, the checksum of the
 951	 * repaired area is verified in order to correctly maintain
 952	 * the statistics.
 953	 */
 954
 955	sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 956				      sizeof(*sblocks_for_recheck), GFP_NOFS);
 957	if (!sblocks_for_recheck) {
 958		spin_lock(&sctx->stat_lock);
 959		sctx->stat.malloc_errors++;
 960		sctx->stat.read_errors++;
 961		sctx->stat.uncorrectable_errors++;
 962		spin_unlock(&sctx->stat_lock);
 963		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 964		goto out;
 965	}
 966
 967	/* setup the context, map the logical blocks and alloc the pages */
 968	ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 969	if (ret) {
 970		spin_lock(&sctx->stat_lock);
 971		sctx->stat.read_errors++;
 972		sctx->stat.uncorrectable_errors++;
 973		spin_unlock(&sctx->stat_lock);
 974		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 975		goto out;
 976	}
 977	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 978	sblock_bad = sblocks_for_recheck + failed_mirror_index;
 979
 980	/* build and submit the bios for the failed mirror, check checksums */
 981	scrub_recheck_block(fs_info, sblock_bad, 1);
 982
 983	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 984	    sblock_bad->no_io_error_seen) {
 985		/*
 986		 * the error disappeared after reading page by page, or
 987		 * the area was part of a huge bio and other parts of the
 988		 * bio caused I/O errors, or the block layer merged several
 989		 * read requests into one and the error is caused by a
 990		 * different bio (usually one of the two latter cases is
 991		 * the cause)
 992		 */
 993		spin_lock(&sctx->stat_lock);
 994		sctx->stat.unverified_errors++;
 995		sblock_to_check->data_corrected = 1;
 996		spin_unlock(&sctx->stat_lock);
 997
 998		if (sctx->is_dev_replace)
 999			scrub_write_block_to_dev_replace(sblock_bad);
1000		goto out;
1001	}
1002
1003	if (!sblock_bad->no_io_error_seen) {
1004		spin_lock(&sctx->stat_lock);
1005		sctx->stat.read_errors++;
1006		spin_unlock(&sctx->stat_lock);
1007		if (__ratelimit(&_rs))
1008			scrub_print_warning("i/o error", sblock_to_check);
1009		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1010	} else if (sblock_bad->checksum_error) {
1011		spin_lock(&sctx->stat_lock);
1012		sctx->stat.csum_errors++;
1013		spin_unlock(&sctx->stat_lock);
1014		if (__ratelimit(&_rs))
1015			scrub_print_warning("checksum error", sblock_to_check);
1016		btrfs_dev_stat_inc_and_print(dev,
1017					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
1018	} else if (sblock_bad->header_error) {
1019		spin_lock(&sctx->stat_lock);
1020		sctx->stat.verify_errors++;
1021		spin_unlock(&sctx->stat_lock);
1022		if (__ratelimit(&_rs))
1023			scrub_print_warning("checksum/header error",
1024					    sblock_to_check);
1025		if (sblock_bad->generation_error)
1026			btrfs_dev_stat_inc_and_print(dev,
1027				BTRFS_DEV_STAT_GENERATION_ERRS);
1028		else
1029			btrfs_dev_stat_inc_and_print(dev,
1030				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1031	}
1032
1033	if (sctx->readonly) {
1034		ASSERT(!sctx->is_dev_replace);
1035		goto out;
1036	}
1037
1038	if (!is_metadata && !have_csum) {
1039		struct scrub_fixup_nodatasum *fixup_nodatasum;
1040
1041		WARN_ON(sctx->is_dev_replace);
1042
1043nodatasum_case:
1044
1045		/*
1046		 * !is_metadata and !have_csum, this means that the data
1047		 * might not be COW'ed, that it might be modified
1048		 * concurrently. The general strategy to work on the
1049		 * commit root does not help in the case when COW is not
1050		 * used.
1051		 */
1052		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1053		if (!fixup_nodatasum)
1054			goto did_not_correct_error;
1055		fixup_nodatasum->sctx = sctx;
1056		fixup_nodatasum->dev = dev;
1057		fixup_nodatasum->logical = logical;
1058		fixup_nodatasum->root = fs_info->extent_root;
1059		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1060		scrub_pending_trans_workers_inc(sctx);
1061		btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1062				scrub_fixup_nodatasum, NULL, NULL);
1063		btrfs_queue_work(fs_info->scrub_workers,
1064				 &fixup_nodatasum->work);
1065		goto out;
1066	}
1067
1068	/*
1069	 * now build and submit the bios for the other mirrors, check
1070	 * checksums.
1071	 * First try to pick the mirror which is completely without I/O
1072	 * errors and also does not have a checksum error.
1073	 * If one is found, and if a checksum is present, the full block
1074	 * that is known to contain an error is rewritten. Afterwards
1075	 * the block is known to be corrected.
1076	 * If a mirror is found which is completely correct, and no
1077	 * checksum is present, only those pages are rewritten that had
1078	 * an I/O error in the block to be repaired, since it cannot be
1079	 * determined, which copy of the other pages is better (and it
1080	 * could happen otherwise that a correct page would be
1081	 * overwritten by a bad one).
1082	 */
1083	for (mirror_index = 0;
1084	     mirror_index < BTRFS_MAX_MIRRORS &&
1085	     sblocks_for_recheck[mirror_index].page_count > 0;
1086	     mirror_index++) {
1087		struct scrub_block *sblock_other;
1088
1089		if (mirror_index == failed_mirror_index)
1090			continue;
1091		sblock_other = sblocks_for_recheck + mirror_index;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1092
1093		/* build and submit the bios, check checksums */
1094		scrub_recheck_block(fs_info, sblock_other, 0);
1095
1096		if (!sblock_other->header_error &&
1097		    !sblock_other->checksum_error &&
1098		    sblock_other->no_io_error_seen) {
1099			if (sctx->is_dev_replace) {
1100				scrub_write_block_to_dev_replace(sblock_other);
1101				goto corrected_error;
1102			} else {
1103				ret = scrub_repair_block_from_good_copy(
1104						sblock_bad, sblock_other);
1105				if (!ret)
1106					goto corrected_error;
1107			}
1108		}
1109	}
1110
1111	if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1112		goto did_not_correct_error;
1113
1114	/*
1115	 * In case of I/O errors in the area that is supposed to be
1116	 * repaired, continue by picking good copies of those pages.
1117	 * Select the good pages from mirrors to rewrite bad pages from
1118	 * the area to fix. Afterwards verify the checksum of the block
1119	 * that is supposed to be repaired. This verification step is
1120	 * only done for the purpose of statistic counting and for the
1121	 * final scrub report, whether errors remain.
1122	 * A perfect algorithm could make use of the checksum and try
1123	 * all possible combinations of pages from the different mirrors
1124	 * until the checksum verification succeeds. For example, when
1125	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1126	 * of mirror #2 is readable but the final checksum test fails,
1127	 * then the 2nd page of mirror #3 could be tried, whether now
1128	 * the final checksum succeedes. But this would be a rare
1129	 * exception and is therefore not implemented. At least it is
1130	 * avoided that the good copy is overwritten.
1131	 * A more useful improvement would be to pick the sectors
1132	 * without I/O error based on sector sizes (512 bytes on legacy
1133	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1134	 * mirror could be repaired by taking 512 byte of a different
1135	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1136	 * area are unreadable.
1137	 */
1138	success = 1;
1139	for (page_num = 0; page_num < sblock_bad->page_count;
1140	     page_num++) {
1141		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1142		struct scrub_block *sblock_other = NULL;
1143
1144		/* skip no-io-error page in scrub */
1145		if (!page_bad->io_error && !sctx->is_dev_replace)
1146			continue;
1147
1148		/* try to find no-io-error page in mirrors */
1149		if (page_bad->io_error) {
 
 
 
 
 
 
 
 
 
1150			for (mirror_index = 0;
1151			     mirror_index < BTRFS_MAX_MIRRORS &&
1152			     sblocks_for_recheck[mirror_index].page_count > 0;
1153			     mirror_index++) {
1154				if (!sblocks_for_recheck[mirror_index].
1155				    pagev[page_num]->io_error) {
1156					sblock_other = sblocks_for_recheck +
1157						       mirror_index;
1158					break;
1159				}
1160			}
1161			if (!sblock_other)
1162				success = 0;
1163		}
1164
1165		if (sctx->is_dev_replace) {
1166			/*
1167			 * did not find a mirror to fetch the page
1168			 * from. scrub_write_page_to_dev_replace()
1169			 * handles this case (page->io_error), by
1170			 * filling the block with zeros before
1171			 * submitting the write request
1172			 */
1173			if (!sblock_other)
1174				sblock_other = sblock_bad;
1175
1176			if (scrub_write_page_to_dev_replace(sblock_other,
1177							    page_num) != 0) {
1178				btrfs_dev_replace_stats_inc(
1179					&sctx->dev_root->
1180					fs_info->dev_replace.
1181					num_write_errors);
1182				success = 0;
1183			}
1184		} else if (sblock_other) {
1185			ret = scrub_repair_page_from_good_copy(sblock_bad,
1186							       sblock_other,
1187							       page_num, 0);
1188			if (0 == ret)
1189				page_bad->io_error = 0;
1190			else
1191				success = 0;
1192		}
1193	}
1194
1195	if (success && !sctx->is_dev_replace) {
1196		if (is_metadata || have_csum) {
1197			/*
1198			 * need to verify the checksum now that all
1199			 * sectors on disk are repaired (the write
1200			 * request for data to be repaired is on its way).
1201			 * Just be lazy and use scrub_recheck_block()
1202			 * which re-reads the data before the checksum
1203			 * is verified, but most likely the data comes out
1204			 * of the page cache.
1205			 */
1206			scrub_recheck_block(fs_info, sblock_bad, 1);
1207			if (!sblock_bad->header_error &&
1208			    !sblock_bad->checksum_error &&
1209			    sblock_bad->no_io_error_seen)
1210				goto corrected_error;
1211			else
1212				goto did_not_correct_error;
1213		} else {
1214corrected_error:
1215			spin_lock(&sctx->stat_lock);
1216			sctx->stat.corrected_errors++;
1217			sblock_to_check->data_corrected = 1;
1218			spin_unlock(&sctx->stat_lock);
1219			btrfs_err_rl_in_rcu(fs_info,
1220				"fixed up error at logical %llu on dev %s",
1221				logical, rcu_str_deref(dev->name));
1222		}
1223	} else {
1224did_not_correct_error:
1225		spin_lock(&sctx->stat_lock);
1226		sctx->stat.uncorrectable_errors++;
1227		spin_unlock(&sctx->stat_lock);
1228		btrfs_err_rl_in_rcu(fs_info,
1229			"unable to fixup (regular) error at logical %llu on dev %s",
1230			logical, rcu_str_deref(dev->name));
1231	}
1232
1233out:
1234	if (sblocks_for_recheck) {
1235		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1236		     mirror_index++) {
1237			struct scrub_block *sblock = sblocks_for_recheck +
1238						     mirror_index;
1239			struct scrub_recover *recover;
1240			int page_index;
1241
1242			for (page_index = 0; page_index < sblock->page_count;
1243			     page_index++) {
1244				sblock->pagev[page_index]->sblock = NULL;
1245				recover = sblock->pagev[page_index]->recover;
1246				if (recover) {
1247					scrub_put_recover(recover);
1248					sblock->pagev[page_index]->recover =
1249									NULL;
1250				}
1251				scrub_page_put(sblock->pagev[page_index]);
1252			}
1253		}
1254		kfree(sblocks_for_recheck);
1255	}
1256
 
 
 
 
1257	return 0;
1258}
1259
1260static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1261{
1262	if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1263		return 2;
1264	else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1265		return 3;
1266	else
1267		return (int)bbio->num_stripes;
1268}
1269
1270static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1271						 u64 *raid_map,
1272						 u64 mapped_length,
1273						 int nstripes, int mirror,
1274						 int *stripe_index,
1275						 u64 *stripe_offset)
1276{
1277	int i;
1278
1279	if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1280		/* RAID5/6 */
1281		for (i = 0; i < nstripes; i++) {
1282			if (raid_map[i] == RAID6_Q_STRIPE ||
1283			    raid_map[i] == RAID5_P_STRIPE)
1284				continue;
1285
1286			if (logical >= raid_map[i] &&
1287			    logical < raid_map[i] + mapped_length)
1288				break;
1289		}
1290
1291		*stripe_index = i;
1292		*stripe_offset = logical - raid_map[i];
1293	} else {
1294		/* The other RAID type */
1295		*stripe_index = mirror;
1296		*stripe_offset = 0;
1297	}
1298}
1299
1300static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1301				     struct scrub_block *sblocks_for_recheck)
1302{
1303	struct scrub_ctx *sctx = original_sblock->sctx;
1304	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
1305	u64 length = original_sblock->page_count * PAGE_SIZE;
1306	u64 logical = original_sblock->pagev[0]->logical;
1307	u64 generation = original_sblock->pagev[0]->generation;
1308	u64 flags = original_sblock->pagev[0]->flags;
1309	u64 have_csum = original_sblock->pagev[0]->have_csum;
1310	struct scrub_recover *recover;
1311	struct btrfs_bio *bbio;
1312	u64 sublen;
1313	u64 mapped_length;
1314	u64 stripe_offset;
1315	int stripe_index;
1316	int page_index = 0;
1317	int mirror_index;
1318	int nmirrors;
1319	int ret;
1320
1321	/*
1322	 * note: the two members refs and outstanding_pages
1323	 * are not used (and not set) in the blocks that are used for
1324	 * the recheck procedure
1325	 */
1326
1327	while (length > 0) {
1328		sublen = min_t(u64, length, PAGE_SIZE);
1329		mapped_length = sublen;
1330		bbio = NULL;
1331
1332		/*
1333		 * with a length of PAGE_SIZE, each returned stripe
1334		 * represents one mirror
1335		 */
1336		ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical,
1337				       &mapped_length, &bbio, 0, 1);
 
1338		if (ret || !bbio || mapped_length < sublen) {
1339			btrfs_put_bbio(bbio);
 
1340			return -EIO;
1341		}
1342
1343		recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1344		if (!recover) {
1345			btrfs_put_bbio(bbio);
 
1346			return -ENOMEM;
1347		}
1348
1349		atomic_set(&recover->refs, 1);
1350		recover->bbio = bbio;
1351		recover->map_length = mapped_length;
1352
1353		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1354
1355		nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1356
1357		for (mirror_index = 0; mirror_index < nmirrors;
1358		     mirror_index++) {
1359			struct scrub_block *sblock;
1360			struct scrub_page *page;
1361
1362			sblock = sblocks_for_recheck + mirror_index;
1363			sblock->sctx = sctx;
1364
1365			page = kzalloc(sizeof(*page), GFP_NOFS);
1366			if (!page) {
1367leave_nomem:
1368				spin_lock(&sctx->stat_lock);
1369				sctx->stat.malloc_errors++;
1370				spin_unlock(&sctx->stat_lock);
1371				scrub_put_recover(recover);
1372				return -ENOMEM;
1373			}
1374			scrub_page_get(page);
1375			sblock->pagev[page_index] = page;
1376			page->sblock = sblock;
1377			page->flags = flags;
1378			page->generation = generation;
1379			page->logical = logical;
1380			page->have_csum = have_csum;
1381			if (have_csum)
1382				memcpy(page->csum,
1383				       original_sblock->pagev[0]->csum,
1384				       sctx->csum_size);
1385
1386			scrub_stripe_index_and_offset(logical,
1387						      bbio->map_type,
1388						      bbio->raid_map,
1389						      mapped_length,
1390						      bbio->num_stripes -
1391						      bbio->num_tgtdevs,
1392						      mirror_index,
1393						      &stripe_index,
1394						      &stripe_offset);
1395			page->physical = bbio->stripes[stripe_index].physical +
1396					 stripe_offset;
1397			page->dev = bbio->stripes[stripe_index].dev;
1398
1399			BUG_ON(page_index >= original_sblock->page_count);
1400			page->physical_for_dev_replace =
1401				original_sblock->pagev[page_index]->
1402				physical_for_dev_replace;
1403			/* for missing devices, dev->bdev is NULL */
1404			page->mirror_num = mirror_index + 1;
1405			sblock->page_count++;
1406			page->page = alloc_page(GFP_NOFS);
1407			if (!page->page)
1408				goto leave_nomem;
1409
1410			scrub_get_recover(recover);
1411			page->recover = recover;
1412		}
1413		scrub_put_recover(recover);
1414		length -= sublen;
1415		logical += sublen;
1416		page_index++;
1417	}
1418
1419	return 0;
1420}
1421
1422struct scrub_bio_ret {
1423	struct completion event;
1424	int error;
1425};
1426
1427static void scrub_bio_wait_endio(struct bio *bio)
1428{
1429	struct scrub_bio_ret *ret = bio->bi_private;
1430
1431	ret->error = bio->bi_error;
1432	complete(&ret->event);
1433}
1434
1435static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1436{
1437	return page->recover &&
1438	       (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1439}
1440
1441static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1442					struct bio *bio,
1443					struct scrub_page *page)
1444{
1445	struct scrub_bio_ret done;
1446	int ret;
 
1447
1448	init_completion(&done.event);
1449	done.error = 0;
1450	bio->bi_iter.bi_sector = page->logical >> 9;
1451	bio->bi_private = &done;
1452	bio->bi_end_io = scrub_bio_wait_endio;
1453
1454	ret = raid56_parity_recover(fs_info->fs_root, bio, page->recover->bbio,
1455				    page->recover->map_length,
1456				    page->mirror_num, 0);
 
1457	if (ret)
1458		return ret;
1459
1460	wait_for_completion(&done.event);
1461	if (done.error)
1462		return -EIO;
1463
1464	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1465}
1466
1467/*
1468 * this function will check the on disk data for checksum errors, header
1469 * errors and read I/O errors. If any I/O errors happen, the exact pages
1470 * which are errored are marked as being bad. The goal is to enable scrub
1471 * to take those pages that are not errored from all the mirrors so that
1472 * the pages that are errored in the just handled mirror can be repaired.
1473 */
1474static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1475				struct scrub_block *sblock,
1476				int retry_failed_mirror)
1477{
1478	int page_num;
1479
1480	sblock->no_io_error_seen = 1;
1481
 
 
 
 
1482	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1483		struct bio *bio;
1484		struct scrub_page *page = sblock->pagev[page_num];
1485
1486		if (page->dev->bdev == NULL) {
1487			page->io_error = 1;
1488			sblock->no_io_error_seen = 0;
1489			continue;
1490		}
1491
1492		WARN_ON(!page->page);
1493		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1494		if (!bio) {
1495			page->io_error = 1;
1496			sblock->no_io_error_seen = 0;
1497			continue;
1498		}
1499		bio->bi_bdev = page->dev->bdev;
1500
1501		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1502		if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1503			if (scrub_submit_raid56_bio_wait(fs_info, bio, page))
1504				sblock->no_io_error_seen = 0;
1505		} else {
1506			bio->bi_iter.bi_sector = page->physical >> 9;
 
1507
1508			if (btrfsic_submit_bio_wait(READ, bio))
1509				sblock->no_io_error_seen = 0;
 
1510		}
1511
1512		bio_put(bio);
1513	}
1514
1515	if (sblock->no_io_error_seen)
1516		scrub_recheck_block_checksum(sblock);
1517}
1518
1519static inline int scrub_check_fsid(u8 fsid[],
1520				   struct scrub_page *spage)
1521{
1522	struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1523	int ret;
1524
1525	ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1526	return !ret;
1527}
1528
1529static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1530{
1531	sblock->header_error = 0;
1532	sblock->checksum_error = 0;
1533	sblock->generation_error = 0;
1534
1535	if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1536		scrub_checksum_data(sblock);
1537	else
1538		scrub_checksum_tree_block(sblock);
1539}
1540
1541static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1542					     struct scrub_block *sblock_good)
1543{
1544	int page_num;
1545	int ret = 0;
1546
1547	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1548		int ret_sub;
1549
1550		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1551							   sblock_good,
1552							   page_num, 1);
1553		if (ret_sub)
1554			ret = ret_sub;
1555	}
1556
1557	return ret;
1558}
1559
1560static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1561					    struct scrub_block *sblock_good,
1562					    int page_num, int force_write)
1563{
1564	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1565	struct scrub_page *page_good = sblock_good->pagev[page_num];
 
 
1566
1567	BUG_ON(page_bad->page == NULL);
1568	BUG_ON(page_good->page == NULL);
1569	if (force_write || sblock_bad->header_error ||
1570	    sblock_bad->checksum_error || page_bad->io_error) {
1571		struct bio *bio;
1572		int ret;
1573
1574		if (!page_bad->dev->bdev) {
1575			btrfs_warn_rl(sblock_bad->sctx->dev_root->fs_info,
1576				"scrub_repair_page_from_good_copy(bdev == NULL) "
1577				"is unexpected");
1578			return -EIO;
1579		}
1580
1581		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1582		if (!bio)
1583			return -EIO;
1584		bio->bi_bdev = page_bad->dev->bdev;
1585		bio->bi_iter.bi_sector = page_bad->physical >> 9;
1586
1587		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1588		if (PAGE_SIZE != ret) {
1589			bio_put(bio);
1590			return -EIO;
1591		}
1592
1593		if (btrfsic_submit_bio_wait(WRITE, bio)) {
1594			btrfs_dev_stat_inc_and_print(page_bad->dev,
1595				BTRFS_DEV_STAT_WRITE_ERRS);
1596			btrfs_dev_replace_stats_inc(
1597				&sblock_bad->sctx->dev_root->fs_info->
1598				dev_replace.num_write_errors);
1599			bio_put(bio);
1600			return -EIO;
1601		}
1602		bio_put(bio);
1603	}
1604
1605	return 0;
1606}
1607
1608static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1609{
 
1610	int page_num;
1611
1612	/*
1613	 * This block is used for the check of the parity on the source device,
1614	 * so the data needn't be written into the destination device.
1615	 */
1616	if (sblock->sparity)
1617		return;
1618
1619	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1620		int ret;
1621
1622		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1623		if (ret)
1624			btrfs_dev_replace_stats_inc(
1625				&sblock->sctx->dev_root->fs_info->dev_replace.
1626				num_write_errors);
1627	}
1628}
1629
1630static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1631					   int page_num)
1632{
1633	struct scrub_page *spage = sblock->pagev[page_num];
1634
1635	BUG_ON(spage->page == NULL);
1636	if (spage->io_error) {
1637		void *mapped_buffer = kmap_atomic(spage->page);
1638
1639		memset(mapped_buffer, 0, PAGE_SIZE);
1640		flush_dcache_page(spage->page);
1641		kunmap_atomic(mapped_buffer);
1642	}
1643	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1644}
1645
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1646static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1647				    struct scrub_page *spage)
1648{
1649	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1650	struct scrub_bio *sbio;
1651	int ret;
 
1652
1653	mutex_lock(&wr_ctx->wr_lock);
1654again:
1655	if (!wr_ctx->wr_curr_bio) {
1656		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1657					      GFP_KERNEL);
1658		if (!wr_ctx->wr_curr_bio) {
1659			mutex_unlock(&wr_ctx->wr_lock);
1660			return -ENOMEM;
1661		}
1662		wr_ctx->wr_curr_bio->sctx = sctx;
1663		wr_ctx->wr_curr_bio->page_count = 0;
1664	}
1665	sbio = wr_ctx->wr_curr_bio;
1666	if (sbio->page_count == 0) {
1667		struct bio *bio;
1668
 
 
 
 
 
 
 
1669		sbio->physical = spage->physical_for_dev_replace;
1670		sbio->logical = spage->logical;
1671		sbio->dev = wr_ctx->tgtdev;
1672		bio = sbio->bio;
1673		if (!bio) {
1674			bio = btrfs_io_bio_alloc(GFP_KERNEL,
1675					wr_ctx->pages_per_wr_bio);
1676			if (!bio) {
1677				mutex_unlock(&wr_ctx->wr_lock);
1678				return -ENOMEM;
1679			}
1680			sbio->bio = bio;
1681		}
1682
1683		bio->bi_private = sbio;
1684		bio->bi_end_io = scrub_wr_bio_end_io;
1685		bio->bi_bdev = sbio->dev->bdev;
1686		bio->bi_iter.bi_sector = sbio->physical >> 9;
1687		sbio->err = 0;
1688	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
 
1689		   spage->physical_for_dev_replace ||
1690		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1691		   spage->logical) {
1692		scrub_wr_submit(sctx);
1693		goto again;
1694	}
1695
1696	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1697	if (ret != PAGE_SIZE) {
1698		if (sbio->page_count < 1) {
1699			bio_put(sbio->bio);
1700			sbio->bio = NULL;
1701			mutex_unlock(&wr_ctx->wr_lock);
1702			return -EIO;
1703		}
1704		scrub_wr_submit(sctx);
1705		goto again;
1706	}
1707
1708	sbio->pagev[sbio->page_count] = spage;
1709	scrub_page_get(spage);
1710	sbio->page_count++;
1711	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1712		scrub_wr_submit(sctx);
1713	mutex_unlock(&wr_ctx->wr_lock);
1714
1715	return 0;
1716}
1717
1718static void scrub_wr_submit(struct scrub_ctx *sctx)
1719{
1720	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1721	struct scrub_bio *sbio;
1722
1723	if (!wr_ctx->wr_curr_bio)
1724		return;
1725
1726	sbio = wr_ctx->wr_curr_bio;
1727	wr_ctx->wr_curr_bio = NULL;
1728	WARN_ON(!sbio->bio->bi_bdev);
1729	scrub_pending_bio_inc(sctx);
1730	/* process all writes in a single worker thread. Then the block layer
1731	 * orders the requests before sending them to the driver which
1732	 * doubled the write performance on spinning disks when measured
1733	 * with Linux 3.5 */
1734	btrfsic_submit_bio(WRITE, sbio->bio);
 
 
 
 
1735}
1736
1737static void scrub_wr_bio_end_io(struct bio *bio)
1738{
1739	struct scrub_bio *sbio = bio->bi_private;
1740	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1741
1742	sbio->err = bio->bi_error;
1743	sbio->bio = bio;
1744
1745	btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1746			 scrub_wr_bio_end_io_worker, NULL, NULL);
1747	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1748}
1749
1750static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1751{
1752	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1753	struct scrub_ctx *sctx = sbio->sctx;
1754	int i;
1755
1756	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1757	if (sbio->err) {
1758		struct btrfs_dev_replace *dev_replace =
1759			&sbio->sctx->dev_root->fs_info->dev_replace;
1760
1761		for (i = 0; i < sbio->page_count; i++) {
1762			struct scrub_page *spage = sbio->pagev[i];
1763
1764			spage->io_error = 1;
1765			btrfs_dev_replace_stats_inc(&dev_replace->
1766						    num_write_errors);
1767		}
1768	}
1769
1770	for (i = 0; i < sbio->page_count; i++)
1771		scrub_page_put(sbio->pagev[i]);
1772
1773	bio_put(sbio->bio);
1774	kfree(sbio);
1775	scrub_pending_bio_dec(sctx);
1776}
1777
1778static int scrub_checksum(struct scrub_block *sblock)
1779{
1780	u64 flags;
1781	int ret;
1782
1783	/*
1784	 * No need to initialize these stats currently,
1785	 * because this function only use return value
1786	 * instead of these stats value.
1787	 *
1788	 * Todo:
1789	 * always use stats
1790	 */
1791	sblock->header_error = 0;
1792	sblock->generation_error = 0;
1793	sblock->checksum_error = 0;
1794
1795	WARN_ON(sblock->page_count < 1);
1796	flags = sblock->pagev[0]->flags;
1797	ret = 0;
1798	if (flags & BTRFS_EXTENT_FLAG_DATA)
1799		ret = scrub_checksum_data(sblock);
1800	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1801		ret = scrub_checksum_tree_block(sblock);
1802	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1803		(void)scrub_checksum_super(sblock);
1804	else
1805		WARN_ON(1);
1806	if (ret)
1807		scrub_handle_errored_block(sblock);
1808
1809	return ret;
1810}
1811
1812static int scrub_checksum_data(struct scrub_block *sblock)
1813{
1814	struct scrub_ctx *sctx = sblock->sctx;
 
 
1815	u8 csum[BTRFS_CSUM_SIZE];
1816	u8 *on_disk_csum;
1817	struct page *page;
1818	void *buffer;
1819	u32 crc = ~(u32)0;
1820	u64 len;
1821	int index;
1822
1823	BUG_ON(sblock->page_count < 1);
1824	if (!sblock->pagev[0]->have_csum)
 
1825		return 0;
1826
1827	on_disk_csum = sblock->pagev[0]->csum;
1828	page = sblock->pagev[0]->page;
1829	buffer = kmap_atomic(page);
1830
1831	len = sctx->sectorsize;
1832	index = 0;
1833	for (;;) {
1834		u64 l = min_t(u64, len, PAGE_SIZE);
1835
1836		crc = btrfs_csum_data(buffer, crc, l);
1837		kunmap_atomic(buffer);
1838		len -= l;
1839		if (len == 0)
1840			break;
1841		index++;
1842		BUG_ON(index >= sblock->page_count);
1843		BUG_ON(!sblock->pagev[index]->page);
1844		page = sblock->pagev[index]->page;
1845		buffer = kmap_atomic(page);
1846	}
1847
1848	btrfs_csum_final(crc, csum);
1849	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1850		sblock->checksum_error = 1;
 
 
1851
 
 
1852	return sblock->checksum_error;
1853}
1854
1855static int scrub_checksum_tree_block(struct scrub_block *sblock)
1856{
1857	struct scrub_ctx *sctx = sblock->sctx;
1858	struct btrfs_header *h;
1859	struct btrfs_root *root = sctx->dev_root;
1860	struct btrfs_fs_info *fs_info = root->fs_info;
1861	u8 calculated_csum[BTRFS_CSUM_SIZE];
1862	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1863	struct page *page;
1864	void *mapped_buffer;
1865	u64 mapped_size;
1866	void *p;
1867	u32 crc = ~(u32)0;
1868	u64 len;
1869	int index;
 
 
 
1870
1871	BUG_ON(sblock->page_count < 1);
1872	page = sblock->pagev[0]->page;
1873	mapped_buffer = kmap_atomic(page);
1874	h = (struct btrfs_header *)mapped_buffer;
1875	memcpy(on_disk_csum, h->csum, sctx->csum_size);
 
 
 
 
1876
1877	/*
1878	 * we don't use the getter functions here, as we
1879	 * a) don't have an extent buffer and
1880	 * b) the page is already kmapped
1881	 */
1882	if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1883		sblock->header_error = 1;
1884
1885	if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1886		sblock->header_error = 1;
1887		sblock->generation_error = 1;
1888	}
1889
1890	if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1891		sblock->header_error = 1;
1892
1893	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1894		   BTRFS_UUID_SIZE))
1895		sblock->header_error = 1;
1896
1897	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1898	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1899	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1900	index = 0;
1901	for (;;) {
1902		u64 l = min_t(u64, len, mapped_size);
1903
1904		crc = btrfs_csum_data(p, crc, l);
1905		kunmap_atomic(mapped_buffer);
1906		len -= l;
1907		if (len == 0)
1908			break;
1909		index++;
1910		BUG_ON(index >= sblock->page_count);
1911		BUG_ON(!sblock->pagev[index]->page);
1912		page = sblock->pagev[index]->page;
1913		mapped_buffer = kmap_atomic(page);
1914		mapped_size = PAGE_SIZE;
1915		p = mapped_buffer;
1916	}
1917
1918	btrfs_csum_final(crc, calculated_csum);
1919	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1920		sblock->checksum_error = 1;
1921
1922	return sblock->header_error || sblock->checksum_error;
1923}
1924
1925static int scrub_checksum_super(struct scrub_block *sblock)
1926{
1927	struct btrfs_super_block *s;
1928	struct scrub_ctx *sctx = sblock->sctx;
 
 
1929	u8 calculated_csum[BTRFS_CSUM_SIZE];
1930	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1931	struct page *page;
1932	void *mapped_buffer;
1933	u64 mapped_size;
1934	void *p;
1935	u32 crc = ~(u32)0;
1936	int fail_gen = 0;
1937	int fail_cor = 0;
1938	u64 len;
1939	int index;
1940
1941	BUG_ON(sblock->page_count < 1);
1942	page = sblock->pagev[0]->page;
1943	mapped_buffer = kmap_atomic(page);
1944	s = (struct btrfs_super_block *)mapped_buffer;
1945	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1946
1947	if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1948		++fail_cor;
1949
1950	if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1951		++fail_gen;
1952
1953	if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1954		++fail_cor;
1955
1956	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1957	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1958	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1959	index = 0;
1960	for (;;) {
1961		u64 l = min_t(u64, len, mapped_size);
1962
1963		crc = btrfs_csum_data(p, crc, l);
1964		kunmap_atomic(mapped_buffer);
1965		len -= l;
1966		if (len == 0)
1967			break;
1968		index++;
1969		BUG_ON(index >= sblock->page_count);
1970		BUG_ON(!sblock->pagev[index]->page);
1971		page = sblock->pagev[index]->page;
1972		mapped_buffer = kmap_atomic(page);
1973		mapped_size = PAGE_SIZE;
1974		p = mapped_buffer;
1975	}
1976
1977	btrfs_csum_final(crc, calculated_csum);
1978	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1979		++fail_cor;
1980
1981	if (fail_cor + fail_gen) {
1982		/*
1983		 * if we find an error in a super block, we just report it.
1984		 * They will get written with the next transaction commit
1985		 * anyway
1986		 */
1987		spin_lock(&sctx->stat_lock);
1988		++sctx->stat.super_errors;
1989		spin_unlock(&sctx->stat_lock);
1990		if (fail_cor)
1991			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1992				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1993		else
1994			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1995				BTRFS_DEV_STAT_GENERATION_ERRS);
1996	}
1997
1998	return fail_cor + fail_gen;
1999}
2000
2001static void scrub_block_get(struct scrub_block *sblock)
2002{
2003	atomic_inc(&sblock->refs);
2004}
2005
2006static void scrub_block_put(struct scrub_block *sblock)
2007{
2008	if (atomic_dec_and_test(&sblock->refs)) {
2009		int i;
2010
2011		if (sblock->sparity)
2012			scrub_parity_put(sblock->sparity);
2013
2014		for (i = 0; i < sblock->page_count; i++)
2015			scrub_page_put(sblock->pagev[i]);
2016		kfree(sblock);
2017	}
2018}
2019
2020static void scrub_page_get(struct scrub_page *spage)
2021{
2022	atomic_inc(&spage->refs);
2023}
2024
2025static void scrub_page_put(struct scrub_page *spage)
2026{
2027	if (atomic_dec_and_test(&spage->refs)) {
2028		if (spage->page)
2029			__free_page(spage->page);
2030		kfree(spage);
2031	}
2032}
2033
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2034static void scrub_submit(struct scrub_ctx *sctx)
2035{
2036	struct scrub_bio *sbio;
2037
2038	if (sctx->curr == -1)
2039		return;
2040
 
 
2041	sbio = sctx->bios[sctx->curr];
2042	sctx->curr = -1;
2043	scrub_pending_bio_inc(sctx);
2044	btrfsic_submit_bio(READ, sbio->bio);
2045}
2046
2047static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2048				    struct scrub_page *spage)
2049{
2050	struct scrub_block *sblock = spage->sblock;
2051	struct scrub_bio *sbio;
 
2052	int ret;
2053
2054again:
2055	/*
2056	 * grab a fresh bio or wait for one to become available
2057	 */
2058	while (sctx->curr == -1) {
2059		spin_lock(&sctx->list_lock);
2060		sctx->curr = sctx->first_free;
2061		if (sctx->curr != -1) {
2062			sctx->first_free = sctx->bios[sctx->curr]->next_free;
2063			sctx->bios[sctx->curr]->next_free = -1;
2064			sctx->bios[sctx->curr]->page_count = 0;
2065			spin_unlock(&sctx->list_lock);
2066		} else {
2067			spin_unlock(&sctx->list_lock);
2068			wait_event(sctx->list_wait, sctx->first_free != -1);
2069		}
2070	}
2071	sbio = sctx->bios[sctx->curr];
2072	if (sbio->page_count == 0) {
2073		struct bio *bio;
2074
2075		sbio->physical = spage->physical;
2076		sbio->logical = spage->logical;
2077		sbio->dev = spage->dev;
2078		bio = sbio->bio;
2079		if (!bio) {
2080			bio = btrfs_io_bio_alloc(GFP_KERNEL,
2081					sctx->pages_per_rd_bio);
2082			if (!bio)
2083				return -ENOMEM;
2084			sbio->bio = bio;
2085		}
2086
2087		bio->bi_private = sbio;
2088		bio->bi_end_io = scrub_bio_end_io;
2089		bio->bi_bdev = sbio->dev->bdev;
2090		bio->bi_iter.bi_sector = sbio->physical >> 9;
2091		sbio->err = 0;
2092	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
 
2093		   spage->physical ||
2094		   sbio->logical + sbio->page_count * PAGE_SIZE !=
2095		   spage->logical ||
2096		   sbio->dev != spage->dev) {
2097		scrub_submit(sctx);
2098		goto again;
2099	}
2100
2101	sbio->pagev[sbio->page_count] = spage;
2102	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2103	if (ret != PAGE_SIZE) {
2104		if (sbio->page_count < 1) {
2105			bio_put(sbio->bio);
2106			sbio->bio = NULL;
2107			return -EIO;
2108		}
2109		scrub_submit(sctx);
2110		goto again;
2111	}
2112
2113	scrub_block_get(sblock); /* one for the page added to the bio */
2114	atomic_inc(&sblock->outstanding_pages);
2115	sbio->page_count++;
2116	if (sbio->page_count == sctx->pages_per_rd_bio)
2117		scrub_submit(sctx);
2118
2119	return 0;
2120}
2121
2122static void scrub_missing_raid56_end_io(struct bio *bio)
2123{
2124	struct scrub_block *sblock = bio->bi_private;
2125	struct btrfs_fs_info *fs_info = sblock->sctx->dev_root->fs_info;
2126
2127	if (bio->bi_error)
2128		sblock->no_io_error_seen = 0;
2129
 
 
2130	btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2131}
2132
2133static void scrub_missing_raid56_worker(struct btrfs_work *work)
2134{
2135	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2136	struct scrub_ctx *sctx = sblock->sctx;
 
2137	u64 logical;
2138	struct btrfs_device *dev;
2139
2140	logical = sblock->pagev[0]->logical;
2141	dev = sblock->pagev[0]->dev;
2142
2143	if (sblock->no_io_error_seen)
2144		scrub_recheck_block_checksum(sblock);
2145
2146	if (!sblock->no_io_error_seen) {
2147		spin_lock(&sctx->stat_lock);
2148		sctx->stat.read_errors++;
2149		spin_unlock(&sctx->stat_lock);
2150		btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2151			"IO error rebuilding logical %llu for dev %s",
2152			logical, rcu_str_deref(dev->name));
2153	} else if (sblock->header_error || sblock->checksum_error) {
2154		spin_lock(&sctx->stat_lock);
2155		sctx->stat.uncorrectable_errors++;
2156		spin_unlock(&sctx->stat_lock);
2157		btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2158			"failed to rebuild valid logical %llu for dev %s",
2159			logical, rcu_str_deref(dev->name));
2160	} else {
2161		scrub_write_block_to_dev_replace(sblock);
2162	}
2163
2164	scrub_block_put(sblock);
2165
2166	if (sctx->is_dev_replace &&
2167	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2168		mutex_lock(&sctx->wr_ctx.wr_lock);
2169		scrub_wr_submit(sctx);
2170		mutex_unlock(&sctx->wr_ctx.wr_lock);
2171	}
2172
 
2173	scrub_pending_bio_dec(sctx);
2174}
2175
2176static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2177{
2178	struct scrub_ctx *sctx = sblock->sctx;
2179	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2180	u64 length = sblock->page_count * PAGE_SIZE;
2181	u64 logical = sblock->pagev[0]->logical;
2182	struct btrfs_bio *bbio;
2183	struct bio *bio;
2184	struct btrfs_raid_bio *rbio;
2185	int ret;
2186	int i;
2187
2188	ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
2189			       &bbio, 0, 1);
 
2190	if (ret || !bbio || !bbio->raid_map)
2191		goto bbio_out;
2192
2193	if (WARN_ON(!sctx->is_dev_replace ||
2194		    !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2195		/*
2196		 * We shouldn't be scrubbing a missing device. Even for dev
2197		 * replace, we should only get here for RAID 5/6. We either
2198		 * managed to mount something with no mirrors remaining or
2199		 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2200		 */
2201		goto bbio_out;
2202	}
2203
2204	bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2205	if (!bio)
2206		goto bbio_out;
2207
2208	bio->bi_iter.bi_sector = logical >> 9;
2209	bio->bi_private = sblock;
2210	bio->bi_end_io = scrub_missing_raid56_end_io;
2211
2212	rbio = raid56_alloc_missing_rbio(sctx->dev_root, bio, bbio, length);
2213	if (!rbio)
2214		goto rbio_out;
2215
2216	for (i = 0; i < sblock->page_count; i++) {
2217		struct scrub_page *spage = sblock->pagev[i];
2218
2219		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2220	}
2221
2222	btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2223			scrub_missing_raid56_worker, NULL, NULL);
2224	scrub_block_get(sblock);
2225	scrub_pending_bio_inc(sctx);
2226	raid56_submit_missing_rbio(rbio);
2227	return;
2228
2229rbio_out:
2230	bio_put(bio);
2231bbio_out:
 
2232	btrfs_put_bbio(bbio);
2233	spin_lock(&sctx->stat_lock);
2234	sctx->stat.malloc_errors++;
2235	spin_unlock(&sctx->stat_lock);
2236}
2237
2238static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2239		       u64 physical, struct btrfs_device *dev, u64 flags,
2240		       u64 gen, int mirror_num, u8 *csum, int force,
2241		       u64 physical_for_dev_replace)
2242{
2243	struct scrub_block *sblock;
 
2244	int index;
2245
2246	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2247	if (!sblock) {
2248		spin_lock(&sctx->stat_lock);
2249		sctx->stat.malloc_errors++;
2250		spin_unlock(&sctx->stat_lock);
2251		return -ENOMEM;
2252	}
2253
2254	/* one ref inside this function, plus one for each page added to
2255	 * a bio later on */
2256	atomic_set(&sblock->refs, 1);
2257	sblock->sctx = sctx;
2258	sblock->no_io_error_seen = 1;
2259
2260	for (index = 0; len > 0; index++) {
2261		struct scrub_page *spage;
2262		u64 l = min_t(u64, len, PAGE_SIZE);
 
 
 
 
 
2263
2264		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2265		if (!spage) {
2266leave_nomem:
2267			spin_lock(&sctx->stat_lock);
2268			sctx->stat.malloc_errors++;
2269			spin_unlock(&sctx->stat_lock);
2270			scrub_block_put(sblock);
2271			return -ENOMEM;
2272		}
2273		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2274		scrub_page_get(spage);
2275		sblock->pagev[index] = spage;
2276		spage->sblock = sblock;
2277		spage->dev = dev;
2278		spage->flags = flags;
2279		spage->generation = gen;
2280		spage->logical = logical;
2281		spage->physical = physical;
2282		spage->physical_for_dev_replace = physical_for_dev_replace;
2283		spage->mirror_num = mirror_num;
2284		if (csum) {
2285			spage->have_csum = 1;
2286			memcpy(spage->csum, csum, sctx->csum_size);
2287		} else {
2288			spage->have_csum = 0;
2289		}
2290		sblock->page_count++;
2291		spage->page = alloc_page(GFP_KERNEL);
2292		if (!spage->page)
2293			goto leave_nomem;
2294		len -= l;
2295		logical += l;
2296		physical += l;
2297		physical_for_dev_replace += l;
2298	}
2299
2300	WARN_ON(sblock->page_count == 0);
2301	if (dev->missing) {
2302		/*
2303		 * This case should only be hit for RAID 5/6 device replace. See
2304		 * the comment in scrub_missing_raid56_pages() for details.
2305		 */
2306		scrub_missing_raid56_pages(sblock);
2307	} else {
2308		for (index = 0; index < sblock->page_count; index++) {
2309			struct scrub_page *spage = sblock->pagev[index];
2310			int ret;
2311
2312			ret = scrub_add_page_to_rd_bio(sctx, spage);
2313			if (ret) {
2314				scrub_block_put(sblock);
2315				return ret;
2316			}
2317		}
2318
2319		if (force)
2320			scrub_submit(sctx);
2321	}
2322
2323	/* last one frees, either here or in bio completion for last page */
2324	scrub_block_put(sblock);
2325	return 0;
2326}
2327
2328static void scrub_bio_end_io(struct bio *bio)
2329{
2330	struct scrub_bio *sbio = bio->bi_private;
2331	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2332
2333	sbio->err = bio->bi_error;
2334	sbio->bio = bio;
2335
2336	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2337}
2338
2339static void scrub_bio_end_io_worker(struct btrfs_work *work)
2340{
2341	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2342	struct scrub_ctx *sctx = sbio->sctx;
2343	int i;
2344
2345	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2346	if (sbio->err) {
2347		for (i = 0; i < sbio->page_count; i++) {
2348			struct scrub_page *spage = sbio->pagev[i];
2349
2350			spage->io_error = 1;
2351			spage->sblock->no_io_error_seen = 0;
2352		}
2353	}
2354
2355	/* now complete the scrub_block items that have all pages completed */
2356	for (i = 0; i < sbio->page_count; i++) {
2357		struct scrub_page *spage = sbio->pagev[i];
2358		struct scrub_block *sblock = spage->sblock;
2359
2360		if (atomic_dec_and_test(&sblock->outstanding_pages))
2361			scrub_block_complete(sblock);
2362		scrub_block_put(sblock);
2363	}
2364
2365	bio_put(sbio->bio);
2366	sbio->bio = NULL;
2367	spin_lock(&sctx->list_lock);
2368	sbio->next_free = sctx->first_free;
2369	sctx->first_free = sbio->index;
2370	spin_unlock(&sctx->list_lock);
2371
2372	if (sctx->is_dev_replace &&
2373	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2374		mutex_lock(&sctx->wr_ctx.wr_lock);
2375		scrub_wr_submit(sctx);
2376		mutex_unlock(&sctx->wr_ctx.wr_lock);
2377	}
2378
2379	scrub_pending_bio_dec(sctx);
2380}
2381
2382static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2383				       unsigned long *bitmap,
2384				       u64 start, u64 len)
2385{
2386	u32 offset;
2387	int nsectors;
2388	int sectorsize = sparity->sctx->dev_root->sectorsize;
2389
2390	if (len >= sparity->stripe_len) {
2391		bitmap_set(bitmap, 0, sparity->nsectors);
2392		return;
2393	}
2394
2395	start -= sparity->logic_start;
2396	start = div_u64_rem(start, sparity->stripe_len, &offset);
2397	offset /= sectorsize;
2398	nsectors = (int)len / sectorsize;
2399
2400	if (offset + nsectors <= sparity->nsectors) {
2401		bitmap_set(bitmap, offset, nsectors);
2402		return;
2403	}
2404
2405	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2406	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2407}
2408
2409static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2410						   u64 start, u64 len)
2411{
2412	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2413}
2414
2415static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2416						  u64 start, u64 len)
2417{
2418	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2419}
2420
2421static void scrub_block_complete(struct scrub_block *sblock)
2422{
2423	int corrupted = 0;
2424
2425	if (!sblock->no_io_error_seen) {
2426		corrupted = 1;
2427		scrub_handle_errored_block(sblock);
2428	} else {
2429		/*
2430		 * if has checksum error, write via repair mechanism in
2431		 * dev replace case, otherwise write here in dev replace
2432		 * case.
2433		 */
2434		corrupted = scrub_checksum(sblock);
2435		if (!corrupted && sblock->sctx->is_dev_replace)
2436			scrub_write_block_to_dev_replace(sblock);
2437	}
2438
2439	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2440		u64 start = sblock->pagev[0]->logical;
2441		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2442			  PAGE_SIZE;
2443
 
2444		scrub_parity_mark_sectors_error(sblock->sparity,
2445						start, end - start);
2446	}
2447}
2448
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2449static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2450{
2451	struct btrfs_ordered_sum *sum = NULL;
2452	unsigned long index;
2453	unsigned long num_sectors;
2454
2455	while (!list_empty(&sctx->csum_list)) {
 
 
 
 
2456		sum = list_first_entry(&sctx->csum_list,
2457				       struct btrfs_ordered_sum, list);
 
2458		if (sum->bytenr > logical)
2459			return 0;
2460		if (sum->bytenr + sum->len > logical)
2461			break;
2462
2463		++sctx->stat.csum_discards;
2464		list_del(&sum->list);
2465		kfree(sum);
2466		sum = NULL;
2467	}
2468	if (!sum)
2469		return 0;
 
 
 
2470
2471	index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2472	num_sectors = sum->len / sctx->sectorsize;
2473	memcpy(csum, sum->sums + index, sctx->csum_size);
2474	if (index == num_sectors - 1) {
2475		list_del(&sum->list);
2476		kfree(sum);
 
 
 
 
 
 
2477	}
 
 
2478	return 1;
2479}
2480
2481/* scrub extent tries to collect up to 64 kB for each bio */
2482static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
 
2483			u64 physical, struct btrfs_device *dev, u64 flags,
2484			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2485{
2486	int ret;
2487	u8 csum[BTRFS_CSUM_SIZE];
2488	u32 blocksize;
2489
2490	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2491		blocksize = sctx->sectorsize;
 
 
 
2492		spin_lock(&sctx->stat_lock);
2493		sctx->stat.data_extents_scrubbed++;
2494		sctx->stat.data_bytes_scrubbed += len;
2495		spin_unlock(&sctx->stat_lock);
2496	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2497		blocksize = sctx->nodesize;
 
 
 
2498		spin_lock(&sctx->stat_lock);
2499		sctx->stat.tree_extents_scrubbed++;
2500		sctx->stat.tree_bytes_scrubbed += len;
2501		spin_unlock(&sctx->stat_lock);
2502	} else {
2503		blocksize = sctx->sectorsize;
2504		WARN_ON(1);
2505	}
2506
2507	while (len) {
2508		u64 l = min_t(u64, len, blocksize);
2509		int have_csum = 0;
2510
2511		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2512			/* push csums to sbio */
2513			have_csum = scrub_find_csum(sctx, logical, csum);
2514			if (have_csum == 0)
2515				++sctx->stat.no_csum;
2516			if (sctx->is_dev_replace && !have_csum) {
2517				ret = copy_nocow_pages(sctx, logical, l,
2518						       mirror_num,
2519						      physical_for_dev_replace);
2520				goto behind_scrub_pages;
2521			}
2522		}
2523		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2524				  mirror_num, have_csum ? csum : NULL, 0,
2525				  physical_for_dev_replace);
2526behind_scrub_pages:
2527		if (ret)
2528			return ret;
2529		len -= l;
2530		logical += l;
2531		physical += l;
2532		physical_for_dev_replace += l;
2533	}
2534	return 0;
2535}
2536
2537static int scrub_pages_for_parity(struct scrub_parity *sparity,
2538				  u64 logical, u64 len,
2539				  u64 physical, struct btrfs_device *dev,
2540				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2541{
2542	struct scrub_ctx *sctx = sparity->sctx;
2543	struct scrub_block *sblock;
 
2544	int index;
2545
 
 
2546	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2547	if (!sblock) {
2548		spin_lock(&sctx->stat_lock);
2549		sctx->stat.malloc_errors++;
2550		spin_unlock(&sctx->stat_lock);
2551		return -ENOMEM;
2552	}
2553
2554	/* one ref inside this function, plus one for each page added to
2555	 * a bio later on */
2556	atomic_set(&sblock->refs, 1);
2557	sblock->sctx = sctx;
2558	sblock->no_io_error_seen = 1;
2559	sblock->sparity = sparity;
2560	scrub_parity_get(sparity);
2561
2562	for (index = 0; len > 0; index++) {
2563		struct scrub_page *spage;
2564		u64 l = min_t(u64, len, PAGE_SIZE);
2565
2566		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2567		if (!spage) {
2568leave_nomem:
2569			spin_lock(&sctx->stat_lock);
2570			sctx->stat.malloc_errors++;
2571			spin_unlock(&sctx->stat_lock);
2572			scrub_block_put(sblock);
2573			return -ENOMEM;
2574		}
2575		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2576		/* For scrub block */
2577		scrub_page_get(spage);
2578		sblock->pagev[index] = spage;
2579		/* For scrub parity */
2580		scrub_page_get(spage);
2581		list_add_tail(&spage->list, &sparity->spages);
2582		spage->sblock = sblock;
2583		spage->dev = dev;
2584		spage->flags = flags;
2585		spage->generation = gen;
2586		spage->logical = logical;
2587		spage->physical = physical;
2588		spage->mirror_num = mirror_num;
2589		if (csum) {
2590			spage->have_csum = 1;
2591			memcpy(spage->csum, csum, sctx->csum_size);
2592		} else {
2593			spage->have_csum = 0;
2594		}
2595		sblock->page_count++;
2596		spage->page = alloc_page(GFP_KERNEL);
2597		if (!spage->page)
2598			goto leave_nomem;
2599		len -= l;
2600		logical += l;
2601		physical += l;
 
 
 
2602	}
2603
2604	WARN_ON(sblock->page_count == 0);
2605	for (index = 0; index < sblock->page_count; index++) {
2606		struct scrub_page *spage = sblock->pagev[index];
2607		int ret;
2608
2609		ret = scrub_add_page_to_rd_bio(sctx, spage);
2610		if (ret) {
2611			scrub_block_put(sblock);
2612			return ret;
2613		}
2614	}
2615
2616	/* last one frees, either here or in bio completion for last page */
2617	scrub_block_put(sblock);
2618	return 0;
2619}
2620
2621static int scrub_extent_for_parity(struct scrub_parity *sparity,
2622				   u64 logical, u64 len,
2623				   u64 physical, struct btrfs_device *dev,
2624				   u64 flags, u64 gen, int mirror_num)
2625{
2626	struct scrub_ctx *sctx = sparity->sctx;
2627	int ret;
2628	u8 csum[BTRFS_CSUM_SIZE];
2629	u32 blocksize;
2630
2631	if (dev->missing) {
2632		scrub_parity_mark_sectors_error(sparity, logical, len);
2633		return 0;
2634	}
2635
2636	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2637		blocksize = sctx->sectorsize;
2638	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2639		blocksize = sctx->nodesize;
2640	} else {
2641		blocksize = sctx->sectorsize;
2642		WARN_ON(1);
2643	}
2644
2645	while (len) {
2646		u64 l = min_t(u64, len, blocksize);
2647		int have_csum = 0;
2648
2649		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2650			/* push csums to sbio */
2651			have_csum = scrub_find_csum(sctx, logical, csum);
2652			if (have_csum == 0)
2653				goto skip;
2654		}
2655		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2656					     flags, gen, mirror_num,
2657					     have_csum ? csum : NULL);
2658		if (ret)
2659			return ret;
2660skip:
2661		len -= l;
2662		logical += l;
2663		physical += l;
2664	}
2665	return 0;
2666}
2667
2668/*
2669 * Given a physical address, this will calculate it's
2670 * logical offset. if this is a parity stripe, it will return
2671 * the most left data stripe's logical offset.
2672 *
2673 * return 0 if it is a data stripe, 1 means parity stripe.
2674 */
2675static int get_raid56_logic_offset(u64 physical, int num,
2676				   struct map_lookup *map, u64 *offset,
2677				   u64 *stripe_start)
2678{
2679	int i;
2680	int j = 0;
2681	u64 stripe_nr;
2682	u64 last_offset;
2683	u32 stripe_index;
2684	u32 rot;
 
2685
2686	last_offset = (physical - map->stripes[num].physical) *
2687		      nr_data_stripes(map);
2688	if (stripe_start)
2689		*stripe_start = last_offset;
2690
2691	*offset = last_offset;
2692	for (i = 0; i < nr_data_stripes(map); i++) {
2693		*offset = last_offset + i * map->stripe_len;
2694
2695		stripe_nr = div_u64(*offset, map->stripe_len);
2696		stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2697
2698		/* Work out the disk rotation on this stripe-set */
2699		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2700		/* calculate which stripe this data locates */
2701		rot += i;
2702		stripe_index = rot % map->num_stripes;
2703		if (stripe_index == num)
2704			return 0;
2705		if (stripe_index < num)
2706			j++;
2707	}
2708	*offset = last_offset + j * map->stripe_len;
2709	return 1;
2710}
2711
2712static void scrub_free_parity(struct scrub_parity *sparity)
2713{
2714	struct scrub_ctx *sctx = sparity->sctx;
2715	struct scrub_page *curr, *next;
2716	int nbits;
2717
2718	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2719	if (nbits) {
2720		spin_lock(&sctx->stat_lock);
2721		sctx->stat.read_errors += nbits;
2722		sctx->stat.uncorrectable_errors += nbits;
2723		spin_unlock(&sctx->stat_lock);
2724	}
2725
2726	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2727		list_del_init(&curr->list);
2728		scrub_page_put(curr);
2729	}
2730
2731	kfree(sparity);
2732}
2733
2734static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2735{
2736	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2737						    work);
2738	struct scrub_ctx *sctx = sparity->sctx;
2739
2740	scrub_free_parity(sparity);
2741	scrub_pending_bio_dec(sctx);
2742}
2743
2744static void scrub_parity_bio_endio(struct bio *bio)
2745{
2746	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
 
2747
2748	if (bio->bi_error)
2749		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2750			  sparity->nsectors);
2751
2752	bio_put(bio);
2753
2754	btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2755			scrub_parity_bio_endio_worker, NULL, NULL);
2756	btrfs_queue_work(sparity->sctx->dev_root->fs_info->scrub_parity_workers,
2757			 &sparity->work);
2758}
2759
2760static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2761{
2762	struct scrub_ctx *sctx = sparity->sctx;
 
2763	struct bio *bio;
2764	struct btrfs_raid_bio *rbio;
2765	struct scrub_page *spage;
2766	struct btrfs_bio *bbio = NULL;
2767	u64 length;
2768	int ret;
2769
2770	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2771			   sparity->nsectors))
2772		goto out;
2773
2774	length = sparity->logic_end - sparity->logic_start;
2775	ret = btrfs_map_sblock(sctx->dev_root->fs_info, WRITE,
2776			       sparity->logic_start,
2777			       &length, &bbio, 0, 1);
2778	if (ret || !bbio || !bbio->raid_map)
2779		goto bbio_out;
2780
2781	bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2782	if (!bio)
 
 
2783		goto bbio_out;
2784
 
2785	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2786	bio->bi_private = sparity;
2787	bio->bi_end_io = scrub_parity_bio_endio;
2788
2789	rbio = raid56_parity_alloc_scrub_rbio(sctx->dev_root, bio, bbio,
2790					      length, sparity->scrub_dev,
2791					      sparity->dbitmap,
2792					      sparity->nsectors);
2793	if (!rbio)
2794		goto rbio_out;
2795
2796	list_for_each_entry(spage, &sparity->spages, list)
2797		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2798
2799	scrub_pending_bio_inc(sctx);
2800	raid56_parity_submit_scrub_rbio(rbio);
2801	return;
2802
2803rbio_out:
2804	bio_put(bio);
2805bbio_out:
 
2806	btrfs_put_bbio(bbio);
2807	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2808		  sparity->nsectors);
2809	spin_lock(&sctx->stat_lock);
2810	sctx->stat.malloc_errors++;
2811	spin_unlock(&sctx->stat_lock);
2812out:
2813	scrub_free_parity(sparity);
2814}
2815
2816static inline int scrub_calc_parity_bitmap_len(int nsectors)
2817{
2818	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2819}
2820
2821static void scrub_parity_get(struct scrub_parity *sparity)
2822{
2823	atomic_inc(&sparity->refs);
2824}
2825
2826static void scrub_parity_put(struct scrub_parity *sparity)
2827{
2828	if (!atomic_dec_and_test(&sparity->refs))
2829		return;
2830
2831	scrub_parity_check_and_repair(sparity);
2832}
2833
2834static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2835						  struct map_lookup *map,
2836						  struct btrfs_device *sdev,
2837						  struct btrfs_path *path,
2838						  u64 logic_start,
2839						  u64 logic_end)
2840{
2841	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2842	struct btrfs_root *root = fs_info->extent_root;
2843	struct btrfs_root *csum_root = fs_info->csum_root;
2844	struct btrfs_extent_item *extent;
2845	struct btrfs_bio *bbio = NULL;
2846	u64 flags;
2847	int ret;
2848	int slot;
2849	struct extent_buffer *l;
2850	struct btrfs_key key;
2851	u64 generation;
2852	u64 extent_logical;
2853	u64 extent_physical;
2854	u64 extent_len;
 
2855	u64 mapped_length;
2856	struct btrfs_device *extent_dev;
2857	struct scrub_parity *sparity;
2858	int nsectors;
2859	int bitmap_len;
2860	int extent_mirror_num;
2861	int stop_loop = 0;
2862
2863	nsectors = map->stripe_len / root->sectorsize;
 
2864	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2865	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2866			  GFP_NOFS);
2867	if (!sparity) {
2868		spin_lock(&sctx->stat_lock);
2869		sctx->stat.malloc_errors++;
2870		spin_unlock(&sctx->stat_lock);
2871		return -ENOMEM;
2872	}
2873
 
2874	sparity->stripe_len = map->stripe_len;
2875	sparity->nsectors = nsectors;
2876	sparity->sctx = sctx;
2877	sparity->scrub_dev = sdev;
2878	sparity->logic_start = logic_start;
2879	sparity->logic_end = logic_end;
2880	atomic_set(&sparity->refs, 1);
2881	INIT_LIST_HEAD(&sparity->spages);
2882	sparity->dbitmap = sparity->bitmap;
2883	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2884
2885	ret = 0;
2886	while (logic_start < logic_end) {
2887		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2888			key.type = BTRFS_METADATA_ITEM_KEY;
2889		else
2890			key.type = BTRFS_EXTENT_ITEM_KEY;
2891		key.objectid = logic_start;
2892		key.offset = (u64)-1;
2893
2894		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2895		if (ret < 0)
2896			goto out;
2897
2898		if (ret > 0) {
2899			ret = btrfs_previous_extent_item(root, path, 0);
2900			if (ret < 0)
2901				goto out;
2902			if (ret > 0) {
2903				btrfs_release_path(path);
2904				ret = btrfs_search_slot(NULL, root, &key,
2905							path, 0, 0);
2906				if (ret < 0)
2907					goto out;
2908			}
2909		}
2910
2911		stop_loop = 0;
2912		while (1) {
2913			u64 bytes;
2914
2915			l = path->nodes[0];
2916			slot = path->slots[0];
2917			if (slot >= btrfs_header_nritems(l)) {
2918				ret = btrfs_next_leaf(root, path);
2919				if (ret == 0)
2920					continue;
2921				if (ret < 0)
2922					goto out;
2923
2924				stop_loop = 1;
2925				break;
2926			}
2927			btrfs_item_key_to_cpu(l, &key, slot);
2928
2929			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2930			    key.type != BTRFS_METADATA_ITEM_KEY)
2931				goto next;
2932
2933			if (key.type == BTRFS_METADATA_ITEM_KEY)
2934				bytes = root->nodesize;
2935			else
2936				bytes = key.offset;
2937
2938			if (key.objectid + bytes <= logic_start)
2939				goto next;
2940
2941			if (key.objectid >= logic_end) {
2942				stop_loop = 1;
2943				break;
2944			}
2945
2946			while (key.objectid >= logic_start + map->stripe_len)
2947				logic_start += map->stripe_len;
2948
2949			extent = btrfs_item_ptr(l, slot,
2950						struct btrfs_extent_item);
2951			flags = btrfs_extent_flags(l, extent);
2952			generation = btrfs_extent_generation(l, extent);
2953
2954			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2955			    (key.objectid < logic_start ||
2956			     key.objectid + bytes >
2957			     logic_start + map->stripe_len)) {
2958				btrfs_err(fs_info, "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
 
2959					  key.objectid, logic_start);
2960				spin_lock(&sctx->stat_lock);
2961				sctx->stat.uncorrectable_errors++;
2962				spin_unlock(&sctx->stat_lock);
2963				goto next;
2964			}
2965again:
2966			extent_logical = key.objectid;
 
2967			extent_len = bytes;
2968
2969			if (extent_logical < logic_start) {
2970				extent_len -= logic_start - extent_logical;
2971				extent_logical = logic_start;
2972			}
2973
2974			if (extent_logical + extent_len >
2975			    logic_start + map->stripe_len)
2976				extent_len = logic_start + map->stripe_len -
2977					     extent_logical;
2978
2979			scrub_parity_mark_sectors_data(sparity, extent_logical,
2980						       extent_len);
2981
2982			mapped_length = extent_len;
2983			ret = btrfs_map_block(fs_info, READ, extent_logical,
2984					      &mapped_length, &bbio, 0);
 
 
2985			if (!ret) {
2986				if (!bbio || mapped_length < extent_len)
2987					ret = -EIO;
2988			}
2989			if (ret) {
2990				btrfs_put_bbio(bbio);
2991				goto out;
2992			}
2993			extent_physical = bbio->stripes[0].physical;
2994			extent_mirror_num = bbio->mirror_num;
2995			extent_dev = bbio->stripes[0].dev;
2996			btrfs_put_bbio(bbio);
2997
2998			ret = btrfs_lookup_csums_range(csum_root,
2999						extent_logical,
3000						extent_logical + extent_len - 1,
3001						&sctx->csum_list, 1);
3002			if (ret)
3003				goto out;
3004
3005			ret = scrub_extent_for_parity(sparity, extent_logical,
3006						      extent_len,
3007						      extent_physical,
3008						      extent_dev, flags,
3009						      generation,
3010						      extent_mirror_num);
3011
3012			scrub_free_csums(sctx);
3013
3014			if (ret)
3015				goto out;
3016
3017			if (extent_logical + extent_len <
3018			    key.objectid + bytes) {
3019				logic_start += map->stripe_len;
3020
3021				if (logic_start >= logic_end) {
3022					stop_loop = 1;
3023					break;
3024				}
3025
3026				if (logic_start < key.objectid + bytes) {
3027					cond_resched();
3028					goto again;
3029				}
3030			}
3031next:
3032			path->slots[0]++;
3033		}
3034
3035		btrfs_release_path(path);
3036
3037		if (stop_loop)
3038			break;
3039
3040		logic_start += map->stripe_len;
3041	}
3042out:
3043	if (ret < 0)
 
3044		scrub_parity_mark_sectors_error(sparity, logic_start,
3045						logic_end - logic_start);
 
3046	scrub_parity_put(sparity);
3047	scrub_submit(sctx);
3048	mutex_lock(&sctx->wr_ctx.wr_lock);
3049	scrub_wr_submit(sctx);
3050	mutex_unlock(&sctx->wr_ctx.wr_lock);
3051
3052	btrfs_release_path(path);
3053	return ret < 0 ? ret : 0;
3054}
3055
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3056static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3057					   struct map_lookup *map,
3058					   struct btrfs_device *scrub_dev,
3059					   int num, u64 base, u64 length,
3060					   int is_dev_replace)
3061{
3062	struct btrfs_path *path, *ppath;
3063	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3064	struct btrfs_root *root = fs_info->extent_root;
3065	struct btrfs_root *csum_root = fs_info->csum_root;
3066	struct btrfs_extent_item *extent;
3067	struct blk_plug plug;
3068	u64 flags;
3069	int ret;
3070	int slot;
3071	u64 nstripes;
3072	struct extent_buffer *l;
3073	struct btrfs_key key;
3074	u64 physical;
3075	u64 logical;
3076	u64 logic_end;
3077	u64 physical_end;
3078	u64 generation;
3079	int mirror_num;
3080	struct reada_control *reada1;
3081	struct reada_control *reada2;
3082	struct btrfs_key key_start;
3083	struct btrfs_key key_end;
3084	u64 increment = map->stripe_len;
3085	u64 offset;
3086	u64 extent_logical;
3087	u64 extent_physical;
3088	u64 extent_len;
 
 
 
 
3089	u64 stripe_logical;
3090	u64 stripe_end;
3091	struct btrfs_device *extent_dev;
3092	int extent_mirror_num;
3093	int stop_loop = 0;
3094
3095	physical = map->stripes[num].physical;
3096	offset = 0;
3097	nstripes = div_u64(length, map->stripe_len);
 
 
3098	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3099		offset = map->stripe_len * num;
3100		increment = map->stripe_len * map->num_stripes;
3101		mirror_num = 1;
3102	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3103		int factor = map->num_stripes / map->sub_stripes;
3104		offset = map->stripe_len * (num / map->sub_stripes);
3105		increment = map->stripe_len * factor;
3106		mirror_num = num % map->sub_stripes + 1;
3107	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3108		increment = map->stripe_len;
3109		mirror_num = num % map->num_stripes + 1;
3110	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3111		increment = map->stripe_len;
3112		mirror_num = num % map->num_stripes + 1;
3113	} else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3114		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3115		increment = map->stripe_len * nr_data_stripes(map);
3116		mirror_num = 1;
3117	} else {
3118		increment = map->stripe_len;
3119		mirror_num = 1;
3120	}
3121
3122	path = btrfs_alloc_path();
3123	if (!path)
3124		return -ENOMEM;
3125
3126	ppath = btrfs_alloc_path();
3127	if (!ppath) {
3128		btrfs_free_path(path);
3129		return -ENOMEM;
3130	}
3131
3132	/*
3133	 * work on commit root. The related disk blocks are static as
3134	 * long as COW is applied. This means, it is save to rewrite
3135	 * them to repair disk errors without any race conditions
3136	 */
3137	path->search_commit_root = 1;
3138	path->skip_locking = 1;
3139
3140	ppath->search_commit_root = 1;
3141	ppath->skip_locking = 1;
3142	/*
3143	 * trigger the readahead for extent tree csum tree and wait for
3144	 * completion. During readahead, the scrub is officially paused
3145	 * to not hold off transaction commits
3146	 */
3147	logical = base + offset;
3148	physical_end = physical + nstripes * map->stripe_len;
3149	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3150		get_raid56_logic_offset(physical_end, num,
3151					map, &logic_end, NULL);
3152		logic_end += base;
3153	} else {
3154		logic_end = logical + increment * nstripes;
3155	}
3156	wait_event(sctx->list_wait,
3157		   atomic_read(&sctx->bios_in_flight) == 0);
3158	scrub_blocked_if_needed(fs_info);
3159
3160	/* FIXME it might be better to start readahead at commit root */
3161	key_start.objectid = logical;
3162	key_start.type = BTRFS_EXTENT_ITEM_KEY;
3163	key_start.offset = (u64)0;
3164	key_end.objectid = logic_end;
3165	key_end.type = BTRFS_METADATA_ITEM_KEY;
3166	key_end.offset = (u64)-1;
3167	reada1 = btrfs_reada_add(root, &key_start, &key_end);
3168
3169	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3170	key_start.type = BTRFS_EXTENT_CSUM_KEY;
3171	key_start.offset = logical;
3172	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3173	key_end.type = BTRFS_EXTENT_CSUM_KEY;
3174	key_end.offset = logic_end;
3175	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
 
 
 
 
3176
3177	if (!IS_ERR(reada1))
3178		btrfs_reada_wait(reada1);
3179	if (!IS_ERR(reada2))
3180		btrfs_reada_wait(reada2);
3181
3182
3183	/*
3184	 * collect all data csums for the stripe to avoid seeking during
3185	 * the scrub. This might currently (crc32) end up to be about 1MB
3186	 */
3187	blk_start_plug(&plug);
3188
 
 
 
 
 
 
 
 
3189	/*
3190	 * now find all extents for each stripe and scrub them
3191	 */
3192	ret = 0;
3193	while (physical < physical_end) {
3194		/*
3195		 * canceled?
3196		 */
3197		if (atomic_read(&fs_info->scrub_cancel_req) ||
3198		    atomic_read(&sctx->cancel_req)) {
3199			ret = -ECANCELED;
3200			goto out;
3201		}
3202		/*
3203		 * check to see if we have to pause
3204		 */
3205		if (atomic_read(&fs_info->scrub_pause_req)) {
3206			/* push queued extents */
3207			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3208			scrub_submit(sctx);
3209			mutex_lock(&sctx->wr_ctx.wr_lock);
3210			scrub_wr_submit(sctx);
3211			mutex_unlock(&sctx->wr_ctx.wr_lock);
3212			wait_event(sctx->list_wait,
3213				   atomic_read(&sctx->bios_in_flight) == 0);
3214			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3215			scrub_blocked_if_needed(fs_info);
3216		}
3217
3218		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3219			ret = get_raid56_logic_offset(physical, num, map,
3220						      &logical,
3221						      &stripe_logical);
3222			logical += base;
3223			if (ret) {
3224				/* it is parity strip */
3225				stripe_logical += base;
3226				stripe_end = stripe_logical + increment;
3227				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3228							  ppath, stripe_logical,
3229							  stripe_end);
3230				if (ret)
3231					goto out;
3232				goto skip;
3233			}
3234		}
3235
3236		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3237			key.type = BTRFS_METADATA_ITEM_KEY;
3238		else
3239			key.type = BTRFS_EXTENT_ITEM_KEY;
3240		key.objectid = logical;
3241		key.offset = (u64)-1;
3242
3243		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3244		if (ret < 0)
3245			goto out;
3246
3247		if (ret > 0) {
3248			ret = btrfs_previous_extent_item(root, path, 0);
3249			if (ret < 0)
3250				goto out;
3251			if (ret > 0) {
3252				/* there's no smaller item, so stick with the
3253				 * larger one */
3254				btrfs_release_path(path);
3255				ret = btrfs_search_slot(NULL, root, &key,
3256							path, 0, 0);
3257				if (ret < 0)
3258					goto out;
3259			}
3260		}
3261
3262		stop_loop = 0;
3263		while (1) {
3264			u64 bytes;
3265
3266			l = path->nodes[0];
3267			slot = path->slots[0];
3268			if (slot >= btrfs_header_nritems(l)) {
3269				ret = btrfs_next_leaf(root, path);
3270				if (ret == 0)
3271					continue;
3272				if (ret < 0)
3273					goto out;
3274
3275				stop_loop = 1;
3276				break;
3277			}
3278			btrfs_item_key_to_cpu(l, &key, slot);
3279
3280			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3281			    key.type != BTRFS_METADATA_ITEM_KEY)
3282				goto next;
3283
3284			if (key.type == BTRFS_METADATA_ITEM_KEY)
3285				bytes = root->nodesize;
3286			else
3287				bytes = key.offset;
3288
3289			if (key.objectid + bytes <= logical)
3290				goto next;
3291
3292			if (key.objectid >= logical + map->stripe_len) {
3293				/* out of this device extent */
3294				if (key.objectid >= logic_end)
3295					stop_loop = 1;
3296				break;
3297			}
3298
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3299			extent = btrfs_item_ptr(l, slot,
3300						struct btrfs_extent_item);
3301			flags = btrfs_extent_flags(l, extent);
3302			generation = btrfs_extent_generation(l, extent);
3303
3304			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3305			    (key.objectid < logical ||
3306			     key.objectid + bytes >
3307			     logical + map->stripe_len)) {
3308				btrfs_err(fs_info,
3309					   "scrub: tree block %llu spanning "
3310					   "stripes, ignored. logical=%llu",
3311				       key.objectid, logical);
3312				spin_lock(&sctx->stat_lock);
3313				sctx->stat.uncorrectable_errors++;
3314				spin_unlock(&sctx->stat_lock);
3315				goto next;
3316			}
3317
3318again:
3319			extent_logical = key.objectid;
 
3320			extent_len = bytes;
3321
3322			/*
3323			 * trim extent to this stripe
3324			 */
3325			if (extent_logical < logical) {
3326				extent_len -= logical - extent_logical;
3327				extent_logical = logical;
3328			}
3329			if (extent_logical + extent_len >
3330			    logical + map->stripe_len) {
3331				extent_len = logical + map->stripe_len -
3332					     extent_logical;
3333			}
3334
3335			extent_physical = extent_logical - logical + physical;
3336			extent_dev = scrub_dev;
3337			extent_mirror_num = mirror_num;
3338			if (is_dev_replace)
3339				scrub_remap_extent(fs_info, extent_logical,
3340						   extent_len, &extent_physical,
3341						   &extent_dev,
3342						   &extent_mirror_num);
3343
3344			ret = btrfs_lookup_csums_range(csum_root,
3345						       extent_logical,
3346						       extent_logical +
3347						       extent_len - 1,
3348						       &sctx->csum_list, 1);
3349			if (ret)
3350				goto out;
 
3351
3352			ret = scrub_extent(sctx, extent_logical, extent_len,
3353					   extent_physical, extent_dev, flags,
3354					   generation, extent_mirror_num,
3355					   extent_logical - logical + physical);
3356
3357			scrub_free_csums(sctx);
3358
3359			if (ret)
3360				goto out;
3361
 
 
 
3362			if (extent_logical + extent_len <
3363			    key.objectid + bytes) {
3364				if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3365					/*
3366					 * loop until we find next data stripe
3367					 * or we have finished all stripes.
3368					 */
3369loop:
3370					physical += map->stripe_len;
3371					ret = get_raid56_logic_offset(physical,
3372							num, map, &logical,
3373							&stripe_logical);
3374					logical += base;
3375
3376					if (ret && physical < physical_end) {
3377						stripe_logical += base;
3378						stripe_end = stripe_logical +
3379								increment;
3380						ret = scrub_raid56_parity(sctx,
3381							map, scrub_dev, ppath,
3382							stripe_logical,
3383							stripe_end);
3384						if (ret)
3385							goto out;
3386						goto loop;
3387					}
3388				} else {
3389					physical += map->stripe_len;
3390					logical += increment;
3391				}
3392				if (logical < key.objectid + bytes) {
3393					cond_resched();
3394					goto again;
3395				}
3396
3397				if (physical >= physical_end) {
3398					stop_loop = 1;
3399					break;
3400				}
3401			}
3402next:
3403			path->slots[0]++;
3404		}
3405		btrfs_release_path(path);
3406skip:
3407		logical += increment;
3408		physical += map->stripe_len;
3409		spin_lock(&sctx->stat_lock);
3410		if (stop_loop)
3411			sctx->stat.last_physical = map->stripes[num].physical +
3412						   length;
3413		else
3414			sctx->stat.last_physical = physical;
3415		spin_unlock(&sctx->stat_lock);
3416		if (stop_loop)
3417			break;
3418	}
3419out:
3420	/* push queued extents */
3421	scrub_submit(sctx);
3422	mutex_lock(&sctx->wr_ctx.wr_lock);
3423	scrub_wr_submit(sctx);
3424	mutex_unlock(&sctx->wr_ctx.wr_lock);
3425
3426	blk_finish_plug(&plug);
3427	btrfs_free_path(path);
3428	btrfs_free_path(ppath);
 
 
 
 
 
 
 
 
 
 
 
3429	return ret < 0 ? ret : 0;
3430}
3431
3432static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3433					  struct btrfs_device *scrub_dev,
3434					  u64 chunk_offset, u64 length,
3435					  u64 dev_offset,
3436					  struct btrfs_block_group_cache *cache,
3437					  int is_dev_replace)
3438{
3439	struct btrfs_mapping_tree *map_tree =
3440		&sctx->dev_root->fs_info->mapping_tree;
3441	struct map_lookup *map;
3442	struct extent_map *em;
3443	int i;
3444	int ret = 0;
3445
3446	read_lock(&map_tree->map_tree.lock);
3447	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3448	read_unlock(&map_tree->map_tree.lock);
3449
3450	if (!em) {
3451		/*
3452		 * Might have been an unused block group deleted by the cleaner
3453		 * kthread or relocation.
3454		 */
3455		spin_lock(&cache->lock);
3456		if (!cache->removed)
3457			ret = -EINVAL;
3458		spin_unlock(&cache->lock);
3459
3460		return ret;
3461	}
3462
3463	map = em->map_lookup;
3464	if (em->start != chunk_offset)
3465		goto out;
3466
3467	if (em->len < length)
3468		goto out;
3469
3470	for (i = 0; i < map->num_stripes; ++i) {
3471		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3472		    map->stripes[i].physical == dev_offset) {
3473			ret = scrub_stripe(sctx, map, scrub_dev, i,
3474					   chunk_offset, length,
3475					   is_dev_replace);
3476			if (ret)
3477				goto out;
3478		}
3479	}
3480out:
3481	free_extent_map(em);
3482
3483	return ret;
3484}
3485
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3486static noinline_for_stack
3487int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3488			   struct btrfs_device *scrub_dev, u64 start, u64 end,
3489			   int is_dev_replace)
3490{
3491	struct btrfs_dev_extent *dev_extent = NULL;
3492	struct btrfs_path *path;
3493	struct btrfs_root *root = sctx->dev_root;
3494	struct btrfs_fs_info *fs_info = root->fs_info;
3495	u64 length;
3496	u64 chunk_offset;
3497	int ret = 0;
3498	int ro_set;
3499	int slot;
3500	struct extent_buffer *l;
3501	struct btrfs_key key;
3502	struct btrfs_key found_key;
3503	struct btrfs_block_group_cache *cache;
3504	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3505
3506	path = btrfs_alloc_path();
3507	if (!path)
3508		return -ENOMEM;
3509
3510	path->reada = READA_FORWARD;
3511	path->search_commit_root = 1;
3512	path->skip_locking = 1;
3513
3514	key.objectid = scrub_dev->devid;
3515	key.offset = 0ull;
3516	key.type = BTRFS_DEV_EXTENT_KEY;
3517
3518	while (1) {
3519		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3520		if (ret < 0)
3521			break;
3522		if (ret > 0) {
3523			if (path->slots[0] >=
3524			    btrfs_header_nritems(path->nodes[0])) {
3525				ret = btrfs_next_leaf(root, path);
3526				if (ret < 0)
3527					break;
3528				if (ret > 0) {
3529					ret = 0;
3530					break;
3531				}
3532			} else {
3533				ret = 0;
3534			}
3535		}
3536
3537		l = path->nodes[0];
3538		slot = path->slots[0];
3539
3540		btrfs_item_key_to_cpu(l, &found_key, slot);
3541
3542		if (found_key.objectid != scrub_dev->devid)
3543			break;
3544
3545		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3546			break;
3547
3548		if (found_key.offset >= end)
3549			break;
3550
3551		if (found_key.offset < key.offset)
3552			break;
3553
3554		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3555		length = btrfs_dev_extent_length(l, dev_extent);
3556
3557		if (found_key.offset + length <= start)
3558			goto skip;
3559
3560		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3561
3562		/*
3563		 * get a reference on the corresponding block group to prevent
3564		 * the chunk from going away while we scrub it
3565		 */
3566		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3567
3568		/* some chunks are removed but not committed to disk yet,
3569		 * continue scrubbing */
3570		if (!cache)
3571			goto skip;
3572
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3573		/*
3574		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3575		 * to avoid deadlock caused by:
3576		 * btrfs_inc_block_group_ro()
3577		 * -> btrfs_wait_for_commit()
3578		 * -> btrfs_commit_transaction()
3579		 * -> btrfs_scrub_pause()
3580		 */
3581		scrub_pause_on(fs_info);
3582		ret = btrfs_inc_block_group_ro(root, cache);
3583		scrub_pause_off(fs_info);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3584
3585		if (ret == 0) {
3586			ro_set = 1;
3587		} else if (ret == -ENOSPC) {
3588			/*
3589			 * btrfs_inc_block_group_ro return -ENOSPC when it
3590			 * failed in creating new chunk for metadata.
3591			 * It is not a problem for scrub/replace, because
3592			 * metadata are always cowed, and our scrub paused
3593			 * commit_transactions.
3594			 */
3595			ro_set = 0;
 
 
 
 
 
 
 
3596		} else {
3597			btrfs_warn(fs_info, "failed setting block group ro, ret=%d\n",
3598				   ret);
 
3599			btrfs_put_block_group(cache);
 
3600			break;
3601		}
3602
 
 
 
 
 
 
 
 
 
 
 
 
 
3603		dev_replace->cursor_right = found_key.offset + length;
3604		dev_replace->cursor_left = found_key.offset;
3605		dev_replace->item_needs_writeback = 1;
 
 
3606		ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3607				  found_key.offset, cache, is_dev_replace);
3608
3609		/*
3610		 * flush, submit all pending read and write bios, afterwards
3611		 * wait for them.
3612		 * Note that in the dev replace case, a read request causes
3613		 * write requests that are submitted in the read completion
3614		 * worker. Therefore in the current situation, it is required
3615		 * that all write requests are flushed, so that all read and
3616		 * write requests are really completed when bios_in_flight
3617		 * changes to 0.
3618		 */
3619		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3620		scrub_submit(sctx);
3621		mutex_lock(&sctx->wr_ctx.wr_lock);
3622		scrub_wr_submit(sctx);
3623		mutex_unlock(&sctx->wr_ctx.wr_lock);
3624
3625		wait_event(sctx->list_wait,
3626			   atomic_read(&sctx->bios_in_flight) == 0);
3627
3628		scrub_pause_on(fs_info);
3629
3630		/*
3631		 * must be called before we decrease @scrub_paused.
3632		 * make sure we don't block transaction commit while
3633		 * we are waiting pending workers finished.
3634		 */
3635		wait_event(sctx->list_wait,
3636			   atomic_read(&sctx->workers_pending) == 0);
3637		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3638
3639		scrub_pause_off(fs_info);
3640
 
 
 
 
 
 
 
 
 
 
3641		if (ro_set)
3642			btrfs_dec_block_group_ro(root, cache);
3643
3644		/*
3645		 * We might have prevented the cleaner kthread from deleting
3646		 * this block group if it was already unused because we raced
3647		 * and set it to RO mode first. So add it back to the unused
3648		 * list, otherwise it might not ever be deleted unless a manual
3649		 * balance is triggered or it becomes used and unused again.
3650		 */
3651		spin_lock(&cache->lock);
3652		if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3653		    btrfs_block_group_used(&cache->item) == 0) {
3654			spin_unlock(&cache->lock);
3655			spin_lock(&fs_info->unused_bgs_lock);
3656			if (list_empty(&cache->bg_list)) {
3657				btrfs_get_block_group(cache);
3658				list_add_tail(&cache->bg_list,
3659					      &fs_info->unused_bgs);
3660			}
3661			spin_unlock(&fs_info->unused_bgs_lock);
3662		} else {
3663			spin_unlock(&cache->lock);
3664		}
3665
 
3666		btrfs_put_block_group(cache);
3667		if (ret)
3668			break;
3669		if (is_dev_replace &&
3670		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3671			ret = -EIO;
3672			break;
3673		}
3674		if (sctx->stat.malloc_errors > 0) {
3675			ret = -ENOMEM;
3676			break;
3677		}
3678
3679		dev_replace->cursor_left = dev_replace->cursor_right;
3680		dev_replace->item_needs_writeback = 1;
3681skip:
3682		key.offset = found_key.offset + length;
3683		btrfs_release_path(path);
3684	}
3685
3686	btrfs_free_path(path);
3687
3688	return ret;
3689}
3690
3691static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3692					   struct btrfs_device *scrub_dev)
3693{
3694	int	i;
3695	u64	bytenr;
3696	u64	gen;
3697	int	ret;
3698	struct btrfs_root *root = sctx->dev_root;
3699
3700	if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
3701		return -EIO;
3702
3703	/* Seed devices of a new filesystem has their own generation. */
3704	if (scrub_dev->fs_devices != root->fs_info->fs_devices)
3705		gen = scrub_dev->generation;
3706	else
3707		gen = root->fs_info->last_trans_committed;
3708
3709	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3710		bytenr = btrfs_sb_offset(i);
3711		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3712		    scrub_dev->commit_total_bytes)
3713			break;
 
 
3714
3715		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3716				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3717				  NULL, 1, bytenr);
3718		if (ret)
3719			return ret;
3720	}
3721	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3722
3723	return 0;
3724}
3725
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3726/*
3727 * get a reference count on fs_info->scrub_workers. start worker if necessary
3728 */
3729static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3730						int is_dev_replace)
3731{
 
 
 
3732	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3733	int max_active = fs_info->thread_pool_size;
 
3734
3735	if (fs_info->scrub_workers_refcnt == 0) {
3736		if (is_dev_replace)
3737			fs_info->scrub_workers =
3738				btrfs_alloc_workqueue("scrub", flags,
3739						      1, 4);
3740		else
3741			fs_info->scrub_workers =
3742				btrfs_alloc_workqueue("scrub", flags,
3743						      max_active, 4);
3744		if (!fs_info->scrub_workers)
3745			goto fail_scrub_workers;
3746
3747		fs_info->scrub_wr_completion_workers =
3748			btrfs_alloc_workqueue("scrubwrc", flags,
3749					      max_active, 2);
3750		if (!fs_info->scrub_wr_completion_workers)
3751			goto fail_scrub_wr_completion_workers;
3752
3753		fs_info->scrub_nocow_workers =
3754			btrfs_alloc_workqueue("scrubnc", flags, 1, 0);
3755		if (!fs_info->scrub_nocow_workers)
3756			goto fail_scrub_nocow_workers;
3757		fs_info->scrub_parity_workers =
3758			btrfs_alloc_workqueue("scrubparity", flags,
3759					      max_active, 2);
3760		if (!fs_info->scrub_parity_workers)
3761			goto fail_scrub_parity_workers;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3762	}
3763	++fs_info->scrub_workers_refcnt;
3764	return 0;
 
3765
 
 
3766fail_scrub_parity_workers:
3767	btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3768fail_scrub_nocow_workers:
3769	btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3770fail_scrub_wr_completion_workers:
3771	btrfs_destroy_workqueue(fs_info->scrub_workers);
3772fail_scrub_workers:
3773	return -ENOMEM;
3774}
3775
3776static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3777{
3778	if (--fs_info->scrub_workers_refcnt == 0) {
3779		btrfs_destroy_workqueue(fs_info->scrub_workers);
3780		btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3781		btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3782		btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3783	}
3784	WARN_ON(fs_info->scrub_workers_refcnt < 0);
3785}
3786
3787int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3788		    u64 end, struct btrfs_scrub_progress *progress,
3789		    int readonly, int is_dev_replace)
3790{
3791	struct scrub_ctx *sctx;
3792	int ret;
3793	struct btrfs_device *dev;
3794	struct rcu_string *name;
3795
3796	if (btrfs_fs_closing(fs_info))
3797		return -EINVAL;
3798
3799	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
3800		/*
3801		 * in this case scrub is unable to calculate the checksum
3802		 * the way scrub is implemented. Do not handle this
3803		 * situation at all because it won't ever happen.
3804		 */
3805		btrfs_err(fs_info,
3806			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3807		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
 
3808		return -EINVAL;
3809	}
3810
3811	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
3812		/* not supported for data w/o checksums */
3813		btrfs_err(fs_info,
3814			   "scrub: size assumption sectorsize != PAGE_SIZE "
3815			   "(%d != %lu) fails",
3816		       fs_info->chunk_root->sectorsize, PAGE_SIZE);
3817		return -EINVAL;
3818	}
3819
3820	if (fs_info->chunk_root->nodesize >
3821	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3822	    fs_info->chunk_root->sectorsize >
3823	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3824		/*
3825		 * would exhaust the array bounds of pagev member in
3826		 * struct scrub_block
3827		 */
3828		btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
3829			   "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3830		       fs_info->chunk_root->nodesize,
3831		       SCRUB_MAX_PAGES_PER_BLOCK,
3832		       fs_info->chunk_root->sectorsize,
3833		       SCRUB_MAX_PAGES_PER_BLOCK);
3834		return -EINVAL;
3835	}
3836
 
 
 
 
 
 
 
 
3837
3838	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3839	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3840	if (!dev || (dev->missing && !is_dev_replace)) {
 
3841		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3842		return -ENODEV;
 
3843	}
3844
3845	if (!is_dev_replace && !readonly && !dev->writeable) {
 
3846		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3847		rcu_read_lock();
3848		name = rcu_dereference(dev->name);
3849		btrfs_err(fs_info, "scrub: device %s is not writable",
3850			  name->str);
3851		rcu_read_unlock();
3852		return -EROFS;
3853	}
3854
3855	mutex_lock(&fs_info->scrub_lock);
3856	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
 
3857		mutex_unlock(&fs_info->scrub_lock);
3858		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3859		return -EIO;
 
3860	}
3861
3862	btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
3863	if (dev->scrub_device ||
3864	    (!is_dev_replace &&
3865	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3866		btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3867		mutex_unlock(&fs_info->scrub_lock);
3868		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3869		return -EINPROGRESS;
3870	}
3871	btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3872
3873	ret = scrub_workers_get(fs_info, is_dev_replace);
3874	if (ret) {
3875		mutex_unlock(&fs_info->scrub_lock);
3876		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3877		return ret;
3878	}
 
3879
3880	sctx = scrub_setup_ctx(dev, is_dev_replace);
3881	if (IS_ERR(sctx)) {
3882		mutex_unlock(&fs_info->scrub_lock);
3883		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3884		scrub_workers_put(fs_info);
3885		return PTR_ERR(sctx);
3886	}
3887	sctx->readonly = readonly;
3888	dev->scrub_device = sctx;
3889	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3890
3891	/*
3892	 * checking @scrub_pause_req here, we can avoid
3893	 * race between committing transaction and scrubbing.
3894	 */
3895	__scrub_blocked_if_needed(fs_info);
3896	atomic_inc(&fs_info->scrubs_running);
3897	mutex_unlock(&fs_info->scrub_lock);
3898
 
 
 
 
 
 
 
 
 
 
3899	if (!is_dev_replace) {
 
3900		/*
3901		 * by holding device list mutex, we can
3902		 * kick off writing super in log tree sync.
3903		 */
3904		mutex_lock(&fs_info->fs_devices->device_list_mutex);
3905		ret = scrub_supers(sctx, dev);
3906		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3907	}
3908
3909	if (!ret)
3910		ret = scrub_enumerate_chunks(sctx, dev, start, end,
3911					     is_dev_replace);
3912
3913	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3914	atomic_dec(&fs_info->scrubs_running);
3915	wake_up(&fs_info->scrub_pause_wait);
3916
3917	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3918
3919	if (progress)
3920		memcpy(progress, &sctx->stat, sizeof(*progress));
3921
 
 
 
 
3922	mutex_lock(&fs_info->scrub_lock);
3923	dev->scrub_device = NULL;
3924	scrub_workers_put(fs_info);
3925	mutex_unlock(&fs_info->scrub_lock);
3926
 
3927	scrub_put_ctx(sctx);
3928
3929	return ret;
 
 
 
 
 
 
3930}
3931
3932void btrfs_scrub_pause(struct btrfs_root *root)
3933{
3934	struct btrfs_fs_info *fs_info = root->fs_info;
3935
3936	mutex_lock(&fs_info->scrub_lock);
3937	atomic_inc(&fs_info->scrub_pause_req);
3938	while (atomic_read(&fs_info->scrubs_paused) !=
3939	       atomic_read(&fs_info->scrubs_running)) {
3940		mutex_unlock(&fs_info->scrub_lock);
3941		wait_event(fs_info->scrub_pause_wait,
3942			   atomic_read(&fs_info->scrubs_paused) ==
3943			   atomic_read(&fs_info->scrubs_running));
3944		mutex_lock(&fs_info->scrub_lock);
3945	}
3946	mutex_unlock(&fs_info->scrub_lock);
3947}
3948
3949void btrfs_scrub_continue(struct btrfs_root *root)
3950{
3951	struct btrfs_fs_info *fs_info = root->fs_info;
3952
3953	atomic_dec(&fs_info->scrub_pause_req);
3954	wake_up(&fs_info->scrub_pause_wait);
3955}
3956
3957int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3958{
3959	mutex_lock(&fs_info->scrub_lock);
3960	if (!atomic_read(&fs_info->scrubs_running)) {
3961		mutex_unlock(&fs_info->scrub_lock);
3962		return -ENOTCONN;
3963	}
3964
3965	atomic_inc(&fs_info->scrub_cancel_req);
3966	while (atomic_read(&fs_info->scrubs_running)) {
3967		mutex_unlock(&fs_info->scrub_lock);
3968		wait_event(fs_info->scrub_pause_wait,
3969			   atomic_read(&fs_info->scrubs_running) == 0);
3970		mutex_lock(&fs_info->scrub_lock);
3971	}
3972	atomic_dec(&fs_info->scrub_cancel_req);
3973	mutex_unlock(&fs_info->scrub_lock);
3974
3975	return 0;
3976}
3977
3978int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3979			   struct btrfs_device *dev)
3980{
 
3981	struct scrub_ctx *sctx;
3982
3983	mutex_lock(&fs_info->scrub_lock);
3984	sctx = dev->scrub_device;
3985	if (!sctx) {
3986		mutex_unlock(&fs_info->scrub_lock);
3987		return -ENOTCONN;
3988	}
3989	atomic_inc(&sctx->cancel_req);
3990	while (dev->scrub_device) {
3991		mutex_unlock(&fs_info->scrub_lock);
3992		wait_event(fs_info->scrub_pause_wait,
3993			   dev->scrub_device == NULL);
3994		mutex_lock(&fs_info->scrub_lock);
3995	}
3996	mutex_unlock(&fs_info->scrub_lock);
3997
3998	return 0;
3999}
4000
4001int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
4002			 struct btrfs_scrub_progress *progress)
4003{
4004	struct btrfs_device *dev;
4005	struct scrub_ctx *sctx = NULL;
4006
4007	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
4008	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
4009	if (dev)
4010		sctx = dev->scrub_device;
4011	if (sctx)
4012		memcpy(progress, &sctx->stat, sizeof(*progress));
4013	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
4014
4015	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4016}
4017
4018static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4019			       u64 extent_logical, u64 extent_len,
4020			       u64 *extent_physical,
4021			       struct btrfs_device **extent_dev,
4022			       int *extent_mirror_num)
4023{
4024	u64 mapped_length;
4025	struct btrfs_bio *bbio = NULL;
4026	int ret;
4027
4028	mapped_length = extent_len;
4029	ret = btrfs_map_block(fs_info, READ, extent_logical,
4030			      &mapped_length, &bbio, 0);
4031	if (ret || !bbio || mapped_length < extent_len ||
4032	    !bbio->stripes[0].dev->bdev) {
4033		btrfs_put_bbio(bbio);
4034		return;
4035	}
4036
4037	*extent_physical = bbio->stripes[0].physical;
4038	*extent_mirror_num = bbio->mirror_num;
4039	*extent_dev = bbio->stripes[0].dev;
4040	btrfs_put_bbio(bbio);
4041}
4042
4043static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
4044			      struct scrub_wr_ctx *wr_ctx,
4045			      struct btrfs_fs_info *fs_info,
4046			      struct btrfs_device *dev,
4047			      int is_dev_replace)
4048{
4049	WARN_ON(wr_ctx->wr_curr_bio != NULL);
4050
4051	mutex_init(&wr_ctx->wr_lock);
4052	wr_ctx->wr_curr_bio = NULL;
4053	if (!is_dev_replace)
4054		return 0;
4055
4056	WARN_ON(!dev->bdev);
4057	wr_ctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
4058	wr_ctx->tgtdev = dev;
4059	atomic_set(&wr_ctx->flush_all_writes, 0);
4060	return 0;
4061}
4062
4063static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
4064{
4065	mutex_lock(&wr_ctx->wr_lock);
4066	kfree(wr_ctx->wr_curr_bio);
4067	wr_ctx->wr_curr_bio = NULL;
4068	mutex_unlock(&wr_ctx->wr_lock);
4069}
4070
4071static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4072			    int mirror_num, u64 physical_for_dev_replace)
4073{
4074	struct scrub_copy_nocow_ctx *nocow_ctx;
4075	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
4076
4077	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4078	if (!nocow_ctx) {
4079		spin_lock(&sctx->stat_lock);
4080		sctx->stat.malloc_errors++;
4081		spin_unlock(&sctx->stat_lock);
4082		return -ENOMEM;
4083	}
4084
4085	scrub_pending_trans_workers_inc(sctx);
4086
4087	nocow_ctx->sctx = sctx;
4088	nocow_ctx->logical = logical;
4089	nocow_ctx->len = len;
4090	nocow_ctx->mirror_num = mirror_num;
4091	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4092	btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4093			copy_nocow_pages_worker, NULL, NULL);
4094	INIT_LIST_HEAD(&nocow_ctx->inodes);
4095	btrfs_queue_work(fs_info->scrub_nocow_workers,
4096			 &nocow_ctx->work);
4097
4098	return 0;
4099}
4100
4101static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4102{
4103	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4104	struct scrub_nocow_inode *nocow_inode;
4105
4106	nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4107	if (!nocow_inode)
4108		return -ENOMEM;
4109	nocow_inode->inum = inum;
4110	nocow_inode->offset = offset;
4111	nocow_inode->root = root;
4112	list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4113	return 0;
4114}
4115
4116#define COPY_COMPLETE 1
4117
4118static void copy_nocow_pages_worker(struct btrfs_work *work)
4119{
4120	struct scrub_copy_nocow_ctx *nocow_ctx =
4121		container_of(work, struct scrub_copy_nocow_ctx, work);
4122	struct scrub_ctx *sctx = nocow_ctx->sctx;
4123	u64 logical = nocow_ctx->logical;
4124	u64 len = nocow_ctx->len;
4125	int mirror_num = nocow_ctx->mirror_num;
4126	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4127	int ret;
4128	struct btrfs_trans_handle *trans = NULL;
4129	struct btrfs_fs_info *fs_info;
4130	struct btrfs_path *path;
4131	struct btrfs_root *root;
4132	int not_written = 0;
4133
4134	fs_info = sctx->dev_root->fs_info;
4135	root = fs_info->extent_root;
4136
4137	path = btrfs_alloc_path();
4138	if (!path) {
4139		spin_lock(&sctx->stat_lock);
4140		sctx->stat.malloc_errors++;
4141		spin_unlock(&sctx->stat_lock);
4142		not_written = 1;
4143		goto out;
4144	}
4145
4146	trans = btrfs_join_transaction(root);
4147	if (IS_ERR(trans)) {
4148		not_written = 1;
4149		goto out;
4150	}
4151
4152	ret = iterate_inodes_from_logical(logical, fs_info, path,
4153					  record_inode_for_nocow, nocow_ctx);
4154	if (ret != 0 && ret != -ENOENT) {
4155		btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
4156			"phys %llu, len %llu, mir %u, ret %d",
4157			logical, physical_for_dev_replace, len, mirror_num,
4158			ret);
4159		not_written = 1;
4160		goto out;
4161	}
4162
4163	btrfs_end_transaction(trans, root);
4164	trans = NULL;
4165	while (!list_empty(&nocow_ctx->inodes)) {
4166		struct scrub_nocow_inode *entry;
4167		entry = list_first_entry(&nocow_ctx->inodes,
4168					 struct scrub_nocow_inode,
4169					 list);
4170		list_del_init(&entry->list);
4171		ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4172						 entry->root, nocow_ctx);
4173		kfree(entry);
4174		if (ret == COPY_COMPLETE) {
4175			ret = 0;
4176			break;
4177		} else if (ret) {
4178			break;
4179		}
4180	}
4181out:
4182	while (!list_empty(&nocow_ctx->inodes)) {
4183		struct scrub_nocow_inode *entry;
4184		entry = list_first_entry(&nocow_ctx->inodes,
4185					 struct scrub_nocow_inode,
4186					 list);
4187		list_del_init(&entry->list);
4188		kfree(entry);
4189	}
4190	if (trans && !IS_ERR(trans))
4191		btrfs_end_transaction(trans, root);
4192	if (not_written)
4193		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4194					    num_uncorrectable_read_errors);
4195
4196	btrfs_free_path(path);
4197	kfree(nocow_ctx);
4198
4199	scrub_pending_trans_workers_dec(sctx);
4200}
4201
4202static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
4203				 u64 logical)
4204{
4205	struct extent_state *cached_state = NULL;
4206	struct btrfs_ordered_extent *ordered;
4207	struct extent_io_tree *io_tree;
4208	struct extent_map *em;
4209	u64 lockstart = start, lockend = start + len - 1;
4210	int ret = 0;
4211
4212	io_tree = &BTRFS_I(inode)->io_tree;
4213
4214	lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4215	ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4216	if (ordered) {
4217		btrfs_put_ordered_extent(ordered);
4218		ret = 1;
4219		goto out_unlock;
4220	}
4221
4222	em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4223	if (IS_ERR(em)) {
4224		ret = PTR_ERR(em);
4225		goto out_unlock;
4226	}
4227
4228	/*
4229	 * This extent does not actually cover the logical extent anymore,
4230	 * move on to the next inode.
4231	 */
4232	if (em->block_start > logical ||
4233	    em->block_start + em->block_len < logical + len) {
4234		free_extent_map(em);
4235		ret = 1;
4236		goto out_unlock;
4237	}
4238	free_extent_map(em);
4239
4240out_unlock:
4241	unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4242			     GFP_NOFS);
4243	return ret;
4244}
4245
4246static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4247				      struct scrub_copy_nocow_ctx *nocow_ctx)
4248{
4249	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
4250	struct btrfs_key key;
4251	struct inode *inode;
4252	struct page *page;
4253	struct btrfs_root *local_root;
4254	struct extent_io_tree *io_tree;
4255	u64 physical_for_dev_replace;
4256	u64 nocow_ctx_logical;
4257	u64 len = nocow_ctx->len;
4258	unsigned long index;
4259	int srcu_index;
4260	int ret = 0;
4261	int err = 0;
4262
4263	key.objectid = root;
4264	key.type = BTRFS_ROOT_ITEM_KEY;
4265	key.offset = (u64)-1;
4266
4267	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4268
4269	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4270	if (IS_ERR(local_root)) {
4271		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4272		return PTR_ERR(local_root);
4273	}
4274
4275	key.type = BTRFS_INODE_ITEM_KEY;
4276	key.objectid = inum;
4277	key.offset = 0;
4278	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4279	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4280	if (IS_ERR(inode))
4281		return PTR_ERR(inode);
4282
4283	/* Avoid truncate/dio/punch hole.. */
4284	inode_lock(inode);
4285	inode_dio_wait(inode);
4286
4287	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4288	io_tree = &BTRFS_I(inode)->io_tree;
4289	nocow_ctx_logical = nocow_ctx->logical;
4290
4291	ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
4292	if (ret) {
4293		ret = ret > 0 ? 0 : ret;
4294		goto out;
4295	}
4296
4297	while (len >= PAGE_SIZE) {
4298		index = offset >> PAGE_SHIFT;
4299again:
4300		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4301		if (!page) {
4302			btrfs_err(fs_info, "find_or_create_page() failed");
4303			ret = -ENOMEM;
4304			goto out;
4305		}
4306
4307		if (PageUptodate(page)) {
4308			if (PageDirty(page))
4309				goto next_page;
4310		} else {
4311			ClearPageError(page);
4312			err = extent_read_full_page(io_tree, page,
4313							   btrfs_get_extent,
4314							   nocow_ctx->mirror_num);
4315			if (err) {
4316				ret = err;
4317				goto next_page;
4318			}
4319
4320			lock_page(page);
4321			/*
4322			 * If the page has been remove from the page cache,
4323			 * the data on it is meaningless, because it may be
4324			 * old one, the new data may be written into the new
4325			 * page in the page cache.
4326			 */
4327			if (page->mapping != inode->i_mapping) {
4328				unlock_page(page);
4329				put_page(page);
4330				goto again;
4331			}
4332			if (!PageUptodate(page)) {
4333				ret = -EIO;
4334				goto next_page;
4335			}
4336		}
4337
4338		ret = check_extent_to_block(inode, offset, len,
4339					    nocow_ctx_logical);
4340		if (ret) {
4341			ret = ret > 0 ? 0 : ret;
4342			goto next_page;
4343		}
4344
4345		err = write_page_nocow(nocow_ctx->sctx,
4346				       physical_for_dev_replace, page);
4347		if (err)
4348			ret = err;
4349next_page:
4350		unlock_page(page);
4351		put_page(page);
4352
4353		if (ret)
4354			break;
4355
4356		offset += PAGE_SIZE;
4357		physical_for_dev_replace += PAGE_SIZE;
4358		nocow_ctx_logical += PAGE_SIZE;
4359		len -= PAGE_SIZE;
4360	}
4361	ret = COPY_COMPLETE;
4362out:
4363	inode_unlock(inode);
4364	iput(inode);
4365	return ret;
4366}
4367
4368static int write_page_nocow(struct scrub_ctx *sctx,
4369			    u64 physical_for_dev_replace, struct page *page)
4370{
4371	struct bio *bio;
4372	struct btrfs_device *dev;
4373	int ret;
4374
4375	dev = sctx->wr_ctx.tgtdev;
4376	if (!dev)
4377		return -EIO;
4378	if (!dev->bdev) {
4379		btrfs_warn_rl(dev->dev_root->fs_info,
4380			"scrub write_page_nocow(bdev == NULL) is unexpected");
4381		return -EIO;
4382	}
4383	bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4384	if (!bio) {
4385		spin_lock(&sctx->stat_lock);
4386		sctx->stat.malloc_errors++;
4387		spin_unlock(&sctx->stat_lock);
4388		return -ENOMEM;
4389	}
4390	bio->bi_iter.bi_size = 0;
4391	bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4392	bio->bi_bdev = dev->bdev;
4393	ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4394	if (ret != PAGE_SIZE) {
4395leave_with_eio:
4396		bio_put(bio);
4397		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4398		return -EIO;
4399	}
4400
4401	if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
4402		goto leave_with_eio;
4403
4404	bio_put(bio);
4405	return 0;
4406}
v5.14.15
   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#include "zoned.h"
  24
  25/*
  26 * This is only the first step towards a full-features scrub. It reads all
  27 * extent and super block and verifies the checksums. In case a bad checksum
  28 * is found or the extent cannot be read, good data will be written back if
  29 * any can be found.
  30 *
  31 * Future enhancements:
  32 *  - In case an unrepairable extent is encountered, track which files are
  33 *    affected and report them
  34 *  - track and record media errors, throw out bad devices
  35 *  - add a mode to also read unallocated space
  36 */
  37
  38struct scrub_block;
  39struct scrub_ctx;
  40
  41/*
  42 * the following three values only influence the performance.
  43 * The last one configures the number of parallel and outstanding I/O
  44 * operations. The first two values configure an upper limit for the number
  45 * of (dynamically allocated) pages that are added to a bio.
  46 */
  47#define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
  48#define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
  49#define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
  50
  51/*
  52 * the following value times PAGE_SIZE needs to be large enough to match the
  53 * largest node/leaf/sector size that shall be supported.
  54 * Values larger than BTRFS_STRIPE_LEN are not supported.
  55 */
  56#define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
  57
  58struct scrub_recover {
  59	refcount_t		refs;
  60	struct btrfs_bio	*bbio;
  61	u64			map_length;
  62};
  63
  64struct scrub_page {
  65	struct scrub_block	*sblock;
  66	struct page		*page;
  67	struct btrfs_device	*dev;
  68	struct list_head	list;
  69	u64			flags;  /* extent flags */
  70	u64			generation;
  71	u64			logical;
  72	u64			physical;
  73	u64			physical_for_dev_replace;
  74	atomic_t		refs;
  75	u8			mirror_num;
  76	int			have_csum:1;
  77	int			io_error:1;
 
 
  78	u8			csum[BTRFS_CSUM_SIZE];
  79
  80	struct scrub_recover	*recover;
  81};
  82
  83struct scrub_bio {
  84	int			index;
  85	struct scrub_ctx	*sctx;
  86	struct btrfs_device	*dev;
  87	struct bio		*bio;
  88	blk_status_t		status;
  89	u64			logical;
  90	u64			physical;
  91#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  92	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
  93#else
  94	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
  95#endif
  96	int			page_count;
  97	int			next_free;
  98	struct btrfs_work	work;
  99};
 100
 101struct scrub_block {
 102	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 103	int			page_count;
 104	atomic_t		outstanding_pages;
 105	refcount_t		refs; /* free mem on transition to zero */
 106	struct scrub_ctx	*sctx;
 107	struct scrub_parity	*sparity;
 108	struct {
 109		unsigned int	header_error:1;
 110		unsigned int	checksum_error:1;
 111		unsigned int	no_io_error_seen:1;
 112		unsigned int	generation_error:1; /* also sets header_error */
 113
 114		/* The following is for the data used to check parity */
 115		/* It is for the data with checksum */
 116		unsigned int	data_corrected:1;
 117	};
 118	struct btrfs_work	work;
 119};
 120
 121/* Used for the chunks with parity stripe such RAID5/6 */
 122struct scrub_parity {
 123	struct scrub_ctx	*sctx;
 124
 125	struct btrfs_device	*scrub_dev;
 126
 127	u64			logic_start;
 128
 129	u64			logic_end;
 130
 131	int			nsectors;
 132
 133	u32			stripe_len;
 134
 135	refcount_t		refs;
 136
 137	struct list_head	spages;
 138
 139	/* Work of parity check and repair */
 140	struct btrfs_work	work;
 141
 142	/* Mark the parity blocks which have data */
 143	unsigned long		*dbitmap;
 144
 145	/*
 146	 * Mark the parity blocks which have data, but errors happen when
 147	 * read data or check data
 148	 */
 149	unsigned long		*ebitmap;
 150
 151	unsigned long		bitmap[];
 
 
 
 
 
 
 
 
 152};
 153
 154struct scrub_ctx {
 155	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
 156	struct btrfs_fs_info	*fs_info;
 157	int			first_free;
 158	int			curr;
 159	atomic_t		bios_in_flight;
 160	atomic_t		workers_pending;
 161	spinlock_t		list_lock;
 162	wait_queue_head_t	list_wait;
 
 163	struct list_head	csum_list;
 164	atomic_t		cancel_req;
 165	int			readonly;
 166	int			pages_per_rd_bio;
 167
 168	/* State of IO submission throttling affecting the associated device */
 169	ktime_t			throttle_deadline;
 170	u64			throttle_sent;
 171
 172	int			is_dev_replace;
 173	u64			write_pointer;
 174
 175	struct scrub_bio        *wr_curr_bio;
 176	struct mutex            wr_lock;
 177	int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 178	struct btrfs_device     *wr_tgtdev;
 179	bool                    flush_all_writes;
 180
 181	/*
 182	 * statistics
 183	 */
 184	struct btrfs_scrub_progress stat;
 185	spinlock_t		stat_lock;
 186
 187	/*
 188	 * Use a ref counter to avoid use-after-free issues. Scrub workers
 189	 * decrement bios_in_flight and workers_pending and then do a wakeup
 190	 * on the list_wait wait queue. We must ensure the main scrub task
 191	 * doesn't free the scrub context before or while the workers are
 192	 * doing the wakeup() call.
 193	 */
 194	refcount_t              refs;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 195};
 196
 197struct scrub_warning {
 198	struct btrfs_path	*path;
 199	u64			extent_item_size;
 200	const char		*errstr;
 201	u64			physical;
 202	u64			logical;
 203	struct btrfs_device	*dev;
 204};
 205
 206struct full_stripe_lock {
 207	struct rb_node node;
 208	u64 logical;
 209	u64 refs;
 210	struct mutex mutex;
 211};
 212
 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_put(struct scrub_block *sblock);
 231static void scrub_page_get(struct scrub_page *spage);
 232static void scrub_page_put(struct scrub_page *spage);
 233static void scrub_parity_get(struct scrub_parity *sparity);
 234static void scrub_parity_put(struct scrub_parity *sparity);
 235static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u32 len,
 
 
 236		       u64 physical, struct btrfs_device *dev, u64 flags,
 237		       u64 gen, int mirror_num, u8 *csum,
 238		       u64 physical_for_dev_replace);
 239static void scrub_bio_end_io(struct bio *bio);
 240static void scrub_bio_end_io_worker(struct btrfs_work *work);
 241static void scrub_block_complete(struct scrub_block *sblock);
 242static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 243			       u64 extent_logical, u32 extent_len,
 244			       u64 *extent_physical,
 245			       struct btrfs_device **extent_dev,
 246			       int *extent_mirror_num);
 
 
 
 
 
 
 247static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 248				    struct scrub_page *spage);
 249static void scrub_wr_submit(struct scrub_ctx *sctx);
 250static void scrub_wr_bio_end_io(struct bio *bio);
 251static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 
 
 
 
 
 
 
 
 
 252static void scrub_put_ctx(struct scrub_ctx *sctx);
 253
 254static inline int scrub_is_page_on_raid56(struct scrub_page *spage)
 255{
 256	return spage->recover &&
 257	       (spage->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 258}
 259
 260static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 261{
 262	refcount_inc(&sctx->refs);
 263	atomic_inc(&sctx->bios_in_flight);
 264}
 265
 266static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 267{
 268	atomic_dec(&sctx->bios_in_flight);
 269	wake_up(&sctx->list_wait);
 270	scrub_put_ctx(sctx);
 271}
 272
 273static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 274{
 275	while (atomic_read(&fs_info->scrub_pause_req)) {
 276		mutex_unlock(&fs_info->scrub_lock);
 277		wait_event(fs_info->scrub_pause_wait,
 278		   atomic_read(&fs_info->scrub_pause_req) == 0);
 279		mutex_lock(&fs_info->scrub_lock);
 280	}
 281}
 282
 283static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 284{
 285	atomic_inc(&fs_info->scrubs_paused);
 286	wake_up(&fs_info->scrub_pause_wait);
 287}
 288
 289static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 290{
 291	mutex_lock(&fs_info->scrub_lock);
 292	__scrub_blocked_if_needed(fs_info);
 293	atomic_dec(&fs_info->scrubs_paused);
 294	mutex_unlock(&fs_info->scrub_lock);
 295
 296	wake_up(&fs_info->scrub_pause_wait);
 297}
 298
 299static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 300{
 301	scrub_pause_on(fs_info);
 302	scrub_pause_off(fs_info);
 303}
 304
 305/*
 306 * Insert new full stripe lock into full stripe locks tree
 307 *
 308 * Return pointer to existing or newly inserted full_stripe_lock structure if
 309 * everything works well.
 310 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 311 *
 312 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 313 * function
 314 */
 315static struct full_stripe_lock *insert_full_stripe_lock(
 316		struct btrfs_full_stripe_locks_tree *locks_root,
 317		u64 fstripe_logical)
 318{
 319	struct rb_node **p;
 320	struct rb_node *parent = NULL;
 321	struct full_stripe_lock *entry;
 322	struct full_stripe_lock *ret;
 323
 324	lockdep_assert_held(&locks_root->lock);
 325
 326	p = &locks_root->root.rb_node;
 327	while (*p) {
 328		parent = *p;
 329		entry = rb_entry(parent, struct full_stripe_lock, node);
 330		if (fstripe_logical < entry->logical) {
 331			p = &(*p)->rb_left;
 332		} else if (fstripe_logical > entry->logical) {
 333			p = &(*p)->rb_right;
 334		} else {
 335			entry->refs++;
 336			return entry;
 337		}
 338	}
 339
 340	/*
 341	 * Insert new lock.
 342	 */
 343	ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 344	if (!ret)
 345		return ERR_PTR(-ENOMEM);
 346	ret->logical = fstripe_logical;
 347	ret->refs = 1;
 348	mutex_init(&ret->mutex);
 349
 350	rb_link_node(&ret->node, parent, p);
 351	rb_insert_color(&ret->node, &locks_root->root);
 352	return ret;
 353}
 354
 355/*
 356 * Search for a full stripe lock of a block group
 357 *
 358 * Return pointer to existing full stripe lock if found
 359 * Return NULL if not found
 360 */
 361static struct full_stripe_lock *search_full_stripe_lock(
 362		struct btrfs_full_stripe_locks_tree *locks_root,
 363		u64 fstripe_logical)
 364{
 365	struct rb_node *node;
 366	struct full_stripe_lock *entry;
 367
 368	lockdep_assert_held(&locks_root->lock);
 369
 370	node = locks_root->root.rb_node;
 371	while (node) {
 372		entry = rb_entry(node, struct full_stripe_lock, node);
 373		if (fstripe_logical < entry->logical)
 374			node = node->rb_left;
 375		else if (fstripe_logical > entry->logical)
 376			node = node->rb_right;
 377		else
 378			return entry;
 379	}
 380	return NULL;
 381}
 382
 383/*
 384 * Helper to get full stripe logical from a normal bytenr.
 385 *
 386 * Caller must ensure @cache is a RAID56 block group.
 387 */
 388static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
 389{
 390	u64 ret;
 391
 
 392	/*
 393	 * Due to chunk item size limit, full stripe length should not be
 394	 * larger than U32_MAX. Just a sanity check here.
 
 
 
 
 
 395	 */
 396	WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 
 
 
 397
 398	/*
 399	 * round_down() can only handle power of 2, while RAID56 full
 400	 * stripe length can be 64KiB * n, so we need to manually round down.
 
 
 
 401	 */
 402	ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 403			cache->full_stripe_len + cache->start;
 404	return ret;
 405}
 406
 407/*
 408 * Lock a full stripe to avoid concurrency of recovery and read
 409 *
 410 * It's only used for profiles with parities (RAID5/6), for other profiles it
 411 * does nothing.
 412 *
 413 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 414 * So caller must call unlock_full_stripe() at the same context.
 415 *
 416 * Return <0 if encounters error.
 417 */
 418static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 419			    bool *locked_ret)
 420{
 421	struct btrfs_block_group *bg_cache;
 422	struct btrfs_full_stripe_locks_tree *locks_root;
 423	struct full_stripe_lock *existing;
 424	u64 fstripe_start;
 425	int ret = 0;
 426
 427	*locked_ret = false;
 428	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 429	if (!bg_cache) {
 430		ASSERT(0);
 431		return -ENOENT;
 432	}
 433
 434	/* Profiles not based on parity don't need full stripe lock */
 435	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 436		goto out;
 437	locks_root = &bg_cache->full_stripe_locks_root;
 438
 439	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 440
 441	/* Now insert the full stripe lock */
 442	mutex_lock(&locks_root->lock);
 443	existing = insert_full_stripe_lock(locks_root, fstripe_start);
 444	mutex_unlock(&locks_root->lock);
 445	if (IS_ERR(existing)) {
 446		ret = PTR_ERR(existing);
 447		goto out;
 448	}
 449	mutex_lock(&existing->mutex);
 450	*locked_ret = true;
 451out:
 452	btrfs_put_block_group(bg_cache);
 453	return ret;
 454}
 455
 456/*
 457 * Unlock a full stripe.
 458 *
 459 * NOTE: Caller must ensure it's the same context calling corresponding
 460 * lock_full_stripe().
 461 *
 462 * Return 0 if we unlock full stripe without problem.
 463 * Return <0 for error
 464 */
 465static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 466			      bool locked)
 467{
 468	struct btrfs_block_group *bg_cache;
 469	struct btrfs_full_stripe_locks_tree *locks_root;
 470	struct full_stripe_lock *fstripe_lock;
 471	u64 fstripe_start;
 472	bool freeit = false;
 473	int ret = 0;
 474
 475	/* If we didn't acquire full stripe lock, no need to continue */
 476	if (!locked)
 477		return 0;
 478
 479	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 480	if (!bg_cache) {
 481		ASSERT(0);
 482		return -ENOENT;
 483	}
 484	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 485		goto out;
 486
 487	locks_root = &bg_cache->full_stripe_locks_root;
 488	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 489
 490	mutex_lock(&locks_root->lock);
 491	fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 492	/* Unpaired unlock_full_stripe() detected */
 493	if (!fstripe_lock) {
 494		WARN_ON(1);
 495		ret = -ENOENT;
 496		mutex_unlock(&locks_root->lock);
 497		goto out;
 498	}
 499
 500	if (fstripe_lock->refs == 0) {
 501		WARN_ON(1);
 502		btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 503			fstripe_lock->logical);
 504	} else {
 505		fstripe_lock->refs--;
 506	}
 507
 508	if (fstripe_lock->refs == 0) {
 509		rb_erase(&fstripe_lock->node, &locks_root->root);
 510		freeit = true;
 511	}
 512	mutex_unlock(&locks_root->lock);
 513
 514	mutex_unlock(&fstripe_lock->mutex);
 515	if (freeit)
 516		kfree(fstripe_lock);
 517out:
 518	btrfs_put_block_group(bg_cache);
 519	return ret;
 520}
 521
 522static void scrub_free_csums(struct scrub_ctx *sctx)
 523{
 524	while (!list_empty(&sctx->csum_list)) {
 525		struct btrfs_ordered_sum *sum;
 526		sum = list_first_entry(&sctx->csum_list,
 527				       struct btrfs_ordered_sum, list);
 528		list_del(&sum->list);
 529		kfree(sum);
 530	}
 531}
 532
 533static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 534{
 535	int i;
 536
 537	if (!sctx)
 538		return;
 539
 
 
 540	/* this can happen when scrub is cancelled */
 541	if (sctx->curr != -1) {
 542		struct scrub_bio *sbio = sctx->bios[sctx->curr];
 543
 544		for (i = 0; i < sbio->page_count; i++) {
 545			WARN_ON(!sbio->pagev[i]->page);
 546			scrub_block_put(sbio->pagev[i]->sblock);
 547		}
 548		bio_put(sbio->bio);
 549	}
 550
 551	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 552		struct scrub_bio *sbio = sctx->bios[i];
 553
 554		if (!sbio)
 555			break;
 556		kfree(sbio);
 557	}
 558
 559	kfree(sctx->wr_curr_bio);
 560	scrub_free_csums(sctx);
 561	kfree(sctx);
 562}
 563
 564static void scrub_put_ctx(struct scrub_ctx *sctx)
 565{
 566	if (refcount_dec_and_test(&sctx->refs))
 567		scrub_free_ctx(sctx);
 568}
 569
 570static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 571		struct btrfs_fs_info *fs_info, int is_dev_replace)
 572{
 573	struct scrub_ctx *sctx;
 574	int		i;
 
 
 575
 576	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 577	if (!sctx)
 578		goto nomem;
 579	refcount_set(&sctx->refs, 1);
 580	sctx->is_dev_replace = is_dev_replace;
 581	sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 582	sctx->curr = -1;
 583	sctx->fs_info = fs_info;
 584	INIT_LIST_HEAD(&sctx->csum_list);
 585	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 586		struct scrub_bio *sbio;
 587
 588		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 589		if (!sbio)
 590			goto nomem;
 591		sctx->bios[i] = sbio;
 592
 593		sbio->index = i;
 594		sbio->sctx = sctx;
 595		sbio->page_count = 0;
 596		btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
 597				NULL);
 598
 599		if (i != SCRUB_BIOS_PER_SCTX - 1)
 600			sctx->bios[i]->next_free = i + 1;
 601		else
 602			sctx->bios[i]->next_free = -1;
 603	}
 604	sctx->first_free = 0;
 
 
 605	atomic_set(&sctx->bios_in_flight, 0);
 606	atomic_set(&sctx->workers_pending, 0);
 607	atomic_set(&sctx->cancel_req, 0);
 
 
 608
 609	spin_lock_init(&sctx->list_lock);
 610	spin_lock_init(&sctx->stat_lock);
 611	init_waitqueue_head(&sctx->list_wait);
 612	sctx->throttle_deadline = 0;
 613
 614	WARN_ON(sctx->wr_curr_bio != NULL);
 615	mutex_init(&sctx->wr_lock);
 616	sctx->wr_curr_bio = NULL;
 617	if (is_dev_replace) {
 618		WARN_ON(!fs_info->dev_replace.tgtdev);
 619		sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 620		sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 621		sctx->flush_all_writes = false;
 622	}
 623
 624	return sctx;
 625
 626nomem:
 627	scrub_free_ctx(sctx);
 628	return ERR_PTR(-ENOMEM);
 629}
 630
 631static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 632				     void *warn_ctx)
 633{
 
 634	u32 nlink;
 635	int ret;
 636	int i;
 637	unsigned nofs_flag;
 638	struct extent_buffer *eb;
 639	struct btrfs_inode_item *inode_item;
 640	struct scrub_warning *swarn = warn_ctx;
 641	struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 642	struct inode_fs_paths *ipath = NULL;
 643	struct btrfs_root *local_root;
 
 644	struct btrfs_key key;
 645
 646	local_root = btrfs_get_fs_root(fs_info, root, true);
 
 
 
 647	if (IS_ERR(local_root)) {
 648		ret = PTR_ERR(local_root);
 649		goto err;
 650	}
 651
 652	/*
 653	 * this makes the path point to (inum INODE_ITEM ioff)
 654	 */
 655	key.objectid = inum;
 656	key.type = BTRFS_INODE_ITEM_KEY;
 657	key.offset = 0;
 658
 659	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 660	if (ret) {
 661		btrfs_put_root(local_root);
 662		btrfs_release_path(swarn->path);
 663		goto err;
 664	}
 665
 666	eb = swarn->path->nodes[0];
 667	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 668					struct btrfs_inode_item);
 
 669	nlink = btrfs_inode_nlink(eb, inode_item);
 670	btrfs_release_path(swarn->path);
 671
 672	/*
 673	 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 674	 * uses GFP_NOFS in this context, so we keep it consistent but it does
 675	 * not seem to be strictly necessary.
 676	 */
 677	nofs_flag = memalloc_nofs_save();
 678	ipath = init_ipath(4096, local_root, swarn->path);
 679	memalloc_nofs_restore(nofs_flag);
 680	if (IS_ERR(ipath)) {
 681		btrfs_put_root(local_root);
 682		ret = PTR_ERR(ipath);
 683		ipath = NULL;
 684		goto err;
 685	}
 686	ret = paths_from_inode(inum, ipath);
 687
 688	if (ret < 0)
 689		goto err;
 690
 691	/*
 692	 * we deliberately ignore the bit ipath might have been too small to
 693	 * hold all of the paths here
 694	 */
 695	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 696		btrfs_warn_in_rcu(fs_info,
 697"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
 698				  swarn->errstr, swarn->logical,
 699				  rcu_str_deref(swarn->dev->name),
 700				  swarn->physical,
 701				  root, inum, offset,
 702				  fs_info->sectorsize, nlink,
 703				  (char *)(unsigned long)ipath->fspath->val[i]);
 704
 705	btrfs_put_root(local_root);
 706	free_ipath(ipath);
 707	return 0;
 708
 709err:
 710	btrfs_warn_in_rcu(fs_info,
 711			  "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 712			  swarn->errstr, swarn->logical,
 713			  rcu_str_deref(swarn->dev->name),
 714			  swarn->physical,
 715			  root, inum, offset, ret);
 716
 717	free_ipath(ipath);
 718	return 0;
 719}
 720
 721static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 722{
 723	struct btrfs_device *dev;
 724	struct btrfs_fs_info *fs_info;
 725	struct btrfs_path *path;
 726	struct btrfs_key found_key;
 727	struct extent_buffer *eb;
 728	struct btrfs_extent_item *ei;
 729	struct scrub_warning swarn;
 730	unsigned long ptr = 0;
 731	u64 extent_item_pos;
 732	u64 flags = 0;
 733	u64 ref_root;
 734	u32 item_size;
 735	u8 ref_level = 0;
 736	int ret;
 737
 738	WARN_ON(sblock->page_count < 1);
 739	dev = sblock->pagev[0]->dev;
 740	fs_info = sblock->sctx->fs_info;
 741
 742	path = btrfs_alloc_path();
 743	if (!path)
 744		return;
 745
 746	swarn.physical = sblock->pagev[0]->physical;
 747	swarn.logical = sblock->pagev[0]->logical;
 748	swarn.errstr = errstr;
 749	swarn.dev = NULL;
 750
 751	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 752				  &flags);
 753	if (ret < 0)
 754		goto out;
 755
 756	extent_item_pos = swarn.logical - found_key.objectid;
 757	swarn.extent_item_size = found_key.offset;
 758
 759	eb = path->nodes[0];
 760	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 761	item_size = btrfs_item_size_nr(eb, path->slots[0]);
 762
 763	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 764		do {
 765			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 766						      item_size, &ref_root,
 767						      &ref_level);
 768			btrfs_warn_in_rcu(fs_info,
 769"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 770				errstr, swarn.logical,
 
 771				rcu_str_deref(dev->name),
 772				swarn.physical,
 773				ref_level ? "node" : "leaf",
 774				ret < 0 ? -1 : ref_level,
 775				ret < 0 ? -1 : ref_root);
 776		} while (ret != 1);
 777		btrfs_release_path(path);
 778	} else {
 779		btrfs_release_path(path);
 780		swarn.path = path;
 781		swarn.dev = dev;
 782		iterate_extent_inodes(fs_info, found_key.objectid,
 783					extent_item_pos, 1,
 784					scrub_print_warning_inode, &swarn, false);
 785	}
 786
 787out:
 788	btrfs_free_path(path);
 789}
 790
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 791static inline void scrub_get_recover(struct scrub_recover *recover)
 792{
 793	refcount_inc(&recover->refs);
 794}
 795
 796static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 797				     struct scrub_recover *recover)
 798{
 799	if (refcount_dec_and_test(&recover->refs)) {
 800		btrfs_bio_counter_dec(fs_info);
 801		btrfs_put_bbio(recover->bbio);
 802		kfree(recover);
 803	}
 804}
 805
 806/*
 807 * scrub_handle_errored_block gets called when either verification of the
 808 * pages failed or the bio failed to read, e.g. with EIO. In the latter
 809 * case, this function handles all pages in the bio, even though only one
 810 * may be bad.
 811 * The goal of this function is to repair the errored block by using the
 812 * contents of one of the mirrors.
 813 */
 814static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 815{
 816	struct scrub_ctx *sctx = sblock_to_check->sctx;
 817	struct btrfs_device *dev;
 818	struct btrfs_fs_info *fs_info;
 
 819	u64 logical;
 820	unsigned int failed_mirror_index;
 821	unsigned int is_metadata;
 822	unsigned int have_csum;
 823	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 824	struct scrub_block *sblock_bad;
 825	int ret;
 826	int mirror_index;
 827	int page_num;
 828	int success;
 829	bool full_stripe_locked;
 830	unsigned int nofs_flag;
 831	static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
 832				      DEFAULT_RATELIMIT_BURST);
 833
 834	BUG_ON(sblock_to_check->page_count < 1);
 835	fs_info = sctx->fs_info;
 836	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 837		/*
 838		 * if we find an error in a super block, we just report it.
 839		 * They will get written with the next transaction commit
 840		 * anyway
 841		 */
 842		spin_lock(&sctx->stat_lock);
 843		++sctx->stat.super_errors;
 844		spin_unlock(&sctx->stat_lock);
 845		return 0;
 846	}
 
 847	logical = sblock_to_check->pagev[0]->logical;
 848	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 849	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 850	is_metadata = !(sblock_to_check->pagev[0]->flags &
 851			BTRFS_EXTENT_FLAG_DATA);
 852	have_csum = sblock_to_check->pagev[0]->have_csum;
 853	dev = sblock_to_check->pagev[0]->dev;
 854
 855	if (btrfs_is_zoned(fs_info) && !sctx->is_dev_replace)
 856		return btrfs_repair_one_zone(fs_info, logical);
 857
 858	/*
 859	 * We must use GFP_NOFS because the scrub task might be waiting for a
 860	 * worker task executing this function and in turn a transaction commit
 861	 * might be waiting the scrub task to pause (which needs to wait for all
 862	 * the worker tasks to complete before pausing).
 863	 * We do allocations in the workers through insert_full_stripe_lock()
 864	 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 865	 * this function.
 866	 */
 867	nofs_flag = memalloc_nofs_save();
 868	/*
 869	 * For RAID5/6, race can happen for a different device scrub thread.
 870	 * For data corruption, Parity and Data threads will both try
 871	 * to recovery the data.
 872	 * Race can lead to doubly added csum error, or even unrecoverable
 873	 * error.
 874	 */
 875	ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 876	if (ret < 0) {
 877		memalloc_nofs_restore(nofs_flag);
 878		spin_lock(&sctx->stat_lock);
 879		if (ret == -ENOMEM)
 880			sctx->stat.malloc_errors++;
 881		sctx->stat.read_errors++;
 882		sctx->stat.uncorrectable_errors++;
 883		spin_unlock(&sctx->stat_lock);
 884		return ret;
 885	}
 886
 887	/*
 888	 * read all mirrors one after the other. This includes to
 889	 * re-read the extent or metadata block that failed (that was
 890	 * the cause that this fixup code is called) another time,
 891	 * sector by sector this time in order to know which sectors
 892	 * caused I/O errors and which ones are good (for all mirrors).
 893	 * It is the goal to handle the situation when more than one
 894	 * mirror contains I/O errors, but the errors do not
 895	 * overlap, i.e. the data can be repaired by selecting the
 896	 * sectors from those mirrors without I/O error on the
 897	 * particular sectors. One example (with blocks >= 2 * sectorsize)
 898	 * would be that mirror #1 has an I/O error on the first sector,
 899	 * the second sector is good, and mirror #2 has an I/O error on
 900	 * the second sector, but the first sector is good.
 901	 * Then the first sector of the first mirror can be repaired by
 902	 * taking the first sector of the second mirror, and the
 903	 * second sector of the second mirror can be repaired by
 904	 * copying the contents of the 2nd sector of the 1st mirror.
 905	 * One more note: if the sectors of one mirror contain I/O
 906	 * errors, the checksum cannot be verified. In order to get
 907	 * the best data for repairing, the first attempt is to find
 908	 * a mirror without I/O errors and with a validated checksum.
 909	 * Only if this is not possible, the sectors are picked from
 910	 * mirrors with I/O errors without considering the checksum.
 911	 * If the latter is the case, at the end, the checksum of the
 912	 * repaired area is verified in order to correctly maintain
 913	 * the statistics.
 914	 */
 915
 916	sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 917				      sizeof(*sblocks_for_recheck), GFP_KERNEL);
 918	if (!sblocks_for_recheck) {
 919		spin_lock(&sctx->stat_lock);
 920		sctx->stat.malloc_errors++;
 921		sctx->stat.read_errors++;
 922		sctx->stat.uncorrectable_errors++;
 923		spin_unlock(&sctx->stat_lock);
 924		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 925		goto out;
 926	}
 927
 928	/* setup the context, map the logical blocks and alloc the pages */
 929	ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 930	if (ret) {
 931		spin_lock(&sctx->stat_lock);
 932		sctx->stat.read_errors++;
 933		sctx->stat.uncorrectable_errors++;
 934		spin_unlock(&sctx->stat_lock);
 935		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 936		goto out;
 937	}
 938	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 939	sblock_bad = sblocks_for_recheck + failed_mirror_index;
 940
 941	/* build and submit the bios for the failed mirror, check checksums */
 942	scrub_recheck_block(fs_info, sblock_bad, 1);
 943
 944	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 945	    sblock_bad->no_io_error_seen) {
 946		/*
 947		 * the error disappeared after reading page by page, or
 948		 * the area was part of a huge bio and other parts of the
 949		 * bio caused I/O errors, or the block layer merged several
 950		 * read requests into one and the error is caused by a
 951		 * different bio (usually one of the two latter cases is
 952		 * the cause)
 953		 */
 954		spin_lock(&sctx->stat_lock);
 955		sctx->stat.unverified_errors++;
 956		sblock_to_check->data_corrected = 1;
 957		spin_unlock(&sctx->stat_lock);
 958
 959		if (sctx->is_dev_replace)
 960			scrub_write_block_to_dev_replace(sblock_bad);
 961		goto out;
 962	}
 963
 964	if (!sblock_bad->no_io_error_seen) {
 965		spin_lock(&sctx->stat_lock);
 966		sctx->stat.read_errors++;
 967		spin_unlock(&sctx->stat_lock);
 968		if (__ratelimit(&rs))
 969			scrub_print_warning("i/o error", sblock_to_check);
 970		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 971	} else if (sblock_bad->checksum_error) {
 972		spin_lock(&sctx->stat_lock);
 973		sctx->stat.csum_errors++;
 974		spin_unlock(&sctx->stat_lock);
 975		if (__ratelimit(&rs))
 976			scrub_print_warning("checksum error", sblock_to_check);
 977		btrfs_dev_stat_inc_and_print(dev,
 978					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
 979	} else if (sblock_bad->header_error) {
 980		spin_lock(&sctx->stat_lock);
 981		sctx->stat.verify_errors++;
 982		spin_unlock(&sctx->stat_lock);
 983		if (__ratelimit(&rs))
 984			scrub_print_warning("checksum/header error",
 985					    sblock_to_check);
 986		if (sblock_bad->generation_error)
 987			btrfs_dev_stat_inc_and_print(dev,
 988				BTRFS_DEV_STAT_GENERATION_ERRS);
 989		else
 990			btrfs_dev_stat_inc_and_print(dev,
 991				BTRFS_DEV_STAT_CORRUPTION_ERRS);
 992	}
 993
 994	if (sctx->readonly) {
 995		ASSERT(!sctx->is_dev_replace);
 996		goto out;
 997	}
 998
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 999	/*
1000	 * now build and submit the bios for the other mirrors, check
1001	 * checksums.
1002	 * First try to pick the mirror which is completely without I/O
1003	 * errors and also does not have a checksum error.
1004	 * If one is found, and if a checksum is present, the full block
1005	 * that is known to contain an error is rewritten. Afterwards
1006	 * the block is known to be corrected.
1007	 * If a mirror is found which is completely correct, and no
1008	 * checksum is present, only those pages are rewritten that had
1009	 * an I/O error in the block to be repaired, since it cannot be
1010	 * determined, which copy of the other pages is better (and it
1011	 * could happen otherwise that a correct page would be
1012	 * overwritten by a bad one).
1013	 */
1014	for (mirror_index = 0; ;mirror_index++) {
 
 
 
1015		struct scrub_block *sblock_other;
1016
1017		if (mirror_index == failed_mirror_index)
1018			continue;
1019
1020		/* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1021		if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1022			if (mirror_index >= BTRFS_MAX_MIRRORS)
1023				break;
1024			if (!sblocks_for_recheck[mirror_index].page_count)
1025				break;
1026
1027			sblock_other = sblocks_for_recheck + mirror_index;
1028		} else {
1029			struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1030			int max_allowed = r->bbio->num_stripes -
1031						r->bbio->num_tgtdevs;
1032
1033			if (mirror_index >= max_allowed)
1034				break;
1035			if (!sblocks_for_recheck[1].page_count)
1036				break;
1037
1038			ASSERT(failed_mirror_index == 0);
1039			sblock_other = sblocks_for_recheck + 1;
1040			sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1041		}
1042
1043		/* build and submit the bios, check checksums */
1044		scrub_recheck_block(fs_info, sblock_other, 0);
1045
1046		if (!sblock_other->header_error &&
1047		    !sblock_other->checksum_error &&
1048		    sblock_other->no_io_error_seen) {
1049			if (sctx->is_dev_replace) {
1050				scrub_write_block_to_dev_replace(sblock_other);
1051				goto corrected_error;
1052			} else {
1053				ret = scrub_repair_block_from_good_copy(
1054						sblock_bad, sblock_other);
1055				if (!ret)
1056					goto corrected_error;
1057			}
1058		}
1059	}
1060
1061	if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1062		goto did_not_correct_error;
1063
1064	/*
1065	 * In case of I/O errors in the area that is supposed to be
1066	 * repaired, continue by picking good copies of those sectors.
1067	 * Select the good sectors from mirrors to rewrite bad sectors from
1068	 * the area to fix. Afterwards verify the checksum of the block
1069	 * that is supposed to be repaired. This verification step is
1070	 * only done for the purpose of statistic counting and for the
1071	 * final scrub report, whether errors remain.
1072	 * A perfect algorithm could make use of the checksum and try
1073	 * all possible combinations of sectors from the different mirrors
1074	 * until the checksum verification succeeds. For example, when
1075	 * the 2nd sector of mirror #1 faces I/O errors, and the 2nd sector
1076	 * of mirror #2 is readable but the final checksum test fails,
1077	 * then the 2nd sector of mirror #3 could be tried, whether now
1078	 * the final checksum succeeds. But this would be a rare
1079	 * exception and is therefore not implemented. At least it is
1080	 * avoided that the good copy is overwritten.
1081	 * A more useful improvement would be to pick the sectors
1082	 * without I/O error based on sector sizes (512 bytes on legacy
1083	 * disks) instead of on sectorsize. Then maybe 512 byte of one
1084	 * mirror could be repaired by taking 512 byte of a different
1085	 * mirror, even if other 512 byte sectors in the same sectorsize
1086	 * area are unreadable.
1087	 */
1088	success = 1;
1089	for (page_num = 0; page_num < sblock_bad->page_count;
1090	     page_num++) {
1091		struct scrub_page *spage_bad = sblock_bad->pagev[page_num];
1092		struct scrub_block *sblock_other = NULL;
1093
1094		/* skip no-io-error page in scrub */
1095		if (!spage_bad->io_error && !sctx->is_dev_replace)
1096			continue;
1097
1098		if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1099			/*
1100			 * In case of dev replace, if raid56 rebuild process
1101			 * didn't work out correct data, then copy the content
1102			 * in sblock_bad to make sure target device is identical
1103			 * to source device, instead of writing garbage data in
1104			 * sblock_for_recheck array to target device.
1105			 */
1106			sblock_other = NULL;
1107		} else if (spage_bad->io_error) {
1108			/* try to find no-io-error page in mirrors */
1109			for (mirror_index = 0;
1110			     mirror_index < BTRFS_MAX_MIRRORS &&
1111			     sblocks_for_recheck[mirror_index].page_count > 0;
1112			     mirror_index++) {
1113				if (!sblocks_for_recheck[mirror_index].
1114				    pagev[page_num]->io_error) {
1115					sblock_other = sblocks_for_recheck +
1116						       mirror_index;
1117					break;
1118				}
1119			}
1120			if (!sblock_other)
1121				success = 0;
1122		}
1123
1124		if (sctx->is_dev_replace) {
1125			/*
1126			 * did not find a mirror to fetch the page
1127			 * from. scrub_write_page_to_dev_replace()
1128			 * handles this case (page->io_error), by
1129			 * filling the block with zeros before
1130			 * submitting the write request
1131			 */
1132			if (!sblock_other)
1133				sblock_other = sblock_bad;
1134
1135			if (scrub_write_page_to_dev_replace(sblock_other,
1136							    page_num) != 0) {
1137				atomic64_inc(
1138					&fs_info->dev_replace.num_write_errors);
 
 
1139				success = 0;
1140			}
1141		} else if (sblock_other) {
1142			ret = scrub_repair_page_from_good_copy(sblock_bad,
1143							       sblock_other,
1144							       page_num, 0);
1145			if (0 == ret)
1146				spage_bad->io_error = 0;
1147			else
1148				success = 0;
1149		}
1150	}
1151
1152	if (success && !sctx->is_dev_replace) {
1153		if (is_metadata || have_csum) {
1154			/*
1155			 * need to verify the checksum now that all
1156			 * sectors on disk are repaired (the write
1157			 * request for data to be repaired is on its way).
1158			 * Just be lazy and use scrub_recheck_block()
1159			 * which re-reads the data before the checksum
1160			 * is verified, but most likely the data comes out
1161			 * of the page cache.
1162			 */
1163			scrub_recheck_block(fs_info, sblock_bad, 1);
1164			if (!sblock_bad->header_error &&
1165			    !sblock_bad->checksum_error &&
1166			    sblock_bad->no_io_error_seen)
1167				goto corrected_error;
1168			else
1169				goto did_not_correct_error;
1170		} else {
1171corrected_error:
1172			spin_lock(&sctx->stat_lock);
1173			sctx->stat.corrected_errors++;
1174			sblock_to_check->data_corrected = 1;
1175			spin_unlock(&sctx->stat_lock);
1176			btrfs_err_rl_in_rcu(fs_info,
1177				"fixed up error at logical %llu on dev %s",
1178				logical, rcu_str_deref(dev->name));
1179		}
1180	} else {
1181did_not_correct_error:
1182		spin_lock(&sctx->stat_lock);
1183		sctx->stat.uncorrectable_errors++;
1184		spin_unlock(&sctx->stat_lock);
1185		btrfs_err_rl_in_rcu(fs_info,
1186			"unable to fixup (regular) error at logical %llu on dev %s",
1187			logical, rcu_str_deref(dev->name));
1188	}
1189
1190out:
1191	if (sblocks_for_recheck) {
1192		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1193		     mirror_index++) {
1194			struct scrub_block *sblock = sblocks_for_recheck +
1195						     mirror_index;
1196			struct scrub_recover *recover;
1197			int page_index;
1198
1199			for (page_index = 0; page_index < sblock->page_count;
1200			     page_index++) {
1201				sblock->pagev[page_index]->sblock = NULL;
1202				recover = sblock->pagev[page_index]->recover;
1203				if (recover) {
1204					scrub_put_recover(fs_info, recover);
1205					sblock->pagev[page_index]->recover =
1206									NULL;
1207				}
1208				scrub_page_put(sblock->pagev[page_index]);
1209			}
1210		}
1211		kfree(sblocks_for_recheck);
1212	}
1213
1214	ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1215	memalloc_nofs_restore(nofs_flag);
1216	if (ret < 0)
1217		return ret;
1218	return 0;
1219}
1220
1221static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1222{
1223	if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1224		return 2;
1225	else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1226		return 3;
1227	else
1228		return (int)bbio->num_stripes;
1229}
1230
1231static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1232						 u64 *raid_map,
1233						 u64 mapped_length,
1234						 int nstripes, int mirror,
1235						 int *stripe_index,
1236						 u64 *stripe_offset)
1237{
1238	int i;
1239
1240	if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1241		/* RAID5/6 */
1242		for (i = 0; i < nstripes; i++) {
1243			if (raid_map[i] == RAID6_Q_STRIPE ||
1244			    raid_map[i] == RAID5_P_STRIPE)
1245				continue;
1246
1247			if (logical >= raid_map[i] &&
1248			    logical < raid_map[i] + mapped_length)
1249				break;
1250		}
1251
1252		*stripe_index = i;
1253		*stripe_offset = logical - raid_map[i];
1254	} else {
1255		/* The other RAID type */
1256		*stripe_index = mirror;
1257		*stripe_offset = 0;
1258	}
1259}
1260
1261static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1262				     struct scrub_block *sblocks_for_recheck)
1263{
1264	struct scrub_ctx *sctx = original_sblock->sctx;
1265	struct btrfs_fs_info *fs_info = sctx->fs_info;
1266	u64 length = original_sblock->page_count * fs_info->sectorsize;
1267	u64 logical = original_sblock->pagev[0]->logical;
1268	u64 generation = original_sblock->pagev[0]->generation;
1269	u64 flags = original_sblock->pagev[0]->flags;
1270	u64 have_csum = original_sblock->pagev[0]->have_csum;
1271	struct scrub_recover *recover;
1272	struct btrfs_bio *bbio;
1273	u64 sublen;
1274	u64 mapped_length;
1275	u64 stripe_offset;
1276	int stripe_index;
1277	int page_index = 0;
1278	int mirror_index;
1279	int nmirrors;
1280	int ret;
1281
1282	/*
1283	 * note: the two members refs and outstanding_pages
1284	 * are not used (and not set) in the blocks that are used for
1285	 * the recheck procedure
1286	 */
1287
1288	while (length > 0) {
1289		sublen = min_t(u64, length, fs_info->sectorsize);
1290		mapped_length = sublen;
1291		bbio = NULL;
1292
1293		/*
1294		 * With a length of sectorsize, each returned stripe represents
1295		 * one mirror
1296		 */
1297		btrfs_bio_counter_inc_blocked(fs_info);
1298		ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1299				logical, &mapped_length, &bbio);
1300		if (ret || !bbio || mapped_length < sublen) {
1301			btrfs_put_bbio(bbio);
1302			btrfs_bio_counter_dec(fs_info);
1303			return -EIO;
1304		}
1305
1306		recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1307		if (!recover) {
1308			btrfs_put_bbio(bbio);
1309			btrfs_bio_counter_dec(fs_info);
1310			return -ENOMEM;
1311		}
1312
1313		refcount_set(&recover->refs, 1);
1314		recover->bbio = bbio;
1315		recover->map_length = mapped_length;
1316
1317		BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1318
1319		nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1320
1321		for (mirror_index = 0; mirror_index < nmirrors;
1322		     mirror_index++) {
1323			struct scrub_block *sblock;
1324			struct scrub_page *spage;
1325
1326			sblock = sblocks_for_recheck + mirror_index;
1327			sblock->sctx = sctx;
1328
1329			spage = kzalloc(sizeof(*spage), GFP_NOFS);
1330			if (!spage) {
1331leave_nomem:
1332				spin_lock(&sctx->stat_lock);
1333				sctx->stat.malloc_errors++;
1334				spin_unlock(&sctx->stat_lock);
1335				scrub_put_recover(fs_info, recover);
1336				return -ENOMEM;
1337			}
1338			scrub_page_get(spage);
1339			sblock->pagev[page_index] = spage;
1340			spage->sblock = sblock;
1341			spage->flags = flags;
1342			spage->generation = generation;
1343			spage->logical = logical;
1344			spage->have_csum = have_csum;
1345			if (have_csum)
1346				memcpy(spage->csum,
1347				       original_sblock->pagev[0]->csum,
1348				       sctx->fs_info->csum_size);
1349
1350			scrub_stripe_index_and_offset(logical,
1351						      bbio->map_type,
1352						      bbio->raid_map,
1353						      mapped_length,
1354						      bbio->num_stripes -
1355						      bbio->num_tgtdevs,
1356						      mirror_index,
1357						      &stripe_index,
1358						      &stripe_offset);
1359			spage->physical = bbio->stripes[stripe_index].physical +
1360					 stripe_offset;
1361			spage->dev = bbio->stripes[stripe_index].dev;
1362
1363			BUG_ON(page_index >= original_sblock->page_count);
1364			spage->physical_for_dev_replace =
1365				original_sblock->pagev[page_index]->
1366				physical_for_dev_replace;
1367			/* for missing devices, dev->bdev is NULL */
1368			spage->mirror_num = mirror_index + 1;
1369			sblock->page_count++;
1370			spage->page = alloc_page(GFP_NOFS);
1371			if (!spage->page)
1372				goto leave_nomem;
1373
1374			scrub_get_recover(recover);
1375			spage->recover = recover;
1376		}
1377		scrub_put_recover(fs_info, recover);
1378		length -= sublen;
1379		logical += sublen;
1380		page_index++;
1381	}
1382
1383	return 0;
1384}
1385
 
 
 
 
 
1386static void scrub_bio_wait_endio(struct bio *bio)
1387{
1388	complete(bio->bi_private);
 
 
 
 
 
 
 
 
 
1389}
1390
1391static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1392					struct bio *bio,
1393					struct scrub_page *spage)
1394{
1395	DECLARE_COMPLETION_ONSTACK(done);
1396	int ret;
1397	int mirror_num;
1398
1399	bio->bi_iter.bi_sector = spage->logical >> 9;
 
 
1400	bio->bi_private = &done;
1401	bio->bi_end_io = scrub_bio_wait_endio;
1402
1403	mirror_num = spage->sblock->pagev[0]->mirror_num;
1404	ret = raid56_parity_recover(fs_info, bio, spage->recover->bbio,
1405				    spage->recover->map_length,
1406				    mirror_num, 0);
1407	if (ret)
1408		return ret;
1409
1410	wait_for_completion_io(&done);
1411	return blk_status_to_errno(bio->bi_status);
1412}
1413
1414static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1415					  struct scrub_block *sblock)
1416{
1417	struct scrub_page *first_page = sblock->pagev[0];
1418	struct bio *bio;
1419	int page_num;
1420
1421	/* All pages in sblock belong to the same stripe on the same device. */
1422	ASSERT(first_page->dev);
1423	if (!first_page->dev->bdev)
1424		goto out;
1425
1426	bio = btrfs_io_bio_alloc(BIO_MAX_VECS);
1427	bio_set_dev(bio, first_page->dev->bdev);
1428
1429	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1430		struct scrub_page *spage = sblock->pagev[page_num];
1431
1432		WARN_ON(!spage->page);
1433		bio_add_page(bio, spage->page, PAGE_SIZE, 0);
1434	}
1435
1436	if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1437		bio_put(bio);
1438		goto out;
1439	}
1440
1441	bio_put(bio);
1442
1443	scrub_recheck_block_checksum(sblock);
1444
1445	return;
1446out:
1447	for (page_num = 0; page_num < sblock->page_count; page_num++)
1448		sblock->pagev[page_num]->io_error = 1;
1449
1450	sblock->no_io_error_seen = 0;
1451}
1452
1453/*
1454 * this function will check the on disk data for checksum errors, header
1455 * errors and read I/O errors. If any I/O errors happen, the exact pages
1456 * which are errored are marked as being bad. The goal is to enable scrub
1457 * to take those pages that are not errored from all the mirrors so that
1458 * the pages that are errored in the just handled mirror can be repaired.
1459 */
1460static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1461				struct scrub_block *sblock,
1462				int retry_failed_mirror)
1463{
1464	int page_num;
1465
1466	sblock->no_io_error_seen = 1;
1467
1468	/* short cut for raid56 */
1469	if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1470		return scrub_recheck_block_on_raid56(fs_info, sblock);
1471
1472	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1473		struct bio *bio;
1474		struct scrub_page *spage = sblock->pagev[page_num];
1475
1476		if (spage->dev->bdev == NULL) {
1477			spage->io_error = 1;
 
 
 
 
 
 
 
 
1478			sblock->no_io_error_seen = 0;
1479			continue;
1480		}
 
1481
1482		WARN_ON(!spage->page);
1483		bio = btrfs_io_bio_alloc(1);
1484		bio_set_dev(bio, spage->dev->bdev);
1485
1486		bio_add_page(bio, spage->page, fs_info->sectorsize, 0);
1487		bio->bi_iter.bi_sector = spage->physical >> 9;
1488		bio->bi_opf = REQ_OP_READ;
1489
1490		if (btrfsic_submit_bio_wait(bio)) {
1491			spage->io_error = 1;
1492			sblock->no_io_error_seen = 0;
1493		}
1494
1495		bio_put(bio);
1496	}
1497
1498	if (sblock->no_io_error_seen)
1499		scrub_recheck_block_checksum(sblock);
1500}
1501
1502static inline int scrub_check_fsid(u8 fsid[],
1503				   struct scrub_page *spage)
1504{
1505	struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1506	int ret;
1507
1508	ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1509	return !ret;
1510}
1511
1512static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1513{
1514	sblock->header_error = 0;
1515	sblock->checksum_error = 0;
1516	sblock->generation_error = 0;
1517
1518	if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1519		scrub_checksum_data(sblock);
1520	else
1521		scrub_checksum_tree_block(sblock);
1522}
1523
1524static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1525					     struct scrub_block *sblock_good)
1526{
1527	int page_num;
1528	int ret = 0;
1529
1530	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1531		int ret_sub;
1532
1533		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1534							   sblock_good,
1535							   page_num, 1);
1536		if (ret_sub)
1537			ret = ret_sub;
1538	}
1539
1540	return ret;
1541}
1542
1543static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1544					    struct scrub_block *sblock_good,
1545					    int page_num, int force_write)
1546{
1547	struct scrub_page *spage_bad = sblock_bad->pagev[page_num];
1548	struct scrub_page *spage_good = sblock_good->pagev[page_num];
1549	struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1550	const u32 sectorsize = fs_info->sectorsize;
1551
1552	BUG_ON(spage_bad->page == NULL);
1553	BUG_ON(spage_good->page == NULL);
1554	if (force_write || sblock_bad->header_error ||
1555	    sblock_bad->checksum_error || spage_bad->io_error) {
1556		struct bio *bio;
1557		int ret;
1558
1559		if (!spage_bad->dev->bdev) {
1560			btrfs_warn_rl(fs_info,
1561				"scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
 
1562			return -EIO;
1563		}
1564
1565		bio = btrfs_io_bio_alloc(1);
1566		bio_set_dev(bio, spage_bad->dev->bdev);
1567		bio->bi_iter.bi_sector = spage_bad->physical >> 9;
1568		bio->bi_opf = REQ_OP_WRITE;
 
1569
1570		ret = bio_add_page(bio, spage_good->page, sectorsize, 0);
1571		if (ret != sectorsize) {
1572			bio_put(bio);
1573			return -EIO;
1574		}
1575
1576		if (btrfsic_submit_bio_wait(bio)) {
1577			btrfs_dev_stat_inc_and_print(spage_bad->dev,
1578				BTRFS_DEV_STAT_WRITE_ERRS);
1579			atomic64_inc(&fs_info->dev_replace.num_write_errors);
 
 
1580			bio_put(bio);
1581			return -EIO;
1582		}
1583		bio_put(bio);
1584	}
1585
1586	return 0;
1587}
1588
1589static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1590{
1591	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1592	int page_num;
1593
1594	/*
1595	 * This block is used for the check of the parity on the source device,
1596	 * so the data needn't be written into the destination device.
1597	 */
1598	if (sblock->sparity)
1599		return;
1600
1601	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1602		int ret;
1603
1604		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1605		if (ret)
1606			atomic64_inc(&fs_info->dev_replace.num_write_errors);
 
 
1607	}
1608}
1609
1610static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1611					   int page_num)
1612{
1613	struct scrub_page *spage = sblock->pagev[page_num];
1614
1615	BUG_ON(spage->page == NULL);
1616	if (spage->io_error)
1617		clear_page(page_address(spage->page));
1618
 
 
 
 
1619	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1620}
1621
1622static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
1623{
1624	int ret = 0;
1625	u64 length;
1626
1627	if (!btrfs_is_zoned(sctx->fs_info))
1628		return 0;
1629
1630	if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1631		return 0;
1632
1633	if (sctx->write_pointer < physical) {
1634		length = physical - sctx->write_pointer;
1635
1636		ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1637						sctx->write_pointer, length);
1638		if (!ret)
1639			sctx->write_pointer = physical;
1640	}
1641	return ret;
1642}
1643
1644static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1645				    struct scrub_page *spage)
1646{
 
1647	struct scrub_bio *sbio;
1648	int ret;
1649	const u32 sectorsize = sctx->fs_info->sectorsize;
1650
1651	mutex_lock(&sctx->wr_lock);
1652again:
1653	if (!sctx->wr_curr_bio) {
1654		sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1655					      GFP_KERNEL);
1656		if (!sctx->wr_curr_bio) {
1657			mutex_unlock(&sctx->wr_lock);
1658			return -ENOMEM;
1659		}
1660		sctx->wr_curr_bio->sctx = sctx;
1661		sctx->wr_curr_bio->page_count = 0;
1662	}
1663	sbio = sctx->wr_curr_bio;
1664	if (sbio->page_count == 0) {
1665		struct bio *bio;
1666
1667		ret = fill_writer_pointer_gap(sctx,
1668					      spage->physical_for_dev_replace);
1669		if (ret) {
1670			mutex_unlock(&sctx->wr_lock);
1671			return ret;
1672		}
1673
1674		sbio->physical = spage->physical_for_dev_replace;
1675		sbio->logical = spage->logical;
1676		sbio->dev = sctx->wr_tgtdev;
1677		bio = sbio->bio;
1678		if (!bio) {
1679			bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
 
 
 
 
 
1680			sbio->bio = bio;
1681		}
1682
1683		bio->bi_private = sbio;
1684		bio->bi_end_io = scrub_wr_bio_end_io;
1685		bio_set_dev(bio, sbio->dev->bdev);
1686		bio->bi_iter.bi_sector = sbio->physical >> 9;
1687		bio->bi_opf = REQ_OP_WRITE;
1688		sbio->status = 0;
1689	} else if (sbio->physical + sbio->page_count * sectorsize !=
1690		   spage->physical_for_dev_replace ||
1691		   sbio->logical + sbio->page_count * sectorsize !=
1692		   spage->logical) {
1693		scrub_wr_submit(sctx);
1694		goto again;
1695	}
1696
1697	ret = bio_add_page(sbio->bio, spage->page, sectorsize, 0);
1698	if (ret != sectorsize) {
1699		if (sbio->page_count < 1) {
1700			bio_put(sbio->bio);
1701			sbio->bio = NULL;
1702			mutex_unlock(&sctx->wr_lock);
1703			return -EIO;
1704		}
1705		scrub_wr_submit(sctx);
1706		goto again;
1707	}
1708
1709	sbio->pagev[sbio->page_count] = spage;
1710	scrub_page_get(spage);
1711	sbio->page_count++;
1712	if (sbio->page_count == sctx->pages_per_wr_bio)
1713		scrub_wr_submit(sctx);
1714	mutex_unlock(&sctx->wr_lock);
1715
1716	return 0;
1717}
1718
1719static void scrub_wr_submit(struct scrub_ctx *sctx)
1720{
 
1721	struct scrub_bio *sbio;
1722
1723	if (!sctx->wr_curr_bio)
1724		return;
1725
1726	sbio = sctx->wr_curr_bio;
1727	sctx->wr_curr_bio = NULL;
1728	WARN_ON(!sbio->bio->bi_bdev);
1729	scrub_pending_bio_inc(sctx);
1730	/* process all writes in a single worker thread. Then the block layer
1731	 * orders the requests before sending them to the driver which
1732	 * doubled the write performance on spinning disks when measured
1733	 * with Linux 3.5 */
1734	btrfsic_submit_bio(sbio->bio);
1735
1736	if (btrfs_is_zoned(sctx->fs_info))
1737		sctx->write_pointer = sbio->physical + sbio->page_count *
1738			sctx->fs_info->sectorsize;
1739}
1740
1741static void scrub_wr_bio_end_io(struct bio *bio)
1742{
1743	struct scrub_bio *sbio = bio->bi_private;
1744	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1745
1746	sbio->status = bio->bi_status;
1747	sbio->bio = bio;
1748
1749	btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
 
1750	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1751}
1752
1753static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1754{
1755	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1756	struct scrub_ctx *sctx = sbio->sctx;
1757	int i;
1758
1759	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1760	if (sbio->status) {
1761		struct btrfs_dev_replace *dev_replace =
1762			&sbio->sctx->fs_info->dev_replace;
1763
1764		for (i = 0; i < sbio->page_count; i++) {
1765			struct scrub_page *spage = sbio->pagev[i];
1766
1767			spage->io_error = 1;
1768			atomic64_inc(&dev_replace->num_write_errors);
 
1769		}
1770	}
1771
1772	for (i = 0; i < sbio->page_count; i++)
1773		scrub_page_put(sbio->pagev[i]);
1774
1775	bio_put(sbio->bio);
1776	kfree(sbio);
1777	scrub_pending_bio_dec(sctx);
1778}
1779
1780static int scrub_checksum(struct scrub_block *sblock)
1781{
1782	u64 flags;
1783	int ret;
1784
1785	/*
1786	 * No need to initialize these stats currently,
1787	 * because this function only use return value
1788	 * instead of these stats value.
1789	 *
1790	 * Todo:
1791	 * always use stats
1792	 */
1793	sblock->header_error = 0;
1794	sblock->generation_error = 0;
1795	sblock->checksum_error = 0;
1796
1797	WARN_ON(sblock->page_count < 1);
1798	flags = sblock->pagev[0]->flags;
1799	ret = 0;
1800	if (flags & BTRFS_EXTENT_FLAG_DATA)
1801		ret = scrub_checksum_data(sblock);
1802	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1803		ret = scrub_checksum_tree_block(sblock);
1804	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1805		(void)scrub_checksum_super(sblock);
1806	else
1807		WARN_ON(1);
1808	if (ret)
1809		scrub_handle_errored_block(sblock);
1810
1811	return ret;
1812}
1813
1814static int scrub_checksum_data(struct scrub_block *sblock)
1815{
1816	struct scrub_ctx *sctx = sblock->sctx;
1817	struct btrfs_fs_info *fs_info = sctx->fs_info;
1818	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1819	u8 csum[BTRFS_CSUM_SIZE];
1820	struct scrub_page *spage;
1821	char *kaddr;
 
 
 
 
1822
1823	BUG_ON(sblock->page_count < 1);
1824	spage = sblock->pagev[0];
1825	if (!spage->have_csum)
1826		return 0;
1827
1828	kaddr = page_address(spage->page);
 
 
 
 
 
 
 
1829
1830	shash->tfm = fs_info->csum_shash;
1831	crypto_shash_init(shash);
 
 
 
 
 
 
 
 
 
1832
1833	/*
1834	 * In scrub_pages() and scrub_pages_for_parity() we ensure each spage
1835	 * only contains one sector of data.
1836	 */
1837	crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
1838
1839	if (memcmp(csum, spage->csum, fs_info->csum_size))
1840		sblock->checksum_error = 1;
1841	return sblock->checksum_error;
1842}
1843
1844static int scrub_checksum_tree_block(struct scrub_block *sblock)
1845{
1846	struct scrub_ctx *sctx = sblock->sctx;
1847	struct btrfs_header *h;
1848	struct btrfs_fs_info *fs_info = sctx->fs_info;
1849	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1850	u8 calculated_csum[BTRFS_CSUM_SIZE];
1851	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1852	/*
1853	 * This is done in sectorsize steps even for metadata as there's a
1854	 * constraint for nodesize to be aligned to sectorsize. This will need
1855	 * to change so we don't misuse data and metadata units like that.
1856	 */
1857	const u32 sectorsize = sctx->fs_info->sectorsize;
1858	const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
1859	int i;
1860	struct scrub_page *spage;
1861	char *kaddr;
1862
1863	BUG_ON(sblock->page_count < 1);
1864
1865	/* Each member in pagev is just one block, not a full page */
1866	ASSERT(sblock->page_count == num_sectors);
1867
1868	spage = sblock->pagev[0];
1869	kaddr = page_address(spage->page);
1870	h = (struct btrfs_header *)kaddr;
1871	memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
1872
1873	/*
1874	 * we don't use the getter functions here, as we
1875	 * a) don't have an extent buffer and
1876	 * b) the page is already kmapped
1877	 */
1878	if (spage->logical != btrfs_stack_header_bytenr(h))
1879		sblock->header_error = 1;
1880
1881	if (spage->generation != btrfs_stack_header_generation(h)) {
1882		sblock->header_error = 1;
1883		sblock->generation_error = 1;
1884	}
1885
1886	if (!scrub_check_fsid(h->fsid, spage))
1887		sblock->header_error = 1;
1888
1889	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1890		   BTRFS_UUID_SIZE))
1891		sblock->header_error = 1;
1892
1893	shash->tfm = fs_info->csum_shash;
1894	crypto_shash_init(shash);
1895	crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
1896			    sectorsize - BTRFS_CSUM_SIZE);
1897
1898	for (i = 1; i < num_sectors; i++) {
1899		kaddr = page_address(sblock->pagev[i]->page);
1900		crypto_shash_update(shash, kaddr, sectorsize);
 
 
 
 
 
 
 
 
 
 
 
1901	}
1902
1903	crypto_shash_final(shash, calculated_csum);
1904	if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size))
1905		sblock->checksum_error = 1;
1906
1907	return sblock->header_error || sblock->checksum_error;
1908}
1909
1910static int scrub_checksum_super(struct scrub_block *sblock)
1911{
1912	struct btrfs_super_block *s;
1913	struct scrub_ctx *sctx = sblock->sctx;
1914	struct btrfs_fs_info *fs_info = sctx->fs_info;
1915	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1916	u8 calculated_csum[BTRFS_CSUM_SIZE];
1917	struct scrub_page *spage;
1918	char *kaddr;
 
 
 
 
1919	int fail_gen = 0;
1920	int fail_cor = 0;
 
 
1921
1922	BUG_ON(sblock->page_count < 1);
1923	spage = sblock->pagev[0];
1924	kaddr = page_address(spage->page);
1925	s = (struct btrfs_super_block *)kaddr;
 
1926
1927	if (spage->logical != btrfs_super_bytenr(s))
1928		++fail_cor;
1929
1930	if (spage->generation != btrfs_super_generation(s))
1931		++fail_gen;
1932
1933	if (!scrub_check_fsid(s->fsid, spage))
1934		++fail_cor;
1935
1936	shash->tfm = fs_info->csum_shash;
1937	crypto_shash_init(shash);
1938	crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
1939			BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1940
1941	if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
 
1942		++fail_cor;
1943
1944	if (fail_cor + fail_gen) {
1945		/*
1946		 * if we find an error in a super block, we just report it.
1947		 * They will get written with the next transaction commit
1948		 * anyway
1949		 */
1950		spin_lock(&sctx->stat_lock);
1951		++sctx->stat.super_errors;
1952		spin_unlock(&sctx->stat_lock);
1953		if (fail_cor)
1954			btrfs_dev_stat_inc_and_print(spage->dev,
1955				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1956		else
1957			btrfs_dev_stat_inc_and_print(spage->dev,
1958				BTRFS_DEV_STAT_GENERATION_ERRS);
1959	}
1960
1961	return fail_cor + fail_gen;
1962}
1963
1964static void scrub_block_get(struct scrub_block *sblock)
1965{
1966	refcount_inc(&sblock->refs);
1967}
1968
1969static void scrub_block_put(struct scrub_block *sblock)
1970{
1971	if (refcount_dec_and_test(&sblock->refs)) {
1972		int i;
1973
1974		if (sblock->sparity)
1975			scrub_parity_put(sblock->sparity);
1976
1977		for (i = 0; i < sblock->page_count; i++)
1978			scrub_page_put(sblock->pagev[i]);
1979		kfree(sblock);
1980	}
1981}
1982
1983static void scrub_page_get(struct scrub_page *spage)
1984{
1985	atomic_inc(&spage->refs);
1986}
1987
1988static void scrub_page_put(struct scrub_page *spage)
1989{
1990	if (atomic_dec_and_test(&spage->refs)) {
1991		if (spage->page)
1992			__free_page(spage->page);
1993		kfree(spage);
1994	}
1995}
1996
1997/*
1998 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1999 * second.  Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
2000 */
2001static void scrub_throttle(struct scrub_ctx *sctx)
2002{
2003	const int time_slice = 1000;
2004	struct scrub_bio *sbio;
2005	struct btrfs_device *device;
2006	s64 delta;
2007	ktime_t now;
2008	u32 div;
2009	u64 bwlimit;
2010
2011	sbio = sctx->bios[sctx->curr];
2012	device = sbio->dev;
2013	bwlimit = READ_ONCE(device->scrub_speed_max);
2014	if (bwlimit == 0)
2015		return;
2016
2017	/*
2018	 * Slice is divided into intervals when the IO is submitted, adjust by
2019	 * bwlimit and maximum of 64 intervals.
2020	 */
2021	div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
2022	div = min_t(u32, 64, div);
2023
2024	/* Start new epoch, set deadline */
2025	now = ktime_get();
2026	if (sctx->throttle_deadline == 0) {
2027		sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
2028		sctx->throttle_sent = 0;
2029	}
2030
2031	/* Still in the time to send? */
2032	if (ktime_before(now, sctx->throttle_deadline)) {
2033		/* If current bio is within the limit, send it */
2034		sctx->throttle_sent += sbio->bio->bi_iter.bi_size;
2035		if (sctx->throttle_sent <= div_u64(bwlimit, div))
2036			return;
2037
2038		/* We're over the limit, sleep until the rest of the slice */
2039		delta = ktime_ms_delta(sctx->throttle_deadline, now);
2040	} else {
2041		/* New request after deadline, start new epoch */
2042		delta = 0;
2043	}
2044
2045	if (delta) {
2046		long timeout;
2047
2048		timeout = div_u64(delta * HZ, 1000);
2049		schedule_timeout_interruptible(timeout);
2050	}
2051
2052	/* Next call will start the deadline period */
2053	sctx->throttle_deadline = 0;
2054}
2055
2056static void scrub_submit(struct scrub_ctx *sctx)
2057{
2058	struct scrub_bio *sbio;
2059
2060	if (sctx->curr == -1)
2061		return;
2062
2063	scrub_throttle(sctx);
2064
2065	sbio = sctx->bios[sctx->curr];
2066	sctx->curr = -1;
2067	scrub_pending_bio_inc(sctx);
2068	btrfsic_submit_bio(sbio->bio);
2069}
2070
2071static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2072				    struct scrub_page *spage)
2073{
2074	struct scrub_block *sblock = spage->sblock;
2075	struct scrub_bio *sbio;
2076	const u32 sectorsize = sctx->fs_info->sectorsize;
2077	int ret;
2078
2079again:
2080	/*
2081	 * grab a fresh bio or wait for one to become available
2082	 */
2083	while (sctx->curr == -1) {
2084		spin_lock(&sctx->list_lock);
2085		sctx->curr = sctx->first_free;
2086		if (sctx->curr != -1) {
2087			sctx->first_free = sctx->bios[sctx->curr]->next_free;
2088			sctx->bios[sctx->curr]->next_free = -1;
2089			sctx->bios[sctx->curr]->page_count = 0;
2090			spin_unlock(&sctx->list_lock);
2091		} else {
2092			spin_unlock(&sctx->list_lock);
2093			wait_event(sctx->list_wait, sctx->first_free != -1);
2094		}
2095	}
2096	sbio = sctx->bios[sctx->curr];
2097	if (sbio->page_count == 0) {
2098		struct bio *bio;
2099
2100		sbio->physical = spage->physical;
2101		sbio->logical = spage->logical;
2102		sbio->dev = spage->dev;
2103		bio = sbio->bio;
2104		if (!bio) {
2105			bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
 
 
 
2106			sbio->bio = bio;
2107		}
2108
2109		bio->bi_private = sbio;
2110		bio->bi_end_io = scrub_bio_end_io;
2111		bio_set_dev(bio, sbio->dev->bdev);
2112		bio->bi_iter.bi_sector = sbio->physical >> 9;
2113		bio->bi_opf = REQ_OP_READ;
2114		sbio->status = 0;
2115	} else if (sbio->physical + sbio->page_count * sectorsize !=
2116		   spage->physical ||
2117		   sbio->logical + sbio->page_count * sectorsize !=
2118		   spage->logical ||
2119		   sbio->dev != spage->dev) {
2120		scrub_submit(sctx);
2121		goto again;
2122	}
2123
2124	sbio->pagev[sbio->page_count] = spage;
2125	ret = bio_add_page(sbio->bio, spage->page, sectorsize, 0);
2126	if (ret != sectorsize) {
2127		if (sbio->page_count < 1) {
2128			bio_put(sbio->bio);
2129			sbio->bio = NULL;
2130			return -EIO;
2131		}
2132		scrub_submit(sctx);
2133		goto again;
2134	}
2135
2136	scrub_block_get(sblock); /* one for the page added to the bio */
2137	atomic_inc(&sblock->outstanding_pages);
2138	sbio->page_count++;
2139	if (sbio->page_count == sctx->pages_per_rd_bio)
2140		scrub_submit(sctx);
2141
2142	return 0;
2143}
2144
2145static void scrub_missing_raid56_end_io(struct bio *bio)
2146{
2147	struct scrub_block *sblock = bio->bi_private;
2148	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2149
2150	if (bio->bi_status)
2151		sblock->no_io_error_seen = 0;
2152
2153	bio_put(bio);
2154
2155	btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2156}
2157
2158static void scrub_missing_raid56_worker(struct btrfs_work *work)
2159{
2160	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2161	struct scrub_ctx *sctx = sblock->sctx;
2162	struct btrfs_fs_info *fs_info = sctx->fs_info;
2163	u64 logical;
2164	struct btrfs_device *dev;
2165
2166	logical = sblock->pagev[0]->logical;
2167	dev = sblock->pagev[0]->dev;
2168
2169	if (sblock->no_io_error_seen)
2170		scrub_recheck_block_checksum(sblock);
2171
2172	if (!sblock->no_io_error_seen) {
2173		spin_lock(&sctx->stat_lock);
2174		sctx->stat.read_errors++;
2175		spin_unlock(&sctx->stat_lock);
2176		btrfs_err_rl_in_rcu(fs_info,
2177			"IO error rebuilding logical %llu for dev %s",
2178			logical, rcu_str_deref(dev->name));
2179	} else if (sblock->header_error || sblock->checksum_error) {
2180		spin_lock(&sctx->stat_lock);
2181		sctx->stat.uncorrectable_errors++;
2182		spin_unlock(&sctx->stat_lock);
2183		btrfs_err_rl_in_rcu(fs_info,
2184			"failed to rebuild valid logical %llu for dev %s",
2185			logical, rcu_str_deref(dev->name));
2186	} else {
2187		scrub_write_block_to_dev_replace(sblock);
2188	}
2189
2190	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2191		mutex_lock(&sctx->wr_lock);
 
 
 
2192		scrub_wr_submit(sctx);
2193		mutex_unlock(&sctx->wr_lock);
2194	}
2195
2196	scrub_block_put(sblock);
2197	scrub_pending_bio_dec(sctx);
2198}
2199
2200static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2201{
2202	struct scrub_ctx *sctx = sblock->sctx;
2203	struct btrfs_fs_info *fs_info = sctx->fs_info;
2204	u64 length = sblock->page_count * PAGE_SIZE;
2205	u64 logical = sblock->pagev[0]->logical;
2206	struct btrfs_bio *bbio = NULL;
2207	struct bio *bio;
2208	struct btrfs_raid_bio *rbio;
2209	int ret;
2210	int i;
2211
2212	btrfs_bio_counter_inc_blocked(fs_info);
2213	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2214			&length, &bbio);
2215	if (ret || !bbio || !bbio->raid_map)
2216		goto bbio_out;
2217
2218	if (WARN_ON(!sctx->is_dev_replace ||
2219		    !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2220		/*
2221		 * We shouldn't be scrubbing a missing device. Even for dev
2222		 * replace, we should only get here for RAID 5/6. We either
2223		 * managed to mount something with no mirrors remaining or
2224		 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2225		 */
2226		goto bbio_out;
2227	}
2228
2229	bio = btrfs_io_bio_alloc(0);
 
 
 
2230	bio->bi_iter.bi_sector = logical >> 9;
2231	bio->bi_private = sblock;
2232	bio->bi_end_io = scrub_missing_raid56_end_io;
2233
2234	rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2235	if (!rbio)
2236		goto rbio_out;
2237
2238	for (i = 0; i < sblock->page_count; i++) {
2239		struct scrub_page *spage = sblock->pagev[i];
2240
2241		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2242	}
2243
2244	btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
 
2245	scrub_block_get(sblock);
2246	scrub_pending_bio_inc(sctx);
2247	raid56_submit_missing_rbio(rbio);
2248	return;
2249
2250rbio_out:
2251	bio_put(bio);
2252bbio_out:
2253	btrfs_bio_counter_dec(fs_info);
2254	btrfs_put_bbio(bbio);
2255	spin_lock(&sctx->stat_lock);
2256	sctx->stat.malloc_errors++;
2257	spin_unlock(&sctx->stat_lock);
2258}
2259
2260static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u32 len,
2261		       u64 physical, struct btrfs_device *dev, u64 flags,
2262		       u64 gen, int mirror_num, u8 *csum,
2263		       u64 physical_for_dev_replace)
2264{
2265	struct scrub_block *sblock;
2266	const u32 sectorsize = sctx->fs_info->sectorsize;
2267	int index;
2268
2269	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2270	if (!sblock) {
2271		spin_lock(&sctx->stat_lock);
2272		sctx->stat.malloc_errors++;
2273		spin_unlock(&sctx->stat_lock);
2274		return -ENOMEM;
2275	}
2276
2277	/* one ref inside this function, plus one for each page added to
2278	 * a bio later on */
2279	refcount_set(&sblock->refs, 1);
2280	sblock->sctx = sctx;
2281	sblock->no_io_error_seen = 1;
2282
2283	for (index = 0; len > 0; index++) {
2284		struct scrub_page *spage;
2285		/*
2286		 * Here we will allocate one page for one sector to scrub.
2287		 * This is fine if PAGE_SIZE == sectorsize, but will cost
2288		 * more memory for PAGE_SIZE > sectorsize case.
2289		 */
2290		u32 l = min(sectorsize, len);
2291
2292		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2293		if (!spage) {
2294leave_nomem:
2295			spin_lock(&sctx->stat_lock);
2296			sctx->stat.malloc_errors++;
2297			spin_unlock(&sctx->stat_lock);
2298			scrub_block_put(sblock);
2299			return -ENOMEM;
2300		}
2301		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2302		scrub_page_get(spage);
2303		sblock->pagev[index] = spage;
2304		spage->sblock = sblock;
2305		spage->dev = dev;
2306		spage->flags = flags;
2307		spage->generation = gen;
2308		spage->logical = logical;
2309		spage->physical = physical;
2310		spage->physical_for_dev_replace = physical_for_dev_replace;
2311		spage->mirror_num = mirror_num;
2312		if (csum) {
2313			spage->have_csum = 1;
2314			memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2315		} else {
2316			spage->have_csum = 0;
2317		}
2318		sblock->page_count++;
2319		spage->page = alloc_page(GFP_KERNEL);
2320		if (!spage->page)
2321			goto leave_nomem;
2322		len -= l;
2323		logical += l;
2324		physical += l;
2325		physical_for_dev_replace += l;
2326	}
2327
2328	WARN_ON(sblock->page_count == 0);
2329	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2330		/*
2331		 * This case should only be hit for RAID 5/6 device replace. See
2332		 * the comment in scrub_missing_raid56_pages() for details.
2333		 */
2334		scrub_missing_raid56_pages(sblock);
2335	} else {
2336		for (index = 0; index < sblock->page_count; index++) {
2337			struct scrub_page *spage = sblock->pagev[index];
2338			int ret;
2339
2340			ret = scrub_add_page_to_rd_bio(sctx, spage);
2341			if (ret) {
2342				scrub_block_put(sblock);
2343				return ret;
2344			}
2345		}
2346
2347		if (flags & BTRFS_EXTENT_FLAG_SUPER)
2348			scrub_submit(sctx);
2349	}
2350
2351	/* last one frees, either here or in bio completion for last page */
2352	scrub_block_put(sblock);
2353	return 0;
2354}
2355
2356static void scrub_bio_end_io(struct bio *bio)
2357{
2358	struct scrub_bio *sbio = bio->bi_private;
2359	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2360
2361	sbio->status = bio->bi_status;
2362	sbio->bio = bio;
2363
2364	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2365}
2366
2367static void scrub_bio_end_io_worker(struct btrfs_work *work)
2368{
2369	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2370	struct scrub_ctx *sctx = sbio->sctx;
2371	int i;
2372
2373	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2374	if (sbio->status) {
2375		for (i = 0; i < sbio->page_count; i++) {
2376			struct scrub_page *spage = sbio->pagev[i];
2377
2378			spage->io_error = 1;
2379			spage->sblock->no_io_error_seen = 0;
2380		}
2381	}
2382
2383	/* now complete the scrub_block items that have all pages completed */
2384	for (i = 0; i < sbio->page_count; i++) {
2385		struct scrub_page *spage = sbio->pagev[i];
2386		struct scrub_block *sblock = spage->sblock;
2387
2388		if (atomic_dec_and_test(&sblock->outstanding_pages))
2389			scrub_block_complete(sblock);
2390		scrub_block_put(sblock);
2391	}
2392
2393	bio_put(sbio->bio);
2394	sbio->bio = NULL;
2395	spin_lock(&sctx->list_lock);
2396	sbio->next_free = sctx->first_free;
2397	sctx->first_free = sbio->index;
2398	spin_unlock(&sctx->list_lock);
2399
2400	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2401		mutex_lock(&sctx->wr_lock);
 
2402		scrub_wr_submit(sctx);
2403		mutex_unlock(&sctx->wr_lock);
2404	}
2405
2406	scrub_pending_bio_dec(sctx);
2407}
2408
2409static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2410				       unsigned long *bitmap,
2411				       u64 start, u32 len)
2412{
2413	u64 offset;
2414	u32 nsectors;
2415	u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2416
2417	if (len >= sparity->stripe_len) {
2418		bitmap_set(bitmap, 0, sparity->nsectors);
2419		return;
2420	}
2421
2422	start -= sparity->logic_start;
2423	start = div64_u64_rem(start, sparity->stripe_len, &offset);
2424	offset = offset >> sectorsize_bits;
2425	nsectors = len >> sectorsize_bits;
2426
2427	if (offset + nsectors <= sparity->nsectors) {
2428		bitmap_set(bitmap, offset, nsectors);
2429		return;
2430	}
2431
2432	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2433	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2434}
2435
2436static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2437						   u64 start, u32 len)
2438{
2439	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2440}
2441
2442static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2443						  u64 start, u32 len)
2444{
2445	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2446}
2447
2448static void scrub_block_complete(struct scrub_block *sblock)
2449{
2450	int corrupted = 0;
2451
2452	if (!sblock->no_io_error_seen) {
2453		corrupted = 1;
2454		scrub_handle_errored_block(sblock);
2455	} else {
2456		/*
2457		 * if has checksum error, write via repair mechanism in
2458		 * dev replace case, otherwise write here in dev replace
2459		 * case.
2460		 */
2461		corrupted = scrub_checksum(sblock);
2462		if (!corrupted && sblock->sctx->is_dev_replace)
2463			scrub_write_block_to_dev_replace(sblock);
2464	}
2465
2466	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2467		u64 start = sblock->pagev[0]->logical;
2468		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2469			  sblock->sctx->fs_info->sectorsize;
2470
2471		ASSERT(end - start <= U32_MAX);
2472		scrub_parity_mark_sectors_error(sblock->sparity,
2473						start, end - start);
2474	}
2475}
2476
2477static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2478{
2479	sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2480	list_del(&sum->list);
2481	kfree(sum);
2482}
2483
2484/*
2485 * Find the desired csum for range [logical, logical + sectorsize), and store
2486 * the csum into @csum.
2487 *
2488 * The search source is sctx->csum_list, which is a pre-populated list
2489 * storing bytenr ordered csum ranges.  We're responsible to cleanup any range
2490 * that is before @logical.
2491 *
2492 * Return 0 if there is no csum for the range.
2493 * Return 1 if there is csum for the range and copied to @csum.
2494 */
2495static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2496{
2497	bool found = false;
 
 
2498
2499	while (!list_empty(&sctx->csum_list)) {
2500		struct btrfs_ordered_sum *sum = NULL;
2501		unsigned long index;
2502		unsigned long num_sectors;
2503
2504		sum = list_first_entry(&sctx->csum_list,
2505				       struct btrfs_ordered_sum, list);
2506		/* The current csum range is beyond our range, no csum found */
2507		if (sum->bytenr > logical)
 
 
2508			break;
2509
2510		/*
2511		 * The current sum is before our bytenr, since scrub is always
2512		 * done in bytenr order, the csum will never be used anymore,
2513		 * clean it up so that later calls won't bother with the range,
2514		 * and continue search the next range.
2515		 */
2516		if (sum->bytenr + sum->len <= logical) {
2517			drop_csum_range(sctx, sum);
2518			continue;
2519		}
2520
2521		/* Now the csum range covers our bytenr, copy the csum */
2522		found = true;
2523		index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2524		num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2525
2526		memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2527		       sctx->fs_info->csum_size);
2528
2529		/* Cleanup the range if we're at the end of the csum range */
2530		if (index == num_sectors - 1)
2531			drop_csum_range(sctx, sum);
2532		break;
2533	}
2534	if (!found)
2535		return 0;
2536	return 1;
2537}
2538
2539/* scrub extent tries to collect up to 64 kB for each bio */
2540static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2541			u64 logical, u32 len,
2542			u64 physical, struct btrfs_device *dev, u64 flags,
2543			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2544{
2545	int ret;
2546	u8 csum[BTRFS_CSUM_SIZE];
2547	u32 blocksize;
2548
2549	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2550		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2551			blocksize = map->stripe_len;
2552		else
2553			blocksize = sctx->fs_info->sectorsize;
2554		spin_lock(&sctx->stat_lock);
2555		sctx->stat.data_extents_scrubbed++;
2556		sctx->stat.data_bytes_scrubbed += len;
2557		spin_unlock(&sctx->stat_lock);
2558	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2559		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2560			blocksize = map->stripe_len;
2561		else
2562			blocksize = sctx->fs_info->nodesize;
2563		spin_lock(&sctx->stat_lock);
2564		sctx->stat.tree_extents_scrubbed++;
2565		sctx->stat.tree_bytes_scrubbed += len;
2566		spin_unlock(&sctx->stat_lock);
2567	} else {
2568		blocksize = sctx->fs_info->sectorsize;
2569		WARN_ON(1);
2570	}
2571
2572	while (len) {
2573		u32 l = min(len, blocksize);
2574		int have_csum = 0;
2575
2576		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2577			/* push csums to sbio */
2578			have_csum = scrub_find_csum(sctx, logical, csum);
2579			if (have_csum == 0)
2580				++sctx->stat.no_csum;
 
 
 
 
 
 
2581		}
2582		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2583				  mirror_num, have_csum ? csum : NULL,
2584				  physical_for_dev_replace);
 
2585		if (ret)
2586			return ret;
2587		len -= l;
2588		logical += l;
2589		physical += l;
2590		physical_for_dev_replace += l;
2591	}
2592	return 0;
2593}
2594
2595static int scrub_pages_for_parity(struct scrub_parity *sparity,
2596				  u64 logical, u32 len,
2597				  u64 physical, struct btrfs_device *dev,
2598				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2599{
2600	struct scrub_ctx *sctx = sparity->sctx;
2601	struct scrub_block *sblock;
2602	const u32 sectorsize = sctx->fs_info->sectorsize;
2603	int index;
2604
2605	ASSERT(IS_ALIGNED(len, sectorsize));
2606
2607	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2608	if (!sblock) {
2609		spin_lock(&sctx->stat_lock);
2610		sctx->stat.malloc_errors++;
2611		spin_unlock(&sctx->stat_lock);
2612		return -ENOMEM;
2613	}
2614
2615	/* one ref inside this function, plus one for each page added to
2616	 * a bio later on */
2617	refcount_set(&sblock->refs, 1);
2618	sblock->sctx = sctx;
2619	sblock->no_io_error_seen = 1;
2620	sblock->sparity = sparity;
2621	scrub_parity_get(sparity);
2622
2623	for (index = 0; len > 0; index++) {
2624		struct scrub_page *spage;
 
2625
2626		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2627		if (!spage) {
2628leave_nomem:
2629			spin_lock(&sctx->stat_lock);
2630			sctx->stat.malloc_errors++;
2631			spin_unlock(&sctx->stat_lock);
2632			scrub_block_put(sblock);
2633			return -ENOMEM;
2634		}
2635		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2636		/* For scrub block */
2637		scrub_page_get(spage);
2638		sblock->pagev[index] = spage;
2639		/* For scrub parity */
2640		scrub_page_get(spage);
2641		list_add_tail(&spage->list, &sparity->spages);
2642		spage->sblock = sblock;
2643		spage->dev = dev;
2644		spage->flags = flags;
2645		spage->generation = gen;
2646		spage->logical = logical;
2647		spage->physical = physical;
2648		spage->mirror_num = mirror_num;
2649		if (csum) {
2650			spage->have_csum = 1;
2651			memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2652		} else {
2653			spage->have_csum = 0;
2654		}
2655		sblock->page_count++;
2656		spage->page = alloc_page(GFP_KERNEL);
2657		if (!spage->page)
2658			goto leave_nomem;
2659
2660
2661		/* Iterate over the stripe range in sectorsize steps */
2662		len -= sectorsize;
2663		logical += sectorsize;
2664		physical += sectorsize;
2665	}
2666
2667	WARN_ON(sblock->page_count == 0);
2668	for (index = 0; index < sblock->page_count; index++) {
2669		struct scrub_page *spage = sblock->pagev[index];
2670		int ret;
2671
2672		ret = scrub_add_page_to_rd_bio(sctx, spage);
2673		if (ret) {
2674			scrub_block_put(sblock);
2675			return ret;
2676		}
2677	}
2678
2679	/* last one frees, either here or in bio completion for last page */
2680	scrub_block_put(sblock);
2681	return 0;
2682}
2683
2684static int scrub_extent_for_parity(struct scrub_parity *sparity,
2685				   u64 logical, u32 len,
2686				   u64 physical, struct btrfs_device *dev,
2687				   u64 flags, u64 gen, int mirror_num)
2688{
2689	struct scrub_ctx *sctx = sparity->sctx;
2690	int ret;
2691	u8 csum[BTRFS_CSUM_SIZE];
2692	u32 blocksize;
2693
2694	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2695		scrub_parity_mark_sectors_error(sparity, logical, len);
2696		return 0;
2697	}
2698
2699	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2700		blocksize = sparity->stripe_len;
2701	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2702		blocksize = sparity->stripe_len;
2703	} else {
2704		blocksize = sctx->fs_info->sectorsize;
2705		WARN_ON(1);
2706	}
2707
2708	while (len) {
2709		u32 l = min(len, blocksize);
2710		int have_csum = 0;
2711
2712		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2713			/* push csums to sbio */
2714			have_csum = scrub_find_csum(sctx, logical, csum);
2715			if (have_csum == 0)
2716				goto skip;
2717		}
2718		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2719					     flags, gen, mirror_num,
2720					     have_csum ? csum : NULL);
2721		if (ret)
2722			return ret;
2723skip:
2724		len -= l;
2725		logical += l;
2726		physical += l;
2727	}
2728	return 0;
2729}
2730
2731/*
2732 * Given a physical address, this will calculate it's
2733 * logical offset. if this is a parity stripe, it will return
2734 * the most left data stripe's logical offset.
2735 *
2736 * return 0 if it is a data stripe, 1 means parity stripe.
2737 */
2738static int get_raid56_logic_offset(u64 physical, int num,
2739				   struct map_lookup *map, u64 *offset,
2740				   u64 *stripe_start)
2741{
2742	int i;
2743	int j = 0;
2744	u64 stripe_nr;
2745	u64 last_offset;
2746	u32 stripe_index;
2747	u32 rot;
2748	const int data_stripes = nr_data_stripes(map);
2749
2750	last_offset = (physical - map->stripes[num].physical) * data_stripes;
 
2751	if (stripe_start)
2752		*stripe_start = last_offset;
2753
2754	*offset = last_offset;
2755	for (i = 0; i < data_stripes; i++) {
2756		*offset = last_offset + i * map->stripe_len;
2757
2758		stripe_nr = div64_u64(*offset, map->stripe_len);
2759		stripe_nr = div_u64(stripe_nr, data_stripes);
2760
2761		/* Work out the disk rotation on this stripe-set */
2762		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2763		/* calculate which stripe this data locates */
2764		rot += i;
2765		stripe_index = rot % map->num_stripes;
2766		if (stripe_index == num)
2767			return 0;
2768		if (stripe_index < num)
2769			j++;
2770	}
2771	*offset = last_offset + j * map->stripe_len;
2772	return 1;
2773}
2774
2775static void scrub_free_parity(struct scrub_parity *sparity)
2776{
2777	struct scrub_ctx *sctx = sparity->sctx;
2778	struct scrub_page *curr, *next;
2779	int nbits;
2780
2781	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2782	if (nbits) {
2783		spin_lock(&sctx->stat_lock);
2784		sctx->stat.read_errors += nbits;
2785		sctx->stat.uncorrectable_errors += nbits;
2786		spin_unlock(&sctx->stat_lock);
2787	}
2788
2789	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2790		list_del_init(&curr->list);
2791		scrub_page_put(curr);
2792	}
2793
2794	kfree(sparity);
2795}
2796
2797static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2798{
2799	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2800						    work);
2801	struct scrub_ctx *sctx = sparity->sctx;
2802
2803	scrub_free_parity(sparity);
2804	scrub_pending_bio_dec(sctx);
2805}
2806
2807static void scrub_parity_bio_endio(struct bio *bio)
2808{
2809	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2810	struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2811
2812	if (bio->bi_status)
2813		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2814			  sparity->nsectors);
2815
2816	bio_put(bio);
2817
2818	btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2819			NULL);
2820	btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
 
2821}
2822
2823static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2824{
2825	struct scrub_ctx *sctx = sparity->sctx;
2826	struct btrfs_fs_info *fs_info = sctx->fs_info;
2827	struct bio *bio;
2828	struct btrfs_raid_bio *rbio;
 
2829	struct btrfs_bio *bbio = NULL;
2830	u64 length;
2831	int ret;
2832
2833	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2834			   sparity->nsectors))
2835		goto out;
2836
2837	length = sparity->logic_end - sparity->logic_start;
 
 
 
 
 
2838
2839	btrfs_bio_counter_inc_blocked(fs_info);
2840	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2841			       &length, &bbio);
2842	if (ret || !bbio || !bbio->raid_map)
2843		goto bbio_out;
2844
2845	bio = btrfs_io_bio_alloc(0);
2846	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2847	bio->bi_private = sparity;
2848	bio->bi_end_io = scrub_parity_bio_endio;
2849
2850	rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2851					      length, sparity->scrub_dev,
2852					      sparity->dbitmap,
2853					      sparity->nsectors);
2854	if (!rbio)
2855		goto rbio_out;
2856
 
 
 
2857	scrub_pending_bio_inc(sctx);
2858	raid56_parity_submit_scrub_rbio(rbio);
2859	return;
2860
2861rbio_out:
2862	bio_put(bio);
2863bbio_out:
2864	btrfs_bio_counter_dec(fs_info);
2865	btrfs_put_bbio(bbio);
2866	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2867		  sparity->nsectors);
2868	spin_lock(&sctx->stat_lock);
2869	sctx->stat.malloc_errors++;
2870	spin_unlock(&sctx->stat_lock);
2871out:
2872	scrub_free_parity(sparity);
2873}
2874
2875static inline int scrub_calc_parity_bitmap_len(int nsectors)
2876{
2877	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2878}
2879
2880static void scrub_parity_get(struct scrub_parity *sparity)
2881{
2882	refcount_inc(&sparity->refs);
2883}
2884
2885static void scrub_parity_put(struct scrub_parity *sparity)
2886{
2887	if (!refcount_dec_and_test(&sparity->refs))
2888		return;
2889
2890	scrub_parity_check_and_repair(sparity);
2891}
2892
2893static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2894						  struct map_lookup *map,
2895						  struct btrfs_device *sdev,
2896						  struct btrfs_path *path,
2897						  u64 logic_start,
2898						  u64 logic_end)
2899{
2900	struct btrfs_fs_info *fs_info = sctx->fs_info;
2901	struct btrfs_root *root = fs_info->extent_root;
2902	struct btrfs_root *csum_root = fs_info->csum_root;
2903	struct btrfs_extent_item *extent;
2904	struct btrfs_bio *bbio = NULL;
2905	u64 flags;
2906	int ret;
2907	int slot;
2908	struct extent_buffer *l;
2909	struct btrfs_key key;
2910	u64 generation;
2911	u64 extent_logical;
2912	u64 extent_physical;
2913	/* Check the comment in scrub_stripe() for why u32 is enough here */
2914	u32 extent_len;
2915	u64 mapped_length;
2916	struct btrfs_device *extent_dev;
2917	struct scrub_parity *sparity;
2918	int nsectors;
2919	int bitmap_len;
2920	int extent_mirror_num;
2921	int stop_loop = 0;
2922
2923	ASSERT(map->stripe_len <= U32_MAX);
2924	nsectors = map->stripe_len >> fs_info->sectorsize_bits;
2925	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2926	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2927			  GFP_NOFS);
2928	if (!sparity) {
2929		spin_lock(&sctx->stat_lock);
2930		sctx->stat.malloc_errors++;
2931		spin_unlock(&sctx->stat_lock);
2932		return -ENOMEM;
2933	}
2934
2935	ASSERT(map->stripe_len <= U32_MAX);
2936	sparity->stripe_len = map->stripe_len;
2937	sparity->nsectors = nsectors;
2938	sparity->sctx = sctx;
2939	sparity->scrub_dev = sdev;
2940	sparity->logic_start = logic_start;
2941	sparity->logic_end = logic_end;
2942	refcount_set(&sparity->refs, 1);
2943	INIT_LIST_HEAD(&sparity->spages);
2944	sparity->dbitmap = sparity->bitmap;
2945	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2946
2947	ret = 0;
2948	while (logic_start < logic_end) {
2949		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2950			key.type = BTRFS_METADATA_ITEM_KEY;
2951		else
2952			key.type = BTRFS_EXTENT_ITEM_KEY;
2953		key.objectid = logic_start;
2954		key.offset = (u64)-1;
2955
2956		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2957		if (ret < 0)
2958			goto out;
2959
2960		if (ret > 0) {
2961			ret = btrfs_previous_extent_item(root, path, 0);
2962			if (ret < 0)
2963				goto out;
2964			if (ret > 0) {
2965				btrfs_release_path(path);
2966				ret = btrfs_search_slot(NULL, root, &key,
2967							path, 0, 0);
2968				if (ret < 0)
2969					goto out;
2970			}
2971		}
2972
2973		stop_loop = 0;
2974		while (1) {
2975			u64 bytes;
2976
2977			l = path->nodes[0];
2978			slot = path->slots[0];
2979			if (slot >= btrfs_header_nritems(l)) {
2980				ret = btrfs_next_leaf(root, path);
2981				if (ret == 0)
2982					continue;
2983				if (ret < 0)
2984					goto out;
2985
2986				stop_loop = 1;
2987				break;
2988			}
2989			btrfs_item_key_to_cpu(l, &key, slot);
2990
2991			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2992			    key.type != BTRFS_METADATA_ITEM_KEY)
2993				goto next;
2994
2995			if (key.type == BTRFS_METADATA_ITEM_KEY)
2996				bytes = fs_info->nodesize;
2997			else
2998				bytes = key.offset;
2999
3000			if (key.objectid + bytes <= logic_start)
3001				goto next;
3002
3003			if (key.objectid >= logic_end) {
3004				stop_loop = 1;
3005				break;
3006			}
3007
3008			while (key.objectid >= logic_start + map->stripe_len)
3009				logic_start += map->stripe_len;
3010
3011			extent = btrfs_item_ptr(l, slot,
3012						struct btrfs_extent_item);
3013			flags = btrfs_extent_flags(l, extent);
3014			generation = btrfs_extent_generation(l, extent);
3015
3016			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3017			    (key.objectid < logic_start ||
3018			     key.objectid + bytes >
3019			     logic_start + map->stripe_len)) {
3020				btrfs_err(fs_info,
3021					  "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3022					  key.objectid, logic_start);
3023				spin_lock(&sctx->stat_lock);
3024				sctx->stat.uncorrectable_errors++;
3025				spin_unlock(&sctx->stat_lock);
3026				goto next;
3027			}
3028again:
3029			extent_logical = key.objectid;
3030			ASSERT(bytes <= U32_MAX);
3031			extent_len = bytes;
3032
3033			if (extent_logical < logic_start) {
3034				extent_len -= logic_start - extent_logical;
3035				extent_logical = logic_start;
3036			}
3037
3038			if (extent_logical + extent_len >
3039			    logic_start + map->stripe_len)
3040				extent_len = logic_start + map->stripe_len -
3041					     extent_logical;
3042
3043			scrub_parity_mark_sectors_data(sparity, extent_logical,
3044						       extent_len);
3045
3046			mapped_length = extent_len;
3047			bbio = NULL;
3048			ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3049					extent_logical, &mapped_length, &bbio,
3050					0);
3051			if (!ret) {
3052				if (!bbio || mapped_length < extent_len)
3053					ret = -EIO;
3054			}
3055			if (ret) {
3056				btrfs_put_bbio(bbio);
3057				goto out;
3058			}
3059			extent_physical = bbio->stripes[0].physical;
3060			extent_mirror_num = bbio->mirror_num;
3061			extent_dev = bbio->stripes[0].dev;
3062			btrfs_put_bbio(bbio);
3063
3064			ret = btrfs_lookup_csums_range(csum_root,
3065						extent_logical,
3066						extent_logical + extent_len - 1,
3067						&sctx->csum_list, 1);
3068			if (ret)
3069				goto out;
3070
3071			ret = scrub_extent_for_parity(sparity, extent_logical,
3072						      extent_len,
3073						      extent_physical,
3074						      extent_dev, flags,
3075						      generation,
3076						      extent_mirror_num);
3077
3078			scrub_free_csums(sctx);
3079
3080			if (ret)
3081				goto out;
3082
3083			if (extent_logical + extent_len <
3084			    key.objectid + bytes) {
3085				logic_start += map->stripe_len;
3086
3087				if (logic_start >= logic_end) {
3088					stop_loop = 1;
3089					break;
3090				}
3091
3092				if (logic_start < key.objectid + bytes) {
3093					cond_resched();
3094					goto again;
3095				}
3096			}
3097next:
3098			path->slots[0]++;
3099		}
3100
3101		btrfs_release_path(path);
3102
3103		if (stop_loop)
3104			break;
3105
3106		logic_start += map->stripe_len;
3107	}
3108out:
3109	if (ret < 0) {
3110		ASSERT(logic_end - logic_start <= U32_MAX);
3111		scrub_parity_mark_sectors_error(sparity, logic_start,
3112						logic_end - logic_start);
3113	}
3114	scrub_parity_put(sparity);
3115	scrub_submit(sctx);
3116	mutex_lock(&sctx->wr_lock);
3117	scrub_wr_submit(sctx);
3118	mutex_unlock(&sctx->wr_lock);
3119
3120	btrfs_release_path(path);
3121	return ret < 0 ? ret : 0;
3122}
3123
3124static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3125{
3126	if (!btrfs_is_zoned(sctx->fs_info))
3127		return;
3128
3129	sctx->flush_all_writes = true;
3130	scrub_submit(sctx);
3131	mutex_lock(&sctx->wr_lock);
3132	scrub_wr_submit(sctx);
3133	mutex_unlock(&sctx->wr_lock);
3134
3135	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3136}
3137
3138static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3139					u64 physical, u64 physical_end)
3140{
3141	struct btrfs_fs_info *fs_info = sctx->fs_info;
3142	int ret = 0;
3143
3144	if (!btrfs_is_zoned(fs_info))
3145		return 0;
3146
3147	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3148
3149	mutex_lock(&sctx->wr_lock);
3150	if (sctx->write_pointer < physical_end) {
3151		ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3152						    physical,
3153						    sctx->write_pointer);
3154		if (ret)
3155			btrfs_err(fs_info,
3156				  "zoned: failed to recover write pointer");
3157	}
3158	mutex_unlock(&sctx->wr_lock);
3159	btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3160
3161	return ret;
3162}
3163
3164static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3165					   struct map_lookup *map,
3166					   struct btrfs_device *scrub_dev,
3167					   int num, u64 base, u64 length,
3168					   struct btrfs_block_group *cache)
3169{
3170	struct btrfs_path *path, *ppath;
3171	struct btrfs_fs_info *fs_info = sctx->fs_info;
3172	struct btrfs_root *root = fs_info->extent_root;
3173	struct btrfs_root *csum_root = fs_info->csum_root;
3174	struct btrfs_extent_item *extent;
3175	struct blk_plug plug;
3176	u64 flags;
3177	int ret;
3178	int slot;
3179	u64 nstripes;
3180	struct extent_buffer *l;
 
3181	u64 physical;
3182	u64 logical;
3183	u64 logic_end;
3184	u64 physical_end;
3185	u64 generation;
3186	int mirror_num;
3187	struct reada_control *reada1;
3188	struct reada_control *reada2;
3189	struct btrfs_key key;
3190	struct btrfs_key key_end;
3191	u64 increment = map->stripe_len;
3192	u64 offset;
3193	u64 extent_logical;
3194	u64 extent_physical;
3195	/*
3196	 * Unlike chunk length, extent length should never go beyond
3197	 * BTRFS_MAX_EXTENT_SIZE, thus u32 is enough here.
3198	 */
3199	u32 extent_len;
3200	u64 stripe_logical;
3201	u64 stripe_end;
3202	struct btrfs_device *extent_dev;
3203	int extent_mirror_num;
3204	int stop_loop = 0;
3205
3206	physical = map->stripes[num].physical;
3207	offset = 0;
3208	nstripes = div64_u64(length, map->stripe_len);
3209	mirror_num = 1;
3210	increment = map->stripe_len;
3211	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3212		offset = map->stripe_len * num;
3213		increment = map->stripe_len * map->num_stripes;
 
3214	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3215		int factor = map->num_stripes / map->sub_stripes;
3216		offset = map->stripe_len * (num / map->sub_stripes);
3217		increment = map->stripe_len * factor;
3218		mirror_num = num % map->sub_stripes + 1;
3219	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
 
3220		mirror_num = num % map->num_stripes + 1;
3221	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
 
3222		mirror_num = num % map->num_stripes + 1;
3223	} else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3224		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3225		increment = map->stripe_len * nr_data_stripes(map);
 
 
 
 
3226	}
3227
3228	path = btrfs_alloc_path();
3229	if (!path)
3230		return -ENOMEM;
3231
3232	ppath = btrfs_alloc_path();
3233	if (!ppath) {
3234		btrfs_free_path(path);
3235		return -ENOMEM;
3236	}
3237
3238	/*
3239	 * work on commit root. The related disk blocks are static as
3240	 * long as COW is applied. This means, it is save to rewrite
3241	 * them to repair disk errors without any race conditions
3242	 */
3243	path->search_commit_root = 1;
3244	path->skip_locking = 1;
3245
3246	ppath->search_commit_root = 1;
3247	ppath->skip_locking = 1;
3248	/*
3249	 * trigger the readahead for extent tree csum tree and wait for
3250	 * completion. During readahead, the scrub is officially paused
3251	 * to not hold off transaction commits
3252	 */
3253	logical = base + offset;
3254	physical_end = physical + nstripes * map->stripe_len;
3255	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3256		get_raid56_logic_offset(physical_end, num,
3257					map, &logic_end, NULL);
3258		logic_end += base;
3259	} else {
3260		logic_end = logical + increment * nstripes;
3261	}
3262	wait_event(sctx->list_wait,
3263		   atomic_read(&sctx->bios_in_flight) == 0);
3264	scrub_blocked_if_needed(fs_info);
3265
3266	/* FIXME it might be better to start readahead at commit root */
3267	key.objectid = logical;
3268	key.type = BTRFS_EXTENT_ITEM_KEY;
3269	key.offset = (u64)0;
3270	key_end.objectid = logic_end;
3271	key_end.type = BTRFS_METADATA_ITEM_KEY;
3272	key_end.offset = (u64)-1;
3273	reada1 = btrfs_reada_add(root, &key, &key_end);
3274
3275	if (cache->flags & BTRFS_BLOCK_GROUP_DATA) {
3276		key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3277		key.type = BTRFS_EXTENT_CSUM_KEY;
3278		key.offset = logical;
3279		key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3280		key_end.type = BTRFS_EXTENT_CSUM_KEY;
3281		key_end.offset = logic_end;
3282		reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3283	} else {
3284		reada2 = NULL;
3285	}
3286
3287	if (!IS_ERR(reada1))
3288		btrfs_reada_wait(reada1);
3289	if (!IS_ERR_OR_NULL(reada2))
3290		btrfs_reada_wait(reada2);
3291
3292
3293	/*
3294	 * collect all data csums for the stripe to avoid seeking during
3295	 * the scrub. This might currently (crc32) end up to be about 1MB
3296	 */
3297	blk_start_plug(&plug);
3298
3299	if (sctx->is_dev_replace &&
3300	    btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3301		mutex_lock(&sctx->wr_lock);
3302		sctx->write_pointer = physical;
3303		mutex_unlock(&sctx->wr_lock);
3304		sctx->flush_all_writes = true;
3305	}
3306
3307	/*
3308	 * now find all extents for each stripe and scrub them
3309	 */
3310	ret = 0;
3311	while (physical < physical_end) {
3312		/*
3313		 * canceled?
3314		 */
3315		if (atomic_read(&fs_info->scrub_cancel_req) ||
3316		    atomic_read(&sctx->cancel_req)) {
3317			ret = -ECANCELED;
3318			goto out;
3319		}
3320		/*
3321		 * check to see if we have to pause
3322		 */
3323		if (atomic_read(&fs_info->scrub_pause_req)) {
3324			/* push queued extents */
3325			sctx->flush_all_writes = true;
3326			scrub_submit(sctx);
3327			mutex_lock(&sctx->wr_lock);
3328			scrub_wr_submit(sctx);
3329			mutex_unlock(&sctx->wr_lock);
3330			wait_event(sctx->list_wait,
3331				   atomic_read(&sctx->bios_in_flight) == 0);
3332			sctx->flush_all_writes = false;
3333			scrub_blocked_if_needed(fs_info);
3334		}
3335
3336		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3337			ret = get_raid56_logic_offset(physical, num, map,
3338						      &logical,
3339						      &stripe_logical);
3340			logical += base;
3341			if (ret) {
3342				/* it is parity strip */
3343				stripe_logical += base;
3344				stripe_end = stripe_logical + increment;
3345				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3346							  ppath, stripe_logical,
3347							  stripe_end);
3348				if (ret)
3349					goto out;
3350				goto skip;
3351			}
3352		}
3353
3354		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3355			key.type = BTRFS_METADATA_ITEM_KEY;
3356		else
3357			key.type = BTRFS_EXTENT_ITEM_KEY;
3358		key.objectid = logical;
3359		key.offset = (u64)-1;
3360
3361		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3362		if (ret < 0)
3363			goto out;
3364
3365		if (ret > 0) {
3366			ret = btrfs_previous_extent_item(root, path, 0);
3367			if (ret < 0)
3368				goto out;
3369			if (ret > 0) {
3370				/* there's no smaller item, so stick with the
3371				 * larger one */
3372				btrfs_release_path(path);
3373				ret = btrfs_search_slot(NULL, root, &key,
3374							path, 0, 0);
3375				if (ret < 0)
3376					goto out;
3377			}
3378		}
3379
3380		stop_loop = 0;
3381		while (1) {
3382			u64 bytes;
3383
3384			l = path->nodes[0];
3385			slot = path->slots[0];
3386			if (slot >= btrfs_header_nritems(l)) {
3387				ret = btrfs_next_leaf(root, path);
3388				if (ret == 0)
3389					continue;
3390				if (ret < 0)
3391					goto out;
3392
3393				stop_loop = 1;
3394				break;
3395			}
3396			btrfs_item_key_to_cpu(l, &key, slot);
3397
3398			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3399			    key.type != BTRFS_METADATA_ITEM_KEY)
3400				goto next;
3401
3402			if (key.type == BTRFS_METADATA_ITEM_KEY)
3403				bytes = fs_info->nodesize;
3404			else
3405				bytes = key.offset;
3406
3407			if (key.objectid + bytes <= logical)
3408				goto next;
3409
3410			if (key.objectid >= logical + map->stripe_len) {
3411				/* out of this device extent */
3412				if (key.objectid >= logic_end)
3413					stop_loop = 1;
3414				break;
3415			}
3416
3417			/*
3418			 * If our block group was removed in the meanwhile, just
3419			 * stop scrubbing since there is no point in continuing.
3420			 * Continuing would prevent reusing its device extents
3421			 * for new block groups for a long time.
3422			 */
3423			spin_lock(&cache->lock);
3424			if (cache->removed) {
3425				spin_unlock(&cache->lock);
3426				ret = 0;
3427				goto out;
3428			}
3429			spin_unlock(&cache->lock);
3430
3431			extent = btrfs_item_ptr(l, slot,
3432						struct btrfs_extent_item);
3433			flags = btrfs_extent_flags(l, extent);
3434			generation = btrfs_extent_generation(l, extent);
3435
3436			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3437			    (key.objectid < logical ||
3438			     key.objectid + bytes >
3439			     logical + map->stripe_len)) {
3440				btrfs_err(fs_info,
3441					   "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
 
3442				       key.objectid, logical);
3443				spin_lock(&sctx->stat_lock);
3444				sctx->stat.uncorrectable_errors++;
3445				spin_unlock(&sctx->stat_lock);
3446				goto next;
3447			}
3448
3449again:
3450			extent_logical = key.objectid;
3451			ASSERT(bytes <= U32_MAX);
3452			extent_len = bytes;
3453
3454			/*
3455			 * trim extent to this stripe
3456			 */
3457			if (extent_logical < logical) {
3458				extent_len -= logical - extent_logical;
3459				extent_logical = logical;
3460			}
3461			if (extent_logical + extent_len >
3462			    logical + map->stripe_len) {
3463				extent_len = logical + map->stripe_len -
3464					     extent_logical;
3465			}
3466
3467			extent_physical = extent_logical - logical + physical;
3468			extent_dev = scrub_dev;
3469			extent_mirror_num = mirror_num;
3470			if (sctx->is_dev_replace)
3471				scrub_remap_extent(fs_info, extent_logical,
3472						   extent_len, &extent_physical,
3473						   &extent_dev,
3474						   &extent_mirror_num);
3475
3476			if (flags & BTRFS_EXTENT_FLAG_DATA) {
3477				ret = btrfs_lookup_csums_range(csum_root,
3478						extent_logical,
3479						extent_logical + extent_len - 1,
3480						&sctx->csum_list, 1);
3481				if (ret)
3482					goto out;
3483			}
3484
3485			ret = scrub_extent(sctx, map, extent_logical, extent_len,
3486					   extent_physical, extent_dev, flags,
3487					   generation, extent_mirror_num,
3488					   extent_logical - logical + physical);
3489
3490			scrub_free_csums(sctx);
3491
3492			if (ret)
3493				goto out;
3494
3495			if (sctx->is_dev_replace)
3496				sync_replace_for_zoned(sctx);
3497
3498			if (extent_logical + extent_len <
3499			    key.objectid + bytes) {
3500				if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3501					/*
3502					 * loop until we find next data stripe
3503					 * or we have finished all stripes.
3504					 */
3505loop:
3506					physical += map->stripe_len;
3507					ret = get_raid56_logic_offset(physical,
3508							num, map, &logical,
3509							&stripe_logical);
3510					logical += base;
3511
3512					if (ret && physical < physical_end) {
3513						stripe_logical += base;
3514						stripe_end = stripe_logical +
3515								increment;
3516						ret = scrub_raid56_parity(sctx,
3517							map, scrub_dev, ppath,
3518							stripe_logical,
3519							stripe_end);
3520						if (ret)
3521							goto out;
3522						goto loop;
3523					}
3524				} else {
3525					physical += map->stripe_len;
3526					logical += increment;
3527				}
3528				if (logical < key.objectid + bytes) {
3529					cond_resched();
3530					goto again;
3531				}
3532
3533				if (physical >= physical_end) {
3534					stop_loop = 1;
3535					break;
3536				}
3537			}
3538next:
3539			path->slots[0]++;
3540		}
3541		btrfs_release_path(path);
3542skip:
3543		logical += increment;
3544		physical += map->stripe_len;
3545		spin_lock(&sctx->stat_lock);
3546		if (stop_loop)
3547			sctx->stat.last_physical = map->stripes[num].physical +
3548						   length;
3549		else
3550			sctx->stat.last_physical = physical;
3551		spin_unlock(&sctx->stat_lock);
3552		if (stop_loop)
3553			break;
3554	}
3555out:
3556	/* push queued extents */
3557	scrub_submit(sctx);
3558	mutex_lock(&sctx->wr_lock);
3559	scrub_wr_submit(sctx);
3560	mutex_unlock(&sctx->wr_lock);
3561
3562	blk_finish_plug(&plug);
3563	btrfs_free_path(path);
3564	btrfs_free_path(ppath);
3565
3566	if (sctx->is_dev_replace && ret >= 0) {
3567		int ret2;
3568
3569		ret2 = sync_write_pointer_for_zoned(sctx, base + offset,
3570						    map->stripes[num].physical,
3571						    physical_end);
3572		if (ret2)
3573			ret = ret2;
3574	}
3575
3576	return ret < 0 ? ret : 0;
3577}
3578
3579static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3580					  struct btrfs_device *scrub_dev,
3581					  u64 chunk_offset, u64 length,
3582					  u64 dev_offset,
3583					  struct btrfs_block_group *cache)
 
3584{
3585	struct btrfs_fs_info *fs_info = sctx->fs_info;
3586	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3587	struct map_lookup *map;
3588	struct extent_map *em;
3589	int i;
3590	int ret = 0;
3591
3592	read_lock(&map_tree->lock);
3593	em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3594	read_unlock(&map_tree->lock);
3595
3596	if (!em) {
3597		/*
3598		 * Might have been an unused block group deleted by the cleaner
3599		 * kthread or relocation.
3600		 */
3601		spin_lock(&cache->lock);
3602		if (!cache->removed)
3603			ret = -EINVAL;
3604		spin_unlock(&cache->lock);
3605
3606		return ret;
3607	}
3608
3609	map = em->map_lookup;
3610	if (em->start != chunk_offset)
3611		goto out;
3612
3613	if (em->len < length)
3614		goto out;
3615
3616	for (i = 0; i < map->num_stripes; ++i) {
3617		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3618		    map->stripes[i].physical == dev_offset) {
3619			ret = scrub_stripe(sctx, map, scrub_dev, i,
3620					   chunk_offset, length, cache);
 
3621			if (ret)
3622				goto out;
3623		}
3624	}
3625out:
3626	free_extent_map(em);
3627
3628	return ret;
3629}
3630
3631static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3632					  struct btrfs_block_group *cache)
3633{
3634	struct btrfs_fs_info *fs_info = cache->fs_info;
3635	struct btrfs_trans_handle *trans;
3636
3637	if (!btrfs_is_zoned(fs_info))
3638		return 0;
3639
3640	btrfs_wait_block_group_reservations(cache);
3641	btrfs_wait_nocow_writers(cache);
3642	btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3643
3644	trans = btrfs_join_transaction(root);
3645	if (IS_ERR(trans))
3646		return PTR_ERR(trans);
3647	return btrfs_commit_transaction(trans);
3648}
3649
3650static noinline_for_stack
3651int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3652			   struct btrfs_device *scrub_dev, u64 start, u64 end)
 
3653{
3654	struct btrfs_dev_extent *dev_extent = NULL;
3655	struct btrfs_path *path;
3656	struct btrfs_fs_info *fs_info = sctx->fs_info;
3657	struct btrfs_root *root = fs_info->dev_root;
3658	u64 length;
3659	u64 chunk_offset;
3660	int ret = 0;
3661	int ro_set;
3662	int slot;
3663	struct extent_buffer *l;
3664	struct btrfs_key key;
3665	struct btrfs_key found_key;
3666	struct btrfs_block_group *cache;
3667	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3668
3669	path = btrfs_alloc_path();
3670	if (!path)
3671		return -ENOMEM;
3672
3673	path->reada = READA_FORWARD;
3674	path->search_commit_root = 1;
3675	path->skip_locking = 1;
3676
3677	key.objectid = scrub_dev->devid;
3678	key.offset = 0ull;
3679	key.type = BTRFS_DEV_EXTENT_KEY;
3680
3681	while (1) {
3682		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3683		if (ret < 0)
3684			break;
3685		if (ret > 0) {
3686			if (path->slots[0] >=
3687			    btrfs_header_nritems(path->nodes[0])) {
3688				ret = btrfs_next_leaf(root, path);
3689				if (ret < 0)
3690					break;
3691				if (ret > 0) {
3692					ret = 0;
3693					break;
3694				}
3695			} else {
3696				ret = 0;
3697			}
3698		}
3699
3700		l = path->nodes[0];
3701		slot = path->slots[0];
3702
3703		btrfs_item_key_to_cpu(l, &found_key, slot);
3704
3705		if (found_key.objectid != scrub_dev->devid)
3706			break;
3707
3708		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3709			break;
3710
3711		if (found_key.offset >= end)
3712			break;
3713
3714		if (found_key.offset < key.offset)
3715			break;
3716
3717		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3718		length = btrfs_dev_extent_length(l, dev_extent);
3719
3720		if (found_key.offset + length <= start)
3721			goto skip;
3722
3723		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3724
3725		/*
3726		 * get a reference on the corresponding block group to prevent
3727		 * the chunk from going away while we scrub it
3728		 */
3729		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3730
3731		/* some chunks are removed but not committed to disk yet,
3732		 * continue scrubbing */
3733		if (!cache)
3734			goto skip;
3735
3736		if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3737			spin_lock(&cache->lock);
3738			if (!cache->to_copy) {
3739				spin_unlock(&cache->lock);
3740				btrfs_put_block_group(cache);
3741				goto skip;
3742			}
3743			spin_unlock(&cache->lock);
3744		}
3745
3746		/*
3747		 * Make sure that while we are scrubbing the corresponding block
3748		 * group doesn't get its logical address and its device extents
3749		 * reused for another block group, which can possibly be of a
3750		 * different type and different profile. We do this to prevent
3751		 * false error detections and crashes due to bogus attempts to
3752		 * repair extents.
3753		 */
3754		spin_lock(&cache->lock);
3755		if (cache->removed) {
3756			spin_unlock(&cache->lock);
3757			btrfs_put_block_group(cache);
3758			goto skip;
3759		}
3760		btrfs_freeze_block_group(cache);
3761		spin_unlock(&cache->lock);
3762
3763		/*
3764		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3765		 * to avoid deadlock caused by:
3766		 * btrfs_inc_block_group_ro()
3767		 * -> btrfs_wait_for_commit()
3768		 * -> btrfs_commit_transaction()
3769		 * -> btrfs_scrub_pause()
3770		 */
3771		scrub_pause_on(fs_info);
3772
3773		/*
3774		 * Don't do chunk preallocation for scrub.
3775		 *
3776		 * This is especially important for SYSTEM bgs, or we can hit
3777		 * -EFBIG from btrfs_finish_chunk_alloc() like:
3778		 * 1. The only SYSTEM bg is marked RO.
3779		 *    Since SYSTEM bg is small, that's pretty common.
3780		 * 2. New SYSTEM bg will be allocated
3781		 *    Due to regular version will allocate new chunk.
3782		 * 3. New SYSTEM bg is empty and will get cleaned up
3783		 *    Before cleanup really happens, it's marked RO again.
3784		 * 4. Empty SYSTEM bg get scrubbed
3785		 *    We go back to 2.
3786		 *
3787		 * This can easily boost the amount of SYSTEM chunks if cleaner
3788		 * thread can't be triggered fast enough, and use up all space
3789		 * of btrfs_super_block::sys_chunk_array
3790		 *
3791		 * While for dev replace, we need to try our best to mark block
3792		 * group RO, to prevent race between:
3793		 * - Write duplication
3794		 *   Contains latest data
3795		 * - Scrub copy
3796		 *   Contains data from commit tree
3797		 *
3798		 * If target block group is not marked RO, nocow writes can
3799		 * be overwritten by scrub copy, causing data corruption.
3800		 * So for dev-replace, it's not allowed to continue if a block
3801		 * group is not RO.
3802		 */
3803		ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3804		if (!ret && sctx->is_dev_replace) {
3805			ret = finish_extent_writes_for_zoned(root, cache);
3806			if (ret) {
3807				btrfs_dec_block_group_ro(cache);
3808				scrub_pause_off(fs_info);
3809				btrfs_put_block_group(cache);
3810				break;
3811			}
3812		}
3813
3814		if (ret == 0) {
3815			ro_set = 1;
3816		} else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3817			/*
3818			 * btrfs_inc_block_group_ro return -ENOSPC when it
3819			 * failed in creating new chunk for metadata.
3820			 * It is not a problem for scrub, because
3821			 * metadata are always cowed, and our scrub paused
3822			 * commit_transactions.
3823			 */
3824			ro_set = 0;
3825		} else if (ret == -ETXTBSY) {
3826			btrfs_warn(fs_info,
3827		   "skipping scrub of block group %llu due to active swapfile",
3828				   cache->start);
3829			scrub_pause_off(fs_info);
3830			ret = 0;
3831			goto skip_unfreeze;
3832		} else {
3833			btrfs_warn(fs_info,
3834				   "failed setting block group ro: %d", ret);
3835			btrfs_unfreeze_block_group(cache);
3836			btrfs_put_block_group(cache);
3837			scrub_pause_off(fs_info);
3838			break;
3839		}
3840
3841		/*
3842		 * Now the target block is marked RO, wait for nocow writes to
3843		 * finish before dev-replace.
3844		 * COW is fine, as COW never overwrites extents in commit tree.
3845		 */
3846		if (sctx->is_dev_replace) {
3847			btrfs_wait_nocow_writers(cache);
3848			btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3849					cache->length);
3850		}
3851
3852		scrub_pause_off(fs_info);
3853		down_write(&dev_replace->rwsem);
3854		dev_replace->cursor_right = found_key.offset + length;
3855		dev_replace->cursor_left = found_key.offset;
3856		dev_replace->item_needs_writeback = 1;
3857		up_write(&dev_replace->rwsem);
3858
3859		ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3860				  found_key.offset, cache);
3861
3862		/*
3863		 * flush, submit all pending read and write bios, afterwards
3864		 * wait for them.
3865		 * Note that in the dev replace case, a read request causes
3866		 * write requests that are submitted in the read completion
3867		 * worker. Therefore in the current situation, it is required
3868		 * that all write requests are flushed, so that all read and
3869		 * write requests are really completed when bios_in_flight
3870		 * changes to 0.
3871		 */
3872		sctx->flush_all_writes = true;
3873		scrub_submit(sctx);
3874		mutex_lock(&sctx->wr_lock);
3875		scrub_wr_submit(sctx);
3876		mutex_unlock(&sctx->wr_lock);
3877
3878		wait_event(sctx->list_wait,
3879			   atomic_read(&sctx->bios_in_flight) == 0);
3880
3881		scrub_pause_on(fs_info);
3882
3883		/*
3884		 * must be called before we decrease @scrub_paused.
3885		 * make sure we don't block transaction commit while
3886		 * we are waiting pending workers finished.
3887		 */
3888		wait_event(sctx->list_wait,
3889			   atomic_read(&sctx->workers_pending) == 0);
3890		sctx->flush_all_writes = false;
3891
3892		scrub_pause_off(fs_info);
3893
3894		if (sctx->is_dev_replace &&
3895		    !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
3896						      cache, found_key.offset))
3897			ro_set = 0;
3898
3899		down_write(&dev_replace->rwsem);
3900		dev_replace->cursor_left = dev_replace->cursor_right;
3901		dev_replace->item_needs_writeback = 1;
3902		up_write(&dev_replace->rwsem);
3903
3904		if (ro_set)
3905			btrfs_dec_block_group_ro(cache);
3906
3907		/*
3908		 * We might have prevented the cleaner kthread from deleting
3909		 * this block group if it was already unused because we raced
3910		 * and set it to RO mode first. So add it back to the unused
3911		 * list, otherwise it might not ever be deleted unless a manual
3912		 * balance is triggered or it becomes used and unused again.
3913		 */
3914		spin_lock(&cache->lock);
3915		if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3916		    cache->used == 0) {
3917			spin_unlock(&cache->lock);
3918			if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3919				btrfs_discard_queue_work(&fs_info->discard_ctl,
3920							 cache);
3921			else
3922				btrfs_mark_bg_unused(cache);
 
 
3923		} else {
3924			spin_unlock(&cache->lock);
3925		}
3926skip_unfreeze:
3927		btrfs_unfreeze_block_group(cache);
3928		btrfs_put_block_group(cache);
3929		if (ret)
3930			break;
3931		if (sctx->is_dev_replace &&
3932		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3933			ret = -EIO;
3934			break;
3935		}
3936		if (sctx->stat.malloc_errors > 0) {
3937			ret = -ENOMEM;
3938			break;
3939		}
 
 
 
3940skip:
3941		key.offset = found_key.offset + length;
3942		btrfs_release_path(path);
3943	}
3944
3945	btrfs_free_path(path);
3946
3947	return ret;
3948}
3949
3950static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3951					   struct btrfs_device *scrub_dev)
3952{
3953	int	i;
3954	u64	bytenr;
3955	u64	gen;
3956	int	ret;
3957	struct btrfs_fs_info *fs_info = sctx->fs_info;
3958
3959	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3960		return -EROFS;
3961
3962	/* Seed devices of a new filesystem has their own generation. */
3963	if (scrub_dev->fs_devices != fs_info->fs_devices)
3964		gen = scrub_dev->generation;
3965	else
3966		gen = fs_info->last_trans_committed;
3967
3968	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3969		bytenr = btrfs_sb_offset(i);
3970		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3971		    scrub_dev->commit_total_bytes)
3972			break;
3973		if (!btrfs_check_super_location(scrub_dev, bytenr))
3974			continue;
3975
3976		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3977				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3978				  NULL, bytenr);
3979		if (ret)
3980			return ret;
3981	}
3982	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3983
3984	return 0;
3985}
3986
3987static void scrub_workers_put(struct btrfs_fs_info *fs_info)
3988{
3989	if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
3990					&fs_info->scrub_lock)) {
3991		struct btrfs_workqueue *scrub_workers = NULL;
3992		struct btrfs_workqueue *scrub_wr_comp = NULL;
3993		struct btrfs_workqueue *scrub_parity = NULL;
3994
3995		scrub_workers = fs_info->scrub_workers;
3996		scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3997		scrub_parity = fs_info->scrub_parity_workers;
3998
3999		fs_info->scrub_workers = NULL;
4000		fs_info->scrub_wr_completion_workers = NULL;
4001		fs_info->scrub_parity_workers = NULL;
4002		mutex_unlock(&fs_info->scrub_lock);
4003
4004		btrfs_destroy_workqueue(scrub_workers);
4005		btrfs_destroy_workqueue(scrub_wr_comp);
4006		btrfs_destroy_workqueue(scrub_parity);
4007	}
4008}
4009
4010/*
4011 * get a reference count on fs_info->scrub_workers. start worker if necessary
4012 */
4013static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4014						int is_dev_replace)
4015{
4016	struct btrfs_workqueue *scrub_workers = NULL;
4017	struct btrfs_workqueue *scrub_wr_comp = NULL;
4018	struct btrfs_workqueue *scrub_parity = NULL;
4019	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4020	int max_active = fs_info->thread_pool_size;
4021	int ret = -ENOMEM;
4022
4023	if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
4024		return 0;
 
 
 
 
 
 
 
 
 
4025
4026	scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub", flags,
4027					      is_dev_replace ? 1 : max_active, 4);
4028	if (!scrub_workers)
4029		goto fail_scrub_workers;
 
4030
4031	scrub_wr_comp = btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
 
 
 
 
 
4032					      max_active, 2);
4033	if (!scrub_wr_comp)
4034		goto fail_scrub_wr_completion_workers;
4035
4036	scrub_parity = btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4037					     max_active, 2);
4038	if (!scrub_parity)
4039		goto fail_scrub_parity_workers;
4040
4041	mutex_lock(&fs_info->scrub_lock);
4042	if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
4043		ASSERT(fs_info->scrub_workers == NULL &&
4044		       fs_info->scrub_wr_completion_workers == NULL &&
4045		       fs_info->scrub_parity_workers == NULL);
4046		fs_info->scrub_workers = scrub_workers;
4047		fs_info->scrub_wr_completion_workers = scrub_wr_comp;
4048		fs_info->scrub_parity_workers = scrub_parity;
4049		refcount_set(&fs_info->scrub_workers_refcnt, 1);
4050		mutex_unlock(&fs_info->scrub_lock);
4051		return 0;
4052	}
4053	/* Other thread raced in and created the workers for us */
4054	refcount_inc(&fs_info->scrub_workers_refcnt);
4055	mutex_unlock(&fs_info->scrub_lock);
4056
4057	ret = 0;
4058	btrfs_destroy_workqueue(scrub_parity);
4059fail_scrub_parity_workers:
4060	btrfs_destroy_workqueue(scrub_wr_comp);
 
 
4061fail_scrub_wr_completion_workers:
4062	btrfs_destroy_workqueue(scrub_workers);
4063fail_scrub_workers:
4064	return ret;
 
 
 
 
 
 
 
 
 
 
 
4065}
4066
4067int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4068		    u64 end, struct btrfs_scrub_progress *progress,
4069		    int readonly, int is_dev_replace)
4070{
4071	struct scrub_ctx *sctx;
4072	int ret;
4073	struct btrfs_device *dev;
4074	unsigned int nofs_flag;
4075
4076	if (btrfs_fs_closing(fs_info))
4077		return -EAGAIN;
4078
4079	if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4080		/*
4081		 * in this case scrub is unable to calculate the checksum
4082		 * the way scrub is implemented. Do not handle this
4083		 * situation at all because it won't ever happen.
4084		 */
4085		btrfs_err(fs_info,
4086			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4087		       fs_info->nodesize,
4088		       BTRFS_STRIPE_LEN);
4089		return -EINVAL;
4090	}
4091
4092	if (fs_info->nodesize >
 
 
 
 
 
 
 
 
 
4093	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4094	    fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
 
4095		/*
4096		 * would exhaust the array bounds of pagev member in
4097		 * struct scrub_block
4098		 */
4099		btrfs_err(fs_info,
4100			  "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4101		       fs_info->nodesize,
4102		       SCRUB_MAX_PAGES_PER_BLOCK,
4103		       fs_info->sectorsize,
4104		       SCRUB_MAX_PAGES_PER_BLOCK);
4105		return -EINVAL;
4106	}
4107
4108	/* Allocate outside of device_list_mutex */
4109	sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4110	if (IS_ERR(sctx))
4111		return PTR_ERR(sctx);
4112
4113	ret = scrub_workers_get(fs_info, is_dev_replace);
4114	if (ret)
4115		goto out_free_ctx;
4116
4117	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4118	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4119	if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4120		     !is_dev_replace)) {
4121		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4122		ret = -ENODEV;
4123		goto out;
4124	}
4125
4126	if (!is_dev_replace && !readonly &&
4127	    !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4128		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4129		btrfs_err_in_rcu(fs_info,
4130			"scrub on devid %llu: filesystem on %s is not writable",
4131				 devid, rcu_str_deref(dev->name));
4132		ret = -EROFS;
4133		goto out;
 
4134	}
4135
4136	mutex_lock(&fs_info->scrub_lock);
4137	if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4138	    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4139		mutex_unlock(&fs_info->scrub_lock);
4140		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4141		ret = -EIO;
4142		goto out;
4143	}
4144
4145	down_read(&fs_info->dev_replace.rwsem);
4146	if (dev->scrub_ctx ||
4147	    (!is_dev_replace &&
4148	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4149		up_read(&fs_info->dev_replace.rwsem);
4150		mutex_unlock(&fs_info->scrub_lock);
4151		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4152		ret = -EINPROGRESS;
4153		goto out;
 
 
 
 
 
 
 
4154	}
4155	up_read(&fs_info->dev_replace.rwsem);
4156
 
 
 
 
 
 
 
4157	sctx->readonly = readonly;
4158	dev->scrub_ctx = sctx;
4159	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4160
4161	/*
4162	 * checking @scrub_pause_req here, we can avoid
4163	 * race between committing transaction and scrubbing.
4164	 */
4165	__scrub_blocked_if_needed(fs_info);
4166	atomic_inc(&fs_info->scrubs_running);
4167	mutex_unlock(&fs_info->scrub_lock);
4168
4169	/*
4170	 * In order to avoid deadlock with reclaim when there is a transaction
4171	 * trying to pause scrub, make sure we use GFP_NOFS for all the
4172	 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
4173	 * invoked by our callees. The pausing request is done when the
4174	 * transaction commit starts, and it blocks the transaction until scrub
4175	 * is paused (done at specific points at scrub_stripe() or right above
4176	 * before incrementing fs_info->scrubs_running).
4177	 */
4178	nofs_flag = memalloc_nofs_save();
4179	if (!is_dev_replace) {
4180		btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4181		/*
4182		 * by holding device list mutex, we can
4183		 * kick off writing super in log tree sync.
4184		 */
4185		mutex_lock(&fs_info->fs_devices->device_list_mutex);
4186		ret = scrub_supers(sctx, dev);
4187		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4188	}
4189
4190	if (!ret)
4191		ret = scrub_enumerate_chunks(sctx, dev, start, end);
4192	memalloc_nofs_restore(nofs_flag);
4193
4194	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4195	atomic_dec(&fs_info->scrubs_running);
4196	wake_up(&fs_info->scrub_pause_wait);
4197
4198	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4199
4200	if (progress)
4201		memcpy(progress, &sctx->stat, sizeof(*progress));
4202
4203	if (!is_dev_replace)
4204		btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4205			ret ? "not finished" : "finished", devid, ret);
4206
4207	mutex_lock(&fs_info->scrub_lock);
4208	dev->scrub_ctx = NULL;
 
4209	mutex_unlock(&fs_info->scrub_lock);
4210
4211	scrub_workers_put(fs_info);
4212	scrub_put_ctx(sctx);
4213
4214	return ret;
4215out:
4216	scrub_workers_put(fs_info);
4217out_free_ctx:
4218	scrub_free_ctx(sctx);
4219
4220	return ret;
4221}
4222
4223void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4224{
 
 
4225	mutex_lock(&fs_info->scrub_lock);
4226	atomic_inc(&fs_info->scrub_pause_req);
4227	while (atomic_read(&fs_info->scrubs_paused) !=
4228	       atomic_read(&fs_info->scrubs_running)) {
4229		mutex_unlock(&fs_info->scrub_lock);
4230		wait_event(fs_info->scrub_pause_wait,
4231			   atomic_read(&fs_info->scrubs_paused) ==
4232			   atomic_read(&fs_info->scrubs_running));
4233		mutex_lock(&fs_info->scrub_lock);
4234	}
4235	mutex_unlock(&fs_info->scrub_lock);
4236}
4237
4238void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4239{
 
 
4240	atomic_dec(&fs_info->scrub_pause_req);
4241	wake_up(&fs_info->scrub_pause_wait);
4242}
4243
4244int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4245{
4246	mutex_lock(&fs_info->scrub_lock);
4247	if (!atomic_read(&fs_info->scrubs_running)) {
4248		mutex_unlock(&fs_info->scrub_lock);
4249		return -ENOTCONN;
4250	}
4251
4252	atomic_inc(&fs_info->scrub_cancel_req);
4253	while (atomic_read(&fs_info->scrubs_running)) {
4254		mutex_unlock(&fs_info->scrub_lock);
4255		wait_event(fs_info->scrub_pause_wait,
4256			   atomic_read(&fs_info->scrubs_running) == 0);
4257		mutex_lock(&fs_info->scrub_lock);
4258	}
4259	atomic_dec(&fs_info->scrub_cancel_req);
4260	mutex_unlock(&fs_info->scrub_lock);
4261
4262	return 0;
4263}
4264
4265int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
 
4266{
4267	struct btrfs_fs_info *fs_info = dev->fs_info;
4268	struct scrub_ctx *sctx;
4269
4270	mutex_lock(&fs_info->scrub_lock);
4271	sctx = dev->scrub_ctx;
4272	if (!sctx) {
4273		mutex_unlock(&fs_info->scrub_lock);
4274		return -ENOTCONN;
4275	}
4276	atomic_inc(&sctx->cancel_req);
4277	while (dev->scrub_ctx) {
4278		mutex_unlock(&fs_info->scrub_lock);
4279		wait_event(fs_info->scrub_pause_wait,
4280			   dev->scrub_ctx == NULL);
4281		mutex_lock(&fs_info->scrub_lock);
4282	}
4283	mutex_unlock(&fs_info->scrub_lock);
4284
4285	return 0;
4286}
4287
4288int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4289			 struct btrfs_scrub_progress *progress)
4290{
4291	struct btrfs_device *dev;
4292	struct scrub_ctx *sctx = NULL;
4293
4294	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4295	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4296	if (dev)
4297		sctx = dev->scrub_ctx;
4298	if (sctx)
4299		memcpy(progress, &sctx->stat, sizeof(*progress));
4300	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4301
4302	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4303}
4304
4305static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4306			       u64 extent_logical, u32 extent_len,
4307			       u64 *extent_physical,
4308			       struct btrfs_device **extent_dev,
4309			       int *extent_mirror_num)
4310{
4311	u64 mapped_length;
4312	struct btrfs_bio *bbio = NULL;
4313	int ret;
4314
4315	mapped_length = extent_len;
4316	ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4317			      &mapped_length, &bbio, 0);
4318	if (ret || !bbio || mapped_length < extent_len ||
4319	    !bbio->stripes[0].dev->bdev) {
4320		btrfs_put_bbio(bbio);
4321		return;
4322	}
4323
4324	*extent_physical = bbio->stripes[0].physical;
4325	*extent_mirror_num = bbio->mirror_num;
4326	*extent_dev = bbio->stripes[0].dev;
4327	btrfs_put_bbio(bbio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4328}