xref: /openbmc/linux/fs/btrfs/scrub.c (revision a20eefae)
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 "volumes.h"
12 #include "disk-io.h"
13 #include "ordered-data.h"
14 #include "transaction.h"
15 #include "backref.h"
16 #include "extent_io.h"
17 #include "dev-replace.h"
18 #include "check-integrity.h"
19 #include "rcu-string.h"
20 #include "raid56.h"
21 
22 /*
23  * This is only the first step towards a full-features scrub. It reads all
24  * extent and super block and verifies the checksums. In case a bad checksum
25  * is found or the extent cannot be read, good data will be written back if
26  * any can be found.
27  *
28  * Future enhancements:
29  *  - In case an unrepairable extent is encountered, track which files are
30  *    affected and report them
31  *  - track and record media errors, throw out bad devices
32  *  - add a mode to also read unallocated space
33  */
34 
35 struct scrub_block;
36 struct scrub_ctx;
37 
38 /*
39  * the following three values only influence the performance.
40  * The last one configures the number of parallel and outstanding I/O
41  * operations. The first two values configure an upper limit for the number
42  * of (dynamically allocated) pages that are added to a bio.
43  */
44 #define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
45 #define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
46 #define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
47 
48 /*
49  * the following value times PAGE_SIZE needs to be large enough to match the
50  * largest node/leaf/sector size that shall be supported.
51  * Values larger than BTRFS_STRIPE_LEN are not supported.
52  */
53 #define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
54 
55 struct scrub_recover {
56 	refcount_t		refs;
57 	struct btrfs_bio	*bbio;
58 	u64			map_length;
59 };
60 
61 struct scrub_page {
62 	struct scrub_block	*sblock;
63 	struct page		*page;
64 	struct btrfs_device	*dev;
65 	struct list_head	list;
66 	u64			flags;  /* extent flags */
67 	u64			generation;
68 	u64			logical;
69 	u64			physical;
70 	u64			physical_for_dev_replace;
71 	atomic_t		refs;
72 	struct {
73 		unsigned int	mirror_num:8;
74 		unsigned int	have_csum:1;
75 		unsigned int	io_error:1;
76 	};
77 	u8			csum[BTRFS_CSUM_SIZE];
78 
79 	struct scrub_recover	*recover;
80 };
81 
82 struct scrub_bio {
83 	int			index;
84 	struct scrub_ctx	*sctx;
85 	struct btrfs_device	*dev;
86 	struct bio		*bio;
87 	blk_status_t		status;
88 	u64			logical;
89 	u64			physical;
90 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
91 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
92 #else
93 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
94 #endif
95 	int			page_count;
96 	int			next_free;
97 	struct btrfs_work	work;
98 };
99 
100 struct scrub_block {
101 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
102 	int			page_count;
103 	atomic_t		outstanding_pages;
104 	refcount_t		refs; /* free mem on transition to zero */
105 	struct scrub_ctx	*sctx;
106 	struct scrub_parity	*sparity;
107 	struct {
108 		unsigned int	header_error:1;
109 		unsigned int	checksum_error:1;
110 		unsigned int	no_io_error_seen:1;
111 		unsigned int	generation_error:1; /* also sets header_error */
112 
113 		/* The following is for the data used to check parity */
114 		/* It is for the data with checksum */
115 		unsigned int	data_corrected:1;
116 	};
117 	struct btrfs_work	work;
118 };
119 
120 /* Used for the chunks with parity stripe such RAID5/6 */
121 struct scrub_parity {
122 	struct scrub_ctx	*sctx;
123 
124 	struct btrfs_device	*scrub_dev;
125 
126 	u64			logic_start;
127 
128 	u64			logic_end;
129 
130 	int			nsectors;
131 
132 	u64			stripe_len;
133 
134 	refcount_t		refs;
135 
136 	struct list_head	spages;
137 
138 	/* Work of parity check and repair */
139 	struct btrfs_work	work;
140 
141 	/* Mark the parity blocks which have data */
142 	unsigned long		*dbitmap;
143 
144 	/*
145 	 * Mark the parity blocks which have data, but errors happen when
146 	 * read data or check data
147 	 */
148 	unsigned long		*ebitmap;
149 
150 	unsigned long		bitmap[0];
151 };
152 
153 struct scrub_ctx {
154 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
155 	struct btrfs_fs_info	*fs_info;
156 	int			first_free;
157 	int			curr;
158 	atomic_t		bios_in_flight;
159 	atomic_t		workers_pending;
160 	spinlock_t		list_lock;
161 	wait_queue_head_t	list_wait;
162 	u16			csum_size;
163 	struct list_head	csum_list;
164 	atomic_t		cancel_req;
165 	int			readonly;
166 	int			pages_per_rd_bio;
167 
168 	int			is_dev_replace;
169 
170 	struct scrub_bio        *wr_curr_bio;
171 	struct mutex            wr_lock;
172 	int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
173 	struct btrfs_device     *wr_tgtdev;
174 	bool                    flush_all_writes;
175 
176 	/*
177 	 * statistics
178 	 */
179 	struct btrfs_scrub_progress stat;
180 	spinlock_t		stat_lock;
181 
182 	/*
183 	 * Use a ref counter to avoid use-after-free issues. Scrub workers
184 	 * decrement bios_in_flight and workers_pending and then do a wakeup
185 	 * on the list_wait wait queue. We must ensure the main scrub task
186 	 * doesn't free the scrub context before or while the workers are
187 	 * doing the wakeup() call.
188 	 */
189 	refcount_t              refs;
190 };
191 
192 struct scrub_warning {
193 	struct btrfs_path	*path;
194 	u64			extent_item_size;
195 	const char		*errstr;
196 	u64			physical;
197 	u64			logical;
198 	struct btrfs_device	*dev;
199 };
200 
201 struct full_stripe_lock {
202 	struct rb_node node;
203 	u64 logical;
204 	u64 refs;
205 	struct mutex mutex;
206 };
207 
208 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
209 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
210 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
211 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
212 				     struct scrub_block *sblocks_for_recheck);
213 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
214 				struct scrub_block *sblock,
215 				int retry_failed_mirror);
216 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
217 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
218 					     struct scrub_block *sblock_good);
219 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
220 					    struct scrub_block *sblock_good,
221 					    int page_num, int force_write);
222 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
223 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
224 					   int page_num);
225 static int scrub_checksum_data(struct scrub_block *sblock);
226 static int scrub_checksum_tree_block(struct scrub_block *sblock);
227 static int scrub_checksum_super(struct scrub_block *sblock);
228 static void scrub_block_get(struct scrub_block *sblock);
229 static void scrub_block_put(struct scrub_block *sblock);
230 static void scrub_page_get(struct scrub_page *spage);
231 static void scrub_page_put(struct scrub_page *spage);
232 static void scrub_parity_get(struct scrub_parity *sparity);
233 static void scrub_parity_put(struct scrub_parity *sparity);
234 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
235 				    struct scrub_page *spage);
236 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
237 		       u64 physical, struct btrfs_device *dev, u64 flags,
238 		       u64 gen, int mirror_num, u8 *csum, int force,
239 		       u64 physical_for_dev_replace);
240 static void scrub_bio_end_io(struct bio *bio);
241 static void scrub_bio_end_io_worker(struct btrfs_work *work);
242 static void scrub_block_complete(struct scrub_block *sblock);
243 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
244 			       u64 extent_logical, u64 extent_len,
245 			       u64 *extent_physical,
246 			       struct btrfs_device **extent_dev,
247 			       int *extent_mirror_num);
248 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
249 				    struct scrub_page *spage);
250 static void scrub_wr_submit(struct scrub_ctx *sctx);
251 static void scrub_wr_bio_end_io(struct bio *bio);
252 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
253 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
254 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
255 static void scrub_put_ctx(struct scrub_ctx *sctx);
256 
257 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
258 {
259 	return page->recover &&
260 	       (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
261 }
262 
263 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
264 {
265 	refcount_inc(&sctx->refs);
266 	atomic_inc(&sctx->bios_in_flight);
267 }
268 
269 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
270 {
271 	atomic_dec(&sctx->bios_in_flight);
272 	wake_up(&sctx->list_wait);
273 	scrub_put_ctx(sctx);
274 }
275 
276 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
277 {
278 	while (atomic_read(&fs_info->scrub_pause_req)) {
279 		mutex_unlock(&fs_info->scrub_lock);
280 		wait_event(fs_info->scrub_pause_wait,
281 		   atomic_read(&fs_info->scrub_pause_req) == 0);
282 		mutex_lock(&fs_info->scrub_lock);
283 	}
284 }
285 
286 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
287 {
288 	atomic_inc(&fs_info->scrubs_paused);
289 	wake_up(&fs_info->scrub_pause_wait);
290 }
291 
292 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
293 {
294 	mutex_lock(&fs_info->scrub_lock);
295 	__scrub_blocked_if_needed(fs_info);
296 	atomic_dec(&fs_info->scrubs_paused);
297 	mutex_unlock(&fs_info->scrub_lock);
298 
299 	wake_up(&fs_info->scrub_pause_wait);
300 }
301 
302 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
303 {
304 	scrub_pause_on(fs_info);
305 	scrub_pause_off(fs_info);
306 }
307 
308 /*
309  * Insert new full stripe lock into full stripe locks tree
310  *
311  * Return pointer to existing or newly inserted full_stripe_lock structure if
312  * everything works well.
313  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
314  *
315  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
316  * function
317  */
318 static struct full_stripe_lock *insert_full_stripe_lock(
319 		struct btrfs_full_stripe_locks_tree *locks_root,
320 		u64 fstripe_logical)
321 {
322 	struct rb_node **p;
323 	struct rb_node *parent = NULL;
324 	struct full_stripe_lock *entry;
325 	struct full_stripe_lock *ret;
326 
327 	lockdep_assert_held(&locks_root->lock);
328 
329 	p = &locks_root->root.rb_node;
330 	while (*p) {
331 		parent = *p;
332 		entry = rb_entry(parent, struct full_stripe_lock, node);
333 		if (fstripe_logical < entry->logical) {
334 			p = &(*p)->rb_left;
335 		} else if (fstripe_logical > entry->logical) {
336 			p = &(*p)->rb_right;
337 		} else {
338 			entry->refs++;
339 			return entry;
340 		}
341 	}
342 
343 	/*
344 	 * Insert new lock.
345 	 */
346 	ret = kmalloc(sizeof(*ret), GFP_KERNEL);
347 	if (!ret)
348 		return ERR_PTR(-ENOMEM);
349 	ret->logical = fstripe_logical;
350 	ret->refs = 1;
351 	mutex_init(&ret->mutex);
352 
353 	rb_link_node(&ret->node, parent, p);
354 	rb_insert_color(&ret->node, &locks_root->root);
355 	return ret;
356 }
357 
358 /*
359  * Search for a full stripe lock of a block group
360  *
361  * Return pointer to existing full stripe lock if found
362  * Return NULL if not found
363  */
364 static struct full_stripe_lock *search_full_stripe_lock(
365 		struct btrfs_full_stripe_locks_tree *locks_root,
366 		u64 fstripe_logical)
367 {
368 	struct rb_node *node;
369 	struct full_stripe_lock *entry;
370 
371 	lockdep_assert_held(&locks_root->lock);
372 
373 	node = locks_root->root.rb_node;
374 	while (node) {
375 		entry = rb_entry(node, struct full_stripe_lock, node);
376 		if (fstripe_logical < entry->logical)
377 			node = node->rb_left;
378 		else if (fstripe_logical > entry->logical)
379 			node = node->rb_right;
380 		else
381 			return entry;
382 	}
383 	return NULL;
384 }
385 
386 /*
387  * Helper to get full stripe logical from a normal bytenr.
388  *
389  * Caller must ensure @cache is a RAID56 block group.
390  */
391 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
392 				   u64 bytenr)
393 {
394 	u64 ret;
395 
396 	/*
397 	 * Due to chunk item size limit, full stripe length should not be
398 	 * larger than U32_MAX. Just a sanity check here.
399 	 */
400 	WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
401 
402 	/*
403 	 * round_down() can only handle power of 2, while RAID56 full
404 	 * stripe length can be 64KiB * n, so we need to manually round down.
405 	 */
406 	ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
407 		cache->full_stripe_len + cache->key.objectid;
408 	return ret;
409 }
410 
411 /*
412  * Lock a full stripe to avoid concurrency of recovery and read
413  *
414  * It's only used for profiles with parities (RAID5/6), for other profiles it
415  * does nothing.
416  *
417  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
418  * So caller must call unlock_full_stripe() at the same context.
419  *
420  * Return <0 if encounters error.
421  */
422 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
423 			    bool *locked_ret)
424 {
425 	struct btrfs_block_group_cache *bg_cache;
426 	struct btrfs_full_stripe_locks_tree *locks_root;
427 	struct full_stripe_lock *existing;
428 	u64 fstripe_start;
429 	int ret = 0;
430 
431 	*locked_ret = false;
432 	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
433 	if (!bg_cache) {
434 		ASSERT(0);
435 		return -ENOENT;
436 	}
437 
438 	/* Profiles not based on parity don't need full stripe lock */
439 	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
440 		goto out;
441 	locks_root = &bg_cache->full_stripe_locks_root;
442 
443 	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
444 
445 	/* Now insert the full stripe lock */
446 	mutex_lock(&locks_root->lock);
447 	existing = insert_full_stripe_lock(locks_root, fstripe_start);
448 	mutex_unlock(&locks_root->lock);
449 	if (IS_ERR(existing)) {
450 		ret = PTR_ERR(existing);
451 		goto out;
452 	}
453 	mutex_lock(&existing->mutex);
454 	*locked_ret = true;
455 out:
456 	btrfs_put_block_group(bg_cache);
457 	return ret;
458 }
459 
460 /*
461  * Unlock a full stripe.
462  *
463  * NOTE: Caller must ensure it's the same context calling corresponding
464  * lock_full_stripe().
465  *
466  * Return 0 if we unlock full stripe without problem.
467  * Return <0 for error
468  */
469 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
470 			      bool locked)
471 {
472 	struct btrfs_block_group_cache *bg_cache;
473 	struct btrfs_full_stripe_locks_tree *locks_root;
474 	struct full_stripe_lock *fstripe_lock;
475 	u64 fstripe_start;
476 	bool freeit = false;
477 	int ret = 0;
478 
479 	/* If we didn't acquire full stripe lock, no need to continue */
480 	if (!locked)
481 		return 0;
482 
483 	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
484 	if (!bg_cache) {
485 		ASSERT(0);
486 		return -ENOENT;
487 	}
488 	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
489 		goto out;
490 
491 	locks_root = &bg_cache->full_stripe_locks_root;
492 	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
493 
494 	mutex_lock(&locks_root->lock);
495 	fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
496 	/* Unpaired unlock_full_stripe() detected */
497 	if (!fstripe_lock) {
498 		WARN_ON(1);
499 		ret = -ENOENT;
500 		mutex_unlock(&locks_root->lock);
501 		goto out;
502 	}
503 
504 	if (fstripe_lock->refs == 0) {
505 		WARN_ON(1);
506 		btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
507 			fstripe_lock->logical);
508 	} else {
509 		fstripe_lock->refs--;
510 	}
511 
512 	if (fstripe_lock->refs == 0) {
513 		rb_erase(&fstripe_lock->node, &locks_root->root);
514 		freeit = true;
515 	}
516 	mutex_unlock(&locks_root->lock);
517 
518 	mutex_unlock(&fstripe_lock->mutex);
519 	if (freeit)
520 		kfree(fstripe_lock);
521 out:
522 	btrfs_put_block_group(bg_cache);
523 	return ret;
524 }
525 
526 static void scrub_free_csums(struct scrub_ctx *sctx)
527 {
528 	while (!list_empty(&sctx->csum_list)) {
529 		struct btrfs_ordered_sum *sum;
530 		sum = list_first_entry(&sctx->csum_list,
531 				       struct btrfs_ordered_sum, list);
532 		list_del(&sum->list);
533 		kfree(sum);
534 	}
535 }
536 
537 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
538 {
539 	int i;
540 
541 	if (!sctx)
542 		return;
543 
544 	/* this can happen when scrub is cancelled */
545 	if (sctx->curr != -1) {
546 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
547 
548 		for (i = 0; i < sbio->page_count; i++) {
549 			WARN_ON(!sbio->pagev[i]->page);
550 			scrub_block_put(sbio->pagev[i]->sblock);
551 		}
552 		bio_put(sbio->bio);
553 	}
554 
555 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
556 		struct scrub_bio *sbio = sctx->bios[i];
557 
558 		if (!sbio)
559 			break;
560 		kfree(sbio);
561 	}
562 
563 	kfree(sctx->wr_curr_bio);
564 	scrub_free_csums(sctx);
565 	kfree(sctx);
566 }
567 
568 static void scrub_put_ctx(struct scrub_ctx *sctx)
569 {
570 	if (refcount_dec_and_test(&sctx->refs))
571 		scrub_free_ctx(sctx);
572 }
573 
574 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
575 		struct btrfs_fs_info *fs_info, int is_dev_replace)
576 {
577 	struct scrub_ctx *sctx;
578 	int		i;
579 
580 	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
581 	if (!sctx)
582 		goto nomem;
583 	refcount_set(&sctx->refs, 1);
584 	sctx->is_dev_replace = is_dev_replace;
585 	sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
586 	sctx->curr = -1;
587 	sctx->fs_info = fs_info;
588 	INIT_LIST_HEAD(&sctx->csum_list);
589 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
590 		struct scrub_bio *sbio;
591 
592 		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
593 		if (!sbio)
594 			goto nomem;
595 		sctx->bios[i] = sbio;
596 
597 		sbio->index = i;
598 		sbio->sctx = sctx;
599 		sbio->page_count = 0;
600 		btrfs_init_work(&sbio->work, btrfs_scrub_helper,
601 				scrub_bio_end_io_worker, NULL, NULL);
602 
603 		if (i != SCRUB_BIOS_PER_SCTX - 1)
604 			sctx->bios[i]->next_free = i + 1;
605 		else
606 			sctx->bios[i]->next_free = -1;
607 	}
608 	sctx->first_free = 0;
609 	atomic_set(&sctx->bios_in_flight, 0);
610 	atomic_set(&sctx->workers_pending, 0);
611 	atomic_set(&sctx->cancel_req, 0);
612 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
613 
614 	spin_lock_init(&sctx->list_lock);
615 	spin_lock_init(&sctx->stat_lock);
616 	init_waitqueue_head(&sctx->list_wait);
617 
618 	WARN_ON(sctx->wr_curr_bio != NULL);
619 	mutex_init(&sctx->wr_lock);
620 	sctx->wr_curr_bio = NULL;
621 	if (is_dev_replace) {
622 		WARN_ON(!fs_info->dev_replace.tgtdev);
623 		sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
624 		sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
625 		sctx->flush_all_writes = false;
626 	}
627 
628 	return sctx;
629 
630 nomem:
631 	scrub_free_ctx(sctx);
632 	return ERR_PTR(-ENOMEM);
633 }
634 
635 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
636 				     void *warn_ctx)
637 {
638 	u64 isize;
639 	u32 nlink;
640 	int ret;
641 	int i;
642 	unsigned nofs_flag;
643 	struct extent_buffer *eb;
644 	struct btrfs_inode_item *inode_item;
645 	struct scrub_warning *swarn = warn_ctx;
646 	struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
647 	struct inode_fs_paths *ipath = NULL;
648 	struct btrfs_root *local_root;
649 	struct btrfs_key root_key;
650 	struct btrfs_key key;
651 
652 	root_key.objectid = root;
653 	root_key.type = BTRFS_ROOT_ITEM_KEY;
654 	root_key.offset = (u64)-1;
655 	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
656 	if (IS_ERR(local_root)) {
657 		ret = PTR_ERR(local_root);
658 		goto err;
659 	}
660 
661 	/*
662 	 * this makes the path point to (inum INODE_ITEM ioff)
663 	 */
664 	key.objectid = inum;
665 	key.type = BTRFS_INODE_ITEM_KEY;
666 	key.offset = 0;
667 
668 	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
669 	if (ret) {
670 		btrfs_release_path(swarn->path);
671 		goto err;
672 	}
673 
674 	eb = swarn->path->nodes[0];
675 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
676 					struct btrfs_inode_item);
677 	isize = btrfs_inode_size(eb, inode_item);
678 	nlink = btrfs_inode_nlink(eb, inode_item);
679 	btrfs_release_path(swarn->path);
680 
681 	/*
682 	 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
683 	 * uses GFP_NOFS in this context, so we keep it consistent but it does
684 	 * not seem to be strictly necessary.
685 	 */
686 	nofs_flag = memalloc_nofs_save();
687 	ipath = init_ipath(4096, local_root, swarn->path);
688 	memalloc_nofs_restore(nofs_flag);
689 	if (IS_ERR(ipath)) {
690 		ret = PTR_ERR(ipath);
691 		ipath = NULL;
692 		goto err;
693 	}
694 	ret = paths_from_inode(inum, ipath);
695 
696 	if (ret < 0)
697 		goto err;
698 
699 	/*
700 	 * we deliberately ignore the bit ipath might have been too small to
701 	 * hold all of the paths here
702 	 */
703 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
704 		btrfs_warn_in_rcu(fs_info,
705 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
706 				  swarn->errstr, swarn->logical,
707 				  rcu_str_deref(swarn->dev->name),
708 				  swarn->physical,
709 				  root, inum, offset,
710 				  min(isize - offset, (u64)PAGE_SIZE), nlink,
711 				  (char *)(unsigned long)ipath->fspath->val[i]);
712 
713 	free_ipath(ipath);
714 	return 0;
715 
716 err:
717 	btrfs_warn_in_rcu(fs_info,
718 			  "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
719 			  swarn->errstr, swarn->logical,
720 			  rcu_str_deref(swarn->dev->name),
721 			  swarn->physical,
722 			  root, inum, offset, ret);
723 
724 	free_ipath(ipath);
725 	return 0;
726 }
727 
728 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
729 {
730 	struct btrfs_device *dev;
731 	struct btrfs_fs_info *fs_info;
732 	struct btrfs_path *path;
733 	struct btrfs_key found_key;
734 	struct extent_buffer *eb;
735 	struct btrfs_extent_item *ei;
736 	struct scrub_warning swarn;
737 	unsigned long ptr = 0;
738 	u64 extent_item_pos;
739 	u64 flags = 0;
740 	u64 ref_root;
741 	u32 item_size;
742 	u8 ref_level = 0;
743 	int ret;
744 
745 	WARN_ON(sblock->page_count < 1);
746 	dev = sblock->pagev[0]->dev;
747 	fs_info = sblock->sctx->fs_info;
748 
749 	path = btrfs_alloc_path();
750 	if (!path)
751 		return;
752 
753 	swarn.physical = sblock->pagev[0]->physical;
754 	swarn.logical = sblock->pagev[0]->logical;
755 	swarn.errstr = errstr;
756 	swarn.dev = NULL;
757 
758 	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
759 				  &flags);
760 	if (ret < 0)
761 		goto out;
762 
763 	extent_item_pos = swarn.logical - found_key.objectid;
764 	swarn.extent_item_size = found_key.offset;
765 
766 	eb = path->nodes[0];
767 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
768 	item_size = btrfs_item_size_nr(eb, path->slots[0]);
769 
770 	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
771 		do {
772 			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
773 						      item_size, &ref_root,
774 						      &ref_level);
775 			btrfs_warn_in_rcu(fs_info,
776 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
777 				errstr, swarn.logical,
778 				rcu_str_deref(dev->name),
779 				swarn.physical,
780 				ref_level ? "node" : "leaf",
781 				ret < 0 ? -1 : ref_level,
782 				ret < 0 ? -1 : ref_root);
783 		} while (ret != 1);
784 		btrfs_release_path(path);
785 	} else {
786 		btrfs_release_path(path);
787 		swarn.path = path;
788 		swarn.dev = dev;
789 		iterate_extent_inodes(fs_info, found_key.objectid,
790 					extent_item_pos, 1,
791 					scrub_print_warning_inode, &swarn, false);
792 	}
793 
794 out:
795 	btrfs_free_path(path);
796 }
797 
798 static inline void scrub_get_recover(struct scrub_recover *recover)
799 {
800 	refcount_inc(&recover->refs);
801 }
802 
803 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
804 				     struct scrub_recover *recover)
805 {
806 	if (refcount_dec_and_test(&recover->refs)) {
807 		btrfs_bio_counter_dec(fs_info);
808 		btrfs_put_bbio(recover->bbio);
809 		kfree(recover);
810 	}
811 }
812 
813 /*
814  * scrub_handle_errored_block gets called when either verification of the
815  * pages failed or the bio failed to read, e.g. with EIO. In the latter
816  * case, this function handles all pages in the bio, even though only one
817  * may be bad.
818  * The goal of this function is to repair the errored block by using the
819  * contents of one of the mirrors.
820  */
821 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
822 {
823 	struct scrub_ctx *sctx = sblock_to_check->sctx;
824 	struct btrfs_device *dev;
825 	struct btrfs_fs_info *fs_info;
826 	u64 logical;
827 	unsigned int failed_mirror_index;
828 	unsigned int is_metadata;
829 	unsigned int have_csum;
830 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
831 	struct scrub_block *sblock_bad;
832 	int ret;
833 	int mirror_index;
834 	int page_num;
835 	int success;
836 	bool full_stripe_locked;
837 	unsigned int nofs_flag;
838 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
839 				      DEFAULT_RATELIMIT_BURST);
840 
841 	BUG_ON(sblock_to_check->page_count < 1);
842 	fs_info = sctx->fs_info;
843 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
844 		/*
845 		 * if we find an error in a super block, we just report it.
846 		 * They will get written with the next transaction commit
847 		 * anyway
848 		 */
849 		spin_lock(&sctx->stat_lock);
850 		++sctx->stat.super_errors;
851 		spin_unlock(&sctx->stat_lock);
852 		return 0;
853 	}
854 	logical = sblock_to_check->pagev[0]->logical;
855 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
856 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
857 	is_metadata = !(sblock_to_check->pagev[0]->flags &
858 			BTRFS_EXTENT_FLAG_DATA);
859 	have_csum = sblock_to_check->pagev[0]->have_csum;
860 	dev = sblock_to_check->pagev[0]->dev;
861 
862 	/*
863 	 * We must use GFP_NOFS because the scrub task might be waiting for a
864 	 * worker task executing this function and in turn a transaction commit
865 	 * might be waiting the scrub task to pause (which needs to wait for all
866 	 * the worker tasks to complete before pausing).
867 	 * We do allocations in the workers through insert_full_stripe_lock()
868 	 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
869 	 * this function.
870 	 */
871 	nofs_flag = memalloc_nofs_save();
872 	/*
873 	 * For RAID5/6, race can happen for a different device scrub thread.
874 	 * For data corruption, Parity and Data threads will both try
875 	 * to recovery the data.
876 	 * Race can lead to doubly added csum error, or even unrecoverable
877 	 * error.
878 	 */
879 	ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
880 	if (ret < 0) {
881 		memalloc_nofs_restore(nofs_flag);
882 		spin_lock(&sctx->stat_lock);
883 		if (ret == -ENOMEM)
884 			sctx->stat.malloc_errors++;
885 		sctx->stat.read_errors++;
886 		sctx->stat.uncorrectable_errors++;
887 		spin_unlock(&sctx->stat_lock);
888 		return ret;
889 	}
890 
891 	/*
892 	 * read all mirrors one after the other. This includes to
893 	 * re-read the extent or metadata block that failed (that was
894 	 * the cause that this fixup code is called) another time,
895 	 * page by page this time in order to know which pages
896 	 * caused I/O errors and which ones are good (for all mirrors).
897 	 * It is the goal to handle the situation when more than one
898 	 * mirror contains I/O errors, but the errors do not
899 	 * overlap, i.e. the data can be repaired by selecting the
900 	 * pages from those mirrors without I/O error on the
901 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
902 	 * would be that mirror #1 has an I/O error on the first page,
903 	 * the second page is good, and mirror #2 has an I/O error on
904 	 * the second page, but the first page is good.
905 	 * Then the first page of the first mirror can be repaired by
906 	 * taking the first page of the second mirror, and the
907 	 * second page of the second mirror can be repaired by
908 	 * copying the contents of the 2nd page of the 1st mirror.
909 	 * One more note: if the pages of one mirror contain I/O
910 	 * errors, the checksum cannot be verified. In order to get
911 	 * the best data for repairing, the first attempt is to find
912 	 * a mirror without I/O errors and with a validated checksum.
913 	 * Only if this is not possible, the pages are picked from
914 	 * mirrors with I/O errors without considering the checksum.
915 	 * If the latter is the case, at the end, the checksum of the
916 	 * repaired area is verified in order to correctly maintain
917 	 * the statistics.
918 	 */
919 
920 	sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
921 				      sizeof(*sblocks_for_recheck), GFP_KERNEL);
922 	if (!sblocks_for_recheck) {
923 		spin_lock(&sctx->stat_lock);
924 		sctx->stat.malloc_errors++;
925 		sctx->stat.read_errors++;
926 		sctx->stat.uncorrectable_errors++;
927 		spin_unlock(&sctx->stat_lock);
928 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
929 		goto out;
930 	}
931 
932 	/* setup the context, map the logical blocks and alloc the pages */
933 	ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
934 	if (ret) {
935 		spin_lock(&sctx->stat_lock);
936 		sctx->stat.read_errors++;
937 		sctx->stat.uncorrectable_errors++;
938 		spin_unlock(&sctx->stat_lock);
939 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
940 		goto out;
941 	}
942 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
943 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
944 
945 	/* build and submit the bios for the failed mirror, check checksums */
946 	scrub_recheck_block(fs_info, sblock_bad, 1);
947 
948 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
949 	    sblock_bad->no_io_error_seen) {
950 		/*
951 		 * the error disappeared after reading page by page, or
952 		 * the area was part of a huge bio and other parts of the
953 		 * bio caused I/O errors, or the block layer merged several
954 		 * read requests into one and the error is caused by a
955 		 * different bio (usually one of the two latter cases is
956 		 * the cause)
957 		 */
958 		spin_lock(&sctx->stat_lock);
959 		sctx->stat.unverified_errors++;
960 		sblock_to_check->data_corrected = 1;
961 		spin_unlock(&sctx->stat_lock);
962 
963 		if (sctx->is_dev_replace)
964 			scrub_write_block_to_dev_replace(sblock_bad);
965 		goto out;
966 	}
967 
968 	if (!sblock_bad->no_io_error_seen) {
969 		spin_lock(&sctx->stat_lock);
970 		sctx->stat.read_errors++;
971 		spin_unlock(&sctx->stat_lock);
972 		if (__ratelimit(&_rs))
973 			scrub_print_warning("i/o error", sblock_to_check);
974 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
975 	} else if (sblock_bad->checksum_error) {
976 		spin_lock(&sctx->stat_lock);
977 		sctx->stat.csum_errors++;
978 		spin_unlock(&sctx->stat_lock);
979 		if (__ratelimit(&_rs))
980 			scrub_print_warning("checksum error", sblock_to_check);
981 		btrfs_dev_stat_inc_and_print(dev,
982 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
983 	} else if (sblock_bad->header_error) {
984 		spin_lock(&sctx->stat_lock);
985 		sctx->stat.verify_errors++;
986 		spin_unlock(&sctx->stat_lock);
987 		if (__ratelimit(&_rs))
988 			scrub_print_warning("checksum/header error",
989 					    sblock_to_check);
990 		if (sblock_bad->generation_error)
991 			btrfs_dev_stat_inc_and_print(dev,
992 				BTRFS_DEV_STAT_GENERATION_ERRS);
993 		else
994 			btrfs_dev_stat_inc_and_print(dev,
995 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
996 	}
997 
998 	if (sctx->readonly) {
999 		ASSERT(!sctx->is_dev_replace);
1000 		goto out;
1001 	}
1002 
1003 	/*
1004 	 * now build and submit the bios for the other mirrors, check
1005 	 * checksums.
1006 	 * First try to pick the mirror which is completely without I/O
1007 	 * errors and also does not have a checksum error.
1008 	 * If one is found, and if a checksum is present, the full block
1009 	 * that is known to contain an error is rewritten. Afterwards
1010 	 * the block is known to be corrected.
1011 	 * If a mirror is found which is completely correct, and no
1012 	 * checksum is present, only those pages are rewritten that had
1013 	 * an I/O error in the block to be repaired, since it cannot be
1014 	 * determined, which copy of the other pages is better (and it
1015 	 * could happen otherwise that a correct page would be
1016 	 * overwritten by a bad one).
1017 	 */
1018 	for (mirror_index = 0; ;mirror_index++) {
1019 		struct scrub_block *sblock_other;
1020 
1021 		if (mirror_index == failed_mirror_index)
1022 			continue;
1023 
1024 		/* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1025 		if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1026 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1027 				break;
1028 			if (!sblocks_for_recheck[mirror_index].page_count)
1029 				break;
1030 
1031 			sblock_other = sblocks_for_recheck + mirror_index;
1032 		} else {
1033 			struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1034 			int max_allowed = r->bbio->num_stripes -
1035 						r->bbio->num_tgtdevs;
1036 
1037 			if (mirror_index >= max_allowed)
1038 				break;
1039 			if (!sblocks_for_recheck[1].page_count)
1040 				break;
1041 
1042 			ASSERT(failed_mirror_index == 0);
1043 			sblock_other = sblocks_for_recheck + 1;
1044 			sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1045 		}
1046 
1047 		/* build and submit the bios, check checksums */
1048 		scrub_recheck_block(fs_info, sblock_other, 0);
1049 
1050 		if (!sblock_other->header_error &&
1051 		    !sblock_other->checksum_error &&
1052 		    sblock_other->no_io_error_seen) {
1053 			if (sctx->is_dev_replace) {
1054 				scrub_write_block_to_dev_replace(sblock_other);
1055 				goto corrected_error;
1056 			} else {
1057 				ret = scrub_repair_block_from_good_copy(
1058 						sblock_bad, sblock_other);
1059 				if (!ret)
1060 					goto corrected_error;
1061 			}
1062 		}
1063 	}
1064 
1065 	if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1066 		goto did_not_correct_error;
1067 
1068 	/*
1069 	 * In case of I/O errors in the area that is supposed to be
1070 	 * repaired, continue by picking good copies of those pages.
1071 	 * Select the good pages from mirrors to rewrite bad pages from
1072 	 * the area to fix. Afterwards verify the checksum of the block
1073 	 * that is supposed to be repaired. This verification step is
1074 	 * only done for the purpose of statistic counting and for the
1075 	 * final scrub report, whether errors remain.
1076 	 * A perfect algorithm could make use of the checksum and try
1077 	 * all possible combinations of pages from the different mirrors
1078 	 * until the checksum verification succeeds. For example, when
1079 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1080 	 * of mirror #2 is readable but the final checksum test fails,
1081 	 * then the 2nd page of mirror #3 could be tried, whether now
1082 	 * the final checksum succeeds. But this would be a rare
1083 	 * exception and is therefore not implemented. At least it is
1084 	 * avoided that the good copy is overwritten.
1085 	 * A more useful improvement would be to pick the sectors
1086 	 * without I/O error based on sector sizes (512 bytes on legacy
1087 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1088 	 * mirror could be repaired by taking 512 byte of a different
1089 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1090 	 * area are unreadable.
1091 	 */
1092 	success = 1;
1093 	for (page_num = 0; page_num < sblock_bad->page_count;
1094 	     page_num++) {
1095 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1096 		struct scrub_block *sblock_other = NULL;
1097 
1098 		/* skip no-io-error page in scrub */
1099 		if (!page_bad->io_error && !sctx->is_dev_replace)
1100 			continue;
1101 
1102 		if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1103 			/*
1104 			 * In case of dev replace, if raid56 rebuild process
1105 			 * didn't work out correct data, then copy the content
1106 			 * in sblock_bad to make sure target device is identical
1107 			 * to source device, instead of writing garbage data in
1108 			 * sblock_for_recheck array to target device.
1109 			 */
1110 			sblock_other = NULL;
1111 		} else if (page_bad->io_error) {
1112 			/* try to find no-io-error page in mirrors */
1113 			for (mirror_index = 0;
1114 			     mirror_index < BTRFS_MAX_MIRRORS &&
1115 			     sblocks_for_recheck[mirror_index].page_count > 0;
1116 			     mirror_index++) {
1117 				if (!sblocks_for_recheck[mirror_index].
1118 				    pagev[page_num]->io_error) {
1119 					sblock_other = sblocks_for_recheck +
1120 						       mirror_index;
1121 					break;
1122 				}
1123 			}
1124 			if (!sblock_other)
1125 				success = 0;
1126 		}
1127 
1128 		if (sctx->is_dev_replace) {
1129 			/*
1130 			 * did not find a mirror to fetch the page
1131 			 * from. scrub_write_page_to_dev_replace()
1132 			 * handles this case (page->io_error), by
1133 			 * filling the block with zeros before
1134 			 * submitting the write request
1135 			 */
1136 			if (!sblock_other)
1137 				sblock_other = sblock_bad;
1138 
1139 			if (scrub_write_page_to_dev_replace(sblock_other,
1140 							    page_num) != 0) {
1141 				atomic64_inc(
1142 					&fs_info->dev_replace.num_write_errors);
1143 				success = 0;
1144 			}
1145 		} else if (sblock_other) {
1146 			ret = scrub_repair_page_from_good_copy(sblock_bad,
1147 							       sblock_other,
1148 							       page_num, 0);
1149 			if (0 == ret)
1150 				page_bad->io_error = 0;
1151 			else
1152 				success = 0;
1153 		}
1154 	}
1155 
1156 	if (success && !sctx->is_dev_replace) {
1157 		if (is_metadata || have_csum) {
1158 			/*
1159 			 * need to verify the checksum now that all
1160 			 * sectors on disk are repaired (the write
1161 			 * request for data to be repaired is on its way).
1162 			 * Just be lazy and use scrub_recheck_block()
1163 			 * which re-reads the data before the checksum
1164 			 * is verified, but most likely the data comes out
1165 			 * of the page cache.
1166 			 */
1167 			scrub_recheck_block(fs_info, sblock_bad, 1);
1168 			if (!sblock_bad->header_error &&
1169 			    !sblock_bad->checksum_error &&
1170 			    sblock_bad->no_io_error_seen)
1171 				goto corrected_error;
1172 			else
1173 				goto did_not_correct_error;
1174 		} else {
1175 corrected_error:
1176 			spin_lock(&sctx->stat_lock);
1177 			sctx->stat.corrected_errors++;
1178 			sblock_to_check->data_corrected = 1;
1179 			spin_unlock(&sctx->stat_lock);
1180 			btrfs_err_rl_in_rcu(fs_info,
1181 				"fixed up error at logical %llu on dev %s",
1182 				logical, rcu_str_deref(dev->name));
1183 		}
1184 	} else {
1185 did_not_correct_error:
1186 		spin_lock(&sctx->stat_lock);
1187 		sctx->stat.uncorrectable_errors++;
1188 		spin_unlock(&sctx->stat_lock);
1189 		btrfs_err_rl_in_rcu(fs_info,
1190 			"unable to fixup (regular) error at logical %llu on dev %s",
1191 			logical, rcu_str_deref(dev->name));
1192 	}
1193 
1194 out:
1195 	if (sblocks_for_recheck) {
1196 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1197 		     mirror_index++) {
1198 			struct scrub_block *sblock = sblocks_for_recheck +
1199 						     mirror_index;
1200 			struct scrub_recover *recover;
1201 			int page_index;
1202 
1203 			for (page_index = 0; page_index < sblock->page_count;
1204 			     page_index++) {
1205 				sblock->pagev[page_index]->sblock = NULL;
1206 				recover = sblock->pagev[page_index]->recover;
1207 				if (recover) {
1208 					scrub_put_recover(fs_info, recover);
1209 					sblock->pagev[page_index]->recover =
1210 									NULL;
1211 				}
1212 				scrub_page_put(sblock->pagev[page_index]);
1213 			}
1214 		}
1215 		kfree(sblocks_for_recheck);
1216 	}
1217 
1218 	ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1219 	memalloc_nofs_restore(nofs_flag);
1220 	if (ret < 0)
1221 		return ret;
1222 	return 0;
1223 }
1224 
1225 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1226 {
1227 	if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1228 		return 2;
1229 	else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1230 		return 3;
1231 	else
1232 		return (int)bbio->num_stripes;
1233 }
1234 
1235 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1236 						 u64 *raid_map,
1237 						 u64 mapped_length,
1238 						 int nstripes, int mirror,
1239 						 int *stripe_index,
1240 						 u64 *stripe_offset)
1241 {
1242 	int i;
1243 
1244 	if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1245 		/* RAID5/6 */
1246 		for (i = 0; i < nstripes; i++) {
1247 			if (raid_map[i] == RAID6_Q_STRIPE ||
1248 			    raid_map[i] == RAID5_P_STRIPE)
1249 				continue;
1250 
1251 			if (logical >= raid_map[i] &&
1252 			    logical < raid_map[i] + mapped_length)
1253 				break;
1254 		}
1255 
1256 		*stripe_index = i;
1257 		*stripe_offset = logical - raid_map[i];
1258 	} else {
1259 		/* The other RAID type */
1260 		*stripe_index = mirror;
1261 		*stripe_offset = 0;
1262 	}
1263 }
1264 
1265 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1266 				     struct scrub_block *sblocks_for_recheck)
1267 {
1268 	struct scrub_ctx *sctx = original_sblock->sctx;
1269 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1270 	u64 length = original_sblock->page_count * PAGE_SIZE;
1271 	u64 logical = original_sblock->pagev[0]->logical;
1272 	u64 generation = original_sblock->pagev[0]->generation;
1273 	u64 flags = original_sblock->pagev[0]->flags;
1274 	u64 have_csum = original_sblock->pagev[0]->have_csum;
1275 	struct scrub_recover *recover;
1276 	struct btrfs_bio *bbio;
1277 	u64 sublen;
1278 	u64 mapped_length;
1279 	u64 stripe_offset;
1280 	int stripe_index;
1281 	int page_index = 0;
1282 	int mirror_index;
1283 	int nmirrors;
1284 	int ret;
1285 
1286 	/*
1287 	 * note: the two members refs and outstanding_pages
1288 	 * are not used (and not set) in the blocks that are used for
1289 	 * the recheck procedure
1290 	 */
1291 
1292 	while (length > 0) {
1293 		sublen = min_t(u64, length, PAGE_SIZE);
1294 		mapped_length = sublen;
1295 		bbio = NULL;
1296 
1297 		/*
1298 		 * with a length of PAGE_SIZE, each returned stripe
1299 		 * represents one mirror
1300 		 */
1301 		btrfs_bio_counter_inc_blocked(fs_info);
1302 		ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1303 				logical, &mapped_length, &bbio);
1304 		if (ret || !bbio || mapped_length < sublen) {
1305 			btrfs_put_bbio(bbio);
1306 			btrfs_bio_counter_dec(fs_info);
1307 			return -EIO;
1308 		}
1309 
1310 		recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1311 		if (!recover) {
1312 			btrfs_put_bbio(bbio);
1313 			btrfs_bio_counter_dec(fs_info);
1314 			return -ENOMEM;
1315 		}
1316 
1317 		refcount_set(&recover->refs, 1);
1318 		recover->bbio = bbio;
1319 		recover->map_length = mapped_length;
1320 
1321 		BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1322 
1323 		nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1324 
1325 		for (mirror_index = 0; mirror_index < nmirrors;
1326 		     mirror_index++) {
1327 			struct scrub_block *sblock;
1328 			struct scrub_page *page;
1329 
1330 			sblock = sblocks_for_recheck + mirror_index;
1331 			sblock->sctx = sctx;
1332 
1333 			page = kzalloc(sizeof(*page), GFP_NOFS);
1334 			if (!page) {
1335 leave_nomem:
1336 				spin_lock(&sctx->stat_lock);
1337 				sctx->stat.malloc_errors++;
1338 				spin_unlock(&sctx->stat_lock);
1339 				scrub_put_recover(fs_info, recover);
1340 				return -ENOMEM;
1341 			}
1342 			scrub_page_get(page);
1343 			sblock->pagev[page_index] = page;
1344 			page->sblock = sblock;
1345 			page->flags = flags;
1346 			page->generation = generation;
1347 			page->logical = logical;
1348 			page->have_csum = have_csum;
1349 			if (have_csum)
1350 				memcpy(page->csum,
1351 				       original_sblock->pagev[0]->csum,
1352 				       sctx->csum_size);
1353 
1354 			scrub_stripe_index_and_offset(logical,
1355 						      bbio->map_type,
1356 						      bbio->raid_map,
1357 						      mapped_length,
1358 						      bbio->num_stripes -
1359 						      bbio->num_tgtdevs,
1360 						      mirror_index,
1361 						      &stripe_index,
1362 						      &stripe_offset);
1363 			page->physical = bbio->stripes[stripe_index].physical +
1364 					 stripe_offset;
1365 			page->dev = bbio->stripes[stripe_index].dev;
1366 
1367 			BUG_ON(page_index >= original_sblock->page_count);
1368 			page->physical_for_dev_replace =
1369 				original_sblock->pagev[page_index]->
1370 				physical_for_dev_replace;
1371 			/* for missing devices, dev->bdev is NULL */
1372 			page->mirror_num = mirror_index + 1;
1373 			sblock->page_count++;
1374 			page->page = alloc_page(GFP_NOFS);
1375 			if (!page->page)
1376 				goto leave_nomem;
1377 
1378 			scrub_get_recover(recover);
1379 			page->recover = recover;
1380 		}
1381 		scrub_put_recover(fs_info, recover);
1382 		length -= sublen;
1383 		logical += sublen;
1384 		page_index++;
1385 	}
1386 
1387 	return 0;
1388 }
1389 
1390 static void scrub_bio_wait_endio(struct bio *bio)
1391 {
1392 	complete(bio->bi_private);
1393 }
1394 
1395 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1396 					struct bio *bio,
1397 					struct scrub_page *page)
1398 {
1399 	DECLARE_COMPLETION_ONSTACK(done);
1400 	int ret;
1401 	int mirror_num;
1402 
1403 	bio->bi_iter.bi_sector = page->logical >> 9;
1404 	bio->bi_private = &done;
1405 	bio->bi_end_io = scrub_bio_wait_endio;
1406 
1407 	mirror_num = page->sblock->pagev[0]->mirror_num;
1408 	ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1409 				    page->recover->map_length,
1410 				    mirror_num, 0);
1411 	if (ret)
1412 		return ret;
1413 
1414 	wait_for_completion_io(&done);
1415 	return blk_status_to_errno(bio->bi_status);
1416 }
1417 
1418 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1419 					  struct scrub_block *sblock)
1420 {
1421 	struct scrub_page *first_page = sblock->pagev[0];
1422 	struct bio *bio;
1423 	int page_num;
1424 
1425 	/* All pages in sblock belong to the same stripe on the same device. */
1426 	ASSERT(first_page->dev);
1427 	if (!first_page->dev->bdev)
1428 		goto out;
1429 
1430 	bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1431 	bio_set_dev(bio, first_page->dev->bdev);
1432 
1433 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1434 		struct scrub_page *page = sblock->pagev[page_num];
1435 
1436 		WARN_ON(!page->page);
1437 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1438 	}
1439 
1440 	if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1441 		bio_put(bio);
1442 		goto out;
1443 	}
1444 
1445 	bio_put(bio);
1446 
1447 	scrub_recheck_block_checksum(sblock);
1448 
1449 	return;
1450 out:
1451 	for (page_num = 0; page_num < sblock->page_count; page_num++)
1452 		sblock->pagev[page_num]->io_error = 1;
1453 
1454 	sblock->no_io_error_seen = 0;
1455 }
1456 
1457 /*
1458  * this function will check the on disk data for checksum errors, header
1459  * errors and read I/O errors. If any I/O errors happen, the exact pages
1460  * which are errored are marked as being bad. The goal is to enable scrub
1461  * to take those pages that are not errored from all the mirrors so that
1462  * the pages that are errored in the just handled mirror can be repaired.
1463  */
1464 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1465 				struct scrub_block *sblock,
1466 				int retry_failed_mirror)
1467 {
1468 	int page_num;
1469 
1470 	sblock->no_io_error_seen = 1;
1471 
1472 	/* short cut for raid56 */
1473 	if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1474 		return scrub_recheck_block_on_raid56(fs_info, sblock);
1475 
1476 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1477 		struct bio *bio;
1478 		struct scrub_page *page = sblock->pagev[page_num];
1479 
1480 		if (page->dev->bdev == NULL) {
1481 			page->io_error = 1;
1482 			sblock->no_io_error_seen = 0;
1483 			continue;
1484 		}
1485 
1486 		WARN_ON(!page->page);
1487 		bio = btrfs_io_bio_alloc(1);
1488 		bio_set_dev(bio, page->dev->bdev);
1489 
1490 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1491 		bio->bi_iter.bi_sector = page->physical >> 9;
1492 		bio->bi_opf = REQ_OP_READ;
1493 
1494 		if (btrfsic_submit_bio_wait(bio)) {
1495 			page->io_error = 1;
1496 			sblock->no_io_error_seen = 0;
1497 		}
1498 
1499 		bio_put(bio);
1500 	}
1501 
1502 	if (sblock->no_io_error_seen)
1503 		scrub_recheck_block_checksum(sblock);
1504 }
1505 
1506 static inline int scrub_check_fsid(u8 fsid[],
1507 				   struct scrub_page *spage)
1508 {
1509 	struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1510 	int ret;
1511 
1512 	ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1513 	return !ret;
1514 }
1515 
1516 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1517 {
1518 	sblock->header_error = 0;
1519 	sblock->checksum_error = 0;
1520 	sblock->generation_error = 0;
1521 
1522 	if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1523 		scrub_checksum_data(sblock);
1524 	else
1525 		scrub_checksum_tree_block(sblock);
1526 }
1527 
1528 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1529 					     struct scrub_block *sblock_good)
1530 {
1531 	int page_num;
1532 	int ret = 0;
1533 
1534 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1535 		int ret_sub;
1536 
1537 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1538 							   sblock_good,
1539 							   page_num, 1);
1540 		if (ret_sub)
1541 			ret = ret_sub;
1542 	}
1543 
1544 	return ret;
1545 }
1546 
1547 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1548 					    struct scrub_block *sblock_good,
1549 					    int page_num, int force_write)
1550 {
1551 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1552 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1553 	struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1554 
1555 	BUG_ON(page_bad->page == NULL);
1556 	BUG_ON(page_good->page == NULL);
1557 	if (force_write || sblock_bad->header_error ||
1558 	    sblock_bad->checksum_error || page_bad->io_error) {
1559 		struct bio *bio;
1560 		int ret;
1561 
1562 		if (!page_bad->dev->bdev) {
1563 			btrfs_warn_rl(fs_info,
1564 				"scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1565 			return -EIO;
1566 		}
1567 
1568 		bio = btrfs_io_bio_alloc(1);
1569 		bio_set_dev(bio, page_bad->dev->bdev);
1570 		bio->bi_iter.bi_sector = page_bad->physical >> 9;
1571 		bio->bi_opf = REQ_OP_WRITE;
1572 
1573 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1574 		if (PAGE_SIZE != ret) {
1575 			bio_put(bio);
1576 			return -EIO;
1577 		}
1578 
1579 		if (btrfsic_submit_bio_wait(bio)) {
1580 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1581 				BTRFS_DEV_STAT_WRITE_ERRS);
1582 			atomic64_inc(&fs_info->dev_replace.num_write_errors);
1583 			bio_put(bio);
1584 			return -EIO;
1585 		}
1586 		bio_put(bio);
1587 	}
1588 
1589 	return 0;
1590 }
1591 
1592 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1593 {
1594 	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1595 	int page_num;
1596 
1597 	/*
1598 	 * This block is used for the check of the parity on the source device,
1599 	 * so the data needn't be written into the destination device.
1600 	 */
1601 	if (sblock->sparity)
1602 		return;
1603 
1604 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1605 		int ret;
1606 
1607 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1608 		if (ret)
1609 			atomic64_inc(&fs_info->dev_replace.num_write_errors);
1610 	}
1611 }
1612 
1613 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1614 					   int page_num)
1615 {
1616 	struct scrub_page *spage = sblock->pagev[page_num];
1617 
1618 	BUG_ON(spage->page == NULL);
1619 	if (spage->io_error) {
1620 		void *mapped_buffer = kmap_atomic(spage->page);
1621 
1622 		clear_page(mapped_buffer);
1623 		flush_dcache_page(spage->page);
1624 		kunmap_atomic(mapped_buffer);
1625 	}
1626 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1627 }
1628 
1629 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1630 				    struct scrub_page *spage)
1631 {
1632 	struct scrub_bio *sbio;
1633 	int ret;
1634 
1635 	mutex_lock(&sctx->wr_lock);
1636 again:
1637 	if (!sctx->wr_curr_bio) {
1638 		sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1639 					      GFP_KERNEL);
1640 		if (!sctx->wr_curr_bio) {
1641 			mutex_unlock(&sctx->wr_lock);
1642 			return -ENOMEM;
1643 		}
1644 		sctx->wr_curr_bio->sctx = sctx;
1645 		sctx->wr_curr_bio->page_count = 0;
1646 	}
1647 	sbio = sctx->wr_curr_bio;
1648 	if (sbio->page_count == 0) {
1649 		struct bio *bio;
1650 
1651 		sbio->physical = spage->physical_for_dev_replace;
1652 		sbio->logical = spage->logical;
1653 		sbio->dev = sctx->wr_tgtdev;
1654 		bio = sbio->bio;
1655 		if (!bio) {
1656 			bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1657 			sbio->bio = bio;
1658 		}
1659 
1660 		bio->bi_private = sbio;
1661 		bio->bi_end_io = scrub_wr_bio_end_io;
1662 		bio_set_dev(bio, sbio->dev->bdev);
1663 		bio->bi_iter.bi_sector = sbio->physical >> 9;
1664 		bio->bi_opf = REQ_OP_WRITE;
1665 		sbio->status = 0;
1666 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1667 		   spage->physical_for_dev_replace ||
1668 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1669 		   spage->logical) {
1670 		scrub_wr_submit(sctx);
1671 		goto again;
1672 	}
1673 
1674 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1675 	if (ret != PAGE_SIZE) {
1676 		if (sbio->page_count < 1) {
1677 			bio_put(sbio->bio);
1678 			sbio->bio = NULL;
1679 			mutex_unlock(&sctx->wr_lock);
1680 			return -EIO;
1681 		}
1682 		scrub_wr_submit(sctx);
1683 		goto again;
1684 	}
1685 
1686 	sbio->pagev[sbio->page_count] = spage;
1687 	scrub_page_get(spage);
1688 	sbio->page_count++;
1689 	if (sbio->page_count == sctx->pages_per_wr_bio)
1690 		scrub_wr_submit(sctx);
1691 	mutex_unlock(&sctx->wr_lock);
1692 
1693 	return 0;
1694 }
1695 
1696 static void scrub_wr_submit(struct scrub_ctx *sctx)
1697 {
1698 	struct scrub_bio *sbio;
1699 
1700 	if (!sctx->wr_curr_bio)
1701 		return;
1702 
1703 	sbio = sctx->wr_curr_bio;
1704 	sctx->wr_curr_bio = NULL;
1705 	WARN_ON(!sbio->bio->bi_disk);
1706 	scrub_pending_bio_inc(sctx);
1707 	/* process all writes in a single worker thread. Then the block layer
1708 	 * orders the requests before sending them to the driver which
1709 	 * doubled the write performance on spinning disks when measured
1710 	 * with Linux 3.5 */
1711 	btrfsic_submit_bio(sbio->bio);
1712 }
1713 
1714 static void scrub_wr_bio_end_io(struct bio *bio)
1715 {
1716 	struct scrub_bio *sbio = bio->bi_private;
1717 	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1718 
1719 	sbio->status = bio->bi_status;
1720 	sbio->bio = bio;
1721 
1722 	btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1723 			 scrub_wr_bio_end_io_worker, NULL, NULL);
1724 	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1725 }
1726 
1727 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1728 {
1729 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1730 	struct scrub_ctx *sctx = sbio->sctx;
1731 	int i;
1732 
1733 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1734 	if (sbio->status) {
1735 		struct btrfs_dev_replace *dev_replace =
1736 			&sbio->sctx->fs_info->dev_replace;
1737 
1738 		for (i = 0; i < sbio->page_count; i++) {
1739 			struct scrub_page *spage = sbio->pagev[i];
1740 
1741 			spage->io_error = 1;
1742 			atomic64_inc(&dev_replace->num_write_errors);
1743 		}
1744 	}
1745 
1746 	for (i = 0; i < sbio->page_count; i++)
1747 		scrub_page_put(sbio->pagev[i]);
1748 
1749 	bio_put(sbio->bio);
1750 	kfree(sbio);
1751 	scrub_pending_bio_dec(sctx);
1752 }
1753 
1754 static int scrub_checksum(struct scrub_block *sblock)
1755 {
1756 	u64 flags;
1757 	int ret;
1758 
1759 	/*
1760 	 * No need to initialize these stats currently,
1761 	 * because this function only use return value
1762 	 * instead of these stats value.
1763 	 *
1764 	 * Todo:
1765 	 * always use stats
1766 	 */
1767 	sblock->header_error = 0;
1768 	sblock->generation_error = 0;
1769 	sblock->checksum_error = 0;
1770 
1771 	WARN_ON(sblock->page_count < 1);
1772 	flags = sblock->pagev[0]->flags;
1773 	ret = 0;
1774 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1775 		ret = scrub_checksum_data(sblock);
1776 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1777 		ret = scrub_checksum_tree_block(sblock);
1778 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1779 		(void)scrub_checksum_super(sblock);
1780 	else
1781 		WARN_ON(1);
1782 	if (ret)
1783 		scrub_handle_errored_block(sblock);
1784 
1785 	return ret;
1786 }
1787 
1788 static int scrub_checksum_data(struct scrub_block *sblock)
1789 {
1790 	struct scrub_ctx *sctx = sblock->sctx;
1791 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1792 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1793 	u8 csum[BTRFS_CSUM_SIZE];
1794 	u8 *on_disk_csum;
1795 	struct page *page;
1796 	void *buffer;
1797 	u64 len;
1798 	int index;
1799 
1800 	BUG_ON(sblock->page_count < 1);
1801 	if (!sblock->pagev[0]->have_csum)
1802 		return 0;
1803 
1804 	shash->tfm = fs_info->csum_shash;
1805 	crypto_shash_init(shash);
1806 
1807 	on_disk_csum = sblock->pagev[0]->csum;
1808 	page = sblock->pagev[0]->page;
1809 	buffer = kmap_atomic(page);
1810 
1811 	len = sctx->fs_info->sectorsize;
1812 	index = 0;
1813 	for (;;) {
1814 		u64 l = min_t(u64, len, PAGE_SIZE);
1815 
1816 		crypto_shash_update(shash, buffer, l);
1817 		kunmap_atomic(buffer);
1818 		len -= l;
1819 		if (len == 0)
1820 			break;
1821 		index++;
1822 		BUG_ON(index >= sblock->page_count);
1823 		BUG_ON(!sblock->pagev[index]->page);
1824 		page = sblock->pagev[index]->page;
1825 		buffer = kmap_atomic(page);
1826 	}
1827 
1828 	crypto_shash_final(shash, csum);
1829 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1830 		sblock->checksum_error = 1;
1831 
1832 	return sblock->checksum_error;
1833 }
1834 
1835 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1836 {
1837 	struct scrub_ctx *sctx = sblock->sctx;
1838 	struct btrfs_header *h;
1839 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1840 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1841 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1842 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1843 	struct page *page;
1844 	void *mapped_buffer;
1845 	u64 mapped_size;
1846 	void *p;
1847 	u64 len;
1848 	int index;
1849 
1850 	shash->tfm = fs_info->csum_shash;
1851 	crypto_shash_init(shash);
1852 
1853 	BUG_ON(sblock->page_count < 1);
1854 	page = sblock->pagev[0]->page;
1855 	mapped_buffer = kmap_atomic(page);
1856 	h = (struct btrfs_header *)mapped_buffer;
1857 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1858 
1859 	/*
1860 	 * we don't use the getter functions here, as we
1861 	 * a) don't have an extent buffer and
1862 	 * b) the page is already kmapped
1863 	 */
1864 	if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1865 		sblock->header_error = 1;
1866 
1867 	if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1868 		sblock->header_error = 1;
1869 		sblock->generation_error = 1;
1870 	}
1871 
1872 	if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1873 		sblock->header_error = 1;
1874 
1875 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1876 		   BTRFS_UUID_SIZE))
1877 		sblock->header_error = 1;
1878 
1879 	len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1880 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1881 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1882 	index = 0;
1883 	for (;;) {
1884 		u64 l = min_t(u64, len, mapped_size);
1885 
1886 		crypto_shash_update(shash, p, l);
1887 		kunmap_atomic(mapped_buffer);
1888 		len -= l;
1889 		if (len == 0)
1890 			break;
1891 		index++;
1892 		BUG_ON(index >= sblock->page_count);
1893 		BUG_ON(!sblock->pagev[index]->page);
1894 		page = sblock->pagev[index]->page;
1895 		mapped_buffer = kmap_atomic(page);
1896 		mapped_size = PAGE_SIZE;
1897 		p = mapped_buffer;
1898 	}
1899 
1900 	crypto_shash_final(shash, calculated_csum);
1901 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1902 		sblock->checksum_error = 1;
1903 
1904 	return sblock->header_error || sblock->checksum_error;
1905 }
1906 
1907 static int scrub_checksum_super(struct scrub_block *sblock)
1908 {
1909 	struct btrfs_super_block *s;
1910 	struct scrub_ctx *sctx = sblock->sctx;
1911 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1912 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1913 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1914 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1915 	struct page *page;
1916 	void *mapped_buffer;
1917 	u64 mapped_size;
1918 	void *p;
1919 	int fail_gen = 0;
1920 	int fail_cor = 0;
1921 	u64 len;
1922 	int index;
1923 
1924 	shash->tfm = fs_info->csum_shash;
1925 	crypto_shash_init(shash);
1926 
1927 	BUG_ON(sblock->page_count < 1);
1928 	page = sblock->pagev[0]->page;
1929 	mapped_buffer = kmap_atomic(page);
1930 	s = (struct btrfs_super_block *)mapped_buffer;
1931 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1932 
1933 	if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1934 		++fail_cor;
1935 
1936 	if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1937 		++fail_gen;
1938 
1939 	if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1940 		++fail_cor;
1941 
1942 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1943 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1944 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1945 	index = 0;
1946 	for (;;) {
1947 		u64 l = min_t(u64, len, mapped_size);
1948 
1949 		crypto_shash_update(shash, p, l);
1950 		kunmap_atomic(mapped_buffer);
1951 		len -= l;
1952 		if (len == 0)
1953 			break;
1954 		index++;
1955 		BUG_ON(index >= sblock->page_count);
1956 		BUG_ON(!sblock->pagev[index]->page);
1957 		page = sblock->pagev[index]->page;
1958 		mapped_buffer = kmap_atomic(page);
1959 		mapped_size = PAGE_SIZE;
1960 		p = mapped_buffer;
1961 	}
1962 
1963 	crypto_shash_final(shash, calculated_csum);
1964 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1965 		++fail_cor;
1966 
1967 	if (fail_cor + fail_gen) {
1968 		/*
1969 		 * if we find an error in a super block, we just report it.
1970 		 * They will get written with the next transaction commit
1971 		 * anyway
1972 		 */
1973 		spin_lock(&sctx->stat_lock);
1974 		++sctx->stat.super_errors;
1975 		spin_unlock(&sctx->stat_lock);
1976 		if (fail_cor)
1977 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1978 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1979 		else
1980 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1981 				BTRFS_DEV_STAT_GENERATION_ERRS);
1982 	}
1983 
1984 	return fail_cor + fail_gen;
1985 }
1986 
1987 static void scrub_block_get(struct scrub_block *sblock)
1988 {
1989 	refcount_inc(&sblock->refs);
1990 }
1991 
1992 static void scrub_block_put(struct scrub_block *sblock)
1993 {
1994 	if (refcount_dec_and_test(&sblock->refs)) {
1995 		int i;
1996 
1997 		if (sblock->sparity)
1998 			scrub_parity_put(sblock->sparity);
1999 
2000 		for (i = 0; i < sblock->page_count; i++)
2001 			scrub_page_put(sblock->pagev[i]);
2002 		kfree(sblock);
2003 	}
2004 }
2005 
2006 static void scrub_page_get(struct scrub_page *spage)
2007 {
2008 	atomic_inc(&spage->refs);
2009 }
2010 
2011 static void scrub_page_put(struct scrub_page *spage)
2012 {
2013 	if (atomic_dec_and_test(&spage->refs)) {
2014 		if (spage->page)
2015 			__free_page(spage->page);
2016 		kfree(spage);
2017 	}
2018 }
2019 
2020 static void scrub_submit(struct scrub_ctx *sctx)
2021 {
2022 	struct scrub_bio *sbio;
2023 
2024 	if (sctx->curr == -1)
2025 		return;
2026 
2027 	sbio = sctx->bios[sctx->curr];
2028 	sctx->curr = -1;
2029 	scrub_pending_bio_inc(sctx);
2030 	btrfsic_submit_bio(sbio->bio);
2031 }
2032 
2033 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2034 				    struct scrub_page *spage)
2035 {
2036 	struct scrub_block *sblock = spage->sblock;
2037 	struct scrub_bio *sbio;
2038 	int ret;
2039 
2040 again:
2041 	/*
2042 	 * grab a fresh bio or wait for one to become available
2043 	 */
2044 	while (sctx->curr == -1) {
2045 		spin_lock(&sctx->list_lock);
2046 		sctx->curr = sctx->first_free;
2047 		if (sctx->curr != -1) {
2048 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
2049 			sctx->bios[sctx->curr]->next_free = -1;
2050 			sctx->bios[sctx->curr]->page_count = 0;
2051 			spin_unlock(&sctx->list_lock);
2052 		} else {
2053 			spin_unlock(&sctx->list_lock);
2054 			wait_event(sctx->list_wait, sctx->first_free != -1);
2055 		}
2056 	}
2057 	sbio = sctx->bios[sctx->curr];
2058 	if (sbio->page_count == 0) {
2059 		struct bio *bio;
2060 
2061 		sbio->physical = spage->physical;
2062 		sbio->logical = spage->logical;
2063 		sbio->dev = spage->dev;
2064 		bio = sbio->bio;
2065 		if (!bio) {
2066 			bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2067 			sbio->bio = bio;
2068 		}
2069 
2070 		bio->bi_private = sbio;
2071 		bio->bi_end_io = scrub_bio_end_io;
2072 		bio_set_dev(bio, sbio->dev->bdev);
2073 		bio->bi_iter.bi_sector = sbio->physical >> 9;
2074 		bio->bi_opf = REQ_OP_READ;
2075 		sbio->status = 0;
2076 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2077 		   spage->physical ||
2078 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
2079 		   spage->logical ||
2080 		   sbio->dev != spage->dev) {
2081 		scrub_submit(sctx);
2082 		goto again;
2083 	}
2084 
2085 	sbio->pagev[sbio->page_count] = spage;
2086 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2087 	if (ret != PAGE_SIZE) {
2088 		if (sbio->page_count < 1) {
2089 			bio_put(sbio->bio);
2090 			sbio->bio = NULL;
2091 			return -EIO;
2092 		}
2093 		scrub_submit(sctx);
2094 		goto again;
2095 	}
2096 
2097 	scrub_block_get(sblock); /* one for the page added to the bio */
2098 	atomic_inc(&sblock->outstanding_pages);
2099 	sbio->page_count++;
2100 	if (sbio->page_count == sctx->pages_per_rd_bio)
2101 		scrub_submit(sctx);
2102 
2103 	return 0;
2104 }
2105 
2106 static void scrub_missing_raid56_end_io(struct bio *bio)
2107 {
2108 	struct scrub_block *sblock = bio->bi_private;
2109 	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2110 
2111 	if (bio->bi_status)
2112 		sblock->no_io_error_seen = 0;
2113 
2114 	bio_put(bio);
2115 
2116 	btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2117 }
2118 
2119 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2120 {
2121 	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2122 	struct scrub_ctx *sctx = sblock->sctx;
2123 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2124 	u64 logical;
2125 	struct btrfs_device *dev;
2126 
2127 	logical = sblock->pagev[0]->logical;
2128 	dev = sblock->pagev[0]->dev;
2129 
2130 	if (sblock->no_io_error_seen)
2131 		scrub_recheck_block_checksum(sblock);
2132 
2133 	if (!sblock->no_io_error_seen) {
2134 		spin_lock(&sctx->stat_lock);
2135 		sctx->stat.read_errors++;
2136 		spin_unlock(&sctx->stat_lock);
2137 		btrfs_err_rl_in_rcu(fs_info,
2138 			"IO error rebuilding logical %llu for dev %s",
2139 			logical, rcu_str_deref(dev->name));
2140 	} else if (sblock->header_error || sblock->checksum_error) {
2141 		spin_lock(&sctx->stat_lock);
2142 		sctx->stat.uncorrectable_errors++;
2143 		spin_unlock(&sctx->stat_lock);
2144 		btrfs_err_rl_in_rcu(fs_info,
2145 			"failed to rebuild valid logical %llu for dev %s",
2146 			logical, rcu_str_deref(dev->name));
2147 	} else {
2148 		scrub_write_block_to_dev_replace(sblock);
2149 	}
2150 
2151 	scrub_block_put(sblock);
2152 
2153 	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2154 		mutex_lock(&sctx->wr_lock);
2155 		scrub_wr_submit(sctx);
2156 		mutex_unlock(&sctx->wr_lock);
2157 	}
2158 
2159 	scrub_pending_bio_dec(sctx);
2160 }
2161 
2162 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2163 {
2164 	struct scrub_ctx *sctx = sblock->sctx;
2165 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2166 	u64 length = sblock->page_count * PAGE_SIZE;
2167 	u64 logical = sblock->pagev[0]->logical;
2168 	struct btrfs_bio *bbio = NULL;
2169 	struct bio *bio;
2170 	struct btrfs_raid_bio *rbio;
2171 	int ret;
2172 	int i;
2173 
2174 	btrfs_bio_counter_inc_blocked(fs_info);
2175 	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2176 			&length, &bbio);
2177 	if (ret || !bbio || !bbio->raid_map)
2178 		goto bbio_out;
2179 
2180 	if (WARN_ON(!sctx->is_dev_replace ||
2181 		    !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2182 		/*
2183 		 * We shouldn't be scrubbing a missing device. Even for dev
2184 		 * replace, we should only get here for RAID 5/6. We either
2185 		 * managed to mount something with no mirrors remaining or
2186 		 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2187 		 */
2188 		goto bbio_out;
2189 	}
2190 
2191 	bio = btrfs_io_bio_alloc(0);
2192 	bio->bi_iter.bi_sector = logical >> 9;
2193 	bio->bi_private = sblock;
2194 	bio->bi_end_io = scrub_missing_raid56_end_io;
2195 
2196 	rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2197 	if (!rbio)
2198 		goto rbio_out;
2199 
2200 	for (i = 0; i < sblock->page_count; i++) {
2201 		struct scrub_page *spage = sblock->pagev[i];
2202 
2203 		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2204 	}
2205 
2206 	btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2207 			scrub_missing_raid56_worker, NULL, NULL);
2208 	scrub_block_get(sblock);
2209 	scrub_pending_bio_inc(sctx);
2210 	raid56_submit_missing_rbio(rbio);
2211 	return;
2212 
2213 rbio_out:
2214 	bio_put(bio);
2215 bbio_out:
2216 	btrfs_bio_counter_dec(fs_info);
2217 	btrfs_put_bbio(bbio);
2218 	spin_lock(&sctx->stat_lock);
2219 	sctx->stat.malloc_errors++;
2220 	spin_unlock(&sctx->stat_lock);
2221 }
2222 
2223 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2224 		       u64 physical, struct btrfs_device *dev, u64 flags,
2225 		       u64 gen, int mirror_num, u8 *csum, int force,
2226 		       u64 physical_for_dev_replace)
2227 {
2228 	struct scrub_block *sblock;
2229 	int index;
2230 
2231 	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2232 	if (!sblock) {
2233 		spin_lock(&sctx->stat_lock);
2234 		sctx->stat.malloc_errors++;
2235 		spin_unlock(&sctx->stat_lock);
2236 		return -ENOMEM;
2237 	}
2238 
2239 	/* one ref inside this function, plus one for each page added to
2240 	 * a bio later on */
2241 	refcount_set(&sblock->refs, 1);
2242 	sblock->sctx = sctx;
2243 	sblock->no_io_error_seen = 1;
2244 
2245 	for (index = 0; len > 0; index++) {
2246 		struct scrub_page *spage;
2247 		u64 l = min_t(u64, len, PAGE_SIZE);
2248 
2249 		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2250 		if (!spage) {
2251 leave_nomem:
2252 			spin_lock(&sctx->stat_lock);
2253 			sctx->stat.malloc_errors++;
2254 			spin_unlock(&sctx->stat_lock);
2255 			scrub_block_put(sblock);
2256 			return -ENOMEM;
2257 		}
2258 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2259 		scrub_page_get(spage);
2260 		sblock->pagev[index] = spage;
2261 		spage->sblock = sblock;
2262 		spage->dev = dev;
2263 		spage->flags = flags;
2264 		spage->generation = gen;
2265 		spage->logical = logical;
2266 		spage->physical = physical;
2267 		spage->physical_for_dev_replace = physical_for_dev_replace;
2268 		spage->mirror_num = mirror_num;
2269 		if (csum) {
2270 			spage->have_csum = 1;
2271 			memcpy(spage->csum, csum, sctx->csum_size);
2272 		} else {
2273 			spage->have_csum = 0;
2274 		}
2275 		sblock->page_count++;
2276 		spage->page = alloc_page(GFP_KERNEL);
2277 		if (!spage->page)
2278 			goto leave_nomem;
2279 		len -= l;
2280 		logical += l;
2281 		physical += l;
2282 		physical_for_dev_replace += l;
2283 	}
2284 
2285 	WARN_ON(sblock->page_count == 0);
2286 	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2287 		/*
2288 		 * This case should only be hit for RAID 5/6 device replace. See
2289 		 * the comment in scrub_missing_raid56_pages() for details.
2290 		 */
2291 		scrub_missing_raid56_pages(sblock);
2292 	} else {
2293 		for (index = 0; index < sblock->page_count; index++) {
2294 			struct scrub_page *spage = sblock->pagev[index];
2295 			int ret;
2296 
2297 			ret = scrub_add_page_to_rd_bio(sctx, spage);
2298 			if (ret) {
2299 				scrub_block_put(sblock);
2300 				return ret;
2301 			}
2302 		}
2303 
2304 		if (force)
2305 			scrub_submit(sctx);
2306 	}
2307 
2308 	/* last one frees, either here or in bio completion for last page */
2309 	scrub_block_put(sblock);
2310 	return 0;
2311 }
2312 
2313 static void scrub_bio_end_io(struct bio *bio)
2314 {
2315 	struct scrub_bio *sbio = bio->bi_private;
2316 	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2317 
2318 	sbio->status = bio->bi_status;
2319 	sbio->bio = bio;
2320 
2321 	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2322 }
2323 
2324 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2325 {
2326 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2327 	struct scrub_ctx *sctx = sbio->sctx;
2328 	int i;
2329 
2330 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2331 	if (sbio->status) {
2332 		for (i = 0; i < sbio->page_count; i++) {
2333 			struct scrub_page *spage = sbio->pagev[i];
2334 
2335 			spage->io_error = 1;
2336 			spage->sblock->no_io_error_seen = 0;
2337 		}
2338 	}
2339 
2340 	/* now complete the scrub_block items that have all pages completed */
2341 	for (i = 0; i < sbio->page_count; i++) {
2342 		struct scrub_page *spage = sbio->pagev[i];
2343 		struct scrub_block *sblock = spage->sblock;
2344 
2345 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2346 			scrub_block_complete(sblock);
2347 		scrub_block_put(sblock);
2348 	}
2349 
2350 	bio_put(sbio->bio);
2351 	sbio->bio = NULL;
2352 	spin_lock(&sctx->list_lock);
2353 	sbio->next_free = sctx->first_free;
2354 	sctx->first_free = sbio->index;
2355 	spin_unlock(&sctx->list_lock);
2356 
2357 	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2358 		mutex_lock(&sctx->wr_lock);
2359 		scrub_wr_submit(sctx);
2360 		mutex_unlock(&sctx->wr_lock);
2361 	}
2362 
2363 	scrub_pending_bio_dec(sctx);
2364 }
2365 
2366 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2367 				       unsigned long *bitmap,
2368 				       u64 start, u64 len)
2369 {
2370 	u64 offset;
2371 	u64 nsectors64;
2372 	u32 nsectors;
2373 	int sectorsize = sparity->sctx->fs_info->sectorsize;
2374 
2375 	if (len >= sparity->stripe_len) {
2376 		bitmap_set(bitmap, 0, sparity->nsectors);
2377 		return;
2378 	}
2379 
2380 	start -= sparity->logic_start;
2381 	start = div64_u64_rem(start, sparity->stripe_len, &offset);
2382 	offset = div_u64(offset, sectorsize);
2383 	nsectors64 = div_u64(len, sectorsize);
2384 
2385 	ASSERT(nsectors64 < UINT_MAX);
2386 	nsectors = (u32)nsectors64;
2387 
2388 	if (offset + nsectors <= sparity->nsectors) {
2389 		bitmap_set(bitmap, offset, nsectors);
2390 		return;
2391 	}
2392 
2393 	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2394 	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2395 }
2396 
2397 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2398 						   u64 start, u64 len)
2399 {
2400 	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2401 }
2402 
2403 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2404 						  u64 start, u64 len)
2405 {
2406 	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2407 }
2408 
2409 static void scrub_block_complete(struct scrub_block *sblock)
2410 {
2411 	int corrupted = 0;
2412 
2413 	if (!sblock->no_io_error_seen) {
2414 		corrupted = 1;
2415 		scrub_handle_errored_block(sblock);
2416 	} else {
2417 		/*
2418 		 * if has checksum error, write via repair mechanism in
2419 		 * dev replace case, otherwise write here in dev replace
2420 		 * case.
2421 		 */
2422 		corrupted = scrub_checksum(sblock);
2423 		if (!corrupted && sblock->sctx->is_dev_replace)
2424 			scrub_write_block_to_dev_replace(sblock);
2425 	}
2426 
2427 	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2428 		u64 start = sblock->pagev[0]->logical;
2429 		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2430 			  PAGE_SIZE;
2431 
2432 		scrub_parity_mark_sectors_error(sblock->sparity,
2433 						start, end - start);
2434 	}
2435 }
2436 
2437 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2438 {
2439 	struct btrfs_ordered_sum *sum = NULL;
2440 	unsigned long index;
2441 	unsigned long num_sectors;
2442 
2443 	while (!list_empty(&sctx->csum_list)) {
2444 		sum = list_first_entry(&sctx->csum_list,
2445 				       struct btrfs_ordered_sum, list);
2446 		if (sum->bytenr > logical)
2447 			return 0;
2448 		if (sum->bytenr + sum->len > logical)
2449 			break;
2450 
2451 		++sctx->stat.csum_discards;
2452 		list_del(&sum->list);
2453 		kfree(sum);
2454 		sum = NULL;
2455 	}
2456 	if (!sum)
2457 		return 0;
2458 
2459 	index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2460 	ASSERT(index < UINT_MAX);
2461 
2462 	num_sectors = sum->len / sctx->fs_info->sectorsize;
2463 	memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2464 	if (index == num_sectors - 1) {
2465 		list_del(&sum->list);
2466 		kfree(sum);
2467 	}
2468 	return 1;
2469 }
2470 
2471 /* scrub extent tries to collect up to 64 kB for each bio */
2472 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2473 			u64 logical, u64 len,
2474 			u64 physical, struct btrfs_device *dev, u64 flags,
2475 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2476 {
2477 	int ret;
2478 	u8 csum[BTRFS_CSUM_SIZE];
2479 	u32 blocksize;
2480 
2481 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2482 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2483 			blocksize = map->stripe_len;
2484 		else
2485 			blocksize = sctx->fs_info->sectorsize;
2486 		spin_lock(&sctx->stat_lock);
2487 		sctx->stat.data_extents_scrubbed++;
2488 		sctx->stat.data_bytes_scrubbed += len;
2489 		spin_unlock(&sctx->stat_lock);
2490 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2491 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2492 			blocksize = map->stripe_len;
2493 		else
2494 			blocksize = sctx->fs_info->nodesize;
2495 		spin_lock(&sctx->stat_lock);
2496 		sctx->stat.tree_extents_scrubbed++;
2497 		sctx->stat.tree_bytes_scrubbed += len;
2498 		spin_unlock(&sctx->stat_lock);
2499 	} else {
2500 		blocksize = sctx->fs_info->sectorsize;
2501 		WARN_ON(1);
2502 	}
2503 
2504 	while (len) {
2505 		u64 l = min_t(u64, len, blocksize);
2506 		int have_csum = 0;
2507 
2508 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2509 			/* push csums to sbio */
2510 			have_csum = scrub_find_csum(sctx, logical, csum);
2511 			if (have_csum == 0)
2512 				++sctx->stat.no_csum;
2513 		}
2514 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2515 				  mirror_num, have_csum ? csum : NULL, 0,
2516 				  physical_for_dev_replace);
2517 		if (ret)
2518 			return ret;
2519 		len -= l;
2520 		logical += l;
2521 		physical += l;
2522 		physical_for_dev_replace += l;
2523 	}
2524 	return 0;
2525 }
2526 
2527 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2528 				  u64 logical, u64 len,
2529 				  u64 physical, struct btrfs_device *dev,
2530 				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2531 {
2532 	struct scrub_ctx *sctx = sparity->sctx;
2533 	struct scrub_block *sblock;
2534 	int index;
2535 
2536 	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2537 	if (!sblock) {
2538 		spin_lock(&sctx->stat_lock);
2539 		sctx->stat.malloc_errors++;
2540 		spin_unlock(&sctx->stat_lock);
2541 		return -ENOMEM;
2542 	}
2543 
2544 	/* one ref inside this function, plus one for each page added to
2545 	 * a bio later on */
2546 	refcount_set(&sblock->refs, 1);
2547 	sblock->sctx = sctx;
2548 	sblock->no_io_error_seen = 1;
2549 	sblock->sparity = sparity;
2550 	scrub_parity_get(sparity);
2551 
2552 	for (index = 0; len > 0; index++) {
2553 		struct scrub_page *spage;
2554 		u64 l = min_t(u64, len, PAGE_SIZE);
2555 
2556 		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2557 		if (!spage) {
2558 leave_nomem:
2559 			spin_lock(&sctx->stat_lock);
2560 			sctx->stat.malloc_errors++;
2561 			spin_unlock(&sctx->stat_lock);
2562 			scrub_block_put(sblock);
2563 			return -ENOMEM;
2564 		}
2565 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2566 		/* For scrub block */
2567 		scrub_page_get(spage);
2568 		sblock->pagev[index] = spage;
2569 		/* For scrub parity */
2570 		scrub_page_get(spage);
2571 		list_add_tail(&spage->list, &sparity->spages);
2572 		spage->sblock = sblock;
2573 		spage->dev = dev;
2574 		spage->flags = flags;
2575 		spage->generation = gen;
2576 		spage->logical = logical;
2577 		spage->physical = physical;
2578 		spage->mirror_num = mirror_num;
2579 		if (csum) {
2580 			spage->have_csum = 1;
2581 			memcpy(spage->csum, csum, sctx->csum_size);
2582 		} else {
2583 			spage->have_csum = 0;
2584 		}
2585 		sblock->page_count++;
2586 		spage->page = alloc_page(GFP_KERNEL);
2587 		if (!spage->page)
2588 			goto leave_nomem;
2589 		len -= l;
2590 		logical += l;
2591 		physical += l;
2592 	}
2593 
2594 	WARN_ON(sblock->page_count == 0);
2595 	for (index = 0; index < sblock->page_count; index++) {
2596 		struct scrub_page *spage = sblock->pagev[index];
2597 		int ret;
2598 
2599 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2600 		if (ret) {
2601 			scrub_block_put(sblock);
2602 			return ret;
2603 		}
2604 	}
2605 
2606 	/* last one frees, either here or in bio completion for last page */
2607 	scrub_block_put(sblock);
2608 	return 0;
2609 }
2610 
2611 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2612 				   u64 logical, u64 len,
2613 				   u64 physical, struct btrfs_device *dev,
2614 				   u64 flags, u64 gen, int mirror_num)
2615 {
2616 	struct scrub_ctx *sctx = sparity->sctx;
2617 	int ret;
2618 	u8 csum[BTRFS_CSUM_SIZE];
2619 	u32 blocksize;
2620 
2621 	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2622 		scrub_parity_mark_sectors_error(sparity, logical, len);
2623 		return 0;
2624 	}
2625 
2626 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2627 		blocksize = sparity->stripe_len;
2628 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2629 		blocksize = sparity->stripe_len;
2630 	} else {
2631 		blocksize = sctx->fs_info->sectorsize;
2632 		WARN_ON(1);
2633 	}
2634 
2635 	while (len) {
2636 		u64 l = min_t(u64, len, blocksize);
2637 		int have_csum = 0;
2638 
2639 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2640 			/* push csums to sbio */
2641 			have_csum = scrub_find_csum(sctx, logical, csum);
2642 			if (have_csum == 0)
2643 				goto skip;
2644 		}
2645 		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2646 					     flags, gen, mirror_num,
2647 					     have_csum ? csum : NULL);
2648 		if (ret)
2649 			return ret;
2650 skip:
2651 		len -= l;
2652 		logical += l;
2653 		physical += l;
2654 	}
2655 	return 0;
2656 }
2657 
2658 /*
2659  * Given a physical address, this will calculate it's
2660  * logical offset. if this is a parity stripe, it will return
2661  * the most left data stripe's logical offset.
2662  *
2663  * return 0 if it is a data stripe, 1 means parity stripe.
2664  */
2665 static int get_raid56_logic_offset(u64 physical, int num,
2666 				   struct map_lookup *map, u64 *offset,
2667 				   u64 *stripe_start)
2668 {
2669 	int i;
2670 	int j = 0;
2671 	u64 stripe_nr;
2672 	u64 last_offset;
2673 	u32 stripe_index;
2674 	u32 rot;
2675 	const int data_stripes = nr_data_stripes(map);
2676 
2677 	last_offset = (physical - map->stripes[num].physical) * data_stripes;
2678 	if (stripe_start)
2679 		*stripe_start = last_offset;
2680 
2681 	*offset = last_offset;
2682 	for (i = 0; i < data_stripes; i++) {
2683 		*offset = last_offset + i * map->stripe_len;
2684 
2685 		stripe_nr = div64_u64(*offset, map->stripe_len);
2686 		stripe_nr = div_u64(stripe_nr, data_stripes);
2687 
2688 		/* Work out the disk rotation on this stripe-set */
2689 		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2690 		/* calculate which stripe this data locates */
2691 		rot += i;
2692 		stripe_index = rot % map->num_stripes;
2693 		if (stripe_index == num)
2694 			return 0;
2695 		if (stripe_index < num)
2696 			j++;
2697 	}
2698 	*offset = last_offset + j * map->stripe_len;
2699 	return 1;
2700 }
2701 
2702 static void scrub_free_parity(struct scrub_parity *sparity)
2703 {
2704 	struct scrub_ctx *sctx = sparity->sctx;
2705 	struct scrub_page *curr, *next;
2706 	int nbits;
2707 
2708 	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2709 	if (nbits) {
2710 		spin_lock(&sctx->stat_lock);
2711 		sctx->stat.read_errors += nbits;
2712 		sctx->stat.uncorrectable_errors += nbits;
2713 		spin_unlock(&sctx->stat_lock);
2714 	}
2715 
2716 	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2717 		list_del_init(&curr->list);
2718 		scrub_page_put(curr);
2719 	}
2720 
2721 	kfree(sparity);
2722 }
2723 
2724 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2725 {
2726 	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2727 						    work);
2728 	struct scrub_ctx *sctx = sparity->sctx;
2729 
2730 	scrub_free_parity(sparity);
2731 	scrub_pending_bio_dec(sctx);
2732 }
2733 
2734 static void scrub_parity_bio_endio(struct bio *bio)
2735 {
2736 	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2737 	struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2738 
2739 	if (bio->bi_status)
2740 		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2741 			  sparity->nsectors);
2742 
2743 	bio_put(bio);
2744 
2745 	btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2746 			scrub_parity_bio_endio_worker, NULL, NULL);
2747 	btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2748 }
2749 
2750 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2751 {
2752 	struct scrub_ctx *sctx = sparity->sctx;
2753 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2754 	struct bio *bio;
2755 	struct btrfs_raid_bio *rbio;
2756 	struct btrfs_bio *bbio = NULL;
2757 	u64 length;
2758 	int ret;
2759 
2760 	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2761 			   sparity->nsectors))
2762 		goto out;
2763 
2764 	length = sparity->logic_end - sparity->logic_start;
2765 
2766 	btrfs_bio_counter_inc_blocked(fs_info);
2767 	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2768 			       &length, &bbio);
2769 	if (ret || !bbio || !bbio->raid_map)
2770 		goto bbio_out;
2771 
2772 	bio = btrfs_io_bio_alloc(0);
2773 	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2774 	bio->bi_private = sparity;
2775 	bio->bi_end_io = scrub_parity_bio_endio;
2776 
2777 	rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2778 					      length, sparity->scrub_dev,
2779 					      sparity->dbitmap,
2780 					      sparity->nsectors);
2781 	if (!rbio)
2782 		goto rbio_out;
2783 
2784 	scrub_pending_bio_inc(sctx);
2785 	raid56_parity_submit_scrub_rbio(rbio);
2786 	return;
2787 
2788 rbio_out:
2789 	bio_put(bio);
2790 bbio_out:
2791 	btrfs_bio_counter_dec(fs_info);
2792 	btrfs_put_bbio(bbio);
2793 	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2794 		  sparity->nsectors);
2795 	spin_lock(&sctx->stat_lock);
2796 	sctx->stat.malloc_errors++;
2797 	spin_unlock(&sctx->stat_lock);
2798 out:
2799 	scrub_free_parity(sparity);
2800 }
2801 
2802 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2803 {
2804 	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2805 }
2806 
2807 static void scrub_parity_get(struct scrub_parity *sparity)
2808 {
2809 	refcount_inc(&sparity->refs);
2810 }
2811 
2812 static void scrub_parity_put(struct scrub_parity *sparity)
2813 {
2814 	if (!refcount_dec_and_test(&sparity->refs))
2815 		return;
2816 
2817 	scrub_parity_check_and_repair(sparity);
2818 }
2819 
2820 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2821 						  struct map_lookup *map,
2822 						  struct btrfs_device *sdev,
2823 						  struct btrfs_path *path,
2824 						  u64 logic_start,
2825 						  u64 logic_end)
2826 {
2827 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2828 	struct btrfs_root *root = fs_info->extent_root;
2829 	struct btrfs_root *csum_root = fs_info->csum_root;
2830 	struct btrfs_extent_item *extent;
2831 	struct btrfs_bio *bbio = NULL;
2832 	u64 flags;
2833 	int ret;
2834 	int slot;
2835 	struct extent_buffer *l;
2836 	struct btrfs_key key;
2837 	u64 generation;
2838 	u64 extent_logical;
2839 	u64 extent_physical;
2840 	u64 extent_len;
2841 	u64 mapped_length;
2842 	struct btrfs_device *extent_dev;
2843 	struct scrub_parity *sparity;
2844 	int nsectors;
2845 	int bitmap_len;
2846 	int extent_mirror_num;
2847 	int stop_loop = 0;
2848 
2849 	nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2850 	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2851 	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2852 			  GFP_NOFS);
2853 	if (!sparity) {
2854 		spin_lock(&sctx->stat_lock);
2855 		sctx->stat.malloc_errors++;
2856 		spin_unlock(&sctx->stat_lock);
2857 		return -ENOMEM;
2858 	}
2859 
2860 	sparity->stripe_len = map->stripe_len;
2861 	sparity->nsectors = nsectors;
2862 	sparity->sctx = sctx;
2863 	sparity->scrub_dev = sdev;
2864 	sparity->logic_start = logic_start;
2865 	sparity->logic_end = logic_end;
2866 	refcount_set(&sparity->refs, 1);
2867 	INIT_LIST_HEAD(&sparity->spages);
2868 	sparity->dbitmap = sparity->bitmap;
2869 	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2870 
2871 	ret = 0;
2872 	while (logic_start < logic_end) {
2873 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2874 			key.type = BTRFS_METADATA_ITEM_KEY;
2875 		else
2876 			key.type = BTRFS_EXTENT_ITEM_KEY;
2877 		key.objectid = logic_start;
2878 		key.offset = (u64)-1;
2879 
2880 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2881 		if (ret < 0)
2882 			goto out;
2883 
2884 		if (ret > 0) {
2885 			ret = btrfs_previous_extent_item(root, path, 0);
2886 			if (ret < 0)
2887 				goto out;
2888 			if (ret > 0) {
2889 				btrfs_release_path(path);
2890 				ret = btrfs_search_slot(NULL, root, &key,
2891 							path, 0, 0);
2892 				if (ret < 0)
2893 					goto out;
2894 			}
2895 		}
2896 
2897 		stop_loop = 0;
2898 		while (1) {
2899 			u64 bytes;
2900 
2901 			l = path->nodes[0];
2902 			slot = path->slots[0];
2903 			if (slot >= btrfs_header_nritems(l)) {
2904 				ret = btrfs_next_leaf(root, path);
2905 				if (ret == 0)
2906 					continue;
2907 				if (ret < 0)
2908 					goto out;
2909 
2910 				stop_loop = 1;
2911 				break;
2912 			}
2913 			btrfs_item_key_to_cpu(l, &key, slot);
2914 
2915 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2916 			    key.type != BTRFS_METADATA_ITEM_KEY)
2917 				goto next;
2918 
2919 			if (key.type == BTRFS_METADATA_ITEM_KEY)
2920 				bytes = fs_info->nodesize;
2921 			else
2922 				bytes = key.offset;
2923 
2924 			if (key.objectid + bytes <= logic_start)
2925 				goto next;
2926 
2927 			if (key.objectid >= logic_end) {
2928 				stop_loop = 1;
2929 				break;
2930 			}
2931 
2932 			while (key.objectid >= logic_start + map->stripe_len)
2933 				logic_start += map->stripe_len;
2934 
2935 			extent = btrfs_item_ptr(l, slot,
2936 						struct btrfs_extent_item);
2937 			flags = btrfs_extent_flags(l, extent);
2938 			generation = btrfs_extent_generation(l, extent);
2939 
2940 			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2941 			    (key.objectid < logic_start ||
2942 			     key.objectid + bytes >
2943 			     logic_start + map->stripe_len)) {
2944 				btrfs_err(fs_info,
2945 					  "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2946 					  key.objectid, logic_start);
2947 				spin_lock(&sctx->stat_lock);
2948 				sctx->stat.uncorrectable_errors++;
2949 				spin_unlock(&sctx->stat_lock);
2950 				goto next;
2951 			}
2952 again:
2953 			extent_logical = key.objectid;
2954 			extent_len = bytes;
2955 
2956 			if (extent_logical < logic_start) {
2957 				extent_len -= logic_start - extent_logical;
2958 				extent_logical = logic_start;
2959 			}
2960 
2961 			if (extent_logical + extent_len >
2962 			    logic_start + map->stripe_len)
2963 				extent_len = logic_start + map->stripe_len -
2964 					     extent_logical;
2965 
2966 			scrub_parity_mark_sectors_data(sparity, extent_logical,
2967 						       extent_len);
2968 
2969 			mapped_length = extent_len;
2970 			bbio = NULL;
2971 			ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2972 					extent_logical, &mapped_length, &bbio,
2973 					0);
2974 			if (!ret) {
2975 				if (!bbio || mapped_length < extent_len)
2976 					ret = -EIO;
2977 			}
2978 			if (ret) {
2979 				btrfs_put_bbio(bbio);
2980 				goto out;
2981 			}
2982 			extent_physical = bbio->stripes[0].physical;
2983 			extent_mirror_num = bbio->mirror_num;
2984 			extent_dev = bbio->stripes[0].dev;
2985 			btrfs_put_bbio(bbio);
2986 
2987 			ret = btrfs_lookup_csums_range(csum_root,
2988 						extent_logical,
2989 						extent_logical + extent_len - 1,
2990 						&sctx->csum_list, 1);
2991 			if (ret)
2992 				goto out;
2993 
2994 			ret = scrub_extent_for_parity(sparity, extent_logical,
2995 						      extent_len,
2996 						      extent_physical,
2997 						      extent_dev, flags,
2998 						      generation,
2999 						      extent_mirror_num);
3000 
3001 			scrub_free_csums(sctx);
3002 
3003 			if (ret)
3004 				goto out;
3005 
3006 			if (extent_logical + extent_len <
3007 			    key.objectid + bytes) {
3008 				logic_start += map->stripe_len;
3009 
3010 				if (logic_start >= logic_end) {
3011 					stop_loop = 1;
3012 					break;
3013 				}
3014 
3015 				if (logic_start < key.objectid + bytes) {
3016 					cond_resched();
3017 					goto again;
3018 				}
3019 			}
3020 next:
3021 			path->slots[0]++;
3022 		}
3023 
3024 		btrfs_release_path(path);
3025 
3026 		if (stop_loop)
3027 			break;
3028 
3029 		logic_start += map->stripe_len;
3030 	}
3031 out:
3032 	if (ret < 0)
3033 		scrub_parity_mark_sectors_error(sparity, logic_start,
3034 						logic_end - logic_start);
3035 	scrub_parity_put(sparity);
3036 	scrub_submit(sctx);
3037 	mutex_lock(&sctx->wr_lock);
3038 	scrub_wr_submit(sctx);
3039 	mutex_unlock(&sctx->wr_lock);
3040 
3041 	btrfs_release_path(path);
3042 	return ret < 0 ? ret : 0;
3043 }
3044 
3045 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3046 					   struct map_lookup *map,
3047 					   struct btrfs_device *scrub_dev,
3048 					   int num, u64 base, u64 length)
3049 {
3050 	struct btrfs_path *path, *ppath;
3051 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3052 	struct btrfs_root *root = fs_info->extent_root;
3053 	struct btrfs_root *csum_root = fs_info->csum_root;
3054 	struct btrfs_extent_item *extent;
3055 	struct blk_plug plug;
3056 	u64 flags;
3057 	int ret;
3058 	int slot;
3059 	u64 nstripes;
3060 	struct extent_buffer *l;
3061 	u64 physical;
3062 	u64 logical;
3063 	u64 logic_end;
3064 	u64 physical_end;
3065 	u64 generation;
3066 	int mirror_num;
3067 	struct reada_control *reada1;
3068 	struct reada_control *reada2;
3069 	struct btrfs_key key;
3070 	struct btrfs_key key_end;
3071 	u64 increment = map->stripe_len;
3072 	u64 offset;
3073 	u64 extent_logical;
3074 	u64 extent_physical;
3075 	u64 extent_len;
3076 	u64 stripe_logical;
3077 	u64 stripe_end;
3078 	struct btrfs_device *extent_dev;
3079 	int extent_mirror_num;
3080 	int stop_loop = 0;
3081 
3082 	physical = map->stripes[num].physical;
3083 	offset = 0;
3084 	nstripes = div64_u64(length, map->stripe_len);
3085 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3086 		offset = map->stripe_len * num;
3087 		increment = map->stripe_len * map->num_stripes;
3088 		mirror_num = 1;
3089 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3090 		int factor = map->num_stripes / map->sub_stripes;
3091 		offset = map->stripe_len * (num / map->sub_stripes);
3092 		increment = map->stripe_len * factor;
3093 		mirror_num = num % map->sub_stripes + 1;
3094 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3095 		increment = map->stripe_len;
3096 		mirror_num = num % map->num_stripes + 1;
3097 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3098 		increment = map->stripe_len;
3099 		mirror_num = num % map->num_stripes + 1;
3100 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3101 		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3102 		increment = map->stripe_len * nr_data_stripes(map);
3103 		mirror_num = 1;
3104 	} else {
3105 		increment = map->stripe_len;
3106 		mirror_num = 1;
3107 	}
3108 
3109 	path = btrfs_alloc_path();
3110 	if (!path)
3111 		return -ENOMEM;
3112 
3113 	ppath = btrfs_alloc_path();
3114 	if (!ppath) {
3115 		btrfs_free_path(path);
3116 		return -ENOMEM;
3117 	}
3118 
3119 	/*
3120 	 * work on commit root. The related disk blocks are static as
3121 	 * long as COW is applied. This means, it is save to rewrite
3122 	 * them to repair disk errors without any race conditions
3123 	 */
3124 	path->search_commit_root = 1;
3125 	path->skip_locking = 1;
3126 
3127 	ppath->search_commit_root = 1;
3128 	ppath->skip_locking = 1;
3129 	/*
3130 	 * trigger the readahead for extent tree csum tree and wait for
3131 	 * completion. During readahead, the scrub is officially paused
3132 	 * to not hold off transaction commits
3133 	 */
3134 	logical = base + offset;
3135 	physical_end = physical + nstripes * map->stripe_len;
3136 	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3137 		get_raid56_logic_offset(physical_end, num,
3138 					map, &logic_end, NULL);
3139 		logic_end += base;
3140 	} else {
3141 		logic_end = logical + increment * nstripes;
3142 	}
3143 	wait_event(sctx->list_wait,
3144 		   atomic_read(&sctx->bios_in_flight) == 0);
3145 	scrub_blocked_if_needed(fs_info);
3146 
3147 	/* FIXME it might be better to start readahead at commit root */
3148 	key.objectid = logical;
3149 	key.type = BTRFS_EXTENT_ITEM_KEY;
3150 	key.offset = (u64)0;
3151 	key_end.objectid = logic_end;
3152 	key_end.type = BTRFS_METADATA_ITEM_KEY;
3153 	key_end.offset = (u64)-1;
3154 	reada1 = btrfs_reada_add(root, &key, &key_end);
3155 
3156 	key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3157 	key.type = BTRFS_EXTENT_CSUM_KEY;
3158 	key.offset = logical;
3159 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3160 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
3161 	key_end.offset = logic_end;
3162 	reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3163 
3164 	if (!IS_ERR(reada1))
3165 		btrfs_reada_wait(reada1);
3166 	if (!IS_ERR(reada2))
3167 		btrfs_reada_wait(reada2);
3168 
3169 
3170 	/*
3171 	 * collect all data csums for the stripe to avoid seeking during
3172 	 * the scrub. This might currently (crc32) end up to be about 1MB
3173 	 */
3174 	blk_start_plug(&plug);
3175 
3176 	/*
3177 	 * now find all extents for each stripe and scrub them
3178 	 */
3179 	ret = 0;
3180 	while (physical < physical_end) {
3181 		/*
3182 		 * canceled?
3183 		 */
3184 		if (atomic_read(&fs_info->scrub_cancel_req) ||
3185 		    atomic_read(&sctx->cancel_req)) {
3186 			ret = -ECANCELED;
3187 			goto out;
3188 		}
3189 		/*
3190 		 * check to see if we have to pause
3191 		 */
3192 		if (atomic_read(&fs_info->scrub_pause_req)) {
3193 			/* push queued extents */
3194 			sctx->flush_all_writes = true;
3195 			scrub_submit(sctx);
3196 			mutex_lock(&sctx->wr_lock);
3197 			scrub_wr_submit(sctx);
3198 			mutex_unlock(&sctx->wr_lock);
3199 			wait_event(sctx->list_wait,
3200 				   atomic_read(&sctx->bios_in_flight) == 0);
3201 			sctx->flush_all_writes = false;
3202 			scrub_blocked_if_needed(fs_info);
3203 		}
3204 
3205 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3206 			ret = get_raid56_logic_offset(physical, num, map,
3207 						      &logical,
3208 						      &stripe_logical);
3209 			logical += base;
3210 			if (ret) {
3211 				/* it is parity strip */
3212 				stripe_logical += base;
3213 				stripe_end = stripe_logical + increment;
3214 				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3215 							  ppath, stripe_logical,
3216 							  stripe_end);
3217 				if (ret)
3218 					goto out;
3219 				goto skip;
3220 			}
3221 		}
3222 
3223 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3224 			key.type = BTRFS_METADATA_ITEM_KEY;
3225 		else
3226 			key.type = BTRFS_EXTENT_ITEM_KEY;
3227 		key.objectid = logical;
3228 		key.offset = (u64)-1;
3229 
3230 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3231 		if (ret < 0)
3232 			goto out;
3233 
3234 		if (ret > 0) {
3235 			ret = btrfs_previous_extent_item(root, path, 0);
3236 			if (ret < 0)
3237 				goto out;
3238 			if (ret > 0) {
3239 				/* there's no smaller item, so stick with the
3240 				 * larger one */
3241 				btrfs_release_path(path);
3242 				ret = btrfs_search_slot(NULL, root, &key,
3243 							path, 0, 0);
3244 				if (ret < 0)
3245 					goto out;
3246 			}
3247 		}
3248 
3249 		stop_loop = 0;
3250 		while (1) {
3251 			u64 bytes;
3252 
3253 			l = path->nodes[0];
3254 			slot = path->slots[0];
3255 			if (slot >= btrfs_header_nritems(l)) {
3256 				ret = btrfs_next_leaf(root, path);
3257 				if (ret == 0)
3258 					continue;
3259 				if (ret < 0)
3260 					goto out;
3261 
3262 				stop_loop = 1;
3263 				break;
3264 			}
3265 			btrfs_item_key_to_cpu(l, &key, slot);
3266 
3267 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3268 			    key.type != BTRFS_METADATA_ITEM_KEY)
3269 				goto next;
3270 
3271 			if (key.type == BTRFS_METADATA_ITEM_KEY)
3272 				bytes = fs_info->nodesize;
3273 			else
3274 				bytes = key.offset;
3275 
3276 			if (key.objectid + bytes <= logical)
3277 				goto next;
3278 
3279 			if (key.objectid >= logical + map->stripe_len) {
3280 				/* out of this device extent */
3281 				if (key.objectid >= logic_end)
3282 					stop_loop = 1;
3283 				break;
3284 			}
3285 
3286 			extent = btrfs_item_ptr(l, slot,
3287 						struct btrfs_extent_item);
3288 			flags = btrfs_extent_flags(l, extent);
3289 			generation = btrfs_extent_generation(l, extent);
3290 
3291 			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3292 			    (key.objectid < logical ||
3293 			     key.objectid + bytes >
3294 			     logical + map->stripe_len)) {
3295 				btrfs_err(fs_info,
3296 					   "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3297 				       key.objectid, logical);
3298 				spin_lock(&sctx->stat_lock);
3299 				sctx->stat.uncorrectable_errors++;
3300 				spin_unlock(&sctx->stat_lock);
3301 				goto next;
3302 			}
3303 
3304 again:
3305 			extent_logical = key.objectid;
3306 			extent_len = bytes;
3307 
3308 			/*
3309 			 * trim extent to this stripe
3310 			 */
3311 			if (extent_logical < logical) {
3312 				extent_len -= logical - extent_logical;
3313 				extent_logical = logical;
3314 			}
3315 			if (extent_logical + extent_len >
3316 			    logical + map->stripe_len) {
3317 				extent_len = logical + map->stripe_len -
3318 					     extent_logical;
3319 			}
3320 
3321 			extent_physical = extent_logical - logical + physical;
3322 			extent_dev = scrub_dev;
3323 			extent_mirror_num = mirror_num;
3324 			if (sctx->is_dev_replace)
3325 				scrub_remap_extent(fs_info, extent_logical,
3326 						   extent_len, &extent_physical,
3327 						   &extent_dev,
3328 						   &extent_mirror_num);
3329 
3330 			ret = btrfs_lookup_csums_range(csum_root,
3331 						       extent_logical,
3332 						       extent_logical +
3333 						       extent_len - 1,
3334 						       &sctx->csum_list, 1);
3335 			if (ret)
3336 				goto out;
3337 
3338 			ret = scrub_extent(sctx, map, extent_logical, extent_len,
3339 					   extent_physical, extent_dev, flags,
3340 					   generation, extent_mirror_num,
3341 					   extent_logical - logical + physical);
3342 
3343 			scrub_free_csums(sctx);
3344 
3345 			if (ret)
3346 				goto out;
3347 
3348 			if (extent_logical + extent_len <
3349 			    key.objectid + bytes) {
3350 				if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3351 					/*
3352 					 * loop until we find next data stripe
3353 					 * or we have finished all stripes.
3354 					 */
3355 loop:
3356 					physical += map->stripe_len;
3357 					ret = get_raid56_logic_offset(physical,
3358 							num, map, &logical,
3359 							&stripe_logical);
3360 					logical += base;
3361 
3362 					if (ret && physical < physical_end) {
3363 						stripe_logical += base;
3364 						stripe_end = stripe_logical +
3365 								increment;
3366 						ret = scrub_raid56_parity(sctx,
3367 							map, scrub_dev, ppath,
3368 							stripe_logical,
3369 							stripe_end);
3370 						if (ret)
3371 							goto out;
3372 						goto loop;
3373 					}
3374 				} else {
3375 					physical += map->stripe_len;
3376 					logical += increment;
3377 				}
3378 				if (logical < key.objectid + bytes) {
3379 					cond_resched();
3380 					goto again;
3381 				}
3382 
3383 				if (physical >= physical_end) {
3384 					stop_loop = 1;
3385 					break;
3386 				}
3387 			}
3388 next:
3389 			path->slots[0]++;
3390 		}
3391 		btrfs_release_path(path);
3392 skip:
3393 		logical += increment;
3394 		physical += map->stripe_len;
3395 		spin_lock(&sctx->stat_lock);
3396 		if (stop_loop)
3397 			sctx->stat.last_physical = map->stripes[num].physical +
3398 						   length;
3399 		else
3400 			sctx->stat.last_physical = physical;
3401 		spin_unlock(&sctx->stat_lock);
3402 		if (stop_loop)
3403 			break;
3404 	}
3405 out:
3406 	/* push queued extents */
3407 	scrub_submit(sctx);
3408 	mutex_lock(&sctx->wr_lock);
3409 	scrub_wr_submit(sctx);
3410 	mutex_unlock(&sctx->wr_lock);
3411 
3412 	blk_finish_plug(&plug);
3413 	btrfs_free_path(path);
3414 	btrfs_free_path(ppath);
3415 	return ret < 0 ? ret : 0;
3416 }
3417 
3418 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3419 					  struct btrfs_device *scrub_dev,
3420 					  u64 chunk_offset, u64 length,
3421 					  u64 dev_offset,
3422 					  struct btrfs_block_group_cache *cache)
3423 {
3424 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3425 	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3426 	struct map_lookup *map;
3427 	struct extent_map *em;
3428 	int i;
3429 	int ret = 0;
3430 
3431 	read_lock(&map_tree->lock);
3432 	em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3433 	read_unlock(&map_tree->lock);
3434 
3435 	if (!em) {
3436 		/*
3437 		 * Might have been an unused block group deleted by the cleaner
3438 		 * kthread or relocation.
3439 		 */
3440 		spin_lock(&cache->lock);
3441 		if (!cache->removed)
3442 			ret = -EINVAL;
3443 		spin_unlock(&cache->lock);
3444 
3445 		return ret;
3446 	}
3447 
3448 	map = em->map_lookup;
3449 	if (em->start != chunk_offset)
3450 		goto out;
3451 
3452 	if (em->len < length)
3453 		goto out;
3454 
3455 	for (i = 0; i < map->num_stripes; ++i) {
3456 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3457 		    map->stripes[i].physical == dev_offset) {
3458 			ret = scrub_stripe(sctx, map, scrub_dev, i,
3459 					   chunk_offset, length);
3460 			if (ret)
3461 				goto out;
3462 		}
3463 	}
3464 out:
3465 	free_extent_map(em);
3466 
3467 	return ret;
3468 }
3469 
3470 static noinline_for_stack
3471 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3472 			   struct btrfs_device *scrub_dev, u64 start, u64 end)
3473 {
3474 	struct btrfs_dev_extent *dev_extent = NULL;
3475 	struct btrfs_path *path;
3476 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3477 	struct btrfs_root *root = fs_info->dev_root;
3478 	u64 length;
3479 	u64 chunk_offset;
3480 	int ret = 0;
3481 	int ro_set;
3482 	int slot;
3483 	struct extent_buffer *l;
3484 	struct btrfs_key key;
3485 	struct btrfs_key found_key;
3486 	struct btrfs_block_group_cache *cache;
3487 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3488 
3489 	path = btrfs_alloc_path();
3490 	if (!path)
3491 		return -ENOMEM;
3492 
3493 	path->reada = READA_FORWARD;
3494 	path->search_commit_root = 1;
3495 	path->skip_locking = 1;
3496 
3497 	key.objectid = scrub_dev->devid;
3498 	key.offset = 0ull;
3499 	key.type = BTRFS_DEV_EXTENT_KEY;
3500 
3501 	while (1) {
3502 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3503 		if (ret < 0)
3504 			break;
3505 		if (ret > 0) {
3506 			if (path->slots[0] >=
3507 			    btrfs_header_nritems(path->nodes[0])) {
3508 				ret = btrfs_next_leaf(root, path);
3509 				if (ret < 0)
3510 					break;
3511 				if (ret > 0) {
3512 					ret = 0;
3513 					break;
3514 				}
3515 			} else {
3516 				ret = 0;
3517 			}
3518 		}
3519 
3520 		l = path->nodes[0];
3521 		slot = path->slots[0];
3522 
3523 		btrfs_item_key_to_cpu(l, &found_key, slot);
3524 
3525 		if (found_key.objectid != scrub_dev->devid)
3526 			break;
3527 
3528 		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3529 			break;
3530 
3531 		if (found_key.offset >= end)
3532 			break;
3533 
3534 		if (found_key.offset < key.offset)
3535 			break;
3536 
3537 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3538 		length = btrfs_dev_extent_length(l, dev_extent);
3539 
3540 		if (found_key.offset + length <= start)
3541 			goto skip;
3542 
3543 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3544 
3545 		/*
3546 		 * get a reference on the corresponding block group to prevent
3547 		 * the chunk from going away while we scrub it
3548 		 */
3549 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3550 
3551 		/* some chunks are removed but not committed to disk yet,
3552 		 * continue scrubbing */
3553 		if (!cache)
3554 			goto skip;
3555 
3556 		/*
3557 		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3558 		 * to avoid deadlock caused by:
3559 		 * btrfs_inc_block_group_ro()
3560 		 * -> btrfs_wait_for_commit()
3561 		 * -> btrfs_commit_transaction()
3562 		 * -> btrfs_scrub_pause()
3563 		 */
3564 		scrub_pause_on(fs_info);
3565 		ret = btrfs_inc_block_group_ro(cache);
3566 		if (!ret && sctx->is_dev_replace) {
3567 			/*
3568 			 * If we are doing a device replace wait for any tasks
3569 			 * that started delalloc right before we set the block
3570 			 * group to RO mode, as they might have just allocated
3571 			 * an extent from it or decided they could do a nocow
3572 			 * write. And if any such tasks did that, wait for their
3573 			 * ordered extents to complete and then commit the
3574 			 * current transaction, so that we can later see the new
3575 			 * extent items in the extent tree - the ordered extents
3576 			 * create delayed data references (for cow writes) when
3577 			 * they complete, which will be run and insert the
3578 			 * corresponding extent items into the extent tree when
3579 			 * we commit the transaction they used when running
3580 			 * inode.c:btrfs_finish_ordered_io(). We later use
3581 			 * the commit root of the extent tree to find extents
3582 			 * to copy from the srcdev into the tgtdev, and we don't
3583 			 * want to miss any new extents.
3584 			 */
3585 			btrfs_wait_block_group_reservations(cache);
3586 			btrfs_wait_nocow_writers(cache);
3587 			ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3588 						       cache->key.objectid,
3589 						       cache->key.offset);
3590 			if (ret > 0) {
3591 				struct btrfs_trans_handle *trans;
3592 
3593 				trans = btrfs_join_transaction(root);
3594 				if (IS_ERR(trans))
3595 					ret = PTR_ERR(trans);
3596 				else
3597 					ret = btrfs_commit_transaction(trans);
3598 				if (ret) {
3599 					scrub_pause_off(fs_info);
3600 					btrfs_put_block_group(cache);
3601 					break;
3602 				}
3603 			}
3604 		}
3605 		scrub_pause_off(fs_info);
3606 
3607 		if (ret == 0) {
3608 			ro_set = 1;
3609 		} else if (ret == -ENOSPC) {
3610 			/*
3611 			 * btrfs_inc_block_group_ro return -ENOSPC when it
3612 			 * failed in creating new chunk for metadata.
3613 			 * It is not a problem for scrub/replace, because
3614 			 * metadata are always cowed, and our scrub paused
3615 			 * commit_transactions.
3616 			 */
3617 			ro_set = 0;
3618 		} else {
3619 			btrfs_warn(fs_info,
3620 				   "failed setting block group ro: %d", ret);
3621 			btrfs_put_block_group(cache);
3622 			break;
3623 		}
3624 
3625 		down_write(&fs_info->dev_replace.rwsem);
3626 		dev_replace->cursor_right = found_key.offset + length;
3627 		dev_replace->cursor_left = found_key.offset;
3628 		dev_replace->item_needs_writeback = 1;
3629 		up_write(&dev_replace->rwsem);
3630 
3631 		ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3632 				  found_key.offset, cache);
3633 
3634 		/*
3635 		 * flush, submit all pending read and write bios, afterwards
3636 		 * wait for them.
3637 		 * Note that in the dev replace case, a read request causes
3638 		 * write requests that are submitted in the read completion
3639 		 * worker. Therefore in the current situation, it is required
3640 		 * that all write requests are flushed, so that all read and
3641 		 * write requests are really completed when bios_in_flight
3642 		 * changes to 0.
3643 		 */
3644 		sctx->flush_all_writes = true;
3645 		scrub_submit(sctx);
3646 		mutex_lock(&sctx->wr_lock);
3647 		scrub_wr_submit(sctx);
3648 		mutex_unlock(&sctx->wr_lock);
3649 
3650 		wait_event(sctx->list_wait,
3651 			   atomic_read(&sctx->bios_in_flight) == 0);
3652 
3653 		scrub_pause_on(fs_info);
3654 
3655 		/*
3656 		 * must be called before we decrease @scrub_paused.
3657 		 * make sure we don't block transaction commit while
3658 		 * we are waiting pending workers finished.
3659 		 */
3660 		wait_event(sctx->list_wait,
3661 			   atomic_read(&sctx->workers_pending) == 0);
3662 		sctx->flush_all_writes = false;
3663 
3664 		scrub_pause_off(fs_info);
3665 
3666 		down_write(&fs_info->dev_replace.rwsem);
3667 		dev_replace->cursor_left = dev_replace->cursor_right;
3668 		dev_replace->item_needs_writeback = 1;
3669 		up_write(&fs_info->dev_replace.rwsem);
3670 
3671 		if (ro_set)
3672 			btrfs_dec_block_group_ro(cache);
3673 
3674 		/*
3675 		 * We might have prevented the cleaner kthread from deleting
3676 		 * this block group if it was already unused because we raced
3677 		 * and set it to RO mode first. So add it back to the unused
3678 		 * list, otherwise it might not ever be deleted unless a manual
3679 		 * balance is triggered or it becomes used and unused again.
3680 		 */
3681 		spin_lock(&cache->lock);
3682 		if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3683 		    btrfs_block_group_used(&cache->item) == 0) {
3684 			spin_unlock(&cache->lock);
3685 			btrfs_mark_bg_unused(cache);
3686 		} else {
3687 			spin_unlock(&cache->lock);
3688 		}
3689 
3690 		btrfs_put_block_group(cache);
3691 		if (ret)
3692 			break;
3693 		if (sctx->is_dev_replace &&
3694 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3695 			ret = -EIO;
3696 			break;
3697 		}
3698 		if (sctx->stat.malloc_errors > 0) {
3699 			ret = -ENOMEM;
3700 			break;
3701 		}
3702 skip:
3703 		key.offset = found_key.offset + length;
3704 		btrfs_release_path(path);
3705 	}
3706 
3707 	btrfs_free_path(path);
3708 
3709 	return ret;
3710 }
3711 
3712 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3713 					   struct btrfs_device *scrub_dev)
3714 {
3715 	int	i;
3716 	u64	bytenr;
3717 	u64	gen;
3718 	int	ret;
3719 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3720 
3721 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3722 		return -EIO;
3723 
3724 	/* Seed devices of a new filesystem has their own generation. */
3725 	if (scrub_dev->fs_devices != fs_info->fs_devices)
3726 		gen = scrub_dev->generation;
3727 	else
3728 		gen = fs_info->last_trans_committed;
3729 
3730 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3731 		bytenr = btrfs_sb_offset(i);
3732 		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3733 		    scrub_dev->commit_total_bytes)
3734 			break;
3735 
3736 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3737 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3738 				  NULL, 1, bytenr);
3739 		if (ret)
3740 			return ret;
3741 	}
3742 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3743 
3744 	return 0;
3745 }
3746 
3747 /*
3748  * get a reference count on fs_info->scrub_workers. start worker if necessary
3749  */
3750 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3751 						int is_dev_replace)
3752 {
3753 	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3754 	int max_active = fs_info->thread_pool_size;
3755 
3756 	lockdep_assert_held(&fs_info->scrub_lock);
3757 
3758 	if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3759 		ASSERT(fs_info->scrub_workers == NULL);
3760 		fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3761 				flags, is_dev_replace ? 1 : max_active, 4);
3762 		if (!fs_info->scrub_workers)
3763 			goto fail_scrub_workers;
3764 
3765 		ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3766 		fs_info->scrub_wr_completion_workers =
3767 			btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3768 					      max_active, 2);
3769 		if (!fs_info->scrub_wr_completion_workers)
3770 			goto fail_scrub_wr_completion_workers;
3771 
3772 		ASSERT(fs_info->scrub_parity_workers == NULL);
3773 		fs_info->scrub_parity_workers =
3774 			btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3775 					      max_active, 2);
3776 		if (!fs_info->scrub_parity_workers)
3777 			goto fail_scrub_parity_workers;
3778 
3779 		refcount_set(&fs_info->scrub_workers_refcnt, 1);
3780 	} else {
3781 		refcount_inc(&fs_info->scrub_workers_refcnt);
3782 	}
3783 	return 0;
3784 
3785 fail_scrub_parity_workers:
3786 	btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3787 fail_scrub_wr_completion_workers:
3788 	btrfs_destroy_workqueue(fs_info->scrub_workers);
3789 fail_scrub_workers:
3790 	return -ENOMEM;
3791 }
3792 
3793 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3794 		    u64 end, struct btrfs_scrub_progress *progress,
3795 		    int readonly, int is_dev_replace)
3796 {
3797 	struct scrub_ctx *sctx;
3798 	int ret;
3799 	struct btrfs_device *dev;
3800 	unsigned int nofs_flag;
3801 	struct btrfs_workqueue *scrub_workers = NULL;
3802 	struct btrfs_workqueue *scrub_wr_comp = NULL;
3803 	struct btrfs_workqueue *scrub_parity = NULL;
3804 
3805 	if (btrfs_fs_closing(fs_info))
3806 		return -EAGAIN;
3807 
3808 	if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3809 		/*
3810 		 * in this case scrub is unable to calculate the checksum
3811 		 * the way scrub is implemented. Do not handle this
3812 		 * situation at all because it won't ever happen.
3813 		 */
3814 		btrfs_err(fs_info,
3815 			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3816 		       fs_info->nodesize,
3817 		       BTRFS_STRIPE_LEN);
3818 		return -EINVAL;
3819 	}
3820 
3821 	if (fs_info->sectorsize != PAGE_SIZE) {
3822 		/* not supported for data w/o checksums */
3823 		btrfs_err_rl(fs_info,
3824 			   "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3825 		       fs_info->sectorsize, PAGE_SIZE);
3826 		return -EINVAL;
3827 	}
3828 
3829 	if (fs_info->nodesize >
3830 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3831 	    fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3832 		/*
3833 		 * would exhaust the array bounds of pagev member in
3834 		 * struct scrub_block
3835 		 */
3836 		btrfs_err(fs_info,
3837 			  "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3838 		       fs_info->nodesize,
3839 		       SCRUB_MAX_PAGES_PER_BLOCK,
3840 		       fs_info->sectorsize,
3841 		       SCRUB_MAX_PAGES_PER_BLOCK);
3842 		return -EINVAL;
3843 	}
3844 
3845 	/* Allocate outside of device_list_mutex */
3846 	sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3847 	if (IS_ERR(sctx))
3848 		return PTR_ERR(sctx);
3849 
3850 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3851 	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3852 	if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3853 		     !is_dev_replace)) {
3854 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3855 		ret = -ENODEV;
3856 		goto out_free_ctx;
3857 	}
3858 
3859 	if (!is_dev_replace && !readonly &&
3860 	    !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3861 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3862 		btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3863 				rcu_str_deref(dev->name));
3864 		ret = -EROFS;
3865 		goto out_free_ctx;
3866 	}
3867 
3868 	mutex_lock(&fs_info->scrub_lock);
3869 	if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3870 	    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3871 		mutex_unlock(&fs_info->scrub_lock);
3872 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3873 		ret = -EIO;
3874 		goto out_free_ctx;
3875 	}
3876 
3877 	down_read(&fs_info->dev_replace.rwsem);
3878 	if (dev->scrub_ctx ||
3879 	    (!is_dev_replace &&
3880 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3881 		up_read(&fs_info->dev_replace.rwsem);
3882 		mutex_unlock(&fs_info->scrub_lock);
3883 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3884 		ret = -EINPROGRESS;
3885 		goto out_free_ctx;
3886 	}
3887 	up_read(&fs_info->dev_replace.rwsem);
3888 
3889 	ret = scrub_workers_get(fs_info, is_dev_replace);
3890 	if (ret) {
3891 		mutex_unlock(&fs_info->scrub_lock);
3892 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3893 		goto out_free_ctx;
3894 	}
3895 
3896 	sctx->readonly = readonly;
3897 	dev->scrub_ctx = sctx;
3898 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3899 
3900 	/*
3901 	 * checking @scrub_pause_req here, we can avoid
3902 	 * race between committing transaction and scrubbing.
3903 	 */
3904 	__scrub_blocked_if_needed(fs_info);
3905 	atomic_inc(&fs_info->scrubs_running);
3906 	mutex_unlock(&fs_info->scrub_lock);
3907 
3908 	/*
3909 	 * In order to avoid deadlock with reclaim when there is a transaction
3910 	 * trying to pause scrub, make sure we use GFP_NOFS for all the
3911 	 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3912 	 * invoked by our callees. The pausing request is done when the
3913 	 * transaction commit starts, and it blocks the transaction until scrub
3914 	 * is paused (done at specific points at scrub_stripe() or right above
3915 	 * before incrementing fs_info->scrubs_running).
3916 	 */
3917 	nofs_flag = memalloc_nofs_save();
3918 	if (!is_dev_replace) {
3919 		btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3920 		/*
3921 		 * by holding device list mutex, we can
3922 		 * kick off writing super in log tree sync.
3923 		 */
3924 		mutex_lock(&fs_info->fs_devices->device_list_mutex);
3925 		ret = scrub_supers(sctx, dev);
3926 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3927 	}
3928 
3929 	if (!ret)
3930 		ret = scrub_enumerate_chunks(sctx, dev, start, end);
3931 	memalloc_nofs_restore(nofs_flag);
3932 
3933 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3934 	atomic_dec(&fs_info->scrubs_running);
3935 	wake_up(&fs_info->scrub_pause_wait);
3936 
3937 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3938 
3939 	if (progress)
3940 		memcpy(progress, &sctx->stat, sizeof(*progress));
3941 
3942 	if (!is_dev_replace)
3943 		btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3944 			ret ? "not finished" : "finished", devid, ret);
3945 
3946 	mutex_lock(&fs_info->scrub_lock);
3947 	dev->scrub_ctx = NULL;
3948 	if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3949 		scrub_workers = fs_info->scrub_workers;
3950 		scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3951 		scrub_parity = fs_info->scrub_parity_workers;
3952 
3953 		fs_info->scrub_workers = NULL;
3954 		fs_info->scrub_wr_completion_workers = NULL;
3955 		fs_info->scrub_parity_workers = NULL;
3956 	}
3957 	mutex_unlock(&fs_info->scrub_lock);
3958 
3959 	btrfs_destroy_workqueue(scrub_workers);
3960 	btrfs_destroy_workqueue(scrub_wr_comp);
3961 	btrfs_destroy_workqueue(scrub_parity);
3962 	scrub_put_ctx(sctx);
3963 
3964 	return ret;
3965 
3966 out_free_ctx:
3967 	scrub_free_ctx(sctx);
3968 
3969 	return ret;
3970 }
3971 
3972 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3973 {
3974 	mutex_lock(&fs_info->scrub_lock);
3975 	atomic_inc(&fs_info->scrub_pause_req);
3976 	while (atomic_read(&fs_info->scrubs_paused) !=
3977 	       atomic_read(&fs_info->scrubs_running)) {
3978 		mutex_unlock(&fs_info->scrub_lock);
3979 		wait_event(fs_info->scrub_pause_wait,
3980 			   atomic_read(&fs_info->scrubs_paused) ==
3981 			   atomic_read(&fs_info->scrubs_running));
3982 		mutex_lock(&fs_info->scrub_lock);
3983 	}
3984 	mutex_unlock(&fs_info->scrub_lock);
3985 }
3986 
3987 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3988 {
3989 	atomic_dec(&fs_info->scrub_pause_req);
3990 	wake_up(&fs_info->scrub_pause_wait);
3991 }
3992 
3993 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3994 {
3995 	mutex_lock(&fs_info->scrub_lock);
3996 	if (!atomic_read(&fs_info->scrubs_running)) {
3997 		mutex_unlock(&fs_info->scrub_lock);
3998 		return -ENOTCONN;
3999 	}
4000 
4001 	atomic_inc(&fs_info->scrub_cancel_req);
4002 	while (atomic_read(&fs_info->scrubs_running)) {
4003 		mutex_unlock(&fs_info->scrub_lock);
4004 		wait_event(fs_info->scrub_pause_wait,
4005 			   atomic_read(&fs_info->scrubs_running) == 0);
4006 		mutex_lock(&fs_info->scrub_lock);
4007 	}
4008 	atomic_dec(&fs_info->scrub_cancel_req);
4009 	mutex_unlock(&fs_info->scrub_lock);
4010 
4011 	return 0;
4012 }
4013 
4014 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4015 {
4016 	struct btrfs_fs_info *fs_info = dev->fs_info;
4017 	struct scrub_ctx *sctx;
4018 
4019 	mutex_lock(&fs_info->scrub_lock);
4020 	sctx = dev->scrub_ctx;
4021 	if (!sctx) {
4022 		mutex_unlock(&fs_info->scrub_lock);
4023 		return -ENOTCONN;
4024 	}
4025 	atomic_inc(&sctx->cancel_req);
4026 	while (dev->scrub_ctx) {
4027 		mutex_unlock(&fs_info->scrub_lock);
4028 		wait_event(fs_info->scrub_pause_wait,
4029 			   dev->scrub_ctx == NULL);
4030 		mutex_lock(&fs_info->scrub_lock);
4031 	}
4032 	mutex_unlock(&fs_info->scrub_lock);
4033 
4034 	return 0;
4035 }
4036 
4037 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4038 			 struct btrfs_scrub_progress *progress)
4039 {
4040 	struct btrfs_device *dev;
4041 	struct scrub_ctx *sctx = NULL;
4042 
4043 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4044 	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4045 	if (dev)
4046 		sctx = dev->scrub_ctx;
4047 	if (sctx)
4048 		memcpy(progress, &sctx->stat, sizeof(*progress));
4049 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4050 
4051 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4052 }
4053 
4054 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4055 			       u64 extent_logical, u64 extent_len,
4056 			       u64 *extent_physical,
4057 			       struct btrfs_device **extent_dev,
4058 			       int *extent_mirror_num)
4059 {
4060 	u64 mapped_length;
4061 	struct btrfs_bio *bbio = NULL;
4062 	int ret;
4063 
4064 	mapped_length = extent_len;
4065 	ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4066 			      &mapped_length, &bbio, 0);
4067 	if (ret || !bbio || mapped_length < extent_len ||
4068 	    !bbio->stripes[0].dev->bdev) {
4069 		btrfs_put_bbio(bbio);
4070 		return;
4071 	}
4072 
4073 	*extent_physical = bbio->stripes[0].physical;
4074 	*extent_mirror_num = bbio->mirror_num;
4075 	*extent_dev = bbio->stripes[0].dev;
4076 	btrfs_put_bbio(bbio);
4077 }
4078