xref: /linux/fs/btrfs/scrub.c (revision 962fad301c33dec69324dc2d9320fd84a119a24c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
4  */
5 
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
9 #include <crypto/hash.h>
10 #include "ctree.h"
11 #include "discard.h"
12 #include "volumes.h"
13 #include "disk-io.h"
14 #include "ordered-data.h"
15 #include "transaction.h"
16 #include "backref.h"
17 #include "extent_io.h"
18 #include "dev-replace.h"
19 #include "check-integrity.h"
20 #include "rcu-string.h"
21 #include "raid56.h"
22 #include "block-group.h"
23 
24 /*
25  * This is only the first step towards a full-features scrub. It reads all
26  * extent and super block and verifies the checksums. In case a bad checksum
27  * is found or the extent cannot be read, good data will be written back if
28  * any can be found.
29  *
30  * Future enhancements:
31  *  - In case an unrepairable extent is encountered, track which files are
32  *    affected and report them
33  *  - track and record media errors, throw out bad devices
34  *  - add a mode to also read unallocated space
35  */
36 
37 struct scrub_block;
38 struct scrub_ctx;
39 
40 /*
41  * the following three values only influence the performance.
42  * The last one configures the number of parallel and outstanding I/O
43  * operations. The first two values configure an upper limit for the number
44  * of (dynamically allocated) pages that are added to a bio.
45  */
46 #define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
47 #define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
48 #define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
49 
50 /*
51  * the following value times PAGE_SIZE needs to be large enough to match the
52  * largest node/leaf/sector size that shall be supported.
53  * Values larger than BTRFS_STRIPE_LEN are not supported.
54  */
55 #define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
56 
57 struct scrub_recover {
58 	refcount_t		refs;
59 	struct btrfs_bio	*bbio;
60 	u64			map_length;
61 };
62 
63 struct scrub_page {
64 	struct scrub_block	*sblock;
65 	struct page		*page;
66 	struct btrfs_device	*dev;
67 	struct list_head	list;
68 	u64			flags;  /* extent flags */
69 	u64			generation;
70 	u64			logical;
71 	u64			physical;
72 	u64			physical_for_dev_replace;
73 	atomic_t		refs;
74 	struct {
75 		unsigned int	mirror_num:8;
76 		unsigned int	have_csum:1;
77 		unsigned int	io_error:1;
78 	};
79 	u8			csum[BTRFS_CSUM_SIZE];
80 
81 	struct scrub_recover	*recover;
82 };
83 
84 struct scrub_bio {
85 	int			index;
86 	struct scrub_ctx	*sctx;
87 	struct btrfs_device	*dev;
88 	struct bio		*bio;
89 	blk_status_t		status;
90 	u64			logical;
91 	u64			physical;
92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
94 #else
95 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
96 #endif
97 	int			page_count;
98 	int			next_free;
99 	struct btrfs_work	work;
100 };
101 
102 struct scrub_block {
103 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 	int			page_count;
105 	atomic_t		outstanding_pages;
106 	refcount_t		refs; /* free mem on transition to zero */
107 	struct scrub_ctx	*sctx;
108 	struct scrub_parity	*sparity;
109 	struct {
110 		unsigned int	header_error:1;
111 		unsigned int	checksum_error:1;
112 		unsigned int	no_io_error_seen:1;
113 		unsigned int	generation_error:1; /* also sets header_error */
114 
115 		/* The following is for the data used to check parity */
116 		/* It is for the data with checksum */
117 		unsigned int	data_corrected:1;
118 	};
119 	struct btrfs_work	work;
120 };
121 
122 /* Used for the chunks with parity stripe such RAID5/6 */
123 struct scrub_parity {
124 	struct scrub_ctx	*sctx;
125 
126 	struct btrfs_device	*scrub_dev;
127 
128 	u64			logic_start;
129 
130 	u64			logic_end;
131 
132 	int			nsectors;
133 
134 	u64			stripe_len;
135 
136 	refcount_t		refs;
137 
138 	struct list_head	spages;
139 
140 	/* Work of parity check and repair */
141 	struct btrfs_work	work;
142 
143 	/* Mark the parity blocks which have data */
144 	unsigned long		*dbitmap;
145 
146 	/*
147 	 * Mark the parity blocks which have data, but errors happen when
148 	 * read data or check data
149 	 */
150 	unsigned long		*ebitmap;
151 
152 	unsigned long		bitmap[];
153 };
154 
155 struct scrub_ctx {
156 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
157 	struct btrfs_fs_info	*fs_info;
158 	int			first_free;
159 	int			curr;
160 	atomic_t		bios_in_flight;
161 	atomic_t		workers_pending;
162 	spinlock_t		list_lock;
163 	wait_queue_head_t	list_wait;
164 	u16			csum_size;
165 	struct list_head	csum_list;
166 	atomic_t		cancel_req;
167 	int			readonly;
168 	int			pages_per_rd_bio;
169 
170 	int			is_dev_replace;
171 
172 	struct scrub_bio        *wr_curr_bio;
173 	struct mutex            wr_lock;
174 	int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
175 	struct btrfs_device     *wr_tgtdev;
176 	bool                    flush_all_writes;
177 
178 	/*
179 	 * statistics
180 	 */
181 	struct btrfs_scrub_progress stat;
182 	spinlock_t		stat_lock;
183 
184 	/*
185 	 * Use a ref counter to avoid use-after-free issues. Scrub workers
186 	 * decrement bios_in_flight and workers_pending and then do a wakeup
187 	 * on the list_wait wait queue. We must ensure the main scrub task
188 	 * doesn't free the scrub context before or while the workers are
189 	 * doing the wakeup() call.
190 	 */
191 	refcount_t              refs;
192 };
193 
194 struct scrub_warning {
195 	struct btrfs_path	*path;
196 	u64			extent_item_size;
197 	const char		*errstr;
198 	u64			physical;
199 	u64			logical;
200 	struct btrfs_device	*dev;
201 };
202 
203 struct full_stripe_lock {
204 	struct rb_node node;
205 	u64 logical;
206 	u64 refs;
207 	struct mutex mutex;
208 };
209 
210 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
211 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
212 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
213 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
214 				     struct scrub_block *sblocks_for_recheck);
215 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
216 				struct scrub_block *sblock,
217 				int retry_failed_mirror);
218 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
219 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
220 					     struct scrub_block *sblock_good);
221 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
222 					    struct scrub_block *sblock_good,
223 					    int page_num, int force_write);
224 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
225 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
226 					   int page_num);
227 static int scrub_checksum_data(struct scrub_block *sblock);
228 static int scrub_checksum_tree_block(struct scrub_block *sblock);
229 static int scrub_checksum_super(struct scrub_block *sblock);
230 static void scrub_block_get(struct scrub_block *sblock);
231 static void scrub_block_put(struct scrub_block *sblock);
232 static void scrub_page_get(struct scrub_page *spage);
233 static void scrub_page_put(struct scrub_page *spage);
234 static void scrub_parity_get(struct scrub_parity *sparity);
235 static void scrub_parity_put(struct scrub_parity *sparity);
236 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
237 				    struct scrub_page *spage);
238 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
239 		       u64 physical, struct btrfs_device *dev, u64 flags,
240 		       u64 gen, int mirror_num, u8 *csum, int force,
241 		       u64 physical_for_dev_replace);
242 static void scrub_bio_end_io(struct bio *bio);
243 static void scrub_bio_end_io_worker(struct btrfs_work *work);
244 static void scrub_block_complete(struct scrub_block *sblock);
245 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
246 			       u64 extent_logical, u64 extent_len,
247 			       u64 *extent_physical,
248 			       struct btrfs_device **extent_dev,
249 			       int *extent_mirror_num);
250 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
251 				    struct scrub_page *spage);
252 static void scrub_wr_submit(struct scrub_ctx *sctx);
253 static void scrub_wr_bio_end_io(struct bio *bio);
254 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
255 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
256 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
257 static void scrub_put_ctx(struct scrub_ctx *sctx);
258 
259 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
260 {
261 	return page->recover &&
262 	       (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
263 }
264 
265 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
266 {
267 	refcount_inc(&sctx->refs);
268 	atomic_inc(&sctx->bios_in_flight);
269 }
270 
271 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
272 {
273 	atomic_dec(&sctx->bios_in_flight);
274 	wake_up(&sctx->list_wait);
275 	scrub_put_ctx(sctx);
276 }
277 
278 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
279 {
280 	while (atomic_read(&fs_info->scrub_pause_req)) {
281 		mutex_unlock(&fs_info->scrub_lock);
282 		wait_event(fs_info->scrub_pause_wait,
283 		   atomic_read(&fs_info->scrub_pause_req) == 0);
284 		mutex_lock(&fs_info->scrub_lock);
285 	}
286 }
287 
288 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
289 {
290 	atomic_inc(&fs_info->scrubs_paused);
291 	wake_up(&fs_info->scrub_pause_wait);
292 }
293 
294 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
295 {
296 	mutex_lock(&fs_info->scrub_lock);
297 	__scrub_blocked_if_needed(fs_info);
298 	atomic_dec(&fs_info->scrubs_paused);
299 	mutex_unlock(&fs_info->scrub_lock);
300 
301 	wake_up(&fs_info->scrub_pause_wait);
302 }
303 
304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
305 {
306 	scrub_pause_on(fs_info);
307 	scrub_pause_off(fs_info);
308 }
309 
310 /*
311  * Insert new full stripe lock into full stripe locks tree
312  *
313  * Return pointer to existing or newly inserted full_stripe_lock structure if
314  * everything works well.
315  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
316  *
317  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
318  * function
319  */
320 static struct full_stripe_lock *insert_full_stripe_lock(
321 		struct btrfs_full_stripe_locks_tree *locks_root,
322 		u64 fstripe_logical)
323 {
324 	struct rb_node **p;
325 	struct rb_node *parent = NULL;
326 	struct full_stripe_lock *entry;
327 	struct full_stripe_lock *ret;
328 
329 	lockdep_assert_held(&locks_root->lock);
330 
331 	p = &locks_root->root.rb_node;
332 	while (*p) {
333 		parent = *p;
334 		entry = rb_entry(parent, struct full_stripe_lock, node);
335 		if (fstripe_logical < entry->logical) {
336 			p = &(*p)->rb_left;
337 		} else if (fstripe_logical > entry->logical) {
338 			p = &(*p)->rb_right;
339 		} else {
340 			entry->refs++;
341 			return entry;
342 		}
343 	}
344 
345 	/*
346 	 * Insert new lock.
347 	 */
348 	ret = kmalloc(sizeof(*ret), GFP_KERNEL);
349 	if (!ret)
350 		return ERR_PTR(-ENOMEM);
351 	ret->logical = fstripe_logical;
352 	ret->refs = 1;
353 	mutex_init(&ret->mutex);
354 
355 	rb_link_node(&ret->node, parent, p);
356 	rb_insert_color(&ret->node, &locks_root->root);
357 	return ret;
358 }
359 
360 /*
361  * Search for a full stripe lock of a block group
362  *
363  * Return pointer to existing full stripe lock if found
364  * Return NULL if not found
365  */
366 static struct full_stripe_lock *search_full_stripe_lock(
367 		struct btrfs_full_stripe_locks_tree *locks_root,
368 		u64 fstripe_logical)
369 {
370 	struct rb_node *node;
371 	struct full_stripe_lock *entry;
372 
373 	lockdep_assert_held(&locks_root->lock);
374 
375 	node = locks_root->root.rb_node;
376 	while (node) {
377 		entry = rb_entry(node, struct full_stripe_lock, node);
378 		if (fstripe_logical < entry->logical)
379 			node = node->rb_left;
380 		else if (fstripe_logical > entry->logical)
381 			node = node->rb_right;
382 		else
383 			return entry;
384 	}
385 	return NULL;
386 }
387 
388 /*
389  * Helper to get full stripe logical from a normal bytenr.
390  *
391  * Caller must ensure @cache is a RAID56 block group.
392  */
393 static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
394 {
395 	u64 ret;
396 
397 	/*
398 	 * Due to chunk item size limit, full stripe length should not be
399 	 * larger than U32_MAX. Just a sanity check here.
400 	 */
401 	WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
402 
403 	/*
404 	 * round_down() can only handle power of 2, while RAID56 full
405 	 * stripe length can be 64KiB * n, so we need to manually round down.
406 	 */
407 	ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
408 			cache->full_stripe_len + cache->start;
409 	return ret;
410 }
411 
412 /*
413  * Lock a full stripe to avoid concurrency of recovery and read
414  *
415  * It's only used for profiles with parities (RAID5/6), for other profiles it
416  * does nothing.
417  *
418  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
419  * So caller must call unlock_full_stripe() at the same context.
420  *
421  * Return <0 if encounters error.
422  */
423 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
424 			    bool *locked_ret)
425 {
426 	struct btrfs_block_group *bg_cache;
427 	struct btrfs_full_stripe_locks_tree *locks_root;
428 	struct full_stripe_lock *existing;
429 	u64 fstripe_start;
430 	int ret = 0;
431 
432 	*locked_ret = false;
433 	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
434 	if (!bg_cache) {
435 		ASSERT(0);
436 		return -ENOENT;
437 	}
438 
439 	/* Profiles not based on parity don't need full stripe lock */
440 	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
441 		goto out;
442 	locks_root = &bg_cache->full_stripe_locks_root;
443 
444 	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
445 
446 	/* Now insert the full stripe lock */
447 	mutex_lock(&locks_root->lock);
448 	existing = insert_full_stripe_lock(locks_root, fstripe_start);
449 	mutex_unlock(&locks_root->lock);
450 	if (IS_ERR(existing)) {
451 		ret = PTR_ERR(existing);
452 		goto out;
453 	}
454 	mutex_lock(&existing->mutex);
455 	*locked_ret = true;
456 out:
457 	btrfs_put_block_group(bg_cache);
458 	return ret;
459 }
460 
461 /*
462  * Unlock a full stripe.
463  *
464  * NOTE: Caller must ensure it's the same context calling corresponding
465  * lock_full_stripe().
466  *
467  * Return 0 if we unlock full stripe without problem.
468  * Return <0 for error
469  */
470 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
471 			      bool locked)
472 {
473 	struct btrfs_block_group *bg_cache;
474 	struct btrfs_full_stripe_locks_tree *locks_root;
475 	struct full_stripe_lock *fstripe_lock;
476 	u64 fstripe_start;
477 	bool freeit = false;
478 	int ret = 0;
479 
480 	/* If we didn't acquire full stripe lock, no need to continue */
481 	if (!locked)
482 		return 0;
483 
484 	bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
485 	if (!bg_cache) {
486 		ASSERT(0);
487 		return -ENOENT;
488 	}
489 	if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
490 		goto out;
491 
492 	locks_root = &bg_cache->full_stripe_locks_root;
493 	fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
494 
495 	mutex_lock(&locks_root->lock);
496 	fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
497 	/* Unpaired unlock_full_stripe() detected */
498 	if (!fstripe_lock) {
499 		WARN_ON(1);
500 		ret = -ENOENT;
501 		mutex_unlock(&locks_root->lock);
502 		goto out;
503 	}
504 
505 	if (fstripe_lock->refs == 0) {
506 		WARN_ON(1);
507 		btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
508 			fstripe_lock->logical);
509 	} else {
510 		fstripe_lock->refs--;
511 	}
512 
513 	if (fstripe_lock->refs == 0) {
514 		rb_erase(&fstripe_lock->node, &locks_root->root);
515 		freeit = true;
516 	}
517 	mutex_unlock(&locks_root->lock);
518 
519 	mutex_unlock(&fstripe_lock->mutex);
520 	if (freeit)
521 		kfree(fstripe_lock);
522 out:
523 	btrfs_put_block_group(bg_cache);
524 	return ret;
525 }
526 
527 static void scrub_free_csums(struct scrub_ctx *sctx)
528 {
529 	while (!list_empty(&sctx->csum_list)) {
530 		struct btrfs_ordered_sum *sum;
531 		sum = list_first_entry(&sctx->csum_list,
532 				       struct btrfs_ordered_sum, list);
533 		list_del(&sum->list);
534 		kfree(sum);
535 	}
536 }
537 
538 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
539 {
540 	int i;
541 
542 	if (!sctx)
543 		return;
544 
545 	/* this can happen when scrub is cancelled */
546 	if (sctx->curr != -1) {
547 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
548 
549 		for (i = 0; i < sbio->page_count; i++) {
550 			WARN_ON(!sbio->pagev[i]->page);
551 			scrub_block_put(sbio->pagev[i]->sblock);
552 		}
553 		bio_put(sbio->bio);
554 	}
555 
556 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
557 		struct scrub_bio *sbio = sctx->bios[i];
558 
559 		if (!sbio)
560 			break;
561 		kfree(sbio);
562 	}
563 
564 	kfree(sctx->wr_curr_bio);
565 	scrub_free_csums(sctx);
566 	kfree(sctx);
567 }
568 
569 static void scrub_put_ctx(struct scrub_ctx *sctx)
570 {
571 	if (refcount_dec_and_test(&sctx->refs))
572 		scrub_free_ctx(sctx);
573 }
574 
575 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
576 		struct btrfs_fs_info *fs_info, int is_dev_replace)
577 {
578 	struct scrub_ctx *sctx;
579 	int		i;
580 
581 	sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
582 	if (!sctx)
583 		goto nomem;
584 	refcount_set(&sctx->refs, 1);
585 	sctx->is_dev_replace = is_dev_replace;
586 	sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
587 	sctx->curr = -1;
588 	sctx->fs_info = fs_info;
589 	INIT_LIST_HEAD(&sctx->csum_list);
590 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
591 		struct scrub_bio *sbio;
592 
593 		sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
594 		if (!sbio)
595 			goto nomem;
596 		sctx->bios[i] = sbio;
597 
598 		sbio->index = i;
599 		sbio->sctx = sctx;
600 		sbio->page_count = 0;
601 		btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
602 				NULL);
603 
604 		if (i != SCRUB_BIOS_PER_SCTX - 1)
605 			sctx->bios[i]->next_free = i + 1;
606 		else
607 			sctx->bios[i]->next_free = -1;
608 	}
609 	sctx->first_free = 0;
610 	atomic_set(&sctx->bios_in_flight, 0);
611 	atomic_set(&sctx->workers_pending, 0);
612 	atomic_set(&sctx->cancel_req, 0);
613 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
614 
615 	spin_lock_init(&sctx->list_lock);
616 	spin_lock_init(&sctx->stat_lock);
617 	init_waitqueue_head(&sctx->list_wait);
618 
619 	WARN_ON(sctx->wr_curr_bio != NULL);
620 	mutex_init(&sctx->wr_lock);
621 	sctx->wr_curr_bio = NULL;
622 	if (is_dev_replace) {
623 		WARN_ON(!fs_info->dev_replace.tgtdev);
624 		sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
625 		sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
626 		sctx->flush_all_writes = false;
627 	}
628 
629 	return sctx;
630 
631 nomem:
632 	scrub_free_ctx(sctx);
633 	return ERR_PTR(-ENOMEM);
634 }
635 
636 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
637 				     void *warn_ctx)
638 {
639 	u64 isize;
640 	u32 nlink;
641 	int ret;
642 	int i;
643 	unsigned nofs_flag;
644 	struct extent_buffer *eb;
645 	struct btrfs_inode_item *inode_item;
646 	struct scrub_warning *swarn = warn_ctx;
647 	struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
648 	struct inode_fs_paths *ipath = NULL;
649 	struct btrfs_root *local_root;
650 	struct btrfs_key key;
651 
652 	local_root = btrfs_get_fs_root(fs_info, root, true);
653 	if (IS_ERR(local_root)) {
654 		ret = PTR_ERR(local_root);
655 		goto err;
656 	}
657 
658 	/*
659 	 * this makes the path point to (inum INODE_ITEM ioff)
660 	 */
661 	key.objectid = inum;
662 	key.type = BTRFS_INODE_ITEM_KEY;
663 	key.offset = 0;
664 
665 	ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
666 	if (ret) {
667 		btrfs_put_root(local_root);
668 		btrfs_release_path(swarn->path);
669 		goto err;
670 	}
671 
672 	eb = swarn->path->nodes[0];
673 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
674 					struct btrfs_inode_item);
675 	isize = btrfs_inode_size(eb, inode_item);
676 	nlink = btrfs_inode_nlink(eb, inode_item);
677 	btrfs_release_path(swarn->path);
678 
679 	/*
680 	 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
681 	 * uses GFP_NOFS in this context, so we keep it consistent but it does
682 	 * not seem to be strictly necessary.
683 	 */
684 	nofs_flag = memalloc_nofs_save();
685 	ipath = init_ipath(4096, local_root, swarn->path);
686 	memalloc_nofs_restore(nofs_flag);
687 	if (IS_ERR(ipath)) {
688 		btrfs_put_root(local_root);
689 		ret = PTR_ERR(ipath);
690 		ipath = NULL;
691 		goto err;
692 	}
693 	ret = paths_from_inode(inum, ipath);
694 
695 	if (ret < 0)
696 		goto err;
697 
698 	/*
699 	 * we deliberately ignore the bit ipath might have been too small to
700 	 * hold all of the paths here
701 	 */
702 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
703 		btrfs_warn_in_rcu(fs_info,
704 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
705 				  swarn->errstr, swarn->logical,
706 				  rcu_str_deref(swarn->dev->name),
707 				  swarn->physical,
708 				  root, inum, offset,
709 				  min(isize - offset, (u64)PAGE_SIZE), nlink,
710 				  (char *)(unsigned long)ipath->fspath->val[i]);
711 
712 	btrfs_put_root(local_root);
713 	free_ipath(ipath);
714 	return 0;
715 
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 		clear_page(page_address(spage->page));
1621 
1622 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1623 }
1624 
1625 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1626 				    struct scrub_page *spage)
1627 {
1628 	struct scrub_bio *sbio;
1629 	int ret;
1630 
1631 	mutex_lock(&sctx->wr_lock);
1632 again:
1633 	if (!sctx->wr_curr_bio) {
1634 		sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1635 					      GFP_KERNEL);
1636 		if (!sctx->wr_curr_bio) {
1637 			mutex_unlock(&sctx->wr_lock);
1638 			return -ENOMEM;
1639 		}
1640 		sctx->wr_curr_bio->sctx = sctx;
1641 		sctx->wr_curr_bio->page_count = 0;
1642 	}
1643 	sbio = sctx->wr_curr_bio;
1644 	if (sbio->page_count == 0) {
1645 		struct bio *bio;
1646 
1647 		sbio->physical = spage->physical_for_dev_replace;
1648 		sbio->logical = spage->logical;
1649 		sbio->dev = sctx->wr_tgtdev;
1650 		bio = sbio->bio;
1651 		if (!bio) {
1652 			bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1653 			sbio->bio = bio;
1654 		}
1655 
1656 		bio->bi_private = sbio;
1657 		bio->bi_end_io = scrub_wr_bio_end_io;
1658 		bio_set_dev(bio, sbio->dev->bdev);
1659 		bio->bi_iter.bi_sector = sbio->physical >> 9;
1660 		bio->bi_opf = REQ_OP_WRITE;
1661 		sbio->status = 0;
1662 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1663 		   spage->physical_for_dev_replace ||
1664 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1665 		   spage->logical) {
1666 		scrub_wr_submit(sctx);
1667 		goto again;
1668 	}
1669 
1670 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1671 	if (ret != PAGE_SIZE) {
1672 		if (sbio->page_count < 1) {
1673 			bio_put(sbio->bio);
1674 			sbio->bio = NULL;
1675 			mutex_unlock(&sctx->wr_lock);
1676 			return -EIO;
1677 		}
1678 		scrub_wr_submit(sctx);
1679 		goto again;
1680 	}
1681 
1682 	sbio->pagev[sbio->page_count] = spage;
1683 	scrub_page_get(spage);
1684 	sbio->page_count++;
1685 	if (sbio->page_count == sctx->pages_per_wr_bio)
1686 		scrub_wr_submit(sctx);
1687 	mutex_unlock(&sctx->wr_lock);
1688 
1689 	return 0;
1690 }
1691 
1692 static void scrub_wr_submit(struct scrub_ctx *sctx)
1693 {
1694 	struct scrub_bio *sbio;
1695 
1696 	if (!sctx->wr_curr_bio)
1697 		return;
1698 
1699 	sbio = sctx->wr_curr_bio;
1700 	sctx->wr_curr_bio = NULL;
1701 	WARN_ON(!sbio->bio->bi_disk);
1702 	scrub_pending_bio_inc(sctx);
1703 	/* process all writes in a single worker thread. Then the block layer
1704 	 * orders the requests before sending them to the driver which
1705 	 * doubled the write performance on spinning disks when measured
1706 	 * with Linux 3.5 */
1707 	btrfsic_submit_bio(sbio->bio);
1708 }
1709 
1710 static void scrub_wr_bio_end_io(struct bio *bio)
1711 {
1712 	struct scrub_bio *sbio = bio->bi_private;
1713 	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1714 
1715 	sbio->status = bio->bi_status;
1716 	sbio->bio = bio;
1717 
1718 	btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1719 	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1720 }
1721 
1722 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1723 {
1724 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1725 	struct scrub_ctx *sctx = sbio->sctx;
1726 	int i;
1727 
1728 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1729 	if (sbio->status) {
1730 		struct btrfs_dev_replace *dev_replace =
1731 			&sbio->sctx->fs_info->dev_replace;
1732 
1733 		for (i = 0; i < sbio->page_count; i++) {
1734 			struct scrub_page *spage = sbio->pagev[i];
1735 
1736 			spage->io_error = 1;
1737 			atomic64_inc(&dev_replace->num_write_errors);
1738 		}
1739 	}
1740 
1741 	for (i = 0; i < sbio->page_count; i++)
1742 		scrub_page_put(sbio->pagev[i]);
1743 
1744 	bio_put(sbio->bio);
1745 	kfree(sbio);
1746 	scrub_pending_bio_dec(sctx);
1747 }
1748 
1749 static int scrub_checksum(struct scrub_block *sblock)
1750 {
1751 	u64 flags;
1752 	int ret;
1753 
1754 	/*
1755 	 * No need to initialize these stats currently,
1756 	 * because this function only use return value
1757 	 * instead of these stats value.
1758 	 *
1759 	 * Todo:
1760 	 * always use stats
1761 	 */
1762 	sblock->header_error = 0;
1763 	sblock->generation_error = 0;
1764 	sblock->checksum_error = 0;
1765 
1766 	WARN_ON(sblock->page_count < 1);
1767 	flags = sblock->pagev[0]->flags;
1768 	ret = 0;
1769 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1770 		ret = scrub_checksum_data(sblock);
1771 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1772 		ret = scrub_checksum_tree_block(sblock);
1773 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1774 		(void)scrub_checksum_super(sblock);
1775 	else
1776 		WARN_ON(1);
1777 	if (ret)
1778 		scrub_handle_errored_block(sblock);
1779 
1780 	return ret;
1781 }
1782 
1783 static int scrub_checksum_data(struct scrub_block *sblock)
1784 {
1785 	struct scrub_ctx *sctx = sblock->sctx;
1786 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1787 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1788 	u8 csum[BTRFS_CSUM_SIZE];
1789 	struct scrub_page *spage;
1790 	char *kaddr;
1791 
1792 	BUG_ON(sblock->page_count < 1);
1793 	spage = sblock->pagev[0];
1794 	if (!spage->have_csum)
1795 		return 0;
1796 
1797 	kaddr = page_address(spage->page);
1798 
1799 	shash->tfm = fs_info->csum_shash;
1800 	crypto_shash_init(shash);
1801 	crypto_shash_digest(shash, kaddr, PAGE_SIZE, csum);
1802 
1803 	if (memcmp(csum, spage->csum, sctx->csum_size))
1804 		sblock->checksum_error = 1;
1805 
1806 	return sblock->checksum_error;
1807 }
1808 
1809 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1810 {
1811 	struct scrub_ctx *sctx = sblock->sctx;
1812 	struct btrfs_header *h;
1813 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1814 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1815 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1816 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1817 	const int num_pages = sctx->fs_info->nodesize >> PAGE_SHIFT;
1818 	int i;
1819 	struct scrub_page *spage;
1820 	char *kaddr;
1821 
1822 	BUG_ON(sblock->page_count < 1);
1823 	spage = sblock->pagev[0];
1824 	kaddr = page_address(spage->page);
1825 	h = (struct btrfs_header *)kaddr;
1826 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1827 
1828 	/*
1829 	 * we don't use the getter functions here, as we
1830 	 * a) don't have an extent buffer and
1831 	 * b) the page is already kmapped
1832 	 */
1833 	if (spage->logical != btrfs_stack_header_bytenr(h))
1834 		sblock->header_error = 1;
1835 
1836 	if (spage->generation != btrfs_stack_header_generation(h)) {
1837 		sblock->header_error = 1;
1838 		sblock->generation_error = 1;
1839 	}
1840 
1841 	if (!scrub_check_fsid(h->fsid, spage))
1842 		sblock->header_error = 1;
1843 
1844 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1845 		   BTRFS_UUID_SIZE))
1846 		sblock->header_error = 1;
1847 
1848 	shash->tfm = fs_info->csum_shash;
1849 	crypto_shash_init(shash);
1850 	crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
1851 			    PAGE_SIZE - BTRFS_CSUM_SIZE);
1852 
1853 	for (i = 1; i < num_pages; i++) {
1854 		kaddr = page_address(sblock->pagev[i]->page);
1855 		crypto_shash_update(shash, kaddr, PAGE_SIZE);
1856 	}
1857 
1858 	crypto_shash_final(shash, calculated_csum);
1859 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1860 		sblock->checksum_error = 1;
1861 
1862 	return sblock->header_error || sblock->checksum_error;
1863 }
1864 
1865 static int scrub_checksum_super(struct scrub_block *sblock)
1866 {
1867 	struct btrfs_super_block *s;
1868 	struct scrub_ctx *sctx = sblock->sctx;
1869 	struct btrfs_fs_info *fs_info = sctx->fs_info;
1870 	SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1871 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1872 	struct scrub_page *spage;
1873 	char *kaddr;
1874 	int fail_gen = 0;
1875 	int fail_cor = 0;
1876 
1877 	BUG_ON(sblock->page_count < 1);
1878 	spage = sblock->pagev[0];
1879 	kaddr = page_address(spage->page);
1880 	s = (struct btrfs_super_block *)kaddr;
1881 
1882 	if (spage->logical != btrfs_super_bytenr(s))
1883 		++fail_cor;
1884 
1885 	if (spage->generation != btrfs_super_generation(s))
1886 		++fail_gen;
1887 
1888 	if (!scrub_check_fsid(s->fsid, spage))
1889 		++fail_cor;
1890 
1891 	shash->tfm = fs_info->csum_shash;
1892 	crypto_shash_init(shash);
1893 	crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
1894 			BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
1895 
1896 	if (memcmp(calculated_csum, s->csum, sctx->csum_size))
1897 		++fail_cor;
1898 
1899 	if (fail_cor + fail_gen) {
1900 		/*
1901 		 * if we find an error in a super block, we just report it.
1902 		 * They will get written with the next transaction commit
1903 		 * anyway
1904 		 */
1905 		spin_lock(&sctx->stat_lock);
1906 		++sctx->stat.super_errors;
1907 		spin_unlock(&sctx->stat_lock);
1908 		if (fail_cor)
1909 			btrfs_dev_stat_inc_and_print(spage->dev,
1910 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1911 		else
1912 			btrfs_dev_stat_inc_and_print(spage->dev,
1913 				BTRFS_DEV_STAT_GENERATION_ERRS);
1914 	}
1915 
1916 	return fail_cor + fail_gen;
1917 }
1918 
1919 static void scrub_block_get(struct scrub_block *sblock)
1920 {
1921 	refcount_inc(&sblock->refs);
1922 }
1923 
1924 static void scrub_block_put(struct scrub_block *sblock)
1925 {
1926 	if (refcount_dec_and_test(&sblock->refs)) {
1927 		int i;
1928 
1929 		if (sblock->sparity)
1930 			scrub_parity_put(sblock->sparity);
1931 
1932 		for (i = 0; i < sblock->page_count; i++)
1933 			scrub_page_put(sblock->pagev[i]);
1934 		kfree(sblock);
1935 	}
1936 }
1937 
1938 static void scrub_page_get(struct scrub_page *spage)
1939 {
1940 	atomic_inc(&spage->refs);
1941 }
1942 
1943 static void scrub_page_put(struct scrub_page *spage)
1944 {
1945 	if (atomic_dec_and_test(&spage->refs)) {
1946 		if (spage->page)
1947 			__free_page(spage->page);
1948 		kfree(spage);
1949 	}
1950 }
1951 
1952 static void scrub_submit(struct scrub_ctx *sctx)
1953 {
1954 	struct scrub_bio *sbio;
1955 
1956 	if (sctx->curr == -1)
1957 		return;
1958 
1959 	sbio = sctx->bios[sctx->curr];
1960 	sctx->curr = -1;
1961 	scrub_pending_bio_inc(sctx);
1962 	btrfsic_submit_bio(sbio->bio);
1963 }
1964 
1965 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1966 				    struct scrub_page *spage)
1967 {
1968 	struct scrub_block *sblock = spage->sblock;
1969 	struct scrub_bio *sbio;
1970 	int ret;
1971 
1972 again:
1973 	/*
1974 	 * grab a fresh bio or wait for one to become available
1975 	 */
1976 	while (sctx->curr == -1) {
1977 		spin_lock(&sctx->list_lock);
1978 		sctx->curr = sctx->first_free;
1979 		if (sctx->curr != -1) {
1980 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1981 			sctx->bios[sctx->curr]->next_free = -1;
1982 			sctx->bios[sctx->curr]->page_count = 0;
1983 			spin_unlock(&sctx->list_lock);
1984 		} else {
1985 			spin_unlock(&sctx->list_lock);
1986 			wait_event(sctx->list_wait, sctx->first_free != -1);
1987 		}
1988 	}
1989 	sbio = sctx->bios[sctx->curr];
1990 	if (sbio->page_count == 0) {
1991 		struct bio *bio;
1992 
1993 		sbio->physical = spage->physical;
1994 		sbio->logical = spage->logical;
1995 		sbio->dev = spage->dev;
1996 		bio = sbio->bio;
1997 		if (!bio) {
1998 			bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
1999 			sbio->bio = bio;
2000 		}
2001 
2002 		bio->bi_private = sbio;
2003 		bio->bi_end_io = scrub_bio_end_io;
2004 		bio_set_dev(bio, sbio->dev->bdev);
2005 		bio->bi_iter.bi_sector = sbio->physical >> 9;
2006 		bio->bi_opf = REQ_OP_READ;
2007 		sbio->status = 0;
2008 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2009 		   spage->physical ||
2010 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
2011 		   spage->logical ||
2012 		   sbio->dev != spage->dev) {
2013 		scrub_submit(sctx);
2014 		goto again;
2015 	}
2016 
2017 	sbio->pagev[sbio->page_count] = spage;
2018 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2019 	if (ret != PAGE_SIZE) {
2020 		if (sbio->page_count < 1) {
2021 			bio_put(sbio->bio);
2022 			sbio->bio = NULL;
2023 			return -EIO;
2024 		}
2025 		scrub_submit(sctx);
2026 		goto again;
2027 	}
2028 
2029 	scrub_block_get(sblock); /* one for the page added to the bio */
2030 	atomic_inc(&sblock->outstanding_pages);
2031 	sbio->page_count++;
2032 	if (sbio->page_count == sctx->pages_per_rd_bio)
2033 		scrub_submit(sctx);
2034 
2035 	return 0;
2036 }
2037 
2038 static void scrub_missing_raid56_end_io(struct bio *bio)
2039 {
2040 	struct scrub_block *sblock = bio->bi_private;
2041 	struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2042 
2043 	if (bio->bi_status)
2044 		sblock->no_io_error_seen = 0;
2045 
2046 	bio_put(bio);
2047 
2048 	btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2049 }
2050 
2051 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2052 {
2053 	struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2054 	struct scrub_ctx *sctx = sblock->sctx;
2055 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2056 	u64 logical;
2057 	struct btrfs_device *dev;
2058 
2059 	logical = sblock->pagev[0]->logical;
2060 	dev = sblock->pagev[0]->dev;
2061 
2062 	if (sblock->no_io_error_seen)
2063 		scrub_recheck_block_checksum(sblock);
2064 
2065 	if (!sblock->no_io_error_seen) {
2066 		spin_lock(&sctx->stat_lock);
2067 		sctx->stat.read_errors++;
2068 		spin_unlock(&sctx->stat_lock);
2069 		btrfs_err_rl_in_rcu(fs_info,
2070 			"IO error rebuilding logical %llu for dev %s",
2071 			logical, rcu_str_deref(dev->name));
2072 	} else if (sblock->header_error || sblock->checksum_error) {
2073 		spin_lock(&sctx->stat_lock);
2074 		sctx->stat.uncorrectable_errors++;
2075 		spin_unlock(&sctx->stat_lock);
2076 		btrfs_err_rl_in_rcu(fs_info,
2077 			"failed to rebuild valid logical %llu for dev %s",
2078 			logical, rcu_str_deref(dev->name));
2079 	} else {
2080 		scrub_write_block_to_dev_replace(sblock);
2081 	}
2082 
2083 	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2084 		mutex_lock(&sctx->wr_lock);
2085 		scrub_wr_submit(sctx);
2086 		mutex_unlock(&sctx->wr_lock);
2087 	}
2088 
2089 	scrub_block_put(sblock);
2090 	scrub_pending_bio_dec(sctx);
2091 }
2092 
2093 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2094 {
2095 	struct scrub_ctx *sctx = sblock->sctx;
2096 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2097 	u64 length = sblock->page_count * PAGE_SIZE;
2098 	u64 logical = sblock->pagev[0]->logical;
2099 	struct btrfs_bio *bbio = NULL;
2100 	struct bio *bio;
2101 	struct btrfs_raid_bio *rbio;
2102 	int ret;
2103 	int i;
2104 
2105 	btrfs_bio_counter_inc_blocked(fs_info);
2106 	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2107 			&length, &bbio);
2108 	if (ret || !bbio || !bbio->raid_map)
2109 		goto bbio_out;
2110 
2111 	if (WARN_ON(!sctx->is_dev_replace ||
2112 		    !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2113 		/*
2114 		 * We shouldn't be scrubbing a missing device. Even for dev
2115 		 * replace, we should only get here for RAID 5/6. We either
2116 		 * managed to mount something with no mirrors remaining or
2117 		 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2118 		 */
2119 		goto bbio_out;
2120 	}
2121 
2122 	bio = btrfs_io_bio_alloc(0);
2123 	bio->bi_iter.bi_sector = logical >> 9;
2124 	bio->bi_private = sblock;
2125 	bio->bi_end_io = scrub_missing_raid56_end_io;
2126 
2127 	rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2128 	if (!rbio)
2129 		goto rbio_out;
2130 
2131 	for (i = 0; i < sblock->page_count; i++) {
2132 		struct scrub_page *spage = sblock->pagev[i];
2133 
2134 		raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2135 	}
2136 
2137 	btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2138 	scrub_block_get(sblock);
2139 	scrub_pending_bio_inc(sctx);
2140 	raid56_submit_missing_rbio(rbio);
2141 	return;
2142 
2143 rbio_out:
2144 	bio_put(bio);
2145 bbio_out:
2146 	btrfs_bio_counter_dec(fs_info);
2147 	btrfs_put_bbio(bbio);
2148 	spin_lock(&sctx->stat_lock);
2149 	sctx->stat.malloc_errors++;
2150 	spin_unlock(&sctx->stat_lock);
2151 }
2152 
2153 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2154 		       u64 physical, struct btrfs_device *dev, u64 flags,
2155 		       u64 gen, int mirror_num, u8 *csum, int force,
2156 		       u64 physical_for_dev_replace)
2157 {
2158 	struct scrub_block *sblock;
2159 	int index;
2160 
2161 	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2162 	if (!sblock) {
2163 		spin_lock(&sctx->stat_lock);
2164 		sctx->stat.malloc_errors++;
2165 		spin_unlock(&sctx->stat_lock);
2166 		return -ENOMEM;
2167 	}
2168 
2169 	/* one ref inside this function, plus one for each page added to
2170 	 * a bio later on */
2171 	refcount_set(&sblock->refs, 1);
2172 	sblock->sctx = sctx;
2173 	sblock->no_io_error_seen = 1;
2174 
2175 	for (index = 0; len > 0; index++) {
2176 		struct scrub_page *spage;
2177 		u64 l = min_t(u64, len, PAGE_SIZE);
2178 
2179 		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2180 		if (!spage) {
2181 leave_nomem:
2182 			spin_lock(&sctx->stat_lock);
2183 			sctx->stat.malloc_errors++;
2184 			spin_unlock(&sctx->stat_lock);
2185 			scrub_block_put(sblock);
2186 			return -ENOMEM;
2187 		}
2188 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2189 		scrub_page_get(spage);
2190 		sblock->pagev[index] = spage;
2191 		spage->sblock = sblock;
2192 		spage->dev = dev;
2193 		spage->flags = flags;
2194 		spage->generation = gen;
2195 		spage->logical = logical;
2196 		spage->physical = physical;
2197 		spage->physical_for_dev_replace = physical_for_dev_replace;
2198 		spage->mirror_num = mirror_num;
2199 		if (csum) {
2200 			spage->have_csum = 1;
2201 			memcpy(spage->csum, csum, sctx->csum_size);
2202 		} else {
2203 			spage->have_csum = 0;
2204 		}
2205 		sblock->page_count++;
2206 		spage->page = alloc_page(GFP_KERNEL);
2207 		if (!spage->page)
2208 			goto leave_nomem;
2209 		len -= l;
2210 		logical += l;
2211 		physical += l;
2212 		physical_for_dev_replace += l;
2213 	}
2214 
2215 	WARN_ON(sblock->page_count == 0);
2216 	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2217 		/*
2218 		 * This case should only be hit for RAID 5/6 device replace. See
2219 		 * the comment in scrub_missing_raid56_pages() for details.
2220 		 */
2221 		scrub_missing_raid56_pages(sblock);
2222 	} else {
2223 		for (index = 0; index < sblock->page_count; index++) {
2224 			struct scrub_page *spage = sblock->pagev[index];
2225 			int ret;
2226 
2227 			ret = scrub_add_page_to_rd_bio(sctx, spage);
2228 			if (ret) {
2229 				scrub_block_put(sblock);
2230 				return ret;
2231 			}
2232 		}
2233 
2234 		if (force)
2235 			scrub_submit(sctx);
2236 	}
2237 
2238 	/* last one frees, either here or in bio completion for last page */
2239 	scrub_block_put(sblock);
2240 	return 0;
2241 }
2242 
2243 static void scrub_bio_end_io(struct bio *bio)
2244 {
2245 	struct scrub_bio *sbio = bio->bi_private;
2246 	struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2247 
2248 	sbio->status = bio->bi_status;
2249 	sbio->bio = bio;
2250 
2251 	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2252 }
2253 
2254 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2255 {
2256 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2257 	struct scrub_ctx *sctx = sbio->sctx;
2258 	int i;
2259 
2260 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2261 	if (sbio->status) {
2262 		for (i = 0; i < sbio->page_count; i++) {
2263 			struct scrub_page *spage = sbio->pagev[i];
2264 
2265 			spage->io_error = 1;
2266 			spage->sblock->no_io_error_seen = 0;
2267 		}
2268 	}
2269 
2270 	/* now complete the scrub_block items that have all pages completed */
2271 	for (i = 0; i < sbio->page_count; i++) {
2272 		struct scrub_page *spage = sbio->pagev[i];
2273 		struct scrub_block *sblock = spage->sblock;
2274 
2275 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2276 			scrub_block_complete(sblock);
2277 		scrub_block_put(sblock);
2278 	}
2279 
2280 	bio_put(sbio->bio);
2281 	sbio->bio = NULL;
2282 	spin_lock(&sctx->list_lock);
2283 	sbio->next_free = sctx->first_free;
2284 	sctx->first_free = sbio->index;
2285 	spin_unlock(&sctx->list_lock);
2286 
2287 	if (sctx->is_dev_replace && sctx->flush_all_writes) {
2288 		mutex_lock(&sctx->wr_lock);
2289 		scrub_wr_submit(sctx);
2290 		mutex_unlock(&sctx->wr_lock);
2291 	}
2292 
2293 	scrub_pending_bio_dec(sctx);
2294 }
2295 
2296 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2297 				       unsigned long *bitmap,
2298 				       u64 start, u64 len)
2299 {
2300 	u64 offset;
2301 	u64 nsectors64;
2302 	u32 nsectors;
2303 	int sectorsize = sparity->sctx->fs_info->sectorsize;
2304 
2305 	if (len >= sparity->stripe_len) {
2306 		bitmap_set(bitmap, 0, sparity->nsectors);
2307 		return;
2308 	}
2309 
2310 	start -= sparity->logic_start;
2311 	start = div64_u64_rem(start, sparity->stripe_len, &offset);
2312 	offset = div_u64(offset, sectorsize);
2313 	nsectors64 = div_u64(len, sectorsize);
2314 
2315 	ASSERT(nsectors64 < UINT_MAX);
2316 	nsectors = (u32)nsectors64;
2317 
2318 	if (offset + nsectors <= sparity->nsectors) {
2319 		bitmap_set(bitmap, offset, nsectors);
2320 		return;
2321 	}
2322 
2323 	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2324 	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2325 }
2326 
2327 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2328 						   u64 start, u64 len)
2329 {
2330 	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2331 }
2332 
2333 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2334 						  u64 start, u64 len)
2335 {
2336 	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2337 }
2338 
2339 static void scrub_block_complete(struct scrub_block *sblock)
2340 {
2341 	int corrupted = 0;
2342 
2343 	if (!sblock->no_io_error_seen) {
2344 		corrupted = 1;
2345 		scrub_handle_errored_block(sblock);
2346 	} else {
2347 		/*
2348 		 * if has checksum error, write via repair mechanism in
2349 		 * dev replace case, otherwise write here in dev replace
2350 		 * case.
2351 		 */
2352 		corrupted = scrub_checksum(sblock);
2353 		if (!corrupted && sblock->sctx->is_dev_replace)
2354 			scrub_write_block_to_dev_replace(sblock);
2355 	}
2356 
2357 	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2358 		u64 start = sblock->pagev[0]->logical;
2359 		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2360 			  PAGE_SIZE;
2361 
2362 		scrub_parity_mark_sectors_error(sblock->sparity,
2363 						start, end - start);
2364 	}
2365 }
2366 
2367 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2368 {
2369 	struct btrfs_ordered_sum *sum = NULL;
2370 	unsigned long index;
2371 	unsigned long num_sectors;
2372 
2373 	while (!list_empty(&sctx->csum_list)) {
2374 		sum = list_first_entry(&sctx->csum_list,
2375 				       struct btrfs_ordered_sum, list);
2376 		if (sum->bytenr > logical)
2377 			return 0;
2378 		if (sum->bytenr + sum->len > logical)
2379 			break;
2380 
2381 		++sctx->stat.csum_discards;
2382 		list_del(&sum->list);
2383 		kfree(sum);
2384 		sum = NULL;
2385 	}
2386 	if (!sum)
2387 		return 0;
2388 
2389 	index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2390 	ASSERT(index < UINT_MAX);
2391 
2392 	num_sectors = sum->len / sctx->fs_info->sectorsize;
2393 	memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2394 	if (index == num_sectors - 1) {
2395 		list_del(&sum->list);
2396 		kfree(sum);
2397 	}
2398 	return 1;
2399 }
2400 
2401 /* scrub extent tries to collect up to 64 kB for each bio */
2402 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2403 			u64 logical, u64 len,
2404 			u64 physical, struct btrfs_device *dev, u64 flags,
2405 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2406 {
2407 	int ret;
2408 	u8 csum[BTRFS_CSUM_SIZE];
2409 	u32 blocksize;
2410 
2411 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2412 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2413 			blocksize = map->stripe_len;
2414 		else
2415 			blocksize = sctx->fs_info->sectorsize;
2416 		spin_lock(&sctx->stat_lock);
2417 		sctx->stat.data_extents_scrubbed++;
2418 		sctx->stat.data_bytes_scrubbed += len;
2419 		spin_unlock(&sctx->stat_lock);
2420 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2421 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2422 			blocksize = map->stripe_len;
2423 		else
2424 			blocksize = sctx->fs_info->nodesize;
2425 		spin_lock(&sctx->stat_lock);
2426 		sctx->stat.tree_extents_scrubbed++;
2427 		sctx->stat.tree_bytes_scrubbed += len;
2428 		spin_unlock(&sctx->stat_lock);
2429 	} else {
2430 		blocksize = sctx->fs_info->sectorsize;
2431 		WARN_ON(1);
2432 	}
2433 
2434 	while (len) {
2435 		u64 l = min_t(u64, len, blocksize);
2436 		int have_csum = 0;
2437 
2438 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2439 			/* push csums to sbio */
2440 			have_csum = scrub_find_csum(sctx, logical, csum);
2441 			if (have_csum == 0)
2442 				++sctx->stat.no_csum;
2443 		}
2444 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2445 				  mirror_num, have_csum ? csum : NULL, 0,
2446 				  physical_for_dev_replace);
2447 		if (ret)
2448 			return ret;
2449 		len -= l;
2450 		logical += l;
2451 		physical += l;
2452 		physical_for_dev_replace += l;
2453 	}
2454 	return 0;
2455 }
2456 
2457 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2458 				  u64 logical, u64 len,
2459 				  u64 physical, struct btrfs_device *dev,
2460 				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2461 {
2462 	struct scrub_ctx *sctx = sparity->sctx;
2463 	struct scrub_block *sblock;
2464 	int index;
2465 
2466 	sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2467 	if (!sblock) {
2468 		spin_lock(&sctx->stat_lock);
2469 		sctx->stat.malloc_errors++;
2470 		spin_unlock(&sctx->stat_lock);
2471 		return -ENOMEM;
2472 	}
2473 
2474 	/* one ref inside this function, plus one for each page added to
2475 	 * a bio later on */
2476 	refcount_set(&sblock->refs, 1);
2477 	sblock->sctx = sctx;
2478 	sblock->no_io_error_seen = 1;
2479 	sblock->sparity = sparity;
2480 	scrub_parity_get(sparity);
2481 
2482 	for (index = 0; len > 0; index++) {
2483 		struct scrub_page *spage;
2484 		u64 l = min_t(u64, len, PAGE_SIZE);
2485 
2486 		spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2487 		if (!spage) {
2488 leave_nomem:
2489 			spin_lock(&sctx->stat_lock);
2490 			sctx->stat.malloc_errors++;
2491 			spin_unlock(&sctx->stat_lock);
2492 			scrub_block_put(sblock);
2493 			return -ENOMEM;
2494 		}
2495 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2496 		/* For scrub block */
2497 		scrub_page_get(spage);
2498 		sblock->pagev[index] = spage;
2499 		/* For scrub parity */
2500 		scrub_page_get(spage);
2501 		list_add_tail(&spage->list, &sparity->spages);
2502 		spage->sblock = sblock;
2503 		spage->dev = dev;
2504 		spage->flags = flags;
2505 		spage->generation = gen;
2506 		spage->logical = logical;
2507 		spage->physical = physical;
2508 		spage->mirror_num = mirror_num;
2509 		if (csum) {
2510 			spage->have_csum = 1;
2511 			memcpy(spage->csum, csum, sctx->csum_size);
2512 		} else {
2513 			spage->have_csum = 0;
2514 		}
2515 		sblock->page_count++;
2516 		spage->page = alloc_page(GFP_KERNEL);
2517 		if (!spage->page)
2518 			goto leave_nomem;
2519 		len -= l;
2520 		logical += l;
2521 		physical += l;
2522 	}
2523 
2524 	WARN_ON(sblock->page_count == 0);
2525 	for (index = 0; index < sblock->page_count; index++) {
2526 		struct scrub_page *spage = sblock->pagev[index];
2527 		int ret;
2528 
2529 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2530 		if (ret) {
2531 			scrub_block_put(sblock);
2532 			return ret;
2533 		}
2534 	}
2535 
2536 	/* last one frees, either here or in bio completion for last page */
2537 	scrub_block_put(sblock);
2538 	return 0;
2539 }
2540 
2541 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2542 				   u64 logical, u64 len,
2543 				   u64 physical, struct btrfs_device *dev,
2544 				   u64 flags, u64 gen, int mirror_num)
2545 {
2546 	struct scrub_ctx *sctx = sparity->sctx;
2547 	int ret;
2548 	u8 csum[BTRFS_CSUM_SIZE];
2549 	u32 blocksize;
2550 
2551 	if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2552 		scrub_parity_mark_sectors_error(sparity, logical, len);
2553 		return 0;
2554 	}
2555 
2556 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2557 		blocksize = sparity->stripe_len;
2558 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2559 		blocksize = sparity->stripe_len;
2560 	} else {
2561 		blocksize = sctx->fs_info->sectorsize;
2562 		WARN_ON(1);
2563 	}
2564 
2565 	while (len) {
2566 		u64 l = min_t(u64, len, blocksize);
2567 		int have_csum = 0;
2568 
2569 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2570 			/* push csums to sbio */
2571 			have_csum = scrub_find_csum(sctx, logical, csum);
2572 			if (have_csum == 0)
2573 				goto skip;
2574 		}
2575 		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2576 					     flags, gen, mirror_num,
2577 					     have_csum ? csum : NULL);
2578 		if (ret)
2579 			return ret;
2580 skip:
2581 		len -= l;
2582 		logical += l;
2583 		physical += l;
2584 	}
2585 	return 0;
2586 }
2587 
2588 /*
2589  * Given a physical address, this will calculate it's
2590  * logical offset. if this is a parity stripe, it will return
2591  * the most left data stripe's logical offset.
2592  *
2593  * return 0 if it is a data stripe, 1 means parity stripe.
2594  */
2595 static int get_raid56_logic_offset(u64 physical, int num,
2596 				   struct map_lookup *map, u64 *offset,
2597 				   u64 *stripe_start)
2598 {
2599 	int i;
2600 	int j = 0;
2601 	u64 stripe_nr;
2602 	u64 last_offset;
2603 	u32 stripe_index;
2604 	u32 rot;
2605 	const int data_stripes = nr_data_stripes(map);
2606 
2607 	last_offset = (physical - map->stripes[num].physical) * data_stripes;
2608 	if (stripe_start)
2609 		*stripe_start = last_offset;
2610 
2611 	*offset = last_offset;
2612 	for (i = 0; i < data_stripes; i++) {
2613 		*offset = last_offset + i * map->stripe_len;
2614 
2615 		stripe_nr = div64_u64(*offset, map->stripe_len);
2616 		stripe_nr = div_u64(stripe_nr, data_stripes);
2617 
2618 		/* Work out the disk rotation on this stripe-set */
2619 		stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2620 		/* calculate which stripe this data locates */
2621 		rot += i;
2622 		stripe_index = rot % map->num_stripes;
2623 		if (stripe_index == num)
2624 			return 0;
2625 		if (stripe_index < num)
2626 			j++;
2627 	}
2628 	*offset = last_offset + j * map->stripe_len;
2629 	return 1;
2630 }
2631 
2632 static void scrub_free_parity(struct scrub_parity *sparity)
2633 {
2634 	struct scrub_ctx *sctx = sparity->sctx;
2635 	struct scrub_page *curr, *next;
2636 	int nbits;
2637 
2638 	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2639 	if (nbits) {
2640 		spin_lock(&sctx->stat_lock);
2641 		sctx->stat.read_errors += nbits;
2642 		sctx->stat.uncorrectable_errors += nbits;
2643 		spin_unlock(&sctx->stat_lock);
2644 	}
2645 
2646 	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2647 		list_del_init(&curr->list);
2648 		scrub_page_put(curr);
2649 	}
2650 
2651 	kfree(sparity);
2652 }
2653 
2654 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2655 {
2656 	struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2657 						    work);
2658 	struct scrub_ctx *sctx = sparity->sctx;
2659 
2660 	scrub_free_parity(sparity);
2661 	scrub_pending_bio_dec(sctx);
2662 }
2663 
2664 static void scrub_parity_bio_endio(struct bio *bio)
2665 {
2666 	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2667 	struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2668 
2669 	if (bio->bi_status)
2670 		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2671 			  sparity->nsectors);
2672 
2673 	bio_put(bio);
2674 
2675 	btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2676 			NULL);
2677 	btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2678 }
2679 
2680 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2681 {
2682 	struct scrub_ctx *sctx = sparity->sctx;
2683 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2684 	struct bio *bio;
2685 	struct btrfs_raid_bio *rbio;
2686 	struct btrfs_bio *bbio = NULL;
2687 	u64 length;
2688 	int ret;
2689 
2690 	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2691 			   sparity->nsectors))
2692 		goto out;
2693 
2694 	length = sparity->logic_end - sparity->logic_start;
2695 
2696 	btrfs_bio_counter_inc_blocked(fs_info);
2697 	ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2698 			       &length, &bbio);
2699 	if (ret || !bbio || !bbio->raid_map)
2700 		goto bbio_out;
2701 
2702 	bio = btrfs_io_bio_alloc(0);
2703 	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2704 	bio->bi_private = sparity;
2705 	bio->bi_end_io = scrub_parity_bio_endio;
2706 
2707 	rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2708 					      length, sparity->scrub_dev,
2709 					      sparity->dbitmap,
2710 					      sparity->nsectors);
2711 	if (!rbio)
2712 		goto rbio_out;
2713 
2714 	scrub_pending_bio_inc(sctx);
2715 	raid56_parity_submit_scrub_rbio(rbio);
2716 	return;
2717 
2718 rbio_out:
2719 	bio_put(bio);
2720 bbio_out:
2721 	btrfs_bio_counter_dec(fs_info);
2722 	btrfs_put_bbio(bbio);
2723 	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2724 		  sparity->nsectors);
2725 	spin_lock(&sctx->stat_lock);
2726 	sctx->stat.malloc_errors++;
2727 	spin_unlock(&sctx->stat_lock);
2728 out:
2729 	scrub_free_parity(sparity);
2730 }
2731 
2732 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2733 {
2734 	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2735 }
2736 
2737 static void scrub_parity_get(struct scrub_parity *sparity)
2738 {
2739 	refcount_inc(&sparity->refs);
2740 }
2741 
2742 static void scrub_parity_put(struct scrub_parity *sparity)
2743 {
2744 	if (!refcount_dec_and_test(&sparity->refs))
2745 		return;
2746 
2747 	scrub_parity_check_and_repair(sparity);
2748 }
2749 
2750 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2751 						  struct map_lookup *map,
2752 						  struct btrfs_device *sdev,
2753 						  struct btrfs_path *path,
2754 						  u64 logic_start,
2755 						  u64 logic_end)
2756 {
2757 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2758 	struct btrfs_root *root = fs_info->extent_root;
2759 	struct btrfs_root *csum_root = fs_info->csum_root;
2760 	struct btrfs_extent_item *extent;
2761 	struct btrfs_bio *bbio = NULL;
2762 	u64 flags;
2763 	int ret;
2764 	int slot;
2765 	struct extent_buffer *l;
2766 	struct btrfs_key key;
2767 	u64 generation;
2768 	u64 extent_logical;
2769 	u64 extent_physical;
2770 	u64 extent_len;
2771 	u64 mapped_length;
2772 	struct btrfs_device *extent_dev;
2773 	struct scrub_parity *sparity;
2774 	int nsectors;
2775 	int bitmap_len;
2776 	int extent_mirror_num;
2777 	int stop_loop = 0;
2778 
2779 	nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2780 	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2781 	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2782 			  GFP_NOFS);
2783 	if (!sparity) {
2784 		spin_lock(&sctx->stat_lock);
2785 		sctx->stat.malloc_errors++;
2786 		spin_unlock(&sctx->stat_lock);
2787 		return -ENOMEM;
2788 	}
2789 
2790 	sparity->stripe_len = map->stripe_len;
2791 	sparity->nsectors = nsectors;
2792 	sparity->sctx = sctx;
2793 	sparity->scrub_dev = sdev;
2794 	sparity->logic_start = logic_start;
2795 	sparity->logic_end = logic_end;
2796 	refcount_set(&sparity->refs, 1);
2797 	INIT_LIST_HEAD(&sparity->spages);
2798 	sparity->dbitmap = sparity->bitmap;
2799 	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2800 
2801 	ret = 0;
2802 	while (logic_start < logic_end) {
2803 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2804 			key.type = BTRFS_METADATA_ITEM_KEY;
2805 		else
2806 			key.type = BTRFS_EXTENT_ITEM_KEY;
2807 		key.objectid = logic_start;
2808 		key.offset = (u64)-1;
2809 
2810 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2811 		if (ret < 0)
2812 			goto out;
2813 
2814 		if (ret > 0) {
2815 			ret = btrfs_previous_extent_item(root, path, 0);
2816 			if (ret < 0)
2817 				goto out;
2818 			if (ret > 0) {
2819 				btrfs_release_path(path);
2820 				ret = btrfs_search_slot(NULL, root, &key,
2821 							path, 0, 0);
2822 				if (ret < 0)
2823 					goto out;
2824 			}
2825 		}
2826 
2827 		stop_loop = 0;
2828 		while (1) {
2829 			u64 bytes;
2830 
2831 			l = path->nodes[0];
2832 			slot = path->slots[0];
2833 			if (slot >= btrfs_header_nritems(l)) {
2834 				ret = btrfs_next_leaf(root, path);
2835 				if (ret == 0)
2836 					continue;
2837 				if (ret < 0)
2838 					goto out;
2839 
2840 				stop_loop = 1;
2841 				break;
2842 			}
2843 			btrfs_item_key_to_cpu(l, &key, slot);
2844 
2845 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2846 			    key.type != BTRFS_METADATA_ITEM_KEY)
2847 				goto next;
2848 
2849 			if (key.type == BTRFS_METADATA_ITEM_KEY)
2850 				bytes = fs_info->nodesize;
2851 			else
2852 				bytes = key.offset;
2853 
2854 			if (key.objectid + bytes <= logic_start)
2855 				goto next;
2856 
2857 			if (key.objectid >= logic_end) {
2858 				stop_loop = 1;
2859 				break;
2860 			}
2861 
2862 			while (key.objectid >= logic_start + map->stripe_len)
2863 				logic_start += map->stripe_len;
2864 
2865 			extent = btrfs_item_ptr(l, slot,
2866 						struct btrfs_extent_item);
2867 			flags = btrfs_extent_flags(l, extent);
2868 			generation = btrfs_extent_generation(l, extent);
2869 
2870 			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2871 			    (key.objectid < logic_start ||
2872 			     key.objectid + bytes >
2873 			     logic_start + map->stripe_len)) {
2874 				btrfs_err(fs_info,
2875 					  "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2876 					  key.objectid, logic_start);
2877 				spin_lock(&sctx->stat_lock);
2878 				sctx->stat.uncorrectable_errors++;
2879 				spin_unlock(&sctx->stat_lock);
2880 				goto next;
2881 			}
2882 again:
2883 			extent_logical = key.objectid;
2884 			extent_len = bytes;
2885 
2886 			if (extent_logical < logic_start) {
2887 				extent_len -= logic_start - extent_logical;
2888 				extent_logical = logic_start;
2889 			}
2890 
2891 			if (extent_logical + extent_len >
2892 			    logic_start + map->stripe_len)
2893 				extent_len = logic_start + map->stripe_len -
2894 					     extent_logical;
2895 
2896 			scrub_parity_mark_sectors_data(sparity, extent_logical,
2897 						       extent_len);
2898 
2899 			mapped_length = extent_len;
2900 			bbio = NULL;
2901 			ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2902 					extent_logical, &mapped_length, &bbio,
2903 					0);
2904 			if (!ret) {
2905 				if (!bbio || mapped_length < extent_len)
2906 					ret = -EIO;
2907 			}
2908 			if (ret) {
2909 				btrfs_put_bbio(bbio);
2910 				goto out;
2911 			}
2912 			extent_physical = bbio->stripes[0].physical;
2913 			extent_mirror_num = bbio->mirror_num;
2914 			extent_dev = bbio->stripes[0].dev;
2915 			btrfs_put_bbio(bbio);
2916 
2917 			ret = btrfs_lookup_csums_range(csum_root,
2918 						extent_logical,
2919 						extent_logical + extent_len - 1,
2920 						&sctx->csum_list, 1);
2921 			if (ret)
2922 				goto out;
2923 
2924 			ret = scrub_extent_for_parity(sparity, extent_logical,
2925 						      extent_len,
2926 						      extent_physical,
2927 						      extent_dev, flags,
2928 						      generation,
2929 						      extent_mirror_num);
2930 
2931 			scrub_free_csums(sctx);
2932 
2933 			if (ret)
2934 				goto out;
2935 
2936 			if (extent_logical + extent_len <
2937 			    key.objectid + bytes) {
2938 				logic_start += map->stripe_len;
2939 
2940 				if (logic_start >= logic_end) {
2941 					stop_loop = 1;
2942 					break;
2943 				}
2944 
2945 				if (logic_start < key.objectid + bytes) {
2946 					cond_resched();
2947 					goto again;
2948 				}
2949 			}
2950 next:
2951 			path->slots[0]++;
2952 		}
2953 
2954 		btrfs_release_path(path);
2955 
2956 		if (stop_loop)
2957 			break;
2958 
2959 		logic_start += map->stripe_len;
2960 	}
2961 out:
2962 	if (ret < 0)
2963 		scrub_parity_mark_sectors_error(sparity, logic_start,
2964 						logic_end - logic_start);
2965 	scrub_parity_put(sparity);
2966 	scrub_submit(sctx);
2967 	mutex_lock(&sctx->wr_lock);
2968 	scrub_wr_submit(sctx);
2969 	mutex_unlock(&sctx->wr_lock);
2970 
2971 	btrfs_release_path(path);
2972 	return ret < 0 ? ret : 0;
2973 }
2974 
2975 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2976 					   struct map_lookup *map,
2977 					   struct btrfs_device *scrub_dev,
2978 					   int num, u64 base, u64 length,
2979 					   struct btrfs_block_group *cache)
2980 {
2981 	struct btrfs_path *path, *ppath;
2982 	struct btrfs_fs_info *fs_info = sctx->fs_info;
2983 	struct btrfs_root *root = fs_info->extent_root;
2984 	struct btrfs_root *csum_root = fs_info->csum_root;
2985 	struct btrfs_extent_item *extent;
2986 	struct blk_plug plug;
2987 	u64 flags;
2988 	int ret;
2989 	int slot;
2990 	u64 nstripes;
2991 	struct extent_buffer *l;
2992 	u64 physical;
2993 	u64 logical;
2994 	u64 logic_end;
2995 	u64 physical_end;
2996 	u64 generation;
2997 	int mirror_num;
2998 	struct reada_control *reada1;
2999 	struct reada_control *reada2;
3000 	struct btrfs_key key;
3001 	struct btrfs_key key_end;
3002 	u64 increment = map->stripe_len;
3003 	u64 offset;
3004 	u64 extent_logical;
3005 	u64 extent_physical;
3006 	u64 extent_len;
3007 	u64 stripe_logical;
3008 	u64 stripe_end;
3009 	struct btrfs_device *extent_dev;
3010 	int extent_mirror_num;
3011 	int stop_loop = 0;
3012 
3013 	physical = map->stripes[num].physical;
3014 	offset = 0;
3015 	nstripes = div64_u64(length, map->stripe_len);
3016 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3017 		offset = map->stripe_len * num;
3018 		increment = map->stripe_len * map->num_stripes;
3019 		mirror_num = 1;
3020 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3021 		int factor = map->num_stripes / map->sub_stripes;
3022 		offset = map->stripe_len * (num / map->sub_stripes);
3023 		increment = map->stripe_len * factor;
3024 		mirror_num = num % map->sub_stripes + 1;
3025 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3026 		increment = map->stripe_len;
3027 		mirror_num = num % map->num_stripes + 1;
3028 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3029 		increment = map->stripe_len;
3030 		mirror_num = num % map->num_stripes + 1;
3031 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3032 		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3033 		increment = map->stripe_len * nr_data_stripes(map);
3034 		mirror_num = 1;
3035 	} else {
3036 		increment = map->stripe_len;
3037 		mirror_num = 1;
3038 	}
3039 
3040 	path = btrfs_alloc_path();
3041 	if (!path)
3042 		return -ENOMEM;
3043 
3044 	ppath = btrfs_alloc_path();
3045 	if (!ppath) {
3046 		btrfs_free_path(path);
3047 		return -ENOMEM;
3048 	}
3049 
3050 	/*
3051 	 * work on commit root. The related disk blocks are static as
3052 	 * long as COW is applied. This means, it is save to rewrite
3053 	 * them to repair disk errors without any race conditions
3054 	 */
3055 	path->search_commit_root = 1;
3056 	path->skip_locking = 1;
3057 
3058 	ppath->search_commit_root = 1;
3059 	ppath->skip_locking = 1;
3060 	/*
3061 	 * trigger the readahead for extent tree csum tree and wait for
3062 	 * completion. During readahead, the scrub is officially paused
3063 	 * to not hold off transaction commits
3064 	 */
3065 	logical = base + offset;
3066 	physical_end = physical + nstripes * map->stripe_len;
3067 	if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3068 		get_raid56_logic_offset(physical_end, num,
3069 					map, &logic_end, NULL);
3070 		logic_end += base;
3071 	} else {
3072 		logic_end = logical + increment * nstripes;
3073 	}
3074 	wait_event(sctx->list_wait,
3075 		   atomic_read(&sctx->bios_in_flight) == 0);
3076 	scrub_blocked_if_needed(fs_info);
3077 
3078 	/* FIXME it might be better to start readahead at commit root */
3079 	key.objectid = logical;
3080 	key.type = BTRFS_EXTENT_ITEM_KEY;
3081 	key.offset = (u64)0;
3082 	key_end.objectid = logic_end;
3083 	key_end.type = BTRFS_METADATA_ITEM_KEY;
3084 	key_end.offset = (u64)-1;
3085 	reada1 = btrfs_reada_add(root, &key, &key_end);
3086 
3087 	key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3088 	key.type = BTRFS_EXTENT_CSUM_KEY;
3089 	key.offset = logical;
3090 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3091 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
3092 	key_end.offset = logic_end;
3093 	reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3094 
3095 	if (!IS_ERR(reada1))
3096 		btrfs_reada_wait(reada1);
3097 	if (!IS_ERR(reada2))
3098 		btrfs_reada_wait(reada2);
3099 
3100 
3101 	/*
3102 	 * collect all data csums for the stripe to avoid seeking during
3103 	 * the scrub. This might currently (crc32) end up to be about 1MB
3104 	 */
3105 	blk_start_plug(&plug);
3106 
3107 	/*
3108 	 * now find all extents for each stripe and scrub them
3109 	 */
3110 	ret = 0;
3111 	while (physical < physical_end) {
3112 		/*
3113 		 * canceled?
3114 		 */
3115 		if (atomic_read(&fs_info->scrub_cancel_req) ||
3116 		    atomic_read(&sctx->cancel_req)) {
3117 			ret = -ECANCELED;
3118 			goto out;
3119 		}
3120 		/*
3121 		 * check to see if we have to pause
3122 		 */
3123 		if (atomic_read(&fs_info->scrub_pause_req)) {
3124 			/* push queued extents */
3125 			sctx->flush_all_writes = true;
3126 			scrub_submit(sctx);
3127 			mutex_lock(&sctx->wr_lock);
3128 			scrub_wr_submit(sctx);
3129 			mutex_unlock(&sctx->wr_lock);
3130 			wait_event(sctx->list_wait,
3131 				   atomic_read(&sctx->bios_in_flight) == 0);
3132 			sctx->flush_all_writes = false;
3133 			scrub_blocked_if_needed(fs_info);
3134 		}
3135 
3136 		if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3137 			ret = get_raid56_logic_offset(physical, num, map,
3138 						      &logical,
3139 						      &stripe_logical);
3140 			logical += base;
3141 			if (ret) {
3142 				/* it is parity strip */
3143 				stripe_logical += base;
3144 				stripe_end = stripe_logical + increment;
3145 				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3146 							  ppath, stripe_logical,
3147 							  stripe_end);
3148 				if (ret)
3149 					goto out;
3150 				goto skip;
3151 			}
3152 		}
3153 
3154 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3155 			key.type = BTRFS_METADATA_ITEM_KEY;
3156 		else
3157 			key.type = BTRFS_EXTENT_ITEM_KEY;
3158 		key.objectid = logical;
3159 		key.offset = (u64)-1;
3160 
3161 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3162 		if (ret < 0)
3163 			goto out;
3164 
3165 		if (ret > 0) {
3166 			ret = btrfs_previous_extent_item(root, path, 0);
3167 			if (ret < 0)
3168 				goto out;
3169 			if (ret > 0) {
3170 				/* there's no smaller item, so stick with the
3171 				 * larger one */
3172 				btrfs_release_path(path);
3173 				ret = btrfs_search_slot(NULL, root, &key,
3174 							path, 0, 0);
3175 				if (ret < 0)
3176 					goto out;
3177 			}
3178 		}
3179 
3180 		stop_loop = 0;
3181 		while (1) {
3182 			u64 bytes;
3183 
3184 			l = path->nodes[0];
3185 			slot = path->slots[0];
3186 			if (slot >= btrfs_header_nritems(l)) {
3187 				ret = btrfs_next_leaf(root, path);
3188 				if (ret == 0)
3189 					continue;
3190 				if (ret < 0)
3191 					goto out;
3192 
3193 				stop_loop = 1;
3194 				break;
3195 			}
3196 			btrfs_item_key_to_cpu(l, &key, slot);
3197 
3198 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3199 			    key.type != BTRFS_METADATA_ITEM_KEY)
3200 				goto next;
3201 
3202 			if (key.type == BTRFS_METADATA_ITEM_KEY)
3203 				bytes = fs_info->nodesize;
3204 			else
3205 				bytes = key.offset;
3206 
3207 			if (key.objectid + bytes <= logical)
3208 				goto next;
3209 
3210 			if (key.objectid >= logical + map->stripe_len) {
3211 				/* out of this device extent */
3212 				if (key.objectid >= logic_end)
3213 					stop_loop = 1;
3214 				break;
3215 			}
3216 
3217 			/*
3218 			 * If our block group was removed in the meanwhile, just
3219 			 * stop scrubbing since there is no point in continuing.
3220 			 * Continuing would prevent reusing its device extents
3221 			 * for new block groups for a long time.
3222 			 */
3223 			spin_lock(&cache->lock);
3224 			if (cache->removed) {
3225 				spin_unlock(&cache->lock);
3226 				ret = 0;
3227 				goto out;
3228 			}
3229 			spin_unlock(&cache->lock);
3230 
3231 			extent = btrfs_item_ptr(l, slot,
3232 						struct btrfs_extent_item);
3233 			flags = btrfs_extent_flags(l, extent);
3234 			generation = btrfs_extent_generation(l, extent);
3235 
3236 			if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3237 			    (key.objectid < logical ||
3238 			     key.objectid + bytes >
3239 			     logical + map->stripe_len)) {
3240 				btrfs_err(fs_info,
3241 					   "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3242 				       key.objectid, logical);
3243 				spin_lock(&sctx->stat_lock);
3244 				sctx->stat.uncorrectable_errors++;
3245 				spin_unlock(&sctx->stat_lock);
3246 				goto next;
3247 			}
3248 
3249 again:
3250 			extent_logical = key.objectid;
3251 			extent_len = bytes;
3252 
3253 			/*
3254 			 * trim extent to this stripe
3255 			 */
3256 			if (extent_logical < logical) {
3257 				extent_len -= logical - extent_logical;
3258 				extent_logical = logical;
3259 			}
3260 			if (extent_logical + extent_len >
3261 			    logical + map->stripe_len) {
3262 				extent_len = logical + map->stripe_len -
3263 					     extent_logical;
3264 			}
3265 
3266 			extent_physical = extent_logical - logical + physical;
3267 			extent_dev = scrub_dev;
3268 			extent_mirror_num = mirror_num;
3269 			if (sctx->is_dev_replace)
3270 				scrub_remap_extent(fs_info, extent_logical,
3271 						   extent_len, &extent_physical,
3272 						   &extent_dev,
3273 						   &extent_mirror_num);
3274 
3275 			if (flags & BTRFS_EXTENT_FLAG_DATA) {
3276 				ret = btrfs_lookup_csums_range(csum_root,
3277 						extent_logical,
3278 						extent_logical + extent_len - 1,
3279 						&sctx->csum_list, 1);
3280 				if (ret)
3281 					goto out;
3282 			}
3283 
3284 			ret = scrub_extent(sctx, map, extent_logical, extent_len,
3285 					   extent_physical, extent_dev, flags,
3286 					   generation, extent_mirror_num,
3287 					   extent_logical - logical + physical);
3288 
3289 			scrub_free_csums(sctx);
3290 
3291 			if (ret)
3292 				goto out;
3293 
3294 			if (extent_logical + extent_len <
3295 			    key.objectid + bytes) {
3296 				if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3297 					/*
3298 					 * loop until we find next data stripe
3299 					 * or we have finished all stripes.
3300 					 */
3301 loop:
3302 					physical += map->stripe_len;
3303 					ret = get_raid56_logic_offset(physical,
3304 							num, map, &logical,
3305 							&stripe_logical);
3306 					logical += base;
3307 
3308 					if (ret && physical < physical_end) {
3309 						stripe_logical += base;
3310 						stripe_end = stripe_logical +
3311 								increment;
3312 						ret = scrub_raid56_parity(sctx,
3313 							map, scrub_dev, ppath,
3314 							stripe_logical,
3315 							stripe_end);
3316 						if (ret)
3317 							goto out;
3318 						goto loop;
3319 					}
3320 				} else {
3321 					physical += map->stripe_len;
3322 					logical += increment;
3323 				}
3324 				if (logical < key.objectid + bytes) {
3325 					cond_resched();
3326 					goto again;
3327 				}
3328 
3329 				if (physical >= physical_end) {
3330 					stop_loop = 1;
3331 					break;
3332 				}
3333 			}
3334 next:
3335 			path->slots[0]++;
3336 		}
3337 		btrfs_release_path(path);
3338 skip:
3339 		logical += increment;
3340 		physical += map->stripe_len;
3341 		spin_lock(&sctx->stat_lock);
3342 		if (stop_loop)
3343 			sctx->stat.last_physical = map->stripes[num].physical +
3344 						   length;
3345 		else
3346 			sctx->stat.last_physical = physical;
3347 		spin_unlock(&sctx->stat_lock);
3348 		if (stop_loop)
3349 			break;
3350 	}
3351 out:
3352 	/* push queued extents */
3353 	scrub_submit(sctx);
3354 	mutex_lock(&sctx->wr_lock);
3355 	scrub_wr_submit(sctx);
3356 	mutex_unlock(&sctx->wr_lock);
3357 
3358 	blk_finish_plug(&plug);
3359 	btrfs_free_path(path);
3360 	btrfs_free_path(ppath);
3361 	return ret < 0 ? ret : 0;
3362 }
3363 
3364 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3365 					  struct btrfs_device *scrub_dev,
3366 					  u64 chunk_offset, u64 length,
3367 					  u64 dev_offset,
3368 					  struct btrfs_block_group *cache)
3369 {
3370 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3371 	struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3372 	struct map_lookup *map;
3373 	struct extent_map *em;
3374 	int i;
3375 	int ret = 0;
3376 
3377 	read_lock(&map_tree->lock);
3378 	em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3379 	read_unlock(&map_tree->lock);
3380 
3381 	if (!em) {
3382 		/*
3383 		 * Might have been an unused block group deleted by the cleaner
3384 		 * kthread or relocation.
3385 		 */
3386 		spin_lock(&cache->lock);
3387 		if (!cache->removed)
3388 			ret = -EINVAL;
3389 		spin_unlock(&cache->lock);
3390 
3391 		return ret;
3392 	}
3393 
3394 	map = em->map_lookup;
3395 	if (em->start != chunk_offset)
3396 		goto out;
3397 
3398 	if (em->len < length)
3399 		goto out;
3400 
3401 	for (i = 0; i < map->num_stripes; ++i) {
3402 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3403 		    map->stripes[i].physical == dev_offset) {
3404 			ret = scrub_stripe(sctx, map, scrub_dev, i,
3405 					   chunk_offset, length, cache);
3406 			if (ret)
3407 				goto out;
3408 		}
3409 	}
3410 out:
3411 	free_extent_map(em);
3412 
3413 	return ret;
3414 }
3415 
3416 static noinline_for_stack
3417 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3418 			   struct btrfs_device *scrub_dev, u64 start, u64 end)
3419 {
3420 	struct btrfs_dev_extent *dev_extent = NULL;
3421 	struct btrfs_path *path;
3422 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3423 	struct btrfs_root *root = fs_info->dev_root;
3424 	u64 length;
3425 	u64 chunk_offset;
3426 	int ret = 0;
3427 	int ro_set;
3428 	int slot;
3429 	struct extent_buffer *l;
3430 	struct btrfs_key key;
3431 	struct btrfs_key found_key;
3432 	struct btrfs_block_group *cache;
3433 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3434 
3435 	path = btrfs_alloc_path();
3436 	if (!path)
3437 		return -ENOMEM;
3438 
3439 	path->reada = READA_FORWARD;
3440 	path->search_commit_root = 1;
3441 	path->skip_locking = 1;
3442 
3443 	key.objectid = scrub_dev->devid;
3444 	key.offset = 0ull;
3445 	key.type = BTRFS_DEV_EXTENT_KEY;
3446 
3447 	while (1) {
3448 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3449 		if (ret < 0)
3450 			break;
3451 		if (ret > 0) {
3452 			if (path->slots[0] >=
3453 			    btrfs_header_nritems(path->nodes[0])) {
3454 				ret = btrfs_next_leaf(root, path);
3455 				if (ret < 0)
3456 					break;
3457 				if (ret > 0) {
3458 					ret = 0;
3459 					break;
3460 				}
3461 			} else {
3462 				ret = 0;
3463 			}
3464 		}
3465 
3466 		l = path->nodes[0];
3467 		slot = path->slots[0];
3468 
3469 		btrfs_item_key_to_cpu(l, &found_key, slot);
3470 
3471 		if (found_key.objectid != scrub_dev->devid)
3472 			break;
3473 
3474 		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3475 			break;
3476 
3477 		if (found_key.offset >= end)
3478 			break;
3479 
3480 		if (found_key.offset < key.offset)
3481 			break;
3482 
3483 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3484 		length = btrfs_dev_extent_length(l, dev_extent);
3485 
3486 		if (found_key.offset + length <= start)
3487 			goto skip;
3488 
3489 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3490 
3491 		/*
3492 		 * get a reference on the corresponding block group to prevent
3493 		 * the chunk from going away while we scrub it
3494 		 */
3495 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3496 
3497 		/* some chunks are removed but not committed to disk yet,
3498 		 * continue scrubbing */
3499 		if (!cache)
3500 			goto skip;
3501 
3502 		/*
3503 		 * Make sure that while we are scrubbing the corresponding block
3504 		 * group doesn't get its logical address and its device extents
3505 		 * reused for another block group, which can possibly be of a
3506 		 * different type and different profile. We do this to prevent
3507 		 * false error detections and crashes due to bogus attempts to
3508 		 * repair extents.
3509 		 */
3510 		spin_lock(&cache->lock);
3511 		if (cache->removed) {
3512 			spin_unlock(&cache->lock);
3513 			btrfs_put_block_group(cache);
3514 			goto skip;
3515 		}
3516 		btrfs_freeze_block_group(cache);
3517 		spin_unlock(&cache->lock);
3518 
3519 		/*
3520 		 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3521 		 * to avoid deadlock caused by:
3522 		 * btrfs_inc_block_group_ro()
3523 		 * -> btrfs_wait_for_commit()
3524 		 * -> btrfs_commit_transaction()
3525 		 * -> btrfs_scrub_pause()
3526 		 */
3527 		scrub_pause_on(fs_info);
3528 
3529 		/*
3530 		 * Don't do chunk preallocation for scrub.
3531 		 *
3532 		 * This is especially important for SYSTEM bgs, or we can hit
3533 		 * -EFBIG from btrfs_finish_chunk_alloc() like:
3534 		 * 1. The only SYSTEM bg is marked RO.
3535 		 *    Since SYSTEM bg is small, that's pretty common.
3536 		 * 2. New SYSTEM bg will be allocated
3537 		 *    Due to regular version will allocate new chunk.
3538 		 * 3. New SYSTEM bg is empty and will get cleaned up
3539 		 *    Before cleanup really happens, it's marked RO again.
3540 		 * 4. Empty SYSTEM bg get scrubbed
3541 		 *    We go back to 2.
3542 		 *
3543 		 * This can easily boost the amount of SYSTEM chunks if cleaner
3544 		 * thread can't be triggered fast enough, and use up all space
3545 		 * of btrfs_super_block::sys_chunk_array
3546 		 *
3547 		 * While for dev replace, we need to try our best to mark block
3548 		 * group RO, to prevent race between:
3549 		 * - Write duplication
3550 		 *   Contains latest data
3551 		 * - Scrub copy
3552 		 *   Contains data from commit tree
3553 		 *
3554 		 * If target block group is not marked RO, nocow writes can
3555 		 * be overwritten by scrub copy, causing data corruption.
3556 		 * So for dev-replace, it's not allowed to continue if a block
3557 		 * group is not RO.
3558 		 */
3559 		ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3560 		if (ret == 0) {
3561 			ro_set = 1;
3562 		} else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3563 			/*
3564 			 * btrfs_inc_block_group_ro return -ENOSPC when it
3565 			 * failed in creating new chunk for metadata.
3566 			 * It is not a problem for scrub, because
3567 			 * metadata are always cowed, and our scrub paused
3568 			 * commit_transactions.
3569 			 */
3570 			ro_set = 0;
3571 		} else {
3572 			btrfs_warn(fs_info,
3573 				   "failed setting block group ro: %d", ret);
3574 			btrfs_unfreeze_block_group(cache);
3575 			btrfs_put_block_group(cache);
3576 			scrub_pause_off(fs_info);
3577 			break;
3578 		}
3579 
3580 		/*
3581 		 * Now the target block is marked RO, wait for nocow writes to
3582 		 * finish before dev-replace.
3583 		 * COW is fine, as COW never overwrites extents in commit tree.
3584 		 */
3585 		if (sctx->is_dev_replace) {
3586 			btrfs_wait_nocow_writers(cache);
3587 			btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3588 					cache->length);
3589 		}
3590 
3591 		scrub_pause_off(fs_info);
3592 		down_write(&dev_replace->rwsem);
3593 		dev_replace->cursor_right = found_key.offset + length;
3594 		dev_replace->cursor_left = found_key.offset;
3595 		dev_replace->item_needs_writeback = 1;
3596 		up_write(&dev_replace->rwsem);
3597 
3598 		ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3599 				  found_key.offset, cache);
3600 
3601 		/*
3602 		 * flush, submit all pending read and write bios, afterwards
3603 		 * wait for them.
3604 		 * Note that in the dev replace case, a read request causes
3605 		 * write requests that are submitted in the read completion
3606 		 * worker. Therefore in the current situation, it is required
3607 		 * that all write requests are flushed, so that all read and
3608 		 * write requests are really completed when bios_in_flight
3609 		 * changes to 0.
3610 		 */
3611 		sctx->flush_all_writes = true;
3612 		scrub_submit(sctx);
3613 		mutex_lock(&sctx->wr_lock);
3614 		scrub_wr_submit(sctx);
3615 		mutex_unlock(&sctx->wr_lock);
3616 
3617 		wait_event(sctx->list_wait,
3618 			   atomic_read(&sctx->bios_in_flight) == 0);
3619 
3620 		scrub_pause_on(fs_info);
3621 
3622 		/*
3623 		 * must be called before we decrease @scrub_paused.
3624 		 * make sure we don't block transaction commit while
3625 		 * we are waiting pending workers finished.
3626 		 */
3627 		wait_event(sctx->list_wait,
3628 			   atomic_read(&sctx->workers_pending) == 0);
3629 		sctx->flush_all_writes = false;
3630 
3631 		scrub_pause_off(fs_info);
3632 
3633 		down_write(&dev_replace->rwsem);
3634 		dev_replace->cursor_left = dev_replace->cursor_right;
3635 		dev_replace->item_needs_writeback = 1;
3636 		up_write(&dev_replace->rwsem);
3637 
3638 		if (ro_set)
3639 			btrfs_dec_block_group_ro(cache);
3640 
3641 		/*
3642 		 * We might have prevented the cleaner kthread from deleting
3643 		 * this block group if it was already unused because we raced
3644 		 * and set it to RO mode first. So add it back to the unused
3645 		 * list, otherwise it might not ever be deleted unless a manual
3646 		 * balance is triggered or it becomes used and unused again.
3647 		 */
3648 		spin_lock(&cache->lock);
3649 		if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3650 		    cache->used == 0) {
3651 			spin_unlock(&cache->lock);
3652 			if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3653 				btrfs_discard_queue_work(&fs_info->discard_ctl,
3654 							 cache);
3655 			else
3656 				btrfs_mark_bg_unused(cache);
3657 		} else {
3658 			spin_unlock(&cache->lock);
3659 		}
3660 
3661 		btrfs_unfreeze_block_group(cache);
3662 		btrfs_put_block_group(cache);
3663 		if (ret)
3664 			break;
3665 		if (sctx->is_dev_replace &&
3666 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3667 			ret = -EIO;
3668 			break;
3669 		}
3670 		if (sctx->stat.malloc_errors > 0) {
3671 			ret = -ENOMEM;
3672 			break;
3673 		}
3674 skip:
3675 		key.offset = found_key.offset + length;
3676 		btrfs_release_path(path);
3677 	}
3678 
3679 	btrfs_free_path(path);
3680 
3681 	return ret;
3682 }
3683 
3684 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3685 					   struct btrfs_device *scrub_dev)
3686 {
3687 	int	i;
3688 	u64	bytenr;
3689 	u64	gen;
3690 	int	ret;
3691 	struct btrfs_fs_info *fs_info = sctx->fs_info;
3692 
3693 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3694 		return -EROFS;
3695 
3696 	/* Seed devices of a new filesystem has their own generation. */
3697 	if (scrub_dev->fs_devices != fs_info->fs_devices)
3698 		gen = scrub_dev->generation;
3699 	else
3700 		gen = fs_info->last_trans_committed;
3701 
3702 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3703 		bytenr = btrfs_sb_offset(i);
3704 		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3705 		    scrub_dev->commit_total_bytes)
3706 			break;
3707 
3708 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3709 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3710 				  NULL, 1, bytenr);
3711 		if (ret)
3712 			return ret;
3713 	}
3714 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3715 
3716 	return 0;
3717 }
3718 
3719 /*
3720  * get a reference count on fs_info->scrub_workers. start worker if necessary
3721  */
3722 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3723 						int is_dev_replace)
3724 {
3725 	unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3726 	int max_active = fs_info->thread_pool_size;
3727 
3728 	lockdep_assert_held(&fs_info->scrub_lock);
3729 
3730 	if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3731 		ASSERT(fs_info->scrub_workers == NULL);
3732 		fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3733 				flags, is_dev_replace ? 1 : max_active, 4);
3734 		if (!fs_info->scrub_workers)
3735 			goto fail_scrub_workers;
3736 
3737 		ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3738 		fs_info->scrub_wr_completion_workers =
3739 			btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3740 					      max_active, 2);
3741 		if (!fs_info->scrub_wr_completion_workers)
3742 			goto fail_scrub_wr_completion_workers;
3743 
3744 		ASSERT(fs_info->scrub_parity_workers == NULL);
3745 		fs_info->scrub_parity_workers =
3746 			btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3747 					      max_active, 2);
3748 		if (!fs_info->scrub_parity_workers)
3749 			goto fail_scrub_parity_workers;
3750 
3751 		refcount_set(&fs_info->scrub_workers_refcnt, 1);
3752 	} else {
3753 		refcount_inc(&fs_info->scrub_workers_refcnt);
3754 	}
3755 	return 0;
3756 
3757 fail_scrub_parity_workers:
3758 	btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3759 fail_scrub_wr_completion_workers:
3760 	btrfs_destroy_workqueue(fs_info->scrub_workers);
3761 fail_scrub_workers:
3762 	return -ENOMEM;
3763 }
3764 
3765 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3766 		    u64 end, struct btrfs_scrub_progress *progress,
3767 		    int readonly, int is_dev_replace)
3768 {
3769 	struct scrub_ctx *sctx;
3770 	int ret;
3771 	struct btrfs_device *dev;
3772 	unsigned int nofs_flag;
3773 	struct btrfs_workqueue *scrub_workers = NULL;
3774 	struct btrfs_workqueue *scrub_wr_comp = NULL;
3775 	struct btrfs_workqueue *scrub_parity = NULL;
3776 
3777 	if (btrfs_fs_closing(fs_info))
3778 		return -EAGAIN;
3779 
3780 	if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3781 		/*
3782 		 * in this case scrub is unable to calculate the checksum
3783 		 * the way scrub is implemented. Do not handle this
3784 		 * situation at all because it won't ever happen.
3785 		 */
3786 		btrfs_err(fs_info,
3787 			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3788 		       fs_info->nodesize,
3789 		       BTRFS_STRIPE_LEN);
3790 		return -EINVAL;
3791 	}
3792 
3793 	if (fs_info->sectorsize != PAGE_SIZE) {
3794 		/* not supported for data w/o checksums */
3795 		btrfs_err_rl(fs_info,
3796 			   "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3797 		       fs_info->sectorsize, PAGE_SIZE);
3798 		return -EINVAL;
3799 	}
3800 
3801 	if (fs_info->nodesize >
3802 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3803 	    fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3804 		/*
3805 		 * would exhaust the array bounds of pagev member in
3806 		 * struct scrub_block
3807 		 */
3808 		btrfs_err(fs_info,
3809 			  "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3810 		       fs_info->nodesize,
3811 		       SCRUB_MAX_PAGES_PER_BLOCK,
3812 		       fs_info->sectorsize,
3813 		       SCRUB_MAX_PAGES_PER_BLOCK);
3814 		return -EINVAL;
3815 	}
3816 
3817 	/* Allocate outside of device_list_mutex */
3818 	sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3819 	if (IS_ERR(sctx))
3820 		return PTR_ERR(sctx);
3821 
3822 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3823 	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3824 	if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3825 		     !is_dev_replace)) {
3826 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3827 		ret = -ENODEV;
3828 		goto out_free_ctx;
3829 	}
3830 
3831 	if (!is_dev_replace && !readonly &&
3832 	    !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3833 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3834 		btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3835 				rcu_str_deref(dev->name));
3836 		ret = -EROFS;
3837 		goto out_free_ctx;
3838 	}
3839 
3840 	mutex_lock(&fs_info->scrub_lock);
3841 	if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3842 	    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3843 		mutex_unlock(&fs_info->scrub_lock);
3844 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3845 		ret = -EIO;
3846 		goto out_free_ctx;
3847 	}
3848 
3849 	down_read(&fs_info->dev_replace.rwsem);
3850 	if (dev->scrub_ctx ||
3851 	    (!is_dev_replace &&
3852 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3853 		up_read(&fs_info->dev_replace.rwsem);
3854 		mutex_unlock(&fs_info->scrub_lock);
3855 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3856 		ret = -EINPROGRESS;
3857 		goto out_free_ctx;
3858 	}
3859 	up_read(&fs_info->dev_replace.rwsem);
3860 
3861 	ret = scrub_workers_get(fs_info, is_dev_replace);
3862 	if (ret) {
3863 		mutex_unlock(&fs_info->scrub_lock);
3864 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3865 		goto out_free_ctx;
3866 	}
3867 
3868 	sctx->readonly = readonly;
3869 	dev->scrub_ctx = sctx;
3870 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3871 
3872 	/*
3873 	 * checking @scrub_pause_req here, we can avoid
3874 	 * race between committing transaction and scrubbing.
3875 	 */
3876 	__scrub_blocked_if_needed(fs_info);
3877 	atomic_inc(&fs_info->scrubs_running);
3878 	mutex_unlock(&fs_info->scrub_lock);
3879 
3880 	/*
3881 	 * In order to avoid deadlock with reclaim when there is a transaction
3882 	 * trying to pause scrub, make sure we use GFP_NOFS for all the
3883 	 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3884 	 * invoked by our callees. The pausing request is done when the
3885 	 * transaction commit starts, and it blocks the transaction until scrub
3886 	 * is paused (done at specific points at scrub_stripe() or right above
3887 	 * before incrementing fs_info->scrubs_running).
3888 	 */
3889 	nofs_flag = memalloc_nofs_save();
3890 	if (!is_dev_replace) {
3891 		btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3892 		/*
3893 		 * by holding device list mutex, we can
3894 		 * kick off writing super in log tree sync.
3895 		 */
3896 		mutex_lock(&fs_info->fs_devices->device_list_mutex);
3897 		ret = scrub_supers(sctx, dev);
3898 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3899 	}
3900 
3901 	if (!ret)
3902 		ret = scrub_enumerate_chunks(sctx, dev, start, end);
3903 	memalloc_nofs_restore(nofs_flag);
3904 
3905 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3906 	atomic_dec(&fs_info->scrubs_running);
3907 	wake_up(&fs_info->scrub_pause_wait);
3908 
3909 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3910 
3911 	if (progress)
3912 		memcpy(progress, &sctx->stat, sizeof(*progress));
3913 
3914 	if (!is_dev_replace)
3915 		btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3916 			ret ? "not finished" : "finished", devid, ret);
3917 
3918 	mutex_lock(&fs_info->scrub_lock);
3919 	dev->scrub_ctx = NULL;
3920 	if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3921 		scrub_workers = fs_info->scrub_workers;
3922 		scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3923 		scrub_parity = fs_info->scrub_parity_workers;
3924 
3925 		fs_info->scrub_workers = NULL;
3926 		fs_info->scrub_wr_completion_workers = NULL;
3927 		fs_info->scrub_parity_workers = NULL;
3928 	}
3929 	mutex_unlock(&fs_info->scrub_lock);
3930 
3931 	btrfs_destroy_workqueue(scrub_workers);
3932 	btrfs_destroy_workqueue(scrub_wr_comp);
3933 	btrfs_destroy_workqueue(scrub_parity);
3934 	scrub_put_ctx(sctx);
3935 
3936 	return ret;
3937 
3938 out_free_ctx:
3939 	scrub_free_ctx(sctx);
3940 
3941 	return ret;
3942 }
3943 
3944 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3945 {
3946 	mutex_lock(&fs_info->scrub_lock);
3947 	atomic_inc(&fs_info->scrub_pause_req);
3948 	while (atomic_read(&fs_info->scrubs_paused) !=
3949 	       atomic_read(&fs_info->scrubs_running)) {
3950 		mutex_unlock(&fs_info->scrub_lock);
3951 		wait_event(fs_info->scrub_pause_wait,
3952 			   atomic_read(&fs_info->scrubs_paused) ==
3953 			   atomic_read(&fs_info->scrubs_running));
3954 		mutex_lock(&fs_info->scrub_lock);
3955 	}
3956 	mutex_unlock(&fs_info->scrub_lock);
3957 }
3958 
3959 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3960 {
3961 	atomic_dec(&fs_info->scrub_pause_req);
3962 	wake_up(&fs_info->scrub_pause_wait);
3963 }
3964 
3965 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3966 {
3967 	mutex_lock(&fs_info->scrub_lock);
3968 	if (!atomic_read(&fs_info->scrubs_running)) {
3969 		mutex_unlock(&fs_info->scrub_lock);
3970 		return -ENOTCONN;
3971 	}
3972 
3973 	atomic_inc(&fs_info->scrub_cancel_req);
3974 	while (atomic_read(&fs_info->scrubs_running)) {
3975 		mutex_unlock(&fs_info->scrub_lock);
3976 		wait_event(fs_info->scrub_pause_wait,
3977 			   atomic_read(&fs_info->scrubs_running) == 0);
3978 		mutex_lock(&fs_info->scrub_lock);
3979 	}
3980 	atomic_dec(&fs_info->scrub_cancel_req);
3981 	mutex_unlock(&fs_info->scrub_lock);
3982 
3983 	return 0;
3984 }
3985 
3986 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
3987 {
3988 	struct btrfs_fs_info *fs_info = dev->fs_info;
3989 	struct scrub_ctx *sctx;
3990 
3991 	mutex_lock(&fs_info->scrub_lock);
3992 	sctx = dev->scrub_ctx;
3993 	if (!sctx) {
3994 		mutex_unlock(&fs_info->scrub_lock);
3995 		return -ENOTCONN;
3996 	}
3997 	atomic_inc(&sctx->cancel_req);
3998 	while (dev->scrub_ctx) {
3999 		mutex_unlock(&fs_info->scrub_lock);
4000 		wait_event(fs_info->scrub_pause_wait,
4001 			   dev->scrub_ctx == NULL);
4002 		mutex_lock(&fs_info->scrub_lock);
4003 	}
4004 	mutex_unlock(&fs_info->scrub_lock);
4005 
4006 	return 0;
4007 }
4008 
4009 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4010 			 struct btrfs_scrub_progress *progress)
4011 {
4012 	struct btrfs_device *dev;
4013 	struct scrub_ctx *sctx = NULL;
4014 
4015 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
4016 	dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4017 	if (dev)
4018 		sctx = dev->scrub_ctx;
4019 	if (sctx)
4020 		memcpy(progress, &sctx->stat, sizeof(*progress));
4021 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4022 
4023 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4024 }
4025 
4026 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4027 			       u64 extent_logical, u64 extent_len,
4028 			       u64 *extent_physical,
4029 			       struct btrfs_device **extent_dev,
4030 			       int *extent_mirror_num)
4031 {
4032 	u64 mapped_length;
4033 	struct btrfs_bio *bbio = NULL;
4034 	int ret;
4035 
4036 	mapped_length = extent_len;
4037 	ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4038 			      &mapped_length, &bbio, 0);
4039 	if (ret || !bbio || mapped_length < extent_len ||
4040 	    !bbio->stripes[0].dev->bdev) {
4041 		btrfs_put_bbio(bbio);
4042 		return;
4043 	}
4044 
4045 	*extent_physical = bbio->stripes[0].physical;
4046 	*extent_mirror_num = bbio->mirror_num;
4047 	*extent_dev = bbio->stripes[0].dev;
4048 	btrfs_put_bbio(bbio);
4049 }
4050