xref: /linux/fs/btrfs/scrub.c (revision 93d90ad708b8da6efc0e487b66111aa9db7f70c7)
1 /*
2  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/blkdev.h>
20 #include <linux/ratelimit.h>
21 #include "ctree.h"
22 #include "volumes.h"
23 #include "disk-io.h"
24 #include "ordered-data.h"
25 #include "transaction.h"
26 #include "backref.h"
27 #include "extent_io.h"
28 #include "dev-replace.h"
29 #include "check-integrity.h"
30 #include "rcu-string.h"
31 #include "raid56.h"
32 
33 /*
34  * This is only the first step towards a full-features scrub. It reads all
35  * extent and super block and verifies the checksums. In case a bad checksum
36  * is found or the extent cannot be read, good data will be written back if
37  * any can be found.
38  *
39  * Future enhancements:
40  *  - In case an unrepairable extent is encountered, track which files are
41  *    affected and report them
42  *  - track and record media errors, throw out bad devices
43  *  - add a mode to also read unallocated space
44  */
45 
46 struct scrub_block;
47 struct scrub_ctx;
48 
49 /*
50  * the following three values only influence the performance.
51  * The last one configures the number of parallel and outstanding I/O
52  * operations. The first two values configure an upper limit for the number
53  * of (dynamically allocated) pages that are added to a bio.
54  */
55 #define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
56 #define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
57 #define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
58 
59 /*
60  * the following value times PAGE_SIZE needs to be large enough to match the
61  * largest node/leaf/sector size that shall be supported.
62  * Values larger than BTRFS_STRIPE_LEN are not supported.
63  */
64 #define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
65 
66 struct scrub_recover {
67 	atomic_t		refs;
68 	struct btrfs_bio	*bbio;
69 	u64			*raid_map;
70 	u64			map_length;
71 };
72 
73 struct scrub_page {
74 	struct scrub_block	*sblock;
75 	struct page		*page;
76 	struct btrfs_device	*dev;
77 	struct list_head	list;
78 	u64			flags;  /* extent flags */
79 	u64			generation;
80 	u64			logical;
81 	u64			physical;
82 	u64			physical_for_dev_replace;
83 	atomic_t		ref_count;
84 	struct {
85 		unsigned int	mirror_num:8;
86 		unsigned int	have_csum:1;
87 		unsigned int	io_error:1;
88 	};
89 	u8			csum[BTRFS_CSUM_SIZE];
90 
91 	struct scrub_recover	*recover;
92 };
93 
94 struct scrub_bio {
95 	int			index;
96 	struct scrub_ctx	*sctx;
97 	struct btrfs_device	*dev;
98 	struct bio		*bio;
99 	int			err;
100 	u64			logical;
101 	u64			physical;
102 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
103 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
104 #else
105 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
106 #endif
107 	int			page_count;
108 	int			next_free;
109 	struct btrfs_work	work;
110 };
111 
112 struct scrub_block {
113 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
114 	int			page_count;
115 	atomic_t		outstanding_pages;
116 	atomic_t		ref_count; /* free mem on transition to zero */
117 	struct scrub_ctx	*sctx;
118 	struct scrub_parity	*sparity;
119 	struct {
120 		unsigned int	header_error:1;
121 		unsigned int	checksum_error:1;
122 		unsigned int	no_io_error_seen:1;
123 		unsigned int	generation_error:1; /* also sets header_error */
124 
125 		/* The following is for the data used to check parity */
126 		/* It is for the data with checksum */
127 		unsigned int	data_corrected:1;
128 	};
129 };
130 
131 /* Used for the chunks with parity stripe such RAID5/6 */
132 struct scrub_parity {
133 	struct scrub_ctx	*sctx;
134 
135 	struct btrfs_device	*scrub_dev;
136 
137 	u64			logic_start;
138 
139 	u64			logic_end;
140 
141 	int			nsectors;
142 
143 	int			stripe_len;
144 
145 	atomic_t		ref_count;
146 
147 	struct list_head	spages;
148 
149 	/* Work of parity check and repair */
150 	struct btrfs_work	work;
151 
152 	/* Mark the parity blocks which have data */
153 	unsigned long		*dbitmap;
154 
155 	/*
156 	 * Mark the parity blocks which have data, but errors happen when
157 	 * read data or check data
158 	 */
159 	unsigned long		*ebitmap;
160 
161 	unsigned long		bitmap[0];
162 };
163 
164 struct scrub_wr_ctx {
165 	struct scrub_bio *wr_curr_bio;
166 	struct btrfs_device *tgtdev;
167 	int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
168 	atomic_t flush_all_writes;
169 	struct mutex wr_lock;
170 };
171 
172 struct scrub_ctx {
173 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
174 	struct btrfs_root	*dev_root;
175 	int			first_free;
176 	int			curr;
177 	atomic_t		bios_in_flight;
178 	atomic_t		workers_pending;
179 	spinlock_t		list_lock;
180 	wait_queue_head_t	list_wait;
181 	u16			csum_size;
182 	struct list_head	csum_list;
183 	atomic_t		cancel_req;
184 	int			readonly;
185 	int			pages_per_rd_bio;
186 	u32			sectorsize;
187 	u32			nodesize;
188 
189 	int			is_dev_replace;
190 	struct scrub_wr_ctx	wr_ctx;
191 
192 	/*
193 	 * statistics
194 	 */
195 	struct btrfs_scrub_progress stat;
196 	spinlock_t		stat_lock;
197 };
198 
199 struct scrub_fixup_nodatasum {
200 	struct scrub_ctx	*sctx;
201 	struct btrfs_device	*dev;
202 	u64			logical;
203 	struct btrfs_root	*root;
204 	struct btrfs_work	work;
205 	int			mirror_num;
206 };
207 
208 struct scrub_nocow_inode {
209 	u64			inum;
210 	u64			offset;
211 	u64			root;
212 	struct list_head	list;
213 };
214 
215 struct scrub_copy_nocow_ctx {
216 	struct scrub_ctx	*sctx;
217 	u64			logical;
218 	u64			len;
219 	int			mirror_num;
220 	u64			physical_for_dev_replace;
221 	struct list_head	inodes;
222 	struct btrfs_work	work;
223 };
224 
225 struct scrub_warning {
226 	struct btrfs_path	*path;
227 	u64			extent_item_size;
228 	const char		*errstr;
229 	sector_t		sector;
230 	u64			logical;
231 	struct btrfs_device	*dev;
232 };
233 
234 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
235 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
236 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
237 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
238 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
239 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
240 				     struct btrfs_fs_info *fs_info,
241 				     struct scrub_block *original_sblock,
242 				     u64 length, u64 logical,
243 				     struct scrub_block *sblocks_for_recheck);
244 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
245 				struct scrub_block *sblock, int is_metadata,
246 				int have_csum, u8 *csum, u64 generation,
247 				u16 csum_size, int retry_failed_mirror);
248 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
249 					 struct scrub_block *sblock,
250 					 int is_metadata, int have_csum,
251 					 const u8 *csum, u64 generation,
252 					 u16 csum_size);
253 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
254 					     struct scrub_block *sblock_good,
255 					     int force_write);
256 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
257 					    struct scrub_block *sblock_good,
258 					    int page_num, int force_write);
259 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
260 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
261 					   int page_num);
262 static int scrub_checksum_data(struct scrub_block *sblock);
263 static int scrub_checksum_tree_block(struct scrub_block *sblock);
264 static int scrub_checksum_super(struct scrub_block *sblock);
265 static void scrub_block_get(struct scrub_block *sblock);
266 static void scrub_block_put(struct scrub_block *sblock);
267 static void scrub_page_get(struct scrub_page *spage);
268 static void scrub_page_put(struct scrub_page *spage);
269 static void scrub_parity_get(struct scrub_parity *sparity);
270 static void scrub_parity_put(struct scrub_parity *sparity);
271 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
272 				    struct scrub_page *spage);
273 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
274 		       u64 physical, struct btrfs_device *dev, u64 flags,
275 		       u64 gen, int mirror_num, u8 *csum, int force,
276 		       u64 physical_for_dev_replace);
277 static void scrub_bio_end_io(struct bio *bio, int err);
278 static void scrub_bio_end_io_worker(struct btrfs_work *work);
279 static void scrub_block_complete(struct scrub_block *sblock);
280 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
281 			       u64 extent_logical, u64 extent_len,
282 			       u64 *extent_physical,
283 			       struct btrfs_device **extent_dev,
284 			       int *extent_mirror_num);
285 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
286 			      struct scrub_wr_ctx *wr_ctx,
287 			      struct btrfs_fs_info *fs_info,
288 			      struct btrfs_device *dev,
289 			      int is_dev_replace);
290 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
291 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
292 				    struct scrub_page *spage);
293 static void scrub_wr_submit(struct scrub_ctx *sctx);
294 static void scrub_wr_bio_end_io(struct bio *bio, int err);
295 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
296 static int write_page_nocow(struct scrub_ctx *sctx,
297 			    u64 physical_for_dev_replace, struct page *page);
298 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
299 				      struct scrub_copy_nocow_ctx *ctx);
300 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
301 			    int mirror_num, u64 physical_for_dev_replace);
302 static void copy_nocow_pages_worker(struct btrfs_work *work);
303 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
305 
306 
307 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
308 {
309 	atomic_inc(&sctx->bios_in_flight);
310 }
311 
312 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
313 {
314 	atomic_dec(&sctx->bios_in_flight);
315 	wake_up(&sctx->list_wait);
316 }
317 
318 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
319 {
320 	while (atomic_read(&fs_info->scrub_pause_req)) {
321 		mutex_unlock(&fs_info->scrub_lock);
322 		wait_event(fs_info->scrub_pause_wait,
323 		   atomic_read(&fs_info->scrub_pause_req) == 0);
324 		mutex_lock(&fs_info->scrub_lock);
325 	}
326 }
327 
328 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
329 {
330 	atomic_inc(&fs_info->scrubs_paused);
331 	wake_up(&fs_info->scrub_pause_wait);
332 
333 	mutex_lock(&fs_info->scrub_lock);
334 	__scrub_blocked_if_needed(fs_info);
335 	atomic_dec(&fs_info->scrubs_paused);
336 	mutex_unlock(&fs_info->scrub_lock);
337 
338 	wake_up(&fs_info->scrub_pause_wait);
339 }
340 
341 /*
342  * used for workers that require transaction commits (i.e., for the
343  * NOCOW case)
344  */
345 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
346 {
347 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
348 
349 	/*
350 	 * increment scrubs_running to prevent cancel requests from
351 	 * completing as long as a worker is running. we must also
352 	 * increment scrubs_paused to prevent deadlocking on pause
353 	 * requests used for transactions commits (as the worker uses a
354 	 * transaction context). it is safe to regard the worker
355 	 * as paused for all matters practical. effectively, we only
356 	 * avoid cancellation requests from completing.
357 	 */
358 	mutex_lock(&fs_info->scrub_lock);
359 	atomic_inc(&fs_info->scrubs_running);
360 	atomic_inc(&fs_info->scrubs_paused);
361 	mutex_unlock(&fs_info->scrub_lock);
362 
363 	/*
364 	 * check if @scrubs_running=@scrubs_paused condition
365 	 * inside wait_event() is not an atomic operation.
366 	 * which means we may inc/dec @scrub_running/paused
367 	 * at any time. Let's wake up @scrub_pause_wait as
368 	 * much as we can to let commit transaction blocked less.
369 	 */
370 	wake_up(&fs_info->scrub_pause_wait);
371 
372 	atomic_inc(&sctx->workers_pending);
373 }
374 
375 /* used for workers that require transaction commits */
376 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
377 {
378 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
379 
380 	/*
381 	 * see scrub_pending_trans_workers_inc() why we're pretending
382 	 * to be paused in the scrub counters
383 	 */
384 	mutex_lock(&fs_info->scrub_lock);
385 	atomic_dec(&fs_info->scrubs_running);
386 	atomic_dec(&fs_info->scrubs_paused);
387 	mutex_unlock(&fs_info->scrub_lock);
388 	atomic_dec(&sctx->workers_pending);
389 	wake_up(&fs_info->scrub_pause_wait);
390 	wake_up(&sctx->list_wait);
391 }
392 
393 static void scrub_free_csums(struct scrub_ctx *sctx)
394 {
395 	while (!list_empty(&sctx->csum_list)) {
396 		struct btrfs_ordered_sum *sum;
397 		sum = list_first_entry(&sctx->csum_list,
398 				       struct btrfs_ordered_sum, list);
399 		list_del(&sum->list);
400 		kfree(sum);
401 	}
402 }
403 
404 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
405 {
406 	int i;
407 
408 	if (!sctx)
409 		return;
410 
411 	scrub_free_wr_ctx(&sctx->wr_ctx);
412 
413 	/* this can happen when scrub is cancelled */
414 	if (sctx->curr != -1) {
415 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
416 
417 		for (i = 0; i < sbio->page_count; i++) {
418 			WARN_ON(!sbio->pagev[i]->page);
419 			scrub_block_put(sbio->pagev[i]->sblock);
420 		}
421 		bio_put(sbio->bio);
422 	}
423 
424 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
425 		struct scrub_bio *sbio = sctx->bios[i];
426 
427 		if (!sbio)
428 			break;
429 		kfree(sbio);
430 	}
431 
432 	scrub_free_csums(sctx);
433 	kfree(sctx);
434 }
435 
436 static noinline_for_stack
437 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
438 {
439 	struct scrub_ctx *sctx;
440 	int		i;
441 	struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
442 	int pages_per_rd_bio;
443 	int ret;
444 
445 	/*
446 	 * the setting of pages_per_rd_bio is correct for scrub but might
447 	 * be wrong for the dev_replace code where we might read from
448 	 * different devices in the initial huge bios. However, that
449 	 * code is able to correctly handle the case when adding a page
450 	 * to a bio fails.
451 	 */
452 	if (dev->bdev)
453 		pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
454 					 bio_get_nr_vecs(dev->bdev));
455 	else
456 		pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
457 	sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
458 	if (!sctx)
459 		goto nomem;
460 	sctx->is_dev_replace = is_dev_replace;
461 	sctx->pages_per_rd_bio = pages_per_rd_bio;
462 	sctx->curr = -1;
463 	sctx->dev_root = dev->dev_root;
464 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
465 		struct scrub_bio *sbio;
466 
467 		sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
468 		if (!sbio)
469 			goto nomem;
470 		sctx->bios[i] = sbio;
471 
472 		sbio->index = i;
473 		sbio->sctx = sctx;
474 		sbio->page_count = 0;
475 		btrfs_init_work(&sbio->work, btrfs_scrub_helper,
476 				scrub_bio_end_io_worker, NULL, NULL);
477 
478 		if (i != SCRUB_BIOS_PER_SCTX - 1)
479 			sctx->bios[i]->next_free = i + 1;
480 		else
481 			sctx->bios[i]->next_free = -1;
482 	}
483 	sctx->first_free = 0;
484 	sctx->nodesize = dev->dev_root->nodesize;
485 	sctx->sectorsize = dev->dev_root->sectorsize;
486 	atomic_set(&sctx->bios_in_flight, 0);
487 	atomic_set(&sctx->workers_pending, 0);
488 	atomic_set(&sctx->cancel_req, 0);
489 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
490 	INIT_LIST_HEAD(&sctx->csum_list);
491 
492 	spin_lock_init(&sctx->list_lock);
493 	spin_lock_init(&sctx->stat_lock);
494 	init_waitqueue_head(&sctx->list_wait);
495 
496 	ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
497 				 fs_info->dev_replace.tgtdev, is_dev_replace);
498 	if (ret) {
499 		scrub_free_ctx(sctx);
500 		return ERR_PTR(ret);
501 	}
502 	return sctx;
503 
504 nomem:
505 	scrub_free_ctx(sctx);
506 	return ERR_PTR(-ENOMEM);
507 }
508 
509 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
510 				     void *warn_ctx)
511 {
512 	u64 isize;
513 	u32 nlink;
514 	int ret;
515 	int i;
516 	struct extent_buffer *eb;
517 	struct btrfs_inode_item *inode_item;
518 	struct scrub_warning *swarn = warn_ctx;
519 	struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
520 	struct inode_fs_paths *ipath = NULL;
521 	struct btrfs_root *local_root;
522 	struct btrfs_key root_key;
523 
524 	root_key.objectid = root;
525 	root_key.type = BTRFS_ROOT_ITEM_KEY;
526 	root_key.offset = (u64)-1;
527 	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
528 	if (IS_ERR(local_root)) {
529 		ret = PTR_ERR(local_root);
530 		goto err;
531 	}
532 
533 	ret = inode_item_info(inum, 0, local_root, swarn->path);
534 	if (ret) {
535 		btrfs_release_path(swarn->path);
536 		goto err;
537 	}
538 
539 	eb = swarn->path->nodes[0];
540 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
541 					struct btrfs_inode_item);
542 	isize = btrfs_inode_size(eb, inode_item);
543 	nlink = btrfs_inode_nlink(eb, inode_item);
544 	btrfs_release_path(swarn->path);
545 
546 	ipath = init_ipath(4096, local_root, swarn->path);
547 	if (IS_ERR(ipath)) {
548 		ret = PTR_ERR(ipath);
549 		ipath = NULL;
550 		goto err;
551 	}
552 	ret = paths_from_inode(inum, ipath);
553 
554 	if (ret < 0)
555 		goto err;
556 
557 	/*
558 	 * we deliberately ignore the bit ipath might have been too small to
559 	 * hold all of the paths here
560 	 */
561 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
562 		printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
563 			"%s, sector %llu, root %llu, inode %llu, offset %llu, "
564 			"length %llu, links %u (path: %s)\n", swarn->errstr,
565 			swarn->logical, rcu_str_deref(swarn->dev->name),
566 			(unsigned long long)swarn->sector, root, inum, offset,
567 			min(isize - offset, (u64)PAGE_SIZE), nlink,
568 			(char *)(unsigned long)ipath->fspath->val[i]);
569 
570 	free_ipath(ipath);
571 	return 0;
572 
573 err:
574 	printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
575 		"%s, sector %llu, root %llu, inode %llu, offset %llu: path "
576 		"resolving failed with ret=%d\n", swarn->errstr,
577 		swarn->logical, rcu_str_deref(swarn->dev->name),
578 		(unsigned long long)swarn->sector, root, inum, offset, ret);
579 
580 	free_ipath(ipath);
581 	return 0;
582 }
583 
584 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
585 {
586 	struct btrfs_device *dev;
587 	struct btrfs_fs_info *fs_info;
588 	struct btrfs_path *path;
589 	struct btrfs_key found_key;
590 	struct extent_buffer *eb;
591 	struct btrfs_extent_item *ei;
592 	struct scrub_warning swarn;
593 	unsigned long ptr = 0;
594 	u64 extent_item_pos;
595 	u64 flags = 0;
596 	u64 ref_root;
597 	u32 item_size;
598 	u8 ref_level;
599 	int ret;
600 
601 	WARN_ON(sblock->page_count < 1);
602 	dev = sblock->pagev[0]->dev;
603 	fs_info = sblock->sctx->dev_root->fs_info;
604 
605 	path = btrfs_alloc_path();
606 	if (!path)
607 		return;
608 
609 	swarn.sector = (sblock->pagev[0]->physical) >> 9;
610 	swarn.logical = sblock->pagev[0]->logical;
611 	swarn.errstr = errstr;
612 	swarn.dev = NULL;
613 
614 	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
615 				  &flags);
616 	if (ret < 0)
617 		goto out;
618 
619 	extent_item_pos = swarn.logical - found_key.objectid;
620 	swarn.extent_item_size = found_key.offset;
621 
622 	eb = path->nodes[0];
623 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
624 	item_size = btrfs_item_size_nr(eb, path->slots[0]);
625 
626 	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
627 		do {
628 			ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
629 						      item_size, &ref_root,
630 						      &ref_level);
631 			printk_in_rcu(KERN_WARNING
632 				"BTRFS: %s at logical %llu on dev %s, "
633 				"sector %llu: metadata %s (level %d) in tree "
634 				"%llu\n", errstr, swarn.logical,
635 				rcu_str_deref(dev->name),
636 				(unsigned long long)swarn.sector,
637 				ref_level ? "node" : "leaf",
638 				ret < 0 ? -1 : ref_level,
639 				ret < 0 ? -1 : ref_root);
640 		} while (ret != 1);
641 		btrfs_release_path(path);
642 	} else {
643 		btrfs_release_path(path);
644 		swarn.path = path;
645 		swarn.dev = dev;
646 		iterate_extent_inodes(fs_info, found_key.objectid,
647 					extent_item_pos, 1,
648 					scrub_print_warning_inode, &swarn);
649 	}
650 
651 out:
652 	btrfs_free_path(path);
653 }
654 
655 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
656 {
657 	struct page *page = NULL;
658 	unsigned long index;
659 	struct scrub_fixup_nodatasum *fixup = fixup_ctx;
660 	int ret;
661 	int corrected = 0;
662 	struct btrfs_key key;
663 	struct inode *inode = NULL;
664 	struct btrfs_fs_info *fs_info;
665 	u64 end = offset + PAGE_SIZE - 1;
666 	struct btrfs_root *local_root;
667 	int srcu_index;
668 
669 	key.objectid = root;
670 	key.type = BTRFS_ROOT_ITEM_KEY;
671 	key.offset = (u64)-1;
672 
673 	fs_info = fixup->root->fs_info;
674 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
675 
676 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
677 	if (IS_ERR(local_root)) {
678 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
679 		return PTR_ERR(local_root);
680 	}
681 
682 	key.type = BTRFS_INODE_ITEM_KEY;
683 	key.objectid = inum;
684 	key.offset = 0;
685 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
686 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
687 	if (IS_ERR(inode))
688 		return PTR_ERR(inode);
689 
690 	index = offset >> PAGE_CACHE_SHIFT;
691 
692 	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
693 	if (!page) {
694 		ret = -ENOMEM;
695 		goto out;
696 	}
697 
698 	if (PageUptodate(page)) {
699 		if (PageDirty(page)) {
700 			/*
701 			 * we need to write the data to the defect sector. the
702 			 * data that was in that sector is not in memory,
703 			 * because the page was modified. we must not write the
704 			 * modified page to that sector.
705 			 *
706 			 * TODO: what could be done here: wait for the delalloc
707 			 *       runner to write out that page (might involve
708 			 *       COW) and see whether the sector is still
709 			 *       referenced afterwards.
710 			 *
711 			 * For the meantime, we'll treat this error
712 			 * incorrectable, although there is a chance that a
713 			 * later scrub will find the bad sector again and that
714 			 * there's no dirty page in memory, then.
715 			 */
716 			ret = -EIO;
717 			goto out;
718 		}
719 		ret = repair_io_failure(inode, offset, PAGE_SIZE,
720 					fixup->logical, page,
721 					offset - page_offset(page),
722 					fixup->mirror_num);
723 		unlock_page(page);
724 		corrected = !ret;
725 	} else {
726 		/*
727 		 * we need to get good data first. the general readpage path
728 		 * will call repair_io_failure for us, we just have to make
729 		 * sure we read the bad mirror.
730 		 */
731 		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
732 					EXTENT_DAMAGED, GFP_NOFS);
733 		if (ret) {
734 			/* set_extent_bits should give proper error */
735 			WARN_ON(ret > 0);
736 			if (ret > 0)
737 				ret = -EFAULT;
738 			goto out;
739 		}
740 
741 		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
742 						btrfs_get_extent,
743 						fixup->mirror_num);
744 		wait_on_page_locked(page);
745 
746 		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
747 						end, EXTENT_DAMAGED, 0, NULL);
748 		if (!corrected)
749 			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
750 						EXTENT_DAMAGED, GFP_NOFS);
751 	}
752 
753 out:
754 	if (page)
755 		put_page(page);
756 
757 	iput(inode);
758 
759 	if (ret < 0)
760 		return ret;
761 
762 	if (ret == 0 && corrected) {
763 		/*
764 		 * we only need to call readpage for one of the inodes belonging
765 		 * to this extent. so make iterate_extent_inodes stop
766 		 */
767 		return 1;
768 	}
769 
770 	return -EIO;
771 }
772 
773 static void scrub_fixup_nodatasum(struct btrfs_work *work)
774 {
775 	int ret;
776 	struct scrub_fixup_nodatasum *fixup;
777 	struct scrub_ctx *sctx;
778 	struct btrfs_trans_handle *trans = NULL;
779 	struct btrfs_path *path;
780 	int uncorrectable = 0;
781 
782 	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
783 	sctx = fixup->sctx;
784 
785 	path = btrfs_alloc_path();
786 	if (!path) {
787 		spin_lock(&sctx->stat_lock);
788 		++sctx->stat.malloc_errors;
789 		spin_unlock(&sctx->stat_lock);
790 		uncorrectable = 1;
791 		goto out;
792 	}
793 
794 	trans = btrfs_join_transaction(fixup->root);
795 	if (IS_ERR(trans)) {
796 		uncorrectable = 1;
797 		goto out;
798 	}
799 
800 	/*
801 	 * the idea is to trigger a regular read through the standard path. we
802 	 * read a page from the (failed) logical address by specifying the
803 	 * corresponding copynum of the failed sector. thus, that readpage is
804 	 * expected to fail.
805 	 * that is the point where on-the-fly error correction will kick in
806 	 * (once it's finished) and rewrite the failed sector if a good copy
807 	 * can be found.
808 	 */
809 	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
810 						path, scrub_fixup_readpage,
811 						fixup);
812 	if (ret < 0) {
813 		uncorrectable = 1;
814 		goto out;
815 	}
816 	WARN_ON(ret != 1);
817 
818 	spin_lock(&sctx->stat_lock);
819 	++sctx->stat.corrected_errors;
820 	spin_unlock(&sctx->stat_lock);
821 
822 out:
823 	if (trans && !IS_ERR(trans))
824 		btrfs_end_transaction(trans, fixup->root);
825 	if (uncorrectable) {
826 		spin_lock(&sctx->stat_lock);
827 		++sctx->stat.uncorrectable_errors;
828 		spin_unlock(&sctx->stat_lock);
829 		btrfs_dev_replace_stats_inc(
830 			&sctx->dev_root->fs_info->dev_replace.
831 			num_uncorrectable_read_errors);
832 		printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
833 		    "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
834 			fixup->logical, rcu_str_deref(fixup->dev->name));
835 	}
836 
837 	btrfs_free_path(path);
838 	kfree(fixup);
839 
840 	scrub_pending_trans_workers_dec(sctx);
841 }
842 
843 static inline void scrub_get_recover(struct scrub_recover *recover)
844 {
845 	atomic_inc(&recover->refs);
846 }
847 
848 static inline void scrub_put_recover(struct scrub_recover *recover)
849 {
850 	if (atomic_dec_and_test(&recover->refs)) {
851 		kfree(recover->bbio);
852 		kfree(recover->raid_map);
853 		kfree(recover);
854 	}
855 }
856 
857 /*
858  * scrub_handle_errored_block gets called when either verification of the
859  * pages failed or the bio failed to read, e.g. with EIO. In the latter
860  * case, this function handles all pages in the bio, even though only one
861  * may be bad.
862  * The goal of this function is to repair the errored block by using the
863  * contents of one of the mirrors.
864  */
865 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
866 {
867 	struct scrub_ctx *sctx = sblock_to_check->sctx;
868 	struct btrfs_device *dev;
869 	struct btrfs_fs_info *fs_info;
870 	u64 length;
871 	u64 logical;
872 	u64 generation;
873 	unsigned int failed_mirror_index;
874 	unsigned int is_metadata;
875 	unsigned int have_csum;
876 	u8 *csum;
877 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
878 	struct scrub_block *sblock_bad;
879 	int ret;
880 	int mirror_index;
881 	int page_num;
882 	int success;
883 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
884 				      DEFAULT_RATELIMIT_BURST);
885 
886 	BUG_ON(sblock_to_check->page_count < 1);
887 	fs_info = sctx->dev_root->fs_info;
888 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
889 		/*
890 		 * if we find an error in a super block, we just report it.
891 		 * They will get written with the next transaction commit
892 		 * anyway
893 		 */
894 		spin_lock(&sctx->stat_lock);
895 		++sctx->stat.super_errors;
896 		spin_unlock(&sctx->stat_lock);
897 		return 0;
898 	}
899 	length = sblock_to_check->page_count * PAGE_SIZE;
900 	logical = sblock_to_check->pagev[0]->logical;
901 	generation = sblock_to_check->pagev[0]->generation;
902 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
903 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
904 	is_metadata = !(sblock_to_check->pagev[0]->flags &
905 			BTRFS_EXTENT_FLAG_DATA);
906 	have_csum = sblock_to_check->pagev[0]->have_csum;
907 	csum = sblock_to_check->pagev[0]->csum;
908 	dev = sblock_to_check->pagev[0]->dev;
909 
910 	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
911 		sblocks_for_recheck = NULL;
912 		goto nodatasum_case;
913 	}
914 
915 	/*
916 	 * read all mirrors one after the other. This includes to
917 	 * re-read the extent or metadata block that failed (that was
918 	 * the cause that this fixup code is called) another time,
919 	 * page by page this time in order to know which pages
920 	 * caused I/O errors and which ones are good (for all mirrors).
921 	 * It is the goal to handle the situation when more than one
922 	 * mirror contains I/O errors, but the errors do not
923 	 * overlap, i.e. the data can be repaired by selecting the
924 	 * pages from those mirrors without I/O error on the
925 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
926 	 * would be that mirror #1 has an I/O error on the first page,
927 	 * the second page is good, and mirror #2 has an I/O error on
928 	 * the second page, but the first page is good.
929 	 * Then the first page of the first mirror can be repaired by
930 	 * taking the first page of the second mirror, and the
931 	 * second page of the second mirror can be repaired by
932 	 * copying the contents of the 2nd page of the 1st mirror.
933 	 * One more note: if the pages of one mirror contain I/O
934 	 * errors, the checksum cannot be verified. In order to get
935 	 * the best data for repairing, the first attempt is to find
936 	 * a mirror without I/O errors and with a validated checksum.
937 	 * Only if this is not possible, the pages are picked from
938 	 * mirrors with I/O errors without considering the checksum.
939 	 * If the latter is the case, at the end, the checksum of the
940 	 * repaired area is verified in order to correctly maintain
941 	 * the statistics.
942 	 */
943 
944 	sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
945 				     sizeof(*sblocks_for_recheck),
946 				     GFP_NOFS);
947 	if (!sblocks_for_recheck) {
948 		spin_lock(&sctx->stat_lock);
949 		sctx->stat.malloc_errors++;
950 		sctx->stat.read_errors++;
951 		sctx->stat.uncorrectable_errors++;
952 		spin_unlock(&sctx->stat_lock);
953 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
954 		goto out;
955 	}
956 
957 	/* setup the context, map the logical blocks and alloc the pages */
958 	ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
959 					logical, sblocks_for_recheck);
960 	if (ret) {
961 		spin_lock(&sctx->stat_lock);
962 		sctx->stat.read_errors++;
963 		sctx->stat.uncorrectable_errors++;
964 		spin_unlock(&sctx->stat_lock);
965 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
966 		goto out;
967 	}
968 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
969 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
970 
971 	/* build and submit the bios for the failed mirror, check checksums */
972 	scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
973 			    csum, generation, sctx->csum_size, 1);
974 
975 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
976 	    sblock_bad->no_io_error_seen) {
977 		/*
978 		 * the error disappeared after reading page by page, or
979 		 * the area was part of a huge bio and other parts of the
980 		 * bio caused I/O errors, or the block layer merged several
981 		 * read requests into one and the error is caused by a
982 		 * different bio (usually one of the two latter cases is
983 		 * the cause)
984 		 */
985 		spin_lock(&sctx->stat_lock);
986 		sctx->stat.unverified_errors++;
987 		sblock_to_check->data_corrected = 1;
988 		spin_unlock(&sctx->stat_lock);
989 
990 		if (sctx->is_dev_replace)
991 			scrub_write_block_to_dev_replace(sblock_bad);
992 		goto out;
993 	}
994 
995 	if (!sblock_bad->no_io_error_seen) {
996 		spin_lock(&sctx->stat_lock);
997 		sctx->stat.read_errors++;
998 		spin_unlock(&sctx->stat_lock);
999 		if (__ratelimit(&_rs))
1000 			scrub_print_warning("i/o error", sblock_to_check);
1001 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1002 	} else if (sblock_bad->checksum_error) {
1003 		spin_lock(&sctx->stat_lock);
1004 		sctx->stat.csum_errors++;
1005 		spin_unlock(&sctx->stat_lock);
1006 		if (__ratelimit(&_rs))
1007 			scrub_print_warning("checksum error", sblock_to_check);
1008 		btrfs_dev_stat_inc_and_print(dev,
1009 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
1010 	} else if (sblock_bad->header_error) {
1011 		spin_lock(&sctx->stat_lock);
1012 		sctx->stat.verify_errors++;
1013 		spin_unlock(&sctx->stat_lock);
1014 		if (__ratelimit(&_rs))
1015 			scrub_print_warning("checksum/header error",
1016 					    sblock_to_check);
1017 		if (sblock_bad->generation_error)
1018 			btrfs_dev_stat_inc_and_print(dev,
1019 				BTRFS_DEV_STAT_GENERATION_ERRS);
1020 		else
1021 			btrfs_dev_stat_inc_and_print(dev,
1022 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1023 	}
1024 
1025 	if (sctx->readonly) {
1026 		ASSERT(!sctx->is_dev_replace);
1027 		goto out;
1028 	}
1029 
1030 	if (!is_metadata && !have_csum) {
1031 		struct scrub_fixup_nodatasum *fixup_nodatasum;
1032 
1033 nodatasum_case:
1034 		WARN_ON(sctx->is_dev_replace);
1035 
1036 		/*
1037 		 * !is_metadata and !have_csum, this means that the data
1038 		 * might not be COW'ed, that it might be modified
1039 		 * concurrently. The general strategy to work on the
1040 		 * commit root does not help in the case when COW is not
1041 		 * used.
1042 		 */
1043 		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1044 		if (!fixup_nodatasum)
1045 			goto did_not_correct_error;
1046 		fixup_nodatasum->sctx = sctx;
1047 		fixup_nodatasum->dev = dev;
1048 		fixup_nodatasum->logical = logical;
1049 		fixup_nodatasum->root = fs_info->extent_root;
1050 		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1051 		scrub_pending_trans_workers_inc(sctx);
1052 		btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1053 				scrub_fixup_nodatasum, NULL, NULL);
1054 		btrfs_queue_work(fs_info->scrub_workers,
1055 				 &fixup_nodatasum->work);
1056 		goto out;
1057 	}
1058 
1059 	/*
1060 	 * now build and submit the bios for the other mirrors, check
1061 	 * checksums.
1062 	 * First try to pick the mirror which is completely without I/O
1063 	 * errors and also does not have a checksum error.
1064 	 * If one is found, and if a checksum is present, the full block
1065 	 * that is known to contain an error is rewritten. Afterwards
1066 	 * the block is known to be corrected.
1067 	 * If a mirror is found which is completely correct, and no
1068 	 * checksum is present, only those pages are rewritten that had
1069 	 * an I/O error in the block to be repaired, since it cannot be
1070 	 * determined, which copy of the other pages is better (and it
1071 	 * could happen otherwise that a correct page would be
1072 	 * overwritten by a bad one).
1073 	 */
1074 	for (mirror_index = 0;
1075 	     mirror_index < BTRFS_MAX_MIRRORS &&
1076 	     sblocks_for_recheck[mirror_index].page_count > 0;
1077 	     mirror_index++) {
1078 		struct scrub_block *sblock_other;
1079 
1080 		if (mirror_index == failed_mirror_index)
1081 			continue;
1082 		sblock_other = sblocks_for_recheck + mirror_index;
1083 
1084 		/* build and submit the bios, check checksums */
1085 		scrub_recheck_block(fs_info, sblock_other, is_metadata,
1086 				    have_csum, csum, generation,
1087 				    sctx->csum_size, 0);
1088 
1089 		if (!sblock_other->header_error &&
1090 		    !sblock_other->checksum_error &&
1091 		    sblock_other->no_io_error_seen) {
1092 			if (sctx->is_dev_replace) {
1093 				scrub_write_block_to_dev_replace(sblock_other);
1094 			} else {
1095 				int force_write = is_metadata || have_csum;
1096 
1097 				ret = scrub_repair_block_from_good_copy(
1098 						sblock_bad, sblock_other,
1099 						force_write);
1100 			}
1101 			if (0 == ret)
1102 				goto corrected_error;
1103 		}
1104 	}
1105 
1106 	/*
1107 	 * for dev_replace, pick good pages and write to the target device.
1108 	 */
1109 	if (sctx->is_dev_replace) {
1110 		success = 1;
1111 		for (page_num = 0; page_num < sblock_bad->page_count;
1112 		     page_num++) {
1113 			int sub_success;
1114 
1115 			sub_success = 0;
1116 			for (mirror_index = 0;
1117 			     mirror_index < BTRFS_MAX_MIRRORS &&
1118 			     sblocks_for_recheck[mirror_index].page_count > 0;
1119 			     mirror_index++) {
1120 				struct scrub_block *sblock_other =
1121 					sblocks_for_recheck + mirror_index;
1122 				struct scrub_page *page_other =
1123 					sblock_other->pagev[page_num];
1124 
1125 				if (!page_other->io_error) {
1126 					ret = scrub_write_page_to_dev_replace(
1127 							sblock_other, page_num);
1128 					if (ret == 0) {
1129 						/* succeeded for this page */
1130 						sub_success = 1;
1131 						break;
1132 					} else {
1133 						btrfs_dev_replace_stats_inc(
1134 							&sctx->dev_root->
1135 							fs_info->dev_replace.
1136 							num_write_errors);
1137 					}
1138 				}
1139 			}
1140 
1141 			if (!sub_success) {
1142 				/*
1143 				 * did not find a mirror to fetch the page
1144 				 * from. scrub_write_page_to_dev_replace()
1145 				 * handles this case (page->io_error), by
1146 				 * filling the block with zeros before
1147 				 * submitting the write request
1148 				 */
1149 				success = 0;
1150 				ret = scrub_write_page_to_dev_replace(
1151 						sblock_bad, page_num);
1152 				if (ret)
1153 					btrfs_dev_replace_stats_inc(
1154 						&sctx->dev_root->fs_info->
1155 						dev_replace.num_write_errors);
1156 			}
1157 		}
1158 
1159 		goto out;
1160 	}
1161 
1162 	/*
1163 	 * for regular scrub, repair those pages that are errored.
1164 	 * In case of I/O errors in the area that is supposed to be
1165 	 * repaired, continue by picking good copies of those pages.
1166 	 * Select the good pages from mirrors to rewrite bad pages from
1167 	 * the area to fix. Afterwards verify the checksum of the block
1168 	 * that is supposed to be repaired. This verification step is
1169 	 * only done for the purpose of statistic counting and for the
1170 	 * final scrub report, whether errors remain.
1171 	 * A perfect algorithm could make use of the checksum and try
1172 	 * all possible combinations of pages from the different mirrors
1173 	 * until the checksum verification succeeds. For example, when
1174 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1175 	 * of mirror #2 is readable but the final checksum test fails,
1176 	 * then the 2nd page of mirror #3 could be tried, whether now
1177 	 * the final checksum succeedes. But this would be a rare
1178 	 * exception and is therefore not implemented. At least it is
1179 	 * avoided that the good copy is overwritten.
1180 	 * A more useful improvement would be to pick the sectors
1181 	 * without I/O error based on sector sizes (512 bytes on legacy
1182 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1183 	 * mirror could be repaired by taking 512 byte of a different
1184 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1185 	 * area are unreadable.
1186 	 */
1187 
1188 	/* can only fix I/O errors from here on */
1189 	if (sblock_bad->no_io_error_seen)
1190 		goto did_not_correct_error;
1191 
1192 	success = 1;
1193 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1194 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1195 
1196 		if (!page_bad->io_error)
1197 			continue;
1198 
1199 		for (mirror_index = 0;
1200 		     mirror_index < BTRFS_MAX_MIRRORS &&
1201 		     sblocks_for_recheck[mirror_index].page_count > 0;
1202 		     mirror_index++) {
1203 			struct scrub_block *sblock_other = sblocks_for_recheck +
1204 							   mirror_index;
1205 			struct scrub_page *page_other = sblock_other->pagev[
1206 							page_num];
1207 
1208 			if (!page_other->io_error) {
1209 				ret = scrub_repair_page_from_good_copy(
1210 					sblock_bad, sblock_other, page_num, 0);
1211 				if (0 == ret) {
1212 					page_bad->io_error = 0;
1213 					break; /* succeeded for this page */
1214 				}
1215 			}
1216 		}
1217 
1218 		if (page_bad->io_error) {
1219 			/* did not find a mirror to copy the page from */
1220 			success = 0;
1221 		}
1222 	}
1223 
1224 	if (success) {
1225 		if (is_metadata || have_csum) {
1226 			/*
1227 			 * need to verify the checksum now that all
1228 			 * sectors on disk are repaired (the write
1229 			 * request for data to be repaired is on its way).
1230 			 * Just be lazy and use scrub_recheck_block()
1231 			 * which re-reads the data before the checksum
1232 			 * is verified, but most likely the data comes out
1233 			 * of the page cache.
1234 			 */
1235 			scrub_recheck_block(fs_info, sblock_bad,
1236 					    is_metadata, have_csum, csum,
1237 					    generation, sctx->csum_size, 1);
1238 			if (!sblock_bad->header_error &&
1239 			    !sblock_bad->checksum_error &&
1240 			    sblock_bad->no_io_error_seen)
1241 				goto corrected_error;
1242 			else
1243 				goto did_not_correct_error;
1244 		} else {
1245 corrected_error:
1246 			spin_lock(&sctx->stat_lock);
1247 			sctx->stat.corrected_errors++;
1248 			sblock_to_check->data_corrected = 1;
1249 			spin_unlock(&sctx->stat_lock);
1250 			printk_ratelimited_in_rcu(KERN_ERR
1251 				"BTRFS: fixed up error at logical %llu on dev %s\n",
1252 				logical, rcu_str_deref(dev->name));
1253 		}
1254 	} else {
1255 did_not_correct_error:
1256 		spin_lock(&sctx->stat_lock);
1257 		sctx->stat.uncorrectable_errors++;
1258 		spin_unlock(&sctx->stat_lock);
1259 		printk_ratelimited_in_rcu(KERN_ERR
1260 			"BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1261 			logical, rcu_str_deref(dev->name));
1262 	}
1263 
1264 out:
1265 	if (sblocks_for_recheck) {
1266 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1267 		     mirror_index++) {
1268 			struct scrub_block *sblock = sblocks_for_recheck +
1269 						     mirror_index;
1270 			struct scrub_recover *recover;
1271 			int page_index;
1272 
1273 			for (page_index = 0; page_index < sblock->page_count;
1274 			     page_index++) {
1275 				sblock->pagev[page_index]->sblock = NULL;
1276 				recover = sblock->pagev[page_index]->recover;
1277 				if (recover) {
1278 					scrub_put_recover(recover);
1279 					sblock->pagev[page_index]->recover =
1280 									NULL;
1281 				}
1282 				scrub_page_put(sblock->pagev[page_index]);
1283 			}
1284 		}
1285 		kfree(sblocks_for_recheck);
1286 	}
1287 
1288 	return 0;
1289 }
1290 
1291 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio, u64 *raid_map)
1292 {
1293 	if (raid_map) {
1294 		if (raid_map[bbio->num_stripes - 1] == RAID6_Q_STRIPE)
1295 			return 3;
1296 		else
1297 			return 2;
1298 	} else {
1299 		return (int)bbio->num_stripes;
1300 	}
1301 }
1302 
1303 static inline void scrub_stripe_index_and_offset(u64 logical, u64 *raid_map,
1304 						 u64 mapped_length,
1305 						 int nstripes, int mirror,
1306 						 int *stripe_index,
1307 						 u64 *stripe_offset)
1308 {
1309 	int i;
1310 
1311 	if (raid_map) {
1312 		/* RAID5/6 */
1313 		for (i = 0; i < nstripes; i++) {
1314 			if (raid_map[i] == RAID6_Q_STRIPE ||
1315 			    raid_map[i] == RAID5_P_STRIPE)
1316 				continue;
1317 
1318 			if (logical >= raid_map[i] &&
1319 			    logical < raid_map[i] + mapped_length)
1320 				break;
1321 		}
1322 
1323 		*stripe_index = i;
1324 		*stripe_offset = logical - raid_map[i];
1325 	} else {
1326 		/* The other RAID type */
1327 		*stripe_index = mirror;
1328 		*stripe_offset = 0;
1329 	}
1330 }
1331 
1332 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1333 				     struct btrfs_fs_info *fs_info,
1334 				     struct scrub_block *original_sblock,
1335 				     u64 length, u64 logical,
1336 				     struct scrub_block *sblocks_for_recheck)
1337 {
1338 	struct scrub_recover *recover;
1339 	struct btrfs_bio *bbio;
1340 	u64 *raid_map;
1341 	u64 sublen;
1342 	u64 mapped_length;
1343 	u64 stripe_offset;
1344 	int stripe_index;
1345 	int page_index;
1346 	int mirror_index;
1347 	int nmirrors;
1348 	int ret;
1349 
1350 	/*
1351 	 * note: the two members ref_count and outstanding_pages
1352 	 * are not used (and not set) in the blocks that are used for
1353 	 * the recheck procedure
1354 	 */
1355 
1356 	page_index = 0;
1357 	while (length > 0) {
1358 		sublen = min_t(u64, length, PAGE_SIZE);
1359 		mapped_length = sublen;
1360 		bbio = NULL;
1361 		raid_map = NULL;
1362 
1363 		/*
1364 		 * with a length of PAGE_SIZE, each returned stripe
1365 		 * represents one mirror
1366 		 */
1367 		ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical,
1368 				       &mapped_length, &bbio, 0, &raid_map);
1369 		if (ret || !bbio || mapped_length < sublen) {
1370 			kfree(bbio);
1371 			kfree(raid_map);
1372 			return -EIO;
1373 		}
1374 
1375 		recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1376 		if (!recover) {
1377 			kfree(bbio);
1378 			kfree(raid_map);
1379 			return -ENOMEM;
1380 		}
1381 
1382 		atomic_set(&recover->refs, 1);
1383 		recover->bbio = bbio;
1384 		recover->raid_map = raid_map;
1385 		recover->map_length = mapped_length;
1386 
1387 		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1388 
1389 		nmirrors = scrub_nr_raid_mirrors(bbio, raid_map);
1390 		for (mirror_index = 0; mirror_index < nmirrors;
1391 		     mirror_index++) {
1392 			struct scrub_block *sblock;
1393 			struct scrub_page *page;
1394 
1395 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1396 				continue;
1397 
1398 			sblock = sblocks_for_recheck + mirror_index;
1399 			sblock->sctx = sctx;
1400 			page = kzalloc(sizeof(*page), GFP_NOFS);
1401 			if (!page) {
1402 leave_nomem:
1403 				spin_lock(&sctx->stat_lock);
1404 				sctx->stat.malloc_errors++;
1405 				spin_unlock(&sctx->stat_lock);
1406 				scrub_put_recover(recover);
1407 				return -ENOMEM;
1408 			}
1409 			scrub_page_get(page);
1410 			sblock->pagev[page_index] = page;
1411 			page->logical = logical;
1412 
1413 			scrub_stripe_index_and_offset(logical, raid_map,
1414 						      mapped_length,
1415 						      bbio->num_stripes,
1416 						      mirror_index,
1417 						      &stripe_index,
1418 						      &stripe_offset);
1419 			page->physical = bbio->stripes[stripe_index].physical +
1420 					 stripe_offset;
1421 			page->dev = bbio->stripes[stripe_index].dev;
1422 
1423 			BUG_ON(page_index >= original_sblock->page_count);
1424 			page->physical_for_dev_replace =
1425 				original_sblock->pagev[page_index]->
1426 				physical_for_dev_replace;
1427 			/* for missing devices, dev->bdev is NULL */
1428 			page->mirror_num = mirror_index + 1;
1429 			sblock->page_count++;
1430 			page->page = alloc_page(GFP_NOFS);
1431 			if (!page->page)
1432 				goto leave_nomem;
1433 
1434 			scrub_get_recover(recover);
1435 			page->recover = recover;
1436 		}
1437 		scrub_put_recover(recover);
1438 		length -= sublen;
1439 		logical += sublen;
1440 		page_index++;
1441 	}
1442 
1443 	return 0;
1444 }
1445 
1446 struct scrub_bio_ret {
1447 	struct completion event;
1448 	int error;
1449 };
1450 
1451 static void scrub_bio_wait_endio(struct bio *bio, int error)
1452 {
1453 	struct scrub_bio_ret *ret = bio->bi_private;
1454 
1455 	ret->error = error;
1456 	complete(&ret->event);
1457 }
1458 
1459 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1460 {
1461 	return page->recover && page->recover->raid_map;
1462 }
1463 
1464 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1465 					struct bio *bio,
1466 					struct scrub_page *page)
1467 {
1468 	struct scrub_bio_ret done;
1469 	int ret;
1470 
1471 	init_completion(&done.event);
1472 	done.error = 0;
1473 	bio->bi_iter.bi_sector = page->logical >> 9;
1474 	bio->bi_private = &done;
1475 	bio->bi_end_io = scrub_bio_wait_endio;
1476 
1477 	ret = raid56_parity_recover(fs_info->fs_root, bio, page->recover->bbio,
1478 				    page->recover->raid_map,
1479 				    page->recover->map_length,
1480 				    page->mirror_num, 0);
1481 	if (ret)
1482 		return ret;
1483 
1484 	wait_for_completion(&done.event);
1485 	if (done.error)
1486 		return -EIO;
1487 
1488 	return 0;
1489 }
1490 
1491 /*
1492  * this function will check the on disk data for checksum errors, header
1493  * errors and read I/O errors. If any I/O errors happen, the exact pages
1494  * which are errored are marked as being bad. The goal is to enable scrub
1495  * to take those pages that are not errored from all the mirrors so that
1496  * the pages that are errored in the just handled mirror can be repaired.
1497  */
1498 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1499 				struct scrub_block *sblock, int is_metadata,
1500 				int have_csum, u8 *csum, u64 generation,
1501 				u16 csum_size, int retry_failed_mirror)
1502 {
1503 	int page_num;
1504 
1505 	sblock->no_io_error_seen = 1;
1506 	sblock->header_error = 0;
1507 	sblock->checksum_error = 0;
1508 
1509 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1510 		struct bio *bio;
1511 		struct scrub_page *page = sblock->pagev[page_num];
1512 
1513 		if (page->dev->bdev == NULL) {
1514 			page->io_error = 1;
1515 			sblock->no_io_error_seen = 0;
1516 			continue;
1517 		}
1518 
1519 		WARN_ON(!page->page);
1520 		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1521 		if (!bio) {
1522 			page->io_error = 1;
1523 			sblock->no_io_error_seen = 0;
1524 			continue;
1525 		}
1526 		bio->bi_bdev = page->dev->bdev;
1527 
1528 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1529 		if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1530 			if (scrub_submit_raid56_bio_wait(fs_info, bio, page))
1531 				sblock->no_io_error_seen = 0;
1532 		} else {
1533 			bio->bi_iter.bi_sector = page->physical >> 9;
1534 
1535 			if (btrfsic_submit_bio_wait(READ, bio))
1536 				sblock->no_io_error_seen = 0;
1537 		}
1538 
1539 		bio_put(bio);
1540 	}
1541 
1542 	if (sblock->no_io_error_seen)
1543 		scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1544 					     have_csum, csum, generation,
1545 					     csum_size);
1546 
1547 	return;
1548 }
1549 
1550 static inline int scrub_check_fsid(u8 fsid[],
1551 				   struct scrub_page *spage)
1552 {
1553 	struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1554 	int ret;
1555 
1556 	ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1557 	return !ret;
1558 }
1559 
1560 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1561 					 struct scrub_block *sblock,
1562 					 int is_metadata, int have_csum,
1563 					 const u8 *csum, u64 generation,
1564 					 u16 csum_size)
1565 {
1566 	int page_num;
1567 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1568 	u32 crc = ~(u32)0;
1569 	void *mapped_buffer;
1570 
1571 	WARN_ON(!sblock->pagev[0]->page);
1572 	if (is_metadata) {
1573 		struct btrfs_header *h;
1574 
1575 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1576 		h = (struct btrfs_header *)mapped_buffer;
1577 
1578 		if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1579 		    !scrub_check_fsid(h->fsid, sblock->pagev[0]) ||
1580 		    memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1581 			   BTRFS_UUID_SIZE)) {
1582 			sblock->header_error = 1;
1583 		} else if (generation != btrfs_stack_header_generation(h)) {
1584 			sblock->header_error = 1;
1585 			sblock->generation_error = 1;
1586 		}
1587 		csum = h->csum;
1588 	} else {
1589 		if (!have_csum)
1590 			return;
1591 
1592 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1593 	}
1594 
1595 	for (page_num = 0;;) {
1596 		if (page_num == 0 && is_metadata)
1597 			crc = btrfs_csum_data(
1598 				((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1599 				crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1600 		else
1601 			crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1602 
1603 		kunmap_atomic(mapped_buffer);
1604 		page_num++;
1605 		if (page_num >= sblock->page_count)
1606 			break;
1607 		WARN_ON(!sblock->pagev[page_num]->page);
1608 
1609 		mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1610 	}
1611 
1612 	btrfs_csum_final(crc, calculated_csum);
1613 	if (memcmp(calculated_csum, csum, csum_size))
1614 		sblock->checksum_error = 1;
1615 }
1616 
1617 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1618 					     struct scrub_block *sblock_good,
1619 					     int force_write)
1620 {
1621 	int page_num;
1622 	int ret = 0;
1623 
1624 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1625 		int ret_sub;
1626 
1627 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1628 							   sblock_good,
1629 							   page_num,
1630 							   force_write);
1631 		if (ret_sub)
1632 			ret = ret_sub;
1633 	}
1634 
1635 	return ret;
1636 }
1637 
1638 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1639 					    struct scrub_block *sblock_good,
1640 					    int page_num, int force_write)
1641 {
1642 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1643 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1644 
1645 	BUG_ON(page_bad->page == NULL);
1646 	BUG_ON(page_good->page == NULL);
1647 	if (force_write || sblock_bad->header_error ||
1648 	    sblock_bad->checksum_error || page_bad->io_error) {
1649 		struct bio *bio;
1650 		int ret;
1651 
1652 		if (!page_bad->dev->bdev) {
1653 			printk_ratelimited(KERN_WARNING "BTRFS: "
1654 				"scrub_repair_page_from_good_copy(bdev == NULL) "
1655 				"is unexpected!\n");
1656 			return -EIO;
1657 		}
1658 
1659 		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1660 		if (!bio)
1661 			return -EIO;
1662 		bio->bi_bdev = page_bad->dev->bdev;
1663 		bio->bi_iter.bi_sector = page_bad->physical >> 9;
1664 
1665 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1666 		if (PAGE_SIZE != ret) {
1667 			bio_put(bio);
1668 			return -EIO;
1669 		}
1670 
1671 		if (btrfsic_submit_bio_wait(WRITE, bio)) {
1672 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1673 				BTRFS_DEV_STAT_WRITE_ERRS);
1674 			btrfs_dev_replace_stats_inc(
1675 				&sblock_bad->sctx->dev_root->fs_info->
1676 				dev_replace.num_write_errors);
1677 			bio_put(bio);
1678 			return -EIO;
1679 		}
1680 		bio_put(bio);
1681 	}
1682 
1683 	return 0;
1684 }
1685 
1686 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1687 {
1688 	int page_num;
1689 
1690 	/*
1691 	 * This block is used for the check of the parity on the source device,
1692 	 * so the data needn't be written into the destination device.
1693 	 */
1694 	if (sblock->sparity)
1695 		return;
1696 
1697 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1698 		int ret;
1699 
1700 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1701 		if (ret)
1702 			btrfs_dev_replace_stats_inc(
1703 				&sblock->sctx->dev_root->fs_info->dev_replace.
1704 				num_write_errors);
1705 	}
1706 }
1707 
1708 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1709 					   int page_num)
1710 {
1711 	struct scrub_page *spage = sblock->pagev[page_num];
1712 
1713 	BUG_ON(spage->page == NULL);
1714 	if (spage->io_error) {
1715 		void *mapped_buffer = kmap_atomic(spage->page);
1716 
1717 		memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1718 		flush_dcache_page(spage->page);
1719 		kunmap_atomic(mapped_buffer);
1720 	}
1721 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1722 }
1723 
1724 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1725 				    struct scrub_page *spage)
1726 {
1727 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1728 	struct scrub_bio *sbio;
1729 	int ret;
1730 
1731 	mutex_lock(&wr_ctx->wr_lock);
1732 again:
1733 	if (!wr_ctx->wr_curr_bio) {
1734 		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1735 					      GFP_NOFS);
1736 		if (!wr_ctx->wr_curr_bio) {
1737 			mutex_unlock(&wr_ctx->wr_lock);
1738 			return -ENOMEM;
1739 		}
1740 		wr_ctx->wr_curr_bio->sctx = sctx;
1741 		wr_ctx->wr_curr_bio->page_count = 0;
1742 	}
1743 	sbio = wr_ctx->wr_curr_bio;
1744 	if (sbio->page_count == 0) {
1745 		struct bio *bio;
1746 
1747 		sbio->physical = spage->physical_for_dev_replace;
1748 		sbio->logical = spage->logical;
1749 		sbio->dev = wr_ctx->tgtdev;
1750 		bio = sbio->bio;
1751 		if (!bio) {
1752 			bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1753 			if (!bio) {
1754 				mutex_unlock(&wr_ctx->wr_lock);
1755 				return -ENOMEM;
1756 			}
1757 			sbio->bio = bio;
1758 		}
1759 
1760 		bio->bi_private = sbio;
1761 		bio->bi_end_io = scrub_wr_bio_end_io;
1762 		bio->bi_bdev = sbio->dev->bdev;
1763 		bio->bi_iter.bi_sector = sbio->physical >> 9;
1764 		sbio->err = 0;
1765 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1766 		   spage->physical_for_dev_replace ||
1767 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1768 		   spage->logical) {
1769 		scrub_wr_submit(sctx);
1770 		goto again;
1771 	}
1772 
1773 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1774 	if (ret != PAGE_SIZE) {
1775 		if (sbio->page_count < 1) {
1776 			bio_put(sbio->bio);
1777 			sbio->bio = NULL;
1778 			mutex_unlock(&wr_ctx->wr_lock);
1779 			return -EIO;
1780 		}
1781 		scrub_wr_submit(sctx);
1782 		goto again;
1783 	}
1784 
1785 	sbio->pagev[sbio->page_count] = spage;
1786 	scrub_page_get(spage);
1787 	sbio->page_count++;
1788 	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1789 		scrub_wr_submit(sctx);
1790 	mutex_unlock(&wr_ctx->wr_lock);
1791 
1792 	return 0;
1793 }
1794 
1795 static void scrub_wr_submit(struct scrub_ctx *sctx)
1796 {
1797 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1798 	struct scrub_bio *sbio;
1799 
1800 	if (!wr_ctx->wr_curr_bio)
1801 		return;
1802 
1803 	sbio = wr_ctx->wr_curr_bio;
1804 	wr_ctx->wr_curr_bio = NULL;
1805 	WARN_ON(!sbio->bio->bi_bdev);
1806 	scrub_pending_bio_inc(sctx);
1807 	/* process all writes in a single worker thread. Then the block layer
1808 	 * orders the requests before sending them to the driver which
1809 	 * doubled the write performance on spinning disks when measured
1810 	 * with Linux 3.5 */
1811 	btrfsic_submit_bio(WRITE, sbio->bio);
1812 }
1813 
1814 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1815 {
1816 	struct scrub_bio *sbio = bio->bi_private;
1817 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1818 
1819 	sbio->err = err;
1820 	sbio->bio = bio;
1821 
1822 	btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1823 			 scrub_wr_bio_end_io_worker, NULL, NULL);
1824 	btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1825 }
1826 
1827 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1828 {
1829 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1830 	struct scrub_ctx *sctx = sbio->sctx;
1831 	int i;
1832 
1833 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1834 	if (sbio->err) {
1835 		struct btrfs_dev_replace *dev_replace =
1836 			&sbio->sctx->dev_root->fs_info->dev_replace;
1837 
1838 		for (i = 0; i < sbio->page_count; i++) {
1839 			struct scrub_page *spage = sbio->pagev[i];
1840 
1841 			spage->io_error = 1;
1842 			btrfs_dev_replace_stats_inc(&dev_replace->
1843 						    num_write_errors);
1844 		}
1845 	}
1846 
1847 	for (i = 0; i < sbio->page_count; i++)
1848 		scrub_page_put(sbio->pagev[i]);
1849 
1850 	bio_put(sbio->bio);
1851 	kfree(sbio);
1852 	scrub_pending_bio_dec(sctx);
1853 }
1854 
1855 static int scrub_checksum(struct scrub_block *sblock)
1856 {
1857 	u64 flags;
1858 	int ret;
1859 
1860 	WARN_ON(sblock->page_count < 1);
1861 	flags = sblock->pagev[0]->flags;
1862 	ret = 0;
1863 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1864 		ret = scrub_checksum_data(sblock);
1865 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1866 		ret = scrub_checksum_tree_block(sblock);
1867 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1868 		(void)scrub_checksum_super(sblock);
1869 	else
1870 		WARN_ON(1);
1871 	if (ret)
1872 		scrub_handle_errored_block(sblock);
1873 
1874 	return ret;
1875 }
1876 
1877 static int scrub_checksum_data(struct scrub_block *sblock)
1878 {
1879 	struct scrub_ctx *sctx = sblock->sctx;
1880 	u8 csum[BTRFS_CSUM_SIZE];
1881 	u8 *on_disk_csum;
1882 	struct page *page;
1883 	void *buffer;
1884 	u32 crc = ~(u32)0;
1885 	int fail = 0;
1886 	u64 len;
1887 	int index;
1888 
1889 	BUG_ON(sblock->page_count < 1);
1890 	if (!sblock->pagev[0]->have_csum)
1891 		return 0;
1892 
1893 	on_disk_csum = sblock->pagev[0]->csum;
1894 	page = sblock->pagev[0]->page;
1895 	buffer = kmap_atomic(page);
1896 
1897 	len = sctx->sectorsize;
1898 	index = 0;
1899 	for (;;) {
1900 		u64 l = min_t(u64, len, PAGE_SIZE);
1901 
1902 		crc = btrfs_csum_data(buffer, crc, l);
1903 		kunmap_atomic(buffer);
1904 		len -= l;
1905 		if (len == 0)
1906 			break;
1907 		index++;
1908 		BUG_ON(index >= sblock->page_count);
1909 		BUG_ON(!sblock->pagev[index]->page);
1910 		page = sblock->pagev[index]->page;
1911 		buffer = kmap_atomic(page);
1912 	}
1913 
1914 	btrfs_csum_final(crc, csum);
1915 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1916 		fail = 1;
1917 
1918 	return fail;
1919 }
1920 
1921 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1922 {
1923 	struct scrub_ctx *sctx = sblock->sctx;
1924 	struct btrfs_header *h;
1925 	struct btrfs_root *root = sctx->dev_root;
1926 	struct btrfs_fs_info *fs_info = root->fs_info;
1927 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1928 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1929 	struct page *page;
1930 	void *mapped_buffer;
1931 	u64 mapped_size;
1932 	void *p;
1933 	u32 crc = ~(u32)0;
1934 	int fail = 0;
1935 	int crc_fail = 0;
1936 	u64 len;
1937 	int index;
1938 
1939 	BUG_ON(sblock->page_count < 1);
1940 	page = sblock->pagev[0]->page;
1941 	mapped_buffer = kmap_atomic(page);
1942 	h = (struct btrfs_header *)mapped_buffer;
1943 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1944 
1945 	/*
1946 	 * we don't use the getter functions here, as we
1947 	 * a) don't have an extent buffer and
1948 	 * b) the page is already kmapped
1949 	 */
1950 
1951 	if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1952 		++fail;
1953 
1954 	if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1955 		++fail;
1956 
1957 	if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1958 		++fail;
1959 
1960 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1961 		   BTRFS_UUID_SIZE))
1962 		++fail;
1963 
1964 	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1965 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1966 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1967 	index = 0;
1968 	for (;;) {
1969 		u64 l = min_t(u64, len, mapped_size);
1970 
1971 		crc = btrfs_csum_data(p, crc, l);
1972 		kunmap_atomic(mapped_buffer);
1973 		len -= l;
1974 		if (len == 0)
1975 			break;
1976 		index++;
1977 		BUG_ON(index >= sblock->page_count);
1978 		BUG_ON(!sblock->pagev[index]->page);
1979 		page = sblock->pagev[index]->page;
1980 		mapped_buffer = kmap_atomic(page);
1981 		mapped_size = PAGE_SIZE;
1982 		p = mapped_buffer;
1983 	}
1984 
1985 	btrfs_csum_final(crc, calculated_csum);
1986 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1987 		++crc_fail;
1988 
1989 	return fail || crc_fail;
1990 }
1991 
1992 static int scrub_checksum_super(struct scrub_block *sblock)
1993 {
1994 	struct btrfs_super_block *s;
1995 	struct scrub_ctx *sctx = sblock->sctx;
1996 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1997 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1998 	struct page *page;
1999 	void *mapped_buffer;
2000 	u64 mapped_size;
2001 	void *p;
2002 	u32 crc = ~(u32)0;
2003 	int fail_gen = 0;
2004 	int fail_cor = 0;
2005 	u64 len;
2006 	int index;
2007 
2008 	BUG_ON(sblock->page_count < 1);
2009 	page = sblock->pagev[0]->page;
2010 	mapped_buffer = kmap_atomic(page);
2011 	s = (struct btrfs_super_block *)mapped_buffer;
2012 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
2013 
2014 	if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2015 		++fail_cor;
2016 
2017 	if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2018 		++fail_gen;
2019 
2020 	if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2021 		++fail_cor;
2022 
2023 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2024 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2025 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2026 	index = 0;
2027 	for (;;) {
2028 		u64 l = min_t(u64, len, mapped_size);
2029 
2030 		crc = btrfs_csum_data(p, crc, l);
2031 		kunmap_atomic(mapped_buffer);
2032 		len -= l;
2033 		if (len == 0)
2034 			break;
2035 		index++;
2036 		BUG_ON(index >= sblock->page_count);
2037 		BUG_ON(!sblock->pagev[index]->page);
2038 		page = sblock->pagev[index]->page;
2039 		mapped_buffer = kmap_atomic(page);
2040 		mapped_size = PAGE_SIZE;
2041 		p = mapped_buffer;
2042 	}
2043 
2044 	btrfs_csum_final(crc, calculated_csum);
2045 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2046 		++fail_cor;
2047 
2048 	if (fail_cor + fail_gen) {
2049 		/*
2050 		 * if we find an error in a super block, we just report it.
2051 		 * They will get written with the next transaction commit
2052 		 * anyway
2053 		 */
2054 		spin_lock(&sctx->stat_lock);
2055 		++sctx->stat.super_errors;
2056 		spin_unlock(&sctx->stat_lock);
2057 		if (fail_cor)
2058 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2059 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
2060 		else
2061 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2062 				BTRFS_DEV_STAT_GENERATION_ERRS);
2063 	}
2064 
2065 	return fail_cor + fail_gen;
2066 }
2067 
2068 static void scrub_block_get(struct scrub_block *sblock)
2069 {
2070 	atomic_inc(&sblock->ref_count);
2071 }
2072 
2073 static void scrub_block_put(struct scrub_block *sblock)
2074 {
2075 	if (atomic_dec_and_test(&sblock->ref_count)) {
2076 		int i;
2077 
2078 		if (sblock->sparity)
2079 			scrub_parity_put(sblock->sparity);
2080 
2081 		for (i = 0; i < sblock->page_count; i++)
2082 			scrub_page_put(sblock->pagev[i]);
2083 		kfree(sblock);
2084 	}
2085 }
2086 
2087 static void scrub_page_get(struct scrub_page *spage)
2088 {
2089 	atomic_inc(&spage->ref_count);
2090 }
2091 
2092 static void scrub_page_put(struct scrub_page *spage)
2093 {
2094 	if (atomic_dec_and_test(&spage->ref_count)) {
2095 		if (spage->page)
2096 			__free_page(spage->page);
2097 		kfree(spage);
2098 	}
2099 }
2100 
2101 static void scrub_submit(struct scrub_ctx *sctx)
2102 {
2103 	struct scrub_bio *sbio;
2104 
2105 	if (sctx->curr == -1)
2106 		return;
2107 
2108 	sbio = sctx->bios[sctx->curr];
2109 	sctx->curr = -1;
2110 	scrub_pending_bio_inc(sctx);
2111 
2112 	if (!sbio->bio->bi_bdev) {
2113 		/*
2114 		 * this case should not happen. If btrfs_map_block() is
2115 		 * wrong, it could happen for dev-replace operations on
2116 		 * missing devices when no mirrors are available, but in
2117 		 * this case it should already fail the mount.
2118 		 * This case is handled correctly (but _very_ slowly).
2119 		 */
2120 		printk_ratelimited(KERN_WARNING
2121 			"BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
2122 		bio_endio(sbio->bio, -EIO);
2123 	} else {
2124 		btrfsic_submit_bio(READ, sbio->bio);
2125 	}
2126 }
2127 
2128 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2129 				    struct scrub_page *spage)
2130 {
2131 	struct scrub_block *sblock = spage->sblock;
2132 	struct scrub_bio *sbio;
2133 	int ret;
2134 
2135 again:
2136 	/*
2137 	 * grab a fresh bio or wait for one to become available
2138 	 */
2139 	while (sctx->curr == -1) {
2140 		spin_lock(&sctx->list_lock);
2141 		sctx->curr = sctx->first_free;
2142 		if (sctx->curr != -1) {
2143 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
2144 			sctx->bios[sctx->curr]->next_free = -1;
2145 			sctx->bios[sctx->curr]->page_count = 0;
2146 			spin_unlock(&sctx->list_lock);
2147 		} else {
2148 			spin_unlock(&sctx->list_lock);
2149 			wait_event(sctx->list_wait, sctx->first_free != -1);
2150 		}
2151 	}
2152 	sbio = sctx->bios[sctx->curr];
2153 	if (sbio->page_count == 0) {
2154 		struct bio *bio;
2155 
2156 		sbio->physical = spage->physical;
2157 		sbio->logical = spage->logical;
2158 		sbio->dev = spage->dev;
2159 		bio = sbio->bio;
2160 		if (!bio) {
2161 			bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
2162 			if (!bio)
2163 				return -ENOMEM;
2164 			sbio->bio = bio;
2165 		}
2166 
2167 		bio->bi_private = sbio;
2168 		bio->bi_end_io = scrub_bio_end_io;
2169 		bio->bi_bdev = sbio->dev->bdev;
2170 		bio->bi_iter.bi_sector = sbio->physical >> 9;
2171 		sbio->err = 0;
2172 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2173 		   spage->physical ||
2174 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
2175 		   spage->logical ||
2176 		   sbio->dev != spage->dev) {
2177 		scrub_submit(sctx);
2178 		goto again;
2179 	}
2180 
2181 	sbio->pagev[sbio->page_count] = spage;
2182 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2183 	if (ret != PAGE_SIZE) {
2184 		if (sbio->page_count < 1) {
2185 			bio_put(sbio->bio);
2186 			sbio->bio = NULL;
2187 			return -EIO;
2188 		}
2189 		scrub_submit(sctx);
2190 		goto again;
2191 	}
2192 
2193 	scrub_block_get(sblock); /* one for the page added to the bio */
2194 	atomic_inc(&sblock->outstanding_pages);
2195 	sbio->page_count++;
2196 	if (sbio->page_count == sctx->pages_per_rd_bio)
2197 		scrub_submit(sctx);
2198 
2199 	return 0;
2200 }
2201 
2202 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2203 		       u64 physical, struct btrfs_device *dev, u64 flags,
2204 		       u64 gen, int mirror_num, u8 *csum, int force,
2205 		       u64 physical_for_dev_replace)
2206 {
2207 	struct scrub_block *sblock;
2208 	int index;
2209 
2210 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2211 	if (!sblock) {
2212 		spin_lock(&sctx->stat_lock);
2213 		sctx->stat.malloc_errors++;
2214 		spin_unlock(&sctx->stat_lock);
2215 		return -ENOMEM;
2216 	}
2217 
2218 	/* one ref inside this function, plus one for each page added to
2219 	 * a bio later on */
2220 	atomic_set(&sblock->ref_count, 1);
2221 	sblock->sctx = sctx;
2222 	sblock->no_io_error_seen = 1;
2223 
2224 	for (index = 0; len > 0; index++) {
2225 		struct scrub_page *spage;
2226 		u64 l = min_t(u64, len, PAGE_SIZE);
2227 
2228 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
2229 		if (!spage) {
2230 leave_nomem:
2231 			spin_lock(&sctx->stat_lock);
2232 			sctx->stat.malloc_errors++;
2233 			spin_unlock(&sctx->stat_lock);
2234 			scrub_block_put(sblock);
2235 			return -ENOMEM;
2236 		}
2237 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2238 		scrub_page_get(spage);
2239 		sblock->pagev[index] = spage;
2240 		spage->sblock = sblock;
2241 		spage->dev = dev;
2242 		spage->flags = flags;
2243 		spage->generation = gen;
2244 		spage->logical = logical;
2245 		spage->physical = physical;
2246 		spage->physical_for_dev_replace = physical_for_dev_replace;
2247 		spage->mirror_num = mirror_num;
2248 		if (csum) {
2249 			spage->have_csum = 1;
2250 			memcpy(spage->csum, csum, sctx->csum_size);
2251 		} else {
2252 			spage->have_csum = 0;
2253 		}
2254 		sblock->page_count++;
2255 		spage->page = alloc_page(GFP_NOFS);
2256 		if (!spage->page)
2257 			goto leave_nomem;
2258 		len -= l;
2259 		logical += l;
2260 		physical += l;
2261 		physical_for_dev_replace += l;
2262 	}
2263 
2264 	WARN_ON(sblock->page_count == 0);
2265 	for (index = 0; index < sblock->page_count; index++) {
2266 		struct scrub_page *spage = sblock->pagev[index];
2267 		int ret;
2268 
2269 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2270 		if (ret) {
2271 			scrub_block_put(sblock);
2272 			return ret;
2273 		}
2274 	}
2275 
2276 	if (force)
2277 		scrub_submit(sctx);
2278 
2279 	/* last one frees, either here or in bio completion for last page */
2280 	scrub_block_put(sblock);
2281 	return 0;
2282 }
2283 
2284 static void scrub_bio_end_io(struct bio *bio, int err)
2285 {
2286 	struct scrub_bio *sbio = bio->bi_private;
2287 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2288 
2289 	sbio->err = err;
2290 	sbio->bio = bio;
2291 
2292 	btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2293 }
2294 
2295 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2296 {
2297 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2298 	struct scrub_ctx *sctx = sbio->sctx;
2299 	int i;
2300 
2301 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2302 	if (sbio->err) {
2303 		for (i = 0; i < sbio->page_count; i++) {
2304 			struct scrub_page *spage = sbio->pagev[i];
2305 
2306 			spage->io_error = 1;
2307 			spage->sblock->no_io_error_seen = 0;
2308 		}
2309 	}
2310 
2311 	/* now complete the scrub_block items that have all pages completed */
2312 	for (i = 0; i < sbio->page_count; i++) {
2313 		struct scrub_page *spage = sbio->pagev[i];
2314 		struct scrub_block *sblock = spage->sblock;
2315 
2316 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2317 			scrub_block_complete(sblock);
2318 		scrub_block_put(sblock);
2319 	}
2320 
2321 	bio_put(sbio->bio);
2322 	sbio->bio = NULL;
2323 	spin_lock(&sctx->list_lock);
2324 	sbio->next_free = sctx->first_free;
2325 	sctx->first_free = sbio->index;
2326 	spin_unlock(&sctx->list_lock);
2327 
2328 	if (sctx->is_dev_replace &&
2329 	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2330 		mutex_lock(&sctx->wr_ctx.wr_lock);
2331 		scrub_wr_submit(sctx);
2332 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2333 	}
2334 
2335 	scrub_pending_bio_dec(sctx);
2336 }
2337 
2338 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2339 				       unsigned long *bitmap,
2340 				       u64 start, u64 len)
2341 {
2342 	int offset;
2343 	int nsectors;
2344 	int sectorsize = sparity->sctx->dev_root->sectorsize;
2345 
2346 	if (len >= sparity->stripe_len) {
2347 		bitmap_set(bitmap, 0, sparity->nsectors);
2348 		return;
2349 	}
2350 
2351 	start -= sparity->logic_start;
2352 	offset = (int)do_div(start, sparity->stripe_len);
2353 	offset /= sectorsize;
2354 	nsectors = (int)len / sectorsize;
2355 
2356 	if (offset + nsectors <= sparity->nsectors) {
2357 		bitmap_set(bitmap, offset, nsectors);
2358 		return;
2359 	}
2360 
2361 	bitmap_set(bitmap, offset, sparity->nsectors - offset);
2362 	bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2363 }
2364 
2365 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2366 						   u64 start, u64 len)
2367 {
2368 	__scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2369 }
2370 
2371 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2372 						  u64 start, u64 len)
2373 {
2374 	__scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2375 }
2376 
2377 static void scrub_block_complete(struct scrub_block *sblock)
2378 {
2379 	int corrupted = 0;
2380 
2381 	if (!sblock->no_io_error_seen) {
2382 		corrupted = 1;
2383 		scrub_handle_errored_block(sblock);
2384 	} else {
2385 		/*
2386 		 * if has checksum error, write via repair mechanism in
2387 		 * dev replace case, otherwise write here in dev replace
2388 		 * case.
2389 		 */
2390 		corrupted = scrub_checksum(sblock);
2391 		if (!corrupted && sblock->sctx->is_dev_replace)
2392 			scrub_write_block_to_dev_replace(sblock);
2393 	}
2394 
2395 	if (sblock->sparity && corrupted && !sblock->data_corrected) {
2396 		u64 start = sblock->pagev[0]->logical;
2397 		u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2398 			  PAGE_SIZE;
2399 
2400 		scrub_parity_mark_sectors_error(sblock->sparity,
2401 						start, end - start);
2402 	}
2403 }
2404 
2405 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2406 			   u8 *csum)
2407 {
2408 	struct btrfs_ordered_sum *sum = NULL;
2409 	unsigned long index;
2410 	unsigned long num_sectors;
2411 
2412 	while (!list_empty(&sctx->csum_list)) {
2413 		sum = list_first_entry(&sctx->csum_list,
2414 				       struct btrfs_ordered_sum, list);
2415 		if (sum->bytenr > logical)
2416 			return 0;
2417 		if (sum->bytenr + sum->len > logical)
2418 			break;
2419 
2420 		++sctx->stat.csum_discards;
2421 		list_del(&sum->list);
2422 		kfree(sum);
2423 		sum = NULL;
2424 	}
2425 	if (!sum)
2426 		return 0;
2427 
2428 	index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2429 	num_sectors = sum->len / sctx->sectorsize;
2430 	memcpy(csum, sum->sums + index, sctx->csum_size);
2431 	if (index == num_sectors - 1) {
2432 		list_del(&sum->list);
2433 		kfree(sum);
2434 	}
2435 	return 1;
2436 }
2437 
2438 /* scrub extent tries to collect up to 64 kB for each bio */
2439 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2440 			u64 physical, struct btrfs_device *dev, u64 flags,
2441 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2442 {
2443 	int ret;
2444 	u8 csum[BTRFS_CSUM_SIZE];
2445 	u32 blocksize;
2446 
2447 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2448 		blocksize = sctx->sectorsize;
2449 		spin_lock(&sctx->stat_lock);
2450 		sctx->stat.data_extents_scrubbed++;
2451 		sctx->stat.data_bytes_scrubbed += len;
2452 		spin_unlock(&sctx->stat_lock);
2453 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2454 		blocksize = sctx->nodesize;
2455 		spin_lock(&sctx->stat_lock);
2456 		sctx->stat.tree_extents_scrubbed++;
2457 		sctx->stat.tree_bytes_scrubbed += len;
2458 		spin_unlock(&sctx->stat_lock);
2459 	} else {
2460 		blocksize = sctx->sectorsize;
2461 		WARN_ON(1);
2462 	}
2463 
2464 	while (len) {
2465 		u64 l = min_t(u64, len, blocksize);
2466 		int have_csum = 0;
2467 
2468 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2469 			/* push csums to sbio */
2470 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2471 			if (have_csum == 0)
2472 				++sctx->stat.no_csum;
2473 			if (sctx->is_dev_replace && !have_csum) {
2474 				ret = copy_nocow_pages(sctx, logical, l,
2475 						       mirror_num,
2476 						      physical_for_dev_replace);
2477 				goto behind_scrub_pages;
2478 			}
2479 		}
2480 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2481 				  mirror_num, have_csum ? csum : NULL, 0,
2482 				  physical_for_dev_replace);
2483 behind_scrub_pages:
2484 		if (ret)
2485 			return ret;
2486 		len -= l;
2487 		logical += l;
2488 		physical += l;
2489 		physical_for_dev_replace += l;
2490 	}
2491 	return 0;
2492 }
2493 
2494 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2495 				  u64 logical, u64 len,
2496 				  u64 physical, struct btrfs_device *dev,
2497 				  u64 flags, u64 gen, int mirror_num, u8 *csum)
2498 {
2499 	struct scrub_ctx *sctx = sparity->sctx;
2500 	struct scrub_block *sblock;
2501 	int index;
2502 
2503 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2504 	if (!sblock) {
2505 		spin_lock(&sctx->stat_lock);
2506 		sctx->stat.malloc_errors++;
2507 		spin_unlock(&sctx->stat_lock);
2508 		return -ENOMEM;
2509 	}
2510 
2511 	/* one ref inside this function, plus one for each page added to
2512 	 * a bio later on */
2513 	atomic_set(&sblock->ref_count, 1);
2514 	sblock->sctx = sctx;
2515 	sblock->no_io_error_seen = 1;
2516 	sblock->sparity = sparity;
2517 	scrub_parity_get(sparity);
2518 
2519 	for (index = 0; len > 0; index++) {
2520 		struct scrub_page *spage;
2521 		u64 l = min_t(u64, len, PAGE_SIZE);
2522 
2523 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
2524 		if (!spage) {
2525 leave_nomem:
2526 			spin_lock(&sctx->stat_lock);
2527 			sctx->stat.malloc_errors++;
2528 			spin_unlock(&sctx->stat_lock);
2529 			scrub_block_put(sblock);
2530 			return -ENOMEM;
2531 		}
2532 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2533 		/* For scrub block */
2534 		scrub_page_get(spage);
2535 		sblock->pagev[index] = spage;
2536 		/* For scrub parity */
2537 		scrub_page_get(spage);
2538 		list_add_tail(&spage->list, &sparity->spages);
2539 		spage->sblock = sblock;
2540 		spage->dev = dev;
2541 		spage->flags = flags;
2542 		spage->generation = gen;
2543 		spage->logical = logical;
2544 		spage->physical = physical;
2545 		spage->mirror_num = mirror_num;
2546 		if (csum) {
2547 			spage->have_csum = 1;
2548 			memcpy(spage->csum, csum, sctx->csum_size);
2549 		} else {
2550 			spage->have_csum = 0;
2551 		}
2552 		sblock->page_count++;
2553 		spage->page = alloc_page(GFP_NOFS);
2554 		if (!spage->page)
2555 			goto leave_nomem;
2556 		len -= l;
2557 		logical += l;
2558 		physical += l;
2559 	}
2560 
2561 	WARN_ON(sblock->page_count == 0);
2562 	for (index = 0; index < sblock->page_count; index++) {
2563 		struct scrub_page *spage = sblock->pagev[index];
2564 		int ret;
2565 
2566 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2567 		if (ret) {
2568 			scrub_block_put(sblock);
2569 			return ret;
2570 		}
2571 	}
2572 
2573 	/* last one frees, either here or in bio completion for last page */
2574 	scrub_block_put(sblock);
2575 	return 0;
2576 }
2577 
2578 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2579 				   u64 logical, u64 len,
2580 				   u64 physical, struct btrfs_device *dev,
2581 				   u64 flags, u64 gen, int mirror_num)
2582 {
2583 	struct scrub_ctx *sctx = sparity->sctx;
2584 	int ret;
2585 	u8 csum[BTRFS_CSUM_SIZE];
2586 	u32 blocksize;
2587 
2588 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2589 		blocksize = sctx->sectorsize;
2590 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2591 		blocksize = sctx->nodesize;
2592 	} else {
2593 		blocksize = sctx->sectorsize;
2594 		WARN_ON(1);
2595 	}
2596 
2597 	while (len) {
2598 		u64 l = min_t(u64, len, blocksize);
2599 		int have_csum = 0;
2600 
2601 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2602 			/* push csums to sbio */
2603 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2604 			if (have_csum == 0)
2605 				goto skip;
2606 		}
2607 		ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2608 					     flags, gen, mirror_num,
2609 					     have_csum ? csum : NULL);
2610 		if (ret)
2611 			return ret;
2612 skip:
2613 		len -= l;
2614 		logical += l;
2615 		physical += l;
2616 	}
2617 	return 0;
2618 }
2619 
2620 /*
2621  * Given a physical address, this will calculate it's
2622  * logical offset. if this is a parity stripe, it will return
2623  * the most left data stripe's logical offset.
2624  *
2625  * return 0 if it is a data stripe, 1 means parity stripe.
2626  */
2627 static int get_raid56_logic_offset(u64 physical, int num,
2628 				   struct map_lookup *map, u64 *offset,
2629 				   u64 *stripe_start)
2630 {
2631 	int i;
2632 	int j = 0;
2633 	u64 stripe_nr;
2634 	u64 last_offset;
2635 	int stripe_index;
2636 	int rot;
2637 
2638 	last_offset = (physical - map->stripes[num].physical) *
2639 		      nr_data_stripes(map);
2640 	if (stripe_start)
2641 		*stripe_start = last_offset;
2642 
2643 	*offset = last_offset;
2644 	for (i = 0; i < nr_data_stripes(map); i++) {
2645 		*offset = last_offset + i * map->stripe_len;
2646 
2647 		stripe_nr = *offset;
2648 		do_div(stripe_nr, map->stripe_len);
2649 		do_div(stripe_nr, nr_data_stripes(map));
2650 
2651 		/* Work out the disk rotation on this stripe-set */
2652 		rot = do_div(stripe_nr, map->num_stripes);
2653 		/* calculate which stripe this data locates */
2654 		rot += i;
2655 		stripe_index = rot % map->num_stripes;
2656 		if (stripe_index == num)
2657 			return 0;
2658 		if (stripe_index < num)
2659 			j++;
2660 	}
2661 	*offset = last_offset + j * map->stripe_len;
2662 	return 1;
2663 }
2664 
2665 static void scrub_free_parity(struct scrub_parity *sparity)
2666 {
2667 	struct scrub_ctx *sctx = sparity->sctx;
2668 	struct scrub_page *curr, *next;
2669 	int nbits;
2670 
2671 	nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2672 	if (nbits) {
2673 		spin_lock(&sctx->stat_lock);
2674 		sctx->stat.read_errors += nbits;
2675 		sctx->stat.uncorrectable_errors += nbits;
2676 		spin_unlock(&sctx->stat_lock);
2677 	}
2678 
2679 	list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2680 		list_del_init(&curr->list);
2681 		scrub_page_put(curr);
2682 	}
2683 
2684 	kfree(sparity);
2685 }
2686 
2687 static void scrub_parity_bio_endio(struct bio *bio, int error)
2688 {
2689 	struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2690 	struct scrub_ctx *sctx = sparity->sctx;
2691 
2692 	if (error)
2693 		bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2694 			  sparity->nsectors);
2695 
2696 	scrub_free_parity(sparity);
2697 	scrub_pending_bio_dec(sctx);
2698 	bio_put(bio);
2699 }
2700 
2701 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2702 {
2703 	struct scrub_ctx *sctx = sparity->sctx;
2704 	struct bio *bio;
2705 	struct btrfs_raid_bio *rbio;
2706 	struct scrub_page *spage;
2707 	struct btrfs_bio *bbio = NULL;
2708 	u64 *raid_map = NULL;
2709 	u64 length;
2710 	int ret;
2711 
2712 	if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2713 			   sparity->nsectors))
2714 		goto out;
2715 
2716 	length = sparity->logic_end - sparity->logic_start + 1;
2717 	ret = btrfs_map_sblock(sctx->dev_root->fs_info, WRITE,
2718 			       sparity->logic_start,
2719 			       &length, &bbio, 0, &raid_map);
2720 	if (ret || !bbio || !raid_map)
2721 		goto bbio_out;
2722 
2723 	bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2724 	if (!bio)
2725 		goto bbio_out;
2726 
2727 	bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2728 	bio->bi_private = sparity;
2729 	bio->bi_end_io = scrub_parity_bio_endio;
2730 
2731 	rbio = raid56_parity_alloc_scrub_rbio(sctx->dev_root, bio, bbio,
2732 					      raid_map, length,
2733 					      sparity->scrub_dev,
2734 					      sparity->dbitmap,
2735 					      sparity->nsectors);
2736 	if (!rbio)
2737 		goto rbio_out;
2738 
2739 	list_for_each_entry(spage, &sparity->spages, list)
2740 		raid56_parity_add_scrub_pages(rbio, spage->page,
2741 					      spage->logical);
2742 
2743 	scrub_pending_bio_inc(sctx);
2744 	raid56_parity_submit_scrub_rbio(rbio);
2745 	return;
2746 
2747 rbio_out:
2748 	bio_put(bio);
2749 bbio_out:
2750 	kfree(bbio);
2751 	kfree(raid_map);
2752 	bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2753 		  sparity->nsectors);
2754 	spin_lock(&sctx->stat_lock);
2755 	sctx->stat.malloc_errors++;
2756 	spin_unlock(&sctx->stat_lock);
2757 out:
2758 	scrub_free_parity(sparity);
2759 }
2760 
2761 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2762 {
2763 	return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * (BITS_PER_LONG / 8);
2764 }
2765 
2766 static void scrub_parity_get(struct scrub_parity *sparity)
2767 {
2768 	atomic_inc(&sparity->ref_count);
2769 }
2770 
2771 static void scrub_parity_put(struct scrub_parity *sparity)
2772 {
2773 	if (!atomic_dec_and_test(&sparity->ref_count))
2774 		return;
2775 
2776 	scrub_parity_check_and_repair(sparity);
2777 }
2778 
2779 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2780 						  struct map_lookup *map,
2781 						  struct btrfs_device *sdev,
2782 						  struct btrfs_path *path,
2783 						  u64 logic_start,
2784 						  u64 logic_end)
2785 {
2786 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2787 	struct btrfs_root *root = fs_info->extent_root;
2788 	struct btrfs_root *csum_root = fs_info->csum_root;
2789 	struct btrfs_extent_item *extent;
2790 	u64 flags;
2791 	int ret;
2792 	int slot;
2793 	struct extent_buffer *l;
2794 	struct btrfs_key key;
2795 	u64 generation;
2796 	u64 extent_logical;
2797 	u64 extent_physical;
2798 	u64 extent_len;
2799 	struct btrfs_device *extent_dev;
2800 	struct scrub_parity *sparity;
2801 	int nsectors;
2802 	int bitmap_len;
2803 	int extent_mirror_num;
2804 	int stop_loop = 0;
2805 
2806 	nsectors = map->stripe_len / root->sectorsize;
2807 	bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2808 	sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2809 			  GFP_NOFS);
2810 	if (!sparity) {
2811 		spin_lock(&sctx->stat_lock);
2812 		sctx->stat.malloc_errors++;
2813 		spin_unlock(&sctx->stat_lock);
2814 		return -ENOMEM;
2815 	}
2816 
2817 	sparity->stripe_len = map->stripe_len;
2818 	sparity->nsectors = nsectors;
2819 	sparity->sctx = sctx;
2820 	sparity->scrub_dev = sdev;
2821 	sparity->logic_start = logic_start;
2822 	sparity->logic_end = logic_end;
2823 	atomic_set(&sparity->ref_count, 1);
2824 	INIT_LIST_HEAD(&sparity->spages);
2825 	sparity->dbitmap = sparity->bitmap;
2826 	sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2827 
2828 	ret = 0;
2829 	while (logic_start < logic_end) {
2830 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2831 			key.type = BTRFS_METADATA_ITEM_KEY;
2832 		else
2833 			key.type = BTRFS_EXTENT_ITEM_KEY;
2834 		key.objectid = logic_start;
2835 		key.offset = (u64)-1;
2836 
2837 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2838 		if (ret < 0)
2839 			goto out;
2840 
2841 		if (ret > 0) {
2842 			ret = btrfs_previous_extent_item(root, path, 0);
2843 			if (ret < 0)
2844 				goto out;
2845 			if (ret > 0) {
2846 				btrfs_release_path(path);
2847 				ret = btrfs_search_slot(NULL, root, &key,
2848 							path, 0, 0);
2849 				if (ret < 0)
2850 					goto out;
2851 			}
2852 		}
2853 
2854 		stop_loop = 0;
2855 		while (1) {
2856 			u64 bytes;
2857 
2858 			l = path->nodes[0];
2859 			slot = path->slots[0];
2860 			if (slot >= btrfs_header_nritems(l)) {
2861 				ret = btrfs_next_leaf(root, path);
2862 				if (ret == 0)
2863 					continue;
2864 				if (ret < 0)
2865 					goto out;
2866 
2867 				stop_loop = 1;
2868 				break;
2869 			}
2870 			btrfs_item_key_to_cpu(l, &key, slot);
2871 
2872 			if (key.type == BTRFS_METADATA_ITEM_KEY)
2873 				bytes = root->nodesize;
2874 			else
2875 				bytes = key.offset;
2876 
2877 			if (key.objectid + bytes <= logic_start)
2878 				goto next;
2879 
2880 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2881 			    key.type != BTRFS_METADATA_ITEM_KEY)
2882 				goto next;
2883 
2884 			if (key.objectid > logic_end) {
2885 				stop_loop = 1;
2886 				break;
2887 			}
2888 
2889 			while (key.objectid >= logic_start + map->stripe_len)
2890 				logic_start += map->stripe_len;
2891 
2892 			extent = btrfs_item_ptr(l, slot,
2893 						struct btrfs_extent_item);
2894 			flags = btrfs_extent_flags(l, extent);
2895 			generation = btrfs_extent_generation(l, extent);
2896 
2897 			if (key.objectid < logic_start &&
2898 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2899 				btrfs_err(fs_info,
2900 					  "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2901 					   key.objectid, logic_start);
2902 				goto next;
2903 			}
2904 again:
2905 			extent_logical = key.objectid;
2906 			extent_len = bytes;
2907 
2908 			if (extent_logical < logic_start) {
2909 				extent_len -= logic_start - extent_logical;
2910 				extent_logical = logic_start;
2911 			}
2912 
2913 			if (extent_logical + extent_len >
2914 			    logic_start + map->stripe_len)
2915 				extent_len = logic_start + map->stripe_len -
2916 					     extent_logical;
2917 
2918 			scrub_parity_mark_sectors_data(sparity, extent_logical,
2919 						       extent_len);
2920 
2921 			scrub_remap_extent(fs_info, extent_logical,
2922 					   extent_len, &extent_physical,
2923 					   &extent_dev,
2924 					   &extent_mirror_num);
2925 
2926 			ret = btrfs_lookup_csums_range(csum_root,
2927 						extent_logical,
2928 						extent_logical + extent_len - 1,
2929 						&sctx->csum_list, 1);
2930 			if (ret)
2931 				goto out;
2932 
2933 			ret = scrub_extent_for_parity(sparity, extent_logical,
2934 						      extent_len,
2935 						      extent_physical,
2936 						      extent_dev, flags,
2937 						      generation,
2938 						      extent_mirror_num);
2939 			if (ret)
2940 				goto out;
2941 
2942 			scrub_free_csums(sctx);
2943 			if (extent_logical + extent_len <
2944 			    key.objectid + bytes) {
2945 				logic_start += map->stripe_len;
2946 
2947 				if (logic_start >= logic_end) {
2948 					stop_loop = 1;
2949 					break;
2950 				}
2951 
2952 				if (logic_start < key.objectid + bytes) {
2953 					cond_resched();
2954 					goto again;
2955 				}
2956 			}
2957 next:
2958 			path->slots[0]++;
2959 		}
2960 
2961 		btrfs_release_path(path);
2962 
2963 		if (stop_loop)
2964 			break;
2965 
2966 		logic_start += map->stripe_len;
2967 	}
2968 out:
2969 	if (ret < 0)
2970 		scrub_parity_mark_sectors_error(sparity, logic_start,
2971 						logic_end - logic_start + 1);
2972 	scrub_parity_put(sparity);
2973 	scrub_submit(sctx);
2974 	mutex_lock(&sctx->wr_ctx.wr_lock);
2975 	scrub_wr_submit(sctx);
2976 	mutex_unlock(&sctx->wr_ctx.wr_lock);
2977 
2978 	btrfs_release_path(path);
2979 	return ret < 0 ? ret : 0;
2980 }
2981 
2982 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2983 					   struct map_lookup *map,
2984 					   struct btrfs_device *scrub_dev,
2985 					   int num, u64 base, u64 length,
2986 					   int is_dev_replace)
2987 {
2988 	struct btrfs_path *path, *ppath;
2989 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2990 	struct btrfs_root *root = fs_info->extent_root;
2991 	struct btrfs_root *csum_root = fs_info->csum_root;
2992 	struct btrfs_extent_item *extent;
2993 	struct blk_plug plug;
2994 	u64 flags;
2995 	int ret;
2996 	int slot;
2997 	u64 nstripes;
2998 	struct extent_buffer *l;
2999 	struct btrfs_key key;
3000 	u64 physical;
3001 	u64 logical;
3002 	u64 logic_end;
3003 	u64 physical_end;
3004 	u64 generation;
3005 	int mirror_num;
3006 	struct reada_control *reada1;
3007 	struct reada_control *reada2;
3008 	struct btrfs_key key_start;
3009 	struct btrfs_key key_end;
3010 	u64 increment = map->stripe_len;
3011 	u64 offset;
3012 	u64 extent_logical;
3013 	u64 extent_physical;
3014 	u64 extent_len;
3015 	u64 stripe_logical;
3016 	u64 stripe_end;
3017 	struct btrfs_device *extent_dev;
3018 	int extent_mirror_num;
3019 	int stop_loop = 0;
3020 
3021 	nstripes = length;
3022 	physical = map->stripes[num].physical;
3023 	offset = 0;
3024 	do_div(nstripes, map->stripe_len);
3025 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3026 		offset = map->stripe_len * num;
3027 		increment = map->stripe_len * map->num_stripes;
3028 		mirror_num = 1;
3029 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3030 		int factor = map->num_stripes / map->sub_stripes;
3031 		offset = map->stripe_len * (num / map->sub_stripes);
3032 		increment = map->stripe_len * factor;
3033 		mirror_num = num % map->sub_stripes + 1;
3034 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3035 		increment = map->stripe_len;
3036 		mirror_num = num % map->num_stripes + 1;
3037 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3038 		increment = map->stripe_len;
3039 		mirror_num = num % map->num_stripes + 1;
3040 	} else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
3041 				BTRFS_BLOCK_GROUP_RAID6)) {
3042 		get_raid56_logic_offset(physical, num, map, &offset, NULL);
3043 		increment = map->stripe_len * nr_data_stripes(map);
3044 		mirror_num = 1;
3045 	} else {
3046 		increment = map->stripe_len;
3047 		mirror_num = 1;
3048 	}
3049 
3050 	path = btrfs_alloc_path();
3051 	if (!path)
3052 		return -ENOMEM;
3053 
3054 	ppath = btrfs_alloc_path();
3055 	if (!ppath) {
3056 		btrfs_free_path(ppath);
3057 		return -ENOMEM;
3058 	}
3059 
3060 	/*
3061 	 * work on commit root. The related disk blocks are static as
3062 	 * long as COW is applied. This means, it is save to rewrite
3063 	 * them to repair disk errors without any race conditions
3064 	 */
3065 	path->search_commit_root = 1;
3066 	path->skip_locking = 1;
3067 
3068 	/*
3069 	 * trigger the readahead for extent tree csum tree and wait for
3070 	 * completion. During readahead, the scrub is officially paused
3071 	 * to not hold off transaction commits
3072 	 */
3073 	logical = base + offset;
3074 	physical_end = physical + nstripes * map->stripe_len;
3075 	if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
3076 			 BTRFS_BLOCK_GROUP_RAID6)) {
3077 		get_raid56_logic_offset(physical_end, num,
3078 					map, &logic_end, NULL);
3079 		logic_end += base;
3080 	} else {
3081 		logic_end = logical + increment * nstripes;
3082 	}
3083 	wait_event(sctx->list_wait,
3084 		   atomic_read(&sctx->bios_in_flight) == 0);
3085 	scrub_blocked_if_needed(fs_info);
3086 
3087 	/* FIXME it might be better to start readahead at commit root */
3088 	key_start.objectid = logical;
3089 	key_start.type = BTRFS_EXTENT_ITEM_KEY;
3090 	key_start.offset = (u64)0;
3091 	key_end.objectid = logic_end;
3092 	key_end.type = BTRFS_METADATA_ITEM_KEY;
3093 	key_end.offset = (u64)-1;
3094 	reada1 = btrfs_reada_add(root, &key_start, &key_end);
3095 
3096 	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3097 	key_start.type = BTRFS_EXTENT_CSUM_KEY;
3098 	key_start.offset = logical;
3099 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3100 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
3101 	key_end.offset = logic_end;
3102 	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
3103 
3104 	if (!IS_ERR(reada1))
3105 		btrfs_reada_wait(reada1);
3106 	if (!IS_ERR(reada2))
3107 		btrfs_reada_wait(reada2);
3108 
3109 
3110 	/*
3111 	 * collect all data csums for the stripe to avoid seeking during
3112 	 * the scrub. This might currently (crc32) end up to be about 1MB
3113 	 */
3114 	blk_start_plug(&plug);
3115 
3116 	/*
3117 	 * now find all extents for each stripe and scrub them
3118 	 */
3119 	ret = 0;
3120 	while (physical < physical_end) {
3121 		/* for raid56, we skip parity stripe */
3122 		if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
3123 				BTRFS_BLOCK_GROUP_RAID6)) {
3124 			ret = get_raid56_logic_offset(physical, num,
3125 					map, &logical, &stripe_logical);
3126 			logical += base;
3127 			if (ret) {
3128 				stripe_logical += base;
3129 				stripe_end = stripe_logical + increment - 1;
3130 				ret = scrub_raid56_parity(sctx, map, scrub_dev,
3131 						ppath, stripe_logical,
3132 						stripe_end);
3133 				if (ret)
3134 					goto out;
3135 				goto skip;
3136 			}
3137 		}
3138 		/*
3139 		 * canceled?
3140 		 */
3141 		if (atomic_read(&fs_info->scrub_cancel_req) ||
3142 		    atomic_read(&sctx->cancel_req)) {
3143 			ret = -ECANCELED;
3144 			goto out;
3145 		}
3146 		/*
3147 		 * check to see if we have to pause
3148 		 */
3149 		if (atomic_read(&fs_info->scrub_pause_req)) {
3150 			/* push queued extents */
3151 			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3152 			scrub_submit(sctx);
3153 			mutex_lock(&sctx->wr_ctx.wr_lock);
3154 			scrub_wr_submit(sctx);
3155 			mutex_unlock(&sctx->wr_ctx.wr_lock);
3156 			wait_event(sctx->list_wait,
3157 				   atomic_read(&sctx->bios_in_flight) == 0);
3158 			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3159 			scrub_blocked_if_needed(fs_info);
3160 		}
3161 
3162 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3163 			key.type = BTRFS_METADATA_ITEM_KEY;
3164 		else
3165 			key.type = BTRFS_EXTENT_ITEM_KEY;
3166 		key.objectid = logical;
3167 		key.offset = (u64)-1;
3168 
3169 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3170 		if (ret < 0)
3171 			goto out;
3172 
3173 		if (ret > 0) {
3174 			ret = btrfs_previous_extent_item(root, path, 0);
3175 			if (ret < 0)
3176 				goto out;
3177 			if (ret > 0) {
3178 				/* there's no smaller item, so stick with the
3179 				 * larger one */
3180 				btrfs_release_path(path);
3181 				ret = btrfs_search_slot(NULL, root, &key,
3182 							path, 0, 0);
3183 				if (ret < 0)
3184 					goto out;
3185 			}
3186 		}
3187 
3188 		stop_loop = 0;
3189 		while (1) {
3190 			u64 bytes;
3191 
3192 			l = path->nodes[0];
3193 			slot = path->slots[0];
3194 			if (slot >= btrfs_header_nritems(l)) {
3195 				ret = btrfs_next_leaf(root, path);
3196 				if (ret == 0)
3197 					continue;
3198 				if (ret < 0)
3199 					goto out;
3200 
3201 				stop_loop = 1;
3202 				break;
3203 			}
3204 			btrfs_item_key_to_cpu(l, &key, slot);
3205 
3206 			if (key.type == BTRFS_METADATA_ITEM_KEY)
3207 				bytes = root->nodesize;
3208 			else
3209 				bytes = key.offset;
3210 
3211 			if (key.objectid + bytes <= logical)
3212 				goto next;
3213 
3214 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3215 			    key.type != BTRFS_METADATA_ITEM_KEY)
3216 				goto next;
3217 
3218 			if (key.objectid >= logical + map->stripe_len) {
3219 				/* out of this device extent */
3220 				if (key.objectid >= logic_end)
3221 					stop_loop = 1;
3222 				break;
3223 			}
3224 
3225 			extent = btrfs_item_ptr(l, slot,
3226 						struct btrfs_extent_item);
3227 			flags = btrfs_extent_flags(l, extent);
3228 			generation = btrfs_extent_generation(l, extent);
3229 
3230 			if (key.objectid < logical &&
3231 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
3232 				btrfs_err(fs_info,
3233 					   "scrub: tree block %llu spanning "
3234 					   "stripes, ignored. logical=%llu",
3235 				       key.objectid, logical);
3236 				goto next;
3237 			}
3238 
3239 again:
3240 			extent_logical = key.objectid;
3241 			extent_len = bytes;
3242 
3243 			/*
3244 			 * trim extent to this stripe
3245 			 */
3246 			if (extent_logical < logical) {
3247 				extent_len -= logical - extent_logical;
3248 				extent_logical = logical;
3249 			}
3250 			if (extent_logical + extent_len >
3251 			    logical + map->stripe_len) {
3252 				extent_len = logical + map->stripe_len -
3253 					     extent_logical;
3254 			}
3255 
3256 			extent_physical = extent_logical - logical + physical;
3257 			extent_dev = scrub_dev;
3258 			extent_mirror_num = mirror_num;
3259 			if (is_dev_replace)
3260 				scrub_remap_extent(fs_info, extent_logical,
3261 						   extent_len, &extent_physical,
3262 						   &extent_dev,
3263 						   &extent_mirror_num);
3264 
3265 			ret = btrfs_lookup_csums_range(csum_root, logical,
3266 						logical + map->stripe_len - 1,
3267 						&sctx->csum_list, 1);
3268 			if (ret)
3269 				goto out;
3270 
3271 			ret = scrub_extent(sctx, extent_logical, extent_len,
3272 					   extent_physical, extent_dev, flags,
3273 					   generation, extent_mirror_num,
3274 					   extent_logical - logical + physical);
3275 			if (ret)
3276 				goto out;
3277 
3278 			scrub_free_csums(sctx);
3279 			if (extent_logical + extent_len <
3280 			    key.objectid + bytes) {
3281 				if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
3282 					BTRFS_BLOCK_GROUP_RAID6)) {
3283 					/*
3284 					 * loop until we find next data stripe
3285 					 * or we have finished all stripes.
3286 					 */
3287 loop:
3288 					physical += map->stripe_len;
3289 					ret = get_raid56_logic_offset(physical,
3290 							num, map, &logical,
3291 							&stripe_logical);
3292 					logical += base;
3293 
3294 					if (ret && physical < physical_end) {
3295 						stripe_logical += base;
3296 						stripe_end = stripe_logical +
3297 								increment - 1;
3298 						ret = scrub_raid56_parity(sctx,
3299 							map, scrub_dev, ppath,
3300 							stripe_logical,
3301 							stripe_end);
3302 						if (ret)
3303 							goto out;
3304 						goto loop;
3305 					}
3306 				} else {
3307 					physical += map->stripe_len;
3308 					logical += increment;
3309 				}
3310 				if (logical < key.objectid + bytes) {
3311 					cond_resched();
3312 					goto again;
3313 				}
3314 
3315 				if (physical >= physical_end) {
3316 					stop_loop = 1;
3317 					break;
3318 				}
3319 			}
3320 next:
3321 			path->slots[0]++;
3322 		}
3323 		btrfs_release_path(path);
3324 skip:
3325 		logical += increment;
3326 		physical += map->stripe_len;
3327 		spin_lock(&sctx->stat_lock);
3328 		if (stop_loop)
3329 			sctx->stat.last_physical = map->stripes[num].physical +
3330 						   length;
3331 		else
3332 			sctx->stat.last_physical = physical;
3333 		spin_unlock(&sctx->stat_lock);
3334 		if (stop_loop)
3335 			break;
3336 	}
3337 out:
3338 	/* push queued extents */
3339 	scrub_submit(sctx);
3340 	mutex_lock(&sctx->wr_ctx.wr_lock);
3341 	scrub_wr_submit(sctx);
3342 	mutex_unlock(&sctx->wr_ctx.wr_lock);
3343 
3344 	blk_finish_plug(&plug);
3345 	btrfs_free_path(path);
3346 	btrfs_free_path(ppath);
3347 	return ret < 0 ? ret : 0;
3348 }
3349 
3350 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3351 					  struct btrfs_device *scrub_dev,
3352 					  u64 chunk_tree, u64 chunk_objectid,
3353 					  u64 chunk_offset, u64 length,
3354 					  u64 dev_offset, int is_dev_replace)
3355 {
3356 	struct btrfs_mapping_tree *map_tree =
3357 		&sctx->dev_root->fs_info->mapping_tree;
3358 	struct map_lookup *map;
3359 	struct extent_map *em;
3360 	int i;
3361 	int ret = 0;
3362 
3363 	read_lock(&map_tree->map_tree.lock);
3364 	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3365 	read_unlock(&map_tree->map_tree.lock);
3366 
3367 	if (!em)
3368 		return -EINVAL;
3369 
3370 	map = (struct map_lookup *)em->bdev;
3371 	if (em->start != chunk_offset)
3372 		goto out;
3373 
3374 	if (em->len < length)
3375 		goto out;
3376 
3377 	for (i = 0; i < map->num_stripes; ++i) {
3378 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3379 		    map->stripes[i].physical == dev_offset) {
3380 			ret = scrub_stripe(sctx, map, scrub_dev, i,
3381 					   chunk_offset, length,
3382 					   is_dev_replace);
3383 			if (ret)
3384 				goto out;
3385 		}
3386 	}
3387 out:
3388 	free_extent_map(em);
3389 
3390 	return ret;
3391 }
3392 
3393 static noinline_for_stack
3394 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3395 			   struct btrfs_device *scrub_dev, u64 start, u64 end,
3396 			   int is_dev_replace)
3397 {
3398 	struct btrfs_dev_extent *dev_extent = NULL;
3399 	struct btrfs_path *path;
3400 	struct btrfs_root *root = sctx->dev_root;
3401 	struct btrfs_fs_info *fs_info = root->fs_info;
3402 	u64 length;
3403 	u64 chunk_tree;
3404 	u64 chunk_objectid;
3405 	u64 chunk_offset;
3406 	int ret;
3407 	int slot;
3408 	struct extent_buffer *l;
3409 	struct btrfs_key key;
3410 	struct btrfs_key found_key;
3411 	struct btrfs_block_group_cache *cache;
3412 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3413 
3414 	path = btrfs_alloc_path();
3415 	if (!path)
3416 		return -ENOMEM;
3417 
3418 	path->reada = 2;
3419 	path->search_commit_root = 1;
3420 	path->skip_locking = 1;
3421 
3422 	key.objectid = scrub_dev->devid;
3423 	key.offset = 0ull;
3424 	key.type = BTRFS_DEV_EXTENT_KEY;
3425 
3426 	while (1) {
3427 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3428 		if (ret < 0)
3429 			break;
3430 		if (ret > 0) {
3431 			if (path->slots[0] >=
3432 			    btrfs_header_nritems(path->nodes[0])) {
3433 				ret = btrfs_next_leaf(root, path);
3434 				if (ret)
3435 					break;
3436 			}
3437 		}
3438 
3439 		l = path->nodes[0];
3440 		slot = path->slots[0];
3441 
3442 		btrfs_item_key_to_cpu(l, &found_key, slot);
3443 
3444 		if (found_key.objectid != scrub_dev->devid)
3445 			break;
3446 
3447 		if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3448 			break;
3449 
3450 		if (found_key.offset >= end)
3451 			break;
3452 
3453 		if (found_key.offset < key.offset)
3454 			break;
3455 
3456 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3457 		length = btrfs_dev_extent_length(l, dev_extent);
3458 
3459 		if (found_key.offset + length <= start)
3460 			goto skip;
3461 
3462 		chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
3463 		chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
3464 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3465 
3466 		/*
3467 		 * get a reference on the corresponding block group to prevent
3468 		 * the chunk from going away while we scrub it
3469 		 */
3470 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3471 
3472 		/* some chunks are removed but not committed to disk yet,
3473 		 * continue scrubbing */
3474 		if (!cache)
3475 			goto skip;
3476 
3477 		dev_replace->cursor_right = found_key.offset + length;
3478 		dev_replace->cursor_left = found_key.offset;
3479 		dev_replace->item_needs_writeback = 1;
3480 		ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
3481 				  chunk_offset, length, found_key.offset,
3482 				  is_dev_replace);
3483 
3484 		/*
3485 		 * flush, submit all pending read and write bios, afterwards
3486 		 * wait for them.
3487 		 * Note that in the dev replace case, a read request causes
3488 		 * write requests that are submitted in the read completion
3489 		 * worker. Therefore in the current situation, it is required
3490 		 * that all write requests are flushed, so that all read and
3491 		 * write requests are really completed when bios_in_flight
3492 		 * changes to 0.
3493 		 */
3494 		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3495 		scrub_submit(sctx);
3496 		mutex_lock(&sctx->wr_ctx.wr_lock);
3497 		scrub_wr_submit(sctx);
3498 		mutex_unlock(&sctx->wr_ctx.wr_lock);
3499 
3500 		wait_event(sctx->list_wait,
3501 			   atomic_read(&sctx->bios_in_flight) == 0);
3502 		atomic_inc(&fs_info->scrubs_paused);
3503 		wake_up(&fs_info->scrub_pause_wait);
3504 
3505 		/*
3506 		 * must be called before we decrease @scrub_paused.
3507 		 * make sure we don't block transaction commit while
3508 		 * we are waiting pending workers finished.
3509 		 */
3510 		wait_event(sctx->list_wait,
3511 			   atomic_read(&sctx->workers_pending) == 0);
3512 		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3513 
3514 		mutex_lock(&fs_info->scrub_lock);
3515 		__scrub_blocked_if_needed(fs_info);
3516 		atomic_dec(&fs_info->scrubs_paused);
3517 		mutex_unlock(&fs_info->scrub_lock);
3518 		wake_up(&fs_info->scrub_pause_wait);
3519 
3520 		btrfs_put_block_group(cache);
3521 		if (ret)
3522 			break;
3523 		if (is_dev_replace &&
3524 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
3525 			ret = -EIO;
3526 			break;
3527 		}
3528 		if (sctx->stat.malloc_errors > 0) {
3529 			ret = -ENOMEM;
3530 			break;
3531 		}
3532 
3533 		dev_replace->cursor_left = dev_replace->cursor_right;
3534 		dev_replace->item_needs_writeback = 1;
3535 skip:
3536 		key.offset = found_key.offset + length;
3537 		btrfs_release_path(path);
3538 	}
3539 
3540 	btrfs_free_path(path);
3541 
3542 	/*
3543 	 * ret can still be 1 from search_slot or next_leaf,
3544 	 * that's not an error
3545 	 */
3546 	return ret < 0 ? ret : 0;
3547 }
3548 
3549 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3550 					   struct btrfs_device *scrub_dev)
3551 {
3552 	int	i;
3553 	u64	bytenr;
3554 	u64	gen;
3555 	int	ret;
3556 	struct btrfs_root *root = sctx->dev_root;
3557 
3558 	if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
3559 		return -EIO;
3560 
3561 	/* Seed devices of a new filesystem has their own generation. */
3562 	if (scrub_dev->fs_devices != root->fs_info->fs_devices)
3563 		gen = scrub_dev->generation;
3564 	else
3565 		gen = root->fs_info->last_trans_committed;
3566 
3567 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3568 		bytenr = btrfs_sb_offset(i);
3569 		if (bytenr + BTRFS_SUPER_INFO_SIZE >
3570 		    scrub_dev->commit_total_bytes)
3571 			break;
3572 
3573 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3574 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3575 				  NULL, 1, bytenr);
3576 		if (ret)
3577 			return ret;
3578 	}
3579 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3580 
3581 	return 0;
3582 }
3583 
3584 /*
3585  * get a reference count on fs_info->scrub_workers. start worker if necessary
3586  */
3587 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3588 						int is_dev_replace)
3589 {
3590 	int ret = 0;
3591 	int flags = WQ_FREEZABLE | WQ_UNBOUND;
3592 	int max_active = fs_info->thread_pool_size;
3593 
3594 	if (fs_info->scrub_workers_refcnt == 0) {
3595 		if (is_dev_replace)
3596 			fs_info->scrub_workers =
3597 				btrfs_alloc_workqueue("btrfs-scrub", flags,
3598 						      1, 4);
3599 		else
3600 			fs_info->scrub_workers =
3601 				btrfs_alloc_workqueue("btrfs-scrub", flags,
3602 						      max_active, 4);
3603 		if (!fs_info->scrub_workers) {
3604 			ret = -ENOMEM;
3605 			goto out;
3606 		}
3607 		fs_info->scrub_wr_completion_workers =
3608 			btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
3609 					      max_active, 2);
3610 		if (!fs_info->scrub_wr_completion_workers) {
3611 			ret = -ENOMEM;
3612 			goto out;
3613 		}
3614 		fs_info->scrub_nocow_workers =
3615 			btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
3616 		if (!fs_info->scrub_nocow_workers) {
3617 			ret = -ENOMEM;
3618 			goto out;
3619 		}
3620 	}
3621 	++fs_info->scrub_workers_refcnt;
3622 out:
3623 	return ret;
3624 }
3625 
3626 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3627 {
3628 	if (--fs_info->scrub_workers_refcnt == 0) {
3629 		btrfs_destroy_workqueue(fs_info->scrub_workers);
3630 		btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3631 		btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3632 	}
3633 	WARN_ON(fs_info->scrub_workers_refcnt < 0);
3634 }
3635 
3636 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3637 		    u64 end, struct btrfs_scrub_progress *progress,
3638 		    int readonly, int is_dev_replace)
3639 {
3640 	struct scrub_ctx *sctx;
3641 	int ret;
3642 	struct btrfs_device *dev;
3643 	struct rcu_string *name;
3644 
3645 	if (btrfs_fs_closing(fs_info))
3646 		return -EINVAL;
3647 
3648 	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
3649 		/*
3650 		 * in this case scrub is unable to calculate the checksum
3651 		 * the way scrub is implemented. Do not handle this
3652 		 * situation at all because it won't ever happen.
3653 		 */
3654 		btrfs_err(fs_info,
3655 			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3656 		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
3657 		return -EINVAL;
3658 	}
3659 
3660 	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
3661 		/* not supported for data w/o checksums */
3662 		btrfs_err(fs_info,
3663 			   "scrub: size assumption sectorsize != PAGE_SIZE "
3664 			   "(%d != %lu) fails",
3665 		       fs_info->chunk_root->sectorsize, PAGE_SIZE);
3666 		return -EINVAL;
3667 	}
3668 
3669 	if (fs_info->chunk_root->nodesize >
3670 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3671 	    fs_info->chunk_root->sectorsize >
3672 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3673 		/*
3674 		 * would exhaust the array bounds of pagev member in
3675 		 * struct scrub_block
3676 		 */
3677 		btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
3678 			   "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3679 		       fs_info->chunk_root->nodesize,
3680 		       SCRUB_MAX_PAGES_PER_BLOCK,
3681 		       fs_info->chunk_root->sectorsize,
3682 		       SCRUB_MAX_PAGES_PER_BLOCK);
3683 		return -EINVAL;
3684 	}
3685 
3686 
3687 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3688 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3689 	if (!dev || (dev->missing && !is_dev_replace)) {
3690 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3691 		return -ENODEV;
3692 	}
3693 
3694 	if (!is_dev_replace && !readonly && !dev->writeable) {
3695 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3696 		rcu_read_lock();
3697 		name = rcu_dereference(dev->name);
3698 		btrfs_err(fs_info, "scrub: device %s is not writable",
3699 			  name->str);
3700 		rcu_read_unlock();
3701 		return -EROFS;
3702 	}
3703 
3704 	mutex_lock(&fs_info->scrub_lock);
3705 	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
3706 		mutex_unlock(&fs_info->scrub_lock);
3707 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3708 		return -EIO;
3709 	}
3710 
3711 	btrfs_dev_replace_lock(&fs_info->dev_replace);
3712 	if (dev->scrub_device ||
3713 	    (!is_dev_replace &&
3714 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3715 		btrfs_dev_replace_unlock(&fs_info->dev_replace);
3716 		mutex_unlock(&fs_info->scrub_lock);
3717 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3718 		return -EINPROGRESS;
3719 	}
3720 	btrfs_dev_replace_unlock(&fs_info->dev_replace);
3721 
3722 	ret = scrub_workers_get(fs_info, is_dev_replace);
3723 	if (ret) {
3724 		mutex_unlock(&fs_info->scrub_lock);
3725 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3726 		return ret;
3727 	}
3728 
3729 	sctx = scrub_setup_ctx(dev, is_dev_replace);
3730 	if (IS_ERR(sctx)) {
3731 		mutex_unlock(&fs_info->scrub_lock);
3732 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3733 		scrub_workers_put(fs_info);
3734 		return PTR_ERR(sctx);
3735 	}
3736 	sctx->readonly = readonly;
3737 	dev->scrub_device = sctx;
3738 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3739 
3740 	/*
3741 	 * checking @scrub_pause_req here, we can avoid
3742 	 * race between committing transaction and scrubbing.
3743 	 */
3744 	__scrub_blocked_if_needed(fs_info);
3745 	atomic_inc(&fs_info->scrubs_running);
3746 	mutex_unlock(&fs_info->scrub_lock);
3747 
3748 	if (!is_dev_replace) {
3749 		/*
3750 		 * by holding device list mutex, we can
3751 		 * kick off writing super in log tree sync.
3752 		 */
3753 		mutex_lock(&fs_info->fs_devices->device_list_mutex);
3754 		ret = scrub_supers(sctx, dev);
3755 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3756 	}
3757 
3758 	if (!ret)
3759 		ret = scrub_enumerate_chunks(sctx, dev, start, end,
3760 					     is_dev_replace);
3761 
3762 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3763 	atomic_dec(&fs_info->scrubs_running);
3764 	wake_up(&fs_info->scrub_pause_wait);
3765 
3766 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3767 
3768 	if (progress)
3769 		memcpy(progress, &sctx->stat, sizeof(*progress));
3770 
3771 	mutex_lock(&fs_info->scrub_lock);
3772 	dev->scrub_device = NULL;
3773 	scrub_workers_put(fs_info);
3774 	mutex_unlock(&fs_info->scrub_lock);
3775 
3776 	scrub_free_ctx(sctx);
3777 
3778 	return ret;
3779 }
3780 
3781 void btrfs_scrub_pause(struct btrfs_root *root)
3782 {
3783 	struct btrfs_fs_info *fs_info = root->fs_info;
3784 
3785 	mutex_lock(&fs_info->scrub_lock);
3786 	atomic_inc(&fs_info->scrub_pause_req);
3787 	while (atomic_read(&fs_info->scrubs_paused) !=
3788 	       atomic_read(&fs_info->scrubs_running)) {
3789 		mutex_unlock(&fs_info->scrub_lock);
3790 		wait_event(fs_info->scrub_pause_wait,
3791 			   atomic_read(&fs_info->scrubs_paused) ==
3792 			   atomic_read(&fs_info->scrubs_running));
3793 		mutex_lock(&fs_info->scrub_lock);
3794 	}
3795 	mutex_unlock(&fs_info->scrub_lock);
3796 }
3797 
3798 void btrfs_scrub_continue(struct btrfs_root *root)
3799 {
3800 	struct btrfs_fs_info *fs_info = root->fs_info;
3801 
3802 	atomic_dec(&fs_info->scrub_pause_req);
3803 	wake_up(&fs_info->scrub_pause_wait);
3804 }
3805 
3806 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3807 {
3808 	mutex_lock(&fs_info->scrub_lock);
3809 	if (!atomic_read(&fs_info->scrubs_running)) {
3810 		mutex_unlock(&fs_info->scrub_lock);
3811 		return -ENOTCONN;
3812 	}
3813 
3814 	atomic_inc(&fs_info->scrub_cancel_req);
3815 	while (atomic_read(&fs_info->scrubs_running)) {
3816 		mutex_unlock(&fs_info->scrub_lock);
3817 		wait_event(fs_info->scrub_pause_wait,
3818 			   atomic_read(&fs_info->scrubs_running) == 0);
3819 		mutex_lock(&fs_info->scrub_lock);
3820 	}
3821 	atomic_dec(&fs_info->scrub_cancel_req);
3822 	mutex_unlock(&fs_info->scrub_lock);
3823 
3824 	return 0;
3825 }
3826 
3827 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3828 			   struct btrfs_device *dev)
3829 {
3830 	struct scrub_ctx *sctx;
3831 
3832 	mutex_lock(&fs_info->scrub_lock);
3833 	sctx = dev->scrub_device;
3834 	if (!sctx) {
3835 		mutex_unlock(&fs_info->scrub_lock);
3836 		return -ENOTCONN;
3837 	}
3838 	atomic_inc(&sctx->cancel_req);
3839 	while (dev->scrub_device) {
3840 		mutex_unlock(&fs_info->scrub_lock);
3841 		wait_event(fs_info->scrub_pause_wait,
3842 			   dev->scrub_device == NULL);
3843 		mutex_lock(&fs_info->scrub_lock);
3844 	}
3845 	mutex_unlock(&fs_info->scrub_lock);
3846 
3847 	return 0;
3848 }
3849 
3850 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3851 			 struct btrfs_scrub_progress *progress)
3852 {
3853 	struct btrfs_device *dev;
3854 	struct scrub_ctx *sctx = NULL;
3855 
3856 	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3857 	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3858 	if (dev)
3859 		sctx = dev->scrub_device;
3860 	if (sctx)
3861 		memcpy(progress, &sctx->stat, sizeof(*progress));
3862 	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3863 
3864 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3865 }
3866 
3867 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3868 			       u64 extent_logical, u64 extent_len,
3869 			       u64 *extent_physical,
3870 			       struct btrfs_device **extent_dev,
3871 			       int *extent_mirror_num)
3872 {
3873 	u64 mapped_length;
3874 	struct btrfs_bio *bbio = NULL;
3875 	int ret;
3876 
3877 	mapped_length = extent_len;
3878 	ret = btrfs_map_block(fs_info, READ, extent_logical,
3879 			      &mapped_length, &bbio, 0);
3880 	if (ret || !bbio || mapped_length < extent_len ||
3881 	    !bbio->stripes[0].dev->bdev) {
3882 		kfree(bbio);
3883 		return;
3884 	}
3885 
3886 	*extent_physical = bbio->stripes[0].physical;
3887 	*extent_mirror_num = bbio->mirror_num;
3888 	*extent_dev = bbio->stripes[0].dev;
3889 	kfree(bbio);
3890 }
3891 
3892 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3893 			      struct scrub_wr_ctx *wr_ctx,
3894 			      struct btrfs_fs_info *fs_info,
3895 			      struct btrfs_device *dev,
3896 			      int is_dev_replace)
3897 {
3898 	WARN_ON(wr_ctx->wr_curr_bio != NULL);
3899 
3900 	mutex_init(&wr_ctx->wr_lock);
3901 	wr_ctx->wr_curr_bio = NULL;
3902 	if (!is_dev_replace)
3903 		return 0;
3904 
3905 	WARN_ON(!dev->bdev);
3906 	wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3907 					 bio_get_nr_vecs(dev->bdev));
3908 	wr_ctx->tgtdev = dev;
3909 	atomic_set(&wr_ctx->flush_all_writes, 0);
3910 	return 0;
3911 }
3912 
3913 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3914 {
3915 	mutex_lock(&wr_ctx->wr_lock);
3916 	kfree(wr_ctx->wr_curr_bio);
3917 	wr_ctx->wr_curr_bio = NULL;
3918 	mutex_unlock(&wr_ctx->wr_lock);
3919 }
3920 
3921 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3922 			    int mirror_num, u64 physical_for_dev_replace)
3923 {
3924 	struct scrub_copy_nocow_ctx *nocow_ctx;
3925 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3926 
3927 	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3928 	if (!nocow_ctx) {
3929 		spin_lock(&sctx->stat_lock);
3930 		sctx->stat.malloc_errors++;
3931 		spin_unlock(&sctx->stat_lock);
3932 		return -ENOMEM;
3933 	}
3934 
3935 	scrub_pending_trans_workers_inc(sctx);
3936 
3937 	nocow_ctx->sctx = sctx;
3938 	nocow_ctx->logical = logical;
3939 	nocow_ctx->len = len;
3940 	nocow_ctx->mirror_num = mirror_num;
3941 	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3942 	btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
3943 			copy_nocow_pages_worker, NULL, NULL);
3944 	INIT_LIST_HEAD(&nocow_ctx->inodes);
3945 	btrfs_queue_work(fs_info->scrub_nocow_workers,
3946 			 &nocow_ctx->work);
3947 
3948 	return 0;
3949 }
3950 
3951 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3952 {
3953 	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3954 	struct scrub_nocow_inode *nocow_inode;
3955 
3956 	nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3957 	if (!nocow_inode)
3958 		return -ENOMEM;
3959 	nocow_inode->inum = inum;
3960 	nocow_inode->offset = offset;
3961 	nocow_inode->root = root;
3962 	list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3963 	return 0;
3964 }
3965 
3966 #define COPY_COMPLETE 1
3967 
3968 static void copy_nocow_pages_worker(struct btrfs_work *work)
3969 {
3970 	struct scrub_copy_nocow_ctx *nocow_ctx =
3971 		container_of(work, struct scrub_copy_nocow_ctx, work);
3972 	struct scrub_ctx *sctx = nocow_ctx->sctx;
3973 	u64 logical = nocow_ctx->logical;
3974 	u64 len = nocow_ctx->len;
3975 	int mirror_num = nocow_ctx->mirror_num;
3976 	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3977 	int ret;
3978 	struct btrfs_trans_handle *trans = NULL;
3979 	struct btrfs_fs_info *fs_info;
3980 	struct btrfs_path *path;
3981 	struct btrfs_root *root;
3982 	int not_written = 0;
3983 
3984 	fs_info = sctx->dev_root->fs_info;
3985 	root = fs_info->extent_root;
3986 
3987 	path = btrfs_alloc_path();
3988 	if (!path) {
3989 		spin_lock(&sctx->stat_lock);
3990 		sctx->stat.malloc_errors++;
3991 		spin_unlock(&sctx->stat_lock);
3992 		not_written = 1;
3993 		goto out;
3994 	}
3995 
3996 	trans = btrfs_join_transaction(root);
3997 	if (IS_ERR(trans)) {
3998 		not_written = 1;
3999 		goto out;
4000 	}
4001 
4002 	ret = iterate_inodes_from_logical(logical, fs_info, path,
4003 					  record_inode_for_nocow, nocow_ctx);
4004 	if (ret != 0 && ret != -ENOENT) {
4005 		btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
4006 			"phys %llu, len %llu, mir %u, ret %d",
4007 			logical, physical_for_dev_replace, len, mirror_num,
4008 			ret);
4009 		not_written = 1;
4010 		goto out;
4011 	}
4012 
4013 	btrfs_end_transaction(trans, root);
4014 	trans = NULL;
4015 	while (!list_empty(&nocow_ctx->inodes)) {
4016 		struct scrub_nocow_inode *entry;
4017 		entry = list_first_entry(&nocow_ctx->inodes,
4018 					 struct scrub_nocow_inode,
4019 					 list);
4020 		list_del_init(&entry->list);
4021 		ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4022 						 entry->root, nocow_ctx);
4023 		kfree(entry);
4024 		if (ret == COPY_COMPLETE) {
4025 			ret = 0;
4026 			break;
4027 		} else if (ret) {
4028 			break;
4029 		}
4030 	}
4031 out:
4032 	while (!list_empty(&nocow_ctx->inodes)) {
4033 		struct scrub_nocow_inode *entry;
4034 		entry = list_first_entry(&nocow_ctx->inodes,
4035 					 struct scrub_nocow_inode,
4036 					 list);
4037 		list_del_init(&entry->list);
4038 		kfree(entry);
4039 	}
4040 	if (trans && !IS_ERR(trans))
4041 		btrfs_end_transaction(trans, root);
4042 	if (not_written)
4043 		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4044 					    num_uncorrectable_read_errors);
4045 
4046 	btrfs_free_path(path);
4047 	kfree(nocow_ctx);
4048 
4049 	scrub_pending_trans_workers_dec(sctx);
4050 }
4051 
4052 static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
4053 				 u64 logical)
4054 {
4055 	struct extent_state *cached_state = NULL;
4056 	struct btrfs_ordered_extent *ordered;
4057 	struct extent_io_tree *io_tree;
4058 	struct extent_map *em;
4059 	u64 lockstart = start, lockend = start + len - 1;
4060 	int ret = 0;
4061 
4062 	io_tree = &BTRFS_I(inode)->io_tree;
4063 
4064 	lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
4065 	ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4066 	if (ordered) {
4067 		btrfs_put_ordered_extent(ordered);
4068 		ret = 1;
4069 		goto out_unlock;
4070 	}
4071 
4072 	em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4073 	if (IS_ERR(em)) {
4074 		ret = PTR_ERR(em);
4075 		goto out_unlock;
4076 	}
4077 
4078 	/*
4079 	 * This extent does not actually cover the logical extent anymore,
4080 	 * move on to the next inode.
4081 	 */
4082 	if (em->block_start > logical ||
4083 	    em->block_start + em->block_len < logical + len) {
4084 		free_extent_map(em);
4085 		ret = 1;
4086 		goto out_unlock;
4087 	}
4088 	free_extent_map(em);
4089 
4090 out_unlock:
4091 	unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4092 			     GFP_NOFS);
4093 	return ret;
4094 }
4095 
4096 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4097 				      struct scrub_copy_nocow_ctx *nocow_ctx)
4098 {
4099 	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
4100 	struct btrfs_key key;
4101 	struct inode *inode;
4102 	struct page *page;
4103 	struct btrfs_root *local_root;
4104 	struct extent_io_tree *io_tree;
4105 	u64 physical_for_dev_replace;
4106 	u64 nocow_ctx_logical;
4107 	u64 len = nocow_ctx->len;
4108 	unsigned long index;
4109 	int srcu_index;
4110 	int ret = 0;
4111 	int err = 0;
4112 
4113 	key.objectid = root;
4114 	key.type = BTRFS_ROOT_ITEM_KEY;
4115 	key.offset = (u64)-1;
4116 
4117 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4118 
4119 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4120 	if (IS_ERR(local_root)) {
4121 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4122 		return PTR_ERR(local_root);
4123 	}
4124 
4125 	key.type = BTRFS_INODE_ITEM_KEY;
4126 	key.objectid = inum;
4127 	key.offset = 0;
4128 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4129 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4130 	if (IS_ERR(inode))
4131 		return PTR_ERR(inode);
4132 
4133 	/* Avoid truncate/dio/punch hole.. */
4134 	mutex_lock(&inode->i_mutex);
4135 	inode_dio_wait(inode);
4136 
4137 	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4138 	io_tree = &BTRFS_I(inode)->io_tree;
4139 	nocow_ctx_logical = nocow_ctx->logical;
4140 
4141 	ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
4142 	if (ret) {
4143 		ret = ret > 0 ? 0 : ret;
4144 		goto out;
4145 	}
4146 
4147 	while (len >= PAGE_CACHE_SIZE) {
4148 		index = offset >> PAGE_CACHE_SHIFT;
4149 again:
4150 		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4151 		if (!page) {
4152 			btrfs_err(fs_info, "find_or_create_page() failed");
4153 			ret = -ENOMEM;
4154 			goto out;
4155 		}
4156 
4157 		if (PageUptodate(page)) {
4158 			if (PageDirty(page))
4159 				goto next_page;
4160 		} else {
4161 			ClearPageError(page);
4162 			err = extent_read_full_page(io_tree, page,
4163 							   btrfs_get_extent,
4164 							   nocow_ctx->mirror_num);
4165 			if (err) {
4166 				ret = err;
4167 				goto next_page;
4168 			}
4169 
4170 			lock_page(page);
4171 			/*
4172 			 * If the page has been remove from the page cache,
4173 			 * the data on it is meaningless, because it may be
4174 			 * old one, the new data may be written into the new
4175 			 * page in the page cache.
4176 			 */
4177 			if (page->mapping != inode->i_mapping) {
4178 				unlock_page(page);
4179 				page_cache_release(page);
4180 				goto again;
4181 			}
4182 			if (!PageUptodate(page)) {
4183 				ret = -EIO;
4184 				goto next_page;
4185 			}
4186 		}
4187 
4188 		ret = check_extent_to_block(inode, offset, len,
4189 					    nocow_ctx_logical);
4190 		if (ret) {
4191 			ret = ret > 0 ? 0 : ret;
4192 			goto next_page;
4193 		}
4194 
4195 		err = write_page_nocow(nocow_ctx->sctx,
4196 				       physical_for_dev_replace, page);
4197 		if (err)
4198 			ret = err;
4199 next_page:
4200 		unlock_page(page);
4201 		page_cache_release(page);
4202 
4203 		if (ret)
4204 			break;
4205 
4206 		offset += PAGE_CACHE_SIZE;
4207 		physical_for_dev_replace += PAGE_CACHE_SIZE;
4208 		nocow_ctx_logical += PAGE_CACHE_SIZE;
4209 		len -= PAGE_CACHE_SIZE;
4210 	}
4211 	ret = COPY_COMPLETE;
4212 out:
4213 	mutex_unlock(&inode->i_mutex);
4214 	iput(inode);
4215 	return ret;
4216 }
4217 
4218 static int write_page_nocow(struct scrub_ctx *sctx,
4219 			    u64 physical_for_dev_replace, struct page *page)
4220 {
4221 	struct bio *bio;
4222 	struct btrfs_device *dev;
4223 	int ret;
4224 
4225 	dev = sctx->wr_ctx.tgtdev;
4226 	if (!dev)
4227 		return -EIO;
4228 	if (!dev->bdev) {
4229 		printk_ratelimited(KERN_WARNING
4230 			"BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
4231 		return -EIO;
4232 	}
4233 	bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4234 	if (!bio) {
4235 		spin_lock(&sctx->stat_lock);
4236 		sctx->stat.malloc_errors++;
4237 		spin_unlock(&sctx->stat_lock);
4238 		return -ENOMEM;
4239 	}
4240 	bio->bi_iter.bi_size = 0;
4241 	bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4242 	bio->bi_bdev = dev->bdev;
4243 	ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
4244 	if (ret != PAGE_CACHE_SIZE) {
4245 leave_with_eio:
4246 		bio_put(bio);
4247 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4248 		return -EIO;
4249 	}
4250 
4251 	if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
4252 		goto leave_with_eio;
4253 
4254 	bio_put(bio);
4255 	return 0;
4256 }
4257