xref: /openbmc/linux/fs/btrfs/scrub.c (revision f6723b56)
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_page {
67 	struct scrub_block	*sblock;
68 	struct page		*page;
69 	struct btrfs_device	*dev;
70 	u64			flags;  /* extent flags */
71 	u64			generation;
72 	u64			logical;
73 	u64			physical;
74 	u64			physical_for_dev_replace;
75 	atomic_t		ref_count;
76 	struct {
77 		unsigned int	mirror_num:8;
78 		unsigned int	have_csum:1;
79 		unsigned int	io_error:1;
80 	};
81 	u8			csum[BTRFS_CSUM_SIZE];
82 };
83 
84 struct scrub_bio {
85 	int			index;
86 	struct scrub_ctx	*sctx;
87 	struct btrfs_device	*dev;
88 	struct bio		*bio;
89 	int			err;
90 	u64			logical;
91 	u64			physical;
92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
94 #else
95 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
96 #endif
97 	int			page_count;
98 	int			next_free;
99 	struct btrfs_work	work;
100 };
101 
102 struct scrub_block {
103 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 	int			page_count;
105 	atomic_t		outstanding_pages;
106 	atomic_t		ref_count; /* free mem on transition to zero */
107 	struct scrub_ctx	*sctx;
108 	struct {
109 		unsigned int	header_error:1;
110 		unsigned int	checksum_error:1;
111 		unsigned int	no_io_error_seen:1;
112 		unsigned int	generation_error:1; /* also sets header_error */
113 	};
114 };
115 
116 struct scrub_wr_ctx {
117 	struct scrub_bio *wr_curr_bio;
118 	struct btrfs_device *tgtdev;
119 	int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
120 	atomic_t flush_all_writes;
121 	struct mutex wr_lock;
122 };
123 
124 struct scrub_ctx {
125 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
126 	struct btrfs_root	*dev_root;
127 	int			first_free;
128 	int			curr;
129 	atomic_t		bios_in_flight;
130 	atomic_t		workers_pending;
131 	spinlock_t		list_lock;
132 	wait_queue_head_t	list_wait;
133 	u16			csum_size;
134 	struct list_head	csum_list;
135 	atomic_t		cancel_req;
136 	int			readonly;
137 	int			pages_per_rd_bio;
138 	u32			sectorsize;
139 	u32			nodesize;
140 	u32			leafsize;
141 
142 	int			is_dev_replace;
143 	struct scrub_wr_ctx	wr_ctx;
144 
145 	/*
146 	 * statistics
147 	 */
148 	struct btrfs_scrub_progress stat;
149 	spinlock_t		stat_lock;
150 };
151 
152 struct scrub_fixup_nodatasum {
153 	struct scrub_ctx	*sctx;
154 	struct btrfs_device	*dev;
155 	u64			logical;
156 	struct btrfs_root	*root;
157 	struct btrfs_work	work;
158 	int			mirror_num;
159 };
160 
161 struct scrub_nocow_inode {
162 	u64			inum;
163 	u64			offset;
164 	u64			root;
165 	struct list_head	list;
166 };
167 
168 struct scrub_copy_nocow_ctx {
169 	struct scrub_ctx	*sctx;
170 	u64			logical;
171 	u64			len;
172 	int			mirror_num;
173 	u64			physical_for_dev_replace;
174 	struct list_head	inodes;
175 	struct btrfs_work	work;
176 };
177 
178 struct scrub_warning {
179 	struct btrfs_path	*path;
180 	u64			extent_item_size;
181 	char			*scratch_buf;
182 	char			*msg_buf;
183 	const char		*errstr;
184 	sector_t		sector;
185 	u64			logical;
186 	struct btrfs_device	*dev;
187 	int			msg_bufsize;
188 	int			scratch_bufsize;
189 };
190 
191 
192 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
193 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
194 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
195 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
196 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
197 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
198 				     struct btrfs_fs_info *fs_info,
199 				     struct scrub_block *original_sblock,
200 				     u64 length, u64 logical,
201 				     struct scrub_block *sblocks_for_recheck);
202 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
203 				struct scrub_block *sblock, int is_metadata,
204 				int have_csum, u8 *csum, u64 generation,
205 				u16 csum_size);
206 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
207 					 struct scrub_block *sblock,
208 					 int is_metadata, int have_csum,
209 					 const u8 *csum, u64 generation,
210 					 u16 csum_size);
211 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
212 					     struct scrub_block *sblock_good,
213 					     int force_write);
214 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
215 					    struct scrub_block *sblock_good,
216 					    int page_num, int force_write);
217 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
218 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
219 					   int page_num);
220 static int scrub_checksum_data(struct scrub_block *sblock);
221 static int scrub_checksum_tree_block(struct scrub_block *sblock);
222 static int scrub_checksum_super(struct scrub_block *sblock);
223 static void scrub_block_get(struct scrub_block *sblock);
224 static void scrub_block_put(struct scrub_block *sblock);
225 static void scrub_page_get(struct scrub_page *spage);
226 static void scrub_page_put(struct scrub_page *spage);
227 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
228 				    struct scrub_page *spage);
229 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
230 		       u64 physical, struct btrfs_device *dev, u64 flags,
231 		       u64 gen, int mirror_num, u8 *csum, int force,
232 		       u64 physical_for_dev_replace);
233 static void scrub_bio_end_io(struct bio *bio, int err);
234 static void scrub_bio_end_io_worker(struct btrfs_work *work);
235 static void scrub_block_complete(struct scrub_block *sblock);
236 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
237 			       u64 extent_logical, u64 extent_len,
238 			       u64 *extent_physical,
239 			       struct btrfs_device **extent_dev,
240 			       int *extent_mirror_num);
241 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
242 			      struct scrub_wr_ctx *wr_ctx,
243 			      struct btrfs_fs_info *fs_info,
244 			      struct btrfs_device *dev,
245 			      int is_dev_replace);
246 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
248 				    struct scrub_page *spage);
249 static void scrub_wr_submit(struct scrub_ctx *sctx);
250 static void scrub_wr_bio_end_io(struct bio *bio, int err);
251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
252 static int write_page_nocow(struct scrub_ctx *sctx,
253 			    u64 physical_for_dev_replace, struct page *page);
254 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
255 				      struct scrub_copy_nocow_ctx *ctx);
256 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
257 			    int mirror_num, u64 physical_for_dev_replace);
258 static void copy_nocow_pages_worker(struct btrfs_work *work);
259 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
260 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
261 
262 
263 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
264 {
265 	atomic_inc(&sctx->bios_in_flight);
266 }
267 
268 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
269 {
270 	atomic_dec(&sctx->bios_in_flight);
271 	wake_up(&sctx->list_wait);
272 }
273 
274 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
275 {
276 	while (atomic_read(&fs_info->scrub_pause_req)) {
277 		mutex_unlock(&fs_info->scrub_lock);
278 		wait_event(fs_info->scrub_pause_wait,
279 		   atomic_read(&fs_info->scrub_pause_req) == 0);
280 		mutex_lock(&fs_info->scrub_lock);
281 	}
282 }
283 
284 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
285 {
286 	atomic_inc(&fs_info->scrubs_paused);
287 	wake_up(&fs_info->scrub_pause_wait);
288 
289 	mutex_lock(&fs_info->scrub_lock);
290 	__scrub_blocked_if_needed(fs_info);
291 	atomic_dec(&fs_info->scrubs_paused);
292 	mutex_unlock(&fs_info->scrub_lock);
293 
294 	wake_up(&fs_info->scrub_pause_wait);
295 }
296 
297 /*
298  * used for workers that require transaction commits (i.e., for the
299  * NOCOW case)
300  */
301 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
302 {
303 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
304 
305 	/*
306 	 * increment scrubs_running to prevent cancel requests from
307 	 * completing as long as a worker is running. we must also
308 	 * increment scrubs_paused to prevent deadlocking on pause
309 	 * requests used for transactions commits (as the worker uses a
310 	 * transaction context). it is safe to regard the worker
311 	 * as paused for all matters practical. effectively, we only
312 	 * avoid cancellation requests from completing.
313 	 */
314 	mutex_lock(&fs_info->scrub_lock);
315 	atomic_inc(&fs_info->scrubs_running);
316 	atomic_inc(&fs_info->scrubs_paused);
317 	mutex_unlock(&fs_info->scrub_lock);
318 	atomic_inc(&sctx->workers_pending);
319 }
320 
321 /* used for workers that require transaction commits */
322 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
323 {
324 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
325 
326 	/*
327 	 * see scrub_pending_trans_workers_inc() why we're pretending
328 	 * to be paused in the scrub counters
329 	 */
330 	mutex_lock(&fs_info->scrub_lock);
331 	atomic_dec(&fs_info->scrubs_running);
332 	atomic_dec(&fs_info->scrubs_paused);
333 	mutex_unlock(&fs_info->scrub_lock);
334 	atomic_dec(&sctx->workers_pending);
335 	wake_up(&fs_info->scrub_pause_wait);
336 	wake_up(&sctx->list_wait);
337 }
338 
339 static void scrub_free_csums(struct scrub_ctx *sctx)
340 {
341 	while (!list_empty(&sctx->csum_list)) {
342 		struct btrfs_ordered_sum *sum;
343 		sum = list_first_entry(&sctx->csum_list,
344 				       struct btrfs_ordered_sum, list);
345 		list_del(&sum->list);
346 		kfree(sum);
347 	}
348 }
349 
350 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
351 {
352 	int i;
353 
354 	if (!sctx)
355 		return;
356 
357 	scrub_free_wr_ctx(&sctx->wr_ctx);
358 
359 	/* this can happen when scrub is cancelled */
360 	if (sctx->curr != -1) {
361 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
362 
363 		for (i = 0; i < sbio->page_count; i++) {
364 			WARN_ON(!sbio->pagev[i]->page);
365 			scrub_block_put(sbio->pagev[i]->sblock);
366 		}
367 		bio_put(sbio->bio);
368 	}
369 
370 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
371 		struct scrub_bio *sbio = sctx->bios[i];
372 
373 		if (!sbio)
374 			break;
375 		kfree(sbio);
376 	}
377 
378 	scrub_free_csums(sctx);
379 	kfree(sctx);
380 }
381 
382 static noinline_for_stack
383 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
384 {
385 	struct scrub_ctx *sctx;
386 	int		i;
387 	struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
388 	int pages_per_rd_bio;
389 	int ret;
390 
391 	/*
392 	 * the setting of pages_per_rd_bio is correct for scrub but might
393 	 * be wrong for the dev_replace code where we might read from
394 	 * different devices in the initial huge bios. However, that
395 	 * code is able to correctly handle the case when adding a page
396 	 * to a bio fails.
397 	 */
398 	if (dev->bdev)
399 		pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
400 					 bio_get_nr_vecs(dev->bdev));
401 	else
402 		pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
403 	sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
404 	if (!sctx)
405 		goto nomem;
406 	sctx->is_dev_replace = is_dev_replace;
407 	sctx->pages_per_rd_bio = pages_per_rd_bio;
408 	sctx->curr = -1;
409 	sctx->dev_root = dev->dev_root;
410 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
411 		struct scrub_bio *sbio;
412 
413 		sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
414 		if (!sbio)
415 			goto nomem;
416 		sctx->bios[i] = sbio;
417 
418 		sbio->index = i;
419 		sbio->sctx = sctx;
420 		sbio->page_count = 0;
421 		sbio->work.func = scrub_bio_end_io_worker;
422 
423 		if (i != SCRUB_BIOS_PER_SCTX - 1)
424 			sctx->bios[i]->next_free = i + 1;
425 		else
426 			sctx->bios[i]->next_free = -1;
427 	}
428 	sctx->first_free = 0;
429 	sctx->nodesize = dev->dev_root->nodesize;
430 	sctx->leafsize = dev->dev_root->leafsize;
431 	sctx->sectorsize = dev->dev_root->sectorsize;
432 	atomic_set(&sctx->bios_in_flight, 0);
433 	atomic_set(&sctx->workers_pending, 0);
434 	atomic_set(&sctx->cancel_req, 0);
435 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
436 	INIT_LIST_HEAD(&sctx->csum_list);
437 
438 	spin_lock_init(&sctx->list_lock);
439 	spin_lock_init(&sctx->stat_lock);
440 	init_waitqueue_head(&sctx->list_wait);
441 
442 	ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
443 				 fs_info->dev_replace.tgtdev, is_dev_replace);
444 	if (ret) {
445 		scrub_free_ctx(sctx);
446 		return ERR_PTR(ret);
447 	}
448 	return sctx;
449 
450 nomem:
451 	scrub_free_ctx(sctx);
452 	return ERR_PTR(-ENOMEM);
453 }
454 
455 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
456 				     void *warn_ctx)
457 {
458 	u64 isize;
459 	u32 nlink;
460 	int ret;
461 	int i;
462 	struct extent_buffer *eb;
463 	struct btrfs_inode_item *inode_item;
464 	struct scrub_warning *swarn = warn_ctx;
465 	struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
466 	struct inode_fs_paths *ipath = NULL;
467 	struct btrfs_root *local_root;
468 	struct btrfs_key root_key;
469 
470 	root_key.objectid = root;
471 	root_key.type = BTRFS_ROOT_ITEM_KEY;
472 	root_key.offset = (u64)-1;
473 	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
474 	if (IS_ERR(local_root)) {
475 		ret = PTR_ERR(local_root);
476 		goto err;
477 	}
478 
479 	ret = inode_item_info(inum, 0, local_root, swarn->path);
480 	if (ret) {
481 		btrfs_release_path(swarn->path);
482 		goto err;
483 	}
484 
485 	eb = swarn->path->nodes[0];
486 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
487 					struct btrfs_inode_item);
488 	isize = btrfs_inode_size(eb, inode_item);
489 	nlink = btrfs_inode_nlink(eb, inode_item);
490 	btrfs_release_path(swarn->path);
491 
492 	ipath = init_ipath(4096, local_root, swarn->path);
493 	if (IS_ERR(ipath)) {
494 		ret = PTR_ERR(ipath);
495 		ipath = NULL;
496 		goto err;
497 	}
498 	ret = paths_from_inode(inum, ipath);
499 
500 	if (ret < 0)
501 		goto err;
502 
503 	/*
504 	 * we deliberately ignore the bit ipath might have been too small to
505 	 * hold all of the paths here
506 	 */
507 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
508 		printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
509 			"%s, sector %llu, root %llu, inode %llu, offset %llu, "
510 			"length %llu, links %u (path: %s)\n", swarn->errstr,
511 			swarn->logical, rcu_str_deref(swarn->dev->name),
512 			(unsigned long long)swarn->sector, root, inum, offset,
513 			min(isize - offset, (u64)PAGE_SIZE), nlink,
514 			(char *)(unsigned long)ipath->fspath->val[i]);
515 
516 	free_ipath(ipath);
517 	return 0;
518 
519 err:
520 	printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
521 		"%s, sector %llu, root %llu, inode %llu, offset %llu: path "
522 		"resolving failed with ret=%d\n", swarn->errstr,
523 		swarn->logical, rcu_str_deref(swarn->dev->name),
524 		(unsigned long long)swarn->sector, root, inum, offset, ret);
525 
526 	free_ipath(ipath);
527 	return 0;
528 }
529 
530 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
531 {
532 	struct btrfs_device *dev;
533 	struct btrfs_fs_info *fs_info;
534 	struct btrfs_path *path;
535 	struct btrfs_key found_key;
536 	struct extent_buffer *eb;
537 	struct btrfs_extent_item *ei;
538 	struct scrub_warning swarn;
539 	unsigned long ptr = 0;
540 	u64 extent_item_pos;
541 	u64 flags = 0;
542 	u64 ref_root;
543 	u32 item_size;
544 	u8 ref_level;
545 	const int bufsize = 4096;
546 	int ret;
547 
548 	WARN_ON(sblock->page_count < 1);
549 	dev = sblock->pagev[0]->dev;
550 	fs_info = sblock->sctx->dev_root->fs_info;
551 
552 	path = btrfs_alloc_path();
553 
554 	swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
555 	swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
556 	swarn.sector = (sblock->pagev[0]->physical) >> 9;
557 	swarn.logical = sblock->pagev[0]->logical;
558 	swarn.errstr = errstr;
559 	swarn.dev = NULL;
560 	swarn.msg_bufsize = bufsize;
561 	swarn.scratch_bufsize = bufsize;
562 
563 	if (!path || !swarn.scratch_buf || !swarn.msg_buf)
564 		goto out;
565 
566 	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
567 				  &flags);
568 	if (ret < 0)
569 		goto out;
570 
571 	extent_item_pos = swarn.logical - found_key.objectid;
572 	swarn.extent_item_size = found_key.offset;
573 
574 	eb = path->nodes[0];
575 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
576 	item_size = btrfs_item_size_nr(eb, path->slots[0]);
577 
578 	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
579 		do {
580 			ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
581 							&ref_root, &ref_level);
582 			printk_in_rcu(KERN_WARNING
583 				"BTRFS: %s at logical %llu on dev %s, "
584 				"sector %llu: metadata %s (level %d) in tree "
585 				"%llu\n", errstr, swarn.logical,
586 				rcu_str_deref(dev->name),
587 				(unsigned long long)swarn.sector,
588 				ref_level ? "node" : "leaf",
589 				ret < 0 ? -1 : ref_level,
590 				ret < 0 ? -1 : ref_root);
591 		} while (ret != 1);
592 		btrfs_release_path(path);
593 	} else {
594 		btrfs_release_path(path);
595 		swarn.path = path;
596 		swarn.dev = dev;
597 		iterate_extent_inodes(fs_info, found_key.objectid,
598 					extent_item_pos, 1,
599 					scrub_print_warning_inode, &swarn);
600 	}
601 
602 out:
603 	btrfs_free_path(path);
604 	kfree(swarn.scratch_buf);
605 	kfree(swarn.msg_buf);
606 }
607 
608 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
609 {
610 	struct page *page = NULL;
611 	unsigned long index;
612 	struct scrub_fixup_nodatasum *fixup = fixup_ctx;
613 	int ret;
614 	int corrected = 0;
615 	struct btrfs_key key;
616 	struct inode *inode = NULL;
617 	struct btrfs_fs_info *fs_info;
618 	u64 end = offset + PAGE_SIZE - 1;
619 	struct btrfs_root *local_root;
620 	int srcu_index;
621 
622 	key.objectid = root;
623 	key.type = BTRFS_ROOT_ITEM_KEY;
624 	key.offset = (u64)-1;
625 
626 	fs_info = fixup->root->fs_info;
627 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
628 
629 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
630 	if (IS_ERR(local_root)) {
631 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
632 		return PTR_ERR(local_root);
633 	}
634 
635 	key.type = BTRFS_INODE_ITEM_KEY;
636 	key.objectid = inum;
637 	key.offset = 0;
638 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
639 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
640 	if (IS_ERR(inode))
641 		return PTR_ERR(inode);
642 
643 	index = offset >> PAGE_CACHE_SHIFT;
644 
645 	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
646 	if (!page) {
647 		ret = -ENOMEM;
648 		goto out;
649 	}
650 
651 	if (PageUptodate(page)) {
652 		if (PageDirty(page)) {
653 			/*
654 			 * we need to write the data to the defect sector. the
655 			 * data that was in that sector is not in memory,
656 			 * because the page was modified. we must not write the
657 			 * modified page to that sector.
658 			 *
659 			 * TODO: what could be done here: wait for the delalloc
660 			 *       runner to write out that page (might involve
661 			 *       COW) and see whether the sector is still
662 			 *       referenced afterwards.
663 			 *
664 			 * For the meantime, we'll treat this error
665 			 * incorrectable, although there is a chance that a
666 			 * later scrub will find the bad sector again and that
667 			 * there's no dirty page in memory, then.
668 			 */
669 			ret = -EIO;
670 			goto out;
671 		}
672 		fs_info = BTRFS_I(inode)->root->fs_info;
673 		ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
674 					fixup->logical, page,
675 					fixup->mirror_num);
676 		unlock_page(page);
677 		corrected = !ret;
678 	} else {
679 		/*
680 		 * we need to get good data first. the general readpage path
681 		 * will call repair_io_failure for us, we just have to make
682 		 * sure we read the bad mirror.
683 		 */
684 		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
685 					EXTENT_DAMAGED, GFP_NOFS);
686 		if (ret) {
687 			/* set_extent_bits should give proper error */
688 			WARN_ON(ret > 0);
689 			if (ret > 0)
690 				ret = -EFAULT;
691 			goto out;
692 		}
693 
694 		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
695 						btrfs_get_extent,
696 						fixup->mirror_num);
697 		wait_on_page_locked(page);
698 
699 		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
700 						end, EXTENT_DAMAGED, 0, NULL);
701 		if (!corrected)
702 			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
703 						EXTENT_DAMAGED, GFP_NOFS);
704 	}
705 
706 out:
707 	if (page)
708 		put_page(page);
709 	if (inode)
710 		iput(inode);
711 
712 	if (ret < 0)
713 		return ret;
714 
715 	if (ret == 0 && corrected) {
716 		/*
717 		 * we only need to call readpage for one of the inodes belonging
718 		 * to this extent. so make iterate_extent_inodes stop
719 		 */
720 		return 1;
721 	}
722 
723 	return -EIO;
724 }
725 
726 static void scrub_fixup_nodatasum(struct btrfs_work *work)
727 {
728 	int ret;
729 	struct scrub_fixup_nodatasum *fixup;
730 	struct scrub_ctx *sctx;
731 	struct btrfs_trans_handle *trans = NULL;
732 	struct btrfs_path *path;
733 	int uncorrectable = 0;
734 
735 	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
736 	sctx = fixup->sctx;
737 
738 	path = btrfs_alloc_path();
739 	if (!path) {
740 		spin_lock(&sctx->stat_lock);
741 		++sctx->stat.malloc_errors;
742 		spin_unlock(&sctx->stat_lock);
743 		uncorrectable = 1;
744 		goto out;
745 	}
746 
747 	trans = btrfs_join_transaction(fixup->root);
748 	if (IS_ERR(trans)) {
749 		uncorrectable = 1;
750 		goto out;
751 	}
752 
753 	/*
754 	 * the idea is to trigger a regular read through the standard path. we
755 	 * read a page from the (failed) logical address by specifying the
756 	 * corresponding copynum of the failed sector. thus, that readpage is
757 	 * expected to fail.
758 	 * that is the point where on-the-fly error correction will kick in
759 	 * (once it's finished) and rewrite the failed sector if a good copy
760 	 * can be found.
761 	 */
762 	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
763 						path, scrub_fixup_readpage,
764 						fixup);
765 	if (ret < 0) {
766 		uncorrectable = 1;
767 		goto out;
768 	}
769 	WARN_ON(ret != 1);
770 
771 	spin_lock(&sctx->stat_lock);
772 	++sctx->stat.corrected_errors;
773 	spin_unlock(&sctx->stat_lock);
774 
775 out:
776 	if (trans && !IS_ERR(trans))
777 		btrfs_end_transaction(trans, fixup->root);
778 	if (uncorrectable) {
779 		spin_lock(&sctx->stat_lock);
780 		++sctx->stat.uncorrectable_errors;
781 		spin_unlock(&sctx->stat_lock);
782 		btrfs_dev_replace_stats_inc(
783 			&sctx->dev_root->fs_info->dev_replace.
784 			num_uncorrectable_read_errors);
785 		printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
786 		    "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
787 			fixup->logical, rcu_str_deref(fixup->dev->name));
788 	}
789 
790 	btrfs_free_path(path);
791 	kfree(fixup);
792 
793 	scrub_pending_trans_workers_dec(sctx);
794 }
795 
796 /*
797  * scrub_handle_errored_block gets called when either verification of the
798  * pages failed or the bio failed to read, e.g. with EIO. In the latter
799  * case, this function handles all pages in the bio, even though only one
800  * may be bad.
801  * The goal of this function is to repair the errored block by using the
802  * contents of one of the mirrors.
803  */
804 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
805 {
806 	struct scrub_ctx *sctx = sblock_to_check->sctx;
807 	struct btrfs_device *dev;
808 	struct btrfs_fs_info *fs_info;
809 	u64 length;
810 	u64 logical;
811 	u64 generation;
812 	unsigned int failed_mirror_index;
813 	unsigned int is_metadata;
814 	unsigned int have_csum;
815 	u8 *csum;
816 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
817 	struct scrub_block *sblock_bad;
818 	int ret;
819 	int mirror_index;
820 	int page_num;
821 	int success;
822 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
823 				      DEFAULT_RATELIMIT_BURST);
824 
825 	BUG_ON(sblock_to_check->page_count < 1);
826 	fs_info = sctx->dev_root->fs_info;
827 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
828 		/*
829 		 * if we find an error in a super block, we just report it.
830 		 * They will get written with the next transaction commit
831 		 * anyway
832 		 */
833 		spin_lock(&sctx->stat_lock);
834 		++sctx->stat.super_errors;
835 		spin_unlock(&sctx->stat_lock);
836 		return 0;
837 	}
838 	length = sblock_to_check->page_count * PAGE_SIZE;
839 	logical = sblock_to_check->pagev[0]->logical;
840 	generation = sblock_to_check->pagev[0]->generation;
841 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
842 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
843 	is_metadata = !(sblock_to_check->pagev[0]->flags &
844 			BTRFS_EXTENT_FLAG_DATA);
845 	have_csum = sblock_to_check->pagev[0]->have_csum;
846 	csum = sblock_to_check->pagev[0]->csum;
847 	dev = sblock_to_check->pagev[0]->dev;
848 
849 	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
850 		sblocks_for_recheck = NULL;
851 		goto nodatasum_case;
852 	}
853 
854 	/*
855 	 * read all mirrors one after the other. This includes to
856 	 * re-read the extent or metadata block that failed (that was
857 	 * the cause that this fixup code is called) another time,
858 	 * page by page this time in order to know which pages
859 	 * caused I/O errors and which ones are good (for all mirrors).
860 	 * It is the goal to handle the situation when more than one
861 	 * mirror contains I/O errors, but the errors do not
862 	 * overlap, i.e. the data can be repaired by selecting the
863 	 * pages from those mirrors without I/O error on the
864 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
865 	 * would be that mirror #1 has an I/O error on the first page,
866 	 * the second page is good, and mirror #2 has an I/O error on
867 	 * the second page, but the first page is good.
868 	 * Then the first page of the first mirror can be repaired by
869 	 * taking the first page of the second mirror, and the
870 	 * second page of the second mirror can be repaired by
871 	 * copying the contents of the 2nd page of the 1st mirror.
872 	 * One more note: if the pages of one mirror contain I/O
873 	 * errors, the checksum cannot be verified. In order to get
874 	 * the best data for repairing, the first attempt is to find
875 	 * a mirror without I/O errors and with a validated checksum.
876 	 * Only if this is not possible, the pages are picked from
877 	 * mirrors with I/O errors without considering the checksum.
878 	 * If the latter is the case, at the end, the checksum of the
879 	 * repaired area is verified in order to correctly maintain
880 	 * the statistics.
881 	 */
882 
883 	sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
884 				     sizeof(*sblocks_for_recheck),
885 				     GFP_NOFS);
886 	if (!sblocks_for_recheck) {
887 		spin_lock(&sctx->stat_lock);
888 		sctx->stat.malloc_errors++;
889 		sctx->stat.read_errors++;
890 		sctx->stat.uncorrectable_errors++;
891 		spin_unlock(&sctx->stat_lock);
892 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
893 		goto out;
894 	}
895 
896 	/* setup the context, map the logical blocks and alloc the pages */
897 	ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
898 					logical, sblocks_for_recheck);
899 	if (ret) {
900 		spin_lock(&sctx->stat_lock);
901 		sctx->stat.read_errors++;
902 		sctx->stat.uncorrectable_errors++;
903 		spin_unlock(&sctx->stat_lock);
904 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
905 		goto out;
906 	}
907 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
908 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
909 
910 	/* build and submit the bios for the failed mirror, check checksums */
911 	scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
912 			    csum, generation, sctx->csum_size);
913 
914 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
915 	    sblock_bad->no_io_error_seen) {
916 		/*
917 		 * the error disappeared after reading page by page, or
918 		 * the area was part of a huge bio and other parts of the
919 		 * bio caused I/O errors, or the block layer merged several
920 		 * read requests into one and the error is caused by a
921 		 * different bio (usually one of the two latter cases is
922 		 * the cause)
923 		 */
924 		spin_lock(&sctx->stat_lock);
925 		sctx->stat.unverified_errors++;
926 		spin_unlock(&sctx->stat_lock);
927 
928 		if (sctx->is_dev_replace)
929 			scrub_write_block_to_dev_replace(sblock_bad);
930 		goto out;
931 	}
932 
933 	if (!sblock_bad->no_io_error_seen) {
934 		spin_lock(&sctx->stat_lock);
935 		sctx->stat.read_errors++;
936 		spin_unlock(&sctx->stat_lock);
937 		if (__ratelimit(&_rs))
938 			scrub_print_warning("i/o error", sblock_to_check);
939 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
940 	} else if (sblock_bad->checksum_error) {
941 		spin_lock(&sctx->stat_lock);
942 		sctx->stat.csum_errors++;
943 		spin_unlock(&sctx->stat_lock);
944 		if (__ratelimit(&_rs))
945 			scrub_print_warning("checksum error", sblock_to_check);
946 		btrfs_dev_stat_inc_and_print(dev,
947 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
948 	} else if (sblock_bad->header_error) {
949 		spin_lock(&sctx->stat_lock);
950 		sctx->stat.verify_errors++;
951 		spin_unlock(&sctx->stat_lock);
952 		if (__ratelimit(&_rs))
953 			scrub_print_warning("checksum/header error",
954 					    sblock_to_check);
955 		if (sblock_bad->generation_error)
956 			btrfs_dev_stat_inc_and_print(dev,
957 				BTRFS_DEV_STAT_GENERATION_ERRS);
958 		else
959 			btrfs_dev_stat_inc_and_print(dev,
960 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
961 	}
962 
963 	if (sctx->readonly) {
964 		ASSERT(!sctx->is_dev_replace);
965 		goto out;
966 	}
967 
968 	if (!is_metadata && !have_csum) {
969 		struct scrub_fixup_nodatasum *fixup_nodatasum;
970 
971 nodatasum_case:
972 		WARN_ON(sctx->is_dev_replace);
973 
974 		/*
975 		 * !is_metadata and !have_csum, this means that the data
976 		 * might not be COW'ed, that it might be modified
977 		 * concurrently. The general strategy to work on the
978 		 * commit root does not help in the case when COW is not
979 		 * used.
980 		 */
981 		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
982 		if (!fixup_nodatasum)
983 			goto did_not_correct_error;
984 		fixup_nodatasum->sctx = sctx;
985 		fixup_nodatasum->dev = dev;
986 		fixup_nodatasum->logical = logical;
987 		fixup_nodatasum->root = fs_info->extent_root;
988 		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
989 		scrub_pending_trans_workers_inc(sctx);
990 		fixup_nodatasum->work.func = scrub_fixup_nodatasum;
991 		btrfs_queue_worker(&fs_info->scrub_workers,
992 				   &fixup_nodatasum->work);
993 		goto out;
994 	}
995 
996 	/*
997 	 * now build and submit the bios for the other mirrors, check
998 	 * checksums.
999 	 * First try to pick the mirror which is completely without I/O
1000 	 * errors and also does not have a checksum error.
1001 	 * If one is found, and if a checksum is present, the full block
1002 	 * that is known to contain an error is rewritten. Afterwards
1003 	 * the block is known to be corrected.
1004 	 * If a mirror is found which is completely correct, and no
1005 	 * checksum is present, only those pages are rewritten that had
1006 	 * an I/O error in the block to be repaired, since it cannot be
1007 	 * determined, which copy of the other pages is better (and it
1008 	 * could happen otherwise that a correct page would be
1009 	 * overwritten by a bad one).
1010 	 */
1011 	for (mirror_index = 0;
1012 	     mirror_index < BTRFS_MAX_MIRRORS &&
1013 	     sblocks_for_recheck[mirror_index].page_count > 0;
1014 	     mirror_index++) {
1015 		struct scrub_block *sblock_other;
1016 
1017 		if (mirror_index == failed_mirror_index)
1018 			continue;
1019 		sblock_other = sblocks_for_recheck + mirror_index;
1020 
1021 		/* build and submit the bios, check checksums */
1022 		scrub_recheck_block(fs_info, sblock_other, is_metadata,
1023 				    have_csum, csum, generation,
1024 				    sctx->csum_size);
1025 
1026 		if (!sblock_other->header_error &&
1027 		    !sblock_other->checksum_error &&
1028 		    sblock_other->no_io_error_seen) {
1029 			if (sctx->is_dev_replace) {
1030 				scrub_write_block_to_dev_replace(sblock_other);
1031 			} else {
1032 				int force_write = is_metadata || have_csum;
1033 
1034 				ret = scrub_repair_block_from_good_copy(
1035 						sblock_bad, sblock_other,
1036 						force_write);
1037 			}
1038 			if (0 == ret)
1039 				goto corrected_error;
1040 		}
1041 	}
1042 
1043 	/*
1044 	 * for dev_replace, pick good pages and write to the target device.
1045 	 */
1046 	if (sctx->is_dev_replace) {
1047 		success = 1;
1048 		for (page_num = 0; page_num < sblock_bad->page_count;
1049 		     page_num++) {
1050 			int sub_success;
1051 
1052 			sub_success = 0;
1053 			for (mirror_index = 0;
1054 			     mirror_index < BTRFS_MAX_MIRRORS &&
1055 			     sblocks_for_recheck[mirror_index].page_count > 0;
1056 			     mirror_index++) {
1057 				struct scrub_block *sblock_other =
1058 					sblocks_for_recheck + mirror_index;
1059 				struct scrub_page *page_other =
1060 					sblock_other->pagev[page_num];
1061 
1062 				if (!page_other->io_error) {
1063 					ret = scrub_write_page_to_dev_replace(
1064 							sblock_other, page_num);
1065 					if (ret == 0) {
1066 						/* succeeded for this page */
1067 						sub_success = 1;
1068 						break;
1069 					} else {
1070 						btrfs_dev_replace_stats_inc(
1071 							&sctx->dev_root->
1072 							fs_info->dev_replace.
1073 							num_write_errors);
1074 					}
1075 				}
1076 			}
1077 
1078 			if (!sub_success) {
1079 				/*
1080 				 * did not find a mirror to fetch the page
1081 				 * from. scrub_write_page_to_dev_replace()
1082 				 * handles this case (page->io_error), by
1083 				 * filling the block with zeros before
1084 				 * submitting the write request
1085 				 */
1086 				success = 0;
1087 				ret = scrub_write_page_to_dev_replace(
1088 						sblock_bad, page_num);
1089 				if (ret)
1090 					btrfs_dev_replace_stats_inc(
1091 						&sctx->dev_root->fs_info->
1092 						dev_replace.num_write_errors);
1093 			}
1094 		}
1095 
1096 		goto out;
1097 	}
1098 
1099 	/*
1100 	 * for regular scrub, repair those pages that are errored.
1101 	 * In case of I/O errors in the area that is supposed to be
1102 	 * repaired, continue by picking good copies of those pages.
1103 	 * Select the good pages from mirrors to rewrite bad pages from
1104 	 * the area to fix. Afterwards verify the checksum of the block
1105 	 * that is supposed to be repaired. This verification step is
1106 	 * only done for the purpose of statistic counting and for the
1107 	 * final scrub report, whether errors remain.
1108 	 * A perfect algorithm could make use of the checksum and try
1109 	 * all possible combinations of pages from the different mirrors
1110 	 * until the checksum verification succeeds. For example, when
1111 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1112 	 * of mirror #2 is readable but the final checksum test fails,
1113 	 * then the 2nd page of mirror #3 could be tried, whether now
1114 	 * the final checksum succeedes. But this would be a rare
1115 	 * exception and is therefore not implemented. At least it is
1116 	 * avoided that the good copy is overwritten.
1117 	 * A more useful improvement would be to pick the sectors
1118 	 * without I/O error based on sector sizes (512 bytes on legacy
1119 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1120 	 * mirror could be repaired by taking 512 byte of a different
1121 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1122 	 * area are unreadable.
1123 	 */
1124 
1125 	/* can only fix I/O errors from here on */
1126 	if (sblock_bad->no_io_error_seen)
1127 		goto did_not_correct_error;
1128 
1129 	success = 1;
1130 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1131 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1132 
1133 		if (!page_bad->io_error)
1134 			continue;
1135 
1136 		for (mirror_index = 0;
1137 		     mirror_index < BTRFS_MAX_MIRRORS &&
1138 		     sblocks_for_recheck[mirror_index].page_count > 0;
1139 		     mirror_index++) {
1140 			struct scrub_block *sblock_other = sblocks_for_recheck +
1141 							   mirror_index;
1142 			struct scrub_page *page_other = sblock_other->pagev[
1143 							page_num];
1144 
1145 			if (!page_other->io_error) {
1146 				ret = scrub_repair_page_from_good_copy(
1147 					sblock_bad, sblock_other, page_num, 0);
1148 				if (0 == ret) {
1149 					page_bad->io_error = 0;
1150 					break; /* succeeded for this page */
1151 				}
1152 			}
1153 		}
1154 
1155 		if (page_bad->io_error) {
1156 			/* did not find a mirror to copy the page from */
1157 			success = 0;
1158 		}
1159 	}
1160 
1161 	if (success) {
1162 		if (is_metadata || have_csum) {
1163 			/*
1164 			 * need to verify the checksum now that all
1165 			 * sectors on disk are repaired (the write
1166 			 * request for data to be repaired is on its way).
1167 			 * Just be lazy and use scrub_recheck_block()
1168 			 * which re-reads the data before the checksum
1169 			 * is verified, but most likely the data comes out
1170 			 * of the page cache.
1171 			 */
1172 			scrub_recheck_block(fs_info, sblock_bad,
1173 					    is_metadata, have_csum, csum,
1174 					    generation, sctx->csum_size);
1175 			if (!sblock_bad->header_error &&
1176 			    !sblock_bad->checksum_error &&
1177 			    sblock_bad->no_io_error_seen)
1178 				goto corrected_error;
1179 			else
1180 				goto did_not_correct_error;
1181 		} else {
1182 corrected_error:
1183 			spin_lock(&sctx->stat_lock);
1184 			sctx->stat.corrected_errors++;
1185 			spin_unlock(&sctx->stat_lock);
1186 			printk_ratelimited_in_rcu(KERN_ERR
1187 				"BTRFS: fixed up error at logical %llu on dev %s\n",
1188 				logical, rcu_str_deref(dev->name));
1189 		}
1190 	} else {
1191 did_not_correct_error:
1192 		spin_lock(&sctx->stat_lock);
1193 		sctx->stat.uncorrectable_errors++;
1194 		spin_unlock(&sctx->stat_lock);
1195 		printk_ratelimited_in_rcu(KERN_ERR
1196 			"BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1197 			logical, rcu_str_deref(dev->name));
1198 	}
1199 
1200 out:
1201 	if (sblocks_for_recheck) {
1202 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1203 		     mirror_index++) {
1204 			struct scrub_block *sblock = sblocks_for_recheck +
1205 						     mirror_index;
1206 			int page_index;
1207 
1208 			for (page_index = 0; page_index < sblock->page_count;
1209 			     page_index++) {
1210 				sblock->pagev[page_index]->sblock = NULL;
1211 				scrub_page_put(sblock->pagev[page_index]);
1212 			}
1213 		}
1214 		kfree(sblocks_for_recheck);
1215 	}
1216 
1217 	return 0;
1218 }
1219 
1220 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1221 				     struct btrfs_fs_info *fs_info,
1222 				     struct scrub_block *original_sblock,
1223 				     u64 length, u64 logical,
1224 				     struct scrub_block *sblocks_for_recheck)
1225 {
1226 	int page_index;
1227 	int mirror_index;
1228 	int ret;
1229 
1230 	/*
1231 	 * note: the two members ref_count and outstanding_pages
1232 	 * are not used (and not set) in the blocks that are used for
1233 	 * the recheck procedure
1234 	 */
1235 
1236 	page_index = 0;
1237 	while (length > 0) {
1238 		u64 sublen = min_t(u64, length, PAGE_SIZE);
1239 		u64 mapped_length = sublen;
1240 		struct btrfs_bio *bbio = NULL;
1241 
1242 		/*
1243 		 * with a length of PAGE_SIZE, each returned stripe
1244 		 * represents one mirror
1245 		 */
1246 		ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1247 				      &mapped_length, &bbio, 0);
1248 		if (ret || !bbio || mapped_length < sublen) {
1249 			kfree(bbio);
1250 			return -EIO;
1251 		}
1252 
1253 		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1254 		for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1255 		     mirror_index++) {
1256 			struct scrub_block *sblock;
1257 			struct scrub_page *page;
1258 
1259 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1260 				continue;
1261 
1262 			sblock = sblocks_for_recheck + mirror_index;
1263 			sblock->sctx = sctx;
1264 			page = kzalloc(sizeof(*page), GFP_NOFS);
1265 			if (!page) {
1266 leave_nomem:
1267 				spin_lock(&sctx->stat_lock);
1268 				sctx->stat.malloc_errors++;
1269 				spin_unlock(&sctx->stat_lock);
1270 				kfree(bbio);
1271 				return -ENOMEM;
1272 			}
1273 			scrub_page_get(page);
1274 			sblock->pagev[page_index] = page;
1275 			page->logical = logical;
1276 			page->physical = bbio->stripes[mirror_index].physical;
1277 			BUG_ON(page_index >= original_sblock->page_count);
1278 			page->physical_for_dev_replace =
1279 				original_sblock->pagev[page_index]->
1280 				physical_for_dev_replace;
1281 			/* for missing devices, dev->bdev is NULL */
1282 			page->dev = bbio->stripes[mirror_index].dev;
1283 			page->mirror_num = mirror_index + 1;
1284 			sblock->page_count++;
1285 			page->page = alloc_page(GFP_NOFS);
1286 			if (!page->page)
1287 				goto leave_nomem;
1288 		}
1289 		kfree(bbio);
1290 		length -= sublen;
1291 		logical += sublen;
1292 		page_index++;
1293 	}
1294 
1295 	return 0;
1296 }
1297 
1298 /*
1299  * this function will check the on disk data for checksum errors, header
1300  * errors and read I/O errors. If any I/O errors happen, the exact pages
1301  * which are errored are marked as being bad. The goal is to enable scrub
1302  * to take those pages that are not errored from all the mirrors so that
1303  * the pages that are errored in the just handled mirror can be repaired.
1304  */
1305 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1306 				struct scrub_block *sblock, int is_metadata,
1307 				int have_csum, u8 *csum, u64 generation,
1308 				u16 csum_size)
1309 {
1310 	int page_num;
1311 
1312 	sblock->no_io_error_seen = 1;
1313 	sblock->header_error = 0;
1314 	sblock->checksum_error = 0;
1315 
1316 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1317 		struct bio *bio;
1318 		struct scrub_page *page = sblock->pagev[page_num];
1319 
1320 		if (page->dev->bdev == NULL) {
1321 			page->io_error = 1;
1322 			sblock->no_io_error_seen = 0;
1323 			continue;
1324 		}
1325 
1326 		WARN_ON(!page->page);
1327 		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1328 		if (!bio) {
1329 			page->io_error = 1;
1330 			sblock->no_io_error_seen = 0;
1331 			continue;
1332 		}
1333 		bio->bi_bdev = page->dev->bdev;
1334 		bio->bi_iter.bi_sector = page->physical >> 9;
1335 
1336 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1337 		if (btrfsic_submit_bio_wait(READ, bio))
1338 			sblock->no_io_error_seen = 0;
1339 
1340 		bio_put(bio);
1341 	}
1342 
1343 	if (sblock->no_io_error_seen)
1344 		scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1345 					     have_csum, csum, generation,
1346 					     csum_size);
1347 
1348 	return;
1349 }
1350 
1351 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1352 					 struct scrub_block *sblock,
1353 					 int is_metadata, int have_csum,
1354 					 const u8 *csum, u64 generation,
1355 					 u16 csum_size)
1356 {
1357 	int page_num;
1358 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1359 	u32 crc = ~(u32)0;
1360 	void *mapped_buffer;
1361 
1362 	WARN_ON(!sblock->pagev[0]->page);
1363 	if (is_metadata) {
1364 		struct btrfs_header *h;
1365 
1366 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1367 		h = (struct btrfs_header *)mapped_buffer;
1368 
1369 		if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1370 		    memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1371 		    memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1372 			   BTRFS_UUID_SIZE)) {
1373 			sblock->header_error = 1;
1374 		} else if (generation != btrfs_stack_header_generation(h)) {
1375 			sblock->header_error = 1;
1376 			sblock->generation_error = 1;
1377 		}
1378 		csum = h->csum;
1379 	} else {
1380 		if (!have_csum)
1381 			return;
1382 
1383 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1384 	}
1385 
1386 	for (page_num = 0;;) {
1387 		if (page_num == 0 && is_metadata)
1388 			crc = btrfs_csum_data(
1389 				((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1390 				crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1391 		else
1392 			crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1393 
1394 		kunmap_atomic(mapped_buffer);
1395 		page_num++;
1396 		if (page_num >= sblock->page_count)
1397 			break;
1398 		WARN_ON(!sblock->pagev[page_num]->page);
1399 
1400 		mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1401 	}
1402 
1403 	btrfs_csum_final(crc, calculated_csum);
1404 	if (memcmp(calculated_csum, csum, csum_size))
1405 		sblock->checksum_error = 1;
1406 }
1407 
1408 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1409 					     struct scrub_block *sblock_good,
1410 					     int force_write)
1411 {
1412 	int page_num;
1413 	int ret = 0;
1414 
1415 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1416 		int ret_sub;
1417 
1418 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1419 							   sblock_good,
1420 							   page_num,
1421 							   force_write);
1422 		if (ret_sub)
1423 			ret = ret_sub;
1424 	}
1425 
1426 	return ret;
1427 }
1428 
1429 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1430 					    struct scrub_block *sblock_good,
1431 					    int page_num, int force_write)
1432 {
1433 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1434 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1435 
1436 	BUG_ON(page_bad->page == NULL);
1437 	BUG_ON(page_good->page == NULL);
1438 	if (force_write || sblock_bad->header_error ||
1439 	    sblock_bad->checksum_error || page_bad->io_error) {
1440 		struct bio *bio;
1441 		int ret;
1442 
1443 		if (!page_bad->dev->bdev) {
1444 			printk_ratelimited(KERN_WARNING "BTRFS: "
1445 				"scrub_repair_page_from_good_copy(bdev == NULL) "
1446 				"is unexpected!\n");
1447 			return -EIO;
1448 		}
1449 
1450 		bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1451 		if (!bio)
1452 			return -EIO;
1453 		bio->bi_bdev = page_bad->dev->bdev;
1454 		bio->bi_iter.bi_sector = page_bad->physical >> 9;
1455 
1456 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1457 		if (PAGE_SIZE != ret) {
1458 			bio_put(bio);
1459 			return -EIO;
1460 		}
1461 
1462 		if (btrfsic_submit_bio_wait(WRITE, bio)) {
1463 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1464 				BTRFS_DEV_STAT_WRITE_ERRS);
1465 			btrfs_dev_replace_stats_inc(
1466 				&sblock_bad->sctx->dev_root->fs_info->
1467 				dev_replace.num_write_errors);
1468 			bio_put(bio);
1469 			return -EIO;
1470 		}
1471 		bio_put(bio);
1472 	}
1473 
1474 	return 0;
1475 }
1476 
1477 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1478 {
1479 	int page_num;
1480 
1481 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1482 		int ret;
1483 
1484 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1485 		if (ret)
1486 			btrfs_dev_replace_stats_inc(
1487 				&sblock->sctx->dev_root->fs_info->dev_replace.
1488 				num_write_errors);
1489 	}
1490 }
1491 
1492 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1493 					   int page_num)
1494 {
1495 	struct scrub_page *spage = sblock->pagev[page_num];
1496 
1497 	BUG_ON(spage->page == NULL);
1498 	if (spage->io_error) {
1499 		void *mapped_buffer = kmap_atomic(spage->page);
1500 
1501 		memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1502 		flush_dcache_page(spage->page);
1503 		kunmap_atomic(mapped_buffer);
1504 	}
1505 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1506 }
1507 
1508 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1509 				    struct scrub_page *spage)
1510 {
1511 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1512 	struct scrub_bio *sbio;
1513 	int ret;
1514 
1515 	mutex_lock(&wr_ctx->wr_lock);
1516 again:
1517 	if (!wr_ctx->wr_curr_bio) {
1518 		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1519 					      GFP_NOFS);
1520 		if (!wr_ctx->wr_curr_bio) {
1521 			mutex_unlock(&wr_ctx->wr_lock);
1522 			return -ENOMEM;
1523 		}
1524 		wr_ctx->wr_curr_bio->sctx = sctx;
1525 		wr_ctx->wr_curr_bio->page_count = 0;
1526 	}
1527 	sbio = wr_ctx->wr_curr_bio;
1528 	if (sbio->page_count == 0) {
1529 		struct bio *bio;
1530 
1531 		sbio->physical = spage->physical_for_dev_replace;
1532 		sbio->logical = spage->logical;
1533 		sbio->dev = wr_ctx->tgtdev;
1534 		bio = sbio->bio;
1535 		if (!bio) {
1536 			bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1537 			if (!bio) {
1538 				mutex_unlock(&wr_ctx->wr_lock);
1539 				return -ENOMEM;
1540 			}
1541 			sbio->bio = bio;
1542 		}
1543 
1544 		bio->bi_private = sbio;
1545 		bio->bi_end_io = scrub_wr_bio_end_io;
1546 		bio->bi_bdev = sbio->dev->bdev;
1547 		bio->bi_iter.bi_sector = sbio->physical >> 9;
1548 		sbio->err = 0;
1549 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1550 		   spage->physical_for_dev_replace ||
1551 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1552 		   spage->logical) {
1553 		scrub_wr_submit(sctx);
1554 		goto again;
1555 	}
1556 
1557 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1558 	if (ret != PAGE_SIZE) {
1559 		if (sbio->page_count < 1) {
1560 			bio_put(sbio->bio);
1561 			sbio->bio = NULL;
1562 			mutex_unlock(&wr_ctx->wr_lock);
1563 			return -EIO;
1564 		}
1565 		scrub_wr_submit(sctx);
1566 		goto again;
1567 	}
1568 
1569 	sbio->pagev[sbio->page_count] = spage;
1570 	scrub_page_get(spage);
1571 	sbio->page_count++;
1572 	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1573 		scrub_wr_submit(sctx);
1574 	mutex_unlock(&wr_ctx->wr_lock);
1575 
1576 	return 0;
1577 }
1578 
1579 static void scrub_wr_submit(struct scrub_ctx *sctx)
1580 {
1581 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1582 	struct scrub_bio *sbio;
1583 
1584 	if (!wr_ctx->wr_curr_bio)
1585 		return;
1586 
1587 	sbio = wr_ctx->wr_curr_bio;
1588 	wr_ctx->wr_curr_bio = NULL;
1589 	WARN_ON(!sbio->bio->bi_bdev);
1590 	scrub_pending_bio_inc(sctx);
1591 	/* process all writes in a single worker thread. Then the block layer
1592 	 * orders the requests before sending them to the driver which
1593 	 * doubled the write performance on spinning disks when measured
1594 	 * with Linux 3.5 */
1595 	btrfsic_submit_bio(WRITE, sbio->bio);
1596 }
1597 
1598 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1599 {
1600 	struct scrub_bio *sbio = bio->bi_private;
1601 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1602 
1603 	sbio->err = err;
1604 	sbio->bio = bio;
1605 
1606 	sbio->work.func = scrub_wr_bio_end_io_worker;
1607 	btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1608 }
1609 
1610 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1611 {
1612 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1613 	struct scrub_ctx *sctx = sbio->sctx;
1614 	int i;
1615 
1616 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1617 	if (sbio->err) {
1618 		struct btrfs_dev_replace *dev_replace =
1619 			&sbio->sctx->dev_root->fs_info->dev_replace;
1620 
1621 		for (i = 0; i < sbio->page_count; i++) {
1622 			struct scrub_page *spage = sbio->pagev[i];
1623 
1624 			spage->io_error = 1;
1625 			btrfs_dev_replace_stats_inc(&dev_replace->
1626 						    num_write_errors);
1627 		}
1628 	}
1629 
1630 	for (i = 0; i < sbio->page_count; i++)
1631 		scrub_page_put(sbio->pagev[i]);
1632 
1633 	bio_put(sbio->bio);
1634 	kfree(sbio);
1635 	scrub_pending_bio_dec(sctx);
1636 }
1637 
1638 static int scrub_checksum(struct scrub_block *sblock)
1639 {
1640 	u64 flags;
1641 	int ret;
1642 
1643 	WARN_ON(sblock->page_count < 1);
1644 	flags = sblock->pagev[0]->flags;
1645 	ret = 0;
1646 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1647 		ret = scrub_checksum_data(sblock);
1648 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1649 		ret = scrub_checksum_tree_block(sblock);
1650 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1651 		(void)scrub_checksum_super(sblock);
1652 	else
1653 		WARN_ON(1);
1654 	if (ret)
1655 		scrub_handle_errored_block(sblock);
1656 
1657 	return ret;
1658 }
1659 
1660 static int scrub_checksum_data(struct scrub_block *sblock)
1661 {
1662 	struct scrub_ctx *sctx = sblock->sctx;
1663 	u8 csum[BTRFS_CSUM_SIZE];
1664 	u8 *on_disk_csum;
1665 	struct page *page;
1666 	void *buffer;
1667 	u32 crc = ~(u32)0;
1668 	int fail = 0;
1669 	u64 len;
1670 	int index;
1671 
1672 	BUG_ON(sblock->page_count < 1);
1673 	if (!sblock->pagev[0]->have_csum)
1674 		return 0;
1675 
1676 	on_disk_csum = sblock->pagev[0]->csum;
1677 	page = sblock->pagev[0]->page;
1678 	buffer = kmap_atomic(page);
1679 
1680 	len = sctx->sectorsize;
1681 	index = 0;
1682 	for (;;) {
1683 		u64 l = min_t(u64, len, PAGE_SIZE);
1684 
1685 		crc = btrfs_csum_data(buffer, crc, l);
1686 		kunmap_atomic(buffer);
1687 		len -= l;
1688 		if (len == 0)
1689 			break;
1690 		index++;
1691 		BUG_ON(index >= sblock->page_count);
1692 		BUG_ON(!sblock->pagev[index]->page);
1693 		page = sblock->pagev[index]->page;
1694 		buffer = kmap_atomic(page);
1695 	}
1696 
1697 	btrfs_csum_final(crc, csum);
1698 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1699 		fail = 1;
1700 
1701 	return fail;
1702 }
1703 
1704 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1705 {
1706 	struct scrub_ctx *sctx = sblock->sctx;
1707 	struct btrfs_header *h;
1708 	struct btrfs_root *root = sctx->dev_root;
1709 	struct btrfs_fs_info *fs_info = root->fs_info;
1710 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1711 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1712 	struct page *page;
1713 	void *mapped_buffer;
1714 	u64 mapped_size;
1715 	void *p;
1716 	u32 crc = ~(u32)0;
1717 	int fail = 0;
1718 	int crc_fail = 0;
1719 	u64 len;
1720 	int index;
1721 
1722 	BUG_ON(sblock->page_count < 1);
1723 	page = sblock->pagev[0]->page;
1724 	mapped_buffer = kmap_atomic(page);
1725 	h = (struct btrfs_header *)mapped_buffer;
1726 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1727 
1728 	/*
1729 	 * we don't use the getter functions here, as we
1730 	 * a) don't have an extent buffer and
1731 	 * b) the page is already kmapped
1732 	 */
1733 
1734 	if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1735 		++fail;
1736 
1737 	if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1738 		++fail;
1739 
1740 	if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1741 		++fail;
1742 
1743 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1744 		   BTRFS_UUID_SIZE))
1745 		++fail;
1746 
1747 	WARN_ON(sctx->nodesize != sctx->leafsize);
1748 	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1749 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1750 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1751 	index = 0;
1752 	for (;;) {
1753 		u64 l = min_t(u64, len, mapped_size);
1754 
1755 		crc = btrfs_csum_data(p, crc, l);
1756 		kunmap_atomic(mapped_buffer);
1757 		len -= l;
1758 		if (len == 0)
1759 			break;
1760 		index++;
1761 		BUG_ON(index >= sblock->page_count);
1762 		BUG_ON(!sblock->pagev[index]->page);
1763 		page = sblock->pagev[index]->page;
1764 		mapped_buffer = kmap_atomic(page);
1765 		mapped_size = PAGE_SIZE;
1766 		p = mapped_buffer;
1767 	}
1768 
1769 	btrfs_csum_final(crc, calculated_csum);
1770 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1771 		++crc_fail;
1772 
1773 	return fail || crc_fail;
1774 }
1775 
1776 static int scrub_checksum_super(struct scrub_block *sblock)
1777 {
1778 	struct btrfs_super_block *s;
1779 	struct scrub_ctx *sctx = sblock->sctx;
1780 	struct btrfs_root *root = sctx->dev_root;
1781 	struct btrfs_fs_info *fs_info = root->fs_info;
1782 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1783 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1784 	struct page *page;
1785 	void *mapped_buffer;
1786 	u64 mapped_size;
1787 	void *p;
1788 	u32 crc = ~(u32)0;
1789 	int fail_gen = 0;
1790 	int fail_cor = 0;
1791 	u64 len;
1792 	int index;
1793 
1794 	BUG_ON(sblock->page_count < 1);
1795 	page = sblock->pagev[0]->page;
1796 	mapped_buffer = kmap_atomic(page);
1797 	s = (struct btrfs_super_block *)mapped_buffer;
1798 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1799 
1800 	if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1801 		++fail_cor;
1802 
1803 	if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1804 		++fail_gen;
1805 
1806 	if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1807 		++fail_cor;
1808 
1809 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1810 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1811 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1812 	index = 0;
1813 	for (;;) {
1814 		u64 l = min_t(u64, len, mapped_size);
1815 
1816 		crc = btrfs_csum_data(p, crc, l);
1817 		kunmap_atomic(mapped_buffer);
1818 		len -= l;
1819 		if (len == 0)
1820 			break;
1821 		index++;
1822 		BUG_ON(index >= sblock->page_count);
1823 		BUG_ON(!sblock->pagev[index]->page);
1824 		page = sblock->pagev[index]->page;
1825 		mapped_buffer = kmap_atomic(page);
1826 		mapped_size = PAGE_SIZE;
1827 		p = mapped_buffer;
1828 	}
1829 
1830 	btrfs_csum_final(crc, calculated_csum);
1831 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1832 		++fail_cor;
1833 
1834 	if (fail_cor + fail_gen) {
1835 		/*
1836 		 * if we find an error in a super block, we just report it.
1837 		 * They will get written with the next transaction commit
1838 		 * anyway
1839 		 */
1840 		spin_lock(&sctx->stat_lock);
1841 		++sctx->stat.super_errors;
1842 		spin_unlock(&sctx->stat_lock);
1843 		if (fail_cor)
1844 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1845 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1846 		else
1847 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1848 				BTRFS_DEV_STAT_GENERATION_ERRS);
1849 	}
1850 
1851 	return fail_cor + fail_gen;
1852 }
1853 
1854 static void scrub_block_get(struct scrub_block *sblock)
1855 {
1856 	atomic_inc(&sblock->ref_count);
1857 }
1858 
1859 static void scrub_block_put(struct scrub_block *sblock)
1860 {
1861 	if (atomic_dec_and_test(&sblock->ref_count)) {
1862 		int i;
1863 
1864 		for (i = 0; i < sblock->page_count; i++)
1865 			scrub_page_put(sblock->pagev[i]);
1866 		kfree(sblock);
1867 	}
1868 }
1869 
1870 static void scrub_page_get(struct scrub_page *spage)
1871 {
1872 	atomic_inc(&spage->ref_count);
1873 }
1874 
1875 static void scrub_page_put(struct scrub_page *spage)
1876 {
1877 	if (atomic_dec_and_test(&spage->ref_count)) {
1878 		if (spage->page)
1879 			__free_page(spage->page);
1880 		kfree(spage);
1881 	}
1882 }
1883 
1884 static void scrub_submit(struct scrub_ctx *sctx)
1885 {
1886 	struct scrub_bio *sbio;
1887 
1888 	if (sctx->curr == -1)
1889 		return;
1890 
1891 	sbio = sctx->bios[sctx->curr];
1892 	sctx->curr = -1;
1893 	scrub_pending_bio_inc(sctx);
1894 
1895 	if (!sbio->bio->bi_bdev) {
1896 		/*
1897 		 * this case should not happen. If btrfs_map_block() is
1898 		 * wrong, it could happen for dev-replace operations on
1899 		 * missing devices when no mirrors are available, but in
1900 		 * this case it should already fail the mount.
1901 		 * This case is handled correctly (but _very_ slowly).
1902 		 */
1903 		printk_ratelimited(KERN_WARNING
1904 			"BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1905 		bio_endio(sbio->bio, -EIO);
1906 	} else {
1907 		btrfsic_submit_bio(READ, sbio->bio);
1908 	}
1909 }
1910 
1911 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1912 				    struct scrub_page *spage)
1913 {
1914 	struct scrub_block *sblock = spage->sblock;
1915 	struct scrub_bio *sbio;
1916 	int ret;
1917 
1918 again:
1919 	/*
1920 	 * grab a fresh bio or wait for one to become available
1921 	 */
1922 	while (sctx->curr == -1) {
1923 		spin_lock(&sctx->list_lock);
1924 		sctx->curr = sctx->first_free;
1925 		if (sctx->curr != -1) {
1926 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1927 			sctx->bios[sctx->curr]->next_free = -1;
1928 			sctx->bios[sctx->curr]->page_count = 0;
1929 			spin_unlock(&sctx->list_lock);
1930 		} else {
1931 			spin_unlock(&sctx->list_lock);
1932 			wait_event(sctx->list_wait, sctx->first_free != -1);
1933 		}
1934 	}
1935 	sbio = sctx->bios[sctx->curr];
1936 	if (sbio->page_count == 0) {
1937 		struct bio *bio;
1938 
1939 		sbio->physical = spage->physical;
1940 		sbio->logical = spage->logical;
1941 		sbio->dev = spage->dev;
1942 		bio = sbio->bio;
1943 		if (!bio) {
1944 			bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1945 			if (!bio)
1946 				return -ENOMEM;
1947 			sbio->bio = bio;
1948 		}
1949 
1950 		bio->bi_private = sbio;
1951 		bio->bi_end_io = scrub_bio_end_io;
1952 		bio->bi_bdev = sbio->dev->bdev;
1953 		bio->bi_iter.bi_sector = sbio->physical >> 9;
1954 		sbio->err = 0;
1955 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1956 		   spage->physical ||
1957 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1958 		   spage->logical ||
1959 		   sbio->dev != spage->dev) {
1960 		scrub_submit(sctx);
1961 		goto again;
1962 	}
1963 
1964 	sbio->pagev[sbio->page_count] = spage;
1965 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1966 	if (ret != PAGE_SIZE) {
1967 		if (sbio->page_count < 1) {
1968 			bio_put(sbio->bio);
1969 			sbio->bio = NULL;
1970 			return -EIO;
1971 		}
1972 		scrub_submit(sctx);
1973 		goto again;
1974 	}
1975 
1976 	scrub_block_get(sblock); /* one for the page added to the bio */
1977 	atomic_inc(&sblock->outstanding_pages);
1978 	sbio->page_count++;
1979 	if (sbio->page_count == sctx->pages_per_rd_bio)
1980 		scrub_submit(sctx);
1981 
1982 	return 0;
1983 }
1984 
1985 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1986 		       u64 physical, struct btrfs_device *dev, u64 flags,
1987 		       u64 gen, int mirror_num, u8 *csum, int force,
1988 		       u64 physical_for_dev_replace)
1989 {
1990 	struct scrub_block *sblock;
1991 	int index;
1992 
1993 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1994 	if (!sblock) {
1995 		spin_lock(&sctx->stat_lock);
1996 		sctx->stat.malloc_errors++;
1997 		spin_unlock(&sctx->stat_lock);
1998 		return -ENOMEM;
1999 	}
2000 
2001 	/* one ref inside this function, plus one for each page added to
2002 	 * a bio later on */
2003 	atomic_set(&sblock->ref_count, 1);
2004 	sblock->sctx = sctx;
2005 	sblock->no_io_error_seen = 1;
2006 
2007 	for (index = 0; len > 0; index++) {
2008 		struct scrub_page *spage;
2009 		u64 l = min_t(u64, len, PAGE_SIZE);
2010 
2011 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
2012 		if (!spage) {
2013 leave_nomem:
2014 			spin_lock(&sctx->stat_lock);
2015 			sctx->stat.malloc_errors++;
2016 			spin_unlock(&sctx->stat_lock);
2017 			scrub_block_put(sblock);
2018 			return -ENOMEM;
2019 		}
2020 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2021 		scrub_page_get(spage);
2022 		sblock->pagev[index] = spage;
2023 		spage->sblock = sblock;
2024 		spage->dev = dev;
2025 		spage->flags = flags;
2026 		spage->generation = gen;
2027 		spage->logical = logical;
2028 		spage->physical = physical;
2029 		spage->physical_for_dev_replace = physical_for_dev_replace;
2030 		spage->mirror_num = mirror_num;
2031 		if (csum) {
2032 			spage->have_csum = 1;
2033 			memcpy(spage->csum, csum, sctx->csum_size);
2034 		} else {
2035 			spage->have_csum = 0;
2036 		}
2037 		sblock->page_count++;
2038 		spage->page = alloc_page(GFP_NOFS);
2039 		if (!spage->page)
2040 			goto leave_nomem;
2041 		len -= l;
2042 		logical += l;
2043 		physical += l;
2044 		physical_for_dev_replace += l;
2045 	}
2046 
2047 	WARN_ON(sblock->page_count == 0);
2048 	for (index = 0; index < sblock->page_count; index++) {
2049 		struct scrub_page *spage = sblock->pagev[index];
2050 		int ret;
2051 
2052 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2053 		if (ret) {
2054 			scrub_block_put(sblock);
2055 			return ret;
2056 		}
2057 	}
2058 
2059 	if (force)
2060 		scrub_submit(sctx);
2061 
2062 	/* last one frees, either here or in bio completion for last page */
2063 	scrub_block_put(sblock);
2064 	return 0;
2065 }
2066 
2067 static void scrub_bio_end_io(struct bio *bio, int err)
2068 {
2069 	struct scrub_bio *sbio = bio->bi_private;
2070 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2071 
2072 	sbio->err = err;
2073 	sbio->bio = bio;
2074 
2075 	btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2076 }
2077 
2078 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2079 {
2080 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2081 	struct scrub_ctx *sctx = sbio->sctx;
2082 	int i;
2083 
2084 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2085 	if (sbio->err) {
2086 		for (i = 0; i < sbio->page_count; i++) {
2087 			struct scrub_page *spage = sbio->pagev[i];
2088 
2089 			spage->io_error = 1;
2090 			spage->sblock->no_io_error_seen = 0;
2091 		}
2092 	}
2093 
2094 	/* now complete the scrub_block items that have all pages completed */
2095 	for (i = 0; i < sbio->page_count; i++) {
2096 		struct scrub_page *spage = sbio->pagev[i];
2097 		struct scrub_block *sblock = spage->sblock;
2098 
2099 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2100 			scrub_block_complete(sblock);
2101 		scrub_block_put(sblock);
2102 	}
2103 
2104 	bio_put(sbio->bio);
2105 	sbio->bio = NULL;
2106 	spin_lock(&sctx->list_lock);
2107 	sbio->next_free = sctx->first_free;
2108 	sctx->first_free = sbio->index;
2109 	spin_unlock(&sctx->list_lock);
2110 
2111 	if (sctx->is_dev_replace &&
2112 	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2113 		mutex_lock(&sctx->wr_ctx.wr_lock);
2114 		scrub_wr_submit(sctx);
2115 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2116 	}
2117 
2118 	scrub_pending_bio_dec(sctx);
2119 }
2120 
2121 static void scrub_block_complete(struct scrub_block *sblock)
2122 {
2123 	if (!sblock->no_io_error_seen) {
2124 		scrub_handle_errored_block(sblock);
2125 	} else {
2126 		/*
2127 		 * if has checksum error, write via repair mechanism in
2128 		 * dev replace case, otherwise write here in dev replace
2129 		 * case.
2130 		 */
2131 		if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2132 			scrub_write_block_to_dev_replace(sblock);
2133 	}
2134 }
2135 
2136 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2137 			   u8 *csum)
2138 {
2139 	struct btrfs_ordered_sum *sum = NULL;
2140 	unsigned long index;
2141 	unsigned long num_sectors;
2142 
2143 	while (!list_empty(&sctx->csum_list)) {
2144 		sum = list_first_entry(&sctx->csum_list,
2145 				       struct btrfs_ordered_sum, list);
2146 		if (sum->bytenr > logical)
2147 			return 0;
2148 		if (sum->bytenr + sum->len > logical)
2149 			break;
2150 
2151 		++sctx->stat.csum_discards;
2152 		list_del(&sum->list);
2153 		kfree(sum);
2154 		sum = NULL;
2155 	}
2156 	if (!sum)
2157 		return 0;
2158 
2159 	index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2160 	num_sectors = sum->len / sctx->sectorsize;
2161 	memcpy(csum, sum->sums + index, sctx->csum_size);
2162 	if (index == num_sectors - 1) {
2163 		list_del(&sum->list);
2164 		kfree(sum);
2165 	}
2166 	return 1;
2167 }
2168 
2169 /* scrub extent tries to collect up to 64 kB for each bio */
2170 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2171 			u64 physical, struct btrfs_device *dev, u64 flags,
2172 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2173 {
2174 	int ret;
2175 	u8 csum[BTRFS_CSUM_SIZE];
2176 	u32 blocksize;
2177 
2178 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2179 		blocksize = sctx->sectorsize;
2180 		spin_lock(&sctx->stat_lock);
2181 		sctx->stat.data_extents_scrubbed++;
2182 		sctx->stat.data_bytes_scrubbed += len;
2183 		spin_unlock(&sctx->stat_lock);
2184 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2185 		WARN_ON(sctx->nodesize != sctx->leafsize);
2186 		blocksize = sctx->nodesize;
2187 		spin_lock(&sctx->stat_lock);
2188 		sctx->stat.tree_extents_scrubbed++;
2189 		sctx->stat.tree_bytes_scrubbed += len;
2190 		spin_unlock(&sctx->stat_lock);
2191 	} else {
2192 		blocksize = sctx->sectorsize;
2193 		WARN_ON(1);
2194 	}
2195 
2196 	while (len) {
2197 		u64 l = min_t(u64, len, blocksize);
2198 		int have_csum = 0;
2199 
2200 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2201 			/* push csums to sbio */
2202 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2203 			if (have_csum == 0)
2204 				++sctx->stat.no_csum;
2205 			if (sctx->is_dev_replace && !have_csum) {
2206 				ret = copy_nocow_pages(sctx, logical, l,
2207 						       mirror_num,
2208 						      physical_for_dev_replace);
2209 				goto behind_scrub_pages;
2210 			}
2211 		}
2212 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2213 				  mirror_num, have_csum ? csum : NULL, 0,
2214 				  physical_for_dev_replace);
2215 behind_scrub_pages:
2216 		if (ret)
2217 			return ret;
2218 		len -= l;
2219 		logical += l;
2220 		physical += l;
2221 		physical_for_dev_replace += l;
2222 	}
2223 	return 0;
2224 }
2225 
2226 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2227 					   struct map_lookup *map,
2228 					   struct btrfs_device *scrub_dev,
2229 					   int num, u64 base, u64 length,
2230 					   int is_dev_replace)
2231 {
2232 	struct btrfs_path *path;
2233 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2234 	struct btrfs_root *root = fs_info->extent_root;
2235 	struct btrfs_root *csum_root = fs_info->csum_root;
2236 	struct btrfs_extent_item *extent;
2237 	struct blk_plug plug;
2238 	u64 flags;
2239 	int ret;
2240 	int slot;
2241 	u64 nstripes;
2242 	struct extent_buffer *l;
2243 	struct btrfs_key key;
2244 	u64 physical;
2245 	u64 logical;
2246 	u64 logic_end;
2247 	u64 generation;
2248 	int mirror_num;
2249 	struct reada_control *reada1;
2250 	struct reada_control *reada2;
2251 	struct btrfs_key key_start;
2252 	struct btrfs_key key_end;
2253 	u64 increment = map->stripe_len;
2254 	u64 offset;
2255 	u64 extent_logical;
2256 	u64 extent_physical;
2257 	u64 extent_len;
2258 	struct btrfs_device *extent_dev;
2259 	int extent_mirror_num;
2260 	int stop_loop;
2261 
2262 	if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2263 			 BTRFS_BLOCK_GROUP_RAID6)) {
2264 		if (num >= nr_data_stripes(map)) {
2265 			return 0;
2266 		}
2267 	}
2268 
2269 	nstripes = length;
2270 	offset = 0;
2271 	do_div(nstripes, map->stripe_len);
2272 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2273 		offset = map->stripe_len * num;
2274 		increment = map->stripe_len * map->num_stripes;
2275 		mirror_num = 1;
2276 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2277 		int factor = map->num_stripes / map->sub_stripes;
2278 		offset = map->stripe_len * (num / map->sub_stripes);
2279 		increment = map->stripe_len * factor;
2280 		mirror_num = num % map->sub_stripes + 1;
2281 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2282 		increment = map->stripe_len;
2283 		mirror_num = num % map->num_stripes + 1;
2284 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2285 		increment = map->stripe_len;
2286 		mirror_num = num % map->num_stripes + 1;
2287 	} else {
2288 		increment = map->stripe_len;
2289 		mirror_num = 1;
2290 	}
2291 
2292 	path = btrfs_alloc_path();
2293 	if (!path)
2294 		return -ENOMEM;
2295 
2296 	/*
2297 	 * work on commit root. The related disk blocks are static as
2298 	 * long as COW is applied. This means, it is save to rewrite
2299 	 * them to repair disk errors without any race conditions
2300 	 */
2301 	path->search_commit_root = 1;
2302 	path->skip_locking = 1;
2303 
2304 	/*
2305 	 * trigger the readahead for extent tree csum tree and wait for
2306 	 * completion. During readahead, the scrub is officially paused
2307 	 * to not hold off transaction commits
2308 	 */
2309 	logical = base + offset;
2310 
2311 	wait_event(sctx->list_wait,
2312 		   atomic_read(&sctx->bios_in_flight) == 0);
2313 	scrub_blocked_if_needed(fs_info);
2314 
2315 	/* FIXME it might be better to start readahead at commit root */
2316 	key_start.objectid = logical;
2317 	key_start.type = BTRFS_EXTENT_ITEM_KEY;
2318 	key_start.offset = (u64)0;
2319 	key_end.objectid = base + offset + nstripes * increment;
2320 	key_end.type = BTRFS_METADATA_ITEM_KEY;
2321 	key_end.offset = (u64)-1;
2322 	reada1 = btrfs_reada_add(root, &key_start, &key_end);
2323 
2324 	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2325 	key_start.type = BTRFS_EXTENT_CSUM_KEY;
2326 	key_start.offset = logical;
2327 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2328 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
2329 	key_end.offset = base + offset + nstripes * increment;
2330 	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2331 
2332 	if (!IS_ERR(reada1))
2333 		btrfs_reada_wait(reada1);
2334 	if (!IS_ERR(reada2))
2335 		btrfs_reada_wait(reada2);
2336 
2337 
2338 	/*
2339 	 * collect all data csums for the stripe to avoid seeking during
2340 	 * the scrub. This might currently (crc32) end up to be about 1MB
2341 	 */
2342 	blk_start_plug(&plug);
2343 
2344 	/*
2345 	 * now find all extents for each stripe and scrub them
2346 	 */
2347 	logical = base + offset;
2348 	physical = map->stripes[num].physical;
2349 	logic_end = logical + increment * nstripes;
2350 	ret = 0;
2351 	while (logical < logic_end) {
2352 		/*
2353 		 * canceled?
2354 		 */
2355 		if (atomic_read(&fs_info->scrub_cancel_req) ||
2356 		    atomic_read(&sctx->cancel_req)) {
2357 			ret = -ECANCELED;
2358 			goto out;
2359 		}
2360 		/*
2361 		 * check to see if we have to pause
2362 		 */
2363 		if (atomic_read(&fs_info->scrub_pause_req)) {
2364 			/* push queued extents */
2365 			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2366 			scrub_submit(sctx);
2367 			mutex_lock(&sctx->wr_ctx.wr_lock);
2368 			scrub_wr_submit(sctx);
2369 			mutex_unlock(&sctx->wr_ctx.wr_lock);
2370 			wait_event(sctx->list_wait,
2371 				   atomic_read(&sctx->bios_in_flight) == 0);
2372 			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2373 			scrub_blocked_if_needed(fs_info);
2374 		}
2375 
2376 		if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2377 			key.type = BTRFS_METADATA_ITEM_KEY;
2378 		else
2379 			key.type = BTRFS_EXTENT_ITEM_KEY;
2380 		key.objectid = logical;
2381 		key.offset = (u64)-1;
2382 
2383 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2384 		if (ret < 0)
2385 			goto out;
2386 
2387 		if (ret > 0) {
2388 			ret = btrfs_previous_extent_item(root, path, 0);
2389 			if (ret < 0)
2390 				goto out;
2391 			if (ret > 0) {
2392 				/* there's no smaller item, so stick with the
2393 				 * larger one */
2394 				btrfs_release_path(path);
2395 				ret = btrfs_search_slot(NULL, root, &key,
2396 							path, 0, 0);
2397 				if (ret < 0)
2398 					goto out;
2399 			}
2400 		}
2401 
2402 		stop_loop = 0;
2403 		while (1) {
2404 			u64 bytes;
2405 
2406 			l = path->nodes[0];
2407 			slot = path->slots[0];
2408 			if (slot >= btrfs_header_nritems(l)) {
2409 				ret = btrfs_next_leaf(root, path);
2410 				if (ret == 0)
2411 					continue;
2412 				if (ret < 0)
2413 					goto out;
2414 
2415 				stop_loop = 1;
2416 				break;
2417 			}
2418 			btrfs_item_key_to_cpu(l, &key, slot);
2419 
2420 			if (key.type == BTRFS_METADATA_ITEM_KEY)
2421 				bytes = root->leafsize;
2422 			else
2423 				bytes = key.offset;
2424 
2425 			if (key.objectid + bytes <= logical)
2426 				goto next;
2427 
2428 			if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2429 			    key.type != BTRFS_METADATA_ITEM_KEY)
2430 				goto next;
2431 
2432 			if (key.objectid >= logical + map->stripe_len) {
2433 				/* out of this device extent */
2434 				if (key.objectid >= logic_end)
2435 					stop_loop = 1;
2436 				break;
2437 			}
2438 
2439 			extent = btrfs_item_ptr(l, slot,
2440 						struct btrfs_extent_item);
2441 			flags = btrfs_extent_flags(l, extent);
2442 			generation = btrfs_extent_generation(l, extent);
2443 
2444 			if (key.objectid < logical &&
2445 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2446 				btrfs_err(fs_info,
2447 					   "scrub: tree block %llu spanning "
2448 					   "stripes, ignored. logical=%llu",
2449 				       key.objectid, logical);
2450 				goto next;
2451 			}
2452 
2453 again:
2454 			extent_logical = key.objectid;
2455 			extent_len = bytes;
2456 
2457 			/*
2458 			 * trim extent to this stripe
2459 			 */
2460 			if (extent_logical < logical) {
2461 				extent_len -= logical - extent_logical;
2462 				extent_logical = logical;
2463 			}
2464 			if (extent_logical + extent_len >
2465 			    logical + map->stripe_len) {
2466 				extent_len = logical + map->stripe_len -
2467 					     extent_logical;
2468 			}
2469 
2470 			extent_physical = extent_logical - logical + physical;
2471 			extent_dev = scrub_dev;
2472 			extent_mirror_num = mirror_num;
2473 			if (is_dev_replace)
2474 				scrub_remap_extent(fs_info, extent_logical,
2475 						   extent_len, &extent_physical,
2476 						   &extent_dev,
2477 						   &extent_mirror_num);
2478 
2479 			ret = btrfs_lookup_csums_range(csum_root, logical,
2480 						logical + map->stripe_len - 1,
2481 						&sctx->csum_list, 1);
2482 			if (ret)
2483 				goto out;
2484 
2485 			ret = scrub_extent(sctx, extent_logical, extent_len,
2486 					   extent_physical, extent_dev, flags,
2487 					   generation, extent_mirror_num,
2488 					   extent_logical - logical + physical);
2489 			if (ret)
2490 				goto out;
2491 
2492 			scrub_free_csums(sctx);
2493 			if (extent_logical + extent_len <
2494 			    key.objectid + bytes) {
2495 				logical += increment;
2496 				physical += map->stripe_len;
2497 
2498 				if (logical < key.objectid + bytes) {
2499 					cond_resched();
2500 					goto again;
2501 				}
2502 
2503 				if (logical >= logic_end) {
2504 					stop_loop = 1;
2505 					break;
2506 				}
2507 			}
2508 next:
2509 			path->slots[0]++;
2510 		}
2511 		btrfs_release_path(path);
2512 		logical += increment;
2513 		physical += map->stripe_len;
2514 		spin_lock(&sctx->stat_lock);
2515 		if (stop_loop)
2516 			sctx->stat.last_physical = map->stripes[num].physical +
2517 						   length;
2518 		else
2519 			sctx->stat.last_physical = physical;
2520 		spin_unlock(&sctx->stat_lock);
2521 		if (stop_loop)
2522 			break;
2523 	}
2524 out:
2525 	/* push queued extents */
2526 	scrub_submit(sctx);
2527 	mutex_lock(&sctx->wr_ctx.wr_lock);
2528 	scrub_wr_submit(sctx);
2529 	mutex_unlock(&sctx->wr_ctx.wr_lock);
2530 
2531 	blk_finish_plug(&plug);
2532 	btrfs_free_path(path);
2533 	return ret < 0 ? ret : 0;
2534 }
2535 
2536 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2537 					  struct btrfs_device *scrub_dev,
2538 					  u64 chunk_tree, u64 chunk_objectid,
2539 					  u64 chunk_offset, u64 length,
2540 					  u64 dev_offset, int is_dev_replace)
2541 {
2542 	struct btrfs_mapping_tree *map_tree =
2543 		&sctx->dev_root->fs_info->mapping_tree;
2544 	struct map_lookup *map;
2545 	struct extent_map *em;
2546 	int i;
2547 	int ret = 0;
2548 
2549 	read_lock(&map_tree->map_tree.lock);
2550 	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2551 	read_unlock(&map_tree->map_tree.lock);
2552 
2553 	if (!em)
2554 		return -EINVAL;
2555 
2556 	map = (struct map_lookup *)em->bdev;
2557 	if (em->start != chunk_offset)
2558 		goto out;
2559 
2560 	if (em->len < length)
2561 		goto out;
2562 
2563 	for (i = 0; i < map->num_stripes; ++i) {
2564 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2565 		    map->stripes[i].physical == dev_offset) {
2566 			ret = scrub_stripe(sctx, map, scrub_dev, i,
2567 					   chunk_offset, length,
2568 					   is_dev_replace);
2569 			if (ret)
2570 				goto out;
2571 		}
2572 	}
2573 out:
2574 	free_extent_map(em);
2575 
2576 	return ret;
2577 }
2578 
2579 static noinline_for_stack
2580 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2581 			   struct btrfs_device *scrub_dev, u64 start, u64 end,
2582 			   int is_dev_replace)
2583 {
2584 	struct btrfs_dev_extent *dev_extent = NULL;
2585 	struct btrfs_path *path;
2586 	struct btrfs_root *root = sctx->dev_root;
2587 	struct btrfs_fs_info *fs_info = root->fs_info;
2588 	u64 length;
2589 	u64 chunk_tree;
2590 	u64 chunk_objectid;
2591 	u64 chunk_offset;
2592 	int ret;
2593 	int slot;
2594 	struct extent_buffer *l;
2595 	struct btrfs_key key;
2596 	struct btrfs_key found_key;
2597 	struct btrfs_block_group_cache *cache;
2598 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2599 
2600 	path = btrfs_alloc_path();
2601 	if (!path)
2602 		return -ENOMEM;
2603 
2604 	path->reada = 2;
2605 	path->search_commit_root = 1;
2606 	path->skip_locking = 1;
2607 
2608 	key.objectid = scrub_dev->devid;
2609 	key.offset = 0ull;
2610 	key.type = BTRFS_DEV_EXTENT_KEY;
2611 
2612 	while (1) {
2613 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2614 		if (ret < 0)
2615 			break;
2616 		if (ret > 0) {
2617 			if (path->slots[0] >=
2618 			    btrfs_header_nritems(path->nodes[0])) {
2619 				ret = btrfs_next_leaf(root, path);
2620 				if (ret)
2621 					break;
2622 			}
2623 		}
2624 
2625 		l = path->nodes[0];
2626 		slot = path->slots[0];
2627 
2628 		btrfs_item_key_to_cpu(l, &found_key, slot);
2629 
2630 		if (found_key.objectid != scrub_dev->devid)
2631 			break;
2632 
2633 		if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2634 			break;
2635 
2636 		if (found_key.offset >= end)
2637 			break;
2638 
2639 		if (found_key.offset < key.offset)
2640 			break;
2641 
2642 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2643 		length = btrfs_dev_extent_length(l, dev_extent);
2644 
2645 		if (found_key.offset + length <= start) {
2646 			key.offset = found_key.offset + length;
2647 			btrfs_release_path(path);
2648 			continue;
2649 		}
2650 
2651 		chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2652 		chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2653 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2654 
2655 		/*
2656 		 * get a reference on the corresponding block group to prevent
2657 		 * the chunk from going away while we scrub it
2658 		 */
2659 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2660 		if (!cache) {
2661 			ret = -ENOENT;
2662 			break;
2663 		}
2664 		dev_replace->cursor_right = found_key.offset + length;
2665 		dev_replace->cursor_left = found_key.offset;
2666 		dev_replace->item_needs_writeback = 1;
2667 		ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2668 				  chunk_offset, length, found_key.offset,
2669 				  is_dev_replace);
2670 
2671 		/*
2672 		 * flush, submit all pending read and write bios, afterwards
2673 		 * wait for them.
2674 		 * Note that in the dev replace case, a read request causes
2675 		 * write requests that are submitted in the read completion
2676 		 * worker. Therefore in the current situation, it is required
2677 		 * that all write requests are flushed, so that all read and
2678 		 * write requests are really completed when bios_in_flight
2679 		 * changes to 0.
2680 		 */
2681 		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2682 		scrub_submit(sctx);
2683 		mutex_lock(&sctx->wr_ctx.wr_lock);
2684 		scrub_wr_submit(sctx);
2685 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2686 
2687 		wait_event(sctx->list_wait,
2688 			   atomic_read(&sctx->bios_in_flight) == 0);
2689 		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2690 		wait_event(sctx->list_wait,
2691 			   atomic_read(&sctx->workers_pending) == 0);
2692 		scrub_blocked_if_needed(fs_info);
2693 
2694 		btrfs_put_block_group(cache);
2695 		if (ret)
2696 			break;
2697 		if (is_dev_replace &&
2698 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
2699 			ret = -EIO;
2700 			break;
2701 		}
2702 		if (sctx->stat.malloc_errors > 0) {
2703 			ret = -ENOMEM;
2704 			break;
2705 		}
2706 
2707 		dev_replace->cursor_left = dev_replace->cursor_right;
2708 		dev_replace->item_needs_writeback = 1;
2709 
2710 		key.offset = found_key.offset + length;
2711 		btrfs_release_path(path);
2712 	}
2713 
2714 	btrfs_free_path(path);
2715 
2716 	/*
2717 	 * ret can still be 1 from search_slot or next_leaf,
2718 	 * that's not an error
2719 	 */
2720 	return ret < 0 ? ret : 0;
2721 }
2722 
2723 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2724 					   struct btrfs_device *scrub_dev)
2725 {
2726 	int	i;
2727 	u64	bytenr;
2728 	u64	gen;
2729 	int	ret;
2730 	struct btrfs_root *root = sctx->dev_root;
2731 
2732 	if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2733 		return -EIO;
2734 
2735 	gen = root->fs_info->last_trans_committed;
2736 
2737 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2738 		bytenr = btrfs_sb_offset(i);
2739 		if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2740 			break;
2741 
2742 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2743 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2744 				  NULL, 1, bytenr);
2745 		if (ret)
2746 			return ret;
2747 	}
2748 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2749 
2750 	return 0;
2751 }
2752 
2753 /*
2754  * get a reference count on fs_info->scrub_workers. start worker if necessary
2755  */
2756 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2757 						int is_dev_replace)
2758 {
2759 	int ret = 0;
2760 
2761 	if (fs_info->scrub_workers_refcnt == 0) {
2762 		if (is_dev_replace)
2763 			btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2764 					&fs_info->generic_worker);
2765 		else
2766 			btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2767 					fs_info->thread_pool_size,
2768 					&fs_info->generic_worker);
2769 		fs_info->scrub_workers.idle_thresh = 4;
2770 		ret = btrfs_start_workers(&fs_info->scrub_workers);
2771 		if (ret)
2772 			goto out;
2773 		btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2774 				   "scrubwrc",
2775 				   fs_info->thread_pool_size,
2776 				   &fs_info->generic_worker);
2777 		fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2778 		ret = btrfs_start_workers(
2779 				&fs_info->scrub_wr_completion_workers);
2780 		if (ret)
2781 			goto out;
2782 		btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2783 				   &fs_info->generic_worker);
2784 		ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2785 		if (ret)
2786 			goto out;
2787 	}
2788 	++fs_info->scrub_workers_refcnt;
2789 out:
2790 	return ret;
2791 }
2792 
2793 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2794 {
2795 	if (--fs_info->scrub_workers_refcnt == 0) {
2796 		btrfs_stop_workers(&fs_info->scrub_workers);
2797 		btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2798 		btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2799 	}
2800 	WARN_ON(fs_info->scrub_workers_refcnt < 0);
2801 }
2802 
2803 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2804 		    u64 end, struct btrfs_scrub_progress *progress,
2805 		    int readonly, int is_dev_replace)
2806 {
2807 	struct scrub_ctx *sctx;
2808 	int ret;
2809 	struct btrfs_device *dev;
2810 
2811 	if (btrfs_fs_closing(fs_info))
2812 		return -EINVAL;
2813 
2814 	/*
2815 	 * check some assumptions
2816 	 */
2817 	if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2818 		btrfs_err(fs_info,
2819 			   "scrub: size assumption nodesize == leafsize (%d == %d) fails",
2820 		       fs_info->chunk_root->nodesize,
2821 		       fs_info->chunk_root->leafsize);
2822 		return -EINVAL;
2823 	}
2824 
2825 	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2826 		/*
2827 		 * in this case scrub is unable to calculate the checksum
2828 		 * the way scrub is implemented. Do not handle this
2829 		 * situation at all because it won't ever happen.
2830 		 */
2831 		btrfs_err(fs_info,
2832 			   "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2833 		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2834 		return -EINVAL;
2835 	}
2836 
2837 	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2838 		/* not supported for data w/o checksums */
2839 		btrfs_err(fs_info,
2840 			   "scrub: size assumption sectorsize != PAGE_SIZE "
2841 			   "(%d != %lu) fails",
2842 		       fs_info->chunk_root->sectorsize, PAGE_SIZE);
2843 		return -EINVAL;
2844 	}
2845 
2846 	if (fs_info->chunk_root->nodesize >
2847 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2848 	    fs_info->chunk_root->sectorsize >
2849 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2850 		/*
2851 		 * would exhaust the array bounds of pagev member in
2852 		 * struct scrub_block
2853 		 */
2854 		btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2855 			   "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2856 		       fs_info->chunk_root->nodesize,
2857 		       SCRUB_MAX_PAGES_PER_BLOCK,
2858 		       fs_info->chunk_root->sectorsize,
2859 		       SCRUB_MAX_PAGES_PER_BLOCK);
2860 		return -EINVAL;
2861 	}
2862 
2863 
2864 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2865 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2866 	if (!dev || (dev->missing && !is_dev_replace)) {
2867 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2868 		return -ENODEV;
2869 	}
2870 
2871 	mutex_lock(&fs_info->scrub_lock);
2872 	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2873 		mutex_unlock(&fs_info->scrub_lock);
2874 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2875 		return -EIO;
2876 	}
2877 
2878 	btrfs_dev_replace_lock(&fs_info->dev_replace);
2879 	if (dev->scrub_device ||
2880 	    (!is_dev_replace &&
2881 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2882 		btrfs_dev_replace_unlock(&fs_info->dev_replace);
2883 		mutex_unlock(&fs_info->scrub_lock);
2884 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2885 		return -EINPROGRESS;
2886 	}
2887 	btrfs_dev_replace_unlock(&fs_info->dev_replace);
2888 
2889 	ret = scrub_workers_get(fs_info, is_dev_replace);
2890 	if (ret) {
2891 		mutex_unlock(&fs_info->scrub_lock);
2892 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2893 		return ret;
2894 	}
2895 
2896 	sctx = scrub_setup_ctx(dev, is_dev_replace);
2897 	if (IS_ERR(sctx)) {
2898 		mutex_unlock(&fs_info->scrub_lock);
2899 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2900 		scrub_workers_put(fs_info);
2901 		return PTR_ERR(sctx);
2902 	}
2903 	sctx->readonly = readonly;
2904 	dev->scrub_device = sctx;
2905 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2906 
2907 	/*
2908 	 * checking @scrub_pause_req here, we can avoid
2909 	 * race between committing transaction and scrubbing.
2910 	 */
2911 	__scrub_blocked_if_needed(fs_info);
2912 	atomic_inc(&fs_info->scrubs_running);
2913 	mutex_unlock(&fs_info->scrub_lock);
2914 
2915 	if (!is_dev_replace) {
2916 		/*
2917 		 * by holding device list mutex, we can
2918 		 * kick off writing super in log tree sync.
2919 		 */
2920 		mutex_lock(&fs_info->fs_devices->device_list_mutex);
2921 		ret = scrub_supers(sctx, dev);
2922 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2923 	}
2924 
2925 	if (!ret)
2926 		ret = scrub_enumerate_chunks(sctx, dev, start, end,
2927 					     is_dev_replace);
2928 
2929 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2930 	atomic_dec(&fs_info->scrubs_running);
2931 	wake_up(&fs_info->scrub_pause_wait);
2932 
2933 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2934 
2935 	if (progress)
2936 		memcpy(progress, &sctx->stat, sizeof(*progress));
2937 
2938 	mutex_lock(&fs_info->scrub_lock);
2939 	dev->scrub_device = NULL;
2940 	scrub_workers_put(fs_info);
2941 	mutex_unlock(&fs_info->scrub_lock);
2942 
2943 	scrub_free_ctx(sctx);
2944 
2945 	return ret;
2946 }
2947 
2948 void btrfs_scrub_pause(struct btrfs_root *root)
2949 {
2950 	struct btrfs_fs_info *fs_info = root->fs_info;
2951 
2952 	mutex_lock(&fs_info->scrub_lock);
2953 	atomic_inc(&fs_info->scrub_pause_req);
2954 	while (atomic_read(&fs_info->scrubs_paused) !=
2955 	       atomic_read(&fs_info->scrubs_running)) {
2956 		mutex_unlock(&fs_info->scrub_lock);
2957 		wait_event(fs_info->scrub_pause_wait,
2958 			   atomic_read(&fs_info->scrubs_paused) ==
2959 			   atomic_read(&fs_info->scrubs_running));
2960 		mutex_lock(&fs_info->scrub_lock);
2961 	}
2962 	mutex_unlock(&fs_info->scrub_lock);
2963 }
2964 
2965 void btrfs_scrub_continue(struct btrfs_root *root)
2966 {
2967 	struct btrfs_fs_info *fs_info = root->fs_info;
2968 
2969 	atomic_dec(&fs_info->scrub_pause_req);
2970 	wake_up(&fs_info->scrub_pause_wait);
2971 }
2972 
2973 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2974 {
2975 	mutex_lock(&fs_info->scrub_lock);
2976 	if (!atomic_read(&fs_info->scrubs_running)) {
2977 		mutex_unlock(&fs_info->scrub_lock);
2978 		return -ENOTCONN;
2979 	}
2980 
2981 	atomic_inc(&fs_info->scrub_cancel_req);
2982 	while (atomic_read(&fs_info->scrubs_running)) {
2983 		mutex_unlock(&fs_info->scrub_lock);
2984 		wait_event(fs_info->scrub_pause_wait,
2985 			   atomic_read(&fs_info->scrubs_running) == 0);
2986 		mutex_lock(&fs_info->scrub_lock);
2987 	}
2988 	atomic_dec(&fs_info->scrub_cancel_req);
2989 	mutex_unlock(&fs_info->scrub_lock);
2990 
2991 	return 0;
2992 }
2993 
2994 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2995 			   struct btrfs_device *dev)
2996 {
2997 	struct scrub_ctx *sctx;
2998 
2999 	mutex_lock(&fs_info->scrub_lock);
3000 	sctx = dev->scrub_device;
3001 	if (!sctx) {
3002 		mutex_unlock(&fs_info->scrub_lock);
3003 		return -ENOTCONN;
3004 	}
3005 	atomic_inc(&sctx->cancel_req);
3006 	while (dev->scrub_device) {
3007 		mutex_unlock(&fs_info->scrub_lock);
3008 		wait_event(fs_info->scrub_pause_wait,
3009 			   dev->scrub_device == NULL);
3010 		mutex_lock(&fs_info->scrub_lock);
3011 	}
3012 	mutex_unlock(&fs_info->scrub_lock);
3013 
3014 	return 0;
3015 }
3016 
3017 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3018 			 struct btrfs_scrub_progress *progress)
3019 {
3020 	struct btrfs_device *dev;
3021 	struct scrub_ctx *sctx = NULL;
3022 
3023 	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3024 	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3025 	if (dev)
3026 		sctx = dev->scrub_device;
3027 	if (sctx)
3028 		memcpy(progress, &sctx->stat, sizeof(*progress));
3029 	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3030 
3031 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3032 }
3033 
3034 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3035 			       u64 extent_logical, u64 extent_len,
3036 			       u64 *extent_physical,
3037 			       struct btrfs_device **extent_dev,
3038 			       int *extent_mirror_num)
3039 {
3040 	u64 mapped_length;
3041 	struct btrfs_bio *bbio = NULL;
3042 	int ret;
3043 
3044 	mapped_length = extent_len;
3045 	ret = btrfs_map_block(fs_info, READ, extent_logical,
3046 			      &mapped_length, &bbio, 0);
3047 	if (ret || !bbio || mapped_length < extent_len ||
3048 	    !bbio->stripes[0].dev->bdev) {
3049 		kfree(bbio);
3050 		return;
3051 	}
3052 
3053 	*extent_physical = bbio->stripes[0].physical;
3054 	*extent_mirror_num = bbio->mirror_num;
3055 	*extent_dev = bbio->stripes[0].dev;
3056 	kfree(bbio);
3057 }
3058 
3059 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3060 			      struct scrub_wr_ctx *wr_ctx,
3061 			      struct btrfs_fs_info *fs_info,
3062 			      struct btrfs_device *dev,
3063 			      int is_dev_replace)
3064 {
3065 	WARN_ON(wr_ctx->wr_curr_bio != NULL);
3066 
3067 	mutex_init(&wr_ctx->wr_lock);
3068 	wr_ctx->wr_curr_bio = NULL;
3069 	if (!is_dev_replace)
3070 		return 0;
3071 
3072 	WARN_ON(!dev->bdev);
3073 	wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3074 					 bio_get_nr_vecs(dev->bdev));
3075 	wr_ctx->tgtdev = dev;
3076 	atomic_set(&wr_ctx->flush_all_writes, 0);
3077 	return 0;
3078 }
3079 
3080 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3081 {
3082 	mutex_lock(&wr_ctx->wr_lock);
3083 	kfree(wr_ctx->wr_curr_bio);
3084 	wr_ctx->wr_curr_bio = NULL;
3085 	mutex_unlock(&wr_ctx->wr_lock);
3086 }
3087 
3088 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3089 			    int mirror_num, u64 physical_for_dev_replace)
3090 {
3091 	struct scrub_copy_nocow_ctx *nocow_ctx;
3092 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3093 
3094 	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3095 	if (!nocow_ctx) {
3096 		spin_lock(&sctx->stat_lock);
3097 		sctx->stat.malloc_errors++;
3098 		spin_unlock(&sctx->stat_lock);
3099 		return -ENOMEM;
3100 	}
3101 
3102 	scrub_pending_trans_workers_inc(sctx);
3103 
3104 	nocow_ctx->sctx = sctx;
3105 	nocow_ctx->logical = logical;
3106 	nocow_ctx->len = len;
3107 	nocow_ctx->mirror_num = mirror_num;
3108 	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3109 	nocow_ctx->work.func = copy_nocow_pages_worker;
3110 	INIT_LIST_HEAD(&nocow_ctx->inodes);
3111 	btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3112 			   &nocow_ctx->work);
3113 
3114 	return 0;
3115 }
3116 
3117 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3118 {
3119 	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3120 	struct scrub_nocow_inode *nocow_inode;
3121 
3122 	nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3123 	if (!nocow_inode)
3124 		return -ENOMEM;
3125 	nocow_inode->inum = inum;
3126 	nocow_inode->offset = offset;
3127 	nocow_inode->root = root;
3128 	list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3129 	return 0;
3130 }
3131 
3132 #define COPY_COMPLETE 1
3133 
3134 static void copy_nocow_pages_worker(struct btrfs_work *work)
3135 {
3136 	struct scrub_copy_nocow_ctx *nocow_ctx =
3137 		container_of(work, struct scrub_copy_nocow_ctx, work);
3138 	struct scrub_ctx *sctx = nocow_ctx->sctx;
3139 	u64 logical = nocow_ctx->logical;
3140 	u64 len = nocow_ctx->len;
3141 	int mirror_num = nocow_ctx->mirror_num;
3142 	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3143 	int ret;
3144 	struct btrfs_trans_handle *trans = NULL;
3145 	struct btrfs_fs_info *fs_info;
3146 	struct btrfs_path *path;
3147 	struct btrfs_root *root;
3148 	int not_written = 0;
3149 
3150 	fs_info = sctx->dev_root->fs_info;
3151 	root = fs_info->extent_root;
3152 
3153 	path = btrfs_alloc_path();
3154 	if (!path) {
3155 		spin_lock(&sctx->stat_lock);
3156 		sctx->stat.malloc_errors++;
3157 		spin_unlock(&sctx->stat_lock);
3158 		not_written = 1;
3159 		goto out;
3160 	}
3161 
3162 	trans = btrfs_join_transaction(root);
3163 	if (IS_ERR(trans)) {
3164 		not_written = 1;
3165 		goto out;
3166 	}
3167 
3168 	ret = iterate_inodes_from_logical(logical, fs_info, path,
3169 					  record_inode_for_nocow, nocow_ctx);
3170 	if (ret != 0 && ret != -ENOENT) {
3171 		btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3172 			"phys %llu, len %llu, mir %u, ret %d",
3173 			logical, physical_for_dev_replace, len, mirror_num,
3174 			ret);
3175 		not_written = 1;
3176 		goto out;
3177 	}
3178 
3179 	btrfs_end_transaction(trans, root);
3180 	trans = NULL;
3181 	while (!list_empty(&nocow_ctx->inodes)) {
3182 		struct scrub_nocow_inode *entry;
3183 		entry = list_first_entry(&nocow_ctx->inodes,
3184 					 struct scrub_nocow_inode,
3185 					 list);
3186 		list_del_init(&entry->list);
3187 		ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3188 						 entry->root, nocow_ctx);
3189 		kfree(entry);
3190 		if (ret == COPY_COMPLETE) {
3191 			ret = 0;
3192 			break;
3193 		} else if (ret) {
3194 			break;
3195 		}
3196 	}
3197 out:
3198 	while (!list_empty(&nocow_ctx->inodes)) {
3199 		struct scrub_nocow_inode *entry;
3200 		entry = list_first_entry(&nocow_ctx->inodes,
3201 					 struct scrub_nocow_inode,
3202 					 list);
3203 		list_del_init(&entry->list);
3204 		kfree(entry);
3205 	}
3206 	if (trans && !IS_ERR(trans))
3207 		btrfs_end_transaction(trans, root);
3208 	if (not_written)
3209 		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3210 					    num_uncorrectable_read_errors);
3211 
3212 	btrfs_free_path(path);
3213 	kfree(nocow_ctx);
3214 
3215 	scrub_pending_trans_workers_dec(sctx);
3216 }
3217 
3218 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3219 				      struct scrub_copy_nocow_ctx *nocow_ctx)
3220 {
3221 	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3222 	struct btrfs_key key;
3223 	struct inode *inode;
3224 	struct page *page;
3225 	struct btrfs_root *local_root;
3226 	struct btrfs_ordered_extent *ordered;
3227 	struct extent_map *em;
3228 	struct extent_state *cached_state = NULL;
3229 	struct extent_io_tree *io_tree;
3230 	u64 physical_for_dev_replace;
3231 	u64 len = nocow_ctx->len;
3232 	u64 lockstart = offset, lockend = offset + len - 1;
3233 	unsigned long index;
3234 	int srcu_index;
3235 	int ret = 0;
3236 	int err = 0;
3237 
3238 	key.objectid = root;
3239 	key.type = BTRFS_ROOT_ITEM_KEY;
3240 	key.offset = (u64)-1;
3241 
3242 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3243 
3244 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3245 	if (IS_ERR(local_root)) {
3246 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3247 		return PTR_ERR(local_root);
3248 	}
3249 
3250 	key.type = BTRFS_INODE_ITEM_KEY;
3251 	key.objectid = inum;
3252 	key.offset = 0;
3253 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3254 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3255 	if (IS_ERR(inode))
3256 		return PTR_ERR(inode);
3257 
3258 	/* Avoid truncate/dio/punch hole.. */
3259 	mutex_lock(&inode->i_mutex);
3260 	inode_dio_wait(inode);
3261 
3262 	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3263 	io_tree = &BTRFS_I(inode)->io_tree;
3264 
3265 	lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3266 	ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3267 	if (ordered) {
3268 		btrfs_put_ordered_extent(ordered);
3269 		goto out_unlock;
3270 	}
3271 
3272 	em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3273 	if (IS_ERR(em)) {
3274 		ret = PTR_ERR(em);
3275 		goto out_unlock;
3276 	}
3277 
3278 	/*
3279 	 * This extent does not actually cover the logical extent anymore,
3280 	 * move on to the next inode.
3281 	 */
3282 	if (em->block_start > nocow_ctx->logical ||
3283 	    em->block_start + em->block_len < nocow_ctx->logical + len) {
3284 		free_extent_map(em);
3285 		goto out_unlock;
3286 	}
3287 	free_extent_map(em);
3288 
3289 	while (len >= PAGE_CACHE_SIZE) {
3290 		index = offset >> PAGE_CACHE_SHIFT;
3291 again:
3292 		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3293 		if (!page) {
3294 			btrfs_err(fs_info, "find_or_create_page() failed");
3295 			ret = -ENOMEM;
3296 			goto out;
3297 		}
3298 
3299 		if (PageUptodate(page)) {
3300 			if (PageDirty(page))
3301 				goto next_page;
3302 		} else {
3303 			ClearPageError(page);
3304 			err = extent_read_full_page_nolock(io_tree, page,
3305 							   btrfs_get_extent,
3306 							   nocow_ctx->mirror_num);
3307 			if (err) {
3308 				ret = err;
3309 				goto next_page;
3310 			}
3311 
3312 			lock_page(page);
3313 			/*
3314 			 * If the page has been remove from the page cache,
3315 			 * the data on it is meaningless, because it may be
3316 			 * old one, the new data may be written into the new
3317 			 * page in the page cache.
3318 			 */
3319 			if (page->mapping != inode->i_mapping) {
3320 				unlock_page(page);
3321 				page_cache_release(page);
3322 				goto again;
3323 			}
3324 			if (!PageUptodate(page)) {
3325 				ret = -EIO;
3326 				goto next_page;
3327 			}
3328 		}
3329 		err = write_page_nocow(nocow_ctx->sctx,
3330 				       physical_for_dev_replace, page);
3331 		if (err)
3332 			ret = err;
3333 next_page:
3334 		unlock_page(page);
3335 		page_cache_release(page);
3336 
3337 		if (ret)
3338 			break;
3339 
3340 		offset += PAGE_CACHE_SIZE;
3341 		physical_for_dev_replace += PAGE_CACHE_SIZE;
3342 		len -= PAGE_CACHE_SIZE;
3343 	}
3344 	ret = COPY_COMPLETE;
3345 out_unlock:
3346 	unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3347 			     GFP_NOFS);
3348 out:
3349 	mutex_unlock(&inode->i_mutex);
3350 	iput(inode);
3351 	return ret;
3352 }
3353 
3354 static int write_page_nocow(struct scrub_ctx *sctx,
3355 			    u64 physical_for_dev_replace, struct page *page)
3356 {
3357 	struct bio *bio;
3358 	struct btrfs_device *dev;
3359 	int ret;
3360 
3361 	dev = sctx->wr_ctx.tgtdev;
3362 	if (!dev)
3363 		return -EIO;
3364 	if (!dev->bdev) {
3365 		printk_ratelimited(KERN_WARNING
3366 			"BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3367 		return -EIO;
3368 	}
3369 	bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3370 	if (!bio) {
3371 		spin_lock(&sctx->stat_lock);
3372 		sctx->stat.malloc_errors++;
3373 		spin_unlock(&sctx->stat_lock);
3374 		return -ENOMEM;
3375 	}
3376 	bio->bi_iter.bi_size = 0;
3377 	bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3378 	bio->bi_bdev = dev->bdev;
3379 	ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3380 	if (ret != PAGE_CACHE_SIZE) {
3381 leave_with_eio:
3382 		bio_put(bio);
3383 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3384 		return -EIO;
3385 	}
3386 
3387 	if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3388 		goto leave_with_eio;
3389 
3390 	bio_put(bio);
3391 	return 0;
3392 }
3393