xref: /linux/fs/btrfs/scrub.c (revision 861e10be08c69808065d755d3e3cab5d520a2d32)
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 
32 /*
33  * This is only the first step towards a full-features scrub. It reads all
34  * extent and super block and verifies the checksums. In case a bad checksum
35  * is found or the extent cannot be read, good data will be written back if
36  * any can be found.
37  *
38  * Future enhancements:
39  *  - In case an unrepairable extent is encountered, track which files are
40  *    affected and report them
41  *  - track and record media errors, throw out bad devices
42  *  - add a mode to also read unallocated space
43  */
44 
45 struct scrub_block;
46 struct scrub_ctx;
47 
48 /*
49  * the following three values only influence the performance.
50  * The last one configures the number of parallel and outstanding I/O
51  * operations. The first two values configure an upper limit for the number
52  * of (dynamically allocated) pages that are added to a bio.
53  */
54 #define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
55 #define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
56 #define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
57 
58 /*
59  * the following value times PAGE_SIZE needs to be large enough to match the
60  * largest node/leaf/sector size that shall be supported.
61  * Values larger than BTRFS_STRIPE_LEN are not supported.
62  */
63 #define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
64 
65 struct scrub_page {
66 	struct scrub_block	*sblock;
67 	struct page		*page;
68 	struct btrfs_device	*dev;
69 	u64			flags;  /* extent flags */
70 	u64			generation;
71 	u64			logical;
72 	u64			physical;
73 	u64			physical_for_dev_replace;
74 	atomic_t		ref_count;
75 	struct {
76 		unsigned int	mirror_num:8;
77 		unsigned int	have_csum:1;
78 		unsigned int	io_error:1;
79 	};
80 	u8			csum[BTRFS_CSUM_SIZE];
81 };
82 
83 struct scrub_bio {
84 	int			index;
85 	struct scrub_ctx	*sctx;
86 	struct btrfs_device	*dev;
87 	struct bio		*bio;
88 	int			err;
89 	u64			logical;
90 	u64			physical;
91 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
92 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
93 #else
94 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
95 #endif
96 	int			page_count;
97 	int			next_free;
98 	struct btrfs_work	work;
99 };
100 
101 struct scrub_block {
102 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
103 	int			page_count;
104 	atomic_t		outstanding_pages;
105 	atomic_t		ref_count; /* free mem on transition to zero */
106 	struct scrub_ctx	*sctx;
107 	struct {
108 		unsigned int	header_error:1;
109 		unsigned int	checksum_error:1;
110 		unsigned int	no_io_error_seen:1;
111 		unsigned int	generation_error:1; /* also sets header_error */
112 	};
113 };
114 
115 struct scrub_wr_ctx {
116 	struct scrub_bio *wr_curr_bio;
117 	struct btrfs_device *tgtdev;
118 	int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
119 	atomic_t flush_all_writes;
120 	struct mutex wr_lock;
121 };
122 
123 struct scrub_ctx {
124 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
125 	struct btrfs_root	*dev_root;
126 	int			first_free;
127 	int			curr;
128 	atomic_t		bios_in_flight;
129 	atomic_t		workers_pending;
130 	spinlock_t		list_lock;
131 	wait_queue_head_t	list_wait;
132 	u16			csum_size;
133 	struct list_head	csum_list;
134 	atomic_t		cancel_req;
135 	int			readonly;
136 	int			pages_per_rd_bio;
137 	u32			sectorsize;
138 	u32			nodesize;
139 	u32			leafsize;
140 
141 	int			is_dev_replace;
142 	struct scrub_wr_ctx	wr_ctx;
143 
144 	/*
145 	 * statistics
146 	 */
147 	struct btrfs_scrub_progress stat;
148 	spinlock_t		stat_lock;
149 };
150 
151 struct scrub_fixup_nodatasum {
152 	struct scrub_ctx	*sctx;
153 	struct btrfs_device	*dev;
154 	u64			logical;
155 	struct btrfs_root	*root;
156 	struct btrfs_work	work;
157 	int			mirror_num;
158 };
159 
160 struct scrub_copy_nocow_ctx {
161 	struct scrub_ctx	*sctx;
162 	u64			logical;
163 	u64			len;
164 	int			mirror_num;
165 	u64			physical_for_dev_replace;
166 	struct btrfs_work	work;
167 };
168 
169 struct scrub_warning {
170 	struct btrfs_path	*path;
171 	u64			extent_item_size;
172 	char			*scratch_buf;
173 	char			*msg_buf;
174 	const char		*errstr;
175 	sector_t		sector;
176 	u64			logical;
177 	struct btrfs_device	*dev;
178 	int			msg_bufsize;
179 	int			scratch_bufsize;
180 };
181 
182 
183 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
184 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
185 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
186 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
187 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
188 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
189 				     struct btrfs_fs_info *fs_info,
190 				     struct scrub_block *original_sblock,
191 				     u64 length, u64 logical,
192 				     struct scrub_block *sblocks_for_recheck);
193 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
194 				struct scrub_block *sblock, int is_metadata,
195 				int have_csum, u8 *csum, u64 generation,
196 				u16 csum_size);
197 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
198 					 struct scrub_block *sblock,
199 					 int is_metadata, int have_csum,
200 					 const u8 *csum, u64 generation,
201 					 u16 csum_size);
202 static void scrub_complete_bio_end_io(struct bio *bio, int err);
203 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
204 					     struct scrub_block *sblock_good,
205 					     int force_write);
206 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
207 					    struct scrub_block *sblock_good,
208 					    int page_num, int force_write);
209 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
210 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
211 					   int page_num);
212 static int scrub_checksum_data(struct scrub_block *sblock);
213 static int scrub_checksum_tree_block(struct scrub_block *sblock);
214 static int scrub_checksum_super(struct scrub_block *sblock);
215 static void scrub_block_get(struct scrub_block *sblock);
216 static void scrub_block_put(struct scrub_block *sblock);
217 static void scrub_page_get(struct scrub_page *spage);
218 static void scrub_page_put(struct scrub_page *spage);
219 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
220 				    struct scrub_page *spage);
221 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
222 		       u64 physical, struct btrfs_device *dev, u64 flags,
223 		       u64 gen, int mirror_num, u8 *csum, int force,
224 		       u64 physical_for_dev_replace);
225 static void scrub_bio_end_io(struct bio *bio, int err);
226 static void scrub_bio_end_io_worker(struct btrfs_work *work);
227 static void scrub_block_complete(struct scrub_block *sblock);
228 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
229 			       u64 extent_logical, u64 extent_len,
230 			       u64 *extent_physical,
231 			       struct btrfs_device **extent_dev,
232 			       int *extent_mirror_num);
233 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
234 			      struct scrub_wr_ctx *wr_ctx,
235 			      struct btrfs_fs_info *fs_info,
236 			      struct btrfs_device *dev,
237 			      int is_dev_replace);
238 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
239 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
240 				    struct scrub_page *spage);
241 static void scrub_wr_submit(struct scrub_ctx *sctx);
242 static void scrub_wr_bio_end_io(struct bio *bio, int err);
243 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
244 static int write_page_nocow(struct scrub_ctx *sctx,
245 			    u64 physical_for_dev_replace, struct page *page);
246 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
247 				      void *ctx);
248 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
249 			    int mirror_num, u64 physical_for_dev_replace);
250 static void copy_nocow_pages_worker(struct btrfs_work *work);
251 
252 
253 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
254 {
255 	atomic_inc(&sctx->bios_in_flight);
256 }
257 
258 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
259 {
260 	atomic_dec(&sctx->bios_in_flight);
261 	wake_up(&sctx->list_wait);
262 }
263 
264 /*
265  * used for workers that require transaction commits (i.e., for the
266  * NOCOW case)
267  */
268 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
269 {
270 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
271 
272 	/*
273 	 * increment scrubs_running to prevent cancel requests from
274 	 * completing as long as a worker is running. we must also
275 	 * increment scrubs_paused to prevent deadlocking on pause
276 	 * requests used for transactions commits (as the worker uses a
277 	 * transaction context). it is safe to regard the worker
278 	 * as paused for all matters practical. effectively, we only
279 	 * avoid cancellation requests from completing.
280 	 */
281 	mutex_lock(&fs_info->scrub_lock);
282 	atomic_inc(&fs_info->scrubs_running);
283 	atomic_inc(&fs_info->scrubs_paused);
284 	mutex_unlock(&fs_info->scrub_lock);
285 	atomic_inc(&sctx->workers_pending);
286 }
287 
288 /* used for workers that require transaction commits */
289 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
290 {
291 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
292 
293 	/*
294 	 * see scrub_pending_trans_workers_inc() why we're pretending
295 	 * to be paused in the scrub counters
296 	 */
297 	mutex_lock(&fs_info->scrub_lock);
298 	atomic_dec(&fs_info->scrubs_running);
299 	atomic_dec(&fs_info->scrubs_paused);
300 	mutex_unlock(&fs_info->scrub_lock);
301 	atomic_dec(&sctx->workers_pending);
302 	wake_up(&fs_info->scrub_pause_wait);
303 	wake_up(&sctx->list_wait);
304 }
305 
306 static void scrub_free_csums(struct scrub_ctx *sctx)
307 {
308 	while (!list_empty(&sctx->csum_list)) {
309 		struct btrfs_ordered_sum *sum;
310 		sum = list_first_entry(&sctx->csum_list,
311 				       struct btrfs_ordered_sum, list);
312 		list_del(&sum->list);
313 		kfree(sum);
314 	}
315 }
316 
317 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
318 {
319 	int i;
320 
321 	if (!sctx)
322 		return;
323 
324 	scrub_free_wr_ctx(&sctx->wr_ctx);
325 
326 	/* this can happen when scrub is cancelled */
327 	if (sctx->curr != -1) {
328 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
329 
330 		for (i = 0; i < sbio->page_count; i++) {
331 			WARN_ON(!sbio->pagev[i]->page);
332 			scrub_block_put(sbio->pagev[i]->sblock);
333 		}
334 		bio_put(sbio->bio);
335 	}
336 
337 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
338 		struct scrub_bio *sbio = sctx->bios[i];
339 
340 		if (!sbio)
341 			break;
342 		kfree(sbio);
343 	}
344 
345 	scrub_free_csums(sctx);
346 	kfree(sctx);
347 }
348 
349 static noinline_for_stack
350 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
351 {
352 	struct scrub_ctx *sctx;
353 	int		i;
354 	struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
355 	int pages_per_rd_bio;
356 	int ret;
357 
358 	/*
359 	 * the setting of pages_per_rd_bio is correct for scrub but might
360 	 * be wrong for the dev_replace code where we might read from
361 	 * different devices in the initial huge bios. However, that
362 	 * code is able to correctly handle the case when adding a page
363 	 * to a bio fails.
364 	 */
365 	if (dev->bdev)
366 		pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
367 					 bio_get_nr_vecs(dev->bdev));
368 	else
369 		pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
370 	sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
371 	if (!sctx)
372 		goto nomem;
373 	sctx->is_dev_replace = is_dev_replace;
374 	sctx->pages_per_rd_bio = pages_per_rd_bio;
375 	sctx->curr = -1;
376 	sctx->dev_root = dev->dev_root;
377 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
378 		struct scrub_bio *sbio;
379 
380 		sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
381 		if (!sbio)
382 			goto nomem;
383 		sctx->bios[i] = sbio;
384 
385 		sbio->index = i;
386 		sbio->sctx = sctx;
387 		sbio->page_count = 0;
388 		sbio->work.func = scrub_bio_end_io_worker;
389 
390 		if (i != SCRUB_BIOS_PER_SCTX - 1)
391 			sctx->bios[i]->next_free = i + 1;
392 		else
393 			sctx->bios[i]->next_free = -1;
394 	}
395 	sctx->first_free = 0;
396 	sctx->nodesize = dev->dev_root->nodesize;
397 	sctx->leafsize = dev->dev_root->leafsize;
398 	sctx->sectorsize = dev->dev_root->sectorsize;
399 	atomic_set(&sctx->bios_in_flight, 0);
400 	atomic_set(&sctx->workers_pending, 0);
401 	atomic_set(&sctx->cancel_req, 0);
402 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
403 	INIT_LIST_HEAD(&sctx->csum_list);
404 
405 	spin_lock_init(&sctx->list_lock);
406 	spin_lock_init(&sctx->stat_lock);
407 	init_waitqueue_head(&sctx->list_wait);
408 
409 	ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
410 				 fs_info->dev_replace.tgtdev, is_dev_replace);
411 	if (ret) {
412 		scrub_free_ctx(sctx);
413 		return ERR_PTR(ret);
414 	}
415 	return sctx;
416 
417 nomem:
418 	scrub_free_ctx(sctx);
419 	return ERR_PTR(-ENOMEM);
420 }
421 
422 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
423 				     void *warn_ctx)
424 {
425 	u64 isize;
426 	u32 nlink;
427 	int ret;
428 	int i;
429 	struct extent_buffer *eb;
430 	struct btrfs_inode_item *inode_item;
431 	struct scrub_warning *swarn = warn_ctx;
432 	struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
433 	struct inode_fs_paths *ipath = NULL;
434 	struct btrfs_root *local_root;
435 	struct btrfs_key root_key;
436 
437 	root_key.objectid = root;
438 	root_key.type = BTRFS_ROOT_ITEM_KEY;
439 	root_key.offset = (u64)-1;
440 	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
441 	if (IS_ERR(local_root)) {
442 		ret = PTR_ERR(local_root);
443 		goto err;
444 	}
445 
446 	ret = inode_item_info(inum, 0, local_root, swarn->path);
447 	if (ret) {
448 		btrfs_release_path(swarn->path);
449 		goto err;
450 	}
451 
452 	eb = swarn->path->nodes[0];
453 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
454 					struct btrfs_inode_item);
455 	isize = btrfs_inode_size(eb, inode_item);
456 	nlink = btrfs_inode_nlink(eb, inode_item);
457 	btrfs_release_path(swarn->path);
458 
459 	ipath = init_ipath(4096, local_root, swarn->path);
460 	if (IS_ERR(ipath)) {
461 		ret = PTR_ERR(ipath);
462 		ipath = NULL;
463 		goto err;
464 	}
465 	ret = paths_from_inode(inum, ipath);
466 
467 	if (ret < 0)
468 		goto err;
469 
470 	/*
471 	 * we deliberately ignore the bit ipath might have been too small to
472 	 * hold all of the paths here
473 	 */
474 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
475 		printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
476 			"%s, sector %llu, root %llu, inode %llu, offset %llu, "
477 			"length %llu, links %u (path: %s)\n", swarn->errstr,
478 			swarn->logical, rcu_str_deref(swarn->dev->name),
479 			(unsigned long long)swarn->sector, root, inum, offset,
480 			min(isize - offset, (u64)PAGE_SIZE), nlink,
481 			(char *)(unsigned long)ipath->fspath->val[i]);
482 
483 	free_ipath(ipath);
484 	return 0;
485 
486 err:
487 	printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
488 		"%s, sector %llu, root %llu, inode %llu, offset %llu: path "
489 		"resolving failed with ret=%d\n", swarn->errstr,
490 		swarn->logical, rcu_str_deref(swarn->dev->name),
491 		(unsigned long long)swarn->sector, root, inum, offset, ret);
492 
493 	free_ipath(ipath);
494 	return 0;
495 }
496 
497 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
498 {
499 	struct btrfs_device *dev;
500 	struct btrfs_fs_info *fs_info;
501 	struct btrfs_path *path;
502 	struct btrfs_key found_key;
503 	struct extent_buffer *eb;
504 	struct btrfs_extent_item *ei;
505 	struct scrub_warning swarn;
506 	unsigned long ptr = 0;
507 	u64 extent_item_pos;
508 	u64 flags = 0;
509 	u64 ref_root;
510 	u32 item_size;
511 	u8 ref_level;
512 	const int bufsize = 4096;
513 	int ret;
514 
515 	WARN_ON(sblock->page_count < 1);
516 	dev = sblock->pagev[0]->dev;
517 	fs_info = sblock->sctx->dev_root->fs_info;
518 
519 	path = btrfs_alloc_path();
520 
521 	swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
522 	swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
523 	swarn.sector = (sblock->pagev[0]->physical) >> 9;
524 	swarn.logical = sblock->pagev[0]->logical;
525 	swarn.errstr = errstr;
526 	swarn.dev = NULL;
527 	swarn.msg_bufsize = bufsize;
528 	swarn.scratch_bufsize = bufsize;
529 
530 	if (!path || !swarn.scratch_buf || !swarn.msg_buf)
531 		goto out;
532 
533 	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
534 				  &flags);
535 	if (ret < 0)
536 		goto out;
537 
538 	extent_item_pos = swarn.logical - found_key.objectid;
539 	swarn.extent_item_size = found_key.offset;
540 
541 	eb = path->nodes[0];
542 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
543 	item_size = btrfs_item_size_nr(eb, path->slots[0]);
544 	btrfs_release_path(path);
545 
546 	if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
547 		do {
548 			ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
549 							&ref_root, &ref_level);
550 			printk_in_rcu(KERN_WARNING
551 				"btrfs: %s at logical %llu on dev %s, "
552 				"sector %llu: metadata %s (level %d) in tree "
553 				"%llu\n", errstr, swarn.logical,
554 				rcu_str_deref(dev->name),
555 				(unsigned long long)swarn.sector,
556 				ref_level ? "node" : "leaf",
557 				ret < 0 ? -1 : ref_level,
558 				ret < 0 ? -1 : ref_root);
559 		} while (ret != 1);
560 	} else {
561 		swarn.path = path;
562 		swarn.dev = dev;
563 		iterate_extent_inodes(fs_info, found_key.objectid,
564 					extent_item_pos, 1,
565 					scrub_print_warning_inode, &swarn);
566 	}
567 
568 out:
569 	btrfs_free_path(path);
570 	kfree(swarn.scratch_buf);
571 	kfree(swarn.msg_buf);
572 }
573 
574 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
575 {
576 	struct page *page = NULL;
577 	unsigned long index;
578 	struct scrub_fixup_nodatasum *fixup = fixup_ctx;
579 	int ret;
580 	int corrected = 0;
581 	struct btrfs_key key;
582 	struct inode *inode = NULL;
583 	struct btrfs_fs_info *fs_info;
584 	u64 end = offset + PAGE_SIZE - 1;
585 	struct btrfs_root *local_root;
586 	int srcu_index;
587 
588 	key.objectid = root;
589 	key.type = BTRFS_ROOT_ITEM_KEY;
590 	key.offset = (u64)-1;
591 
592 	fs_info = fixup->root->fs_info;
593 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
594 
595 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
596 	if (IS_ERR(local_root)) {
597 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
598 		return PTR_ERR(local_root);
599 	}
600 
601 	key.type = BTRFS_INODE_ITEM_KEY;
602 	key.objectid = inum;
603 	key.offset = 0;
604 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
605 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
606 	if (IS_ERR(inode))
607 		return PTR_ERR(inode);
608 
609 	index = offset >> PAGE_CACHE_SHIFT;
610 
611 	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
612 	if (!page) {
613 		ret = -ENOMEM;
614 		goto out;
615 	}
616 
617 	if (PageUptodate(page)) {
618 		if (PageDirty(page)) {
619 			/*
620 			 * we need to write the data to the defect sector. the
621 			 * data that was in that sector is not in memory,
622 			 * because the page was modified. we must not write the
623 			 * modified page to that sector.
624 			 *
625 			 * TODO: what could be done here: wait for the delalloc
626 			 *       runner to write out that page (might involve
627 			 *       COW) and see whether the sector is still
628 			 *       referenced afterwards.
629 			 *
630 			 * For the meantime, we'll treat this error
631 			 * incorrectable, although there is a chance that a
632 			 * later scrub will find the bad sector again and that
633 			 * there's no dirty page in memory, then.
634 			 */
635 			ret = -EIO;
636 			goto out;
637 		}
638 		fs_info = BTRFS_I(inode)->root->fs_info;
639 		ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
640 					fixup->logical, page,
641 					fixup->mirror_num);
642 		unlock_page(page);
643 		corrected = !ret;
644 	} else {
645 		/*
646 		 * we need to get good data first. the general readpage path
647 		 * will call repair_io_failure for us, we just have to make
648 		 * sure we read the bad mirror.
649 		 */
650 		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
651 					EXTENT_DAMAGED, GFP_NOFS);
652 		if (ret) {
653 			/* set_extent_bits should give proper error */
654 			WARN_ON(ret > 0);
655 			if (ret > 0)
656 				ret = -EFAULT;
657 			goto out;
658 		}
659 
660 		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
661 						btrfs_get_extent,
662 						fixup->mirror_num);
663 		wait_on_page_locked(page);
664 
665 		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
666 						end, EXTENT_DAMAGED, 0, NULL);
667 		if (!corrected)
668 			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
669 						EXTENT_DAMAGED, GFP_NOFS);
670 	}
671 
672 out:
673 	if (page)
674 		put_page(page);
675 	if (inode)
676 		iput(inode);
677 
678 	if (ret < 0)
679 		return ret;
680 
681 	if (ret == 0 && corrected) {
682 		/*
683 		 * we only need to call readpage for one of the inodes belonging
684 		 * to this extent. so make iterate_extent_inodes stop
685 		 */
686 		return 1;
687 	}
688 
689 	return -EIO;
690 }
691 
692 static void scrub_fixup_nodatasum(struct btrfs_work *work)
693 {
694 	int ret;
695 	struct scrub_fixup_nodatasum *fixup;
696 	struct scrub_ctx *sctx;
697 	struct btrfs_trans_handle *trans = NULL;
698 	struct btrfs_fs_info *fs_info;
699 	struct btrfs_path *path;
700 	int uncorrectable = 0;
701 
702 	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
703 	sctx = fixup->sctx;
704 	fs_info = fixup->root->fs_info;
705 
706 	path = btrfs_alloc_path();
707 	if (!path) {
708 		spin_lock(&sctx->stat_lock);
709 		++sctx->stat.malloc_errors;
710 		spin_unlock(&sctx->stat_lock);
711 		uncorrectable = 1;
712 		goto out;
713 	}
714 
715 	trans = btrfs_join_transaction(fixup->root);
716 	if (IS_ERR(trans)) {
717 		uncorrectable = 1;
718 		goto out;
719 	}
720 
721 	/*
722 	 * the idea is to trigger a regular read through the standard path. we
723 	 * read a page from the (failed) logical address by specifying the
724 	 * corresponding copynum of the failed sector. thus, that readpage is
725 	 * expected to fail.
726 	 * that is the point where on-the-fly error correction will kick in
727 	 * (once it's finished) and rewrite the failed sector if a good copy
728 	 * can be found.
729 	 */
730 	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
731 						path, scrub_fixup_readpage,
732 						fixup);
733 	if (ret < 0) {
734 		uncorrectable = 1;
735 		goto out;
736 	}
737 	WARN_ON(ret != 1);
738 
739 	spin_lock(&sctx->stat_lock);
740 	++sctx->stat.corrected_errors;
741 	spin_unlock(&sctx->stat_lock);
742 
743 out:
744 	if (trans && !IS_ERR(trans))
745 		btrfs_end_transaction(trans, fixup->root);
746 	if (uncorrectable) {
747 		spin_lock(&sctx->stat_lock);
748 		++sctx->stat.uncorrectable_errors;
749 		spin_unlock(&sctx->stat_lock);
750 		btrfs_dev_replace_stats_inc(
751 			&sctx->dev_root->fs_info->dev_replace.
752 			num_uncorrectable_read_errors);
753 		printk_ratelimited_in_rcu(KERN_ERR
754 			"btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
755 			(unsigned long long)fixup->logical,
756 			rcu_str_deref(fixup->dev->name));
757 	}
758 
759 	btrfs_free_path(path);
760 	kfree(fixup);
761 
762 	scrub_pending_trans_workers_dec(sctx);
763 }
764 
765 /*
766  * scrub_handle_errored_block gets called when either verification of the
767  * pages failed or the bio failed to read, e.g. with EIO. In the latter
768  * case, this function handles all pages in the bio, even though only one
769  * may be bad.
770  * The goal of this function is to repair the errored block by using the
771  * contents of one of the mirrors.
772  */
773 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
774 {
775 	struct scrub_ctx *sctx = sblock_to_check->sctx;
776 	struct btrfs_device *dev;
777 	struct btrfs_fs_info *fs_info;
778 	u64 length;
779 	u64 logical;
780 	u64 generation;
781 	unsigned int failed_mirror_index;
782 	unsigned int is_metadata;
783 	unsigned int have_csum;
784 	u8 *csum;
785 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
786 	struct scrub_block *sblock_bad;
787 	int ret;
788 	int mirror_index;
789 	int page_num;
790 	int success;
791 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
792 				      DEFAULT_RATELIMIT_BURST);
793 
794 	BUG_ON(sblock_to_check->page_count < 1);
795 	fs_info = sctx->dev_root->fs_info;
796 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
797 		/*
798 		 * if we find an error in a super block, we just report it.
799 		 * They will get written with the next transaction commit
800 		 * anyway
801 		 */
802 		spin_lock(&sctx->stat_lock);
803 		++sctx->stat.super_errors;
804 		spin_unlock(&sctx->stat_lock);
805 		return 0;
806 	}
807 	length = sblock_to_check->page_count * PAGE_SIZE;
808 	logical = sblock_to_check->pagev[0]->logical;
809 	generation = sblock_to_check->pagev[0]->generation;
810 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
811 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
812 	is_metadata = !(sblock_to_check->pagev[0]->flags &
813 			BTRFS_EXTENT_FLAG_DATA);
814 	have_csum = sblock_to_check->pagev[0]->have_csum;
815 	csum = sblock_to_check->pagev[0]->csum;
816 	dev = sblock_to_check->pagev[0]->dev;
817 
818 	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
819 		sblocks_for_recheck = NULL;
820 		goto nodatasum_case;
821 	}
822 
823 	/*
824 	 * read all mirrors one after the other. This includes to
825 	 * re-read the extent or metadata block that failed (that was
826 	 * the cause that this fixup code is called) another time,
827 	 * page by page this time in order to know which pages
828 	 * caused I/O errors and which ones are good (for all mirrors).
829 	 * It is the goal to handle the situation when more than one
830 	 * mirror contains I/O errors, but the errors do not
831 	 * overlap, i.e. the data can be repaired by selecting the
832 	 * pages from those mirrors without I/O error on the
833 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
834 	 * would be that mirror #1 has an I/O error on the first page,
835 	 * the second page is good, and mirror #2 has an I/O error on
836 	 * the second page, but the first page is good.
837 	 * Then the first page of the first mirror can be repaired by
838 	 * taking the first page of the second mirror, and the
839 	 * second page of the second mirror can be repaired by
840 	 * copying the contents of the 2nd page of the 1st mirror.
841 	 * One more note: if the pages of one mirror contain I/O
842 	 * errors, the checksum cannot be verified. In order to get
843 	 * the best data for repairing, the first attempt is to find
844 	 * a mirror without I/O errors and with a validated checksum.
845 	 * Only if this is not possible, the pages are picked from
846 	 * mirrors with I/O errors without considering the checksum.
847 	 * If the latter is the case, at the end, the checksum of the
848 	 * repaired area is verified in order to correctly maintain
849 	 * the statistics.
850 	 */
851 
852 	sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
853 				     sizeof(*sblocks_for_recheck),
854 				     GFP_NOFS);
855 	if (!sblocks_for_recheck) {
856 		spin_lock(&sctx->stat_lock);
857 		sctx->stat.malloc_errors++;
858 		sctx->stat.read_errors++;
859 		sctx->stat.uncorrectable_errors++;
860 		spin_unlock(&sctx->stat_lock);
861 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
862 		goto out;
863 	}
864 
865 	/* setup the context, map the logical blocks and alloc the pages */
866 	ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
867 					logical, sblocks_for_recheck);
868 	if (ret) {
869 		spin_lock(&sctx->stat_lock);
870 		sctx->stat.read_errors++;
871 		sctx->stat.uncorrectable_errors++;
872 		spin_unlock(&sctx->stat_lock);
873 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
874 		goto out;
875 	}
876 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
877 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
878 
879 	/* build and submit the bios for the failed mirror, check checksums */
880 	scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
881 			    csum, generation, sctx->csum_size);
882 
883 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
884 	    sblock_bad->no_io_error_seen) {
885 		/*
886 		 * the error disappeared after reading page by page, or
887 		 * the area was part of a huge bio and other parts of the
888 		 * bio caused I/O errors, or the block layer merged several
889 		 * read requests into one and the error is caused by a
890 		 * different bio (usually one of the two latter cases is
891 		 * the cause)
892 		 */
893 		spin_lock(&sctx->stat_lock);
894 		sctx->stat.unverified_errors++;
895 		spin_unlock(&sctx->stat_lock);
896 
897 		if (sctx->is_dev_replace)
898 			scrub_write_block_to_dev_replace(sblock_bad);
899 		goto out;
900 	}
901 
902 	if (!sblock_bad->no_io_error_seen) {
903 		spin_lock(&sctx->stat_lock);
904 		sctx->stat.read_errors++;
905 		spin_unlock(&sctx->stat_lock);
906 		if (__ratelimit(&_rs))
907 			scrub_print_warning("i/o error", sblock_to_check);
908 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
909 	} else if (sblock_bad->checksum_error) {
910 		spin_lock(&sctx->stat_lock);
911 		sctx->stat.csum_errors++;
912 		spin_unlock(&sctx->stat_lock);
913 		if (__ratelimit(&_rs))
914 			scrub_print_warning("checksum error", sblock_to_check);
915 		btrfs_dev_stat_inc_and_print(dev,
916 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
917 	} else if (sblock_bad->header_error) {
918 		spin_lock(&sctx->stat_lock);
919 		sctx->stat.verify_errors++;
920 		spin_unlock(&sctx->stat_lock);
921 		if (__ratelimit(&_rs))
922 			scrub_print_warning("checksum/header error",
923 					    sblock_to_check);
924 		if (sblock_bad->generation_error)
925 			btrfs_dev_stat_inc_and_print(dev,
926 				BTRFS_DEV_STAT_GENERATION_ERRS);
927 		else
928 			btrfs_dev_stat_inc_and_print(dev,
929 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
930 	}
931 
932 	if (sctx->readonly && !sctx->is_dev_replace)
933 		goto did_not_correct_error;
934 
935 	if (!is_metadata && !have_csum) {
936 		struct scrub_fixup_nodatasum *fixup_nodatasum;
937 
938 nodatasum_case:
939 		WARN_ON(sctx->is_dev_replace);
940 
941 		/*
942 		 * !is_metadata and !have_csum, this means that the data
943 		 * might not be COW'ed, that it might be modified
944 		 * concurrently. The general strategy to work on the
945 		 * commit root does not help in the case when COW is not
946 		 * used.
947 		 */
948 		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
949 		if (!fixup_nodatasum)
950 			goto did_not_correct_error;
951 		fixup_nodatasum->sctx = sctx;
952 		fixup_nodatasum->dev = dev;
953 		fixup_nodatasum->logical = logical;
954 		fixup_nodatasum->root = fs_info->extent_root;
955 		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
956 		scrub_pending_trans_workers_inc(sctx);
957 		fixup_nodatasum->work.func = scrub_fixup_nodatasum;
958 		btrfs_queue_worker(&fs_info->scrub_workers,
959 				   &fixup_nodatasum->work);
960 		goto out;
961 	}
962 
963 	/*
964 	 * now build and submit the bios for the other mirrors, check
965 	 * checksums.
966 	 * First try to pick the mirror which is completely without I/O
967 	 * errors and also does not have a checksum error.
968 	 * If one is found, and if a checksum is present, the full block
969 	 * that is known to contain an error is rewritten. Afterwards
970 	 * the block is known to be corrected.
971 	 * If a mirror is found which is completely correct, and no
972 	 * checksum is present, only those pages are rewritten that had
973 	 * an I/O error in the block to be repaired, since it cannot be
974 	 * determined, which copy of the other pages is better (and it
975 	 * could happen otherwise that a correct page would be
976 	 * overwritten by a bad one).
977 	 */
978 	for (mirror_index = 0;
979 	     mirror_index < BTRFS_MAX_MIRRORS &&
980 	     sblocks_for_recheck[mirror_index].page_count > 0;
981 	     mirror_index++) {
982 		struct scrub_block *sblock_other;
983 
984 		if (mirror_index == failed_mirror_index)
985 			continue;
986 		sblock_other = sblocks_for_recheck + mirror_index;
987 
988 		/* build and submit the bios, check checksums */
989 		scrub_recheck_block(fs_info, sblock_other, is_metadata,
990 				    have_csum, csum, generation,
991 				    sctx->csum_size);
992 
993 		if (!sblock_other->header_error &&
994 		    !sblock_other->checksum_error &&
995 		    sblock_other->no_io_error_seen) {
996 			if (sctx->is_dev_replace) {
997 				scrub_write_block_to_dev_replace(sblock_other);
998 			} else {
999 				int force_write = is_metadata || have_csum;
1000 
1001 				ret = scrub_repair_block_from_good_copy(
1002 						sblock_bad, sblock_other,
1003 						force_write);
1004 			}
1005 			if (0 == ret)
1006 				goto corrected_error;
1007 		}
1008 	}
1009 
1010 	/*
1011 	 * for dev_replace, pick good pages and write to the target device.
1012 	 */
1013 	if (sctx->is_dev_replace) {
1014 		success = 1;
1015 		for (page_num = 0; page_num < sblock_bad->page_count;
1016 		     page_num++) {
1017 			int sub_success;
1018 
1019 			sub_success = 0;
1020 			for (mirror_index = 0;
1021 			     mirror_index < BTRFS_MAX_MIRRORS &&
1022 			     sblocks_for_recheck[mirror_index].page_count > 0;
1023 			     mirror_index++) {
1024 				struct scrub_block *sblock_other =
1025 					sblocks_for_recheck + mirror_index;
1026 				struct scrub_page *page_other =
1027 					sblock_other->pagev[page_num];
1028 
1029 				if (!page_other->io_error) {
1030 					ret = scrub_write_page_to_dev_replace(
1031 							sblock_other, page_num);
1032 					if (ret == 0) {
1033 						/* succeeded for this page */
1034 						sub_success = 1;
1035 						break;
1036 					} else {
1037 						btrfs_dev_replace_stats_inc(
1038 							&sctx->dev_root->
1039 							fs_info->dev_replace.
1040 							num_write_errors);
1041 					}
1042 				}
1043 			}
1044 
1045 			if (!sub_success) {
1046 				/*
1047 				 * did not find a mirror to fetch the page
1048 				 * from. scrub_write_page_to_dev_replace()
1049 				 * handles this case (page->io_error), by
1050 				 * filling the block with zeros before
1051 				 * submitting the write request
1052 				 */
1053 				success = 0;
1054 				ret = scrub_write_page_to_dev_replace(
1055 						sblock_bad, page_num);
1056 				if (ret)
1057 					btrfs_dev_replace_stats_inc(
1058 						&sctx->dev_root->fs_info->
1059 						dev_replace.num_write_errors);
1060 			}
1061 		}
1062 
1063 		goto out;
1064 	}
1065 
1066 	/*
1067 	 * for regular scrub, repair those pages that are errored.
1068 	 * In case of I/O errors in the area that is supposed to be
1069 	 * repaired, continue by picking good copies of those pages.
1070 	 * Select the good pages from mirrors to rewrite bad pages from
1071 	 * the area to fix. Afterwards verify the checksum of the block
1072 	 * that is supposed to be repaired. This verification step is
1073 	 * only done for the purpose of statistic counting and for the
1074 	 * final scrub report, whether errors remain.
1075 	 * A perfect algorithm could make use of the checksum and try
1076 	 * all possible combinations of pages from the different mirrors
1077 	 * until the checksum verification succeeds. For example, when
1078 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1079 	 * of mirror #2 is readable but the final checksum test fails,
1080 	 * then the 2nd page of mirror #3 could be tried, whether now
1081 	 * the final checksum succeedes. But this would be a rare
1082 	 * exception and is therefore not implemented. At least it is
1083 	 * avoided that the good copy is overwritten.
1084 	 * A more useful improvement would be to pick the sectors
1085 	 * without I/O error based on sector sizes (512 bytes on legacy
1086 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1087 	 * mirror could be repaired by taking 512 byte of a different
1088 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1089 	 * area are unreadable.
1090 	 */
1091 
1092 	/* can only fix I/O errors from here on */
1093 	if (sblock_bad->no_io_error_seen)
1094 		goto did_not_correct_error;
1095 
1096 	success = 1;
1097 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1098 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1099 
1100 		if (!page_bad->io_error)
1101 			continue;
1102 
1103 		for (mirror_index = 0;
1104 		     mirror_index < BTRFS_MAX_MIRRORS &&
1105 		     sblocks_for_recheck[mirror_index].page_count > 0;
1106 		     mirror_index++) {
1107 			struct scrub_block *sblock_other = sblocks_for_recheck +
1108 							   mirror_index;
1109 			struct scrub_page *page_other = sblock_other->pagev[
1110 							page_num];
1111 
1112 			if (!page_other->io_error) {
1113 				ret = scrub_repair_page_from_good_copy(
1114 					sblock_bad, sblock_other, page_num, 0);
1115 				if (0 == ret) {
1116 					page_bad->io_error = 0;
1117 					break; /* succeeded for this page */
1118 				}
1119 			}
1120 		}
1121 
1122 		if (page_bad->io_error) {
1123 			/* did not find a mirror to copy the page from */
1124 			success = 0;
1125 		}
1126 	}
1127 
1128 	if (success) {
1129 		if (is_metadata || have_csum) {
1130 			/*
1131 			 * need to verify the checksum now that all
1132 			 * sectors on disk are repaired (the write
1133 			 * request for data to be repaired is on its way).
1134 			 * Just be lazy and use scrub_recheck_block()
1135 			 * which re-reads the data before the checksum
1136 			 * is verified, but most likely the data comes out
1137 			 * of the page cache.
1138 			 */
1139 			scrub_recheck_block(fs_info, sblock_bad,
1140 					    is_metadata, have_csum, csum,
1141 					    generation, sctx->csum_size);
1142 			if (!sblock_bad->header_error &&
1143 			    !sblock_bad->checksum_error &&
1144 			    sblock_bad->no_io_error_seen)
1145 				goto corrected_error;
1146 			else
1147 				goto did_not_correct_error;
1148 		} else {
1149 corrected_error:
1150 			spin_lock(&sctx->stat_lock);
1151 			sctx->stat.corrected_errors++;
1152 			spin_unlock(&sctx->stat_lock);
1153 			printk_ratelimited_in_rcu(KERN_ERR
1154 				"btrfs: fixed up error at logical %llu on dev %s\n",
1155 				(unsigned long long)logical,
1156 				rcu_str_deref(dev->name));
1157 		}
1158 	} else {
1159 did_not_correct_error:
1160 		spin_lock(&sctx->stat_lock);
1161 		sctx->stat.uncorrectable_errors++;
1162 		spin_unlock(&sctx->stat_lock);
1163 		printk_ratelimited_in_rcu(KERN_ERR
1164 			"btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1165 			(unsigned long long)logical,
1166 			rcu_str_deref(dev->name));
1167 	}
1168 
1169 out:
1170 	if (sblocks_for_recheck) {
1171 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1172 		     mirror_index++) {
1173 			struct scrub_block *sblock = sblocks_for_recheck +
1174 						     mirror_index;
1175 			int page_index;
1176 
1177 			for (page_index = 0; page_index < sblock->page_count;
1178 			     page_index++) {
1179 				sblock->pagev[page_index]->sblock = NULL;
1180 				scrub_page_put(sblock->pagev[page_index]);
1181 			}
1182 		}
1183 		kfree(sblocks_for_recheck);
1184 	}
1185 
1186 	return 0;
1187 }
1188 
1189 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1190 				     struct btrfs_fs_info *fs_info,
1191 				     struct scrub_block *original_sblock,
1192 				     u64 length, u64 logical,
1193 				     struct scrub_block *sblocks_for_recheck)
1194 {
1195 	int page_index;
1196 	int mirror_index;
1197 	int ret;
1198 
1199 	/*
1200 	 * note: the two members ref_count and outstanding_pages
1201 	 * are not used (and not set) in the blocks that are used for
1202 	 * the recheck procedure
1203 	 */
1204 
1205 	page_index = 0;
1206 	while (length > 0) {
1207 		u64 sublen = min_t(u64, length, PAGE_SIZE);
1208 		u64 mapped_length = sublen;
1209 		struct btrfs_bio *bbio = NULL;
1210 
1211 		/*
1212 		 * with a length of PAGE_SIZE, each returned stripe
1213 		 * represents one mirror
1214 		 */
1215 		ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1216 				      &mapped_length, &bbio, 0);
1217 		if (ret || !bbio || mapped_length < sublen) {
1218 			kfree(bbio);
1219 			return -EIO;
1220 		}
1221 
1222 		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1223 		for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1224 		     mirror_index++) {
1225 			struct scrub_block *sblock;
1226 			struct scrub_page *page;
1227 
1228 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1229 				continue;
1230 
1231 			sblock = sblocks_for_recheck + mirror_index;
1232 			sblock->sctx = sctx;
1233 			page = kzalloc(sizeof(*page), GFP_NOFS);
1234 			if (!page) {
1235 leave_nomem:
1236 				spin_lock(&sctx->stat_lock);
1237 				sctx->stat.malloc_errors++;
1238 				spin_unlock(&sctx->stat_lock);
1239 				kfree(bbio);
1240 				return -ENOMEM;
1241 			}
1242 			scrub_page_get(page);
1243 			sblock->pagev[page_index] = page;
1244 			page->logical = logical;
1245 			page->physical = bbio->stripes[mirror_index].physical;
1246 			BUG_ON(page_index >= original_sblock->page_count);
1247 			page->physical_for_dev_replace =
1248 				original_sblock->pagev[page_index]->
1249 				physical_for_dev_replace;
1250 			/* for missing devices, dev->bdev is NULL */
1251 			page->dev = bbio->stripes[mirror_index].dev;
1252 			page->mirror_num = mirror_index + 1;
1253 			sblock->page_count++;
1254 			page->page = alloc_page(GFP_NOFS);
1255 			if (!page->page)
1256 				goto leave_nomem;
1257 		}
1258 		kfree(bbio);
1259 		length -= sublen;
1260 		logical += sublen;
1261 		page_index++;
1262 	}
1263 
1264 	return 0;
1265 }
1266 
1267 /*
1268  * this function will check the on disk data for checksum errors, header
1269  * errors and read I/O errors. If any I/O errors happen, the exact pages
1270  * which are errored are marked as being bad. The goal is to enable scrub
1271  * to take those pages that are not errored from all the mirrors so that
1272  * the pages that are errored in the just handled mirror can be repaired.
1273  */
1274 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1275 				struct scrub_block *sblock, int is_metadata,
1276 				int have_csum, u8 *csum, u64 generation,
1277 				u16 csum_size)
1278 {
1279 	int page_num;
1280 
1281 	sblock->no_io_error_seen = 1;
1282 	sblock->header_error = 0;
1283 	sblock->checksum_error = 0;
1284 
1285 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1286 		struct bio *bio;
1287 		struct scrub_page *page = sblock->pagev[page_num];
1288 		DECLARE_COMPLETION_ONSTACK(complete);
1289 
1290 		if (page->dev->bdev == NULL) {
1291 			page->io_error = 1;
1292 			sblock->no_io_error_seen = 0;
1293 			continue;
1294 		}
1295 
1296 		WARN_ON(!page->page);
1297 		bio = bio_alloc(GFP_NOFS, 1);
1298 		if (!bio) {
1299 			page->io_error = 1;
1300 			sblock->no_io_error_seen = 0;
1301 			continue;
1302 		}
1303 		bio->bi_bdev = page->dev->bdev;
1304 		bio->bi_sector = page->physical >> 9;
1305 		bio->bi_end_io = scrub_complete_bio_end_io;
1306 		bio->bi_private = &complete;
1307 
1308 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1309 		btrfsic_submit_bio(READ, bio);
1310 
1311 		/* this will also unplug the queue */
1312 		wait_for_completion(&complete);
1313 
1314 		page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1315 		if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1316 			sblock->no_io_error_seen = 0;
1317 		bio_put(bio);
1318 	}
1319 
1320 	if (sblock->no_io_error_seen)
1321 		scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1322 					     have_csum, csum, generation,
1323 					     csum_size);
1324 
1325 	return;
1326 }
1327 
1328 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1329 					 struct scrub_block *sblock,
1330 					 int is_metadata, int have_csum,
1331 					 const u8 *csum, u64 generation,
1332 					 u16 csum_size)
1333 {
1334 	int page_num;
1335 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1336 	u32 crc = ~(u32)0;
1337 	struct btrfs_root *root = fs_info->extent_root;
1338 	void *mapped_buffer;
1339 
1340 	WARN_ON(!sblock->pagev[0]->page);
1341 	if (is_metadata) {
1342 		struct btrfs_header *h;
1343 
1344 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1345 		h = (struct btrfs_header *)mapped_buffer;
1346 
1347 		if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1348 		    memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1349 		    memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1350 			   BTRFS_UUID_SIZE)) {
1351 			sblock->header_error = 1;
1352 		} else if (generation != le64_to_cpu(h->generation)) {
1353 			sblock->header_error = 1;
1354 			sblock->generation_error = 1;
1355 		}
1356 		csum = h->csum;
1357 	} else {
1358 		if (!have_csum)
1359 			return;
1360 
1361 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1362 	}
1363 
1364 	for (page_num = 0;;) {
1365 		if (page_num == 0 && is_metadata)
1366 			crc = btrfs_csum_data(root,
1367 				((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1368 				crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1369 		else
1370 			crc = btrfs_csum_data(root, mapped_buffer, crc,
1371 					      PAGE_SIZE);
1372 
1373 		kunmap_atomic(mapped_buffer);
1374 		page_num++;
1375 		if (page_num >= sblock->page_count)
1376 			break;
1377 		WARN_ON(!sblock->pagev[page_num]->page);
1378 
1379 		mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1380 	}
1381 
1382 	btrfs_csum_final(crc, calculated_csum);
1383 	if (memcmp(calculated_csum, csum, csum_size))
1384 		sblock->checksum_error = 1;
1385 }
1386 
1387 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1388 {
1389 	complete((struct completion *)bio->bi_private);
1390 }
1391 
1392 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1393 					     struct scrub_block *sblock_good,
1394 					     int force_write)
1395 {
1396 	int page_num;
1397 	int ret = 0;
1398 
1399 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1400 		int ret_sub;
1401 
1402 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1403 							   sblock_good,
1404 							   page_num,
1405 							   force_write);
1406 		if (ret_sub)
1407 			ret = ret_sub;
1408 	}
1409 
1410 	return ret;
1411 }
1412 
1413 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1414 					    struct scrub_block *sblock_good,
1415 					    int page_num, int force_write)
1416 {
1417 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1418 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1419 
1420 	BUG_ON(page_bad->page == NULL);
1421 	BUG_ON(page_good->page == NULL);
1422 	if (force_write || sblock_bad->header_error ||
1423 	    sblock_bad->checksum_error || page_bad->io_error) {
1424 		struct bio *bio;
1425 		int ret;
1426 		DECLARE_COMPLETION_ONSTACK(complete);
1427 
1428 		if (!page_bad->dev->bdev) {
1429 			printk_ratelimited(KERN_WARNING
1430 				"btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1431 			return -EIO;
1432 		}
1433 
1434 		bio = bio_alloc(GFP_NOFS, 1);
1435 		if (!bio)
1436 			return -EIO;
1437 		bio->bi_bdev = page_bad->dev->bdev;
1438 		bio->bi_sector = page_bad->physical >> 9;
1439 		bio->bi_end_io = scrub_complete_bio_end_io;
1440 		bio->bi_private = &complete;
1441 
1442 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1443 		if (PAGE_SIZE != ret) {
1444 			bio_put(bio);
1445 			return -EIO;
1446 		}
1447 		btrfsic_submit_bio(WRITE, bio);
1448 
1449 		/* this will also unplug the queue */
1450 		wait_for_completion(&complete);
1451 		if (!bio_flagged(bio, BIO_UPTODATE)) {
1452 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1453 				BTRFS_DEV_STAT_WRITE_ERRS);
1454 			btrfs_dev_replace_stats_inc(
1455 				&sblock_bad->sctx->dev_root->fs_info->
1456 				dev_replace.num_write_errors);
1457 			bio_put(bio);
1458 			return -EIO;
1459 		}
1460 		bio_put(bio);
1461 	}
1462 
1463 	return 0;
1464 }
1465 
1466 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1467 {
1468 	int page_num;
1469 
1470 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1471 		int ret;
1472 
1473 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1474 		if (ret)
1475 			btrfs_dev_replace_stats_inc(
1476 				&sblock->sctx->dev_root->fs_info->dev_replace.
1477 				num_write_errors);
1478 	}
1479 }
1480 
1481 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1482 					   int page_num)
1483 {
1484 	struct scrub_page *spage = sblock->pagev[page_num];
1485 
1486 	BUG_ON(spage->page == NULL);
1487 	if (spage->io_error) {
1488 		void *mapped_buffer = kmap_atomic(spage->page);
1489 
1490 		memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1491 		flush_dcache_page(spage->page);
1492 		kunmap_atomic(mapped_buffer);
1493 	}
1494 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1495 }
1496 
1497 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1498 				    struct scrub_page *spage)
1499 {
1500 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1501 	struct scrub_bio *sbio;
1502 	int ret;
1503 
1504 	mutex_lock(&wr_ctx->wr_lock);
1505 again:
1506 	if (!wr_ctx->wr_curr_bio) {
1507 		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1508 					      GFP_NOFS);
1509 		if (!wr_ctx->wr_curr_bio) {
1510 			mutex_unlock(&wr_ctx->wr_lock);
1511 			return -ENOMEM;
1512 		}
1513 		wr_ctx->wr_curr_bio->sctx = sctx;
1514 		wr_ctx->wr_curr_bio->page_count = 0;
1515 	}
1516 	sbio = wr_ctx->wr_curr_bio;
1517 	if (sbio->page_count == 0) {
1518 		struct bio *bio;
1519 
1520 		sbio->physical = spage->physical_for_dev_replace;
1521 		sbio->logical = spage->logical;
1522 		sbio->dev = wr_ctx->tgtdev;
1523 		bio = sbio->bio;
1524 		if (!bio) {
1525 			bio = bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1526 			if (!bio) {
1527 				mutex_unlock(&wr_ctx->wr_lock);
1528 				return -ENOMEM;
1529 			}
1530 			sbio->bio = bio;
1531 		}
1532 
1533 		bio->bi_private = sbio;
1534 		bio->bi_end_io = scrub_wr_bio_end_io;
1535 		bio->bi_bdev = sbio->dev->bdev;
1536 		bio->bi_sector = sbio->physical >> 9;
1537 		sbio->err = 0;
1538 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1539 		   spage->physical_for_dev_replace ||
1540 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1541 		   spage->logical) {
1542 		scrub_wr_submit(sctx);
1543 		goto again;
1544 	}
1545 
1546 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1547 	if (ret != PAGE_SIZE) {
1548 		if (sbio->page_count < 1) {
1549 			bio_put(sbio->bio);
1550 			sbio->bio = NULL;
1551 			mutex_unlock(&wr_ctx->wr_lock);
1552 			return -EIO;
1553 		}
1554 		scrub_wr_submit(sctx);
1555 		goto again;
1556 	}
1557 
1558 	sbio->pagev[sbio->page_count] = spage;
1559 	scrub_page_get(spage);
1560 	sbio->page_count++;
1561 	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1562 		scrub_wr_submit(sctx);
1563 	mutex_unlock(&wr_ctx->wr_lock);
1564 
1565 	return 0;
1566 }
1567 
1568 static void scrub_wr_submit(struct scrub_ctx *sctx)
1569 {
1570 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1571 	struct scrub_bio *sbio;
1572 
1573 	if (!wr_ctx->wr_curr_bio)
1574 		return;
1575 
1576 	sbio = wr_ctx->wr_curr_bio;
1577 	wr_ctx->wr_curr_bio = NULL;
1578 	WARN_ON(!sbio->bio->bi_bdev);
1579 	scrub_pending_bio_inc(sctx);
1580 	/* process all writes in a single worker thread. Then the block layer
1581 	 * orders the requests before sending them to the driver which
1582 	 * doubled the write performance on spinning disks when measured
1583 	 * with Linux 3.5 */
1584 	btrfsic_submit_bio(WRITE, sbio->bio);
1585 }
1586 
1587 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1588 {
1589 	struct scrub_bio *sbio = bio->bi_private;
1590 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1591 
1592 	sbio->err = err;
1593 	sbio->bio = bio;
1594 
1595 	sbio->work.func = scrub_wr_bio_end_io_worker;
1596 	btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1597 }
1598 
1599 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1600 {
1601 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1602 	struct scrub_ctx *sctx = sbio->sctx;
1603 	int i;
1604 
1605 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1606 	if (sbio->err) {
1607 		struct btrfs_dev_replace *dev_replace =
1608 			&sbio->sctx->dev_root->fs_info->dev_replace;
1609 
1610 		for (i = 0; i < sbio->page_count; i++) {
1611 			struct scrub_page *spage = sbio->pagev[i];
1612 
1613 			spage->io_error = 1;
1614 			btrfs_dev_replace_stats_inc(&dev_replace->
1615 						    num_write_errors);
1616 		}
1617 	}
1618 
1619 	for (i = 0; i < sbio->page_count; i++)
1620 		scrub_page_put(sbio->pagev[i]);
1621 
1622 	bio_put(sbio->bio);
1623 	kfree(sbio);
1624 	scrub_pending_bio_dec(sctx);
1625 }
1626 
1627 static int scrub_checksum(struct scrub_block *sblock)
1628 {
1629 	u64 flags;
1630 	int ret;
1631 
1632 	WARN_ON(sblock->page_count < 1);
1633 	flags = sblock->pagev[0]->flags;
1634 	ret = 0;
1635 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1636 		ret = scrub_checksum_data(sblock);
1637 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1638 		ret = scrub_checksum_tree_block(sblock);
1639 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1640 		(void)scrub_checksum_super(sblock);
1641 	else
1642 		WARN_ON(1);
1643 	if (ret)
1644 		scrub_handle_errored_block(sblock);
1645 
1646 	return ret;
1647 }
1648 
1649 static int scrub_checksum_data(struct scrub_block *sblock)
1650 {
1651 	struct scrub_ctx *sctx = sblock->sctx;
1652 	u8 csum[BTRFS_CSUM_SIZE];
1653 	u8 *on_disk_csum;
1654 	struct page *page;
1655 	void *buffer;
1656 	u32 crc = ~(u32)0;
1657 	int fail = 0;
1658 	struct btrfs_root *root = sctx->dev_root;
1659 	u64 len;
1660 	int index;
1661 
1662 	BUG_ON(sblock->page_count < 1);
1663 	if (!sblock->pagev[0]->have_csum)
1664 		return 0;
1665 
1666 	on_disk_csum = sblock->pagev[0]->csum;
1667 	page = sblock->pagev[0]->page;
1668 	buffer = kmap_atomic(page);
1669 
1670 	len = sctx->sectorsize;
1671 	index = 0;
1672 	for (;;) {
1673 		u64 l = min_t(u64, len, PAGE_SIZE);
1674 
1675 		crc = btrfs_csum_data(root, buffer, crc, l);
1676 		kunmap_atomic(buffer);
1677 		len -= l;
1678 		if (len == 0)
1679 			break;
1680 		index++;
1681 		BUG_ON(index >= sblock->page_count);
1682 		BUG_ON(!sblock->pagev[index]->page);
1683 		page = sblock->pagev[index]->page;
1684 		buffer = kmap_atomic(page);
1685 	}
1686 
1687 	btrfs_csum_final(crc, csum);
1688 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1689 		fail = 1;
1690 
1691 	return fail;
1692 }
1693 
1694 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1695 {
1696 	struct scrub_ctx *sctx = sblock->sctx;
1697 	struct btrfs_header *h;
1698 	struct btrfs_root *root = sctx->dev_root;
1699 	struct btrfs_fs_info *fs_info = root->fs_info;
1700 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1701 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1702 	struct page *page;
1703 	void *mapped_buffer;
1704 	u64 mapped_size;
1705 	void *p;
1706 	u32 crc = ~(u32)0;
1707 	int fail = 0;
1708 	int crc_fail = 0;
1709 	u64 len;
1710 	int index;
1711 
1712 	BUG_ON(sblock->page_count < 1);
1713 	page = sblock->pagev[0]->page;
1714 	mapped_buffer = kmap_atomic(page);
1715 	h = (struct btrfs_header *)mapped_buffer;
1716 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1717 
1718 	/*
1719 	 * we don't use the getter functions here, as we
1720 	 * a) don't have an extent buffer and
1721 	 * b) the page is already kmapped
1722 	 */
1723 
1724 	if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1725 		++fail;
1726 
1727 	if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1728 		++fail;
1729 
1730 	if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1731 		++fail;
1732 
1733 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1734 		   BTRFS_UUID_SIZE))
1735 		++fail;
1736 
1737 	WARN_ON(sctx->nodesize != sctx->leafsize);
1738 	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1739 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1740 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1741 	index = 0;
1742 	for (;;) {
1743 		u64 l = min_t(u64, len, mapped_size);
1744 
1745 		crc = btrfs_csum_data(root, p, crc, l);
1746 		kunmap_atomic(mapped_buffer);
1747 		len -= l;
1748 		if (len == 0)
1749 			break;
1750 		index++;
1751 		BUG_ON(index >= sblock->page_count);
1752 		BUG_ON(!sblock->pagev[index]->page);
1753 		page = sblock->pagev[index]->page;
1754 		mapped_buffer = kmap_atomic(page);
1755 		mapped_size = PAGE_SIZE;
1756 		p = mapped_buffer;
1757 	}
1758 
1759 	btrfs_csum_final(crc, calculated_csum);
1760 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1761 		++crc_fail;
1762 
1763 	return fail || crc_fail;
1764 }
1765 
1766 static int scrub_checksum_super(struct scrub_block *sblock)
1767 {
1768 	struct btrfs_super_block *s;
1769 	struct scrub_ctx *sctx = sblock->sctx;
1770 	struct btrfs_root *root = sctx->dev_root;
1771 	struct btrfs_fs_info *fs_info = root->fs_info;
1772 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1773 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1774 	struct page *page;
1775 	void *mapped_buffer;
1776 	u64 mapped_size;
1777 	void *p;
1778 	u32 crc = ~(u32)0;
1779 	int fail_gen = 0;
1780 	int fail_cor = 0;
1781 	u64 len;
1782 	int index;
1783 
1784 	BUG_ON(sblock->page_count < 1);
1785 	page = sblock->pagev[0]->page;
1786 	mapped_buffer = kmap_atomic(page);
1787 	s = (struct btrfs_super_block *)mapped_buffer;
1788 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1789 
1790 	if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1791 		++fail_cor;
1792 
1793 	if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1794 		++fail_gen;
1795 
1796 	if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1797 		++fail_cor;
1798 
1799 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1800 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1801 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1802 	index = 0;
1803 	for (;;) {
1804 		u64 l = min_t(u64, len, mapped_size);
1805 
1806 		crc = btrfs_csum_data(root, p, crc, l);
1807 		kunmap_atomic(mapped_buffer);
1808 		len -= l;
1809 		if (len == 0)
1810 			break;
1811 		index++;
1812 		BUG_ON(index >= sblock->page_count);
1813 		BUG_ON(!sblock->pagev[index]->page);
1814 		page = sblock->pagev[index]->page;
1815 		mapped_buffer = kmap_atomic(page);
1816 		mapped_size = PAGE_SIZE;
1817 		p = mapped_buffer;
1818 	}
1819 
1820 	btrfs_csum_final(crc, calculated_csum);
1821 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1822 		++fail_cor;
1823 
1824 	if (fail_cor + fail_gen) {
1825 		/*
1826 		 * if we find an error in a super block, we just report it.
1827 		 * They will get written with the next transaction commit
1828 		 * anyway
1829 		 */
1830 		spin_lock(&sctx->stat_lock);
1831 		++sctx->stat.super_errors;
1832 		spin_unlock(&sctx->stat_lock);
1833 		if (fail_cor)
1834 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1835 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1836 		else
1837 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1838 				BTRFS_DEV_STAT_GENERATION_ERRS);
1839 	}
1840 
1841 	return fail_cor + fail_gen;
1842 }
1843 
1844 static void scrub_block_get(struct scrub_block *sblock)
1845 {
1846 	atomic_inc(&sblock->ref_count);
1847 }
1848 
1849 static void scrub_block_put(struct scrub_block *sblock)
1850 {
1851 	if (atomic_dec_and_test(&sblock->ref_count)) {
1852 		int i;
1853 
1854 		for (i = 0; i < sblock->page_count; i++)
1855 			scrub_page_put(sblock->pagev[i]);
1856 		kfree(sblock);
1857 	}
1858 }
1859 
1860 static void scrub_page_get(struct scrub_page *spage)
1861 {
1862 	atomic_inc(&spage->ref_count);
1863 }
1864 
1865 static void scrub_page_put(struct scrub_page *spage)
1866 {
1867 	if (atomic_dec_and_test(&spage->ref_count)) {
1868 		if (spage->page)
1869 			__free_page(spage->page);
1870 		kfree(spage);
1871 	}
1872 }
1873 
1874 static void scrub_submit(struct scrub_ctx *sctx)
1875 {
1876 	struct scrub_bio *sbio;
1877 
1878 	if (sctx->curr == -1)
1879 		return;
1880 
1881 	sbio = sctx->bios[sctx->curr];
1882 	sctx->curr = -1;
1883 	scrub_pending_bio_inc(sctx);
1884 
1885 	if (!sbio->bio->bi_bdev) {
1886 		/*
1887 		 * this case should not happen. If btrfs_map_block() is
1888 		 * wrong, it could happen for dev-replace operations on
1889 		 * missing devices when no mirrors are available, but in
1890 		 * this case it should already fail the mount.
1891 		 * This case is handled correctly (but _very_ slowly).
1892 		 */
1893 		printk_ratelimited(KERN_WARNING
1894 			"btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1895 		bio_endio(sbio->bio, -EIO);
1896 	} else {
1897 		btrfsic_submit_bio(READ, sbio->bio);
1898 	}
1899 }
1900 
1901 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1902 				    struct scrub_page *spage)
1903 {
1904 	struct scrub_block *sblock = spage->sblock;
1905 	struct scrub_bio *sbio;
1906 	int ret;
1907 
1908 again:
1909 	/*
1910 	 * grab a fresh bio or wait for one to become available
1911 	 */
1912 	while (sctx->curr == -1) {
1913 		spin_lock(&sctx->list_lock);
1914 		sctx->curr = sctx->first_free;
1915 		if (sctx->curr != -1) {
1916 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1917 			sctx->bios[sctx->curr]->next_free = -1;
1918 			sctx->bios[sctx->curr]->page_count = 0;
1919 			spin_unlock(&sctx->list_lock);
1920 		} else {
1921 			spin_unlock(&sctx->list_lock);
1922 			wait_event(sctx->list_wait, sctx->first_free != -1);
1923 		}
1924 	}
1925 	sbio = sctx->bios[sctx->curr];
1926 	if (sbio->page_count == 0) {
1927 		struct bio *bio;
1928 
1929 		sbio->physical = spage->physical;
1930 		sbio->logical = spage->logical;
1931 		sbio->dev = spage->dev;
1932 		bio = sbio->bio;
1933 		if (!bio) {
1934 			bio = bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1935 			if (!bio)
1936 				return -ENOMEM;
1937 			sbio->bio = bio;
1938 		}
1939 
1940 		bio->bi_private = sbio;
1941 		bio->bi_end_io = scrub_bio_end_io;
1942 		bio->bi_bdev = sbio->dev->bdev;
1943 		bio->bi_sector = sbio->physical >> 9;
1944 		sbio->err = 0;
1945 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1946 		   spage->physical ||
1947 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1948 		   spage->logical ||
1949 		   sbio->dev != spage->dev) {
1950 		scrub_submit(sctx);
1951 		goto again;
1952 	}
1953 
1954 	sbio->pagev[sbio->page_count] = spage;
1955 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1956 	if (ret != PAGE_SIZE) {
1957 		if (sbio->page_count < 1) {
1958 			bio_put(sbio->bio);
1959 			sbio->bio = NULL;
1960 			return -EIO;
1961 		}
1962 		scrub_submit(sctx);
1963 		goto again;
1964 	}
1965 
1966 	scrub_block_get(sblock); /* one for the page added to the bio */
1967 	atomic_inc(&sblock->outstanding_pages);
1968 	sbio->page_count++;
1969 	if (sbio->page_count == sctx->pages_per_rd_bio)
1970 		scrub_submit(sctx);
1971 
1972 	return 0;
1973 }
1974 
1975 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1976 		       u64 physical, struct btrfs_device *dev, u64 flags,
1977 		       u64 gen, int mirror_num, u8 *csum, int force,
1978 		       u64 physical_for_dev_replace)
1979 {
1980 	struct scrub_block *sblock;
1981 	int index;
1982 
1983 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1984 	if (!sblock) {
1985 		spin_lock(&sctx->stat_lock);
1986 		sctx->stat.malloc_errors++;
1987 		spin_unlock(&sctx->stat_lock);
1988 		return -ENOMEM;
1989 	}
1990 
1991 	/* one ref inside this function, plus one for each page added to
1992 	 * a bio later on */
1993 	atomic_set(&sblock->ref_count, 1);
1994 	sblock->sctx = sctx;
1995 	sblock->no_io_error_seen = 1;
1996 
1997 	for (index = 0; len > 0; index++) {
1998 		struct scrub_page *spage;
1999 		u64 l = min_t(u64, len, PAGE_SIZE);
2000 
2001 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
2002 		if (!spage) {
2003 leave_nomem:
2004 			spin_lock(&sctx->stat_lock);
2005 			sctx->stat.malloc_errors++;
2006 			spin_unlock(&sctx->stat_lock);
2007 			scrub_block_put(sblock);
2008 			return -ENOMEM;
2009 		}
2010 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2011 		scrub_page_get(spage);
2012 		sblock->pagev[index] = spage;
2013 		spage->sblock = sblock;
2014 		spage->dev = dev;
2015 		spage->flags = flags;
2016 		spage->generation = gen;
2017 		spage->logical = logical;
2018 		spage->physical = physical;
2019 		spage->physical_for_dev_replace = physical_for_dev_replace;
2020 		spage->mirror_num = mirror_num;
2021 		if (csum) {
2022 			spage->have_csum = 1;
2023 			memcpy(spage->csum, csum, sctx->csum_size);
2024 		} else {
2025 			spage->have_csum = 0;
2026 		}
2027 		sblock->page_count++;
2028 		spage->page = alloc_page(GFP_NOFS);
2029 		if (!spage->page)
2030 			goto leave_nomem;
2031 		len -= l;
2032 		logical += l;
2033 		physical += l;
2034 		physical_for_dev_replace += l;
2035 	}
2036 
2037 	WARN_ON(sblock->page_count == 0);
2038 	for (index = 0; index < sblock->page_count; index++) {
2039 		struct scrub_page *spage = sblock->pagev[index];
2040 		int ret;
2041 
2042 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2043 		if (ret) {
2044 			scrub_block_put(sblock);
2045 			return ret;
2046 		}
2047 	}
2048 
2049 	if (force)
2050 		scrub_submit(sctx);
2051 
2052 	/* last one frees, either here or in bio completion for last page */
2053 	scrub_block_put(sblock);
2054 	return 0;
2055 }
2056 
2057 static void scrub_bio_end_io(struct bio *bio, int err)
2058 {
2059 	struct scrub_bio *sbio = bio->bi_private;
2060 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2061 
2062 	sbio->err = err;
2063 	sbio->bio = bio;
2064 
2065 	btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2066 }
2067 
2068 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2069 {
2070 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2071 	struct scrub_ctx *sctx = sbio->sctx;
2072 	int i;
2073 
2074 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2075 	if (sbio->err) {
2076 		for (i = 0; i < sbio->page_count; i++) {
2077 			struct scrub_page *spage = sbio->pagev[i];
2078 
2079 			spage->io_error = 1;
2080 			spage->sblock->no_io_error_seen = 0;
2081 		}
2082 	}
2083 
2084 	/* now complete the scrub_block items that have all pages completed */
2085 	for (i = 0; i < sbio->page_count; i++) {
2086 		struct scrub_page *spage = sbio->pagev[i];
2087 		struct scrub_block *sblock = spage->sblock;
2088 
2089 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2090 			scrub_block_complete(sblock);
2091 		scrub_block_put(sblock);
2092 	}
2093 
2094 	bio_put(sbio->bio);
2095 	sbio->bio = NULL;
2096 	spin_lock(&sctx->list_lock);
2097 	sbio->next_free = sctx->first_free;
2098 	sctx->first_free = sbio->index;
2099 	spin_unlock(&sctx->list_lock);
2100 
2101 	if (sctx->is_dev_replace &&
2102 	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2103 		mutex_lock(&sctx->wr_ctx.wr_lock);
2104 		scrub_wr_submit(sctx);
2105 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2106 	}
2107 
2108 	scrub_pending_bio_dec(sctx);
2109 }
2110 
2111 static void scrub_block_complete(struct scrub_block *sblock)
2112 {
2113 	if (!sblock->no_io_error_seen) {
2114 		scrub_handle_errored_block(sblock);
2115 	} else {
2116 		/*
2117 		 * if has checksum error, write via repair mechanism in
2118 		 * dev replace case, otherwise write here in dev replace
2119 		 * case.
2120 		 */
2121 		if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2122 			scrub_write_block_to_dev_replace(sblock);
2123 	}
2124 }
2125 
2126 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2127 			   u8 *csum)
2128 {
2129 	struct btrfs_ordered_sum *sum = NULL;
2130 	int ret = 0;
2131 	unsigned long i;
2132 	unsigned long num_sectors;
2133 
2134 	while (!list_empty(&sctx->csum_list)) {
2135 		sum = list_first_entry(&sctx->csum_list,
2136 				       struct btrfs_ordered_sum, list);
2137 		if (sum->bytenr > logical)
2138 			return 0;
2139 		if (sum->bytenr + sum->len > logical)
2140 			break;
2141 
2142 		++sctx->stat.csum_discards;
2143 		list_del(&sum->list);
2144 		kfree(sum);
2145 		sum = NULL;
2146 	}
2147 	if (!sum)
2148 		return 0;
2149 
2150 	num_sectors = sum->len / sctx->sectorsize;
2151 	for (i = 0; i < num_sectors; ++i) {
2152 		if (sum->sums[i].bytenr == logical) {
2153 			memcpy(csum, &sum->sums[i].sum, sctx->csum_size);
2154 			ret = 1;
2155 			break;
2156 		}
2157 	}
2158 	if (ret && i == num_sectors - 1) {
2159 		list_del(&sum->list);
2160 		kfree(sum);
2161 	}
2162 	return ret;
2163 }
2164 
2165 /* scrub extent tries to collect up to 64 kB for each bio */
2166 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2167 			u64 physical, struct btrfs_device *dev, u64 flags,
2168 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2169 {
2170 	int ret;
2171 	u8 csum[BTRFS_CSUM_SIZE];
2172 	u32 blocksize;
2173 
2174 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2175 		blocksize = sctx->sectorsize;
2176 		spin_lock(&sctx->stat_lock);
2177 		sctx->stat.data_extents_scrubbed++;
2178 		sctx->stat.data_bytes_scrubbed += len;
2179 		spin_unlock(&sctx->stat_lock);
2180 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2181 		WARN_ON(sctx->nodesize != sctx->leafsize);
2182 		blocksize = sctx->nodesize;
2183 		spin_lock(&sctx->stat_lock);
2184 		sctx->stat.tree_extents_scrubbed++;
2185 		sctx->stat.tree_bytes_scrubbed += len;
2186 		spin_unlock(&sctx->stat_lock);
2187 	} else {
2188 		blocksize = sctx->sectorsize;
2189 		WARN_ON(1);
2190 	}
2191 
2192 	while (len) {
2193 		u64 l = min_t(u64, len, blocksize);
2194 		int have_csum = 0;
2195 
2196 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2197 			/* push csums to sbio */
2198 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2199 			if (have_csum == 0)
2200 				++sctx->stat.no_csum;
2201 			if (sctx->is_dev_replace && !have_csum) {
2202 				ret = copy_nocow_pages(sctx, logical, l,
2203 						       mirror_num,
2204 						      physical_for_dev_replace);
2205 				goto behind_scrub_pages;
2206 			}
2207 		}
2208 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2209 				  mirror_num, have_csum ? csum : NULL, 0,
2210 				  physical_for_dev_replace);
2211 behind_scrub_pages:
2212 		if (ret)
2213 			return ret;
2214 		len -= l;
2215 		logical += l;
2216 		physical += l;
2217 		physical_for_dev_replace += l;
2218 	}
2219 	return 0;
2220 }
2221 
2222 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2223 					   struct map_lookup *map,
2224 					   struct btrfs_device *scrub_dev,
2225 					   int num, u64 base, u64 length,
2226 					   int is_dev_replace)
2227 {
2228 	struct btrfs_path *path;
2229 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2230 	struct btrfs_root *root = fs_info->extent_root;
2231 	struct btrfs_root *csum_root = fs_info->csum_root;
2232 	struct btrfs_extent_item *extent;
2233 	struct blk_plug plug;
2234 	u64 flags;
2235 	int ret;
2236 	int slot;
2237 	int i;
2238 	u64 nstripes;
2239 	struct extent_buffer *l;
2240 	struct btrfs_key key;
2241 	u64 physical;
2242 	u64 logical;
2243 	u64 generation;
2244 	int mirror_num;
2245 	struct reada_control *reada1;
2246 	struct reada_control *reada2;
2247 	struct btrfs_key key_start;
2248 	struct btrfs_key key_end;
2249 	u64 increment = map->stripe_len;
2250 	u64 offset;
2251 	u64 extent_logical;
2252 	u64 extent_physical;
2253 	u64 extent_len;
2254 	struct btrfs_device *extent_dev;
2255 	int extent_mirror_num;
2256 
2257 	nstripes = length;
2258 	offset = 0;
2259 	do_div(nstripes, map->stripe_len);
2260 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2261 		offset = map->stripe_len * num;
2262 		increment = map->stripe_len * map->num_stripes;
2263 		mirror_num = 1;
2264 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2265 		int factor = map->num_stripes / map->sub_stripes;
2266 		offset = map->stripe_len * (num / map->sub_stripes);
2267 		increment = map->stripe_len * factor;
2268 		mirror_num = num % map->sub_stripes + 1;
2269 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2270 		increment = map->stripe_len;
2271 		mirror_num = num % map->num_stripes + 1;
2272 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2273 		increment = map->stripe_len;
2274 		mirror_num = num % map->num_stripes + 1;
2275 	} else {
2276 		increment = map->stripe_len;
2277 		mirror_num = 1;
2278 	}
2279 
2280 	path = btrfs_alloc_path();
2281 	if (!path)
2282 		return -ENOMEM;
2283 
2284 	/*
2285 	 * work on commit root. The related disk blocks are static as
2286 	 * long as COW is applied. This means, it is save to rewrite
2287 	 * them to repair disk errors without any race conditions
2288 	 */
2289 	path->search_commit_root = 1;
2290 	path->skip_locking = 1;
2291 
2292 	/*
2293 	 * trigger the readahead for extent tree csum tree and wait for
2294 	 * completion. During readahead, the scrub is officially paused
2295 	 * to not hold off transaction commits
2296 	 */
2297 	logical = base + offset;
2298 
2299 	wait_event(sctx->list_wait,
2300 		   atomic_read(&sctx->bios_in_flight) == 0);
2301 	atomic_inc(&fs_info->scrubs_paused);
2302 	wake_up(&fs_info->scrub_pause_wait);
2303 
2304 	/* FIXME it might be better to start readahead at commit root */
2305 	key_start.objectid = logical;
2306 	key_start.type = BTRFS_EXTENT_ITEM_KEY;
2307 	key_start.offset = (u64)0;
2308 	key_end.objectid = base + offset + nstripes * increment;
2309 	key_end.type = BTRFS_EXTENT_ITEM_KEY;
2310 	key_end.offset = (u64)0;
2311 	reada1 = btrfs_reada_add(root, &key_start, &key_end);
2312 
2313 	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2314 	key_start.type = BTRFS_EXTENT_CSUM_KEY;
2315 	key_start.offset = logical;
2316 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2317 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
2318 	key_end.offset = base + offset + nstripes * increment;
2319 	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2320 
2321 	if (!IS_ERR(reada1))
2322 		btrfs_reada_wait(reada1);
2323 	if (!IS_ERR(reada2))
2324 		btrfs_reada_wait(reada2);
2325 
2326 	mutex_lock(&fs_info->scrub_lock);
2327 	while (atomic_read(&fs_info->scrub_pause_req)) {
2328 		mutex_unlock(&fs_info->scrub_lock);
2329 		wait_event(fs_info->scrub_pause_wait,
2330 		   atomic_read(&fs_info->scrub_pause_req) == 0);
2331 		mutex_lock(&fs_info->scrub_lock);
2332 	}
2333 	atomic_dec(&fs_info->scrubs_paused);
2334 	mutex_unlock(&fs_info->scrub_lock);
2335 	wake_up(&fs_info->scrub_pause_wait);
2336 
2337 	/*
2338 	 * collect all data csums for the stripe to avoid seeking during
2339 	 * the scrub. This might currently (crc32) end up to be about 1MB
2340 	 */
2341 	blk_start_plug(&plug);
2342 
2343 	/*
2344 	 * now find all extents for each stripe and scrub them
2345 	 */
2346 	logical = base + offset;
2347 	physical = map->stripes[num].physical;
2348 	ret = 0;
2349 	for (i = 0; i < nstripes; ++i) {
2350 		/*
2351 		 * canceled?
2352 		 */
2353 		if (atomic_read(&fs_info->scrub_cancel_req) ||
2354 		    atomic_read(&sctx->cancel_req)) {
2355 			ret = -ECANCELED;
2356 			goto out;
2357 		}
2358 		/*
2359 		 * check to see if we have to pause
2360 		 */
2361 		if (atomic_read(&fs_info->scrub_pause_req)) {
2362 			/* push queued extents */
2363 			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2364 			scrub_submit(sctx);
2365 			mutex_lock(&sctx->wr_ctx.wr_lock);
2366 			scrub_wr_submit(sctx);
2367 			mutex_unlock(&sctx->wr_ctx.wr_lock);
2368 			wait_event(sctx->list_wait,
2369 				   atomic_read(&sctx->bios_in_flight) == 0);
2370 			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2371 			atomic_inc(&fs_info->scrubs_paused);
2372 			wake_up(&fs_info->scrub_pause_wait);
2373 			mutex_lock(&fs_info->scrub_lock);
2374 			while (atomic_read(&fs_info->scrub_pause_req)) {
2375 				mutex_unlock(&fs_info->scrub_lock);
2376 				wait_event(fs_info->scrub_pause_wait,
2377 				   atomic_read(&fs_info->scrub_pause_req) == 0);
2378 				mutex_lock(&fs_info->scrub_lock);
2379 			}
2380 			atomic_dec(&fs_info->scrubs_paused);
2381 			mutex_unlock(&fs_info->scrub_lock);
2382 			wake_up(&fs_info->scrub_pause_wait);
2383 		}
2384 
2385 		ret = btrfs_lookup_csums_range(csum_root, logical,
2386 					       logical + map->stripe_len - 1,
2387 					       &sctx->csum_list, 1);
2388 		if (ret)
2389 			goto out;
2390 
2391 		key.objectid = logical;
2392 		key.type = BTRFS_EXTENT_ITEM_KEY;
2393 		key.offset = (u64)0;
2394 
2395 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2396 		if (ret < 0)
2397 			goto out;
2398 		if (ret > 0) {
2399 			ret = btrfs_previous_item(root, path, 0,
2400 						  BTRFS_EXTENT_ITEM_KEY);
2401 			if (ret < 0)
2402 				goto out;
2403 			if (ret > 0) {
2404 				/* there's no smaller item, so stick with the
2405 				 * larger one */
2406 				btrfs_release_path(path);
2407 				ret = btrfs_search_slot(NULL, root, &key,
2408 							path, 0, 0);
2409 				if (ret < 0)
2410 					goto out;
2411 			}
2412 		}
2413 
2414 		while (1) {
2415 			l = path->nodes[0];
2416 			slot = path->slots[0];
2417 			if (slot >= btrfs_header_nritems(l)) {
2418 				ret = btrfs_next_leaf(root, path);
2419 				if (ret == 0)
2420 					continue;
2421 				if (ret < 0)
2422 					goto out;
2423 
2424 				break;
2425 			}
2426 			btrfs_item_key_to_cpu(l, &key, slot);
2427 
2428 			if (key.objectid + key.offset <= logical)
2429 				goto next;
2430 
2431 			if (key.objectid >= logical + map->stripe_len)
2432 				break;
2433 
2434 			if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
2435 				goto next;
2436 
2437 			extent = btrfs_item_ptr(l, slot,
2438 						struct btrfs_extent_item);
2439 			flags = btrfs_extent_flags(l, extent);
2440 			generation = btrfs_extent_generation(l, extent);
2441 
2442 			if (key.objectid < logical &&
2443 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2444 				printk(KERN_ERR
2445 				       "btrfs scrub: tree block %llu spanning "
2446 				       "stripes, ignored. logical=%llu\n",
2447 				       (unsigned long long)key.objectid,
2448 				       (unsigned long long)logical);
2449 				goto next;
2450 			}
2451 
2452 			/*
2453 			 * trim extent to this stripe
2454 			 */
2455 			if (key.objectid < logical) {
2456 				key.offset -= logical - key.objectid;
2457 				key.objectid = logical;
2458 			}
2459 			if (key.objectid + key.offset >
2460 			    logical + map->stripe_len) {
2461 				key.offset = logical + map->stripe_len -
2462 					     key.objectid;
2463 			}
2464 
2465 			extent_logical = key.objectid;
2466 			extent_physical = key.objectid - logical + physical;
2467 			extent_len = key.offset;
2468 			extent_dev = scrub_dev;
2469 			extent_mirror_num = mirror_num;
2470 			if (is_dev_replace)
2471 				scrub_remap_extent(fs_info, extent_logical,
2472 						   extent_len, &extent_physical,
2473 						   &extent_dev,
2474 						   &extent_mirror_num);
2475 			ret = scrub_extent(sctx, extent_logical, extent_len,
2476 					   extent_physical, extent_dev, flags,
2477 					   generation, extent_mirror_num,
2478 					   key.objectid - logical + physical);
2479 			if (ret)
2480 				goto out;
2481 
2482 next:
2483 			path->slots[0]++;
2484 		}
2485 		btrfs_release_path(path);
2486 		logical += increment;
2487 		physical += map->stripe_len;
2488 		spin_lock(&sctx->stat_lock);
2489 		sctx->stat.last_physical = physical;
2490 		spin_unlock(&sctx->stat_lock);
2491 	}
2492 out:
2493 	/* push queued extents */
2494 	scrub_submit(sctx);
2495 	mutex_lock(&sctx->wr_ctx.wr_lock);
2496 	scrub_wr_submit(sctx);
2497 	mutex_unlock(&sctx->wr_ctx.wr_lock);
2498 
2499 	blk_finish_plug(&plug);
2500 	btrfs_free_path(path);
2501 	return ret < 0 ? ret : 0;
2502 }
2503 
2504 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2505 					  struct btrfs_device *scrub_dev,
2506 					  u64 chunk_tree, u64 chunk_objectid,
2507 					  u64 chunk_offset, u64 length,
2508 					  u64 dev_offset, int is_dev_replace)
2509 {
2510 	struct btrfs_mapping_tree *map_tree =
2511 		&sctx->dev_root->fs_info->mapping_tree;
2512 	struct map_lookup *map;
2513 	struct extent_map *em;
2514 	int i;
2515 	int ret = 0;
2516 
2517 	read_lock(&map_tree->map_tree.lock);
2518 	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2519 	read_unlock(&map_tree->map_tree.lock);
2520 
2521 	if (!em)
2522 		return -EINVAL;
2523 
2524 	map = (struct map_lookup *)em->bdev;
2525 	if (em->start != chunk_offset)
2526 		goto out;
2527 
2528 	if (em->len < length)
2529 		goto out;
2530 
2531 	for (i = 0; i < map->num_stripes; ++i) {
2532 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2533 		    map->stripes[i].physical == dev_offset) {
2534 			ret = scrub_stripe(sctx, map, scrub_dev, i,
2535 					   chunk_offset, length,
2536 					   is_dev_replace);
2537 			if (ret)
2538 				goto out;
2539 		}
2540 	}
2541 out:
2542 	free_extent_map(em);
2543 
2544 	return ret;
2545 }
2546 
2547 static noinline_for_stack
2548 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2549 			   struct btrfs_device *scrub_dev, u64 start, u64 end,
2550 			   int is_dev_replace)
2551 {
2552 	struct btrfs_dev_extent *dev_extent = NULL;
2553 	struct btrfs_path *path;
2554 	struct btrfs_root *root = sctx->dev_root;
2555 	struct btrfs_fs_info *fs_info = root->fs_info;
2556 	u64 length;
2557 	u64 chunk_tree;
2558 	u64 chunk_objectid;
2559 	u64 chunk_offset;
2560 	int ret;
2561 	int slot;
2562 	struct extent_buffer *l;
2563 	struct btrfs_key key;
2564 	struct btrfs_key found_key;
2565 	struct btrfs_block_group_cache *cache;
2566 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2567 
2568 	path = btrfs_alloc_path();
2569 	if (!path)
2570 		return -ENOMEM;
2571 
2572 	path->reada = 2;
2573 	path->search_commit_root = 1;
2574 	path->skip_locking = 1;
2575 
2576 	key.objectid = scrub_dev->devid;
2577 	key.offset = 0ull;
2578 	key.type = BTRFS_DEV_EXTENT_KEY;
2579 
2580 	while (1) {
2581 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2582 		if (ret < 0)
2583 			break;
2584 		if (ret > 0) {
2585 			if (path->slots[0] >=
2586 			    btrfs_header_nritems(path->nodes[0])) {
2587 				ret = btrfs_next_leaf(root, path);
2588 				if (ret)
2589 					break;
2590 			}
2591 		}
2592 
2593 		l = path->nodes[0];
2594 		slot = path->slots[0];
2595 
2596 		btrfs_item_key_to_cpu(l, &found_key, slot);
2597 
2598 		if (found_key.objectid != scrub_dev->devid)
2599 			break;
2600 
2601 		if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2602 			break;
2603 
2604 		if (found_key.offset >= end)
2605 			break;
2606 
2607 		if (found_key.offset < key.offset)
2608 			break;
2609 
2610 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2611 		length = btrfs_dev_extent_length(l, dev_extent);
2612 
2613 		if (found_key.offset + length <= start) {
2614 			key.offset = found_key.offset + length;
2615 			btrfs_release_path(path);
2616 			continue;
2617 		}
2618 
2619 		chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2620 		chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2621 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2622 
2623 		/*
2624 		 * get a reference on the corresponding block group to prevent
2625 		 * the chunk from going away while we scrub it
2626 		 */
2627 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2628 		if (!cache) {
2629 			ret = -ENOENT;
2630 			break;
2631 		}
2632 		dev_replace->cursor_right = found_key.offset + length;
2633 		dev_replace->cursor_left = found_key.offset;
2634 		dev_replace->item_needs_writeback = 1;
2635 		ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2636 				  chunk_offset, length, found_key.offset,
2637 				  is_dev_replace);
2638 
2639 		/*
2640 		 * flush, submit all pending read and write bios, afterwards
2641 		 * wait for them.
2642 		 * Note that in the dev replace case, a read request causes
2643 		 * write requests that are submitted in the read completion
2644 		 * worker. Therefore in the current situation, it is required
2645 		 * that all write requests are flushed, so that all read and
2646 		 * write requests are really completed when bios_in_flight
2647 		 * changes to 0.
2648 		 */
2649 		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2650 		scrub_submit(sctx);
2651 		mutex_lock(&sctx->wr_ctx.wr_lock);
2652 		scrub_wr_submit(sctx);
2653 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2654 
2655 		wait_event(sctx->list_wait,
2656 			   atomic_read(&sctx->bios_in_flight) == 0);
2657 		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2658 		atomic_inc(&fs_info->scrubs_paused);
2659 		wake_up(&fs_info->scrub_pause_wait);
2660 		wait_event(sctx->list_wait,
2661 			   atomic_read(&sctx->workers_pending) == 0);
2662 
2663 		mutex_lock(&fs_info->scrub_lock);
2664 		while (atomic_read(&fs_info->scrub_pause_req)) {
2665 			mutex_unlock(&fs_info->scrub_lock);
2666 			wait_event(fs_info->scrub_pause_wait,
2667 			   atomic_read(&fs_info->scrub_pause_req) == 0);
2668 			mutex_lock(&fs_info->scrub_lock);
2669 		}
2670 		atomic_dec(&fs_info->scrubs_paused);
2671 		mutex_unlock(&fs_info->scrub_lock);
2672 		wake_up(&fs_info->scrub_pause_wait);
2673 
2674 		dev_replace->cursor_left = dev_replace->cursor_right;
2675 		dev_replace->item_needs_writeback = 1;
2676 		btrfs_put_block_group(cache);
2677 		if (ret)
2678 			break;
2679 		if (is_dev_replace &&
2680 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
2681 			ret = -EIO;
2682 			break;
2683 		}
2684 		if (sctx->stat.malloc_errors > 0) {
2685 			ret = -ENOMEM;
2686 			break;
2687 		}
2688 
2689 		key.offset = found_key.offset + length;
2690 		btrfs_release_path(path);
2691 	}
2692 
2693 	btrfs_free_path(path);
2694 
2695 	/*
2696 	 * ret can still be 1 from search_slot or next_leaf,
2697 	 * that's not an error
2698 	 */
2699 	return ret < 0 ? ret : 0;
2700 }
2701 
2702 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2703 					   struct btrfs_device *scrub_dev)
2704 {
2705 	int	i;
2706 	u64	bytenr;
2707 	u64	gen;
2708 	int	ret;
2709 	struct btrfs_root *root = sctx->dev_root;
2710 
2711 	if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2712 		return -EIO;
2713 
2714 	gen = root->fs_info->last_trans_committed;
2715 
2716 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2717 		bytenr = btrfs_sb_offset(i);
2718 		if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2719 			break;
2720 
2721 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2722 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2723 				  NULL, 1, bytenr);
2724 		if (ret)
2725 			return ret;
2726 	}
2727 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2728 
2729 	return 0;
2730 }
2731 
2732 /*
2733  * get a reference count on fs_info->scrub_workers. start worker if necessary
2734  */
2735 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2736 						int is_dev_replace)
2737 {
2738 	int ret = 0;
2739 
2740 	mutex_lock(&fs_info->scrub_lock);
2741 	if (fs_info->scrub_workers_refcnt == 0) {
2742 		if (is_dev_replace)
2743 			btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2744 					&fs_info->generic_worker);
2745 		else
2746 			btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2747 					fs_info->thread_pool_size,
2748 					&fs_info->generic_worker);
2749 		fs_info->scrub_workers.idle_thresh = 4;
2750 		ret = btrfs_start_workers(&fs_info->scrub_workers);
2751 		if (ret)
2752 			goto out;
2753 		btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2754 				   "scrubwrc",
2755 				   fs_info->thread_pool_size,
2756 				   &fs_info->generic_worker);
2757 		fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2758 		ret = btrfs_start_workers(
2759 				&fs_info->scrub_wr_completion_workers);
2760 		if (ret)
2761 			goto out;
2762 		btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2763 				   &fs_info->generic_worker);
2764 		ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2765 		if (ret)
2766 			goto out;
2767 	}
2768 	++fs_info->scrub_workers_refcnt;
2769 out:
2770 	mutex_unlock(&fs_info->scrub_lock);
2771 
2772 	return ret;
2773 }
2774 
2775 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2776 {
2777 	mutex_lock(&fs_info->scrub_lock);
2778 	if (--fs_info->scrub_workers_refcnt == 0) {
2779 		btrfs_stop_workers(&fs_info->scrub_workers);
2780 		btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2781 		btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2782 	}
2783 	WARN_ON(fs_info->scrub_workers_refcnt < 0);
2784 	mutex_unlock(&fs_info->scrub_lock);
2785 }
2786 
2787 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2788 		    u64 end, struct btrfs_scrub_progress *progress,
2789 		    int readonly, int is_dev_replace)
2790 {
2791 	struct scrub_ctx *sctx;
2792 	int ret;
2793 	struct btrfs_device *dev;
2794 
2795 	if (btrfs_fs_closing(fs_info))
2796 		return -EINVAL;
2797 
2798 	/*
2799 	 * check some assumptions
2800 	 */
2801 	if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2802 		printk(KERN_ERR
2803 		       "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2804 		       fs_info->chunk_root->nodesize,
2805 		       fs_info->chunk_root->leafsize);
2806 		return -EINVAL;
2807 	}
2808 
2809 	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2810 		/*
2811 		 * in this case scrub is unable to calculate the checksum
2812 		 * the way scrub is implemented. Do not handle this
2813 		 * situation at all because it won't ever happen.
2814 		 */
2815 		printk(KERN_ERR
2816 		       "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2817 		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2818 		return -EINVAL;
2819 	}
2820 
2821 	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2822 		/* not supported for data w/o checksums */
2823 		printk(KERN_ERR
2824 		       "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2825 		       fs_info->chunk_root->sectorsize,
2826 		       (unsigned long long)PAGE_SIZE);
2827 		return -EINVAL;
2828 	}
2829 
2830 	if (fs_info->chunk_root->nodesize >
2831 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2832 	    fs_info->chunk_root->sectorsize >
2833 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2834 		/*
2835 		 * would exhaust the array bounds of pagev member in
2836 		 * struct scrub_block
2837 		 */
2838 		pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2839 		       fs_info->chunk_root->nodesize,
2840 		       SCRUB_MAX_PAGES_PER_BLOCK,
2841 		       fs_info->chunk_root->sectorsize,
2842 		       SCRUB_MAX_PAGES_PER_BLOCK);
2843 		return -EINVAL;
2844 	}
2845 
2846 	ret = scrub_workers_get(fs_info, is_dev_replace);
2847 	if (ret)
2848 		return ret;
2849 
2850 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2851 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2852 	if (!dev || (dev->missing && !is_dev_replace)) {
2853 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2854 		scrub_workers_put(fs_info);
2855 		return -ENODEV;
2856 	}
2857 	mutex_lock(&fs_info->scrub_lock);
2858 
2859 	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2860 		mutex_unlock(&fs_info->scrub_lock);
2861 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2862 		scrub_workers_put(fs_info);
2863 		return -EIO;
2864 	}
2865 
2866 	btrfs_dev_replace_lock(&fs_info->dev_replace);
2867 	if (dev->scrub_device ||
2868 	    (!is_dev_replace &&
2869 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2870 		btrfs_dev_replace_unlock(&fs_info->dev_replace);
2871 		mutex_unlock(&fs_info->scrub_lock);
2872 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2873 		scrub_workers_put(fs_info);
2874 		return -EINPROGRESS;
2875 	}
2876 	btrfs_dev_replace_unlock(&fs_info->dev_replace);
2877 	sctx = scrub_setup_ctx(dev, is_dev_replace);
2878 	if (IS_ERR(sctx)) {
2879 		mutex_unlock(&fs_info->scrub_lock);
2880 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2881 		scrub_workers_put(fs_info);
2882 		return PTR_ERR(sctx);
2883 	}
2884 	sctx->readonly = readonly;
2885 	dev->scrub_device = sctx;
2886 
2887 	atomic_inc(&fs_info->scrubs_running);
2888 	mutex_unlock(&fs_info->scrub_lock);
2889 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2890 
2891 	if (!is_dev_replace) {
2892 		down_read(&fs_info->scrub_super_lock);
2893 		ret = scrub_supers(sctx, dev);
2894 		up_read(&fs_info->scrub_super_lock);
2895 	}
2896 
2897 	if (!ret)
2898 		ret = scrub_enumerate_chunks(sctx, dev, start, end,
2899 					     is_dev_replace);
2900 
2901 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2902 	atomic_dec(&fs_info->scrubs_running);
2903 	wake_up(&fs_info->scrub_pause_wait);
2904 
2905 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2906 
2907 	if (progress)
2908 		memcpy(progress, &sctx->stat, sizeof(*progress));
2909 
2910 	mutex_lock(&fs_info->scrub_lock);
2911 	dev->scrub_device = NULL;
2912 	mutex_unlock(&fs_info->scrub_lock);
2913 
2914 	scrub_free_ctx(sctx);
2915 	scrub_workers_put(fs_info);
2916 
2917 	return ret;
2918 }
2919 
2920 void btrfs_scrub_pause(struct btrfs_root *root)
2921 {
2922 	struct btrfs_fs_info *fs_info = root->fs_info;
2923 
2924 	mutex_lock(&fs_info->scrub_lock);
2925 	atomic_inc(&fs_info->scrub_pause_req);
2926 	while (atomic_read(&fs_info->scrubs_paused) !=
2927 	       atomic_read(&fs_info->scrubs_running)) {
2928 		mutex_unlock(&fs_info->scrub_lock);
2929 		wait_event(fs_info->scrub_pause_wait,
2930 			   atomic_read(&fs_info->scrubs_paused) ==
2931 			   atomic_read(&fs_info->scrubs_running));
2932 		mutex_lock(&fs_info->scrub_lock);
2933 	}
2934 	mutex_unlock(&fs_info->scrub_lock);
2935 }
2936 
2937 void btrfs_scrub_continue(struct btrfs_root *root)
2938 {
2939 	struct btrfs_fs_info *fs_info = root->fs_info;
2940 
2941 	atomic_dec(&fs_info->scrub_pause_req);
2942 	wake_up(&fs_info->scrub_pause_wait);
2943 }
2944 
2945 void btrfs_scrub_pause_super(struct btrfs_root *root)
2946 {
2947 	down_write(&root->fs_info->scrub_super_lock);
2948 }
2949 
2950 void btrfs_scrub_continue_super(struct btrfs_root *root)
2951 {
2952 	up_write(&root->fs_info->scrub_super_lock);
2953 }
2954 
2955 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2956 {
2957 	mutex_lock(&fs_info->scrub_lock);
2958 	if (!atomic_read(&fs_info->scrubs_running)) {
2959 		mutex_unlock(&fs_info->scrub_lock);
2960 		return -ENOTCONN;
2961 	}
2962 
2963 	atomic_inc(&fs_info->scrub_cancel_req);
2964 	while (atomic_read(&fs_info->scrubs_running)) {
2965 		mutex_unlock(&fs_info->scrub_lock);
2966 		wait_event(fs_info->scrub_pause_wait,
2967 			   atomic_read(&fs_info->scrubs_running) == 0);
2968 		mutex_lock(&fs_info->scrub_lock);
2969 	}
2970 	atomic_dec(&fs_info->scrub_cancel_req);
2971 	mutex_unlock(&fs_info->scrub_lock);
2972 
2973 	return 0;
2974 }
2975 
2976 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2977 			   struct btrfs_device *dev)
2978 {
2979 	struct scrub_ctx *sctx;
2980 
2981 	mutex_lock(&fs_info->scrub_lock);
2982 	sctx = dev->scrub_device;
2983 	if (!sctx) {
2984 		mutex_unlock(&fs_info->scrub_lock);
2985 		return -ENOTCONN;
2986 	}
2987 	atomic_inc(&sctx->cancel_req);
2988 	while (dev->scrub_device) {
2989 		mutex_unlock(&fs_info->scrub_lock);
2990 		wait_event(fs_info->scrub_pause_wait,
2991 			   dev->scrub_device == NULL);
2992 		mutex_lock(&fs_info->scrub_lock);
2993 	}
2994 	mutex_unlock(&fs_info->scrub_lock);
2995 
2996 	return 0;
2997 }
2998 
2999 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
3000 {
3001 	struct btrfs_fs_info *fs_info = root->fs_info;
3002 	struct btrfs_device *dev;
3003 	int ret;
3004 
3005 	/*
3006 	 * we have to hold the device_list_mutex here so the device
3007 	 * does not go away in cancel_dev. FIXME: find a better solution
3008 	 */
3009 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3010 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3011 	if (!dev) {
3012 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3013 		return -ENODEV;
3014 	}
3015 	ret = btrfs_scrub_cancel_dev(fs_info, dev);
3016 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3017 
3018 	return ret;
3019 }
3020 
3021 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3022 			 struct btrfs_scrub_progress *progress)
3023 {
3024 	struct btrfs_device *dev;
3025 	struct scrub_ctx *sctx = NULL;
3026 
3027 	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3028 	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3029 	if (dev)
3030 		sctx = dev->scrub_device;
3031 	if (sctx)
3032 		memcpy(progress, &sctx->stat, sizeof(*progress));
3033 	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3034 
3035 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3036 }
3037 
3038 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3039 			       u64 extent_logical, u64 extent_len,
3040 			       u64 *extent_physical,
3041 			       struct btrfs_device **extent_dev,
3042 			       int *extent_mirror_num)
3043 {
3044 	u64 mapped_length;
3045 	struct btrfs_bio *bbio = NULL;
3046 	int ret;
3047 
3048 	mapped_length = extent_len;
3049 	ret = btrfs_map_block(fs_info, READ, extent_logical,
3050 			      &mapped_length, &bbio, 0);
3051 	if (ret || !bbio || mapped_length < extent_len ||
3052 	    !bbio->stripes[0].dev->bdev) {
3053 		kfree(bbio);
3054 		return;
3055 	}
3056 
3057 	*extent_physical = bbio->stripes[0].physical;
3058 	*extent_mirror_num = bbio->mirror_num;
3059 	*extent_dev = bbio->stripes[0].dev;
3060 	kfree(bbio);
3061 }
3062 
3063 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3064 			      struct scrub_wr_ctx *wr_ctx,
3065 			      struct btrfs_fs_info *fs_info,
3066 			      struct btrfs_device *dev,
3067 			      int is_dev_replace)
3068 {
3069 	WARN_ON(wr_ctx->wr_curr_bio != NULL);
3070 
3071 	mutex_init(&wr_ctx->wr_lock);
3072 	wr_ctx->wr_curr_bio = NULL;
3073 	if (!is_dev_replace)
3074 		return 0;
3075 
3076 	WARN_ON(!dev->bdev);
3077 	wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3078 					 bio_get_nr_vecs(dev->bdev));
3079 	wr_ctx->tgtdev = dev;
3080 	atomic_set(&wr_ctx->flush_all_writes, 0);
3081 	return 0;
3082 }
3083 
3084 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3085 {
3086 	mutex_lock(&wr_ctx->wr_lock);
3087 	kfree(wr_ctx->wr_curr_bio);
3088 	wr_ctx->wr_curr_bio = NULL;
3089 	mutex_unlock(&wr_ctx->wr_lock);
3090 }
3091 
3092 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3093 			    int mirror_num, u64 physical_for_dev_replace)
3094 {
3095 	struct scrub_copy_nocow_ctx *nocow_ctx;
3096 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3097 
3098 	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3099 	if (!nocow_ctx) {
3100 		spin_lock(&sctx->stat_lock);
3101 		sctx->stat.malloc_errors++;
3102 		spin_unlock(&sctx->stat_lock);
3103 		return -ENOMEM;
3104 	}
3105 
3106 	scrub_pending_trans_workers_inc(sctx);
3107 
3108 	nocow_ctx->sctx = sctx;
3109 	nocow_ctx->logical = logical;
3110 	nocow_ctx->len = len;
3111 	nocow_ctx->mirror_num = mirror_num;
3112 	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3113 	nocow_ctx->work.func = copy_nocow_pages_worker;
3114 	btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3115 			   &nocow_ctx->work);
3116 
3117 	return 0;
3118 }
3119 
3120 static void copy_nocow_pages_worker(struct btrfs_work *work)
3121 {
3122 	struct scrub_copy_nocow_ctx *nocow_ctx =
3123 		container_of(work, struct scrub_copy_nocow_ctx, work);
3124 	struct scrub_ctx *sctx = nocow_ctx->sctx;
3125 	u64 logical = nocow_ctx->logical;
3126 	u64 len = nocow_ctx->len;
3127 	int mirror_num = nocow_ctx->mirror_num;
3128 	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3129 	int ret;
3130 	struct btrfs_trans_handle *trans = NULL;
3131 	struct btrfs_fs_info *fs_info;
3132 	struct btrfs_path *path;
3133 	struct btrfs_root *root;
3134 	int not_written = 0;
3135 
3136 	fs_info = sctx->dev_root->fs_info;
3137 	root = fs_info->extent_root;
3138 
3139 	path = btrfs_alloc_path();
3140 	if (!path) {
3141 		spin_lock(&sctx->stat_lock);
3142 		sctx->stat.malloc_errors++;
3143 		spin_unlock(&sctx->stat_lock);
3144 		not_written = 1;
3145 		goto out;
3146 	}
3147 
3148 	trans = btrfs_join_transaction(root);
3149 	if (IS_ERR(trans)) {
3150 		not_written = 1;
3151 		goto out;
3152 	}
3153 
3154 	ret = iterate_inodes_from_logical(logical, fs_info, path,
3155 					  copy_nocow_pages_for_inode,
3156 					  nocow_ctx);
3157 	if (ret != 0 && ret != -ENOENT) {
3158 		pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3159 			(unsigned long long)logical,
3160 			(unsigned long long)physical_for_dev_replace,
3161 			(unsigned long long)len,
3162 			(unsigned long long)mirror_num, ret);
3163 		not_written = 1;
3164 		goto out;
3165 	}
3166 
3167 out:
3168 	if (trans && !IS_ERR(trans))
3169 		btrfs_end_transaction(trans, root);
3170 	if (not_written)
3171 		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3172 					    num_uncorrectable_read_errors);
3173 
3174 	btrfs_free_path(path);
3175 	kfree(nocow_ctx);
3176 
3177 	scrub_pending_trans_workers_dec(sctx);
3178 }
3179 
3180 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3181 {
3182 	unsigned long index;
3183 	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3184 	int ret = 0;
3185 	struct btrfs_key key;
3186 	struct inode *inode = NULL;
3187 	struct btrfs_root *local_root;
3188 	u64 physical_for_dev_replace;
3189 	u64 len;
3190 	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3191 	int srcu_index;
3192 
3193 	key.objectid = root;
3194 	key.type = BTRFS_ROOT_ITEM_KEY;
3195 	key.offset = (u64)-1;
3196 
3197 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3198 
3199 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3200 	if (IS_ERR(local_root)) {
3201 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3202 		return PTR_ERR(local_root);
3203 	}
3204 
3205 	key.type = BTRFS_INODE_ITEM_KEY;
3206 	key.objectid = inum;
3207 	key.offset = 0;
3208 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3209 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3210 	if (IS_ERR(inode))
3211 		return PTR_ERR(inode);
3212 
3213 	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3214 	len = nocow_ctx->len;
3215 	while (len >= PAGE_CACHE_SIZE) {
3216 		struct page *page = NULL;
3217 		int ret_sub;
3218 
3219 		index = offset >> PAGE_CACHE_SHIFT;
3220 
3221 		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3222 		if (!page) {
3223 			pr_err("find_or_create_page() failed\n");
3224 			ret = -ENOMEM;
3225 			goto next_page;
3226 		}
3227 
3228 		if (PageUptodate(page)) {
3229 			if (PageDirty(page))
3230 				goto next_page;
3231 		} else {
3232 			ClearPageError(page);
3233 			ret_sub = extent_read_full_page(&BTRFS_I(inode)->
3234 							 io_tree,
3235 							page, btrfs_get_extent,
3236 							nocow_ctx->mirror_num);
3237 			if (ret_sub) {
3238 				ret = ret_sub;
3239 				goto next_page;
3240 			}
3241 			wait_on_page_locked(page);
3242 			if (!PageUptodate(page)) {
3243 				ret = -EIO;
3244 				goto next_page;
3245 			}
3246 		}
3247 		ret_sub = write_page_nocow(nocow_ctx->sctx,
3248 					   physical_for_dev_replace, page);
3249 		if (ret_sub) {
3250 			ret = ret_sub;
3251 			goto next_page;
3252 		}
3253 
3254 next_page:
3255 		if (page) {
3256 			unlock_page(page);
3257 			put_page(page);
3258 		}
3259 		offset += PAGE_CACHE_SIZE;
3260 		physical_for_dev_replace += PAGE_CACHE_SIZE;
3261 		len -= PAGE_CACHE_SIZE;
3262 	}
3263 
3264 	if (inode)
3265 		iput(inode);
3266 	return ret;
3267 }
3268 
3269 static int write_page_nocow(struct scrub_ctx *sctx,
3270 			    u64 physical_for_dev_replace, struct page *page)
3271 {
3272 	struct bio *bio;
3273 	struct btrfs_device *dev;
3274 	int ret;
3275 	DECLARE_COMPLETION_ONSTACK(compl);
3276 
3277 	dev = sctx->wr_ctx.tgtdev;
3278 	if (!dev)
3279 		return -EIO;
3280 	if (!dev->bdev) {
3281 		printk_ratelimited(KERN_WARNING
3282 			"btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3283 		return -EIO;
3284 	}
3285 	bio = bio_alloc(GFP_NOFS, 1);
3286 	if (!bio) {
3287 		spin_lock(&sctx->stat_lock);
3288 		sctx->stat.malloc_errors++;
3289 		spin_unlock(&sctx->stat_lock);
3290 		return -ENOMEM;
3291 	}
3292 	bio->bi_private = &compl;
3293 	bio->bi_end_io = scrub_complete_bio_end_io;
3294 	bio->bi_size = 0;
3295 	bio->bi_sector = physical_for_dev_replace >> 9;
3296 	bio->bi_bdev = dev->bdev;
3297 	ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3298 	if (ret != PAGE_CACHE_SIZE) {
3299 leave_with_eio:
3300 		bio_put(bio);
3301 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3302 		return -EIO;
3303 	}
3304 	btrfsic_submit_bio(WRITE_SYNC, bio);
3305 	wait_for_completion(&compl);
3306 
3307 	if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3308 		goto leave_with_eio;
3309 
3310 	bio_put(bio);
3311 	return 0;
3312 }
3313