xref: /linux/fs/btrfs/scrub.c (revision b889fcf63cb62e7fdb7816565e28f44dbe4a76a5)
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 	u64 end = offset + PAGE_SIZE - 1;
584 	struct btrfs_root *local_root;
585 
586 	key.objectid = root;
587 	key.type = BTRFS_ROOT_ITEM_KEY;
588 	key.offset = (u64)-1;
589 	local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
590 	if (IS_ERR(local_root))
591 		return PTR_ERR(local_root);
592 
593 	key.type = BTRFS_INODE_ITEM_KEY;
594 	key.objectid = inum;
595 	key.offset = 0;
596 	inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
597 	if (IS_ERR(inode))
598 		return PTR_ERR(inode);
599 
600 	index = offset >> PAGE_CACHE_SHIFT;
601 
602 	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
603 	if (!page) {
604 		ret = -ENOMEM;
605 		goto out;
606 	}
607 
608 	if (PageUptodate(page)) {
609 		struct btrfs_fs_info *fs_info;
610 		if (PageDirty(page)) {
611 			/*
612 			 * we need to write the data to the defect sector. the
613 			 * data that was in that sector is not in memory,
614 			 * because the page was modified. we must not write the
615 			 * modified page to that sector.
616 			 *
617 			 * TODO: what could be done here: wait for the delalloc
618 			 *       runner to write out that page (might involve
619 			 *       COW) and see whether the sector is still
620 			 *       referenced afterwards.
621 			 *
622 			 * For the meantime, we'll treat this error
623 			 * incorrectable, although there is a chance that a
624 			 * later scrub will find the bad sector again and that
625 			 * there's no dirty page in memory, then.
626 			 */
627 			ret = -EIO;
628 			goto out;
629 		}
630 		fs_info = BTRFS_I(inode)->root->fs_info;
631 		ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
632 					fixup->logical, page,
633 					fixup->mirror_num);
634 		unlock_page(page);
635 		corrected = !ret;
636 	} else {
637 		/*
638 		 * we need to get good data first. the general readpage path
639 		 * will call repair_io_failure for us, we just have to make
640 		 * sure we read the bad mirror.
641 		 */
642 		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
643 					EXTENT_DAMAGED, GFP_NOFS);
644 		if (ret) {
645 			/* set_extent_bits should give proper error */
646 			WARN_ON(ret > 0);
647 			if (ret > 0)
648 				ret = -EFAULT;
649 			goto out;
650 		}
651 
652 		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
653 						btrfs_get_extent,
654 						fixup->mirror_num);
655 		wait_on_page_locked(page);
656 
657 		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
658 						end, EXTENT_DAMAGED, 0, NULL);
659 		if (!corrected)
660 			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
661 						EXTENT_DAMAGED, GFP_NOFS);
662 	}
663 
664 out:
665 	if (page)
666 		put_page(page);
667 	if (inode)
668 		iput(inode);
669 
670 	if (ret < 0)
671 		return ret;
672 
673 	if (ret == 0 && corrected) {
674 		/*
675 		 * we only need to call readpage for one of the inodes belonging
676 		 * to this extent. so make iterate_extent_inodes stop
677 		 */
678 		return 1;
679 	}
680 
681 	return -EIO;
682 }
683 
684 static void scrub_fixup_nodatasum(struct btrfs_work *work)
685 {
686 	int ret;
687 	struct scrub_fixup_nodatasum *fixup;
688 	struct scrub_ctx *sctx;
689 	struct btrfs_trans_handle *trans = NULL;
690 	struct btrfs_fs_info *fs_info;
691 	struct btrfs_path *path;
692 	int uncorrectable = 0;
693 
694 	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
695 	sctx = fixup->sctx;
696 	fs_info = fixup->root->fs_info;
697 
698 	path = btrfs_alloc_path();
699 	if (!path) {
700 		spin_lock(&sctx->stat_lock);
701 		++sctx->stat.malloc_errors;
702 		spin_unlock(&sctx->stat_lock);
703 		uncorrectable = 1;
704 		goto out;
705 	}
706 
707 	trans = btrfs_join_transaction(fixup->root);
708 	if (IS_ERR(trans)) {
709 		uncorrectable = 1;
710 		goto out;
711 	}
712 
713 	/*
714 	 * the idea is to trigger a regular read through the standard path. we
715 	 * read a page from the (failed) logical address by specifying the
716 	 * corresponding copynum of the failed sector. thus, that readpage is
717 	 * expected to fail.
718 	 * that is the point where on-the-fly error correction will kick in
719 	 * (once it's finished) and rewrite the failed sector if a good copy
720 	 * can be found.
721 	 */
722 	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
723 						path, scrub_fixup_readpage,
724 						fixup);
725 	if (ret < 0) {
726 		uncorrectable = 1;
727 		goto out;
728 	}
729 	WARN_ON(ret != 1);
730 
731 	spin_lock(&sctx->stat_lock);
732 	++sctx->stat.corrected_errors;
733 	spin_unlock(&sctx->stat_lock);
734 
735 out:
736 	if (trans && !IS_ERR(trans))
737 		btrfs_end_transaction(trans, fixup->root);
738 	if (uncorrectable) {
739 		spin_lock(&sctx->stat_lock);
740 		++sctx->stat.uncorrectable_errors;
741 		spin_unlock(&sctx->stat_lock);
742 		btrfs_dev_replace_stats_inc(
743 			&sctx->dev_root->fs_info->dev_replace.
744 			num_uncorrectable_read_errors);
745 		printk_ratelimited_in_rcu(KERN_ERR
746 			"btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
747 			(unsigned long long)fixup->logical,
748 			rcu_str_deref(fixup->dev->name));
749 	}
750 
751 	btrfs_free_path(path);
752 	kfree(fixup);
753 
754 	scrub_pending_trans_workers_dec(sctx);
755 }
756 
757 /*
758  * scrub_handle_errored_block gets called when either verification of the
759  * pages failed or the bio failed to read, e.g. with EIO. In the latter
760  * case, this function handles all pages in the bio, even though only one
761  * may be bad.
762  * The goal of this function is to repair the errored block by using the
763  * contents of one of the mirrors.
764  */
765 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
766 {
767 	struct scrub_ctx *sctx = sblock_to_check->sctx;
768 	struct btrfs_device *dev;
769 	struct btrfs_fs_info *fs_info;
770 	u64 length;
771 	u64 logical;
772 	u64 generation;
773 	unsigned int failed_mirror_index;
774 	unsigned int is_metadata;
775 	unsigned int have_csum;
776 	u8 *csum;
777 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
778 	struct scrub_block *sblock_bad;
779 	int ret;
780 	int mirror_index;
781 	int page_num;
782 	int success;
783 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
784 				      DEFAULT_RATELIMIT_BURST);
785 
786 	BUG_ON(sblock_to_check->page_count < 1);
787 	fs_info = sctx->dev_root->fs_info;
788 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
789 		/*
790 		 * if we find an error in a super block, we just report it.
791 		 * They will get written with the next transaction commit
792 		 * anyway
793 		 */
794 		spin_lock(&sctx->stat_lock);
795 		++sctx->stat.super_errors;
796 		spin_unlock(&sctx->stat_lock);
797 		return 0;
798 	}
799 	length = sblock_to_check->page_count * PAGE_SIZE;
800 	logical = sblock_to_check->pagev[0]->logical;
801 	generation = sblock_to_check->pagev[0]->generation;
802 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
803 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
804 	is_metadata = !(sblock_to_check->pagev[0]->flags &
805 			BTRFS_EXTENT_FLAG_DATA);
806 	have_csum = sblock_to_check->pagev[0]->have_csum;
807 	csum = sblock_to_check->pagev[0]->csum;
808 	dev = sblock_to_check->pagev[0]->dev;
809 
810 	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
811 		sblocks_for_recheck = NULL;
812 		goto nodatasum_case;
813 	}
814 
815 	/*
816 	 * read all mirrors one after the other. This includes to
817 	 * re-read the extent or metadata block that failed (that was
818 	 * the cause that this fixup code is called) another time,
819 	 * page by page this time in order to know which pages
820 	 * caused I/O errors and which ones are good (for all mirrors).
821 	 * It is the goal to handle the situation when more than one
822 	 * mirror contains I/O errors, but the errors do not
823 	 * overlap, i.e. the data can be repaired by selecting the
824 	 * pages from those mirrors without I/O error on the
825 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
826 	 * would be that mirror #1 has an I/O error on the first page,
827 	 * the second page is good, and mirror #2 has an I/O error on
828 	 * the second page, but the first page is good.
829 	 * Then the first page of the first mirror can be repaired by
830 	 * taking the first page of the second mirror, and the
831 	 * second page of the second mirror can be repaired by
832 	 * copying the contents of the 2nd page of the 1st mirror.
833 	 * One more note: if the pages of one mirror contain I/O
834 	 * errors, the checksum cannot be verified. In order to get
835 	 * the best data for repairing, the first attempt is to find
836 	 * a mirror without I/O errors and with a validated checksum.
837 	 * Only if this is not possible, the pages are picked from
838 	 * mirrors with I/O errors without considering the checksum.
839 	 * If the latter is the case, at the end, the checksum of the
840 	 * repaired area is verified in order to correctly maintain
841 	 * the statistics.
842 	 */
843 
844 	sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
845 				     sizeof(*sblocks_for_recheck),
846 				     GFP_NOFS);
847 	if (!sblocks_for_recheck) {
848 		spin_lock(&sctx->stat_lock);
849 		sctx->stat.malloc_errors++;
850 		sctx->stat.read_errors++;
851 		sctx->stat.uncorrectable_errors++;
852 		spin_unlock(&sctx->stat_lock);
853 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
854 		goto out;
855 	}
856 
857 	/* setup the context, map the logical blocks and alloc the pages */
858 	ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
859 					logical, sblocks_for_recheck);
860 	if (ret) {
861 		spin_lock(&sctx->stat_lock);
862 		sctx->stat.read_errors++;
863 		sctx->stat.uncorrectable_errors++;
864 		spin_unlock(&sctx->stat_lock);
865 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
866 		goto out;
867 	}
868 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
869 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
870 
871 	/* build and submit the bios for the failed mirror, check checksums */
872 	scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
873 			    csum, generation, sctx->csum_size);
874 
875 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
876 	    sblock_bad->no_io_error_seen) {
877 		/*
878 		 * the error disappeared after reading page by page, or
879 		 * the area was part of a huge bio and other parts of the
880 		 * bio caused I/O errors, or the block layer merged several
881 		 * read requests into one and the error is caused by a
882 		 * different bio (usually one of the two latter cases is
883 		 * the cause)
884 		 */
885 		spin_lock(&sctx->stat_lock);
886 		sctx->stat.unverified_errors++;
887 		spin_unlock(&sctx->stat_lock);
888 
889 		if (sctx->is_dev_replace)
890 			scrub_write_block_to_dev_replace(sblock_bad);
891 		goto out;
892 	}
893 
894 	if (!sblock_bad->no_io_error_seen) {
895 		spin_lock(&sctx->stat_lock);
896 		sctx->stat.read_errors++;
897 		spin_unlock(&sctx->stat_lock);
898 		if (__ratelimit(&_rs))
899 			scrub_print_warning("i/o error", sblock_to_check);
900 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
901 	} else if (sblock_bad->checksum_error) {
902 		spin_lock(&sctx->stat_lock);
903 		sctx->stat.csum_errors++;
904 		spin_unlock(&sctx->stat_lock);
905 		if (__ratelimit(&_rs))
906 			scrub_print_warning("checksum error", sblock_to_check);
907 		btrfs_dev_stat_inc_and_print(dev,
908 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
909 	} else if (sblock_bad->header_error) {
910 		spin_lock(&sctx->stat_lock);
911 		sctx->stat.verify_errors++;
912 		spin_unlock(&sctx->stat_lock);
913 		if (__ratelimit(&_rs))
914 			scrub_print_warning("checksum/header error",
915 					    sblock_to_check);
916 		if (sblock_bad->generation_error)
917 			btrfs_dev_stat_inc_and_print(dev,
918 				BTRFS_DEV_STAT_GENERATION_ERRS);
919 		else
920 			btrfs_dev_stat_inc_and_print(dev,
921 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
922 	}
923 
924 	if (sctx->readonly && !sctx->is_dev_replace)
925 		goto did_not_correct_error;
926 
927 	if (!is_metadata && !have_csum) {
928 		struct scrub_fixup_nodatasum *fixup_nodatasum;
929 
930 nodatasum_case:
931 		WARN_ON(sctx->is_dev_replace);
932 
933 		/*
934 		 * !is_metadata and !have_csum, this means that the data
935 		 * might not be COW'ed, that it might be modified
936 		 * concurrently. The general strategy to work on the
937 		 * commit root does not help in the case when COW is not
938 		 * used.
939 		 */
940 		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
941 		if (!fixup_nodatasum)
942 			goto did_not_correct_error;
943 		fixup_nodatasum->sctx = sctx;
944 		fixup_nodatasum->dev = dev;
945 		fixup_nodatasum->logical = logical;
946 		fixup_nodatasum->root = fs_info->extent_root;
947 		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
948 		scrub_pending_trans_workers_inc(sctx);
949 		fixup_nodatasum->work.func = scrub_fixup_nodatasum;
950 		btrfs_queue_worker(&fs_info->scrub_workers,
951 				   &fixup_nodatasum->work);
952 		goto out;
953 	}
954 
955 	/*
956 	 * now build and submit the bios for the other mirrors, check
957 	 * checksums.
958 	 * First try to pick the mirror which is completely without I/O
959 	 * errors and also does not have a checksum error.
960 	 * If one is found, and if a checksum is present, the full block
961 	 * that is known to contain an error is rewritten. Afterwards
962 	 * the block is known to be corrected.
963 	 * If a mirror is found which is completely correct, and no
964 	 * checksum is present, only those pages are rewritten that had
965 	 * an I/O error in the block to be repaired, since it cannot be
966 	 * determined, which copy of the other pages is better (and it
967 	 * could happen otherwise that a correct page would be
968 	 * overwritten by a bad one).
969 	 */
970 	for (mirror_index = 0;
971 	     mirror_index < BTRFS_MAX_MIRRORS &&
972 	     sblocks_for_recheck[mirror_index].page_count > 0;
973 	     mirror_index++) {
974 		struct scrub_block *sblock_other;
975 
976 		if (mirror_index == failed_mirror_index)
977 			continue;
978 		sblock_other = sblocks_for_recheck + mirror_index;
979 
980 		/* build and submit the bios, check checksums */
981 		scrub_recheck_block(fs_info, sblock_other, is_metadata,
982 				    have_csum, csum, generation,
983 				    sctx->csum_size);
984 
985 		if (!sblock_other->header_error &&
986 		    !sblock_other->checksum_error &&
987 		    sblock_other->no_io_error_seen) {
988 			if (sctx->is_dev_replace) {
989 				scrub_write_block_to_dev_replace(sblock_other);
990 			} else {
991 				int force_write = is_metadata || have_csum;
992 
993 				ret = scrub_repair_block_from_good_copy(
994 						sblock_bad, sblock_other,
995 						force_write);
996 			}
997 			if (0 == ret)
998 				goto corrected_error;
999 		}
1000 	}
1001 
1002 	/*
1003 	 * for dev_replace, pick good pages and write to the target device.
1004 	 */
1005 	if (sctx->is_dev_replace) {
1006 		success = 1;
1007 		for (page_num = 0; page_num < sblock_bad->page_count;
1008 		     page_num++) {
1009 			int sub_success;
1010 
1011 			sub_success = 0;
1012 			for (mirror_index = 0;
1013 			     mirror_index < BTRFS_MAX_MIRRORS &&
1014 			     sblocks_for_recheck[mirror_index].page_count > 0;
1015 			     mirror_index++) {
1016 				struct scrub_block *sblock_other =
1017 					sblocks_for_recheck + mirror_index;
1018 				struct scrub_page *page_other =
1019 					sblock_other->pagev[page_num];
1020 
1021 				if (!page_other->io_error) {
1022 					ret = scrub_write_page_to_dev_replace(
1023 							sblock_other, page_num);
1024 					if (ret == 0) {
1025 						/* succeeded for this page */
1026 						sub_success = 1;
1027 						break;
1028 					} else {
1029 						btrfs_dev_replace_stats_inc(
1030 							&sctx->dev_root->
1031 							fs_info->dev_replace.
1032 							num_write_errors);
1033 					}
1034 				}
1035 			}
1036 
1037 			if (!sub_success) {
1038 				/*
1039 				 * did not find a mirror to fetch the page
1040 				 * from. scrub_write_page_to_dev_replace()
1041 				 * handles this case (page->io_error), by
1042 				 * filling the block with zeros before
1043 				 * submitting the write request
1044 				 */
1045 				success = 0;
1046 				ret = scrub_write_page_to_dev_replace(
1047 						sblock_bad, page_num);
1048 				if (ret)
1049 					btrfs_dev_replace_stats_inc(
1050 						&sctx->dev_root->fs_info->
1051 						dev_replace.num_write_errors);
1052 			}
1053 		}
1054 
1055 		goto out;
1056 	}
1057 
1058 	/*
1059 	 * for regular scrub, repair those pages that are errored.
1060 	 * In case of I/O errors in the area that is supposed to be
1061 	 * repaired, continue by picking good copies of those pages.
1062 	 * Select the good pages from mirrors to rewrite bad pages from
1063 	 * the area to fix. Afterwards verify the checksum of the block
1064 	 * that is supposed to be repaired. This verification step is
1065 	 * only done for the purpose of statistic counting and for the
1066 	 * final scrub report, whether errors remain.
1067 	 * A perfect algorithm could make use of the checksum and try
1068 	 * all possible combinations of pages from the different mirrors
1069 	 * until the checksum verification succeeds. For example, when
1070 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1071 	 * of mirror #2 is readable but the final checksum test fails,
1072 	 * then the 2nd page of mirror #3 could be tried, whether now
1073 	 * the final checksum succeedes. But this would be a rare
1074 	 * exception and is therefore not implemented. At least it is
1075 	 * avoided that the good copy is overwritten.
1076 	 * A more useful improvement would be to pick the sectors
1077 	 * without I/O error based on sector sizes (512 bytes on legacy
1078 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1079 	 * mirror could be repaired by taking 512 byte of a different
1080 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1081 	 * area are unreadable.
1082 	 */
1083 
1084 	/* can only fix I/O errors from here on */
1085 	if (sblock_bad->no_io_error_seen)
1086 		goto did_not_correct_error;
1087 
1088 	success = 1;
1089 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1090 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1091 
1092 		if (!page_bad->io_error)
1093 			continue;
1094 
1095 		for (mirror_index = 0;
1096 		     mirror_index < BTRFS_MAX_MIRRORS &&
1097 		     sblocks_for_recheck[mirror_index].page_count > 0;
1098 		     mirror_index++) {
1099 			struct scrub_block *sblock_other = sblocks_for_recheck +
1100 							   mirror_index;
1101 			struct scrub_page *page_other = sblock_other->pagev[
1102 							page_num];
1103 
1104 			if (!page_other->io_error) {
1105 				ret = scrub_repair_page_from_good_copy(
1106 					sblock_bad, sblock_other, page_num, 0);
1107 				if (0 == ret) {
1108 					page_bad->io_error = 0;
1109 					break; /* succeeded for this page */
1110 				}
1111 			}
1112 		}
1113 
1114 		if (page_bad->io_error) {
1115 			/* did not find a mirror to copy the page from */
1116 			success = 0;
1117 		}
1118 	}
1119 
1120 	if (success) {
1121 		if (is_metadata || have_csum) {
1122 			/*
1123 			 * need to verify the checksum now that all
1124 			 * sectors on disk are repaired (the write
1125 			 * request for data to be repaired is on its way).
1126 			 * Just be lazy and use scrub_recheck_block()
1127 			 * which re-reads the data before the checksum
1128 			 * is verified, but most likely the data comes out
1129 			 * of the page cache.
1130 			 */
1131 			scrub_recheck_block(fs_info, sblock_bad,
1132 					    is_metadata, have_csum, csum,
1133 					    generation, sctx->csum_size);
1134 			if (!sblock_bad->header_error &&
1135 			    !sblock_bad->checksum_error &&
1136 			    sblock_bad->no_io_error_seen)
1137 				goto corrected_error;
1138 			else
1139 				goto did_not_correct_error;
1140 		} else {
1141 corrected_error:
1142 			spin_lock(&sctx->stat_lock);
1143 			sctx->stat.corrected_errors++;
1144 			spin_unlock(&sctx->stat_lock);
1145 			printk_ratelimited_in_rcu(KERN_ERR
1146 				"btrfs: fixed up error at logical %llu on dev %s\n",
1147 				(unsigned long long)logical,
1148 				rcu_str_deref(dev->name));
1149 		}
1150 	} else {
1151 did_not_correct_error:
1152 		spin_lock(&sctx->stat_lock);
1153 		sctx->stat.uncorrectable_errors++;
1154 		spin_unlock(&sctx->stat_lock);
1155 		printk_ratelimited_in_rcu(KERN_ERR
1156 			"btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1157 			(unsigned long long)logical,
1158 			rcu_str_deref(dev->name));
1159 	}
1160 
1161 out:
1162 	if (sblocks_for_recheck) {
1163 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1164 		     mirror_index++) {
1165 			struct scrub_block *sblock = sblocks_for_recheck +
1166 						     mirror_index;
1167 			int page_index;
1168 
1169 			for (page_index = 0; page_index < sblock->page_count;
1170 			     page_index++) {
1171 				sblock->pagev[page_index]->sblock = NULL;
1172 				scrub_page_put(sblock->pagev[page_index]);
1173 			}
1174 		}
1175 		kfree(sblocks_for_recheck);
1176 	}
1177 
1178 	return 0;
1179 }
1180 
1181 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1182 				     struct btrfs_fs_info *fs_info,
1183 				     struct scrub_block *original_sblock,
1184 				     u64 length, u64 logical,
1185 				     struct scrub_block *sblocks_for_recheck)
1186 {
1187 	int page_index;
1188 	int mirror_index;
1189 	int ret;
1190 
1191 	/*
1192 	 * note: the two members ref_count and outstanding_pages
1193 	 * are not used (and not set) in the blocks that are used for
1194 	 * the recheck procedure
1195 	 */
1196 
1197 	page_index = 0;
1198 	while (length > 0) {
1199 		u64 sublen = min_t(u64, length, PAGE_SIZE);
1200 		u64 mapped_length = sublen;
1201 		struct btrfs_bio *bbio = NULL;
1202 
1203 		/*
1204 		 * with a length of PAGE_SIZE, each returned stripe
1205 		 * represents one mirror
1206 		 */
1207 		ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1208 				      &mapped_length, &bbio, 0);
1209 		if (ret || !bbio || mapped_length < sublen) {
1210 			kfree(bbio);
1211 			return -EIO;
1212 		}
1213 
1214 		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1215 		for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1216 		     mirror_index++) {
1217 			struct scrub_block *sblock;
1218 			struct scrub_page *page;
1219 
1220 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1221 				continue;
1222 
1223 			sblock = sblocks_for_recheck + mirror_index;
1224 			sblock->sctx = sctx;
1225 			page = kzalloc(sizeof(*page), GFP_NOFS);
1226 			if (!page) {
1227 leave_nomem:
1228 				spin_lock(&sctx->stat_lock);
1229 				sctx->stat.malloc_errors++;
1230 				spin_unlock(&sctx->stat_lock);
1231 				kfree(bbio);
1232 				return -ENOMEM;
1233 			}
1234 			scrub_page_get(page);
1235 			sblock->pagev[page_index] = page;
1236 			page->logical = logical;
1237 			page->physical = bbio->stripes[mirror_index].physical;
1238 			BUG_ON(page_index >= original_sblock->page_count);
1239 			page->physical_for_dev_replace =
1240 				original_sblock->pagev[page_index]->
1241 				physical_for_dev_replace;
1242 			/* for missing devices, dev->bdev is NULL */
1243 			page->dev = bbio->stripes[mirror_index].dev;
1244 			page->mirror_num = mirror_index + 1;
1245 			sblock->page_count++;
1246 			page->page = alloc_page(GFP_NOFS);
1247 			if (!page->page)
1248 				goto leave_nomem;
1249 		}
1250 		kfree(bbio);
1251 		length -= sublen;
1252 		logical += sublen;
1253 		page_index++;
1254 	}
1255 
1256 	return 0;
1257 }
1258 
1259 /*
1260  * this function will check the on disk data for checksum errors, header
1261  * errors and read I/O errors. If any I/O errors happen, the exact pages
1262  * which are errored are marked as being bad. The goal is to enable scrub
1263  * to take those pages that are not errored from all the mirrors so that
1264  * the pages that are errored in the just handled mirror can be repaired.
1265  */
1266 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1267 				struct scrub_block *sblock, int is_metadata,
1268 				int have_csum, u8 *csum, u64 generation,
1269 				u16 csum_size)
1270 {
1271 	int page_num;
1272 
1273 	sblock->no_io_error_seen = 1;
1274 	sblock->header_error = 0;
1275 	sblock->checksum_error = 0;
1276 
1277 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1278 		struct bio *bio;
1279 		struct scrub_page *page = sblock->pagev[page_num];
1280 		DECLARE_COMPLETION_ONSTACK(complete);
1281 
1282 		if (page->dev->bdev == NULL) {
1283 			page->io_error = 1;
1284 			sblock->no_io_error_seen = 0;
1285 			continue;
1286 		}
1287 
1288 		WARN_ON(!page->page);
1289 		bio = bio_alloc(GFP_NOFS, 1);
1290 		if (!bio) {
1291 			page->io_error = 1;
1292 			sblock->no_io_error_seen = 0;
1293 			continue;
1294 		}
1295 		bio->bi_bdev = page->dev->bdev;
1296 		bio->bi_sector = page->physical >> 9;
1297 		bio->bi_end_io = scrub_complete_bio_end_io;
1298 		bio->bi_private = &complete;
1299 
1300 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1301 		btrfsic_submit_bio(READ, bio);
1302 
1303 		/* this will also unplug the queue */
1304 		wait_for_completion(&complete);
1305 
1306 		page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1307 		if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1308 			sblock->no_io_error_seen = 0;
1309 		bio_put(bio);
1310 	}
1311 
1312 	if (sblock->no_io_error_seen)
1313 		scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1314 					     have_csum, csum, generation,
1315 					     csum_size);
1316 
1317 	return;
1318 }
1319 
1320 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1321 					 struct scrub_block *sblock,
1322 					 int is_metadata, int have_csum,
1323 					 const u8 *csum, u64 generation,
1324 					 u16 csum_size)
1325 {
1326 	int page_num;
1327 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1328 	u32 crc = ~(u32)0;
1329 	struct btrfs_root *root = fs_info->extent_root;
1330 	void *mapped_buffer;
1331 
1332 	WARN_ON(!sblock->pagev[0]->page);
1333 	if (is_metadata) {
1334 		struct btrfs_header *h;
1335 
1336 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1337 		h = (struct btrfs_header *)mapped_buffer;
1338 
1339 		if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1340 		    memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1341 		    memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1342 			   BTRFS_UUID_SIZE)) {
1343 			sblock->header_error = 1;
1344 		} else if (generation != le64_to_cpu(h->generation)) {
1345 			sblock->header_error = 1;
1346 			sblock->generation_error = 1;
1347 		}
1348 		csum = h->csum;
1349 	} else {
1350 		if (!have_csum)
1351 			return;
1352 
1353 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1354 	}
1355 
1356 	for (page_num = 0;;) {
1357 		if (page_num == 0 && is_metadata)
1358 			crc = btrfs_csum_data(root,
1359 				((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1360 				crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1361 		else
1362 			crc = btrfs_csum_data(root, mapped_buffer, crc,
1363 					      PAGE_SIZE);
1364 
1365 		kunmap_atomic(mapped_buffer);
1366 		page_num++;
1367 		if (page_num >= sblock->page_count)
1368 			break;
1369 		WARN_ON(!sblock->pagev[page_num]->page);
1370 
1371 		mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1372 	}
1373 
1374 	btrfs_csum_final(crc, calculated_csum);
1375 	if (memcmp(calculated_csum, csum, csum_size))
1376 		sblock->checksum_error = 1;
1377 }
1378 
1379 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1380 {
1381 	complete((struct completion *)bio->bi_private);
1382 }
1383 
1384 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1385 					     struct scrub_block *sblock_good,
1386 					     int force_write)
1387 {
1388 	int page_num;
1389 	int ret = 0;
1390 
1391 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1392 		int ret_sub;
1393 
1394 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1395 							   sblock_good,
1396 							   page_num,
1397 							   force_write);
1398 		if (ret_sub)
1399 			ret = ret_sub;
1400 	}
1401 
1402 	return ret;
1403 }
1404 
1405 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1406 					    struct scrub_block *sblock_good,
1407 					    int page_num, int force_write)
1408 {
1409 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1410 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1411 
1412 	BUG_ON(page_bad->page == NULL);
1413 	BUG_ON(page_good->page == NULL);
1414 	if (force_write || sblock_bad->header_error ||
1415 	    sblock_bad->checksum_error || page_bad->io_error) {
1416 		struct bio *bio;
1417 		int ret;
1418 		DECLARE_COMPLETION_ONSTACK(complete);
1419 
1420 		if (!page_bad->dev->bdev) {
1421 			printk_ratelimited(KERN_WARNING
1422 				"btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1423 			return -EIO;
1424 		}
1425 
1426 		bio = bio_alloc(GFP_NOFS, 1);
1427 		if (!bio)
1428 			return -EIO;
1429 		bio->bi_bdev = page_bad->dev->bdev;
1430 		bio->bi_sector = page_bad->physical >> 9;
1431 		bio->bi_end_io = scrub_complete_bio_end_io;
1432 		bio->bi_private = &complete;
1433 
1434 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1435 		if (PAGE_SIZE != ret) {
1436 			bio_put(bio);
1437 			return -EIO;
1438 		}
1439 		btrfsic_submit_bio(WRITE, bio);
1440 
1441 		/* this will also unplug the queue */
1442 		wait_for_completion(&complete);
1443 		if (!bio_flagged(bio, BIO_UPTODATE)) {
1444 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1445 				BTRFS_DEV_STAT_WRITE_ERRS);
1446 			btrfs_dev_replace_stats_inc(
1447 				&sblock_bad->sctx->dev_root->fs_info->
1448 				dev_replace.num_write_errors);
1449 			bio_put(bio);
1450 			return -EIO;
1451 		}
1452 		bio_put(bio);
1453 	}
1454 
1455 	return 0;
1456 }
1457 
1458 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1459 {
1460 	int page_num;
1461 
1462 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1463 		int ret;
1464 
1465 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1466 		if (ret)
1467 			btrfs_dev_replace_stats_inc(
1468 				&sblock->sctx->dev_root->fs_info->dev_replace.
1469 				num_write_errors);
1470 	}
1471 }
1472 
1473 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1474 					   int page_num)
1475 {
1476 	struct scrub_page *spage = sblock->pagev[page_num];
1477 
1478 	BUG_ON(spage->page == NULL);
1479 	if (spage->io_error) {
1480 		void *mapped_buffer = kmap_atomic(spage->page);
1481 
1482 		memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1483 		flush_dcache_page(spage->page);
1484 		kunmap_atomic(mapped_buffer);
1485 	}
1486 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1487 }
1488 
1489 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1490 				    struct scrub_page *spage)
1491 {
1492 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1493 	struct scrub_bio *sbio;
1494 	int ret;
1495 
1496 	mutex_lock(&wr_ctx->wr_lock);
1497 again:
1498 	if (!wr_ctx->wr_curr_bio) {
1499 		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1500 					      GFP_NOFS);
1501 		if (!wr_ctx->wr_curr_bio) {
1502 			mutex_unlock(&wr_ctx->wr_lock);
1503 			return -ENOMEM;
1504 		}
1505 		wr_ctx->wr_curr_bio->sctx = sctx;
1506 		wr_ctx->wr_curr_bio->page_count = 0;
1507 	}
1508 	sbio = wr_ctx->wr_curr_bio;
1509 	if (sbio->page_count == 0) {
1510 		struct bio *bio;
1511 
1512 		sbio->physical = spage->physical_for_dev_replace;
1513 		sbio->logical = spage->logical;
1514 		sbio->dev = wr_ctx->tgtdev;
1515 		bio = sbio->bio;
1516 		if (!bio) {
1517 			bio = bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1518 			if (!bio) {
1519 				mutex_unlock(&wr_ctx->wr_lock);
1520 				return -ENOMEM;
1521 			}
1522 			sbio->bio = bio;
1523 		}
1524 
1525 		bio->bi_private = sbio;
1526 		bio->bi_end_io = scrub_wr_bio_end_io;
1527 		bio->bi_bdev = sbio->dev->bdev;
1528 		bio->bi_sector = sbio->physical >> 9;
1529 		sbio->err = 0;
1530 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1531 		   spage->physical_for_dev_replace ||
1532 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1533 		   spage->logical) {
1534 		scrub_wr_submit(sctx);
1535 		goto again;
1536 	}
1537 
1538 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1539 	if (ret != PAGE_SIZE) {
1540 		if (sbio->page_count < 1) {
1541 			bio_put(sbio->bio);
1542 			sbio->bio = NULL;
1543 			mutex_unlock(&wr_ctx->wr_lock);
1544 			return -EIO;
1545 		}
1546 		scrub_wr_submit(sctx);
1547 		goto again;
1548 	}
1549 
1550 	sbio->pagev[sbio->page_count] = spage;
1551 	scrub_page_get(spage);
1552 	sbio->page_count++;
1553 	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1554 		scrub_wr_submit(sctx);
1555 	mutex_unlock(&wr_ctx->wr_lock);
1556 
1557 	return 0;
1558 }
1559 
1560 static void scrub_wr_submit(struct scrub_ctx *sctx)
1561 {
1562 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1563 	struct scrub_bio *sbio;
1564 
1565 	if (!wr_ctx->wr_curr_bio)
1566 		return;
1567 
1568 	sbio = wr_ctx->wr_curr_bio;
1569 	wr_ctx->wr_curr_bio = NULL;
1570 	WARN_ON(!sbio->bio->bi_bdev);
1571 	scrub_pending_bio_inc(sctx);
1572 	/* process all writes in a single worker thread. Then the block layer
1573 	 * orders the requests before sending them to the driver which
1574 	 * doubled the write performance on spinning disks when measured
1575 	 * with Linux 3.5 */
1576 	btrfsic_submit_bio(WRITE, sbio->bio);
1577 }
1578 
1579 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1580 {
1581 	struct scrub_bio *sbio = bio->bi_private;
1582 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1583 
1584 	sbio->err = err;
1585 	sbio->bio = bio;
1586 
1587 	sbio->work.func = scrub_wr_bio_end_io_worker;
1588 	btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1589 }
1590 
1591 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1592 {
1593 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1594 	struct scrub_ctx *sctx = sbio->sctx;
1595 	int i;
1596 
1597 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1598 	if (sbio->err) {
1599 		struct btrfs_dev_replace *dev_replace =
1600 			&sbio->sctx->dev_root->fs_info->dev_replace;
1601 
1602 		for (i = 0; i < sbio->page_count; i++) {
1603 			struct scrub_page *spage = sbio->pagev[i];
1604 
1605 			spage->io_error = 1;
1606 			btrfs_dev_replace_stats_inc(&dev_replace->
1607 						    num_write_errors);
1608 		}
1609 	}
1610 
1611 	for (i = 0; i < sbio->page_count; i++)
1612 		scrub_page_put(sbio->pagev[i]);
1613 
1614 	bio_put(sbio->bio);
1615 	kfree(sbio);
1616 	scrub_pending_bio_dec(sctx);
1617 }
1618 
1619 static int scrub_checksum(struct scrub_block *sblock)
1620 {
1621 	u64 flags;
1622 	int ret;
1623 
1624 	WARN_ON(sblock->page_count < 1);
1625 	flags = sblock->pagev[0]->flags;
1626 	ret = 0;
1627 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1628 		ret = scrub_checksum_data(sblock);
1629 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1630 		ret = scrub_checksum_tree_block(sblock);
1631 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1632 		(void)scrub_checksum_super(sblock);
1633 	else
1634 		WARN_ON(1);
1635 	if (ret)
1636 		scrub_handle_errored_block(sblock);
1637 
1638 	return ret;
1639 }
1640 
1641 static int scrub_checksum_data(struct scrub_block *sblock)
1642 {
1643 	struct scrub_ctx *sctx = sblock->sctx;
1644 	u8 csum[BTRFS_CSUM_SIZE];
1645 	u8 *on_disk_csum;
1646 	struct page *page;
1647 	void *buffer;
1648 	u32 crc = ~(u32)0;
1649 	int fail = 0;
1650 	struct btrfs_root *root = sctx->dev_root;
1651 	u64 len;
1652 	int index;
1653 
1654 	BUG_ON(sblock->page_count < 1);
1655 	if (!sblock->pagev[0]->have_csum)
1656 		return 0;
1657 
1658 	on_disk_csum = sblock->pagev[0]->csum;
1659 	page = sblock->pagev[0]->page;
1660 	buffer = kmap_atomic(page);
1661 
1662 	len = sctx->sectorsize;
1663 	index = 0;
1664 	for (;;) {
1665 		u64 l = min_t(u64, len, PAGE_SIZE);
1666 
1667 		crc = btrfs_csum_data(root, buffer, crc, l);
1668 		kunmap_atomic(buffer);
1669 		len -= l;
1670 		if (len == 0)
1671 			break;
1672 		index++;
1673 		BUG_ON(index >= sblock->page_count);
1674 		BUG_ON(!sblock->pagev[index]->page);
1675 		page = sblock->pagev[index]->page;
1676 		buffer = kmap_atomic(page);
1677 	}
1678 
1679 	btrfs_csum_final(crc, csum);
1680 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1681 		fail = 1;
1682 
1683 	return fail;
1684 }
1685 
1686 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1687 {
1688 	struct scrub_ctx *sctx = sblock->sctx;
1689 	struct btrfs_header *h;
1690 	struct btrfs_root *root = sctx->dev_root;
1691 	struct btrfs_fs_info *fs_info = root->fs_info;
1692 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1693 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1694 	struct page *page;
1695 	void *mapped_buffer;
1696 	u64 mapped_size;
1697 	void *p;
1698 	u32 crc = ~(u32)0;
1699 	int fail = 0;
1700 	int crc_fail = 0;
1701 	u64 len;
1702 	int index;
1703 
1704 	BUG_ON(sblock->page_count < 1);
1705 	page = sblock->pagev[0]->page;
1706 	mapped_buffer = kmap_atomic(page);
1707 	h = (struct btrfs_header *)mapped_buffer;
1708 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1709 
1710 	/*
1711 	 * we don't use the getter functions here, as we
1712 	 * a) don't have an extent buffer and
1713 	 * b) the page is already kmapped
1714 	 */
1715 
1716 	if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1717 		++fail;
1718 
1719 	if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1720 		++fail;
1721 
1722 	if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1723 		++fail;
1724 
1725 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1726 		   BTRFS_UUID_SIZE))
1727 		++fail;
1728 
1729 	WARN_ON(sctx->nodesize != sctx->leafsize);
1730 	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1731 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1732 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1733 	index = 0;
1734 	for (;;) {
1735 		u64 l = min_t(u64, len, mapped_size);
1736 
1737 		crc = btrfs_csum_data(root, p, crc, l);
1738 		kunmap_atomic(mapped_buffer);
1739 		len -= l;
1740 		if (len == 0)
1741 			break;
1742 		index++;
1743 		BUG_ON(index >= sblock->page_count);
1744 		BUG_ON(!sblock->pagev[index]->page);
1745 		page = sblock->pagev[index]->page;
1746 		mapped_buffer = kmap_atomic(page);
1747 		mapped_size = PAGE_SIZE;
1748 		p = mapped_buffer;
1749 	}
1750 
1751 	btrfs_csum_final(crc, calculated_csum);
1752 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1753 		++crc_fail;
1754 
1755 	return fail || crc_fail;
1756 }
1757 
1758 static int scrub_checksum_super(struct scrub_block *sblock)
1759 {
1760 	struct btrfs_super_block *s;
1761 	struct scrub_ctx *sctx = sblock->sctx;
1762 	struct btrfs_root *root = sctx->dev_root;
1763 	struct btrfs_fs_info *fs_info = root->fs_info;
1764 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1765 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1766 	struct page *page;
1767 	void *mapped_buffer;
1768 	u64 mapped_size;
1769 	void *p;
1770 	u32 crc = ~(u32)0;
1771 	int fail_gen = 0;
1772 	int fail_cor = 0;
1773 	u64 len;
1774 	int index;
1775 
1776 	BUG_ON(sblock->page_count < 1);
1777 	page = sblock->pagev[0]->page;
1778 	mapped_buffer = kmap_atomic(page);
1779 	s = (struct btrfs_super_block *)mapped_buffer;
1780 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1781 
1782 	if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1783 		++fail_cor;
1784 
1785 	if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1786 		++fail_gen;
1787 
1788 	if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1789 		++fail_cor;
1790 
1791 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1792 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1793 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1794 	index = 0;
1795 	for (;;) {
1796 		u64 l = min_t(u64, len, mapped_size);
1797 
1798 		crc = btrfs_csum_data(root, p, crc, l);
1799 		kunmap_atomic(mapped_buffer);
1800 		len -= l;
1801 		if (len == 0)
1802 			break;
1803 		index++;
1804 		BUG_ON(index >= sblock->page_count);
1805 		BUG_ON(!sblock->pagev[index]->page);
1806 		page = sblock->pagev[index]->page;
1807 		mapped_buffer = kmap_atomic(page);
1808 		mapped_size = PAGE_SIZE;
1809 		p = mapped_buffer;
1810 	}
1811 
1812 	btrfs_csum_final(crc, calculated_csum);
1813 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1814 		++fail_cor;
1815 
1816 	if (fail_cor + fail_gen) {
1817 		/*
1818 		 * if we find an error in a super block, we just report it.
1819 		 * They will get written with the next transaction commit
1820 		 * anyway
1821 		 */
1822 		spin_lock(&sctx->stat_lock);
1823 		++sctx->stat.super_errors;
1824 		spin_unlock(&sctx->stat_lock);
1825 		if (fail_cor)
1826 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1827 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1828 		else
1829 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1830 				BTRFS_DEV_STAT_GENERATION_ERRS);
1831 	}
1832 
1833 	return fail_cor + fail_gen;
1834 }
1835 
1836 static void scrub_block_get(struct scrub_block *sblock)
1837 {
1838 	atomic_inc(&sblock->ref_count);
1839 }
1840 
1841 static void scrub_block_put(struct scrub_block *sblock)
1842 {
1843 	if (atomic_dec_and_test(&sblock->ref_count)) {
1844 		int i;
1845 
1846 		for (i = 0; i < sblock->page_count; i++)
1847 			scrub_page_put(sblock->pagev[i]);
1848 		kfree(sblock);
1849 	}
1850 }
1851 
1852 static void scrub_page_get(struct scrub_page *spage)
1853 {
1854 	atomic_inc(&spage->ref_count);
1855 }
1856 
1857 static void scrub_page_put(struct scrub_page *spage)
1858 {
1859 	if (atomic_dec_and_test(&spage->ref_count)) {
1860 		if (spage->page)
1861 			__free_page(spage->page);
1862 		kfree(spage);
1863 	}
1864 }
1865 
1866 static void scrub_submit(struct scrub_ctx *sctx)
1867 {
1868 	struct scrub_bio *sbio;
1869 
1870 	if (sctx->curr == -1)
1871 		return;
1872 
1873 	sbio = sctx->bios[sctx->curr];
1874 	sctx->curr = -1;
1875 	scrub_pending_bio_inc(sctx);
1876 
1877 	if (!sbio->bio->bi_bdev) {
1878 		/*
1879 		 * this case should not happen. If btrfs_map_block() is
1880 		 * wrong, it could happen for dev-replace operations on
1881 		 * missing devices when no mirrors are available, but in
1882 		 * this case it should already fail the mount.
1883 		 * This case is handled correctly (but _very_ slowly).
1884 		 */
1885 		printk_ratelimited(KERN_WARNING
1886 			"btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1887 		bio_endio(sbio->bio, -EIO);
1888 	} else {
1889 		btrfsic_submit_bio(READ, sbio->bio);
1890 	}
1891 }
1892 
1893 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1894 				    struct scrub_page *spage)
1895 {
1896 	struct scrub_block *sblock = spage->sblock;
1897 	struct scrub_bio *sbio;
1898 	int ret;
1899 
1900 again:
1901 	/*
1902 	 * grab a fresh bio or wait for one to become available
1903 	 */
1904 	while (sctx->curr == -1) {
1905 		spin_lock(&sctx->list_lock);
1906 		sctx->curr = sctx->first_free;
1907 		if (sctx->curr != -1) {
1908 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1909 			sctx->bios[sctx->curr]->next_free = -1;
1910 			sctx->bios[sctx->curr]->page_count = 0;
1911 			spin_unlock(&sctx->list_lock);
1912 		} else {
1913 			spin_unlock(&sctx->list_lock);
1914 			wait_event(sctx->list_wait, sctx->first_free != -1);
1915 		}
1916 	}
1917 	sbio = sctx->bios[sctx->curr];
1918 	if (sbio->page_count == 0) {
1919 		struct bio *bio;
1920 
1921 		sbio->physical = spage->physical;
1922 		sbio->logical = spage->logical;
1923 		sbio->dev = spage->dev;
1924 		bio = sbio->bio;
1925 		if (!bio) {
1926 			bio = bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1927 			if (!bio)
1928 				return -ENOMEM;
1929 			sbio->bio = bio;
1930 		}
1931 
1932 		bio->bi_private = sbio;
1933 		bio->bi_end_io = scrub_bio_end_io;
1934 		bio->bi_bdev = sbio->dev->bdev;
1935 		bio->bi_sector = sbio->physical >> 9;
1936 		sbio->err = 0;
1937 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1938 		   spage->physical ||
1939 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1940 		   spage->logical ||
1941 		   sbio->dev != spage->dev) {
1942 		scrub_submit(sctx);
1943 		goto again;
1944 	}
1945 
1946 	sbio->pagev[sbio->page_count] = spage;
1947 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1948 	if (ret != PAGE_SIZE) {
1949 		if (sbio->page_count < 1) {
1950 			bio_put(sbio->bio);
1951 			sbio->bio = NULL;
1952 			return -EIO;
1953 		}
1954 		scrub_submit(sctx);
1955 		goto again;
1956 	}
1957 
1958 	scrub_block_get(sblock); /* one for the page added to the bio */
1959 	atomic_inc(&sblock->outstanding_pages);
1960 	sbio->page_count++;
1961 	if (sbio->page_count == sctx->pages_per_rd_bio)
1962 		scrub_submit(sctx);
1963 
1964 	return 0;
1965 }
1966 
1967 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1968 		       u64 physical, struct btrfs_device *dev, u64 flags,
1969 		       u64 gen, int mirror_num, u8 *csum, int force,
1970 		       u64 physical_for_dev_replace)
1971 {
1972 	struct scrub_block *sblock;
1973 	int index;
1974 
1975 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1976 	if (!sblock) {
1977 		spin_lock(&sctx->stat_lock);
1978 		sctx->stat.malloc_errors++;
1979 		spin_unlock(&sctx->stat_lock);
1980 		return -ENOMEM;
1981 	}
1982 
1983 	/* one ref inside this function, plus one for each page added to
1984 	 * a bio later on */
1985 	atomic_set(&sblock->ref_count, 1);
1986 	sblock->sctx = sctx;
1987 	sblock->no_io_error_seen = 1;
1988 
1989 	for (index = 0; len > 0; index++) {
1990 		struct scrub_page *spage;
1991 		u64 l = min_t(u64, len, PAGE_SIZE);
1992 
1993 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
1994 		if (!spage) {
1995 leave_nomem:
1996 			spin_lock(&sctx->stat_lock);
1997 			sctx->stat.malloc_errors++;
1998 			spin_unlock(&sctx->stat_lock);
1999 			scrub_block_put(sblock);
2000 			return -ENOMEM;
2001 		}
2002 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2003 		scrub_page_get(spage);
2004 		sblock->pagev[index] = spage;
2005 		spage->sblock = sblock;
2006 		spage->dev = dev;
2007 		spage->flags = flags;
2008 		spage->generation = gen;
2009 		spage->logical = logical;
2010 		spage->physical = physical;
2011 		spage->physical_for_dev_replace = physical_for_dev_replace;
2012 		spage->mirror_num = mirror_num;
2013 		if (csum) {
2014 			spage->have_csum = 1;
2015 			memcpy(spage->csum, csum, sctx->csum_size);
2016 		} else {
2017 			spage->have_csum = 0;
2018 		}
2019 		sblock->page_count++;
2020 		spage->page = alloc_page(GFP_NOFS);
2021 		if (!spage->page)
2022 			goto leave_nomem;
2023 		len -= l;
2024 		logical += l;
2025 		physical += l;
2026 		physical_for_dev_replace += l;
2027 	}
2028 
2029 	WARN_ON(sblock->page_count == 0);
2030 	for (index = 0; index < sblock->page_count; index++) {
2031 		struct scrub_page *spage = sblock->pagev[index];
2032 		int ret;
2033 
2034 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2035 		if (ret) {
2036 			scrub_block_put(sblock);
2037 			return ret;
2038 		}
2039 	}
2040 
2041 	if (force)
2042 		scrub_submit(sctx);
2043 
2044 	/* last one frees, either here or in bio completion for last page */
2045 	scrub_block_put(sblock);
2046 	return 0;
2047 }
2048 
2049 static void scrub_bio_end_io(struct bio *bio, int err)
2050 {
2051 	struct scrub_bio *sbio = bio->bi_private;
2052 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2053 
2054 	sbio->err = err;
2055 	sbio->bio = bio;
2056 
2057 	btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2058 }
2059 
2060 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2061 {
2062 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2063 	struct scrub_ctx *sctx = sbio->sctx;
2064 	int i;
2065 
2066 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2067 	if (sbio->err) {
2068 		for (i = 0; i < sbio->page_count; i++) {
2069 			struct scrub_page *spage = sbio->pagev[i];
2070 
2071 			spage->io_error = 1;
2072 			spage->sblock->no_io_error_seen = 0;
2073 		}
2074 	}
2075 
2076 	/* now complete the scrub_block items that have all pages completed */
2077 	for (i = 0; i < sbio->page_count; i++) {
2078 		struct scrub_page *spage = sbio->pagev[i];
2079 		struct scrub_block *sblock = spage->sblock;
2080 
2081 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2082 			scrub_block_complete(sblock);
2083 		scrub_block_put(sblock);
2084 	}
2085 
2086 	bio_put(sbio->bio);
2087 	sbio->bio = NULL;
2088 	spin_lock(&sctx->list_lock);
2089 	sbio->next_free = sctx->first_free;
2090 	sctx->first_free = sbio->index;
2091 	spin_unlock(&sctx->list_lock);
2092 
2093 	if (sctx->is_dev_replace &&
2094 	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2095 		mutex_lock(&sctx->wr_ctx.wr_lock);
2096 		scrub_wr_submit(sctx);
2097 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2098 	}
2099 
2100 	scrub_pending_bio_dec(sctx);
2101 }
2102 
2103 static void scrub_block_complete(struct scrub_block *sblock)
2104 {
2105 	if (!sblock->no_io_error_seen) {
2106 		scrub_handle_errored_block(sblock);
2107 	} else {
2108 		/*
2109 		 * if has checksum error, write via repair mechanism in
2110 		 * dev replace case, otherwise write here in dev replace
2111 		 * case.
2112 		 */
2113 		if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2114 			scrub_write_block_to_dev_replace(sblock);
2115 	}
2116 }
2117 
2118 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2119 			   u8 *csum)
2120 {
2121 	struct btrfs_ordered_sum *sum = NULL;
2122 	int ret = 0;
2123 	unsigned long i;
2124 	unsigned long num_sectors;
2125 
2126 	while (!list_empty(&sctx->csum_list)) {
2127 		sum = list_first_entry(&sctx->csum_list,
2128 				       struct btrfs_ordered_sum, list);
2129 		if (sum->bytenr > logical)
2130 			return 0;
2131 		if (sum->bytenr + sum->len > logical)
2132 			break;
2133 
2134 		++sctx->stat.csum_discards;
2135 		list_del(&sum->list);
2136 		kfree(sum);
2137 		sum = NULL;
2138 	}
2139 	if (!sum)
2140 		return 0;
2141 
2142 	num_sectors = sum->len / sctx->sectorsize;
2143 	for (i = 0; i < num_sectors; ++i) {
2144 		if (sum->sums[i].bytenr == logical) {
2145 			memcpy(csum, &sum->sums[i].sum, sctx->csum_size);
2146 			ret = 1;
2147 			break;
2148 		}
2149 	}
2150 	if (ret && i == num_sectors - 1) {
2151 		list_del(&sum->list);
2152 		kfree(sum);
2153 	}
2154 	return ret;
2155 }
2156 
2157 /* scrub extent tries to collect up to 64 kB for each bio */
2158 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2159 			u64 physical, struct btrfs_device *dev, u64 flags,
2160 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2161 {
2162 	int ret;
2163 	u8 csum[BTRFS_CSUM_SIZE];
2164 	u32 blocksize;
2165 
2166 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2167 		blocksize = sctx->sectorsize;
2168 		spin_lock(&sctx->stat_lock);
2169 		sctx->stat.data_extents_scrubbed++;
2170 		sctx->stat.data_bytes_scrubbed += len;
2171 		spin_unlock(&sctx->stat_lock);
2172 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2173 		WARN_ON(sctx->nodesize != sctx->leafsize);
2174 		blocksize = sctx->nodesize;
2175 		spin_lock(&sctx->stat_lock);
2176 		sctx->stat.tree_extents_scrubbed++;
2177 		sctx->stat.tree_bytes_scrubbed += len;
2178 		spin_unlock(&sctx->stat_lock);
2179 	} else {
2180 		blocksize = sctx->sectorsize;
2181 		WARN_ON(1);
2182 	}
2183 
2184 	while (len) {
2185 		u64 l = min_t(u64, len, blocksize);
2186 		int have_csum = 0;
2187 
2188 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2189 			/* push csums to sbio */
2190 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2191 			if (have_csum == 0)
2192 				++sctx->stat.no_csum;
2193 			if (sctx->is_dev_replace && !have_csum) {
2194 				ret = copy_nocow_pages(sctx, logical, l,
2195 						       mirror_num,
2196 						      physical_for_dev_replace);
2197 				goto behind_scrub_pages;
2198 			}
2199 		}
2200 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2201 				  mirror_num, have_csum ? csum : NULL, 0,
2202 				  physical_for_dev_replace);
2203 behind_scrub_pages:
2204 		if (ret)
2205 			return ret;
2206 		len -= l;
2207 		logical += l;
2208 		physical += l;
2209 		physical_for_dev_replace += l;
2210 	}
2211 	return 0;
2212 }
2213 
2214 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2215 					   struct map_lookup *map,
2216 					   struct btrfs_device *scrub_dev,
2217 					   int num, u64 base, u64 length,
2218 					   int is_dev_replace)
2219 {
2220 	struct btrfs_path *path;
2221 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2222 	struct btrfs_root *root = fs_info->extent_root;
2223 	struct btrfs_root *csum_root = fs_info->csum_root;
2224 	struct btrfs_extent_item *extent;
2225 	struct blk_plug plug;
2226 	u64 flags;
2227 	int ret;
2228 	int slot;
2229 	int i;
2230 	u64 nstripes;
2231 	struct extent_buffer *l;
2232 	struct btrfs_key key;
2233 	u64 physical;
2234 	u64 logical;
2235 	u64 generation;
2236 	int mirror_num;
2237 	struct reada_control *reada1;
2238 	struct reada_control *reada2;
2239 	struct btrfs_key key_start;
2240 	struct btrfs_key key_end;
2241 	u64 increment = map->stripe_len;
2242 	u64 offset;
2243 	u64 extent_logical;
2244 	u64 extent_physical;
2245 	u64 extent_len;
2246 	struct btrfs_device *extent_dev;
2247 	int extent_mirror_num;
2248 
2249 	nstripes = length;
2250 	offset = 0;
2251 	do_div(nstripes, map->stripe_len);
2252 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2253 		offset = map->stripe_len * num;
2254 		increment = map->stripe_len * map->num_stripes;
2255 		mirror_num = 1;
2256 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2257 		int factor = map->num_stripes / map->sub_stripes;
2258 		offset = map->stripe_len * (num / map->sub_stripes);
2259 		increment = map->stripe_len * factor;
2260 		mirror_num = num % map->sub_stripes + 1;
2261 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2262 		increment = map->stripe_len;
2263 		mirror_num = num % map->num_stripes + 1;
2264 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2265 		increment = map->stripe_len;
2266 		mirror_num = num % map->num_stripes + 1;
2267 	} else {
2268 		increment = map->stripe_len;
2269 		mirror_num = 1;
2270 	}
2271 
2272 	path = btrfs_alloc_path();
2273 	if (!path)
2274 		return -ENOMEM;
2275 
2276 	/*
2277 	 * work on commit root. The related disk blocks are static as
2278 	 * long as COW is applied. This means, it is save to rewrite
2279 	 * them to repair disk errors without any race conditions
2280 	 */
2281 	path->search_commit_root = 1;
2282 	path->skip_locking = 1;
2283 
2284 	/*
2285 	 * trigger the readahead for extent tree csum tree and wait for
2286 	 * completion. During readahead, the scrub is officially paused
2287 	 * to not hold off transaction commits
2288 	 */
2289 	logical = base + offset;
2290 
2291 	wait_event(sctx->list_wait,
2292 		   atomic_read(&sctx->bios_in_flight) == 0);
2293 	atomic_inc(&fs_info->scrubs_paused);
2294 	wake_up(&fs_info->scrub_pause_wait);
2295 
2296 	/* FIXME it might be better to start readahead at commit root */
2297 	key_start.objectid = logical;
2298 	key_start.type = BTRFS_EXTENT_ITEM_KEY;
2299 	key_start.offset = (u64)0;
2300 	key_end.objectid = base + offset + nstripes * increment;
2301 	key_end.type = BTRFS_EXTENT_ITEM_KEY;
2302 	key_end.offset = (u64)0;
2303 	reada1 = btrfs_reada_add(root, &key_start, &key_end);
2304 
2305 	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2306 	key_start.type = BTRFS_EXTENT_CSUM_KEY;
2307 	key_start.offset = logical;
2308 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2309 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
2310 	key_end.offset = base + offset + nstripes * increment;
2311 	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2312 
2313 	if (!IS_ERR(reada1))
2314 		btrfs_reada_wait(reada1);
2315 	if (!IS_ERR(reada2))
2316 		btrfs_reada_wait(reada2);
2317 
2318 	mutex_lock(&fs_info->scrub_lock);
2319 	while (atomic_read(&fs_info->scrub_pause_req)) {
2320 		mutex_unlock(&fs_info->scrub_lock);
2321 		wait_event(fs_info->scrub_pause_wait,
2322 		   atomic_read(&fs_info->scrub_pause_req) == 0);
2323 		mutex_lock(&fs_info->scrub_lock);
2324 	}
2325 	atomic_dec(&fs_info->scrubs_paused);
2326 	mutex_unlock(&fs_info->scrub_lock);
2327 	wake_up(&fs_info->scrub_pause_wait);
2328 
2329 	/*
2330 	 * collect all data csums for the stripe to avoid seeking during
2331 	 * the scrub. This might currently (crc32) end up to be about 1MB
2332 	 */
2333 	blk_start_plug(&plug);
2334 
2335 	/*
2336 	 * now find all extents for each stripe and scrub them
2337 	 */
2338 	logical = base + offset;
2339 	physical = map->stripes[num].physical;
2340 	ret = 0;
2341 	for (i = 0; i < nstripes; ++i) {
2342 		/*
2343 		 * canceled?
2344 		 */
2345 		if (atomic_read(&fs_info->scrub_cancel_req) ||
2346 		    atomic_read(&sctx->cancel_req)) {
2347 			ret = -ECANCELED;
2348 			goto out;
2349 		}
2350 		/*
2351 		 * check to see if we have to pause
2352 		 */
2353 		if (atomic_read(&fs_info->scrub_pause_req)) {
2354 			/* push queued extents */
2355 			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2356 			scrub_submit(sctx);
2357 			mutex_lock(&sctx->wr_ctx.wr_lock);
2358 			scrub_wr_submit(sctx);
2359 			mutex_unlock(&sctx->wr_ctx.wr_lock);
2360 			wait_event(sctx->list_wait,
2361 				   atomic_read(&sctx->bios_in_flight) == 0);
2362 			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2363 			atomic_inc(&fs_info->scrubs_paused);
2364 			wake_up(&fs_info->scrub_pause_wait);
2365 			mutex_lock(&fs_info->scrub_lock);
2366 			while (atomic_read(&fs_info->scrub_pause_req)) {
2367 				mutex_unlock(&fs_info->scrub_lock);
2368 				wait_event(fs_info->scrub_pause_wait,
2369 				   atomic_read(&fs_info->scrub_pause_req) == 0);
2370 				mutex_lock(&fs_info->scrub_lock);
2371 			}
2372 			atomic_dec(&fs_info->scrubs_paused);
2373 			mutex_unlock(&fs_info->scrub_lock);
2374 			wake_up(&fs_info->scrub_pause_wait);
2375 		}
2376 
2377 		ret = btrfs_lookup_csums_range(csum_root, logical,
2378 					       logical + map->stripe_len - 1,
2379 					       &sctx->csum_list, 1);
2380 		if (ret)
2381 			goto out;
2382 
2383 		key.objectid = logical;
2384 		key.type = BTRFS_EXTENT_ITEM_KEY;
2385 		key.offset = (u64)0;
2386 
2387 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2388 		if (ret < 0)
2389 			goto out;
2390 		if (ret > 0) {
2391 			ret = btrfs_previous_item(root, path, 0,
2392 						  BTRFS_EXTENT_ITEM_KEY);
2393 			if (ret < 0)
2394 				goto out;
2395 			if (ret > 0) {
2396 				/* there's no smaller item, so stick with the
2397 				 * larger one */
2398 				btrfs_release_path(path);
2399 				ret = btrfs_search_slot(NULL, root, &key,
2400 							path, 0, 0);
2401 				if (ret < 0)
2402 					goto out;
2403 			}
2404 		}
2405 
2406 		while (1) {
2407 			l = path->nodes[0];
2408 			slot = path->slots[0];
2409 			if (slot >= btrfs_header_nritems(l)) {
2410 				ret = btrfs_next_leaf(root, path);
2411 				if (ret == 0)
2412 					continue;
2413 				if (ret < 0)
2414 					goto out;
2415 
2416 				break;
2417 			}
2418 			btrfs_item_key_to_cpu(l, &key, slot);
2419 
2420 			if (key.objectid + key.offset <= logical)
2421 				goto next;
2422 
2423 			if (key.objectid >= logical + map->stripe_len)
2424 				break;
2425 
2426 			if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
2427 				goto next;
2428 
2429 			extent = btrfs_item_ptr(l, slot,
2430 						struct btrfs_extent_item);
2431 			flags = btrfs_extent_flags(l, extent);
2432 			generation = btrfs_extent_generation(l, extent);
2433 
2434 			if (key.objectid < logical &&
2435 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2436 				printk(KERN_ERR
2437 				       "btrfs scrub: tree block %llu spanning "
2438 				       "stripes, ignored. logical=%llu\n",
2439 				       (unsigned long long)key.objectid,
2440 				       (unsigned long long)logical);
2441 				goto next;
2442 			}
2443 
2444 			/*
2445 			 * trim extent to this stripe
2446 			 */
2447 			if (key.objectid < logical) {
2448 				key.offset -= logical - key.objectid;
2449 				key.objectid = logical;
2450 			}
2451 			if (key.objectid + key.offset >
2452 			    logical + map->stripe_len) {
2453 				key.offset = logical + map->stripe_len -
2454 					     key.objectid;
2455 			}
2456 
2457 			extent_logical = key.objectid;
2458 			extent_physical = key.objectid - logical + physical;
2459 			extent_len = key.offset;
2460 			extent_dev = scrub_dev;
2461 			extent_mirror_num = mirror_num;
2462 			if (is_dev_replace)
2463 				scrub_remap_extent(fs_info, extent_logical,
2464 						   extent_len, &extent_physical,
2465 						   &extent_dev,
2466 						   &extent_mirror_num);
2467 			ret = scrub_extent(sctx, extent_logical, extent_len,
2468 					   extent_physical, extent_dev, flags,
2469 					   generation, extent_mirror_num,
2470 					   key.objectid - logical + physical);
2471 			if (ret)
2472 				goto out;
2473 
2474 next:
2475 			path->slots[0]++;
2476 		}
2477 		btrfs_release_path(path);
2478 		logical += increment;
2479 		physical += map->stripe_len;
2480 		spin_lock(&sctx->stat_lock);
2481 		sctx->stat.last_physical = physical;
2482 		spin_unlock(&sctx->stat_lock);
2483 	}
2484 out:
2485 	/* push queued extents */
2486 	scrub_submit(sctx);
2487 	mutex_lock(&sctx->wr_ctx.wr_lock);
2488 	scrub_wr_submit(sctx);
2489 	mutex_unlock(&sctx->wr_ctx.wr_lock);
2490 
2491 	blk_finish_plug(&plug);
2492 	btrfs_free_path(path);
2493 	return ret < 0 ? ret : 0;
2494 }
2495 
2496 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2497 					  struct btrfs_device *scrub_dev,
2498 					  u64 chunk_tree, u64 chunk_objectid,
2499 					  u64 chunk_offset, u64 length,
2500 					  u64 dev_offset, int is_dev_replace)
2501 {
2502 	struct btrfs_mapping_tree *map_tree =
2503 		&sctx->dev_root->fs_info->mapping_tree;
2504 	struct map_lookup *map;
2505 	struct extent_map *em;
2506 	int i;
2507 	int ret = 0;
2508 
2509 	read_lock(&map_tree->map_tree.lock);
2510 	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2511 	read_unlock(&map_tree->map_tree.lock);
2512 
2513 	if (!em)
2514 		return -EINVAL;
2515 
2516 	map = (struct map_lookup *)em->bdev;
2517 	if (em->start != chunk_offset)
2518 		goto out;
2519 
2520 	if (em->len < length)
2521 		goto out;
2522 
2523 	for (i = 0; i < map->num_stripes; ++i) {
2524 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2525 		    map->stripes[i].physical == dev_offset) {
2526 			ret = scrub_stripe(sctx, map, scrub_dev, i,
2527 					   chunk_offset, length,
2528 					   is_dev_replace);
2529 			if (ret)
2530 				goto out;
2531 		}
2532 	}
2533 out:
2534 	free_extent_map(em);
2535 
2536 	return ret;
2537 }
2538 
2539 static noinline_for_stack
2540 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2541 			   struct btrfs_device *scrub_dev, u64 start, u64 end,
2542 			   int is_dev_replace)
2543 {
2544 	struct btrfs_dev_extent *dev_extent = NULL;
2545 	struct btrfs_path *path;
2546 	struct btrfs_root *root = sctx->dev_root;
2547 	struct btrfs_fs_info *fs_info = root->fs_info;
2548 	u64 length;
2549 	u64 chunk_tree;
2550 	u64 chunk_objectid;
2551 	u64 chunk_offset;
2552 	int ret;
2553 	int slot;
2554 	struct extent_buffer *l;
2555 	struct btrfs_key key;
2556 	struct btrfs_key found_key;
2557 	struct btrfs_block_group_cache *cache;
2558 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2559 
2560 	path = btrfs_alloc_path();
2561 	if (!path)
2562 		return -ENOMEM;
2563 
2564 	path->reada = 2;
2565 	path->search_commit_root = 1;
2566 	path->skip_locking = 1;
2567 
2568 	key.objectid = scrub_dev->devid;
2569 	key.offset = 0ull;
2570 	key.type = BTRFS_DEV_EXTENT_KEY;
2571 
2572 	while (1) {
2573 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2574 		if (ret < 0)
2575 			break;
2576 		if (ret > 0) {
2577 			if (path->slots[0] >=
2578 			    btrfs_header_nritems(path->nodes[0])) {
2579 				ret = btrfs_next_leaf(root, path);
2580 				if (ret)
2581 					break;
2582 			}
2583 		}
2584 
2585 		l = path->nodes[0];
2586 		slot = path->slots[0];
2587 
2588 		btrfs_item_key_to_cpu(l, &found_key, slot);
2589 
2590 		if (found_key.objectid != scrub_dev->devid)
2591 			break;
2592 
2593 		if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2594 			break;
2595 
2596 		if (found_key.offset >= end)
2597 			break;
2598 
2599 		if (found_key.offset < key.offset)
2600 			break;
2601 
2602 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2603 		length = btrfs_dev_extent_length(l, dev_extent);
2604 
2605 		if (found_key.offset + length <= start) {
2606 			key.offset = found_key.offset + length;
2607 			btrfs_release_path(path);
2608 			continue;
2609 		}
2610 
2611 		chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2612 		chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2613 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2614 
2615 		/*
2616 		 * get a reference on the corresponding block group to prevent
2617 		 * the chunk from going away while we scrub it
2618 		 */
2619 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2620 		if (!cache) {
2621 			ret = -ENOENT;
2622 			break;
2623 		}
2624 		dev_replace->cursor_right = found_key.offset + length;
2625 		dev_replace->cursor_left = found_key.offset;
2626 		dev_replace->item_needs_writeback = 1;
2627 		ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2628 				  chunk_offset, length, found_key.offset,
2629 				  is_dev_replace);
2630 
2631 		/*
2632 		 * flush, submit all pending read and write bios, afterwards
2633 		 * wait for them.
2634 		 * Note that in the dev replace case, a read request causes
2635 		 * write requests that are submitted in the read completion
2636 		 * worker. Therefore in the current situation, it is required
2637 		 * that all write requests are flushed, so that all read and
2638 		 * write requests are really completed when bios_in_flight
2639 		 * changes to 0.
2640 		 */
2641 		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2642 		scrub_submit(sctx);
2643 		mutex_lock(&sctx->wr_ctx.wr_lock);
2644 		scrub_wr_submit(sctx);
2645 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2646 
2647 		wait_event(sctx->list_wait,
2648 			   atomic_read(&sctx->bios_in_flight) == 0);
2649 		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2650 		atomic_inc(&fs_info->scrubs_paused);
2651 		wake_up(&fs_info->scrub_pause_wait);
2652 		wait_event(sctx->list_wait,
2653 			   atomic_read(&sctx->workers_pending) == 0);
2654 
2655 		mutex_lock(&fs_info->scrub_lock);
2656 		while (atomic_read(&fs_info->scrub_pause_req)) {
2657 			mutex_unlock(&fs_info->scrub_lock);
2658 			wait_event(fs_info->scrub_pause_wait,
2659 			   atomic_read(&fs_info->scrub_pause_req) == 0);
2660 			mutex_lock(&fs_info->scrub_lock);
2661 		}
2662 		atomic_dec(&fs_info->scrubs_paused);
2663 		mutex_unlock(&fs_info->scrub_lock);
2664 		wake_up(&fs_info->scrub_pause_wait);
2665 
2666 		dev_replace->cursor_left = dev_replace->cursor_right;
2667 		dev_replace->item_needs_writeback = 1;
2668 		btrfs_put_block_group(cache);
2669 		if (ret)
2670 			break;
2671 		if (is_dev_replace &&
2672 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
2673 			ret = -EIO;
2674 			break;
2675 		}
2676 		if (sctx->stat.malloc_errors > 0) {
2677 			ret = -ENOMEM;
2678 			break;
2679 		}
2680 
2681 		key.offset = found_key.offset + length;
2682 		btrfs_release_path(path);
2683 	}
2684 
2685 	btrfs_free_path(path);
2686 
2687 	/*
2688 	 * ret can still be 1 from search_slot or next_leaf,
2689 	 * that's not an error
2690 	 */
2691 	return ret < 0 ? ret : 0;
2692 }
2693 
2694 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2695 					   struct btrfs_device *scrub_dev)
2696 {
2697 	int	i;
2698 	u64	bytenr;
2699 	u64	gen;
2700 	int	ret;
2701 	struct btrfs_root *root = sctx->dev_root;
2702 
2703 	if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2704 		return -EIO;
2705 
2706 	gen = root->fs_info->last_trans_committed;
2707 
2708 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2709 		bytenr = btrfs_sb_offset(i);
2710 		if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2711 			break;
2712 
2713 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2714 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2715 				  NULL, 1, bytenr);
2716 		if (ret)
2717 			return ret;
2718 	}
2719 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2720 
2721 	return 0;
2722 }
2723 
2724 /*
2725  * get a reference count on fs_info->scrub_workers. start worker if necessary
2726  */
2727 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2728 						int is_dev_replace)
2729 {
2730 	int ret = 0;
2731 
2732 	mutex_lock(&fs_info->scrub_lock);
2733 	if (fs_info->scrub_workers_refcnt == 0) {
2734 		if (is_dev_replace)
2735 			btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2736 					&fs_info->generic_worker);
2737 		else
2738 			btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2739 					fs_info->thread_pool_size,
2740 					&fs_info->generic_worker);
2741 		fs_info->scrub_workers.idle_thresh = 4;
2742 		ret = btrfs_start_workers(&fs_info->scrub_workers);
2743 		if (ret)
2744 			goto out;
2745 		btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2746 				   "scrubwrc",
2747 				   fs_info->thread_pool_size,
2748 				   &fs_info->generic_worker);
2749 		fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2750 		ret = btrfs_start_workers(
2751 				&fs_info->scrub_wr_completion_workers);
2752 		if (ret)
2753 			goto out;
2754 		btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2755 				   &fs_info->generic_worker);
2756 		ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2757 		if (ret)
2758 			goto out;
2759 	}
2760 	++fs_info->scrub_workers_refcnt;
2761 out:
2762 	mutex_unlock(&fs_info->scrub_lock);
2763 
2764 	return ret;
2765 }
2766 
2767 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2768 {
2769 	mutex_lock(&fs_info->scrub_lock);
2770 	if (--fs_info->scrub_workers_refcnt == 0) {
2771 		btrfs_stop_workers(&fs_info->scrub_workers);
2772 		btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2773 		btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2774 	}
2775 	WARN_ON(fs_info->scrub_workers_refcnt < 0);
2776 	mutex_unlock(&fs_info->scrub_lock);
2777 }
2778 
2779 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2780 		    u64 end, struct btrfs_scrub_progress *progress,
2781 		    int readonly, int is_dev_replace)
2782 {
2783 	struct scrub_ctx *sctx;
2784 	int ret;
2785 	struct btrfs_device *dev;
2786 
2787 	if (btrfs_fs_closing(fs_info))
2788 		return -EINVAL;
2789 
2790 	/*
2791 	 * check some assumptions
2792 	 */
2793 	if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2794 		printk(KERN_ERR
2795 		       "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2796 		       fs_info->chunk_root->nodesize,
2797 		       fs_info->chunk_root->leafsize);
2798 		return -EINVAL;
2799 	}
2800 
2801 	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2802 		/*
2803 		 * in this case scrub is unable to calculate the checksum
2804 		 * the way scrub is implemented. Do not handle this
2805 		 * situation at all because it won't ever happen.
2806 		 */
2807 		printk(KERN_ERR
2808 		       "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2809 		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2810 		return -EINVAL;
2811 	}
2812 
2813 	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2814 		/* not supported for data w/o checksums */
2815 		printk(KERN_ERR
2816 		       "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2817 		       fs_info->chunk_root->sectorsize,
2818 		       (unsigned long long)PAGE_SIZE);
2819 		return -EINVAL;
2820 	}
2821 
2822 	if (fs_info->chunk_root->nodesize >
2823 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2824 	    fs_info->chunk_root->sectorsize >
2825 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2826 		/*
2827 		 * would exhaust the array bounds of pagev member in
2828 		 * struct scrub_block
2829 		 */
2830 		pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2831 		       fs_info->chunk_root->nodesize,
2832 		       SCRUB_MAX_PAGES_PER_BLOCK,
2833 		       fs_info->chunk_root->sectorsize,
2834 		       SCRUB_MAX_PAGES_PER_BLOCK);
2835 		return -EINVAL;
2836 	}
2837 
2838 	ret = scrub_workers_get(fs_info, is_dev_replace);
2839 	if (ret)
2840 		return ret;
2841 
2842 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2843 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2844 	if (!dev || (dev->missing && !is_dev_replace)) {
2845 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2846 		scrub_workers_put(fs_info);
2847 		return -ENODEV;
2848 	}
2849 	mutex_lock(&fs_info->scrub_lock);
2850 
2851 	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2852 		mutex_unlock(&fs_info->scrub_lock);
2853 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2854 		scrub_workers_put(fs_info);
2855 		return -EIO;
2856 	}
2857 
2858 	btrfs_dev_replace_lock(&fs_info->dev_replace);
2859 	if (dev->scrub_device ||
2860 	    (!is_dev_replace &&
2861 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2862 		btrfs_dev_replace_unlock(&fs_info->dev_replace);
2863 		mutex_unlock(&fs_info->scrub_lock);
2864 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2865 		scrub_workers_put(fs_info);
2866 		return -EINPROGRESS;
2867 	}
2868 	btrfs_dev_replace_unlock(&fs_info->dev_replace);
2869 	sctx = scrub_setup_ctx(dev, is_dev_replace);
2870 	if (IS_ERR(sctx)) {
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 PTR_ERR(sctx);
2875 	}
2876 	sctx->readonly = readonly;
2877 	dev->scrub_device = sctx;
2878 
2879 	atomic_inc(&fs_info->scrubs_running);
2880 	mutex_unlock(&fs_info->scrub_lock);
2881 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2882 
2883 	if (!is_dev_replace) {
2884 		down_read(&fs_info->scrub_super_lock);
2885 		ret = scrub_supers(sctx, dev);
2886 		up_read(&fs_info->scrub_super_lock);
2887 	}
2888 
2889 	if (!ret)
2890 		ret = scrub_enumerate_chunks(sctx, dev, start, end,
2891 					     is_dev_replace);
2892 
2893 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2894 	atomic_dec(&fs_info->scrubs_running);
2895 	wake_up(&fs_info->scrub_pause_wait);
2896 
2897 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2898 
2899 	if (progress)
2900 		memcpy(progress, &sctx->stat, sizeof(*progress));
2901 
2902 	mutex_lock(&fs_info->scrub_lock);
2903 	dev->scrub_device = NULL;
2904 	mutex_unlock(&fs_info->scrub_lock);
2905 
2906 	scrub_free_ctx(sctx);
2907 	scrub_workers_put(fs_info);
2908 
2909 	return ret;
2910 }
2911 
2912 void btrfs_scrub_pause(struct btrfs_root *root)
2913 {
2914 	struct btrfs_fs_info *fs_info = root->fs_info;
2915 
2916 	mutex_lock(&fs_info->scrub_lock);
2917 	atomic_inc(&fs_info->scrub_pause_req);
2918 	while (atomic_read(&fs_info->scrubs_paused) !=
2919 	       atomic_read(&fs_info->scrubs_running)) {
2920 		mutex_unlock(&fs_info->scrub_lock);
2921 		wait_event(fs_info->scrub_pause_wait,
2922 			   atomic_read(&fs_info->scrubs_paused) ==
2923 			   atomic_read(&fs_info->scrubs_running));
2924 		mutex_lock(&fs_info->scrub_lock);
2925 	}
2926 	mutex_unlock(&fs_info->scrub_lock);
2927 }
2928 
2929 void btrfs_scrub_continue(struct btrfs_root *root)
2930 {
2931 	struct btrfs_fs_info *fs_info = root->fs_info;
2932 
2933 	atomic_dec(&fs_info->scrub_pause_req);
2934 	wake_up(&fs_info->scrub_pause_wait);
2935 }
2936 
2937 void btrfs_scrub_pause_super(struct btrfs_root *root)
2938 {
2939 	down_write(&root->fs_info->scrub_super_lock);
2940 }
2941 
2942 void btrfs_scrub_continue_super(struct btrfs_root *root)
2943 {
2944 	up_write(&root->fs_info->scrub_super_lock);
2945 }
2946 
2947 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2948 {
2949 	mutex_lock(&fs_info->scrub_lock);
2950 	if (!atomic_read(&fs_info->scrubs_running)) {
2951 		mutex_unlock(&fs_info->scrub_lock);
2952 		return -ENOTCONN;
2953 	}
2954 
2955 	atomic_inc(&fs_info->scrub_cancel_req);
2956 	while (atomic_read(&fs_info->scrubs_running)) {
2957 		mutex_unlock(&fs_info->scrub_lock);
2958 		wait_event(fs_info->scrub_pause_wait,
2959 			   atomic_read(&fs_info->scrubs_running) == 0);
2960 		mutex_lock(&fs_info->scrub_lock);
2961 	}
2962 	atomic_dec(&fs_info->scrub_cancel_req);
2963 	mutex_unlock(&fs_info->scrub_lock);
2964 
2965 	return 0;
2966 }
2967 
2968 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2969 			   struct btrfs_device *dev)
2970 {
2971 	struct scrub_ctx *sctx;
2972 
2973 	mutex_lock(&fs_info->scrub_lock);
2974 	sctx = dev->scrub_device;
2975 	if (!sctx) {
2976 		mutex_unlock(&fs_info->scrub_lock);
2977 		return -ENOTCONN;
2978 	}
2979 	atomic_inc(&sctx->cancel_req);
2980 	while (dev->scrub_device) {
2981 		mutex_unlock(&fs_info->scrub_lock);
2982 		wait_event(fs_info->scrub_pause_wait,
2983 			   dev->scrub_device == NULL);
2984 		mutex_lock(&fs_info->scrub_lock);
2985 	}
2986 	mutex_unlock(&fs_info->scrub_lock);
2987 
2988 	return 0;
2989 }
2990 
2991 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2992 {
2993 	struct btrfs_fs_info *fs_info = root->fs_info;
2994 	struct btrfs_device *dev;
2995 	int ret;
2996 
2997 	/*
2998 	 * we have to hold the device_list_mutex here so the device
2999 	 * does not go away in cancel_dev. FIXME: find a better solution
3000 	 */
3001 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3002 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3003 	if (!dev) {
3004 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3005 		return -ENODEV;
3006 	}
3007 	ret = btrfs_scrub_cancel_dev(fs_info, dev);
3008 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3009 
3010 	return ret;
3011 }
3012 
3013 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3014 			 struct btrfs_scrub_progress *progress)
3015 {
3016 	struct btrfs_device *dev;
3017 	struct scrub_ctx *sctx = NULL;
3018 
3019 	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3020 	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3021 	if (dev)
3022 		sctx = dev->scrub_device;
3023 	if (sctx)
3024 		memcpy(progress, &sctx->stat, sizeof(*progress));
3025 	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3026 
3027 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3028 }
3029 
3030 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3031 			       u64 extent_logical, u64 extent_len,
3032 			       u64 *extent_physical,
3033 			       struct btrfs_device **extent_dev,
3034 			       int *extent_mirror_num)
3035 {
3036 	u64 mapped_length;
3037 	struct btrfs_bio *bbio = NULL;
3038 	int ret;
3039 
3040 	mapped_length = extent_len;
3041 	ret = btrfs_map_block(fs_info, READ, extent_logical,
3042 			      &mapped_length, &bbio, 0);
3043 	if (ret || !bbio || mapped_length < extent_len ||
3044 	    !bbio->stripes[0].dev->bdev) {
3045 		kfree(bbio);
3046 		return;
3047 	}
3048 
3049 	*extent_physical = bbio->stripes[0].physical;
3050 	*extent_mirror_num = bbio->mirror_num;
3051 	*extent_dev = bbio->stripes[0].dev;
3052 	kfree(bbio);
3053 }
3054 
3055 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3056 			      struct scrub_wr_ctx *wr_ctx,
3057 			      struct btrfs_fs_info *fs_info,
3058 			      struct btrfs_device *dev,
3059 			      int is_dev_replace)
3060 {
3061 	WARN_ON(wr_ctx->wr_curr_bio != NULL);
3062 
3063 	mutex_init(&wr_ctx->wr_lock);
3064 	wr_ctx->wr_curr_bio = NULL;
3065 	if (!is_dev_replace)
3066 		return 0;
3067 
3068 	WARN_ON(!dev->bdev);
3069 	wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3070 					 bio_get_nr_vecs(dev->bdev));
3071 	wr_ctx->tgtdev = dev;
3072 	atomic_set(&wr_ctx->flush_all_writes, 0);
3073 	return 0;
3074 }
3075 
3076 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3077 {
3078 	mutex_lock(&wr_ctx->wr_lock);
3079 	kfree(wr_ctx->wr_curr_bio);
3080 	wr_ctx->wr_curr_bio = NULL;
3081 	mutex_unlock(&wr_ctx->wr_lock);
3082 }
3083 
3084 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3085 			    int mirror_num, u64 physical_for_dev_replace)
3086 {
3087 	struct scrub_copy_nocow_ctx *nocow_ctx;
3088 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3089 
3090 	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3091 	if (!nocow_ctx) {
3092 		spin_lock(&sctx->stat_lock);
3093 		sctx->stat.malloc_errors++;
3094 		spin_unlock(&sctx->stat_lock);
3095 		return -ENOMEM;
3096 	}
3097 
3098 	scrub_pending_trans_workers_inc(sctx);
3099 
3100 	nocow_ctx->sctx = sctx;
3101 	nocow_ctx->logical = logical;
3102 	nocow_ctx->len = len;
3103 	nocow_ctx->mirror_num = mirror_num;
3104 	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3105 	nocow_ctx->work.func = copy_nocow_pages_worker;
3106 	btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3107 			   &nocow_ctx->work);
3108 
3109 	return 0;
3110 }
3111 
3112 static void copy_nocow_pages_worker(struct btrfs_work *work)
3113 {
3114 	struct scrub_copy_nocow_ctx *nocow_ctx =
3115 		container_of(work, struct scrub_copy_nocow_ctx, work);
3116 	struct scrub_ctx *sctx = nocow_ctx->sctx;
3117 	u64 logical = nocow_ctx->logical;
3118 	u64 len = nocow_ctx->len;
3119 	int mirror_num = nocow_ctx->mirror_num;
3120 	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3121 	int ret;
3122 	struct btrfs_trans_handle *trans = NULL;
3123 	struct btrfs_fs_info *fs_info;
3124 	struct btrfs_path *path;
3125 	struct btrfs_root *root;
3126 	int not_written = 0;
3127 
3128 	fs_info = sctx->dev_root->fs_info;
3129 	root = fs_info->extent_root;
3130 
3131 	path = btrfs_alloc_path();
3132 	if (!path) {
3133 		spin_lock(&sctx->stat_lock);
3134 		sctx->stat.malloc_errors++;
3135 		spin_unlock(&sctx->stat_lock);
3136 		not_written = 1;
3137 		goto out;
3138 	}
3139 
3140 	trans = btrfs_join_transaction(root);
3141 	if (IS_ERR(trans)) {
3142 		not_written = 1;
3143 		goto out;
3144 	}
3145 
3146 	ret = iterate_inodes_from_logical(logical, fs_info, path,
3147 					  copy_nocow_pages_for_inode,
3148 					  nocow_ctx);
3149 	if (ret != 0 && ret != -ENOENT) {
3150 		pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3151 			(unsigned long long)logical,
3152 			(unsigned long long)physical_for_dev_replace,
3153 			(unsigned long long)len,
3154 			(unsigned long long)mirror_num, ret);
3155 		not_written = 1;
3156 		goto out;
3157 	}
3158 
3159 out:
3160 	if (trans && !IS_ERR(trans))
3161 		btrfs_end_transaction(trans, root);
3162 	if (not_written)
3163 		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3164 					    num_uncorrectable_read_errors);
3165 
3166 	btrfs_free_path(path);
3167 	kfree(nocow_ctx);
3168 
3169 	scrub_pending_trans_workers_dec(sctx);
3170 }
3171 
3172 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3173 {
3174 	unsigned long index;
3175 	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3176 	int ret = 0;
3177 	struct btrfs_key key;
3178 	struct inode *inode = NULL;
3179 	struct btrfs_root *local_root;
3180 	u64 physical_for_dev_replace;
3181 	u64 len;
3182 	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3183 
3184 	key.objectid = root;
3185 	key.type = BTRFS_ROOT_ITEM_KEY;
3186 	key.offset = (u64)-1;
3187 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3188 	if (IS_ERR(local_root))
3189 		return PTR_ERR(local_root);
3190 
3191 	key.type = BTRFS_INODE_ITEM_KEY;
3192 	key.objectid = inum;
3193 	key.offset = 0;
3194 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3195 	if (IS_ERR(inode))
3196 		return PTR_ERR(inode);
3197 
3198 	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3199 	len = nocow_ctx->len;
3200 	while (len >= PAGE_CACHE_SIZE) {
3201 		struct page *page = NULL;
3202 		int ret_sub;
3203 
3204 		index = offset >> PAGE_CACHE_SHIFT;
3205 
3206 		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3207 		if (!page) {
3208 			pr_err("find_or_create_page() failed\n");
3209 			ret = -ENOMEM;
3210 			goto next_page;
3211 		}
3212 
3213 		if (PageUptodate(page)) {
3214 			if (PageDirty(page))
3215 				goto next_page;
3216 		} else {
3217 			ClearPageError(page);
3218 			ret_sub = extent_read_full_page(&BTRFS_I(inode)->
3219 							 io_tree,
3220 							page, btrfs_get_extent,
3221 							nocow_ctx->mirror_num);
3222 			if (ret_sub) {
3223 				ret = ret_sub;
3224 				goto next_page;
3225 			}
3226 			wait_on_page_locked(page);
3227 			if (!PageUptodate(page)) {
3228 				ret = -EIO;
3229 				goto next_page;
3230 			}
3231 		}
3232 		ret_sub = write_page_nocow(nocow_ctx->sctx,
3233 					   physical_for_dev_replace, page);
3234 		if (ret_sub) {
3235 			ret = ret_sub;
3236 			goto next_page;
3237 		}
3238 
3239 next_page:
3240 		if (page) {
3241 			unlock_page(page);
3242 			put_page(page);
3243 		}
3244 		offset += PAGE_CACHE_SIZE;
3245 		physical_for_dev_replace += PAGE_CACHE_SIZE;
3246 		len -= PAGE_CACHE_SIZE;
3247 	}
3248 
3249 	if (inode)
3250 		iput(inode);
3251 	return ret;
3252 }
3253 
3254 static int write_page_nocow(struct scrub_ctx *sctx,
3255 			    u64 physical_for_dev_replace, struct page *page)
3256 {
3257 	struct bio *bio;
3258 	struct btrfs_device *dev;
3259 	int ret;
3260 	DECLARE_COMPLETION_ONSTACK(compl);
3261 
3262 	dev = sctx->wr_ctx.tgtdev;
3263 	if (!dev)
3264 		return -EIO;
3265 	if (!dev->bdev) {
3266 		printk_ratelimited(KERN_WARNING
3267 			"btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3268 		return -EIO;
3269 	}
3270 	bio = bio_alloc(GFP_NOFS, 1);
3271 	if (!bio) {
3272 		spin_lock(&sctx->stat_lock);
3273 		sctx->stat.malloc_errors++;
3274 		spin_unlock(&sctx->stat_lock);
3275 		return -ENOMEM;
3276 	}
3277 	bio->bi_private = &compl;
3278 	bio->bi_end_io = scrub_complete_bio_end_io;
3279 	bio->bi_size = 0;
3280 	bio->bi_sector = physical_for_dev_replace >> 9;
3281 	bio->bi_bdev = dev->bdev;
3282 	ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3283 	if (ret != PAGE_CACHE_SIZE) {
3284 leave_with_eio:
3285 		bio_put(bio);
3286 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3287 		return -EIO;
3288 	}
3289 	btrfsic_submit_bio(WRITE_SYNC, bio);
3290 	wait_for_completion(&compl);
3291 
3292 	if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3293 		goto leave_with_eio;
3294 
3295 	bio_put(bio);
3296 	return 0;
3297 }
3298