xref: /linux/fs/btrfs/scrub.c (revision c75c5ab575af7db707689cdbb5a5c458e9a034bb)
1 /*
2  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/blkdev.h>
20 #include <linux/ratelimit.h>
21 #include "ctree.h"
22 #include "volumes.h"
23 #include "disk-io.h"
24 #include "ordered-data.h"
25 #include "transaction.h"
26 #include "backref.h"
27 #include "extent_io.h"
28 #include "dev-replace.h"
29 #include "check-integrity.h"
30 #include "rcu-string.h"
31 #include "raid56.h"
32 
33 /*
34  * This is only the first step towards a full-features scrub. It reads all
35  * extent and super block and verifies the checksums. In case a bad checksum
36  * is found or the extent cannot be read, good data will be written back if
37  * any can be found.
38  *
39  * Future enhancements:
40  *  - In case an unrepairable extent is encountered, track which files are
41  *    affected and report them
42  *  - track and record media errors, throw out bad devices
43  *  - add a mode to also read unallocated space
44  */
45 
46 struct scrub_block;
47 struct scrub_ctx;
48 
49 /*
50  * the following three values only influence the performance.
51  * The last one configures the number of parallel and outstanding I/O
52  * operations. The first two values configure an upper limit for the number
53  * of (dynamically allocated) pages that are added to a bio.
54  */
55 #define SCRUB_PAGES_PER_RD_BIO	32	/* 128k per bio */
56 #define SCRUB_PAGES_PER_WR_BIO	32	/* 128k per bio */
57 #define SCRUB_BIOS_PER_SCTX	64	/* 8MB per device in flight */
58 
59 /*
60  * the following value times PAGE_SIZE needs to be large enough to match the
61  * largest node/leaf/sector size that shall be supported.
62  * Values larger than BTRFS_STRIPE_LEN are not supported.
63  */
64 #define SCRUB_MAX_PAGES_PER_BLOCK	16	/* 64k per node/leaf/sector */
65 
66 struct scrub_page {
67 	struct scrub_block	*sblock;
68 	struct page		*page;
69 	struct btrfs_device	*dev;
70 	u64			flags;  /* extent flags */
71 	u64			generation;
72 	u64			logical;
73 	u64			physical;
74 	u64			physical_for_dev_replace;
75 	atomic_t		ref_count;
76 	struct {
77 		unsigned int	mirror_num:8;
78 		unsigned int	have_csum:1;
79 		unsigned int	io_error:1;
80 	};
81 	u8			csum[BTRFS_CSUM_SIZE];
82 };
83 
84 struct scrub_bio {
85 	int			index;
86 	struct scrub_ctx	*sctx;
87 	struct btrfs_device	*dev;
88 	struct bio		*bio;
89 	int			err;
90 	u64			logical;
91 	u64			physical;
92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 	struct scrub_page	*pagev[SCRUB_PAGES_PER_WR_BIO];
94 #else
95 	struct scrub_page	*pagev[SCRUB_PAGES_PER_RD_BIO];
96 #endif
97 	int			page_count;
98 	int			next_free;
99 	struct btrfs_work	work;
100 };
101 
102 struct scrub_block {
103 	struct scrub_page	*pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 	int			page_count;
105 	atomic_t		outstanding_pages;
106 	atomic_t		ref_count; /* free mem on transition to zero */
107 	struct scrub_ctx	*sctx;
108 	struct {
109 		unsigned int	header_error:1;
110 		unsigned int	checksum_error:1;
111 		unsigned int	no_io_error_seen:1;
112 		unsigned int	generation_error:1; /* also sets header_error */
113 	};
114 };
115 
116 struct scrub_wr_ctx {
117 	struct scrub_bio *wr_curr_bio;
118 	struct btrfs_device *tgtdev;
119 	int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
120 	atomic_t flush_all_writes;
121 	struct mutex wr_lock;
122 };
123 
124 struct scrub_ctx {
125 	struct scrub_bio	*bios[SCRUB_BIOS_PER_SCTX];
126 	struct btrfs_root	*dev_root;
127 	int			first_free;
128 	int			curr;
129 	atomic_t		bios_in_flight;
130 	atomic_t		workers_pending;
131 	spinlock_t		list_lock;
132 	wait_queue_head_t	list_wait;
133 	u16			csum_size;
134 	struct list_head	csum_list;
135 	atomic_t		cancel_req;
136 	int			readonly;
137 	int			pages_per_rd_bio;
138 	u32			sectorsize;
139 	u32			nodesize;
140 	u32			leafsize;
141 
142 	int			is_dev_replace;
143 	struct scrub_wr_ctx	wr_ctx;
144 
145 	/*
146 	 * statistics
147 	 */
148 	struct btrfs_scrub_progress stat;
149 	spinlock_t		stat_lock;
150 };
151 
152 struct scrub_fixup_nodatasum {
153 	struct scrub_ctx	*sctx;
154 	struct btrfs_device	*dev;
155 	u64			logical;
156 	struct btrfs_root	*root;
157 	struct btrfs_work	work;
158 	int			mirror_num;
159 };
160 
161 struct scrub_copy_nocow_ctx {
162 	struct scrub_ctx	*sctx;
163 	u64			logical;
164 	u64			len;
165 	int			mirror_num;
166 	u64			physical_for_dev_replace;
167 	struct btrfs_work	work;
168 };
169 
170 struct scrub_warning {
171 	struct btrfs_path	*path;
172 	u64			extent_item_size;
173 	char			*scratch_buf;
174 	char			*msg_buf;
175 	const char		*errstr;
176 	sector_t		sector;
177 	u64			logical;
178 	struct btrfs_device	*dev;
179 	int			msg_bufsize;
180 	int			scratch_bufsize;
181 };
182 
183 
184 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
185 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
186 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
187 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
188 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
189 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
190 				     struct btrfs_fs_info *fs_info,
191 				     struct scrub_block *original_sblock,
192 				     u64 length, u64 logical,
193 				     struct scrub_block *sblocks_for_recheck);
194 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
195 				struct scrub_block *sblock, int is_metadata,
196 				int have_csum, u8 *csum, u64 generation,
197 				u16 csum_size);
198 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
199 					 struct scrub_block *sblock,
200 					 int is_metadata, int have_csum,
201 					 const u8 *csum, u64 generation,
202 					 u16 csum_size);
203 static void scrub_complete_bio_end_io(struct bio *bio, int err);
204 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
205 					     struct scrub_block *sblock_good,
206 					     int force_write);
207 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
208 					    struct scrub_block *sblock_good,
209 					    int page_num, int force_write);
210 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
211 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
212 					   int page_num);
213 static int scrub_checksum_data(struct scrub_block *sblock);
214 static int scrub_checksum_tree_block(struct scrub_block *sblock);
215 static int scrub_checksum_super(struct scrub_block *sblock);
216 static void scrub_block_get(struct scrub_block *sblock);
217 static void scrub_block_put(struct scrub_block *sblock);
218 static void scrub_page_get(struct scrub_page *spage);
219 static void scrub_page_put(struct scrub_page *spage);
220 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
221 				    struct scrub_page *spage);
222 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
223 		       u64 physical, struct btrfs_device *dev, u64 flags,
224 		       u64 gen, int mirror_num, u8 *csum, int force,
225 		       u64 physical_for_dev_replace);
226 static void scrub_bio_end_io(struct bio *bio, int err);
227 static void scrub_bio_end_io_worker(struct btrfs_work *work);
228 static void scrub_block_complete(struct scrub_block *sblock);
229 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
230 			       u64 extent_logical, u64 extent_len,
231 			       u64 *extent_physical,
232 			       struct btrfs_device **extent_dev,
233 			       int *extent_mirror_num);
234 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
235 			      struct scrub_wr_ctx *wr_ctx,
236 			      struct btrfs_fs_info *fs_info,
237 			      struct btrfs_device *dev,
238 			      int is_dev_replace);
239 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
240 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
241 				    struct scrub_page *spage);
242 static void scrub_wr_submit(struct scrub_ctx *sctx);
243 static void scrub_wr_bio_end_io(struct bio *bio, int err);
244 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
245 static int write_page_nocow(struct scrub_ctx *sctx,
246 			    u64 physical_for_dev_replace, struct page *page);
247 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
248 				      void *ctx);
249 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
250 			    int mirror_num, u64 physical_for_dev_replace);
251 static void copy_nocow_pages_worker(struct btrfs_work *work);
252 
253 
254 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
255 {
256 	atomic_inc(&sctx->bios_in_flight);
257 }
258 
259 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
260 {
261 	atomic_dec(&sctx->bios_in_flight);
262 	wake_up(&sctx->list_wait);
263 }
264 
265 /*
266  * used for workers that require transaction commits (i.e., for the
267  * NOCOW case)
268  */
269 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
270 {
271 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
272 
273 	/*
274 	 * increment scrubs_running to prevent cancel requests from
275 	 * completing as long as a worker is running. we must also
276 	 * increment scrubs_paused to prevent deadlocking on pause
277 	 * requests used for transactions commits (as the worker uses a
278 	 * transaction context). it is safe to regard the worker
279 	 * as paused for all matters practical. effectively, we only
280 	 * avoid cancellation requests from completing.
281 	 */
282 	mutex_lock(&fs_info->scrub_lock);
283 	atomic_inc(&fs_info->scrubs_running);
284 	atomic_inc(&fs_info->scrubs_paused);
285 	mutex_unlock(&fs_info->scrub_lock);
286 	atomic_inc(&sctx->workers_pending);
287 }
288 
289 /* used for workers that require transaction commits */
290 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
291 {
292 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
293 
294 	/*
295 	 * see scrub_pending_trans_workers_inc() why we're pretending
296 	 * to be paused in the scrub counters
297 	 */
298 	mutex_lock(&fs_info->scrub_lock);
299 	atomic_dec(&fs_info->scrubs_running);
300 	atomic_dec(&fs_info->scrubs_paused);
301 	mutex_unlock(&fs_info->scrub_lock);
302 	atomic_dec(&sctx->workers_pending);
303 	wake_up(&fs_info->scrub_pause_wait);
304 	wake_up(&sctx->list_wait);
305 }
306 
307 static void scrub_free_csums(struct scrub_ctx *sctx)
308 {
309 	while (!list_empty(&sctx->csum_list)) {
310 		struct btrfs_ordered_sum *sum;
311 		sum = list_first_entry(&sctx->csum_list,
312 				       struct btrfs_ordered_sum, list);
313 		list_del(&sum->list);
314 		kfree(sum);
315 	}
316 }
317 
318 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
319 {
320 	int i;
321 
322 	if (!sctx)
323 		return;
324 
325 	scrub_free_wr_ctx(&sctx->wr_ctx);
326 
327 	/* this can happen when scrub is cancelled */
328 	if (sctx->curr != -1) {
329 		struct scrub_bio *sbio = sctx->bios[sctx->curr];
330 
331 		for (i = 0; i < sbio->page_count; i++) {
332 			WARN_ON(!sbio->pagev[i]->page);
333 			scrub_block_put(sbio->pagev[i]->sblock);
334 		}
335 		bio_put(sbio->bio);
336 	}
337 
338 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
339 		struct scrub_bio *sbio = sctx->bios[i];
340 
341 		if (!sbio)
342 			break;
343 		kfree(sbio);
344 	}
345 
346 	scrub_free_csums(sctx);
347 	kfree(sctx);
348 }
349 
350 static noinline_for_stack
351 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
352 {
353 	struct scrub_ctx *sctx;
354 	int		i;
355 	struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
356 	int pages_per_rd_bio;
357 	int ret;
358 
359 	/*
360 	 * the setting of pages_per_rd_bio is correct for scrub but might
361 	 * be wrong for the dev_replace code where we might read from
362 	 * different devices in the initial huge bios. However, that
363 	 * code is able to correctly handle the case when adding a page
364 	 * to a bio fails.
365 	 */
366 	if (dev->bdev)
367 		pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
368 					 bio_get_nr_vecs(dev->bdev));
369 	else
370 		pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
371 	sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
372 	if (!sctx)
373 		goto nomem;
374 	sctx->is_dev_replace = is_dev_replace;
375 	sctx->pages_per_rd_bio = pages_per_rd_bio;
376 	sctx->curr = -1;
377 	sctx->dev_root = dev->dev_root;
378 	for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
379 		struct scrub_bio *sbio;
380 
381 		sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
382 		if (!sbio)
383 			goto nomem;
384 		sctx->bios[i] = sbio;
385 
386 		sbio->index = i;
387 		sbio->sctx = sctx;
388 		sbio->page_count = 0;
389 		sbio->work.func = scrub_bio_end_io_worker;
390 
391 		if (i != SCRUB_BIOS_PER_SCTX - 1)
392 			sctx->bios[i]->next_free = i + 1;
393 		else
394 			sctx->bios[i]->next_free = -1;
395 	}
396 	sctx->first_free = 0;
397 	sctx->nodesize = dev->dev_root->nodesize;
398 	sctx->leafsize = dev->dev_root->leafsize;
399 	sctx->sectorsize = dev->dev_root->sectorsize;
400 	atomic_set(&sctx->bios_in_flight, 0);
401 	atomic_set(&sctx->workers_pending, 0);
402 	atomic_set(&sctx->cancel_req, 0);
403 	sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
404 	INIT_LIST_HEAD(&sctx->csum_list);
405 
406 	spin_lock_init(&sctx->list_lock);
407 	spin_lock_init(&sctx->stat_lock);
408 	init_waitqueue_head(&sctx->list_wait);
409 
410 	ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
411 				 fs_info->dev_replace.tgtdev, is_dev_replace);
412 	if (ret) {
413 		scrub_free_ctx(sctx);
414 		return ERR_PTR(ret);
415 	}
416 	return sctx;
417 
418 nomem:
419 	scrub_free_ctx(sctx);
420 	return ERR_PTR(-ENOMEM);
421 }
422 
423 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
424 				     void *warn_ctx)
425 {
426 	u64 isize;
427 	u32 nlink;
428 	int ret;
429 	int i;
430 	struct extent_buffer *eb;
431 	struct btrfs_inode_item *inode_item;
432 	struct scrub_warning *swarn = warn_ctx;
433 	struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
434 	struct inode_fs_paths *ipath = NULL;
435 	struct btrfs_root *local_root;
436 	struct btrfs_key root_key;
437 
438 	root_key.objectid = root;
439 	root_key.type = BTRFS_ROOT_ITEM_KEY;
440 	root_key.offset = (u64)-1;
441 	local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
442 	if (IS_ERR(local_root)) {
443 		ret = PTR_ERR(local_root);
444 		goto err;
445 	}
446 
447 	ret = inode_item_info(inum, 0, local_root, swarn->path);
448 	if (ret) {
449 		btrfs_release_path(swarn->path);
450 		goto err;
451 	}
452 
453 	eb = swarn->path->nodes[0];
454 	inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
455 					struct btrfs_inode_item);
456 	isize = btrfs_inode_size(eb, inode_item);
457 	nlink = btrfs_inode_nlink(eb, inode_item);
458 	btrfs_release_path(swarn->path);
459 
460 	ipath = init_ipath(4096, local_root, swarn->path);
461 	if (IS_ERR(ipath)) {
462 		ret = PTR_ERR(ipath);
463 		ipath = NULL;
464 		goto err;
465 	}
466 	ret = paths_from_inode(inum, ipath);
467 
468 	if (ret < 0)
469 		goto err;
470 
471 	/*
472 	 * we deliberately ignore the bit ipath might have been too small to
473 	 * hold all of the paths here
474 	 */
475 	for (i = 0; i < ipath->fspath->elem_cnt; ++i)
476 		printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
477 			"%s, sector %llu, root %llu, inode %llu, offset %llu, "
478 			"length %llu, links %u (path: %s)\n", swarn->errstr,
479 			swarn->logical, rcu_str_deref(swarn->dev->name),
480 			(unsigned long long)swarn->sector, root, inum, offset,
481 			min(isize - offset, (u64)PAGE_SIZE), nlink,
482 			(char *)(unsigned long)ipath->fspath->val[i]);
483 
484 	free_ipath(ipath);
485 	return 0;
486 
487 err:
488 	printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
489 		"%s, sector %llu, root %llu, inode %llu, offset %llu: path "
490 		"resolving failed with ret=%d\n", swarn->errstr,
491 		swarn->logical, rcu_str_deref(swarn->dev->name),
492 		(unsigned long long)swarn->sector, root, inum, offset, ret);
493 
494 	free_ipath(ipath);
495 	return 0;
496 }
497 
498 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
499 {
500 	struct btrfs_device *dev;
501 	struct btrfs_fs_info *fs_info;
502 	struct btrfs_path *path;
503 	struct btrfs_key found_key;
504 	struct extent_buffer *eb;
505 	struct btrfs_extent_item *ei;
506 	struct scrub_warning swarn;
507 	unsigned long ptr = 0;
508 	u64 extent_item_pos;
509 	u64 flags = 0;
510 	u64 ref_root;
511 	u32 item_size;
512 	u8 ref_level;
513 	const int bufsize = 4096;
514 	int ret;
515 
516 	WARN_ON(sblock->page_count < 1);
517 	dev = sblock->pagev[0]->dev;
518 	fs_info = sblock->sctx->dev_root->fs_info;
519 
520 	path = btrfs_alloc_path();
521 
522 	swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
523 	swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
524 	swarn.sector = (sblock->pagev[0]->physical) >> 9;
525 	swarn.logical = sblock->pagev[0]->logical;
526 	swarn.errstr = errstr;
527 	swarn.dev = NULL;
528 	swarn.msg_bufsize = bufsize;
529 	swarn.scratch_bufsize = bufsize;
530 
531 	if (!path || !swarn.scratch_buf || !swarn.msg_buf)
532 		goto out;
533 
534 	ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
535 				  &flags);
536 	if (ret < 0)
537 		goto out;
538 
539 	extent_item_pos = swarn.logical - found_key.objectid;
540 	swarn.extent_item_size = found_key.offset;
541 
542 	eb = path->nodes[0];
543 	ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
544 	item_size = btrfs_item_size_nr(eb, path->slots[0]);
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 		btrfs_release_path(path);
561 	} else {
562 		btrfs_release_path(path);
563 		swarn.path = path;
564 		swarn.dev = dev;
565 		iterate_extent_inodes(fs_info, found_key.objectid,
566 					extent_item_pos, 1,
567 					scrub_print_warning_inode, &swarn);
568 	}
569 
570 out:
571 	btrfs_free_path(path);
572 	kfree(swarn.scratch_buf);
573 	kfree(swarn.msg_buf);
574 }
575 
576 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
577 {
578 	struct page *page = NULL;
579 	unsigned long index;
580 	struct scrub_fixup_nodatasum *fixup = fixup_ctx;
581 	int ret;
582 	int corrected = 0;
583 	struct btrfs_key key;
584 	struct inode *inode = NULL;
585 	struct btrfs_fs_info *fs_info;
586 	u64 end = offset + PAGE_SIZE - 1;
587 	struct btrfs_root *local_root;
588 	int srcu_index;
589 
590 	key.objectid = root;
591 	key.type = BTRFS_ROOT_ITEM_KEY;
592 	key.offset = (u64)-1;
593 
594 	fs_info = fixup->root->fs_info;
595 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
596 
597 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
598 	if (IS_ERR(local_root)) {
599 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
600 		return PTR_ERR(local_root);
601 	}
602 
603 	key.type = BTRFS_INODE_ITEM_KEY;
604 	key.objectid = inum;
605 	key.offset = 0;
606 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
607 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
608 	if (IS_ERR(inode))
609 		return PTR_ERR(inode);
610 
611 	index = offset >> PAGE_CACHE_SHIFT;
612 
613 	page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
614 	if (!page) {
615 		ret = -ENOMEM;
616 		goto out;
617 	}
618 
619 	if (PageUptodate(page)) {
620 		if (PageDirty(page)) {
621 			/*
622 			 * we need to write the data to the defect sector. the
623 			 * data that was in that sector is not in memory,
624 			 * because the page was modified. we must not write the
625 			 * modified page to that sector.
626 			 *
627 			 * TODO: what could be done here: wait for the delalloc
628 			 *       runner to write out that page (might involve
629 			 *       COW) and see whether the sector is still
630 			 *       referenced afterwards.
631 			 *
632 			 * For the meantime, we'll treat this error
633 			 * incorrectable, although there is a chance that a
634 			 * later scrub will find the bad sector again and that
635 			 * there's no dirty page in memory, then.
636 			 */
637 			ret = -EIO;
638 			goto out;
639 		}
640 		fs_info = BTRFS_I(inode)->root->fs_info;
641 		ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
642 					fixup->logical, page,
643 					fixup->mirror_num);
644 		unlock_page(page);
645 		corrected = !ret;
646 	} else {
647 		/*
648 		 * we need to get good data first. the general readpage path
649 		 * will call repair_io_failure for us, we just have to make
650 		 * sure we read the bad mirror.
651 		 */
652 		ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
653 					EXTENT_DAMAGED, GFP_NOFS);
654 		if (ret) {
655 			/* set_extent_bits should give proper error */
656 			WARN_ON(ret > 0);
657 			if (ret > 0)
658 				ret = -EFAULT;
659 			goto out;
660 		}
661 
662 		ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
663 						btrfs_get_extent,
664 						fixup->mirror_num);
665 		wait_on_page_locked(page);
666 
667 		corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
668 						end, EXTENT_DAMAGED, 0, NULL);
669 		if (!corrected)
670 			clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
671 						EXTENT_DAMAGED, GFP_NOFS);
672 	}
673 
674 out:
675 	if (page)
676 		put_page(page);
677 	if (inode)
678 		iput(inode);
679 
680 	if (ret < 0)
681 		return ret;
682 
683 	if (ret == 0 && corrected) {
684 		/*
685 		 * we only need to call readpage for one of the inodes belonging
686 		 * to this extent. so make iterate_extent_inodes stop
687 		 */
688 		return 1;
689 	}
690 
691 	return -EIO;
692 }
693 
694 static void scrub_fixup_nodatasum(struct btrfs_work *work)
695 {
696 	int ret;
697 	struct scrub_fixup_nodatasum *fixup;
698 	struct scrub_ctx *sctx;
699 	struct btrfs_trans_handle *trans = NULL;
700 	struct btrfs_fs_info *fs_info;
701 	struct btrfs_path *path;
702 	int uncorrectable = 0;
703 
704 	fixup = container_of(work, struct scrub_fixup_nodatasum, work);
705 	sctx = fixup->sctx;
706 	fs_info = fixup->root->fs_info;
707 
708 	path = btrfs_alloc_path();
709 	if (!path) {
710 		spin_lock(&sctx->stat_lock);
711 		++sctx->stat.malloc_errors;
712 		spin_unlock(&sctx->stat_lock);
713 		uncorrectable = 1;
714 		goto out;
715 	}
716 
717 	trans = btrfs_join_transaction(fixup->root);
718 	if (IS_ERR(trans)) {
719 		uncorrectable = 1;
720 		goto out;
721 	}
722 
723 	/*
724 	 * the idea is to trigger a regular read through the standard path. we
725 	 * read a page from the (failed) logical address by specifying the
726 	 * corresponding copynum of the failed sector. thus, that readpage is
727 	 * expected to fail.
728 	 * that is the point where on-the-fly error correction will kick in
729 	 * (once it's finished) and rewrite the failed sector if a good copy
730 	 * can be found.
731 	 */
732 	ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
733 						path, scrub_fixup_readpage,
734 						fixup);
735 	if (ret < 0) {
736 		uncorrectable = 1;
737 		goto out;
738 	}
739 	WARN_ON(ret != 1);
740 
741 	spin_lock(&sctx->stat_lock);
742 	++sctx->stat.corrected_errors;
743 	spin_unlock(&sctx->stat_lock);
744 
745 out:
746 	if (trans && !IS_ERR(trans))
747 		btrfs_end_transaction(trans, fixup->root);
748 	if (uncorrectable) {
749 		spin_lock(&sctx->stat_lock);
750 		++sctx->stat.uncorrectable_errors;
751 		spin_unlock(&sctx->stat_lock);
752 		btrfs_dev_replace_stats_inc(
753 			&sctx->dev_root->fs_info->dev_replace.
754 			num_uncorrectable_read_errors);
755 		printk_ratelimited_in_rcu(KERN_ERR
756 			"btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
757 			(unsigned long long)fixup->logical,
758 			rcu_str_deref(fixup->dev->name));
759 	}
760 
761 	btrfs_free_path(path);
762 	kfree(fixup);
763 
764 	scrub_pending_trans_workers_dec(sctx);
765 }
766 
767 /*
768  * scrub_handle_errored_block gets called when either verification of the
769  * pages failed or the bio failed to read, e.g. with EIO. In the latter
770  * case, this function handles all pages in the bio, even though only one
771  * may be bad.
772  * The goal of this function is to repair the errored block by using the
773  * contents of one of the mirrors.
774  */
775 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
776 {
777 	struct scrub_ctx *sctx = sblock_to_check->sctx;
778 	struct btrfs_device *dev;
779 	struct btrfs_fs_info *fs_info;
780 	u64 length;
781 	u64 logical;
782 	u64 generation;
783 	unsigned int failed_mirror_index;
784 	unsigned int is_metadata;
785 	unsigned int have_csum;
786 	u8 *csum;
787 	struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
788 	struct scrub_block *sblock_bad;
789 	int ret;
790 	int mirror_index;
791 	int page_num;
792 	int success;
793 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
794 				      DEFAULT_RATELIMIT_BURST);
795 
796 	BUG_ON(sblock_to_check->page_count < 1);
797 	fs_info = sctx->dev_root->fs_info;
798 	if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
799 		/*
800 		 * if we find an error in a super block, we just report it.
801 		 * They will get written with the next transaction commit
802 		 * anyway
803 		 */
804 		spin_lock(&sctx->stat_lock);
805 		++sctx->stat.super_errors;
806 		spin_unlock(&sctx->stat_lock);
807 		return 0;
808 	}
809 	length = sblock_to_check->page_count * PAGE_SIZE;
810 	logical = sblock_to_check->pagev[0]->logical;
811 	generation = sblock_to_check->pagev[0]->generation;
812 	BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
813 	failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
814 	is_metadata = !(sblock_to_check->pagev[0]->flags &
815 			BTRFS_EXTENT_FLAG_DATA);
816 	have_csum = sblock_to_check->pagev[0]->have_csum;
817 	csum = sblock_to_check->pagev[0]->csum;
818 	dev = sblock_to_check->pagev[0]->dev;
819 
820 	if (sctx->is_dev_replace && !is_metadata && !have_csum) {
821 		sblocks_for_recheck = NULL;
822 		goto nodatasum_case;
823 	}
824 
825 	/*
826 	 * read all mirrors one after the other. This includes to
827 	 * re-read the extent or metadata block that failed (that was
828 	 * the cause that this fixup code is called) another time,
829 	 * page by page this time in order to know which pages
830 	 * caused I/O errors and which ones are good (for all mirrors).
831 	 * It is the goal to handle the situation when more than one
832 	 * mirror contains I/O errors, but the errors do not
833 	 * overlap, i.e. the data can be repaired by selecting the
834 	 * pages from those mirrors without I/O error on the
835 	 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
836 	 * would be that mirror #1 has an I/O error on the first page,
837 	 * the second page is good, and mirror #2 has an I/O error on
838 	 * the second page, but the first page is good.
839 	 * Then the first page of the first mirror can be repaired by
840 	 * taking the first page of the second mirror, and the
841 	 * second page of the second mirror can be repaired by
842 	 * copying the contents of the 2nd page of the 1st mirror.
843 	 * One more note: if the pages of one mirror contain I/O
844 	 * errors, the checksum cannot be verified. In order to get
845 	 * the best data for repairing, the first attempt is to find
846 	 * a mirror without I/O errors and with a validated checksum.
847 	 * Only if this is not possible, the pages are picked from
848 	 * mirrors with I/O errors without considering the checksum.
849 	 * If the latter is the case, at the end, the checksum of the
850 	 * repaired area is verified in order to correctly maintain
851 	 * the statistics.
852 	 */
853 
854 	sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
855 				     sizeof(*sblocks_for_recheck),
856 				     GFP_NOFS);
857 	if (!sblocks_for_recheck) {
858 		spin_lock(&sctx->stat_lock);
859 		sctx->stat.malloc_errors++;
860 		sctx->stat.read_errors++;
861 		sctx->stat.uncorrectable_errors++;
862 		spin_unlock(&sctx->stat_lock);
863 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
864 		goto out;
865 	}
866 
867 	/* setup the context, map the logical blocks and alloc the pages */
868 	ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
869 					logical, sblocks_for_recheck);
870 	if (ret) {
871 		spin_lock(&sctx->stat_lock);
872 		sctx->stat.read_errors++;
873 		sctx->stat.uncorrectable_errors++;
874 		spin_unlock(&sctx->stat_lock);
875 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
876 		goto out;
877 	}
878 	BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
879 	sblock_bad = sblocks_for_recheck + failed_mirror_index;
880 
881 	/* build and submit the bios for the failed mirror, check checksums */
882 	scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
883 			    csum, generation, sctx->csum_size);
884 
885 	if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
886 	    sblock_bad->no_io_error_seen) {
887 		/*
888 		 * the error disappeared after reading page by page, or
889 		 * the area was part of a huge bio and other parts of the
890 		 * bio caused I/O errors, or the block layer merged several
891 		 * read requests into one and the error is caused by a
892 		 * different bio (usually one of the two latter cases is
893 		 * the cause)
894 		 */
895 		spin_lock(&sctx->stat_lock);
896 		sctx->stat.unverified_errors++;
897 		spin_unlock(&sctx->stat_lock);
898 
899 		if (sctx->is_dev_replace)
900 			scrub_write_block_to_dev_replace(sblock_bad);
901 		goto out;
902 	}
903 
904 	if (!sblock_bad->no_io_error_seen) {
905 		spin_lock(&sctx->stat_lock);
906 		sctx->stat.read_errors++;
907 		spin_unlock(&sctx->stat_lock);
908 		if (__ratelimit(&_rs))
909 			scrub_print_warning("i/o error", sblock_to_check);
910 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
911 	} else if (sblock_bad->checksum_error) {
912 		spin_lock(&sctx->stat_lock);
913 		sctx->stat.csum_errors++;
914 		spin_unlock(&sctx->stat_lock);
915 		if (__ratelimit(&_rs))
916 			scrub_print_warning("checksum error", sblock_to_check);
917 		btrfs_dev_stat_inc_and_print(dev,
918 					     BTRFS_DEV_STAT_CORRUPTION_ERRS);
919 	} else if (sblock_bad->header_error) {
920 		spin_lock(&sctx->stat_lock);
921 		sctx->stat.verify_errors++;
922 		spin_unlock(&sctx->stat_lock);
923 		if (__ratelimit(&_rs))
924 			scrub_print_warning("checksum/header error",
925 					    sblock_to_check);
926 		if (sblock_bad->generation_error)
927 			btrfs_dev_stat_inc_and_print(dev,
928 				BTRFS_DEV_STAT_GENERATION_ERRS);
929 		else
930 			btrfs_dev_stat_inc_and_print(dev,
931 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
932 	}
933 
934 	if (sctx->readonly && !sctx->is_dev_replace)
935 		goto did_not_correct_error;
936 
937 	if (!is_metadata && !have_csum) {
938 		struct scrub_fixup_nodatasum *fixup_nodatasum;
939 
940 nodatasum_case:
941 		WARN_ON(sctx->is_dev_replace);
942 
943 		/*
944 		 * !is_metadata and !have_csum, this means that the data
945 		 * might not be COW'ed, that it might be modified
946 		 * concurrently. The general strategy to work on the
947 		 * commit root does not help in the case when COW is not
948 		 * used.
949 		 */
950 		fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
951 		if (!fixup_nodatasum)
952 			goto did_not_correct_error;
953 		fixup_nodatasum->sctx = sctx;
954 		fixup_nodatasum->dev = dev;
955 		fixup_nodatasum->logical = logical;
956 		fixup_nodatasum->root = fs_info->extent_root;
957 		fixup_nodatasum->mirror_num = failed_mirror_index + 1;
958 		scrub_pending_trans_workers_inc(sctx);
959 		fixup_nodatasum->work.func = scrub_fixup_nodatasum;
960 		btrfs_queue_worker(&fs_info->scrub_workers,
961 				   &fixup_nodatasum->work);
962 		goto out;
963 	}
964 
965 	/*
966 	 * now build and submit the bios for the other mirrors, check
967 	 * checksums.
968 	 * First try to pick the mirror which is completely without I/O
969 	 * errors and also does not have a checksum error.
970 	 * If one is found, and if a checksum is present, the full block
971 	 * that is known to contain an error is rewritten. Afterwards
972 	 * the block is known to be corrected.
973 	 * If a mirror is found which is completely correct, and no
974 	 * checksum is present, only those pages are rewritten that had
975 	 * an I/O error in the block to be repaired, since it cannot be
976 	 * determined, which copy of the other pages is better (and it
977 	 * could happen otherwise that a correct page would be
978 	 * overwritten by a bad one).
979 	 */
980 	for (mirror_index = 0;
981 	     mirror_index < BTRFS_MAX_MIRRORS &&
982 	     sblocks_for_recheck[mirror_index].page_count > 0;
983 	     mirror_index++) {
984 		struct scrub_block *sblock_other;
985 
986 		if (mirror_index == failed_mirror_index)
987 			continue;
988 		sblock_other = sblocks_for_recheck + mirror_index;
989 
990 		/* build and submit the bios, check checksums */
991 		scrub_recheck_block(fs_info, sblock_other, is_metadata,
992 				    have_csum, csum, generation,
993 				    sctx->csum_size);
994 
995 		if (!sblock_other->header_error &&
996 		    !sblock_other->checksum_error &&
997 		    sblock_other->no_io_error_seen) {
998 			if (sctx->is_dev_replace) {
999 				scrub_write_block_to_dev_replace(sblock_other);
1000 			} else {
1001 				int force_write = is_metadata || have_csum;
1002 
1003 				ret = scrub_repair_block_from_good_copy(
1004 						sblock_bad, sblock_other,
1005 						force_write);
1006 			}
1007 			if (0 == ret)
1008 				goto corrected_error;
1009 		}
1010 	}
1011 
1012 	/*
1013 	 * for dev_replace, pick good pages and write to the target device.
1014 	 */
1015 	if (sctx->is_dev_replace) {
1016 		success = 1;
1017 		for (page_num = 0; page_num < sblock_bad->page_count;
1018 		     page_num++) {
1019 			int sub_success;
1020 
1021 			sub_success = 0;
1022 			for (mirror_index = 0;
1023 			     mirror_index < BTRFS_MAX_MIRRORS &&
1024 			     sblocks_for_recheck[mirror_index].page_count > 0;
1025 			     mirror_index++) {
1026 				struct scrub_block *sblock_other =
1027 					sblocks_for_recheck + mirror_index;
1028 				struct scrub_page *page_other =
1029 					sblock_other->pagev[page_num];
1030 
1031 				if (!page_other->io_error) {
1032 					ret = scrub_write_page_to_dev_replace(
1033 							sblock_other, page_num);
1034 					if (ret == 0) {
1035 						/* succeeded for this page */
1036 						sub_success = 1;
1037 						break;
1038 					} else {
1039 						btrfs_dev_replace_stats_inc(
1040 							&sctx->dev_root->
1041 							fs_info->dev_replace.
1042 							num_write_errors);
1043 					}
1044 				}
1045 			}
1046 
1047 			if (!sub_success) {
1048 				/*
1049 				 * did not find a mirror to fetch the page
1050 				 * from. scrub_write_page_to_dev_replace()
1051 				 * handles this case (page->io_error), by
1052 				 * filling the block with zeros before
1053 				 * submitting the write request
1054 				 */
1055 				success = 0;
1056 				ret = scrub_write_page_to_dev_replace(
1057 						sblock_bad, page_num);
1058 				if (ret)
1059 					btrfs_dev_replace_stats_inc(
1060 						&sctx->dev_root->fs_info->
1061 						dev_replace.num_write_errors);
1062 			}
1063 		}
1064 
1065 		goto out;
1066 	}
1067 
1068 	/*
1069 	 * for regular scrub, repair those pages that are errored.
1070 	 * In case of I/O errors in the area that is supposed to be
1071 	 * repaired, continue by picking good copies of those pages.
1072 	 * Select the good pages from mirrors to rewrite bad pages from
1073 	 * the area to fix. Afterwards verify the checksum of the block
1074 	 * that is supposed to be repaired. This verification step is
1075 	 * only done for the purpose of statistic counting and for the
1076 	 * final scrub report, whether errors remain.
1077 	 * A perfect algorithm could make use of the checksum and try
1078 	 * all possible combinations of pages from the different mirrors
1079 	 * until the checksum verification succeeds. For example, when
1080 	 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1081 	 * of mirror #2 is readable but the final checksum test fails,
1082 	 * then the 2nd page of mirror #3 could be tried, whether now
1083 	 * the final checksum succeedes. But this would be a rare
1084 	 * exception and is therefore not implemented. At least it is
1085 	 * avoided that the good copy is overwritten.
1086 	 * A more useful improvement would be to pick the sectors
1087 	 * without I/O error based on sector sizes (512 bytes on legacy
1088 	 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1089 	 * mirror could be repaired by taking 512 byte of a different
1090 	 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1091 	 * area are unreadable.
1092 	 */
1093 
1094 	/* can only fix I/O errors from here on */
1095 	if (sblock_bad->no_io_error_seen)
1096 		goto did_not_correct_error;
1097 
1098 	success = 1;
1099 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1100 		struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1101 
1102 		if (!page_bad->io_error)
1103 			continue;
1104 
1105 		for (mirror_index = 0;
1106 		     mirror_index < BTRFS_MAX_MIRRORS &&
1107 		     sblocks_for_recheck[mirror_index].page_count > 0;
1108 		     mirror_index++) {
1109 			struct scrub_block *sblock_other = sblocks_for_recheck +
1110 							   mirror_index;
1111 			struct scrub_page *page_other = sblock_other->pagev[
1112 							page_num];
1113 
1114 			if (!page_other->io_error) {
1115 				ret = scrub_repair_page_from_good_copy(
1116 					sblock_bad, sblock_other, page_num, 0);
1117 				if (0 == ret) {
1118 					page_bad->io_error = 0;
1119 					break; /* succeeded for this page */
1120 				}
1121 			}
1122 		}
1123 
1124 		if (page_bad->io_error) {
1125 			/* did not find a mirror to copy the page from */
1126 			success = 0;
1127 		}
1128 	}
1129 
1130 	if (success) {
1131 		if (is_metadata || have_csum) {
1132 			/*
1133 			 * need to verify the checksum now that all
1134 			 * sectors on disk are repaired (the write
1135 			 * request for data to be repaired is on its way).
1136 			 * Just be lazy and use scrub_recheck_block()
1137 			 * which re-reads the data before the checksum
1138 			 * is verified, but most likely the data comes out
1139 			 * of the page cache.
1140 			 */
1141 			scrub_recheck_block(fs_info, sblock_bad,
1142 					    is_metadata, have_csum, csum,
1143 					    generation, sctx->csum_size);
1144 			if (!sblock_bad->header_error &&
1145 			    !sblock_bad->checksum_error &&
1146 			    sblock_bad->no_io_error_seen)
1147 				goto corrected_error;
1148 			else
1149 				goto did_not_correct_error;
1150 		} else {
1151 corrected_error:
1152 			spin_lock(&sctx->stat_lock);
1153 			sctx->stat.corrected_errors++;
1154 			spin_unlock(&sctx->stat_lock);
1155 			printk_ratelimited_in_rcu(KERN_ERR
1156 				"btrfs: fixed up error at logical %llu on dev %s\n",
1157 				(unsigned long long)logical,
1158 				rcu_str_deref(dev->name));
1159 		}
1160 	} else {
1161 did_not_correct_error:
1162 		spin_lock(&sctx->stat_lock);
1163 		sctx->stat.uncorrectable_errors++;
1164 		spin_unlock(&sctx->stat_lock);
1165 		printk_ratelimited_in_rcu(KERN_ERR
1166 			"btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1167 			(unsigned long long)logical,
1168 			rcu_str_deref(dev->name));
1169 	}
1170 
1171 out:
1172 	if (sblocks_for_recheck) {
1173 		for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1174 		     mirror_index++) {
1175 			struct scrub_block *sblock = sblocks_for_recheck +
1176 						     mirror_index;
1177 			int page_index;
1178 
1179 			for (page_index = 0; page_index < sblock->page_count;
1180 			     page_index++) {
1181 				sblock->pagev[page_index]->sblock = NULL;
1182 				scrub_page_put(sblock->pagev[page_index]);
1183 			}
1184 		}
1185 		kfree(sblocks_for_recheck);
1186 	}
1187 
1188 	return 0;
1189 }
1190 
1191 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1192 				     struct btrfs_fs_info *fs_info,
1193 				     struct scrub_block *original_sblock,
1194 				     u64 length, u64 logical,
1195 				     struct scrub_block *sblocks_for_recheck)
1196 {
1197 	int page_index;
1198 	int mirror_index;
1199 	int ret;
1200 
1201 	/*
1202 	 * note: the two members ref_count and outstanding_pages
1203 	 * are not used (and not set) in the blocks that are used for
1204 	 * the recheck procedure
1205 	 */
1206 
1207 	page_index = 0;
1208 	while (length > 0) {
1209 		u64 sublen = min_t(u64, length, PAGE_SIZE);
1210 		u64 mapped_length = sublen;
1211 		struct btrfs_bio *bbio = NULL;
1212 
1213 		/*
1214 		 * with a length of PAGE_SIZE, each returned stripe
1215 		 * represents one mirror
1216 		 */
1217 		ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1218 				      &mapped_length, &bbio, 0);
1219 		if (ret || !bbio || mapped_length < sublen) {
1220 			kfree(bbio);
1221 			return -EIO;
1222 		}
1223 
1224 		BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1225 		for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1226 		     mirror_index++) {
1227 			struct scrub_block *sblock;
1228 			struct scrub_page *page;
1229 
1230 			if (mirror_index >= BTRFS_MAX_MIRRORS)
1231 				continue;
1232 
1233 			sblock = sblocks_for_recheck + mirror_index;
1234 			sblock->sctx = sctx;
1235 			page = kzalloc(sizeof(*page), GFP_NOFS);
1236 			if (!page) {
1237 leave_nomem:
1238 				spin_lock(&sctx->stat_lock);
1239 				sctx->stat.malloc_errors++;
1240 				spin_unlock(&sctx->stat_lock);
1241 				kfree(bbio);
1242 				return -ENOMEM;
1243 			}
1244 			scrub_page_get(page);
1245 			sblock->pagev[page_index] = page;
1246 			page->logical = logical;
1247 			page->physical = bbio->stripes[mirror_index].physical;
1248 			BUG_ON(page_index >= original_sblock->page_count);
1249 			page->physical_for_dev_replace =
1250 				original_sblock->pagev[page_index]->
1251 				physical_for_dev_replace;
1252 			/* for missing devices, dev->bdev is NULL */
1253 			page->dev = bbio->stripes[mirror_index].dev;
1254 			page->mirror_num = mirror_index + 1;
1255 			sblock->page_count++;
1256 			page->page = alloc_page(GFP_NOFS);
1257 			if (!page->page)
1258 				goto leave_nomem;
1259 		}
1260 		kfree(bbio);
1261 		length -= sublen;
1262 		logical += sublen;
1263 		page_index++;
1264 	}
1265 
1266 	return 0;
1267 }
1268 
1269 /*
1270  * this function will check the on disk data for checksum errors, header
1271  * errors and read I/O errors. If any I/O errors happen, the exact pages
1272  * which are errored are marked as being bad. The goal is to enable scrub
1273  * to take those pages that are not errored from all the mirrors so that
1274  * the pages that are errored in the just handled mirror can be repaired.
1275  */
1276 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1277 				struct scrub_block *sblock, int is_metadata,
1278 				int have_csum, u8 *csum, u64 generation,
1279 				u16 csum_size)
1280 {
1281 	int page_num;
1282 
1283 	sblock->no_io_error_seen = 1;
1284 	sblock->header_error = 0;
1285 	sblock->checksum_error = 0;
1286 
1287 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1288 		struct bio *bio;
1289 		struct scrub_page *page = sblock->pagev[page_num];
1290 		DECLARE_COMPLETION_ONSTACK(complete);
1291 
1292 		if (page->dev->bdev == NULL) {
1293 			page->io_error = 1;
1294 			sblock->no_io_error_seen = 0;
1295 			continue;
1296 		}
1297 
1298 		WARN_ON(!page->page);
1299 		bio = bio_alloc(GFP_NOFS, 1);
1300 		if (!bio) {
1301 			page->io_error = 1;
1302 			sblock->no_io_error_seen = 0;
1303 			continue;
1304 		}
1305 		bio->bi_bdev = page->dev->bdev;
1306 		bio->bi_sector = page->physical >> 9;
1307 		bio->bi_end_io = scrub_complete_bio_end_io;
1308 		bio->bi_private = &complete;
1309 
1310 		bio_add_page(bio, page->page, PAGE_SIZE, 0);
1311 		btrfsic_submit_bio(READ, bio);
1312 
1313 		/* this will also unplug the queue */
1314 		wait_for_completion(&complete);
1315 
1316 		page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1317 		if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1318 			sblock->no_io_error_seen = 0;
1319 		bio_put(bio);
1320 	}
1321 
1322 	if (sblock->no_io_error_seen)
1323 		scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1324 					     have_csum, csum, generation,
1325 					     csum_size);
1326 
1327 	return;
1328 }
1329 
1330 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1331 					 struct scrub_block *sblock,
1332 					 int is_metadata, int have_csum,
1333 					 const u8 *csum, u64 generation,
1334 					 u16 csum_size)
1335 {
1336 	int page_num;
1337 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1338 	u32 crc = ~(u32)0;
1339 	struct btrfs_root *root = fs_info->extent_root;
1340 	void *mapped_buffer;
1341 
1342 	WARN_ON(!sblock->pagev[0]->page);
1343 	if (is_metadata) {
1344 		struct btrfs_header *h;
1345 
1346 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1347 		h = (struct btrfs_header *)mapped_buffer;
1348 
1349 		if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1350 		    memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1351 		    memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1352 			   BTRFS_UUID_SIZE)) {
1353 			sblock->header_error = 1;
1354 		} else if (generation != le64_to_cpu(h->generation)) {
1355 			sblock->header_error = 1;
1356 			sblock->generation_error = 1;
1357 		}
1358 		csum = h->csum;
1359 	} else {
1360 		if (!have_csum)
1361 			return;
1362 
1363 		mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1364 	}
1365 
1366 	for (page_num = 0;;) {
1367 		if (page_num == 0 && is_metadata)
1368 			crc = btrfs_csum_data(root,
1369 				((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1370 				crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1371 		else
1372 			crc = btrfs_csum_data(root, mapped_buffer, crc,
1373 					      PAGE_SIZE);
1374 
1375 		kunmap_atomic(mapped_buffer);
1376 		page_num++;
1377 		if (page_num >= sblock->page_count)
1378 			break;
1379 		WARN_ON(!sblock->pagev[page_num]->page);
1380 
1381 		mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1382 	}
1383 
1384 	btrfs_csum_final(crc, calculated_csum);
1385 	if (memcmp(calculated_csum, csum, csum_size))
1386 		sblock->checksum_error = 1;
1387 }
1388 
1389 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1390 {
1391 	complete((struct completion *)bio->bi_private);
1392 }
1393 
1394 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1395 					     struct scrub_block *sblock_good,
1396 					     int force_write)
1397 {
1398 	int page_num;
1399 	int ret = 0;
1400 
1401 	for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1402 		int ret_sub;
1403 
1404 		ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1405 							   sblock_good,
1406 							   page_num,
1407 							   force_write);
1408 		if (ret_sub)
1409 			ret = ret_sub;
1410 	}
1411 
1412 	return ret;
1413 }
1414 
1415 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1416 					    struct scrub_block *sblock_good,
1417 					    int page_num, int force_write)
1418 {
1419 	struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1420 	struct scrub_page *page_good = sblock_good->pagev[page_num];
1421 
1422 	BUG_ON(page_bad->page == NULL);
1423 	BUG_ON(page_good->page == NULL);
1424 	if (force_write || sblock_bad->header_error ||
1425 	    sblock_bad->checksum_error || page_bad->io_error) {
1426 		struct bio *bio;
1427 		int ret;
1428 		DECLARE_COMPLETION_ONSTACK(complete);
1429 
1430 		if (!page_bad->dev->bdev) {
1431 			printk_ratelimited(KERN_WARNING
1432 				"btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1433 			return -EIO;
1434 		}
1435 
1436 		bio = bio_alloc(GFP_NOFS, 1);
1437 		if (!bio)
1438 			return -EIO;
1439 		bio->bi_bdev = page_bad->dev->bdev;
1440 		bio->bi_sector = page_bad->physical >> 9;
1441 		bio->bi_end_io = scrub_complete_bio_end_io;
1442 		bio->bi_private = &complete;
1443 
1444 		ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1445 		if (PAGE_SIZE != ret) {
1446 			bio_put(bio);
1447 			return -EIO;
1448 		}
1449 		btrfsic_submit_bio(WRITE, bio);
1450 
1451 		/* this will also unplug the queue */
1452 		wait_for_completion(&complete);
1453 		if (!bio_flagged(bio, BIO_UPTODATE)) {
1454 			btrfs_dev_stat_inc_and_print(page_bad->dev,
1455 				BTRFS_DEV_STAT_WRITE_ERRS);
1456 			btrfs_dev_replace_stats_inc(
1457 				&sblock_bad->sctx->dev_root->fs_info->
1458 				dev_replace.num_write_errors);
1459 			bio_put(bio);
1460 			return -EIO;
1461 		}
1462 		bio_put(bio);
1463 	}
1464 
1465 	return 0;
1466 }
1467 
1468 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1469 {
1470 	int page_num;
1471 
1472 	for (page_num = 0; page_num < sblock->page_count; page_num++) {
1473 		int ret;
1474 
1475 		ret = scrub_write_page_to_dev_replace(sblock, page_num);
1476 		if (ret)
1477 			btrfs_dev_replace_stats_inc(
1478 				&sblock->sctx->dev_root->fs_info->dev_replace.
1479 				num_write_errors);
1480 	}
1481 }
1482 
1483 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1484 					   int page_num)
1485 {
1486 	struct scrub_page *spage = sblock->pagev[page_num];
1487 
1488 	BUG_ON(spage->page == NULL);
1489 	if (spage->io_error) {
1490 		void *mapped_buffer = kmap_atomic(spage->page);
1491 
1492 		memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1493 		flush_dcache_page(spage->page);
1494 		kunmap_atomic(mapped_buffer);
1495 	}
1496 	return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1497 }
1498 
1499 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1500 				    struct scrub_page *spage)
1501 {
1502 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1503 	struct scrub_bio *sbio;
1504 	int ret;
1505 
1506 	mutex_lock(&wr_ctx->wr_lock);
1507 again:
1508 	if (!wr_ctx->wr_curr_bio) {
1509 		wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1510 					      GFP_NOFS);
1511 		if (!wr_ctx->wr_curr_bio) {
1512 			mutex_unlock(&wr_ctx->wr_lock);
1513 			return -ENOMEM;
1514 		}
1515 		wr_ctx->wr_curr_bio->sctx = sctx;
1516 		wr_ctx->wr_curr_bio->page_count = 0;
1517 	}
1518 	sbio = wr_ctx->wr_curr_bio;
1519 	if (sbio->page_count == 0) {
1520 		struct bio *bio;
1521 
1522 		sbio->physical = spage->physical_for_dev_replace;
1523 		sbio->logical = spage->logical;
1524 		sbio->dev = wr_ctx->tgtdev;
1525 		bio = sbio->bio;
1526 		if (!bio) {
1527 			bio = bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1528 			if (!bio) {
1529 				mutex_unlock(&wr_ctx->wr_lock);
1530 				return -ENOMEM;
1531 			}
1532 			sbio->bio = bio;
1533 		}
1534 
1535 		bio->bi_private = sbio;
1536 		bio->bi_end_io = scrub_wr_bio_end_io;
1537 		bio->bi_bdev = sbio->dev->bdev;
1538 		bio->bi_sector = sbio->physical >> 9;
1539 		sbio->err = 0;
1540 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1541 		   spage->physical_for_dev_replace ||
1542 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1543 		   spage->logical) {
1544 		scrub_wr_submit(sctx);
1545 		goto again;
1546 	}
1547 
1548 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1549 	if (ret != PAGE_SIZE) {
1550 		if (sbio->page_count < 1) {
1551 			bio_put(sbio->bio);
1552 			sbio->bio = NULL;
1553 			mutex_unlock(&wr_ctx->wr_lock);
1554 			return -EIO;
1555 		}
1556 		scrub_wr_submit(sctx);
1557 		goto again;
1558 	}
1559 
1560 	sbio->pagev[sbio->page_count] = spage;
1561 	scrub_page_get(spage);
1562 	sbio->page_count++;
1563 	if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1564 		scrub_wr_submit(sctx);
1565 	mutex_unlock(&wr_ctx->wr_lock);
1566 
1567 	return 0;
1568 }
1569 
1570 static void scrub_wr_submit(struct scrub_ctx *sctx)
1571 {
1572 	struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1573 	struct scrub_bio *sbio;
1574 
1575 	if (!wr_ctx->wr_curr_bio)
1576 		return;
1577 
1578 	sbio = wr_ctx->wr_curr_bio;
1579 	wr_ctx->wr_curr_bio = NULL;
1580 	WARN_ON(!sbio->bio->bi_bdev);
1581 	scrub_pending_bio_inc(sctx);
1582 	/* process all writes in a single worker thread. Then the block layer
1583 	 * orders the requests before sending them to the driver which
1584 	 * doubled the write performance on spinning disks when measured
1585 	 * with Linux 3.5 */
1586 	btrfsic_submit_bio(WRITE, sbio->bio);
1587 }
1588 
1589 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1590 {
1591 	struct scrub_bio *sbio = bio->bi_private;
1592 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1593 
1594 	sbio->err = err;
1595 	sbio->bio = bio;
1596 
1597 	sbio->work.func = scrub_wr_bio_end_io_worker;
1598 	btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1599 }
1600 
1601 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1602 {
1603 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1604 	struct scrub_ctx *sctx = sbio->sctx;
1605 	int i;
1606 
1607 	WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1608 	if (sbio->err) {
1609 		struct btrfs_dev_replace *dev_replace =
1610 			&sbio->sctx->dev_root->fs_info->dev_replace;
1611 
1612 		for (i = 0; i < sbio->page_count; i++) {
1613 			struct scrub_page *spage = sbio->pagev[i];
1614 
1615 			spage->io_error = 1;
1616 			btrfs_dev_replace_stats_inc(&dev_replace->
1617 						    num_write_errors);
1618 		}
1619 	}
1620 
1621 	for (i = 0; i < sbio->page_count; i++)
1622 		scrub_page_put(sbio->pagev[i]);
1623 
1624 	bio_put(sbio->bio);
1625 	kfree(sbio);
1626 	scrub_pending_bio_dec(sctx);
1627 }
1628 
1629 static int scrub_checksum(struct scrub_block *sblock)
1630 {
1631 	u64 flags;
1632 	int ret;
1633 
1634 	WARN_ON(sblock->page_count < 1);
1635 	flags = sblock->pagev[0]->flags;
1636 	ret = 0;
1637 	if (flags & BTRFS_EXTENT_FLAG_DATA)
1638 		ret = scrub_checksum_data(sblock);
1639 	else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1640 		ret = scrub_checksum_tree_block(sblock);
1641 	else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1642 		(void)scrub_checksum_super(sblock);
1643 	else
1644 		WARN_ON(1);
1645 	if (ret)
1646 		scrub_handle_errored_block(sblock);
1647 
1648 	return ret;
1649 }
1650 
1651 static int scrub_checksum_data(struct scrub_block *sblock)
1652 {
1653 	struct scrub_ctx *sctx = sblock->sctx;
1654 	u8 csum[BTRFS_CSUM_SIZE];
1655 	u8 *on_disk_csum;
1656 	struct page *page;
1657 	void *buffer;
1658 	u32 crc = ~(u32)0;
1659 	int fail = 0;
1660 	struct btrfs_root *root = sctx->dev_root;
1661 	u64 len;
1662 	int index;
1663 
1664 	BUG_ON(sblock->page_count < 1);
1665 	if (!sblock->pagev[0]->have_csum)
1666 		return 0;
1667 
1668 	on_disk_csum = sblock->pagev[0]->csum;
1669 	page = sblock->pagev[0]->page;
1670 	buffer = kmap_atomic(page);
1671 
1672 	len = sctx->sectorsize;
1673 	index = 0;
1674 	for (;;) {
1675 		u64 l = min_t(u64, len, PAGE_SIZE);
1676 
1677 		crc = btrfs_csum_data(root, buffer, crc, l);
1678 		kunmap_atomic(buffer);
1679 		len -= l;
1680 		if (len == 0)
1681 			break;
1682 		index++;
1683 		BUG_ON(index >= sblock->page_count);
1684 		BUG_ON(!sblock->pagev[index]->page);
1685 		page = sblock->pagev[index]->page;
1686 		buffer = kmap_atomic(page);
1687 	}
1688 
1689 	btrfs_csum_final(crc, csum);
1690 	if (memcmp(csum, on_disk_csum, sctx->csum_size))
1691 		fail = 1;
1692 
1693 	return fail;
1694 }
1695 
1696 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1697 {
1698 	struct scrub_ctx *sctx = sblock->sctx;
1699 	struct btrfs_header *h;
1700 	struct btrfs_root *root = sctx->dev_root;
1701 	struct btrfs_fs_info *fs_info = root->fs_info;
1702 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1703 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1704 	struct page *page;
1705 	void *mapped_buffer;
1706 	u64 mapped_size;
1707 	void *p;
1708 	u32 crc = ~(u32)0;
1709 	int fail = 0;
1710 	int crc_fail = 0;
1711 	u64 len;
1712 	int index;
1713 
1714 	BUG_ON(sblock->page_count < 1);
1715 	page = sblock->pagev[0]->page;
1716 	mapped_buffer = kmap_atomic(page);
1717 	h = (struct btrfs_header *)mapped_buffer;
1718 	memcpy(on_disk_csum, h->csum, sctx->csum_size);
1719 
1720 	/*
1721 	 * we don't use the getter functions here, as we
1722 	 * a) don't have an extent buffer and
1723 	 * b) the page is already kmapped
1724 	 */
1725 
1726 	if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1727 		++fail;
1728 
1729 	if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1730 		++fail;
1731 
1732 	if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1733 		++fail;
1734 
1735 	if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1736 		   BTRFS_UUID_SIZE))
1737 		++fail;
1738 
1739 	WARN_ON(sctx->nodesize != sctx->leafsize);
1740 	len = sctx->nodesize - BTRFS_CSUM_SIZE;
1741 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1742 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1743 	index = 0;
1744 	for (;;) {
1745 		u64 l = min_t(u64, len, mapped_size);
1746 
1747 		crc = btrfs_csum_data(root, p, crc, l);
1748 		kunmap_atomic(mapped_buffer);
1749 		len -= l;
1750 		if (len == 0)
1751 			break;
1752 		index++;
1753 		BUG_ON(index >= sblock->page_count);
1754 		BUG_ON(!sblock->pagev[index]->page);
1755 		page = sblock->pagev[index]->page;
1756 		mapped_buffer = kmap_atomic(page);
1757 		mapped_size = PAGE_SIZE;
1758 		p = mapped_buffer;
1759 	}
1760 
1761 	btrfs_csum_final(crc, calculated_csum);
1762 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1763 		++crc_fail;
1764 
1765 	return fail || crc_fail;
1766 }
1767 
1768 static int scrub_checksum_super(struct scrub_block *sblock)
1769 {
1770 	struct btrfs_super_block *s;
1771 	struct scrub_ctx *sctx = sblock->sctx;
1772 	struct btrfs_root *root = sctx->dev_root;
1773 	struct btrfs_fs_info *fs_info = root->fs_info;
1774 	u8 calculated_csum[BTRFS_CSUM_SIZE];
1775 	u8 on_disk_csum[BTRFS_CSUM_SIZE];
1776 	struct page *page;
1777 	void *mapped_buffer;
1778 	u64 mapped_size;
1779 	void *p;
1780 	u32 crc = ~(u32)0;
1781 	int fail_gen = 0;
1782 	int fail_cor = 0;
1783 	u64 len;
1784 	int index;
1785 
1786 	BUG_ON(sblock->page_count < 1);
1787 	page = sblock->pagev[0]->page;
1788 	mapped_buffer = kmap_atomic(page);
1789 	s = (struct btrfs_super_block *)mapped_buffer;
1790 	memcpy(on_disk_csum, s->csum, sctx->csum_size);
1791 
1792 	if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1793 		++fail_cor;
1794 
1795 	if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1796 		++fail_gen;
1797 
1798 	if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1799 		++fail_cor;
1800 
1801 	len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1802 	mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1803 	p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1804 	index = 0;
1805 	for (;;) {
1806 		u64 l = min_t(u64, len, mapped_size);
1807 
1808 		crc = btrfs_csum_data(root, p, crc, l);
1809 		kunmap_atomic(mapped_buffer);
1810 		len -= l;
1811 		if (len == 0)
1812 			break;
1813 		index++;
1814 		BUG_ON(index >= sblock->page_count);
1815 		BUG_ON(!sblock->pagev[index]->page);
1816 		page = sblock->pagev[index]->page;
1817 		mapped_buffer = kmap_atomic(page);
1818 		mapped_size = PAGE_SIZE;
1819 		p = mapped_buffer;
1820 	}
1821 
1822 	btrfs_csum_final(crc, calculated_csum);
1823 	if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1824 		++fail_cor;
1825 
1826 	if (fail_cor + fail_gen) {
1827 		/*
1828 		 * if we find an error in a super block, we just report it.
1829 		 * They will get written with the next transaction commit
1830 		 * anyway
1831 		 */
1832 		spin_lock(&sctx->stat_lock);
1833 		++sctx->stat.super_errors;
1834 		spin_unlock(&sctx->stat_lock);
1835 		if (fail_cor)
1836 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1837 				BTRFS_DEV_STAT_CORRUPTION_ERRS);
1838 		else
1839 			btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1840 				BTRFS_DEV_STAT_GENERATION_ERRS);
1841 	}
1842 
1843 	return fail_cor + fail_gen;
1844 }
1845 
1846 static void scrub_block_get(struct scrub_block *sblock)
1847 {
1848 	atomic_inc(&sblock->ref_count);
1849 }
1850 
1851 static void scrub_block_put(struct scrub_block *sblock)
1852 {
1853 	if (atomic_dec_and_test(&sblock->ref_count)) {
1854 		int i;
1855 
1856 		for (i = 0; i < sblock->page_count; i++)
1857 			scrub_page_put(sblock->pagev[i]);
1858 		kfree(sblock);
1859 	}
1860 }
1861 
1862 static void scrub_page_get(struct scrub_page *spage)
1863 {
1864 	atomic_inc(&spage->ref_count);
1865 }
1866 
1867 static void scrub_page_put(struct scrub_page *spage)
1868 {
1869 	if (atomic_dec_and_test(&spage->ref_count)) {
1870 		if (spage->page)
1871 			__free_page(spage->page);
1872 		kfree(spage);
1873 	}
1874 }
1875 
1876 static void scrub_submit(struct scrub_ctx *sctx)
1877 {
1878 	struct scrub_bio *sbio;
1879 
1880 	if (sctx->curr == -1)
1881 		return;
1882 
1883 	sbio = sctx->bios[sctx->curr];
1884 	sctx->curr = -1;
1885 	scrub_pending_bio_inc(sctx);
1886 
1887 	if (!sbio->bio->bi_bdev) {
1888 		/*
1889 		 * this case should not happen. If btrfs_map_block() is
1890 		 * wrong, it could happen for dev-replace operations on
1891 		 * missing devices when no mirrors are available, but in
1892 		 * this case it should already fail the mount.
1893 		 * This case is handled correctly (but _very_ slowly).
1894 		 */
1895 		printk_ratelimited(KERN_WARNING
1896 			"btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1897 		bio_endio(sbio->bio, -EIO);
1898 	} else {
1899 		btrfsic_submit_bio(READ, sbio->bio);
1900 	}
1901 }
1902 
1903 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1904 				    struct scrub_page *spage)
1905 {
1906 	struct scrub_block *sblock = spage->sblock;
1907 	struct scrub_bio *sbio;
1908 	int ret;
1909 
1910 again:
1911 	/*
1912 	 * grab a fresh bio or wait for one to become available
1913 	 */
1914 	while (sctx->curr == -1) {
1915 		spin_lock(&sctx->list_lock);
1916 		sctx->curr = sctx->first_free;
1917 		if (sctx->curr != -1) {
1918 			sctx->first_free = sctx->bios[sctx->curr]->next_free;
1919 			sctx->bios[sctx->curr]->next_free = -1;
1920 			sctx->bios[sctx->curr]->page_count = 0;
1921 			spin_unlock(&sctx->list_lock);
1922 		} else {
1923 			spin_unlock(&sctx->list_lock);
1924 			wait_event(sctx->list_wait, sctx->first_free != -1);
1925 		}
1926 	}
1927 	sbio = sctx->bios[sctx->curr];
1928 	if (sbio->page_count == 0) {
1929 		struct bio *bio;
1930 
1931 		sbio->physical = spage->physical;
1932 		sbio->logical = spage->logical;
1933 		sbio->dev = spage->dev;
1934 		bio = sbio->bio;
1935 		if (!bio) {
1936 			bio = bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1937 			if (!bio)
1938 				return -ENOMEM;
1939 			sbio->bio = bio;
1940 		}
1941 
1942 		bio->bi_private = sbio;
1943 		bio->bi_end_io = scrub_bio_end_io;
1944 		bio->bi_bdev = sbio->dev->bdev;
1945 		bio->bi_sector = sbio->physical >> 9;
1946 		sbio->err = 0;
1947 	} else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1948 		   spage->physical ||
1949 		   sbio->logical + sbio->page_count * PAGE_SIZE !=
1950 		   spage->logical ||
1951 		   sbio->dev != spage->dev) {
1952 		scrub_submit(sctx);
1953 		goto again;
1954 	}
1955 
1956 	sbio->pagev[sbio->page_count] = spage;
1957 	ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1958 	if (ret != PAGE_SIZE) {
1959 		if (sbio->page_count < 1) {
1960 			bio_put(sbio->bio);
1961 			sbio->bio = NULL;
1962 			return -EIO;
1963 		}
1964 		scrub_submit(sctx);
1965 		goto again;
1966 	}
1967 
1968 	scrub_block_get(sblock); /* one for the page added to the bio */
1969 	atomic_inc(&sblock->outstanding_pages);
1970 	sbio->page_count++;
1971 	if (sbio->page_count == sctx->pages_per_rd_bio)
1972 		scrub_submit(sctx);
1973 
1974 	return 0;
1975 }
1976 
1977 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1978 		       u64 physical, struct btrfs_device *dev, u64 flags,
1979 		       u64 gen, int mirror_num, u8 *csum, int force,
1980 		       u64 physical_for_dev_replace)
1981 {
1982 	struct scrub_block *sblock;
1983 	int index;
1984 
1985 	sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1986 	if (!sblock) {
1987 		spin_lock(&sctx->stat_lock);
1988 		sctx->stat.malloc_errors++;
1989 		spin_unlock(&sctx->stat_lock);
1990 		return -ENOMEM;
1991 	}
1992 
1993 	/* one ref inside this function, plus one for each page added to
1994 	 * a bio later on */
1995 	atomic_set(&sblock->ref_count, 1);
1996 	sblock->sctx = sctx;
1997 	sblock->no_io_error_seen = 1;
1998 
1999 	for (index = 0; len > 0; index++) {
2000 		struct scrub_page *spage;
2001 		u64 l = min_t(u64, len, PAGE_SIZE);
2002 
2003 		spage = kzalloc(sizeof(*spage), GFP_NOFS);
2004 		if (!spage) {
2005 leave_nomem:
2006 			spin_lock(&sctx->stat_lock);
2007 			sctx->stat.malloc_errors++;
2008 			spin_unlock(&sctx->stat_lock);
2009 			scrub_block_put(sblock);
2010 			return -ENOMEM;
2011 		}
2012 		BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2013 		scrub_page_get(spage);
2014 		sblock->pagev[index] = spage;
2015 		spage->sblock = sblock;
2016 		spage->dev = dev;
2017 		spage->flags = flags;
2018 		spage->generation = gen;
2019 		spage->logical = logical;
2020 		spage->physical = physical;
2021 		spage->physical_for_dev_replace = physical_for_dev_replace;
2022 		spage->mirror_num = mirror_num;
2023 		if (csum) {
2024 			spage->have_csum = 1;
2025 			memcpy(spage->csum, csum, sctx->csum_size);
2026 		} else {
2027 			spage->have_csum = 0;
2028 		}
2029 		sblock->page_count++;
2030 		spage->page = alloc_page(GFP_NOFS);
2031 		if (!spage->page)
2032 			goto leave_nomem;
2033 		len -= l;
2034 		logical += l;
2035 		physical += l;
2036 		physical_for_dev_replace += l;
2037 	}
2038 
2039 	WARN_ON(sblock->page_count == 0);
2040 	for (index = 0; index < sblock->page_count; index++) {
2041 		struct scrub_page *spage = sblock->pagev[index];
2042 		int ret;
2043 
2044 		ret = scrub_add_page_to_rd_bio(sctx, spage);
2045 		if (ret) {
2046 			scrub_block_put(sblock);
2047 			return ret;
2048 		}
2049 	}
2050 
2051 	if (force)
2052 		scrub_submit(sctx);
2053 
2054 	/* last one frees, either here or in bio completion for last page */
2055 	scrub_block_put(sblock);
2056 	return 0;
2057 }
2058 
2059 static void scrub_bio_end_io(struct bio *bio, int err)
2060 {
2061 	struct scrub_bio *sbio = bio->bi_private;
2062 	struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2063 
2064 	sbio->err = err;
2065 	sbio->bio = bio;
2066 
2067 	btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2068 }
2069 
2070 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2071 {
2072 	struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2073 	struct scrub_ctx *sctx = sbio->sctx;
2074 	int i;
2075 
2076 	BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2077 	if (sbio->err) {
2078 		for (i = 0; i < sbio->page_count; i++) {
2079 			struct scrub_page *spage = sbio->pagev[i];
2080 
2081 			spage->io_error = 1;
2082 			spage->sblock->no_io_error_seen = 0;
2083 		}
2084 	}
2085 
2086 	/* now complete the scrub_block items that have all pages completed */
2087 	for (i = 0; i < sbio->page_count; i++) {
2088 		struct scrub_page *spage = sbio->pagev[i];
2089 		struct scrub_block *sblock = spage->sblock;
2090 
2091 		if (atomic_dec_and_test(&sblock->outstanding_pages))
2092 			scrub_block_complete(sblock);
2093 		scrub_block_put(sblock);
2094 	}
2095 
2096 	bio_put(sbio->bio);
2097 	sbio->bio = NULL;
2098 	spin_lock(&sctx->list_lock);
2099 	sbio->next_free = sctx->first_free;
2100 	sctx->first_free = sbio->index;
2101 	spin_unlock(&sctx->list_lock);
2102 
2103 	if (sctx->is_dev_replace &&
2104 	    atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2105 		mutex_lock(&sctx->wr_ctx.wr_lock);
2106 		scrub_wr_submit(sctx);
2107 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2108 	}
2109 
2110 	scrub_pending_bio_dec(sctx);
2111 }
2112 
2113 static void scrub_block_complete(struct scrub_block *sblock)
2114 {
2115 	if (!sblock->no_io_error_seen) {
2116 		scrub_handle_errored_block(sblock);
2117 	} else {
2118 		/*
2119 		 * if has checksum error, write via repair mechanism in
2120 		 * dev replace case, otherwise write here in dev replace
2121 		 * case.
2122 		 */
2123 		if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2124 			scrub_write_block_to_dev_replace(sblock);
2125 	}
2126 }
2127 
2128 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2129 			   u8 *csum)
2130 {
2131 	struct btrfs_ordered_sum *sum = NULL;
2132 	int ret = 0;
2133 	unsigned long i;
2134 	unsigned long num_sectors;
2135 
2136 	while (!list_empty(&sctx->csum_list)) {
2137 		sum = list_first_entry(&sctx->csum_list,
2138 				       struct btrfs_ordered_sum, list);
2139 		if (sum->bytenr > logical)
2140 			return 0;
2141 		if (sum->bytenr + sum->len > logical)
2142 			break;
2143 
2144 		++sctx->stat.csum_discards;
2145 		list_del(&sum->list);
2146 		kfree(sum);
2147 		sum = NULL;
2148 	}
2149 	if (!sum)
2150 		return 0;
2151 
2152 	num_sectors = sum->len / sctx->sectorsize;
2153 	for (i = 0; i < num_sectors; ++i) {
2154 		if (sum->sums[i].bytenr == logical) {
2155 			memcpy(csum, &sum->sums[i].sum, sctx->csum_size);
2156 			ret = 1;
2157 			break;
2158 		}
2159 	}
2160 	if (ret && i == num_sectors - 1) {
2161 		list_del(&sum->list);
2162 		kfree(sum);
2163 	}
2164 	return ret;
2165 }
2166 
2167 /* scrub extent tries to collect up to 64 kB for each bio */
2168 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2169 			u64 physical, struct btrfs_device *dev, u64 flags,
2170 			u64 gen, int mirror_num, u64 physical_for_dev_replace)
2171 {
2172 	int ret;
2173 	u8 csum[BTRFS_CSUM_SIZE];
2174 	u32 blocksize;
2175 
2176 	if (flags & BTRFS_EXTENT_FLAG_DATA) {
2177 		blocksize = sctx->sectorsize;
2178 		spin_lock(&sctx->stat_lock);
2179 		sctx->stat.data_extents_scrubbed++;
2180 		sctx->stat.data_bytes_scrubbed += len;
2181 		spin_unlock(&sctx->stat_lock);
2182 	} else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2183 		WARN_ON(sctx->nodesize != sctx->leafsize);
2184 		blocksize = sctx->nodesize;
2185 		spin_lock(&sctx->stat_lock);
2186 		sctx->stat.tree_extents_scrubbed++;
2187 		sctx->stat.tree_bytes_scrubbed += len;
2188 		spin_unlock(&sctx->stat_lock);
2189 	} else {
2190 		blocksize = sctx->sectorsize;
2191 		WARN_ON(1);
2192 	}
2193 
2194 	while (len) {
2195 		u64 l = min_t(u64, len, blocksize);
2196 		int have_csum = 0;
2197 
2198 		if (flags & BTRFS_EXTENT_FLAG_DATA) {
2199 			/* push csums to sbio */
2200 			have_csum = scrub_find_csum(sctx, logical, l, csum);
2201 			if (have_csum == 0)
2202 				++sctx->stat.no_csum;
2203 			if (sctx->is_dev_replace && !have_csum) {
2204 				ret = copy_nocow_pages(sctx, logical, l,
2205 						       mirror_num,
2206 						      physical_for_dev_replace);
2207 				goto behind_scrub_pages;
2208 			}
2209 		}
2210 		ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2211 				  mirror_num, have_csum ? csum : NULL, 0,
2212 				  physical_for_dev_replace);
2213 behind_scrub_pages:
2214 		if (ret)
2215 			return ret;
2216 		len -= l;
2217 		logical += l;
2218 		physical += l;
2219 		physical_for_dev_replace += l;
2220 	}
2221 	return 0;
2222 }
2223 
2224 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2225 					   struct map_lookup *map,
2226 					   struct btrfs_device *scrub_dev,
2227 					   int num, u64 base, u64 length,
2228 					   int is_dev_replace)
2229 {
2230 	struct btrfs_path *path;
2231 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2232 	struct btrfs_root *root = fs_info->extent_root;
2233 	struct btrfs_root *csum_root = fs_info->csum_root;
2234 	struct btrfs_extent_item *extent;
2235 	struct blk_plug plug;
2236 	u64 flags;
2237 	int ret;
2238 	int slot;
2239 	int i;
2240 	u64 nstripes;
2241 	struct extent_buffer *l;
2242 	struct btrfs_key key;
2243 	u64 physical;
2244 	u64 logical;
2245 	u64 generation;
2246 	int mirror_num;
2247 	struct reada_control *reada1;
2248 	struct reada_control *reada2;
2249 	struct btrfs_key key_start;
2250 	struct btrfs_key key_end;
2251 	u64 increment = map->stripe_len;
2252 	u64 offset;
2253 	u64 extent_logical;
2254 	u64 extent_physical;
2255 	u64 extent_len;
2256 	struct btrfs_device *extent_dev;
2257 	int extent_mirror_num;
2258 
2259 	if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2260 			 BTRFS_BLOCK_GROUP_RAID6)) {
2261 		if (num >= nr_data_stripes(map)) {
2262 			return 0;
2263 		}
2264 	}
2265 
2266 	nstripes = length;
2267 	offset = 0;
2268 	do_div(nstripes, map->stripe_len);
2269 	if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2270 		offset = map->stripe_len * num;
2271 		increment = map->stripe_len * map->num_stripes;
2272 		mirror_num = 1;
2273 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2274 		int factor = map->num_stripes / map->sub_stripes;
2275 		offset = map->stripe_len * (num / map->sub_stripes);
2276 		increment = map->stripe_len * factor;
2277 		mirror_num = num % map->sub_stripes + 1;
2278 	} else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2279 		increment = map->stripe_len;
2280 		mirror_num = num % map->num_stripes + 1;
2281 	} else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2282 		increment = map->stripe_len;
2283 		mirror_num = num % map->num_stripes + 1;
2284 	} else {
2285 		increment = map->stripe_len;
2286 		mirror_num = 1;
2287 	}
2288 
2289 	path = btrfs_alloc_path();
2290 	if (!path)
2291 		return -ENOMEM;
2292 
2293 	/*
2294 	 * work on commit root. The related disk blocks are static as
2295 	 * long as COW is applied. This means, it is save to rewrite
2296 	 * them to repair disk errors without any race conditions
2297 	 */
2298 	path->search_commit_root = 1;
2299 	path->skip_locking = 1;
2300 
2301 	/*
2302 	 * trigger the readahead for extent tree csum tree and wait for
2303 	 * completion. During readahead, the scrub is officially paused
2304 	 * to not hold off transaction commits
2305 	 */
2306 	logical = base + offset;
2307 
2308 	wait_event(sctx->list_wait,
2309 		   atomic_read(&sctx->bios_in_flight) == 0);
2310 	atomic_inc(&fs_info->scrubs_paused);
2311 	wake_up(&fs_info->scrub_pause_wait);
2312 
2313 	/* FIXME it might be better to start readahead at commit root */
2314 	key_start.objectid = logical;
2315 	key_start.type = BTRFS_EXTENT_ITEM_KEY;
2316 	key_start.offset = (u64)0;
2317 	key_end.objectid = base + offset + nstripes * increment;
2318 	key_end.type = BTRFS_EXTENT_ITEM_KEY;
2319 	key_end.offset = (u64)0;
2320 	reada1 = btrfs_reada_add(root, &key_start, &key_end);
2321 
2322 	key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2323 	key_start.type = BTRFS_EXTENT_CSUM_KEY;
2324 	key_start.offset = logical;
2325 	key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2326 	key_end.type = BTRFS_EXTENT_CSUM_KEY;
2327 	key_end.offset = base + offset + nstripes * increment;
2328 	reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2329 
2330 	if (!IS_ERR(reada1))
2331 		btrfs_reada_wait(reada1);
2332 	if (!IS_ERR(reada2))
2333 		btrfs_reada_wait(reada2);
2334 
2335 	mutex_lock(&fs_info->scrub_lock);
2336 	while (atomic_read(&fs_info->scrub_pause_req)) {
2337 		mutex_unlock(&fs_info->scrub_lock);
2338 		wait_event(fs_info->scrub_pause_wait,
2339 		   atomic_read(&fs_info->scrub_pause_req) == 0);
2340 		mutex_lock(&fs_info->scrub_lock);
2341 	}
2342 	atomic_dec(&fs_info->scrubs_paused);
2343 	mutex_unlock(&fs_info->scrub_lock);
2344 	wake_up(&fs_info->scrub_pause_wait);
2345 
2346 	/*
2347 	 * collect all data csums for the stripe to avoid seeking during
2348 	 * the scrub. This might currently (crc32) end up to be about 1MB
2349 	 */
2350 	blk_start_plug(&plug);
2351 
2352 	/*
2353 	 * now find all extents for each stripe and scrub them
2354 	 */
2355 	logical = base + offset;
2356 	physical = map->stripes[num].physical;
2357 	ret = 0;
2358 	for (i = 0; i < nstripes; ++i) {
2359 		/*
2360 		 * canceled?
2361 		 */
2362 		if (atomic_read(&fs_info->scrub_cancel_req) ||
2363 		    atomic_read(&sctx->cancel_req)) {
2364 			ret = -ECANCELED;
2365 			goto out;
2366 		}
2367 		/*
2368 		 * check to see if we have to pause
2369 		 */
2370 		if (atomic_read(&fs_info->scrub_pause_req)) {
2371 			/* push queued extents */
2372 			atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2373 			scrub_submit(sctx);
2374 			mutex_lock(&sctx->wr_ctx.wr_lock);
2375 			scrub_wr_submit(sctx);
2376 			mutex_unlock(&sctx->wr_ctx.wr_lock);
2377 			wait_event(sctx->list_wait,
2378 				   atomic_read(&sctx->bios_in_flight) == 0);
2379 			atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2380 			atomic_inc(&fs_info->scrubs_paused);
2381 			wake_up(&fs_info->scrub_pause_wait);
2382 			mutex_lock(&fs_info->scrub_lock);
2383 			while (atomic_read(&fs_info->scrub_pause_req)) {
2384 				mutex_unlock(&fs_info->scrub_lock);
2385 				wait_event(fs_info->scrub_pause_wait,
2386 				   atomic_read(&fs_info->scrub_pause_req) == 0);
2387 				mutex_lock(&fs_info->scrub_lock);
2388 			}
2389 			atomic_dec(&fs_info->scrubs_paused);
2390 			mutex_unlock(&fs_info->scrub_lock);
2391 			wake_up(&fs_info->scrub_pause_wait);
2392 		}
2393 
2394 		ret = btrfs_lookup_csums_range(csum_root, logical,
2395 					       logical + map->stripe_len - 1,
2396 					       &sctx->csum_list, 1);
2397 		if (ret)
2398 			goto out;
2399 
2400 		key.objectid = logical;
2401 		key.type = BTRFS_EXTENT_ITEM_KEY;
2402 		key.offset = (u64)0;
2403 
2404 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2405 		if (ret < 0)
2406 			goto out;
2407 		if (ret > 0) {
2408 			ret = btrfs_previous_item(root, path, 0,
2409 						  BTRFS_EXTENT_ITEM_KEY);
2410 			if (ret < 0)
2411 				goto out;
2412 			if (ret > 0) {
2413 				/* there's no smaller item, so stick with the
2414 				 * larger one */
2415 				btrfs_release_path(path);
2416 				ret = btrfs_search_slot(NULL, root, &key,
2417 							path, 0, 0);
2418 				if (ret < 0)
2419 					goto out;
2420 			}
2421 		}
2422 
2423 		while (1) {
2424 			l = path->nodes[0];
2425 			slot = path->slots[0];
2426 			if (slot >= btrfs_header_nritems(l)) {
2427 				ret = btrfs_next_leaf(root, path);
2428 				if (ret == 0)
2429 					continue;
2430 				if (ret < 0)
2431 					goto out;
2432 
2433 				break;
2434 			}
2435 			btrfs_item_key_to_cpu(l, &key, slot);
2436 
2437 			if (key.objectid + key.offset <= logical)
2438 				goto next;
2439 
2440 			if (key.objectid >= logical + map->stripe_len)
2441 				break;
2442 
2443 			if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
2444 				goto next;
2445 
2446 			extent = btrfs_item_ptr(l, slot,
2447 						struct btrfs_extent_item);
2448 			flags = btrfs_extent_flags(l, extent);
2449 			generation = btrfs_extent_generation(l, extent);
2450 
2451 			if (key.objectid < logical &&
2452 			    (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2453 				printk(KERN_ERR
2454 				       "btrfs scrub: tree block %llu spanning "
2455 				       "stripes, ignored. logical=%llu\n",
2456 				       (unsigned long long)key.objectid,
2457 				       (unsigned long long)logical);
2458 				goto next;
2459 			}
2460 
2461 			/*
2462 			 * trim extent to this stripe
2463 			 */
2464 			if (key.objectid < logical) {
2465 				key.offset -= logical - key.objectid;
2466 				key.objectid = logical;
2467 			}
2468 			if (key.objectid + key.offset >
2469 			    logical + map->stripe_len) {
2470 				key.offset = logical + map->stripe_len -
2471 					     key.objectid;
2472 			}
2473 
2474 			extent_logical = key.objectid;
2475 			extent_physical = key.objectid - logical + physical;
2476 			extent_len = key.offset;
2477 			extent_dev = scrub_dev;
2478 			extent_mirror_num = mirror_num;
2479 			if (is_dev_replace)
2480 				scrub_remap_extent(fs_info, extent_logical,
2481 						   extent_len, &extent_physical,
2482 						   &extent_dev,
2483 						   &extent_mirror_num);
2484 			ret = scrub_extent(sctx, extent_logical, extent_len,
2485 					   extent_physical, extent_dev, flags,
2486 					   generation, extent_mirror_num,
2487 					   key.objectid - logical + physical);
2488 			if (ret)
2489 				goto out;
2490 
2491 next:
2492 			path->slots[0]++;
2493 		}
2494 		btrfs_release_path(path);
2495 		logical += increment;
2496 		physical += map->stripe_len;
2497 		spin_lock(&sctx->stat_lock);
2498 		sctx->stat.last_physical = physical;
2499 		spin_unlock(&sctx->stat_lock);
2500 	}
2501 out:
2502 	/* push queued extents */
2503 	scrub_submit(sctx);
2504 	mutex_lock(&sctx->wr_ctx.wr_lock);
2505 	scrub_wr_submit(sctx);
2506 	mutex_unlock(&sctx->wr_ctx.wr_lock);
2507 
2508 	blk_finish_plug(&plug);
2509 	btrfs_free_path(path);
2510 	return ret < 0 ? ret : 0;
2511 }
2512 
2513 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2514 					  struct btrfs_device *scrub_dev,
2515 					  u64 chunk_tree, u64 chunk_objectid,
2516 					  u64 chunk_offset, u64 length,
2517 					  u64 dev_offset, int is_dev_replace)
2518 {
2519 	struct btrfs_mapping_tree *map_tree =
2520 		&sctx->dev_root->fs_info->mapping_tree;
2521 	struct map_lookup *map;
2522 	struct extent_map *em;
2523 	int i;
2524 	int ret = 0;
2525 
2526 	read_lock(&map_tree->map_tree.lock);
2527 	em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2528 	read_unlock(&map_tree->map_tree.lock);
2529 
2530 	if (!em)
2531 		return -EINVAL;
2532 
2533 	map = (struct map_lookup *)em->bdev;
2534 	if (em->start != chunk_offset)
2535 		goto out;
2536 
2537 	if (em->len < length)
2538 		goto out;
2539 
2540 	for (i = 0; i < map->num_stripes; ++i) {
2541 		if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2542 		    map->stripes[i].physical == dev_offset) {
2543 			ret = scrub_stripe(sctx, map, scrub_dev, i,
2544 					   chunk_offset, length,
2545 					   is_dev_replace);
2546 			if (ret)
2547 				goto out;
2548 		}
2549 	}
2550 out:
2551 	free_extent_map(em);
2552 
2553 	return ret;
2554 }
2555 
2556 static noinline_for_stack
2557 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2558 			   struct btrfs_device *scrub_dev, u64 start, u64 end,
2559 			   int is_dev_replace)
2560 {
2561 	struct btrfs_dev_extent *dev_extent = NULL;
2562 	struct btrfs_path *path;
2563 	struct btrfs_root *root = sctx->dev_root;
2564 	struct btrfs_fs_info *fs_info = root->fs_info;
2565 	u64 length;
2566 	u64 chunk_tree;
2567 	u64 chunk_objectid;
2568 	u64 chunk_offset;
2569 	int ret;
2570 	int slot;
2571 	struct extent_buffer *l;
2572 	struct btrfs_key key;
2573 	struct btrfs_key found_key;
2574 	struct btrfs_block_group_cache *cache;
2575 	struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2576 
2577 	path = btrfs_alloc_path();
2578 	if (!path)
2579 		return -ENOMEM;
2580 
2581 	path->reada = 2;
2582 	path->search_commit_root = 1;
2583 	path->skip_locking = 1;
2584 
2585 	key.objectid = scrub_dev->devid;
2586 	key.offset = 0ull;
2587 	key.type = BTRFS_DEV_EXTENT_KEY;
2588 
2589 	while (1) {
2590 		ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2591 		if (ret < 0)
2592 			break;
2593 		if (ret > 0) {
2594 			if (path->slots[0] >=
2595 			    btrfs_header_nritems(path->nodes[0])) {
2596 				ret = btrfs_next_leaf(root, path);
2597 				if (ret)
2598 					break;
2599 			}
2600 		}
2601 
2602 		l = path->nodes[0];
2603 		slot = path->slots[0];
2604 
2605 		btrfs_item_key_to_cpu(l, &found_key, slot);
2606 
2607 		if (found_key.objectid != scrub_dev->devid)
2608 			break;
2609 
2610 		if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2611 			break;
2612 
2613 		if (found_key.offset >= end)
2614 			break;
2615 
2616 		if (found_key.offset < key.offset)
2617 			break;
2618 
2619 		dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2620 		length = btrfs_dev_extent_length(l, dev_extent);
2621 
2622 		if (found_key.offset + length <= start) {
2623 			key.offset = found_key.offset + length;
2624 			btrfs_release_path(path);
2625 			continue;
2626 		}
2627 
2628 		chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2629 		chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2630 		chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2631 
2632 		/*
2633 		 * get a reference on the corresponding block group to prevent
2634 		 * the chunk from going away while we scrub it
2635 		 */
2636 		cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2637 		if (!cache) {
2638 			ret = -ENOENT;
2639 			break;
2640 		}
2641 		dev_replace->cursor_right = found_key.offset + length;
2642 		dev_replace->cursor_left = found_key.offset;
2643 		dev_replace->item_needs_writeback = 1;
2644 		ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2645 				  chunk_offset, length, found_key.offset,
2646 				  is_dev_replace);
2647 
2648 		/*
2649 		 * flush, submit all pending read and write bios, afterwards
2650 		 * wait for them.
2651 		 * Note that in the dev replace case, a read request causes
2652 		 * write requests that are submitted in the read completion
2653 		 * worker. Therefore in the current situation, it is required
2654 		 * that all write requests are flushed, so that all read and
2655 		 * write requests are really completed when bios_in_flight
2656 		 * changes to 0.
2657 		 */
2658 		atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2659 		scrub_submit(sctx);
2660 		mutex_lock(&sctx->wr_ctx.wr_lock);
2661 		scrub_wr_submit(sctx);
2662 		mutex_unlock(&sctx->wr_ctx.wr_lock);
2663 
2664 		wait_event(sctx->list_wait,
2665 			   atomic_read(&sctx->bios_in_flight) == 0);
2666 		atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2667 		atomic_inc(&fs_info->scrubs_paused);
2668 		wake_up(&fs_info->scrub_pause_wait);
2669 		wait_event(sctx->list_wait,
2670 			   atomic_read(&sctx->workers_pending) == 0);
2671 
2672 		mutex_lock(&fs_info->scrub_lock);
2673 		while (atomic_read(&fs_info->scrub_pause_req)) {
2674 			mutex_unlock(&fs_info->scrub_lock);
2675 			wait_event(fs_info->scrub_pause_wait,
2676 			   atomic_read(&fs_info->scrub_pause_req) == 0);
2677 			mutex_lock(&fs_info->scrub_lock);
2678 		}
2679 		atomic_dec(&fs_info->scrubs_paused);
2680 		mutex_unlock(&fs_info->scrub_lock);
2681 		wake_up(&fs_info->scrub_pause_wait);
2682 
2683 		dev_replace->cursor_left = dev_replace->cursor_right;
2684 		dev_replace->item_needs_writeback = 1;
2685 		btrfs_put_block_group(cache);
2686 		if (ret)
2687 			break;
2688 		if (is_dev_replace &&
2689 		    atomic64_read(&dev_replace->num_write_errors) > 0) {
2690 			ret = -EIO;
2691 			break;
2692 		}
2693 		if (sctx->stat.malloc_errors > 0) {
2694 			ret = -ENOMEM;
2695 			break;
2696 		}
2697 
2698 		key.offset = found_key.offset + length;
2699 		btrfs_release_path(path);
2700 	}
2701 
2702 	btrfs_free_path(path);
2703 
2704 	/*
2705 	 * ret can still be 1 from search_slot or next_leaf,
2706 	 * that's not an error
2707 	 */
2708 	return ret < 0 ? ret : 0;
2709 }
2710 
2711 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2712 					   struct btrfs_device *scrub_dev)
2713 {
2714 	int	i;
2715 	u64	bytenr;
2716 	u64	gen;
2717 	int	ret;
2718 	struct btrfs_root *root = sctx->dev_root;
2719 
2720 	if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2721 		return -EIO;
2722 
2723 	gen = root->fs_info->last_trans_committed;
2724 
2725 	for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2726 		bytenr = btrfs_sb_offset(i);
2727 		if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2728 			break;
2729 
2730 		ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2731 				  scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2732 				  NULL, 1, bytenr);
2733 		if (ret)
2734 			return ret;
2735 	}
2736 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2737 
2738 	return 0;
2739 }
2740 
2741 /*
2742  * get a reference count on fs_info->scrub_workers. start worker if necessary
2743  */
2744 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2745 						int is_dev_replace)
2746 {
2747 	int ret = 0;
2748 
2749 	mutex_lock(&fs_info->scrub_lock);
2750 	if (fs_info->scrub_workers_refcnt == 0) {
2751 		if (is_dev_replace)
2752 			btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2753 					&fs_info->generic_worker);
2754 		else
2755 			btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2756 					fs_info->thread_pool_size,
2757 					&fs_info->generic_worker);
2758 		fs_info->scrub_workers.idle_thresh = 4;
2759 		ret = btrfs_start_workers(&fs_info->scrub_workers);
2760 		if (ret)
2761 			goto out;
2762 		btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2763 				   "scrubwrc",
2764 				   fs_info->thread_pool_size,
2765 				   &fs_info->generic_worker);
2766 		fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2767 		ret = btrfs_start_workers(
2768 				&fs_info->scrub_wr_completion_workers);
2769 		if (ret)
2770 			goto out;
2771 		btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2772 				   &fs_info->generic_worker);
2773 		ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2774 		if (ret)
2775 			goto out;
2776 	}
2777 	++fs_info->scrub_workers_refcnt;
2778 out:
2779 	mutex_unlock(&fs_info->scrub_lock);
2780 
2781 	return ret;
2782 }
2783 
2784 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2785 {
2786 	mutex_lock(&fs_info->scrub_lock);
2787 	if (--fs_info->scrub_workers_refcnt == 0) {
2788 		btrfs_stop_workers(&fs_info->scrub_workers);
2789 		btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2790 		btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2791 	}
2792 	WARN_ON(fs_info->scrub_workers_refcnt < 0);
2793 	mutex_unlock(&fs_info->scrub_lock);
2794 }
2795 
2796 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2797 		    u64 end, struct btrfs_scrub_progress *progress,
2798 		    int readonly, int is_dev_replace)
2799 {
2800 	struct scrub_ctx *sctx;
2801 	int ret;
2802 	struct btrfs_device *dev;
2803 
2804 	if (btrfs_fs_closing(fs_info))
2805 		return -EINVAL;
2806 
2807 	/*
2808 	 * check some assumptions
2809 	 */
2810 	if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2811 		printk(KERN_ERR
2812 		       "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2813 		       fs_info->chunk_root->nodesize,
2814 		       fs_info->chunk_root->leafsize);
2815 		return -EINVAL;
2816 	}
2817 
2818 	if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2819 		/*
2820 		 * in this case scrub is unable to calculate the checksum
2821 		 * the way scrub is implemented. Do not handle this
2822 		 * situation at all because it won't ever happen.
2823 		 */
2824 		printk(KERN_ERR
2825 		       "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2826 		       fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2827 		return -EINVAL;
2828 	}
2829 
2830 	if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2831 		/* not supported for data w/o checksums */
2832 		printk(KERN_ERR
2833 		       "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2834 		       fs_info->chunk_root->sectorsize,
2835 		       (unsigned long long)PAGE_SIZE);
2836 		return -EINVAL;
2837 	}
2838 
2839 	if (fs_info->chunk_root->nodesize >
2840 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2841 	    fs_info->chunk_root->sectorsize >
2842 	    PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2843 		/*
2844 		 * would exhaust the array bounds of pagev member in
2845 		 * struct scrub_block
2846 		 */
2847 		pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2848 		       fs_info->chunk_root->nodesize,
2849 		       SCRUB_MAX_PAGES_PER_BLOCK,
2850 		       fs_info->chunk_root->sectorsize,
2851 		       SCRUB_MAX_PAGES_PER_BLOCK);
2852 		return -EINVAL;
2853 	}
2854 
2855 	ret = scrub_workers_get(fs_info, is_dev_replace);
2856 	if (ret)
2857 		return ret;
2858 
2859 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
2860 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2861 	if (!dev || (dev->missing && !is_dev_replace)) {
2862 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2863 		scrub_workers_put(fs_info);
2864 		return -ENODEV;
2865 	}
2866 	mutex_lock(&fs_info->scrub_lock);
2867 
2868 	if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2869 		mutex_unlock(&fs_info->scrub_lock);
2870 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2871 		scrub_workers_put(fs_info);
2872 		return -EIO;
2873 	}
2874 
2875 	btrfs_dev_replace_lock(&fs_info->dev_replace);
2876 	if (dev->scrub_device ||
2877 	    (!is_dev_replace &&
2878 	     btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2879 		btrfs_dev_replace_unlock(&fs_info->dev_replace);
2880 		mutex_unlock(&fs_info->scrub_lock);
2881 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2882 		scrub_workers_put(fs_info);
2883 		return -EINPROGRESS;
2884 	}
2885 	btrfs_dev_replace_unlock(&fs_info->dev_replace);
2886 	sctx = scrub_setup_ctx(dev, is_dev_replace);
2887 	if (IS_ERR(sctx)) {
2888 		mutex_unlock(&fs_info->scrub_lock);
2889 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2890 		scrub_workers_put(fs_info);
2891 		return PTR_ERR(sctx);
2892 	}
2893 	sctx->readonly = readonly;
2894 	dev->scrub_device = sctx;
2895 
2896 	atomic_inc(&fs_info->scrubs_running);
2897 	mutex_unlock(&fs_info->scrub_lock);
2898 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2899 
2900 	if (!is_dev_replace) {
2901 		down_read(&fs_info->scrub_super_lock);
2902 		ret = scrub_supers(sctx, dev);
2903 		up_read(&fs_info->scrub_super_lock);
2904 	}
2905 
2906 	if (!ret)
2907 		ret = scrub_enumerate_chunks(sctx, dev, start, end,
2908 					     is_dev_replace);
2909 
2910 	wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2911 	atomic_dec(&fs_info->scrubs_running);
2912 	wake_up(&fs_info->scrub_pause_wait);
2913 
2914 	wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2915 
2916 	if (progress)
2917 		memcpy(progress, &sctx->stat, sizeof(*progress));
2918 
2919 	mutex_lock(&fs_info->scrub_lock);
2920 	dev->scrub_device = NULL;
2921 	mutex_unlock(&fs_info->scrub_lock);
2922 
2923 	scrub_free_ctx(sctx);
2924 	scrub_workers_put(fs_info);
2925 
2926 	return ret;
2927 }
2928 
2929 void btrfs_scrub_pause(struct btrfs_root *root)
2930 {
2931 	struct btrfs_fs_info *fs_info = root->fs_info;
2932 
2933 	mutex_lock(&fs_info->scrub_lock);
2934 	atomic_inc(&fs_info->scrub_pause_req);
2935 	while (atomic_read(&fs_info->scrubs_paused) !=
2936 	       atomic_read(&fs_info->scrubs_running)) {
2937 		mutex_unlock(&fs_info->scrub_lock);
2938 		wait_event(fs_info->scrub_pause_wait,
2939 			   atomic_read(&fs_info->scrubs_paused) ==
2940 			   atomic_read(&fs_info->scrubs_running));
2941 		mutex_lock(&fs_info->scrub_lock);
2942 	}
2943 	mutex_unlock(&fs_info->scrub_lock);
2944 }
2945 
2946 void btrfs_scrub_continue(struct btrfs_root *root)
2947 {
2948 	struct btrfs_fs_info *fs_info = root->fs_info;
2949 
2950 	atomic_dec(&fs_info->scrub_pause_req);
2951 	wake_up(&fs_info->scrub_pause_wait);
2952 }
2953 
2954 void btrfs_scrub_pause_super(struct btrfs_root *root)
2955 {
2956 	down_write(&root->fs_info->scrub_super_lock);
2957 }
2958 
2959 void btrfs_scrub_continue_super(struct btrfs_root *root)
2960 {
2961 	up_write(&root->fs_info->scrub_super_lock);
2962 }
2963 
2964 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2965 {
2966 	mutex_lock(&fs_info->scrub_lock);
2967 	if (!atomic_read(&fs_info->scrubs_running)) {
2968 		mutex_unlock(&fs_info->scrub_lock);
2969 		return -ENOTCONN;
2970 	}
2971 
2972 	atomic_inc(&fs_info->scrub_cancel_req);
2973 	while (atomic_read(&fs_info->scrubs_running)) {
2974 		mutex_unlock(&fs_info->scrub_lock);
2975 		wait_event(fs_info->scrub_pause_wait,
2976 			   atomic_read(&fs_info->scrubs_running) == 0);
2977 		mutex_lock(&fs_info->scrub_lock);
2978 	}
2979 	atomic_dec(&fs_info->scrub_cancel_req);
2980 	mutex_unlock(&fs_info->scrub_lock);
2981 
2982 	return 0;
2983 }
2984 
2985 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2986 			   struct btrfs_device *dev)
2987 {
2988 	struct scrub_ctx *sctx;
2989 
2990 	mutex_lock(&fs_info->scrub_lock);
2991 	sctx = dev->scrub_device;
2992 	if (!sctx) {
2993 		mutex_unlock(&fs_info->scrub_lock);
2994 		return -ENOTCONN;
2995 	}
2996 	atomic_inc(&sctx->cancel_req);
2997 	while (dev->scrub_device) {
2998 		mutex_unlock(&fs_info->scrub_lock);
2999 		wait_event(fs_info->scrub_pause_wait,
3000 			   dev->scrub_device == NULL);
3001 		mutex_lock(&fs_info->scrub_lock);
3002 	}
3003 	mutex_unlock(&fs_info->scrub_lock);
3004 
3005 	return 0;
3006 }
3007 
3008 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
3009 {
3010 	struct btrfs_fs_info *fs_info = root->fs_info;
3011 	struct btrfs_device *dev;
3012 	int ret;
3013 
3014 	/*
3015 	 * we have to hold the device_list_mutex here so the device
3016 	 * does not go away in cancel_dev. FIXME: find a better solution
3017 	 */
3018 	mutex_lock(&fs_info->fs_devices->device_list_mutex);
3019 	dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3020 	if (!dev) {
3021 		mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3022 		return -ENODEV;
3023 	}
3024 	ret = btrfs_scrub_cancel_dev(fs_info, dev);
3025 	mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3026 
3027 	return ret;
3028 }
3029 
3030 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3031 			 struct btrfs_scrub_progress *progress)
3032 {
3033 	struct btrfs_device *dev;
3034 	struct scrub_ctx *sctx = NULL;
3035 
3036 	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3037 	dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3038 	if (dev)
3039 		sctx = dev->scrub_device;
3040 	if (sctx)
3041 		memcpy(progress, &sctx->stat, sizeof(*progress));
3042 	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3043 
3044 	return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3045 }
3046 
3047 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3048 			       u64 extent_logical, u64 extent_len,
3049 			       u64 *extent_physical,
3050 			       struct btrfs_device **extent_dev,
3051 			       int *extent_mirror_num)
3052 {
3053 	u64 mapped_length;
3054 	struct btrfs_bio *bbio = NULL;
3055 	int ret;
3056 
3057 	mapped_length = extent_len;
3058 	ret = btrfs_map_block(fs_info, READ, extent_logical,
3059 			      &mapped_length, &bbio, 0);
3060 	if (ret || !bbio || mapped_length < extent_len ||
3061 	    !bbio->stripes[0].dev->bdev) {
3062 		kfree(bbio);
3063 		return;
3064 	}
3065 
3066 	*extent_physical = bbio->stripes[0].physical;
3067 	*extent_mirror_num = bbio->mirror_num;
3068 	*extent_dev = bbio->stripes[0].dev;
3069 	kfree(bbio);
3070 }
3071 
3072 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3073 			      struct scrub_wr_ctx *wr_ctx,
3074 			      struct btrfs_fs_info *fs_info,
3075 			      struct btrfs_device *dev,
3076 			      int is_dev_replace)
3077 {
3078 	WARN_ON(wr_ctx->wr_curr_bio != NULL);
3079 
3080 	mutex_init(&wr_ctx->wr_lock);
3081 	wr_ctx->wr_curr_bio = NULL;
3082 	if (!is_dev_replace)
3083 		return 0;
3084 
3085 	WARN_ON(!dev->bdev);
3086 	wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3087 					 bio_get_nr_vecs(dev->bdev));
3088 	wr_ctx->tgtdev = dev;
3089 	atomic_set(&wr_ctx->flush_all_writes, 0);
3090 	return 0;
3091 }
3092 
3093 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3094 {
3095 	mutex_lock(&wr_ctx->wr_lock);
3096 	kfree(wr_ctx->wr_curr_bio);
3097 	wr_ctx->wr_curr_bio = NULL;
3098 	mutex_unlock(&wr_ctx->wr_lock);
3099 }
3100 
3101 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3102 			    int mirror_num, u64 physical_for_dev_replace)
3103 {
3104 	struct scrub_copy_nocow_ctx *nocow_ctx;
3105 	struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3106 
3107 	nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3108 	if (!nocow_ctx) {
3109 		spin_lock(&sctx->stat_lock);
3110 		sctx->stat.malloc_errors++;
3111 		spin_unlock(&sctx->stat_lock);
3112 		return -ENOMEM;
3113 	}
3114 
3115 	scrub_pending_trans_workers_inc(sctx);
3116 
3117 	nocow_ctx->sctx = sctx;
3118 	nocow_ctx->logical = logical;
3119 	nocow_ctx->len = len;
3120 	nocow_ctx->mirror_num = mirror_num;
3121 	nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3122 	nocow_ctx->work.func = copy_nocow_pages_worker;
3123 	btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3124 			   &nocow_ctx->work);
3125 
3126 	return 0;
3127 }
3128 
3129 static void copy_nocow_pages_worker(struct btrfs_work *work)
3130 {
3131 	struct scrub_copy_nocow_ctx *nocow_ctx =
3132 		container_of(work, struct scrub_copy_nocow_ctx, work);
3133 	struct scrub_ctx *sctx = nocow_ctx->sctx;
3134 	u64 logical = nocow_ctx->logical;
3135 	u64 len = nocow_ctx->len;
3136 	int mirror_num = nocow_ctx->mirror_num;
3137 	u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3138 	int ret;
3139 	struct btrfs_trans_handle *trans = NULL;
3140 	struct btrfs_fs_info *fs_info;
3141 	struct btrfs_path *path;
3142 	struct btrfs_root *root;
3143 	int not_written = 0;
3144 
3145 	fs_info = sctx->dev_root->fs_info;
3146 	root = fs_info->extent_root;
3147 
3148 	path = btrfs_alloc_path();
3149 	if (!path) {
3150 		spin_lock(&sctx->stat_lock);
3151 		sctx->stat.malloc_errors++;
3152 		spin_unlock(&sctx->stat_lock);
3153 		not_written = 1;
3154 		goto out;
3155 	}
3156 
3157 	trans = btrfs_join_transaction(root);
3158 	if (IS_ERR(trans)) {
3159 		not_written = 1;
3160 		goto out;
3161 	}
3162 
3163 	ret = iterate_inodes_from_logical(logical, fs_info, path,
3164 					  copy_nocow_pages_for_inode,
3165 					  nocow_ctx);
3166 	if (ret != 0 && ret != -ENOENT) {
3167 		pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3168 			(unsigned long long)logical,
3169 			(unsigned long long)physical_for_dev_replace,
3170 			(unsigned long long)len,
3171 			(unsigned long long)mirror_num, ret);
3172 		not_written = 1;
3173 		goto out;
3174 	}
3175 
3176 out:
3177 	if (trans && !IS_ERR(trans))
3178 		btrfs_end_transaction(trans, root);
3179 	if (not_written)
3180 		btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3181 					    num_uncorrectable_read_errors);
3182 
3183 	btrfs_free_path(path);
3184 	kfree(nocow_ctx);
3185 
3186 	scrub_pending_trans_workers_dec(sctx);
3187 }
3188 
3189 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3190 {
3191 	unsigned long index;
3192 	struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3193 	int ret = 0;
3194 	struct btrfs_key key;
3195 	struct inode *inode = NULL;
3196 	struct btrfs_root *local_root;
3197 	u64 physical_for_dev_replace;
3198 	u64 len;
3199 	struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3200 	int srcu_index;
3201 
3202 	key.objectid = root;
3203 	key.type = BTRFS_ROOT_ITEM_KEY;
3204 	key.offset = (u64)-1;
3205 
3206 	srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3207 
3208 	local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3209 	if (IS_ERR(local_root)) {
3210 		srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3211 		return PTR_ERR(local_root);
3212 	}
3213 
3214 	key.type = BTRFS_INODE_ITEM_KEY;
3215 	key.objectid = inum;
3216 	key.offset = 0;
3217 	inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3218 	srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3219 	if (IS_ERR(inode))
3220 		return PTR_ERR(inode);
3221 
3222 	physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3223 	len = nocow_ctx->len;
3224 	while (len >= PAGE_CACHE_SIZE) {
3225 		struct page *page = NULL;
3226 		int ret_sub;
3227 
3228 		index = offset >> PAGE_CACHE_SHIFT;
3229 
3230 		page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3231 		if (!page) {
3232 			pr_err("find_or_create_page() failed\n");
3233 			ret = -ENOMEM;
3234 			goto next_page;
3235 		}
3236 
3237 		if (PageUptodate(page)) {
3238 			if (PageDirty(page))
3239 				goto next_page;
3240 		} else {
3241 			ClearPageError(page);
3242 			ret_sub = extent_read_full_page(&BTRFS_I(inode)->
3243 							 io_tree,
3244 							page, btrfs_get_extent,
3245 							nocow_ctx->mirror_num);
3246 			if (ret_sub) {
3247 				ret = ret_sub;
3248 				goto next_page;
3249 			}
3250 			wait_on_page_locked(page);
3251 			if (!PageUptodate(page)) {
3252 				ret = -EIO;
3253 				goto next_page;
3254 			}
3255 		}
3256 		ret_sub = write_page_nocow(nocow_ctx->sctx,
3257 					   physical_for_dev_replace, page);
3258 		if (ret_sub) {
3259 			ret = ret_sub;
3260 			goto next_page;
3261 		}
3262 
3263 next_page:
3264 		if (page) {
3265 			unlock_page(page);
3266 			put_page(page);
3267 		}
3268 		offset += PAGE_CACHE_SIZE;
3269 		physical_for_dev_replace += PAGE_CACHE_SIZE;
3270 		len -= PAGE_CACHE_SIZE;
3271 	}
3272 
3273 	if (inode)
3274 		iput(inode);
3275 	return ret;
3276 }
3277 
3278 static int write_page_nocow(struct scrub_ctx *sctx,
3279 			    u64 physical_for_dev_replace, struct page *page)
3280 {
3281 	struct bio *bio;
3282 	struct btrfs_device *dev;
3283 	int ret;
3284 	DECLARE_COMPLETION_ONSTACK(compl);
3285 
3286 	dev = sctx->wr_ctx.tgtdev;
3287 	if (!dev)
3288 		return -EIO;
3289 	if (!dev->bdev) {
3290 		printk_ratelimited(KERN_WARNING
3291 			"btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3292 		return -EIO;
3293 	}
3294 	bio = bio_alloc(GFP_NOFS, 1);
3295 	if (!bio) {
3296 		spin_lock(&sctx->stat_lock);
3297 		sctx->stat.malloc_errors++;
3298 		spin_unlock(&sctx->stat_lock);
3299 		return -ENOMEM;
3300 	}
3301 	bio->bi_private = &compl;
3302 	bio->bi_end_io = scrub_complete_bio_end_io;
3303 	bio->bi_size = 0;
3304 	bio->bi_sector = physical_for_dev_replace >> 9;
3305 	bio->bi_bdev = dev->bdev;
3306 	ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3307 	if (ret != PAGE_CACHE_SIZE) {
3308 leave_with_eio:
3309 		bio_put(bio);
3310 		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3311 		return -EIO;
3312 	}
3313 	btrfsic_submit_bio(WRITE_SYNC, bio);
3314 	wait_for_completion(&compl);
3315 
3316 	if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3317 		goto leave_with_eio;
3318 
3319 	bio_put(bio);
3320 	return 0;
3321 }
3322