Btrfs: update the checks for mixed block groups with big metadata blocks
[~shefty/rdma-dev.git] / fs / btrfs / scrub.c
1 /*
2  * Copyright (C) 2011 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 "check-integrity.h"
29
30 /*
31  * This is only the first step towards a full-features scrub. It reads all
32  * extent and super block and verifies the checksums. In case a bad checksum
33  * is found or the extent cannot be read, good data will be written back if
34  * any can be found.
35  *
36  * Future enhancements:
37  *  - In case an unrepairable extent is encountered, track which files are
38  *    affected and report them
39  *  - track and record media errors, throw out bad devices
40  *  - add a mode to also read unallocated space
41  */
42
43 struct scrub_block;
44 struct scrub_dev;
45
46 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
47 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
48 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
49
50 struct scrub_page {
51         struct scrub_block      *sblock;
52         struct page             *page;
53         struct block_device     *bdev;
54         u64                     flags;  /* extent flags */
55         u64                     generation;
56         u64                     logical;
57         u64                     physical;
58         struct {
59                 unsigned int    mirror_num:8;
60                 unsigned int    have_csum:1;
61                 unsigned int    io_error:1;
62         };
63         u8                      csum[BTRFS_CSUM_SIZE];
64 };
65
66 struct scrub_bio {
67         int                     index;
68         struct scrub_dev        *sdev;
69         struct bio              *bio;
70         int                     err;
71         u64                     logical;
72         u64                     physical;
73         struct scrub_page       *pagev[SCRUB_PAGES_PER_BIO];
74         int                     page_count;
75         int                     next_free;
76         struct btrfs_work       work;
77 };
78
79 struct scrub_block {
80         struct scrub_page       pagev[SCRUB_MAX_PAGES_PER_BLOCK];
81         int                     page_count;
82         atomic_t                outstanding_pages;
83         atomic_t                ref_count; /* free mem on transition to zero */
84         struct scrub_dev        *sdev;
85         struct {
86                 unsigned int    header_error:1;
87                 unsigned int    checksum_error:1;
88                 unsigned int    no_io_error_seen:1;
89         };
90 };
91
92 struct scrub_dev {
93         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
94         struct btrfs_device     *dev;
95         int                     first_free;
96         int                     curr;
97         atomic_t                in_flight;
98         atomic_t                fixup_cnt;
99         spinlock_t              list_lock;
100         wait_queue_head_t       list_wait;
101         u16                     csum_size;
102         struct list_head        csum_list;
103         atomic_t                cancel_req;
104         int                     readonly;
105         int                     pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */
106         u32                     sectorsize;
107         u32                     nodesize;
108         u32                     leafsize;
109         /*
110          * statistics
111          */
112         struct btrfs_scrub_progress stat;
113         spinlock_t              stat_lock;
114 };
115
116 struct scrub_fixup_nodatasum {
117         struct scrub_dev        *sdev;
118         u64                     logical;
119         struct btrfs_root       *root;
120         struct btrfs_work       work;
121         int                     mirror_num;
122 };
123
124 struct scrub_warning {
125         struct btrfs_path       *path;
126         u64                     extent_item_size;
127         char                    *scratch_buf;
128         char                    *msg_buf;
129         const char              *errstr;
130         sector_t                sector;
131         u64                     logical;
132         struct btrfs_device     *dev;
133         int                     msg_bufsize;
134         int                     scratch_bufsize;
135 };
136
137
138 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
139 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
140                                      struct btrfs_mapping_tree *map_tree,
141                                      u64 length, u64 logical,
142                                      struct scrub_block *sblock);
143 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
144                                struct scrub_block *sblock, int is_metadata,
145                                int have_csum, u8 *csum, u64 generation,
146                                u16 csum_size);
147 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
148                                          struct scrub_block *sblock,
149                                          int is_metadata, int have_csum,
150                                          const u8 *csum, u64 generation,
151                                          u16 csum_size);
152 static void scrub_complete_bio_end_io(struct bio *bio, int err);
153 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
154                                              struct scrub_block *sblock_good,
155                                              int force_write);
156 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
157                                             struct scrub_block *sblock_good,
158                                             int page_num, int force_write);
159 static int scrub_checksum_data(struct scrub_block *sblock);
160 static int scrub_checksum_tree_block(struct scrub_block *sblock);
161 static int scrub_checksum_super(struct scrub_block *sblock);
162 static void scrub_block_get(struct scrub_block *sblock);
163 static void scrub_block_put(struct scrub_block *sblock);
164 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
165                                  struct scrub_page *spage);
166 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
167                        u64 physical, u64 flags, u64 gen, int mirror_num,
168                        u8 *csum, int force);
169 static void scrub_bio_end_io(struct bio *bio, int err);
170 static void scrub_bio_end_io_worker(struct btrfs_work *work);
171 static void scrub_block_complete(struct scrub_block *sblock);
172
173
174 static void scrub_free_csums(struct scrub_dev *sdev)
175 {
176         while (!list_empty(&sdev->csum_list)) {
177                 struct btrfs_ordered_sum *sum;
178                 sum = list_first_entry(&sdev->csum_list,
179                                        struct btrfs_ordered_sum, list);
180                 list_del(&sum->list);
181                 kfree(sum);
182         }
183 }
184
185 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
186 {
187         int i;
188
189         if (!sdev)
190                 return;
191
192         /* this can happen when scrub is cancelled */
193         if (sdev->curr != -1) {
194                 struct scrub_bio *sbio = sdev->bios[sdev->curr];
195
196                 for (i = 0; i < sbio->page_count; i++) {
197                         BUG_ON(!sbio->pagev[i]);
198                         BUG_ON(!sbio->pagev[i]->page);
199                         scrub_block_put(sbio->pagev[i]->sblock);
200                 }
201                 bio_put(sbio->bio);
202         }
203
204         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
205                 struct scrub_bio *sbio = sdev->bios[i];
206
207                 if (!sbio)
208                         break;
209                 kfree(sbio);
210         }
211
212         scrub_free_csums(sdev);
213         kfree(sdev);
214 }
215
216 static noinline_for_stack
217 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
218 {
219         struct scrub_dev *sdev;
220         int             i;
221         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
222         int pages_per_bio;
223
224         pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO,
225                               bio_get_nr_vecs(dev->bdev));
226         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
227         if (!sdev)
228                 goto nomem;
229         sdev->dev = dev;
230         sdev->pages_per_bio = pages_per_bio;
231         sdev->curr = -1;
232         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
233                 struct scrub_bio *sbio;
234
235                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
236                 if (!sbio)
237                         goto nomem;
238                 sdev->bios[i] = sbio;
239
240                 sbio->index = i;
241                 sbio->sdev = sdev;
242                 sbio->page_count = 0;
243                 sbio->work.func = scrub_bio_end_io_worker;
244
245                 if (i != SCRUB_BIOS_PER_DEV-1)
246                         sdev->bios[i]->next_free = i + 1;
247                 else
248                         sdev->bios[i]->next_free = -1;
249         }
250         sdev->first_free = 0;
251         sdev->nodesize = dev->dev_root->nodesize;
252         sdev->leafsize = dev->dev_root->leafsize;
253         sdev->sectorsize = dev->dev_root->sectorsize;
254         atomic_set(&sdev->in_flight, 0);
255         atomic_set(&sdev->fixup_cnt, 0);
256         atomic_set(&sdev->cancel_req, 0);
257         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
258         INIT_LIST_HEAD(&sdev->csum_list);
259
260         spin_lock_init(&sdev->list_lock);
261         spin_lock_init(&sdev->stat_lock);
262         init_waitqueue_head(&sdev->list_wait);
263         return sdev;
264
265 nomem:
266         scrub_free_dev(sdev);
267         return ERR_PTR(-ENOMEM);
268 }
269
270 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
271 {
272         u64 isize;
273         u32 nlink;
274         int ret;
275         int i;
276         struct extent_buffer *eb;
277         struct btrfs_inode_item *inode_item;
278         struct scrub_warning *swarn = ctx;
279         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
280         struct inode_fs_paths *ipath = NULL;
281         struct btrfs_root *local_root;
282         struct btrfs_key root_key;
283
284         root_key.objectid = root;
285         root_key.type = BTRFS_ROOT_ITEM_KEY;
286         root_key.offset = (u64)-1;
287         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
288         if (IS_ERR(local_root)) {
289                 ret = PTR_ERR(local_root);
290                 goto err;
291         }
292
293         ret = inode_item_info(inum, 0, local_root, swarn->path);
294         if (ret) {
295                 btrfs_release_path(swarn->path);
296                 goto err;
297         }
298
299         eb = swarn->path->nodes[0];
300         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
301                                         struct btrfs_inode_item);
302         isize = btrfs_inode_size(eb, inode_item);
303         nlink = btrfs_inode_nlink(eb, inode_item);
304         btrfs_release_path(swarn->path);
305
306         ipath = init_ipath(4096, local_root, swarn->path);
307         if (IS_ERR(ipath)) {
308                 ret = PTR_ERR(ipath);
309                 ipath = NULL;
310                 goto err;
311         }
312         ret = paths_from_inode(inum, ipath);
313
314         if (ret < 0)
315                 goto err;
316
317         /*
318          * we deliberately ignore the bit ipath might have been too small to
319          * hold all of the paths here
320          */
321         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
322                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
323                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
324                         "length %llu, links %u (path: %s)\n", swarn->errstr,
325                         swarn->logical, swarn->dev->name,
326                         (unsigned long long)swarn->sector, root, inum, offset,
327                         min(isize - offset, (u64)PAGE_SIZE), nlink,
328                         (char *)(unsigned long)ipath->fspath->val[i]);
329
330         free_ipath(ipath);
331         return 0;
332
333 err:
334         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
335                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
336                 "resolving failed with ret=%d\n", swarn->errstr,
337                 swarn->logical, swarn->dev->name,
338                 (unsigned long long)swarn->sector, root, inum, offset, ret);
339
340         free_ipath(ipath);
341         return 0;
342 }
343
344 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
345 {
346         struct btrfs_device *dev = sblock->sdev->dev;
347         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
348         struct btrfs_path *path;
349         struct btrfs_key found_key;
350         struct extent_buffer *eb;
351         struct btrfs_extent_item *ei;
352         struct scrub_warning swarn;
353         u32 item_size;
354         int ret;
355         u64 ref_root;
356         u8 ref_level;
357         unsigned long ptr = 0;
358         const int bufsize = 4096;
359         u64 extent_item_pos;
360
361         path = btrfs_alloc_path();
362
363         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
364         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
365         BUG_ON(sblock->page_count < 1);
366         swarn.sector = (sblock->pagev[0].physical) >> 9;
367         swarn.logical = sblock->pagev[0].logical;
368         swarn.errstr = errstr;
369         swarn.dev = dev;
370         swarn.msg_bufsize = bufsize;
371         swarn.scratch_bufsize = bufsize;
372
373         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
374                 goto out;
375
376         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
377         if (ret < 0)
378                 goto out;
379
380         extent_item_pos = swarn.logical - found_key.objectid;
381         swarn.extent_item_size = found_key.offset;
382
383         eb = path->nodes[0];
384         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
385         item_size = btrfs_item_size_nr(eb, path->slots[0]);
386         btrfs_release_path(path);
387
388         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
389                 do {
390                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
391                                                         &ref_root, &ref_level);
392                         printk(KERN_WARNING
393                                 "btrfs: %s at logical %llu on dev %s, "
394                                 "sector %llu: metadata %s (level %d) in tree "
395                                 "%llu\n", errstr, swarn.logical, dev->name,
396                                 (unsigned long long)swarn.sector,
397                                 ref_level ? "node" : "leaf",
398                                 ret < 0 ? -1 : ref_level,
399                                 ret < 0 ? -1 : ref_root);
400                 } while (ret != 1);
401         } else {
402                 swarn.path = path;
403                 iterate_extent_inodes(fs_info, found_key.objectid,
404                                         extent_item_pos, 1,
405                                         scrub_print_warning_inode, &swarn);
406         }
407
408 out:
409         btrfs_free_path(path);
410         kfree(swarn.scratch_buf);
411         kfree(swarn.msg_buf);
412 }
413
414 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
415 {
416         struct page *page = NULL;
417         unsigned long index;
418         struct scrub_fixup_nodatasum *fixup = ctx;
419         int ret;
420         int corrected = 0;
421         struct btrfs_key key;
422         struct inode *inode = NULL;
423         u64 end = offset + PAGE_SIZE - 1;
424         struct btrfs_root *local_root;
425
426         key.objectid = root;
427         key.type = BTRFS_ROOT_ITEM_KEY;
428         key.offset = (u64)-1;
429         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
430         if (IS_ERR(local_root))
431                 return PTR_ERR(local_root);
432
433         key.type = BTRFS_INODE_ITEM_KEY;
434         key.objectid = inum;
435         key.offset = 0;
436         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
437         if (IS_ERR(inode))
438                 return PTR_ERR(inode);
439
440         index = offset >> PAGE_CACHE_SHIFT;
441
442         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
443         if (!page) {
444                 ret = -ENOMEM;
445                 goto out;
446         }
447
448         if (PageUptodate(page)) {
449                 struct btrfs_mapping_tree *map_tree;
450                 if (PageDirty(page)) {
451                         /*
452                          * we need to write the data to the defect sector. the
453                          * data that was in that sector is not in memory,
454                          * because the page was modified. we must not write the
455                          * modified page to that sector.
456                          *
457                          * TODO: what could be done here: wait for the delalloc
458                          *       runner to write out that page (might involve
459                          *       COW) and see whether the sector is still
460                          *       referenced afterwards.
461                          *
462                          * For the meantime, we'll treat this error
463                          * incorrectable, although there is a chance that a
464                          * later scrub will find the bad sector again and that
465                          * there's no dirty page in memory, then.
466                          */
467                         ret = -EIO;
468                         goto out;
469                 }
470                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
471                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
472                                         fixup->logical, page,
473                                         fixup->mirror_num);
474                 unlock_page(page);
475                 corrected = !ret;
476         } else {
477                 /*
478                  * we need to get good data first. the general readpage path
479                  * will call repair_io_failure for us, we just have to make
480                  * sure we read the bad mirror.
481                  */
482                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
483                                         EXTENT_DAMAGED, GFP_NOFS);
484                 if (ret) {
485                         /* set_extent_bits should give proper error */
486                         WARN_ON(ret > 0);
487                         if (ret > 0)
488                                 ret = -EFAULT;
489                         goto out;
490                 }
491
492                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
493                                                 btrfs_get_extent,
494                                                 fixup->mirror_num);
495                 wait_on_page_locked(page);
496
497                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
498                                                 end, EXTENT_DAMAGED, 0, NULL);
499                 if (!corrected)
500                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
501                                                 EXTENT_DAMAGED, GFP_NOFS);
502         }
503
504 out:
505         if (page)
506                 put_page(page);
507         if (inode)
508                 iput(inode);
509
510         if (ret < 0)
511                 return ret;
512
513         if (ret == 0 && corrected) {
514                 /*
515                  * we only need to call readpage for one of the inodes belonging
516                  * to this extent. so make iterate_extent_inodes stop
517                  */
518                 return 1;
519         }
520
521         return -EIO;
522 }
523
524 static void scrub_fixup_nodatasum(struct btrfs_work *work)
525 {
526         int ret;
527         struct scrub_fixup_nodatasum *fixup;
528         struct scrub_dev *sdev;
529         struct btrfs_trans_handle *trans = NULL;
530         struct btrfs_fs_info *fs_info;
531         struct btrfs_path *path;
532         int uncorrectable = 0;
533
534         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
535         sdev = fixup->sdev;
536         fs_info = fixup->root->fs_info;
537
538         path = btrfs_alloc_path();
539         if (!path) {
540                 spin_lock(&sdev->stat_lock);
541                 ++sdev->stat.malloc_errors;
542                 spin_unlock(&sdev->stat_lock);
543                 uncorrectable = 1;
544                 goto out;
545         }
546
547         trans = btrfs_join_transaction(fixup->root);
548         if (IS_ERR(trans)) {
549                 uncorrectable = 1;
550                 goto out;
551         }
552
553         /*
554          * the idea is to trigger a regular read through the standard path. we
555          * read a page from the (failed) logical address by specifying the
556          * corresponding copynum of the failed sector. thus, that readpage is
557          * expected to fail.
558          * that is the point where on-the-fly error correction will kick in
559          * (once it's finished) and rewrite the failed sector if a good copy
560          * can be found.
561          */
562         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
563                                                 path, scrub_fixup_readpage,
564                                                 fixup);
565         if (ret < 0) {
566                 uncorrectable = 1;
567                 goto out;
568         }
569         WARN_ON(ret != 1);
570
571         spin_lock(&sdev->stat_lock);
572         ++sdev->stat.corrected_errors;
573         spin_unlock(&sdev->stat_lock);
574
575 out:
576         if (trans && !IS_ERR(trans))
577                 btrfs_end_transaction(trans, fixup->root);
578         if (uncorrectable) {
579                 spin_lock(&sdev->stat_lock);
580                 ++sdev->stat.uncorrectable_errors;
581                 spin_unlock(&sdev->stat_lock);
582                 printk_ratelimited(KERN_ERR
583                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
584                         (unsigned long long)fixup->logical, sdev->dev->name);
585         }
586
587         btrfs_free_path(path);
588         kfree(fixup);
589
590         /* see caller why we're pretending to be paused in the scrub counters */
591         mutex_lock(&fs_info->scrub_lock);
592         atomic_dec(&fs_info->scrubs_running);
593         atomic_dec(&fs_info->scrubs_paused);
594         mutex_unlock(&fs_info->scrub_lock);
595         atomic_dec(&sdev->fixup_cnt);
596         wake_up(&fs_info->scrub_pause_wait);
597         wake_up(&sdev->list_wait);
598 }
599
600 /*
601  * scrub_handle_errored_block gets called when either verification of the
602  * pages failed or the bio failed to read, e.g. with EIO. In the latter
603  * case, this function handles all pages in the bio, even though only one
604  * may be bad.
605  * The goal of this function is to repair the errored block by using the
606  * contents of one of the mirrors.
607  */
608 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
609 {
610         struct scrub_dev *sdev = sblock_to_check->sdev;
611         struct btrfs_fs_info *fs_info;
612         u64 length;
613         u64 logical;
614         u64 generation;
615         unsigned int failed_mirror_index;
616         unsigned int is_metadata;
617         unsigned int have_csum;
618         u8 *csum;
619         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
620         struct scrub_block *sblock_bad;
621         int ret;
622         int mirror_index;
623         int page_num;
624         int success;
625         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
626                                       DEFAULT_RATELIMIT_BURST);
627
628         BUG_ON(sblock_to_check->page_count < 1);
629         fs_info = sdev->dev->dev_root->fs_info;
630         length = sblock_to_check->page_count * PAGE_SIZE;
631         logical = sblock_to_check->pagev[0].logical;
632         generation = sblock_to_check->pagev[0].generation;
633         BUG_ON(sblock_to_check->pagev[0].mirror_num < 1);
634         failed_mirror_index = sblock_to_check->pagev[0].mirror_num - 1;
635         is_metadata = !(sblock_to_check->pagev[0].flags &
636                         BTRFS_EXTENT_FLAG_DATA);
637         have_csum = sblock_to_check->pagev[0].have_csum;
638         csum = sblock_to_check->pagev[0].csum;
639
640         /*
641          * read all mirrors one after the other. This includes to
642          * re-read the extent or metadata block that failed (that was
643          * the cause that this fixup code is called) another time,
644          * page by page this time in order to know which pages
645          * caused I/O errors and which ones are good (for all mirrors).
646          * It is the goal to handle the situation when more than one
647          * mirror contains I/O errors, but the errors do not
648          * overlap, i.e. the data can be repaired by selecting the
649          * pages from those mirrors without I/O error on the
650          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
651          * would be that mirror #1 has an I/O error on the first page,
652          * the second page is good, and mirror #2 has an I/O error on
653          * the second page, but the first page is good.
654          * Then the first page of the first mirror can be repaired by
655          * taking the first page of the second mirror, and the
656          * second page of the second mirror can be repaired by
657          * copying the contents of the 2nd page of the 1st mirror.
658          * One more note: if the pages of one mirror contain I/O
659          * errors, the checksum cannot be verified. In order to get
660          * the best data for repairing, the first attempt is to find
661          * a mirror without I/O errors and with a validated checksum.
662          * Only if this is not possible, the pages are picked from
663          * mirrors with I/O errors without considering the checksum.
664          * If the latter is the case, at the end, the checksum of the
665          * repaired area is verified in order to correctly maintain
666          * the statistics.
667          */
668
669         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
670                                      sizeof(*sblocks_for_recheck),
671                                      GFP_NOFS);
672         if (!sblocks_for_recheck) {
673                 spin_lock(&sdev->stat_lock);
674                 sdev->stat.malloc_errors++;
675                 sdev->stat.read_errors++;
676                 sdev->stat.uncorrectable_errors++;
677                 spin_unlock(&sdev->stat_lock);
678                 goto out;
679         }
680
681         /* setup the context, map the logical blocks and alloc the pages */
682         ret = scrub_setup_recheck_block(sdev, &fs_info->mapping_tree, length,
683                                         logical, sblocks_for_recheck);
684         if (ret) {
685                 spin_lock(&sdev->stat_lock);
686                 sdev->stat.read_errors++;
687                 sdev->stat.uncorrectable_errors++;
688                 spin_unlock(&sdev->stat_lock);
689                 goto out;
690         }
691         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
692         sblock_bad = sblocks_for_recheck + failed_mirror_index;
693
694         /* build and submit the bios for the failed mirror, check checksums */
695         ret = scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
696                                   csum, generation, sdev->csum_size);
697         if (ret) {
698                 spin_lock(&sdev->stat_lock);
699                 sdev->stat.read_errors++;
700                 sdev->stat.uncorrectable_errors++;
701                 spin_unlock(&sdev->stat_lock);
702                 goto out;
703         }
704
705         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
706             sblock_bad->no_io_error_seen) {
707                 /*
708                  * the error disappeared after reading page by page, or
709                  * the area was part of a huge bio and other parts of the
710                  * bio caused I/O errors, or the block layer merged several
711                  * read requests into one and the error is caused by a
712                  * different bio (usually one of the two latter cases is
713                  * the cause)
714                  */
715                 spin_lock(&sdev->stat_lock);
716                 sdev->stat.unverified_errors++;
717                 spin_unlock(&sdev->stat_lock);
718
719                 goto out;
720         }
721
722         if (!sblock_bad->no_io_error_seen) {
723                 spin_lock(&sdev->stat_lock);
724                 sdev->stat.read_errors++;
725                 spin_unlock(&sdev->stat_lock);
726                 if (__ratelimit(&_rs))
727                         scrub_print_warning("i/o error", sblock_to_check);
728         } else if (sblock_bad->checksum_error) {
729                 spin_lock(&sdev->stat_lock);
730                 sdev->stat.csum_errors++;
731                 spin_unlock(&sdev->stat_lock);
732                 if (__ratelimit(&_rs))
733                         scrub_print_warning("checksum error", sblock_to_check);
734         } else if (sblock_bad->header_error) {
735                 spin_lock(&sdev->stat_lock);
736                 sdev->stat.verify_errors++;
737                 spin_unlock(&sdev->stat_lock);
738                 if (__ratelimit(&_rs))
739                         scrub_print_warning("checksum/header error",
740                                             sblock_to_check);
741         }
742
743         if (sdev->readonly)
744                 goto did_not_correct_error;
745
746         if (!is_metadata && !have_csum) {
747                 struct scrub_fixup_nodatasum *fixup_nodatasum;
748
749                 /*
750                  * !is_metadata and !have_csum, this means that the data
751                  * might not be COW'ed, that it might be modified
752                  * concurrently. The general strategy to work on the
753                  * commit root does not help in the case when COW is not
754                  * used.
755                  */
756                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
757                 if (!fixup_nodatasum)
758                         goto did_not_correct_error;
759                 fixup_nodatasum->sdev = sdev;
760                 fixup_nodatasum->logical = logical;
761                 fixup_nodatasum->root = fs_info->extent_root;
762                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
763                 /*
764                  * increment scrubs_running to prevent cancel requests from
765                  * completing as long as a fixup worker is running. we must also
766                  * increment scrubs_paused to prevent deadlocking on pause
767                  * requests used for transactions commits (as the worker uses a
768                  * transaction context). it is safe to regard the fixup worker
769                  * as paused for all matters practical. effectively, we only
770                  * avoid cancellation requests from completing.
771                  */
772                 mutex_lock(&fs_info->scrub_lock);
773                 atomic_inc(&fs_info->scrubs_running);
774                 atomic_inc(&fs_info->scrubs_paused);
775                 mutex_unlock(&fs_info->scrub_lock);
776                 atomic_inc(&sdev->fixup_cnt);
777                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
778                 btrfs_queue_worker(&fs_info->scrub_workers,
779                                    &fixup_nodatasum->work);
780                 goto out;
781         }
782
783         /*
784          * now build and submit the bios for the other mirrors, check
785          * checksums
786          */
787         for (mirror_index = 0;
788              mirror_index < BTRFS_MAX_MIRRORS &&
789              sblocks_for_recheck[mirror_index].page_count > 0;
790              mirror_index++) {
791                 if (mirror_index == failed_mirror_index)
792                         continue;
793
794                 /* build and submit the bios, check checksums */
795                 ret = scrub_recheck_block(fs_info,
796                                           sblocks_for_recheck + mirror_index,
797                                           is_metadata, have_csum, csum,
798                                           generation, sdev->csum_size);
799                 if (ret)
800                         goto did_not_correct_error;
801         }
802
803         /*
804          * first try to pick the mirror which is completely without I/O
805          * errors and also does not have a checksum error.
806          * If one is found, and if a checksum is present, the full block
807          * that is known to contain an error is rewritten. Afterwards
808          * the block is known to be corrected.
809          * If a mirror is found which is completely correct, and no
810          * checksum is present, only those pages are rewritten that had
811          * an I/O error in the block to be repaired, since it cannot be
812          * determined, which copy of the other pages is better (and it
813          * could happen otherwise that a correct page would be
814          * overwritten by a bad one).
815          */
816         for (mirror_index = 0;
817              mirror_index < BTRFS_MAX_MIRRORS &&
818              sblocks_for_recheck[mirror_index].page_count > 0;
819              mirror_index++) {
820                 struct scrub_block *sblock_other = sblocks_for_recheck +
821                                                    mirror_index;
822
823                 if (!sblock_other->header_error &&
824                     !sblock_other->checksum_error &&
825                     sblock_other->no_io_error_seen) {
826                         int force_write = is_metadata || have_csum;
827
828                         ret = scrub_repair_block_from_good_copy(sblock_bad,
829                                                                 sblock_other,
830                                                                 force_write);
831                         if (0 == ret)
832                                 goto corrected_error;
833                 }
834         }
835
836         /*
837          * in case of I/O errors in the area that is supposed to be
838          * repaired, continue by picking good copies of those pages.
839          * Select the good pages from mirrors to rewrite bad pages from
840          * the area to fix. Afterwards verify the checksum of the block
841          * that is supposed to be repaired. This verification step is
842          * only done for the purpose of statistic counting and for the
843          * final scrub report, whether errors remain.
844          * A perfect algorithm could make use of the checksum and try
845          * all possible combinations of pages from the different mirrors
846          * until the checksum verification succeeds. For example, when
847          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
848          * of mirror #2 is readable but the final checksum test fails,
849          * then the 2nd page of mirror #3 could be tried, whether now
850          * the final checksum succeedes. But this would be a rare
851          * exception and is therefore not implemented. At least it is
852          * avoided that the good copy is overwritten.
853          * A more useful improvement would be to pick the sectors
854          * without I/O error based on sector sizes (512 bytes on legacy
855          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
856          * mirror could be repaired by taking 512 byte of a different
857          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
858          * area are unreadable.
859          */
860
861         /* can only fix I/O errors from here on */
862         if (sblock_bad->no_io_error_seen)
863                 goto did_not_correct_error;
864
865         success = 1;
866         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
867                 struct scrub_page *page_bad = sblock_bad->pagev + page_num;
868
869                 if (!page_bad->io_error)
870                         continue;
871
872                 for (mirror_index = 0;
873                      mirror_index < BTRFS_MAX_MIRRORS &&
874                      sblocks_for_recheck[mirror_index].page_count > 0;
875                      mirror_index++) {
876                         struct scrub_block *sblock_other = sblocks_for_recheck +
877                                                            mirror_index;
878                         struct scrub_page *page_other = sblock_other->pagev +
879                                                         page_num;
880
881                         if (!page_other->io_error) {
882                                 ret = scrub_repair_page_from_good_copy(
883                                         sblock_bad, sblock_other, page_num, 0);
884                                 if (0 == ret) {
885                                         page_bad->io_error = 0;
886                                         break; /* succeeded for this page */
887                                 }
888                         }
889                 }
890
891                 if (page_bad->io_error) {
892                         /* did not find a mirror to copy the page from */
893                         success = 0;
894                 }
895         }
896
897         if (success) {
898                 if (is_metadata || have_csum) {
899                         /*
900                          * need to verify the checksum now that all
901                          * sectors on disk are repaired (the write
902                          * request for data to be repaired is on its way).
903                          * Just be lazy and use scrub_recheck_block()
904                          * which re-reads the data before the checksum
905                          * is verified, but most likely the data comes out
906                          * of the page cache.
907                          */
908                         ret = scrub_recheck_block(fs_info, sblock_bad,
909                                                   is_metadata, have_csum, csum,
910                                                   generation, sdev->csum_size);
911                         if (!ret && !sblock_bad->header_error &&
912                             !sblock_bad->checksum_error &&
913                             sblock_bad->no_io_error_seen)
914                                 goto corrected_error;
915                         else
916                                 goto did_not_correct_error;
917                 } else {
918 corrected_error:
919                         spin_lock(&sdev->stat_lock);
920                         sdev->stat.corrected_errors++;
921                         spin_unlock(&sdev->stat_lock);
922                         printk_ratelimited(KERN_ERR
923                                 "btrfs: fixed up error at logical %llu on dev %s\n",
924                                 (unsigned long long)logical, sdev->dev->name);
925                 }
926         } else {
927 did_not_correct_error:
928                 spin_lock(&sdev->stat_lock);
929                 sdev->stat.uncorrectable_errors++;
930                 spin_unlock(&sdev->stat_lock);
931                 printk_ratelimited(KERN_ERR
932                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
933                         (unsigned long long)logical, sdev->dev->name);
934         }
935
936 out:
937         if (sblocks_for_recheck) {
938                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
939                      mirror_index++) {
940                         struct scrub_block *sblock = sblocks_for_recheck +
941                                                      mirror_index;
942                         int page_index;
943
944                         for (page_index = 0; page_index < SCRUB_PAGES_PER_BIO;
945                              page_index++)
946                                 if (sblock->pagev[page_index].page)
947                                         __free_page(
948                                                 sblock->pagev[page_index].page);
949                 }
950                 kfree(sblocks_for_recheck);
951         }
952
953         return 0;
954 }
955
956 static int scrub_setup_recheck_block(struct scrub_dev *sdev,
957                                      struct btrfs_mapping_tree *map_tree,
958                                      u64 length, u64 logical,
959                                      struct scrub_block *sblocks_for_recheck)
960 {
961         int page_index;
962         int mirror_index;
963         int ret;
964
965         /*
966          * note: the three members sdev, ref_count and outstanding_pages
967          * are not used (and not set) in the blocks that are used for
968          * the recheck procedure
969          */
970
971         page_index = 0;
972         while (length > 0) {
973                 u64 sublen = min_t(u64, length, PAGE_SIZE);
974                 u64 mapped_length = sublen;
975                 struct btrfs_bio *bbio = NULL;
976
977                 /*
978                  * with a length of PAGE_SIZE, each returned stripe
979                  * represents one mirror
980                  */
981                 ret = btrfs_map_block(map_tree, WRITE, logical, &mapped_length,
982                                       &bbio, 0);
983                 if (ret || !bbio || mapped_length < sublen) {
984                         kfree(bbio);
985                         return -EIO;
986                 }
987
988                 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO);
989                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
990                      mirror_index++) {
991                         struct scrub_block *sblock;
992                         struct scrub_page *page;
993
994                         if (mirror_index >= BTRFS_MAX_MIRRORS)
995                                 continue;
996
997                         sblock = sblocks_for_recheck + mirror_index;
998                         page = sblock->pagev + page_index;
999                         page->logical = logical;
1000                         page->physical = bbio->stripes[mirror_index].physical;
1001                         page->bdev = bbio->stripes[mirror_index].dev->bdev;
1002                         page->mirror_num = mirror_index + 1;
1003                         page->page = alloc_page(GFP_NOFS);
1004                         if (!page->page) {
1005                                 spin_lock(&sdev->stat_lock);
1006                                 sdev->stat.malloc_errors++;
1007                                 spin_unlock(&sdev->stat_lock);
1008                                 return -ENOMEM;
1009                         }
1010                         sblock->page_count++;
1011                 }
1012                 kfree(bbio);
1013                 length -= sublen;
1014                 logical += sublen;
1015                 page_index++;
1016         }
1017
1018         return 0;
1019 }
1020
1021 /*
1022  * this function will check the on disk data for checksum errors, header
1023  * errors and read I/O errors. If any I/O errors happen, the exact pages
1024  * which are errored are marked as being bad. The goal is to enable scrub
1025  * to take those pages that are not errored from all the mirrors so that
1026  * the pages that are errored in the just handled mirror can be repaired.
1027  */
1028 static int scrub_recheck_block(struct btrfs_fs_info *fs_info,
1029                                struct scrub_block *sblock, int is_metadata,
1030                                int have_csum, u8 *csum, u64 generation,
1031                                u16 csum_size)
1032 {
1033         int page_num;
1034
1035         sblock->no_io_error_seen = 1;
1036         sblock->header_error = 0;
1037         sblock->checksum_error = 0;
1038
1039         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1040                 struct bio *bio;
1041                 int ret;
1042                 struct scrub_page *page = sblock->pagev + page_num;
1043                 DECLARE_COMPLETION_ONSTACK(complete);
1044
1045                 BUG_ON(!page->page);
1046                 bio = bio_alloc(GFP_NOFS, 1);
1047                 bio->bi_bdev = page->bdev;
1048                 bio->bi_sector = page->physical >> 9;
1049                 bio->bi_end_io = scrub_complete_bio_end_io;
1050                 bio->bi_private = &complete;
1051
1052                 ret = bio_add_page(bio, page->page, PAGE_SIZE, 0);
1053                 if (PAGE_SIZE != ret) {
1054                         bio_put(bio);
1055                         return -EIO;
1056                 }
1057                 btrfsic_submit_bio(READ, bio);
1058
1059                 /* this will also unplug the queue */
1060                 wait_for_completion(&complete);
1061
1062                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1063                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1064                         sblock->no_io_error_seen = 0;
1065                 bio_put(bio);
1066         }
1067
1068         if (sblock->no_io_error_seen)
1069                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1070                                              have_csum, csum, generation,
1071                                              csum_size);
1072
1073         return 0;
1074 }
1075
1076 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1077                                          struct scrub_block *sblock,
1078                                          int is_metadata, int have_csum,
1079                                          const u8 *csum, u64 generation,
1080                                          u16 csum_size)
1081 {
1082         int page_num;
1083         u8 calculated_csum[BTRFS_CSUM_SIZE];
1084         u32 crc = ~(u32)0;
1085         struct btrfs_root *root = fs_info->extent_root;
1086         void *mapped_buffer;
1087
1088         BUG_ON(!sblock->pagev[0].page);
1089         if (is_metadata) {
1090                 struct btrfs_header *h;
1091
1092                 mapped_buffer = kmap_atomic(sblock->pagev[0].page, KM_USER0);
1093                 h = (struct btrfs_header *)mapped_buffer;
1094
1095                 if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr) ||
1096                     generation != le64_to_cpu(h->generation) ||
1097                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1098                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1099                            BTRFS_UUID_SIZE))
1100                         sblock->header_error = 1;
1101                 csum = h->csum;
1102         } else {
1103                 if (!have_csum)
1104                         return;
1105
1106                 mapped_buffer = kmap_atomic(sblock->pagev[0].page, KM_USER0);
1107         }
1108
1109         for (page_num = 0;;) {
1110                 if (page_num == 0 && is_metadata)
1111                         crc = btrfs_csum_data(root,
1112                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1113                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1114                 else
1115                         crc = btrfs_csum_data(root, mapped_buffer, crc,
1116                                               PAGE_SIZE);
1117
1118                 kunmap_atomic(mapped_buffer, KM_USER0);
1119                 page_num++;
1120                 if (page_num >= sblock->page_count)
1121                         break;
1122                 BUG_ON(!sblock->pagev[page_num].page);
1123
1124                 mapped_buffer = kmap_atomic(sblock->pagev[page_num].page,
1125                                             KM_USER0);
1126         }
1127
1128         btrfs_csum_final(crc, calculated_csum);
1129         if (memcmp(calculated_csum, csum, csum_size))
1130                 sblock->checksum_error = 1;
1131 }
1132
1133 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1134 {
1135         complete((struct completion *)bio->bi_private);
1136 }
1137
1138 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1139                                              struct scrub_block *sblock_good,
1140                                              int force_write)
1141 {
1142         int page_num;
1143         int ret = 0;
1144
1145         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1146                 int ret_sub;
1147
1148                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1149                                                            sblock_good,
1150                                                            page_num,
1151                                                            force_write);
1152                 if (ret_sub)
1153                         ret = ret_sub;
1154         }
1155
1156         return ret;
1157 }
1158
1159 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1160                                             struct scrub_block *sblock_good,
1161                                             int page_num, int force_write)
1162 {
1163         struct scrub_page *page_bad = sblock_bad->pagev + page_num;
1164         struct scrub_page *page_good = sblock_good->pagev + page_num;
1165
1166         BUG_ON(sblock_bad->pagev[page_num].page == NULL);
1167         BUG_ON(sblock_good->pagev[page_num].page == NULL);
1168         if (force_write || sblock_bad->header_error ||
1169             sblock_bad->checksum_error || page_bad->io_error) {
1170                 struct bio *bio;
1171                 int ret;
1172                 DECLARE_COMPLETION_ONSTACK(complete);
1173
1174                 bio = bio_alloc(GFP_NOFS, 1);
1175                 bio->bi_bdev = page_bad->bdev;
1176                 bio->bi_sector = page_bad->physical >> 9;
1177                 bio->bi_end_io = scrub_complete_bio_end_io;
1178                 bio->bi_private = &complete;
1179
1180                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1181                 if (PAGE_SIZE != ret) {
1182                         bio_put(bio);
1183                         return -EIO;
1184                 }
1185                 btrfsic_submit_bio(WRITE, bio);
1186
1187                 /* this will also unplug the queue */
1188                 wait_for_completion(&complete);
1189                 bio_put(bio);
1190         }
1191
1192         return 0;
1193 }
1194
1195 static void scrub_checksum(struct scrub_block *sblock)
1196 {
1197         u64 flags;
1198         int ret;
1199
1200         BUG_ON(sblock->page_count < 1);
1201         flags = sblock->pagev[0].flags;
1202         ret = 0;
1203         if (flags & BTRFS_EXTENT_FLAG_DATA)
1204                 ret = scrub_checksum_data(sblock);
1205         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1206                 ret = scrub_checksum_tree_block(sblock);
1207         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1208                 (void)scrub_checksum_super(sblock);
1209         else
1210                 WARN_ON(1);
1211         if (ret)
1212                 scrub_handle_errored_block(sblock);
1213 }
1214
1215 static int scrub_checksum_data(struct scrub_block *sblock)
1216 {
1217         struct scrub_dev *sdev = sblock->sdev;
1218         u8 csum[BTRFS_CSUM_SIZE];
1219         u8 *on_disk_csum;
1220         struct page *page;
1221         void *buffer;
1222         u32 crc = ~(u32)0;
1223         int fail = 0;
1224         struct btrfs_root *root = sdev->dev->dev_root;
1225         u64 len;
1226         int index;
1227
1228         BUG_ON(sblock->page_count < 1);
1229         if (!sblock->pagev[0].have_csum)
1230                 return 0;
1231
1232         on_disk_csum = sblock->pagev[0].csum;
1233         page = sblock->pagev[0].page;
1234         buffer = kmap_atomic(page, KM_USER0);
1235
1236         len = sdev->sectorsize;
1237         index = 0;
1238         for (;;) {
1239                 u64 l = min_t(u64, len, PAGE_SIZE);
1240
1241                 crc = btrfs_csum_data(root, buffer, crc, l);
1242                 kunmap_atomic(buffer, KM_USER0);
1243                 len -= l;
1244                 if (len == 0)
1245                         break;
1246                 index++;
1247                 BUG_ON(index >= sblock->page_count);
1248                 BUG_ON(!sblock->pagev[index].page);
1249                 page = sblock->pagev[index].page;
1250                 buffer = kmap_atomic(page, KM_USER0);
1251         }
1252
1253         btrfs_csum_final(crc, csum);
1254         if (memcmp(csum, on_disk_csum, sdev->csum_size))
1255                 fail = 1;
1256
1257         if (fail) {
1258                 spin_lock(&sdev->stat_lock);
1259                 ++sdev->stat.csum_errors;
1260                 spin_unlock(&sdev->stat_lock);
1261         }
1262
1263         return fail;
1264 }
1265
1266 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1267 {
1268         struct scrub_dev *sdev = sblock->sdev;
1269         struct btrfs_header *h;
1270         struct btrfs_root *root = sdev->dev->dev_root;
1271         struct btrfs_fs_info *fs_info = root->fs_info;
1272         u8 calculated_csum[BTRFS_CSUM_SIZE];
1273         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1274         struct page *page;
1275         void *mapped_buffer;
1276         u64 mapped_size;
1277         void *p;
1278         u32 crc = ~(u32)0;
1279         int fail = 0;
1280         int crc_fail = 0;
1281         u64 len;
1282         int index;
1283
1284         BUG_ON(sblock->page_count < 1);
1285         page = sblock->pagev[0].page;
1286         mapped_buffer = kmap_atomic(page, KM_USER0);
1287         h = (struct btrfs_header *)mapped_buffer;
1288         memcpy(on_disk_csum, h->csum, sdev->csum_size);
1289
1290         /*
1291          * we don't use the getter functions here, as we
1292          * a) don't have an extent buffer and
1293          * b) the page is already kmapped
1294          */
1295
1296         if (sblock->pagev[0].logical != le64_to_cpu(h->bytenr))
1297                 ++fail;
1298
1299         if (sblock->pagev[0].generation != le64_to_cpu(h->generation))
1300                 ++fail;
1301
1302         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1303                 ++fail;
1304
1305         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1306                    BTRFS_UUID_SIZE))
1307                 ++fail;
1308
1309         BUG_ON(sdev->nodesize != sdev->leafsize);
1310         len = sdev->nodesize - BTRFS_CSUM_SIZE;
1311         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1312         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1313         index = 0;
1314         for (;;) {
1315                 u64 l = min_t(u64, len, mapped_size);
1316
1317                 crc = btrfs_csum_data(root, p, crc, l);
1318                 kunmap_atomic(mapped_buffer, KM_USER0);
1319                 len -= l;
1320                 if (len == 0)
1321                         break;
1322                 index++;
1323                 BUG_ON(index >= sblock->page_count);
1324                 BUG_ON(!sblock->pagev[index].page);
1325                 page = sblock->pagev[index].page;
1326                 mapped_buffer = kmap_atomic(page, KM_USER0);
1327                 mapped_size = PAGE_SIZE;
1328                 p = mapped_buffer;
1329         }
1330
1331         btrfs_csum_final(crc, calculated_csum);
1332         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1333                 ++crc_fail;
1334
1335         if (crc_fail || fail) {
1336                 spin_lock(&sdev->stat_lock);
1337                 if (crc_fail)
1338                         ++sdev->stat.csum_errors;
1339                 if (fail)
1340                         ++sdev->stat.verify_errors;
1341                 spin_unlock(&sdev->stat_lock);
1342         }
1343
1344         return fail || crc_fail;
1345 }
1346
1347 static int scrub_checksum_super(struct scrub_block *sblock)
1348 {
1349         struct btrfs_super_block *s;
1350         struct scrub_dev *sdev = sblock->sdev;
1351         struct btrfs_root *root = sdev->dev->dev_root;
1352         struct btrfs_fs_info *fs_info = root->fs_info;
1353         u8 calculated_csum[BTRFS_CSUM_SIZE];
1354         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1355         struct page *page;
1356         void *mapped_buffer;
1357         u64 mapped_size;
1358         void *p;
1359         u32 crc = ~(u32)0;
1360         int fail = 0;
1361         u64 len;
1362         int index;
1363
1364         BUG_ON(sblock->page_count < 1);
1365         page = sblock->pagev[0].page;
1366         mapped_buffer = kmap_atomic(page, KM_USER0);
1367         s = (struct btrfs_super_block *)mapped_buffer;
1368         memcpy(on_disk_csum, s->csum, sdev->csum_size);
1369
1370         if (sblock->pagev[0].logical != le64_to_cpu(s->bytenr))
1371                 ++fail;
1372
1373         if (sblock->pagev[0].generation != le64_to_cpu(s->generation))
1374                 ++fail;
1375
1376         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1377                 ++fail;
1378
1379         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1380         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1381         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1382         index = 0;
1383         for (;;) {
1384                 u64 l = min_t(u64, len, mapped_size);
1385
1386                 crc = btrfs_csum_data(root, p, crc, l);
1387                 kunmap_atomic(mapped_buffer, KM_USER0);
1388                 len -= l;
1389                 if (len == 0)
1390                         break;
1391                 index++;
1392                 BUG_ON(index >= sblock->page_count);
1393                 BUG_ON(!sblock->pagev[index].page);
1394                 page = sblock->pagev[index].page;
1395                 mapped_buffer = kmap_atomic(page, KM_USER0);
1396                 mapped_size = PAGE_SIZE;
1397                 p = mapped_buffer;
1398         }
1399
1400         btrfs_csum_final(crc, calculated_csum);
1401         if (memcmp(calculated_csum, on_disk_csum, sdev->csum_size))
1402                 ++fail;
1403
1404         if (fail) {
1405                 /*
1406                  * if we find an error in a super block, we just report it.
1407                  * They will get written with the next transaction commit
1408                  * anyway
1409                  */
1410                 spin_lock(&sdev->stat_lock);
1411                 ++sdev->stat.super_errors;
1412                 spin_unlock(&sdev->stat_lock);
1413         }
1414
1415         return fail;
1416 }
1417
1418 static void scrub_block_get(struct scrub_block *sblock)
1419 {
1420         atomic_inc(&sblock->ref_count);
1421 }
1422
1423 static void scrub_block_put(struct scrub_block *sblock)
1424 {
1425         if (atomic_dec_and_test(&sblock->ref_count)) {
1426                 int i;
1427
1428                 for (i = 0; i < sblock->page_count; i++)
1429                         if (sblock->pagev[i].page)
1430                                 __free_page(sblock->pagev[i].page);
1431                 kfree(sblock);
1432         }
1433 }
1434
1435 static void scrub_submit(struct scrub_dev *sdev)
1436 {
1437         struct scrub_bio *sbio;
1438
1439         if (sdev->curr == -1)
1440                 return;
1441
1442         sbio = sdev->bios[sdev->curr];
1443         sdev->curr = -1;
1444         atomic_inc(&sdev->in_flight);
1445
1446         btrfsic_submit_bio(READ, sbio->bio);
1447 }
1448
1449 static int scrub_add_page_to_bio(struct scrub_dev *sdev,
1450                                  struct scrub_page *spage)
1451 {
1452         struct scrub_block *sblock = spage->sblock;
1453         struct scrub_bio *sbio;
1454         int ret;
1455
1456 again:
1457         /*
1458          * grab a fresh bio or wait for one to become available
1459          */
1460         while (sdev->curr == -1) {
1461                 spin_lock(&sdev->list_lock);
1462                 sdev->curr = sdev->first_free;
1463                 if (sdev->curr != -1) {
1464                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
1465                         sdev->bios[sdev->curr]->next_free = -1;
1466                         sdev->bios[sdev->curr]->page_count = 0;
1467                         spin_unlock(&sdev->list_lock);
1468                 } else {
1469                         spin_unlock(&sdev->list_lock);
1470                         wait_event(sdev->list_wait, sdev->first_free != -1);
1471                 }
1472         }
1473         sbio = sdev->bios[sdev->curr];
1474         if (sbio->page_count == 0) {
1475                 struct bio *bio;
1476
1477                 sbio->physical = spage->physical;
1478                 sbio->logical = spage->logical;
1479                 bio = sbio->bio;
1480                 if (!bio) {
1481                         bio = bio_alloc(GFP_NOFS, sdev->pages_per_bio);
1482                         if (!bio)
1483                                 return -ENOMEM;
1484                         sbio->bio = bio;
1485                 }
1486
1487                 bio->bi_private = sbio;
1488                 bio->bi_end_io = scrub_bio_end_io;
1489                 bio->bi_bdev = sdev->dev->bdev;
1490                 bio->bi_sector = spage->physical >> 9;
1491                 sbio->err = 0;
1492         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1493                    spage->physical ||
1494                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1495                    spage->logical) {
1496                 scrub_submit(sdev);
1497                 goto again;
1498         }
1499
1500         sbio->pagev[sbio->page_count] = spage;
1501         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1502         if (ret != PAGE_SIZE) {
1503                 if (sbio->page_count < 1) {
1504                         bio_put(sbio->bio);
1505                         sbio->bio = NULL;
1506                         return -EIO;
1507                 }
1508                 scrub_submit(sdev);
1509                 goto again;
1510         }
1511
1512         scrub_block_get(sblock); /* one for the added page */
1513         atomic_inc(&sblock->outstanding_pages);
1514         sbio->page_count++;
1515         if (sbio->page_count == sdev->pages_per_bio)
1516                 scrub_submit(sdev);
1517
1518         return 0;
1519 }
1520
1521 static int scrub_pages(struct scrub_dev *sdev, u64 logical, u64 len,
1522                        u64 physical, u64 flags, u64 gen, int mirror_num,
1523                        u8 *csum, int force)
1524 {
1525         struct scrub_block *sblock;
1526         int index;
1527
1528         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1529         if (!sblock) {
1530                 spin_lock(&sdev->stat_lock);
1531                 sdev->stat.malloc_errors++;
1532                 spin_unlock(&sdev->stat_lock);
1533                 return -ENOMEM;
1534         }
1535
1536         /* one ref inside this function, plus one for each page later on */
1537         atomic_set(&sblock->ref_count, 1);
1538         sblock->sdev = sdev;
1539         sblock->no_io_error_seen = 1;
1540
1541         for (index = 0; len > 0; index++) {
1542                 struct scrub_page *spage = sblock->pagev + index;
1543                 u64 l = min_t(u64, len, PAGE_SIZE);
1544
1545                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1546                 spage->page = alloc_page(GFP_NOFS);
1547                 if (!spage->page) {
1548                         spin_lock(&sdev->stat_lock);
1549                         sdev->stat.malloc_errors++;
1550                         spin_unlock(&sdev->stat_lock);
1551                         while (index > 0) {
1552                                 index--;
1553                                 __free_page(sblock->pagev[index].page);
1554                         }
1555                         kfree(sblock);
1556                         return -ENOMEM;
1557                 }
1558                 spage->sblock = sblock;
1559                 spage->bdev = sdev->dev->bdev;
1560                 spage->flags = flags;
1561                 spage->generation = gen;
1562                 spage->logical = logical;
1563                 spage->physical = physical;
1564                 spage->mirror_num = mirror_num;
1565                 if (csum) {
1566                         spage->have_csum = 1;
1567                         memcpy(spage->csum, csum, sdev->csum_size);
1568                 } else {
1569                         spage->have_csum = 0;
1570                 }
1571                 sblock->page_count++;
1572                 len -= l;
1573                 logical += l;
1574                 physical += l;
1575         }
1576
1577         BUG_ON(sblock->page_count == 0);
1578         for (index = 0; index < sblock->page_count; index++) {
1579                 struct scrub_page *spage = sblock->pagev + index;
1580                 int ret;
1581
1582                 ret = scrub_add_page_to_bio(sdev, spage);
1583                 if (ret) {
1584                         scrub_block_put(sblock);
1585                         return ret;
1586                 }
1587         }
1588
1589         if (force)
1590                 scrub_submit(sdev);
1591
1592         /* last one frees, either here or in bio completion for last page */
1593         scrub_block_put(sblock);
1594         return 0;
1595 }
1596
1597 static void scrub_bio_end_io(struct bio *bio, int err)
1598 {
1599         struct scrub_bio *sbio = bio->bi_private;
1600         struct scrub_dev *sdev = sbio->sdev;
1601         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1602
1603         sbio->err = err;
1604         sbio->bio = bio;
1605
1606         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
1607 }
1608
1609 static void scrub_bio_end_io_worker(struct btrfs_work *work)
1610 {
1611         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1612         struct scrub_dev *sdev = sbio->sdev;
1613         int i;
1614
1615         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO);
1616         if (sbio->err) {
1617                 for (i = 0; i < sbio->page_count; i++) {
1618                         struct scrub_page *spage = sbio->pagev[i];
1619
1620                         spage->io_error = 1;
1621                         spage->sblock->no_io_error_seen = 0;
1622                 }
1623         }
1624
1625         /* now complete the scrub_block items that have all pages completed */
1626         for (i = 0; i < sbio->page_count; i++) {
1627                 struct scrub_page *spage = sbio->pagev[i];
1628                 struct scrub_block *sblock = spage->sblock;
1629
1630                 if (atomic_dec_and_test(&sblock->outstanding_pages))
1631                         scrub_block_complete(sblock);
1632                 scrub_block_put(sblock);
1633         }
1634
1635         if (sbio->err) {
1636                 /* what is this good for??? */
1637                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1638                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
1639                 sbio->bio->bi_phys_segments = 0;
1640                 sbio->bio->bi_idx = 0;
1641
1642                 for (i = 0; i < sbio->page_count; i++) {
1643                         struct bio_vec *bi;
1644                         bi = &sbio->bio->bi_io_vec[i];
1645                         bi->bv_offset = 0;
1646                         bi->bv_len = PAGE_SIZE;
1647                 }
1648         }
1649
1650         bio_put(sbio->bio);
1651         sbio->bio = NULL;
1652         spin_lock(&sdev->list_lock);
1653         sbio->next_free = sdev->first_free;
1654         sdev->first_free = sbio->index;
1655         spin_unlock(&sdev->list_lock);
1656         atomic_dec(&sdev->in_flight);
1657         wake_up(&sdev->list_wait);
1658 }
1659
1660 static void scrub_block_complete(struct scrub_block *sblock)
1661 {
1662         if (!sblock->no_io_error_seen)
1663                 scrub_handle_errored_block(sblock);
1664         else
1665                 scrub_checksum(sblock);
1666 }
1667
1668 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1669                            u8 *csum)
1670 {
1671         struct btrfs_ordered_sum *sum = NULL;
1672         int ret = 0;
1673         unsigned long i;
1674         unsigned long num_sectors;
1675
1676         while (!list_empty(&sdev->csum_list)) {
1677                 sum = list_first_entry(&sdev->csum_list,
1678                                        struct btrfs_ordered_sum, list);
1679                 if (sum->bytenr > logical)
1680                         return 0;
1681                 if (sum->bytenr + sum->len > logical)
1682                         break;
1683
1684                 ++sdev->stat.csum_discards;
1685                 list_del(&sum->list);
1686                 kfree(sum);
1687                 sum = NULL;
1688         }
1689         if (!sum)
1690                 return 0;
1691
1692         num_sectors = sum->len / sdev->sectorsize;
1693         for (i = 0; i < num_sectors; ++i) {
1694                 if (sum->sums[i].bytenr == logical) {
1695                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1696                         ret = 1;
1697                         break;
1698                 }
1699         }
1700         if (ret && i == num_sectors - 1) {
1701                 list_del(&sum->list);
1702                 kfree(sum);
1703         }
1704         return ret;
1705 }
1706
1707 /* scrub extent tries to collect up to 64 kB for each bio */
1708 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1709                         u64 physical, u64 flags, u64 gen, int mirror_num)
1710 {
1711         int ret;
1712         u8 csum[BTRFS_CSUM_SIZE];
1713         u32 blocksize;
1714
1715         if (flags & BTRFS_EXTENT_FLAG_DATA) {
1716                 blocksize = sdev->sectorsize;
1717                 spin_lock(&sdev->stat_lock);
1718                 sdev->stat.data_extents_scrubbed++;
1719                 sdev->stat.data_bytes_scrubbed += len;
1720                 spin_unlock(&sdev->stat_lock);
1721         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1722                 BUG_ON(sdev->nodesize != sdev->leafsize);
1723                 blocksize = sdev->nodesize;
1724                 spin_lock(&sdev->stat_lock);
1725                 sdev->stat.tree_extents_scrubbed++;
1726                 sdev->stat.tree_bytes_scrubbed += len;
1727                 spin_unlock(&sdev->stat_lock);
1728         } else {
1729                 blocksize = sdev->sectorsize;
1730                 BUG_ON(1);
1731         }
1732
1733         while (len) {
1734                 u64 l = min_t(u64, len, blocksize);
1735                 int have_csum = 0;
1736
1737                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1738                         /* push csums to sbio */
1739                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1740                         if (have_csum == 0)
1741                                 ++sdev->stat.no_csum;
1742                 }
1743                 ret = scrub_pages(sdev, logical, l, physical, flags, gen,
1744                                   mirror_num, have_csum ? csum : NULL, 0);
1745                 if (ret)
1746                         return ret;
1747                 len -= l;
1748                 logical += l;
1749                 physical += l;
1750         }
1751         return 0;
1752 }
1753
1754 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1755         struct map_lookup *map, int num, u64 base, u64 length)
1756 {
1757         struct btrfs_path *path;
1758         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1759         struct btrfs_root *root = fs_info->extent_root;
1760         struct btrfs_root *csum_root = fs_info->csum_root;
1761         struct btrfs_extent_item *extent;
1762         struct blk_plug plug;
1763         u64 flags;
1764         int ret;
1765         int slot;
1766         int i;
1767         u64 nstripes;
1768         struct extent_buffer *l;
1769         struct btrfs_key key;
1770         u64 physical;
1771         u64 logical;
1772         u64 generation;
1773         int mirror_num;
1774         struct reada_control *reada1;
1775         struct reada_control *reada2;
1776         struct btrfs_key key_start;
1777         struct btrfs_key key_end;
1778
1779         u64 increment = map->stripe_len;
1780         u64 offset;
1781
1782         nstripes = length;
1783         offset = 0;
1784         do_div(nstripes, map->stripe_len);
1785         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1786                 offset = map->stripe_len * num;
1787                 increment = map->stripe_len * map->num_stripes;
1788                 mirror_num = 1;
1789         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1790                 int factor = map->num_stripes / map->sub_stripes;
1791                 offset = map->stripe_len * (num / map->sub_stripes);
1792                 increment = map->stripe_len * factor;
1793                 mirror_num = num % map->sub_stripes + 1;
1794         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1795                 increment = map->stripe_len;
1796                 mirror_num = num % map->num_stripes + 1;
1797         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1798                 increment = map->stripe_len;
1799                 mirror_num = num % map->num_stripes + 1;
1800         } else {
1801                 increment = map->stripe_len;
1802                 mirror_num = 1;
1803         }
1804
1805         path = btrfs_alloc_path();
1806         if (!path)
1807                 return -ENOMEM;
1808
1809         /*
1810          * work on commit root. The related disk blocks are static as
1811          * long as COW is applied. This means, it is save to rewrite
1812          * them to repair disk errors without any race conditions
1813          */
1814         path->search_commit_root = 1;
1815         path->skip_locking = 1;
1816
1817         /*
1818          * trigger the readahead for extent tree csum tree and wait for
1819          * completion. During readahead, the scrub is officially paused
1820          * to not hold off transaction commits
1821          */
1822         logical = base + offset;
1823
1824         wait_event(sdev->list_wait,
1825                    atomic_read(&sdev->in_flight) == 0);
1826         atomic_inc(&fs_info->scrubs_paused);
1827         wake_up(&fs_info->scrub_pause_wait);
1828
1829         /* FIXME it might be better to start readahead at commit root */
1830         key_start.objectid = logical;
1831         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1832         key_start.offset = (u64)0;
1833         key_end.objectid = base + offset + nstripes * increment;
1834         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1835         key_end.offset = (u64)0;
1836         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1837
1838         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1839         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1840         key_start.offset = logical;
1841         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1842         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1843         key_end.offset = base + offset + nstripes * increment;
1844         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1845
1846         if (!IS_ERR(reada1))
1847                 btrfs_reada_wait(reada1);
1848         if (!IS_ERR(reada2))
1849                 btrfs_reada_wait(reada2);
1850
1851         mutex_lock(&fs_info->scrub_lock);
1852         while (atomic_read(&fs_info->scrub_pause_req)) {
1853                 mutex_unlock(&fs_info->scrub_lock);
1854                 wait_event(fs_info->scrub_pause_wait,
1855                    atomic_read(&fs_info->scrub_pause_req) == 0);
1856                 mutex_lock(&fs_info->scrub_lock);
1857         }
1858         atomic_dec(&fs_info->scrubs_paused);
1859         mutex_unlock(&fs_info->scrub_lock);
1860         wake_up(&fs_info->scrub_pause_wait);
1861
1862         /*
1863          * collect all data csums for the stripe to avoid seeking during
1864          * the scrub. This might currently (crc32) end up to be about 1MB
1865          */
1866         blk_start_plug(&plug);
1867
1868         /*
1869          * now find all extents for each stripe and scrub them
1870          */
1871         logical = base + offset;
1872         physical = map->stripes[num].physical;
1873         ret = 0;
1874         for (i = 0; i < nstripes; ++i) {
1875                 /*
1876                  * canceled?
1877                  */
1878                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1879                     atomic_read(&sdev->cancel_req)) {
1880                         ret = -ECANCELED;
1881                         goto out;
1882                 }
1883                 /*
1884                  * check to see if we have to pause
1885                  */
1886                 if (atomic_read(&fs_info->scrub_pause_req)) {
1887                         /* push queued extents */
1888                         scrub_submit(sdev);
1889                         wait_event(sdev->list_wait,
1890                                    atomic_read(&sdev->in_flight) == 0);
1891                         atomic_inc(&fs_info->scrubs_paused);
1892                         wake_up(&fs_info->scrub_pause_wait);
1893                         mutex_lock(&fs_info->scrub_lock);
1894                         while (atomic_read(&fs_info->scrub_pause_req)) {
1895                                 mutex_unlock(&fs_info->scrub_lock);
1896                                 wait_event(fs_info->scrub_pause_wait,
1897                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1898                                 mutex_lock(&fs_info->scrub_lock);
1899                         }
1900                         atomic_dec(&fs_info->scrubs_paused);
1901                         mutex_unlock(&fs_info->scrub_lock);
1902                         wake_up(&fs_info->scrub_pause_wait);
1903                 }
1904
1905                 ret = btrfs_lookup_csums_range(csum_root, logical,
1906                                                logical + map->stripe_len - 1,
1907                                                &sdev->csum_list, 1);
1908                 if (ret)
1909                         goto out;
1910
1911                 key.objectid = logical;
1912                 key.type = BTRFS_EXTENT_ITEM_KEY;
1913                 key.offset = (u64)0;
1914
1915                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1916                 if (ret < 0)
1917                         goto out;
1918                 if (ret > 0) {
1919                         ret = btrfs_previous_item(root, path, 0,
1920                                                   BTRFS_EXTENT_ITEM_KEY);
1921                         if (ret < 0)
1922                                 goto out;
1923                         if (ret > 0) {
1924                                 /* there's no smaller item, so stick with the
1925                                  * larger one */
1926                                 btrfs_release_path(path);
1927                                 ret = btrfs_search_slot(NULL, root, &key,
1928                                                         path, 0, 0);
1929                                 if (ret < 0)
1930                                         goto out;
1931                         }
1932                 }
1933
1934                 while (1) {
1935                         l = path->nodes[0];
1936                         slot = path->slots[0];
1937                         if (slot >= btrfs_header_nritems(l)) {
1938                                 ret = btrfs_next_leaf(root, path);
1939                                 if (ret == 0)
1940                                         continue;
1941                                 if (ret < 0)
1942                                         goto out;
1943
1944                                 break;
1945                         }
1946                         btrfs_item_key_to_cpu(l, &key, slot);
1947
1948                         if (key.objectid + key.offset <= logical)
1949                                 goto next;
1950
1951                         if (key.objectid >= logical + map->stripe_len)
1952                                 break;
1953
1954                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1955                                 goto next;
1956
1957                         extent = btrfs_item_ptr(l, slot,
1958                                                 struct btrfs_extent_item);
1959                         flags = btrfs_extent_flags(l, extent);
1960                         generation = btrfs_extent_generation(l, extent);
1961
1962                         if (key.objectid < logical &&
1963                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1964                                 printk(KERN_ERR
1965                                        "btrfs scrub: tree block %llu spanning "
1966                                        "stripes, ignored. logical=%llu\n",
1967                                        (unsigned long long)key.objectid,
1968                                        (unsigned long long)logical);
1969                                 goto next;
1970                         }
1971
1972                         /*
1973                          * trim extent to this stripe
1974                          */
1975                         if (key.objectid < logical) {
1976                                 key.offset -= logical - key.objectid;
1977                                 key.objectid = logical;
1978                         }
1979                         if (key.objectid + key.offset >
1980                             logical + map->stripe_len) {
1981                                 key.offset = logical + map->stripe_len -
1982                                              key.objectid;
1983                         }
1984
1985                         ret = scrub_extent(sdev, key.objectid, key.offset,
1986                                            key.objectid - logical + physical,
1987                                            flags, generation, mirror_num);
1988                         if (ret)
1989                                 goto out;
1990
1991 next:
1992                         path->slots[0]++;
1993                 }
1994                 btrfs_release_path(path);
1995                 logical += increment;
1996                 physical += map->stripe_len;
1997                 spin_lock(&sdev->stat_lock);
1998                 sdev->stat.last_physical = physical;
1999                 spin_unlock(&sdev->stat_lock);
2000         }
2001         /* push queued extents */
2002         scrub_submit(sdev);
2003
2004 out:
2005         blk_finish_plug(&plug);
2006         btrfs_free_path(path);
2007         return ret < 0 ? ret : 0;
2008 }
2009
2010 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
2011         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
2012         u64 dev_offset)
2013 {
2014         struct btrfs_mapping_tree *map_tree =
2015                 &sdev->dev->dev_root->fs_info->mapping_tree;
2016         struct map_lookup *map;
2017         struct extent_map *em;
2018         int i;
2019         int ret = -EINVAL;
2020
2021         read_lock(&map_tree->map_tree.lock);
2022         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2023         read_unlock(&map_tree->map_tree.lock);
2024
2025         if (!em)
2026                 return -EINVAL;
2027
2028         map = (struct map_lookup *)em->bdev;
2029         if (em->start != chunk_offset)
2030                 goto out;
2031
2032         if (em->len < length)
2033                 goto out;
2034
2035         for (i = 0; i < map->num_stripes; ++i) {
2036                 if (map->stripes[i].dev == sdev->dev &&
2037                     map->stripes[i].physical == dev_offset) {
2038                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
2039                         if (ret)
2040                                 goto out;
2041                 }
2042         }
2043 out:
2044         free_extent_map(em);
2045
2046         return ret;
2047 }
2048
2049 static noinline_for_stack
2050 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
2051 {
2052         struct btrfs_dev_extent *dev_extent = NULL;
2053         struct btrfs_path *path;
2054         struct btrfs_root *root = sdev->dev->dev_root;
2055         struct btrfs_fs_info *fs_info = root->fs_info;
2056         u64 length;
2057         u64 chunk_tree;
2058         u64 chunk_objectid;
2059         u64 chunk_offset;
2060         int ret;
2061         int slot;
2062         struct extent_buffer *l;
2063         struct btrfs_key key;
2064         struct btrfs_key found_key;
2065         struct btrfs_block_group_cache *cache;
2066
2067         path = btrfs_alloc_path();
2068         if (!path)
2069                 return -ENOMEM;
2070
2071         path->reada = 2;
2072         path->search_commit_root = 1;
2073         path->skip_locking = 1;
2074
2075         key.objectid = sdev->dev->devid;
2076         key.offset = 0ull;
2077         key.type = BTRFS_DEV_EXTENT_KEY;
2078
2079
2080         while (1) {
2081                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2082                 if (ret < 0)
2083                         break;
2084                 if (ret > 0) {
2085                         if (path->slots[0] >=
2086                             btrfs_header_nritems(path->nodes[0])) {
2087                                 ret = btrfs_next_leaf(root, path);
2088                                 if (ret)
2089                                         break;
2090                         }
2091                 }
2092
2093                 l = path->nodes[0];
2094                 slot = path->slots[0];
2095
2096                 btrfs_item_key_to_cpu(l, &found_key, slot);
2097
2098                 if (found_key.objectid != sdev->dev->devid)
2099                         break;
2100
2101                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2102                         break;
2103
2104                 if (found_key.offset >= end)
2105                         break;
2106
2107                 if (found_key.offset < key.offset)
2108                         break;
2109
2110                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2111                 length = btrfs_dev_extent_length(l, dev_extent);
2112
2113                 if (found_key.offset + length <= start) {
2114                         key.offset = found_key.offset + length;
2115                         btrfs_release_path(path);
2116                         continue;
2117                 }
2118
2119                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2120                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2121                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2122
2123                 /*
2124                  * get a reference on the corresponding block group to prevent
2125                  * the chunk from going away while we scrub it
2126                  */
2127                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2128                 if (!cache) {
2129                         ret = -ENOENT;
2130                         break;
2131                 }
2132                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
2133                                   chunk_offset, length, found_key.offset);
2134                 btrfs_put_block_group(cache);
2135                 if (ret)
2136                         break;
2137
2138                 key.offset = found_key.offset + length;
2139                 btrfs_release_path(path);
2140         }
2141
2142         btrfs_free_path(path);
2143
2144         /*
2145          * ret can still be 1 from search_slot or next_leaf,
2146          * that's not an error
2147          */
2148         return ret < 0 ? ret : 0;
2149 }
2150
2151 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
2152 {
2153         int     i;
2154         u64     bytenr;
2155         u64     gen;
2156         int     ret;
2157         struct btrfs_device *device = sdev->dev;
2158         struct btrfs_root *root = device->dev_root;
2159
2160         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2161                 return -EIO;
2162
2163         gen = root->fs_info->last_trans_committed;
2164
2165         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2166                 bytenr = btrfs_sb_offset(i);
2167                 if (bytenr + BTRFS_SUPER_INFO_SIZE > device->total_bytes)
2168                         break;
2169
2170                 ret = scrub_pages(sdev, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2171                                      BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
2172                 if (ret)
2173                         return ret;
2174         }
2175         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2176
2177         return 0;
2178 }
2179
2180 /*
2181  * get a reference count on fs_info->scrub_workers. start worker if necessary
2182  */
2183 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
2184 {
2185         struct btrfs_fs_info *fs_info = root->fs_info;
2186         int ret = 0;
2187
2188         mutex_lock(&fs_info->scrub_lock);
2189         if (fs_info->scrub_workers_refcnt == 0) {
2190                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2191                            fs_info->thread_pool_size, &fs_info->generic_worker);
2192                 fs_info->scrub_workers.idle_thresh = 4;
2193                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2194                 if (ret)
2195                         goto out;
2196         }
2197         ++fs_info->scrub_workers_refcnt;
2198 out:
2199         mutex_unlock(&fs_info->scrub_lock);
2200
2201         return ret;
2202 }
2203
2204 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
2205 {
2206         struct btrfs_fs_info *fs_info = root->fs_info;
2207
2208         mutex_lock(&fs_info->scrub_lock);
2209         if (--fs_info->scrub_workers_refcnt == 0)
2210                 btrfs_stop_workers(&fs_info->scrub_workers);
2211         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2212         mutex_unlock(&fs_info->scrub_lock);
2213 }
2214
2215
2216 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
2217                     struct btrfs_scrub_progress *progress, int readonly)
2218 {
2219         struct scrub_dev *sdev;
2220         struct btrfs_fs_info *fs_info = root->fs_info;
2221         int ret;
2222         struct btrfs_device *dev;
2223
2224         if (btrfs_fs_closing(root->fs_info))
2225                 return -EINVAL;
2226
2227         /*
2228          * check some assumptions
2229          */
2230         if (root->nodesize != root->leafsize) {
2231                 printk(KERN_ERR
2232                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2233                        root->nodesize, root->leafsize);
2234                 return -EINVAL;
2235         }
2236
2237         if (root->nodesize > BTRFS_STRIPE_LEN) {
2238                 /*
2239                  * in this case scrub is unable to calculate the checksum
2240                  * the way scrub is implemented. Do not handle this
2241                  * situation at all because it won't ever happen.
2242                  */
2243                 printk(KERN_ERR
2244                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2245                        root->nodesize, BTRFS_STRIPE_LEN);
2246                 return -EINVAL;
2247         }
2248
2249         if (root->sectorsize != PAGE_SIZE) {
2250                 /* not supported for data w/o checksums */
2251                 printk(KERN_ERR
2252                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2253                        root->sectorsize, (unsigned long long)PAGE_SIZE);
2254                 return -EINVAL;
2255         }
2256
2257         ret = scrub_workers_get(root);
2258         if (ret)
2259                 return ret;
2260
2261         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2262         dev = btrfs_find_device(root, devid, NULL, NULL);
2263         if (!dev || dev->missing) {
2264                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2265                 scrub_workers_put(root);
2266                 return -ENODEV;
2267         }
2268         mutex_lock(&fs_info->scrub_lock);
2269
2270         if (!dev->in_fs_metadata) {
2271                 mutex_unlock(&fs_info->scrub_lock);
2272                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2273                 scrub_workers_put(root);
2274                 return -ENODEV;
2275         }
2276
2277         if (dev->scrub_device) {
2278                 mutex_unlock(&fs_info->scrub_lock);
2279                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2280                 scrub_workers_put(root);
2281                 return -EINPROGRESS;
2282         }
2283         sdev = scrub_setup_dev(dev);
2284         if (IS_ERR(sdev)) {
2285                 mutex_unlock(&fs_info->scrub_lock);
2286                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2287                 scrub_workers_put(root);
2288                 return PTR_ERR(sdev);
2289         }
2290         sdev->readonly = readonly;
2291         dev->scrub_device = sdev;
2292
2293         atomic_inc(&fs_info->scrubs_running);
2294         mutex_unlock(&fs_info->scrub_lock);
2295         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2296
2297         down_read(&fs_info->scrub_super_lock);
2298         ret = scrub_supers(sdev);
2299         up_read(&fs_info->scrub_super_lock);
2300
2301         if (!ret)
2302                 ret = scrub_enumerate_chunks(sdev, start, end);
2303
2304         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
2305         atomic_dec(&fs_info->scrubs_running);
2306         wake_up(&fs_info->scrub_pause_wait);
2307
2308         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
2309
2310         if (progress)
2311                 memcpy(progress, &sdev->stat, sizeof(*progress));
2312
2313         mutex_lock(&fs_info->scrub_lock);
2314         dev->scrub_device = NULL;
2315         mutex_unlock(&fs_info->scrub_lock);
2316
2317         scrub_free_dev(sdev);
2318         scrub_workers_put(root);
2319
2320         return ret;
2321 }
2322
2323 void btrfs_scrub_pause(struct btrfs_root *root)
2324 {
2325         struct btrfs_fs_info *fs_info = root->fs_info;
2326
2327         mutex_lock(&fs_info->scrub_lock);
2328         atomic_inc(&fs_info->scrub_pause_req);
2329         while (atomic_read(&fs_info->scrubs_paused) !=
2330                atomic_read(&fs_info->scrubs_running)) {
2331                 mutex_unlock(&fs_info->scrub_lock);
2332                 wait_event(fs_info->scrub_pause_wait,
2333                            atomic_read(&fs_info->scrubs_paused) ==
2334                            atomic_read(&fs_info->scrubs_running));
2335                 mutex_lock(&fs_info->scrub_lock);
2336         }
2337         mutex_unlock(&fs_info->scrub_lock);
2338 }
2339
2340 void btrfs_scrub_continue(struct btrfs_root *root)
2341 {
2342         struct btrfs_fs_info *fs_info = root->fs_info;
2343
2344         atomic_dec(&fs_info->scrub_pause_req);
2345         wake_up(&fs_info->scrub_pause_wait);
2346 }
2347
2348 void btrfs_scrub_pause_super(struct btrfs_root *root)
2349 {
2350         down_write(&root->fs_info->scrub_super_lock);
2351 }
2352
2353 void btrfs_scrub_continue_super(struct btrfs_root *root)
2354 {
2355         up_write(&root->fs_info->scrub_super_lock);
2356 }
2357
2358 int __btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2359 {
2360
2361         mutex_lock(&fs_info->scrub_lock);
2362         if (!atomic_read(&fs_info->scrubs_running)) {
2363                 mutex_unlock(&fs_info->scrub_lock);
2364                 return -ENOTCONN;
2365         }
2366
2367         atomic_inc(&fs_info->scrub_cancel_req);
2368         while (atomic_read(&fs_info->scrubs_running)) {
2369                 mutex_unlock(&fs_info->scrub_lock);
2370                 wait_event(fs_info->scrub_pause_wait,
2371                            atomic_read(&fs_info->scrubs_running) == 0);
2372                 mutex_lock(&fs_info->scrub_lock);
2373         }
2374         atomic_dec(&fs_info->scrub_cancel_req);
2375         mutex_unlock(&fs_info->scrub_lock);
2376
2377         return 0;
2378 }
2379
2380 int btrfs_scrub_cancel(struct btrfs_root *root)
2381 {
2382         return __btrfs_scrub_cancel(root->fs_info);
2383 }
2384
2385 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
2386 {
2387         struct btrfs_fs_info *fs_info = root->fs_info;
2388         struct scrub_dev *sdev;
2389
2390         mutex_lock(&fs_info->scrub_lock);
2391         sdev = dev->scrub_device;
2392         if (!sdev) {
2393                 mutex_unlock(&fs_info->scrub_lock);
2394                 return -ENOTCONN;
2395         }
2396         atomic_inc(&sdev->cancel_req);
2397         while (dev->scrub_device) {
2398                 mutex_unlock(&fs_info->scrub_lock);
2399                 wait_event(fs_info->scrub_pause_wait,
2400                            dev->scrub_device == NULL);
2401                 mutex_lock(&fs_info->scrub_lock);
2402         }
2403         mutex_unlock(&fs_info->scrub_lock);
2404
2405         return 0;
2406 }
2407
2408 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2409 {
2410         struct btrfs_fs_info *fs_info = root->fs_info;
2411         struct btrfs_device *dev;
2412         int ret;
2413
2414         /*
2415          * we have to hold the device_list_mutex here so the device
2416          * does not go away in cancel_dev. FIXME: find a better solution
2417          */
2418         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2419         dev = btrfs_find_device(root, devid, NULL, NULL);
2420         if (!dev) {
2421                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2422                 return -ENODEV;
2423         }
2424         ret = btrfs_scrub_cancel_dev(root, dev);
2425         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2426
2427         return ret;
2428 }
2429
2430 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
2431                          struct btrfs_scrub_progress *progress)
2432 {
2433         struct btrfs_device *dev;
2434         struct scrub_dev *sdev = NULL;
2435
2436         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
2437         dev = btrfs_find_device(root, devid, NULL, NULL);
2438         if (dev)
2439                 sdev = dev->scrub_device;
2440         if (sdev)
2441                 memcpy(progress, &sdev->stat, sizeof(*progress));
2442         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
2443
2444         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
2445 }