ia64: Fixup asm/cmpxchg.h
[~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  *  - In case of a read error on files with nodatasum, map the file and read
40  *    the extent to trigger a writeback of the good copy
41  *  - track and record media errors, throw out bad devices
42  *  - add a mode to also read unallocated space
43  */
44
45 struct scrub_bio;
46 struct scrub_page;
47 struct scrub_dev;
48 static void scrub_bio_end_io(struct bio *bio, int err);
49 static void scrub_checksum(struct btrfs_work *work);
50 static int scrub_checksum_data(struct scrub_dev *sdev,
51                                struct scrub_page *spag, void *buffer);
52 static int scrub_checksum_tree_block(struct scrub_dev *sdev,
53                                      struct scrub_page *spag, u64 logical,
54                                      void *buffer);
55 static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer);
56 static int scrub_fixup_check(struct scrub_bio *sbio, int ix);
57 static void scrub_fixup_end_io(struct bio *bio, int err);
58 static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
59                           struct page *page);
60 static void scrub_fixup(struct scrub_bio *sbio, int ix);
61
62 #define SCRUB_PAGES_PER_BIO     16      /* 64k per bio */
63 #define SCRUB_BIOS_PER_DEV      16      /* 1 MB per device in flight */
64
65 struct scrub_page {
66         u64                     flags;  /* extent flags */
67         u64                     generation;
68         int                     mirror_num;
69         int                     have_csum;
70         u8                      csum[BTRFS_CSUM_SIZE];
71 };
72
73 struct scrub_bio {
74         int                     index;
75         struct scrub_dev        *sdev;
76         struct bio              *bio;
77         int                     err;
78         u64                     logical;
79         u64                     physical;
80         struct scrub_page       spag[SCRUB_PAGES_PER_BIO];
81         u64                     count;
82         int                     next_free;
83         struct btrfs_work       work;
84 };
85
86 struct scrub_dev {
87         struct scrub_bio        *bios[SCRUB_BIOS_PER_DEV];
88         struct btrfs_device     *dev;
89         int                     first_free;
90         int                     curr;
91         atomic_t                in_flight;
92         atomic_t                fixup_cnt;
93         spinlock_t              list_lock;
94         wait_queue_head_t       list_wait;
95         u16                     csum_size;
96         struct list_head        csum_list;
97         atomic_t                cancel_req;
98         int                     readonly;
99         /*
100          * statistics
101          */
102         struct btrfs_scrub_progress stat;
103         spinlock_t              stat_lock;
104 };
105
106 struct scrub_fixup_nodatasum {
107         struct scrub_dev        *sdev;
108         u64                     logical;
109         struct btrfs_root       *root;
110         struct btrfs_work       work;
111         int                     mirror_num;
112 };
113
114 struct scrub_warning {
115         struct btrfs_path       *path;
116         u64                     extent_item_size;
117         char                    *scratch_buf;
118         char                    *msg_buf;
119         const char              *errstr;
120         sector_t                sector;
121         u64                     logical;
122         struct btrfs_device     *dev;
123         int                     msg_bufsize;
124         int                     scratch_bufsize;
125 };
126
127 static void scrub_free_csums(struct scrub_dev *sdev)
128 {
129         while (!list_empty(&sdev->csum_list)) {
130                 struct btrfs_ordered_sum *sum;
131                 sum = list_first_entry(&sdev->csum_list,
132                                        struct btrfs_ordered_sum, list);
133                 list_del(&sum->list);
134                 kfree(sum);
135         }
136 }
137
138 static void scrub_free_bio(struct bio *bio)
139 {
140         int i;
141         struct page *last_page = NULL;
142
143         if (!bio)
144                 return;
145
146         for (i = 0; i < bio->bi_vcnt; ++i) {
147                 if (bio->bi_io_vec[i].bv_page == last_page)
148                         continue;
149                 last_page = bio->bi_io_vec[i].bv_page;
150                 __free_page(last_page);
151         }
152         bio_put(bio);
153 }
154
155 static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
156 {
157         int i;
158
159         if (!sdev)
160                 return;
161
162         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
163                 struct scrub_bio *sbio = sdev->bios[i];
164
165                 if (!sbio)
166                         break;
167
168                 scrub_free_bio(sbio->bio);
169                 kfree(sbio);
170         }
171
172         scrub_free_csums(sdev);
173         kfree(sdev);
174 }
175
176 static noinline_for_stack
177 struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
178 {
179         struct scrub_dev *sdev;
180         int             i;
181         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
182
183         sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
184         if (!sdev)
185                 goto nomem;
186         sdev->dev = dev;
187         for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
188                 struct scrub_bio *sbio;
189
190                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
191                 if (!sbio)
192                         goto nomem;
193                 sdev->bios[i] = sbio;
194
195                 sbio->index = i;
196                 sbio->sdev = sdev;
197                 sbio->count = 0;
198                 sbio->work.func = scrub_checksum;
199
200                 if (i != SCRUB_BIOS_PER_DEV-1)
201                         sdev->bios[i]->next_free = i + 1;
202                 else
203                         sdev->bios[i]->next_free = -1;
204         }
205         sdev->first_free = 0;
206         sdev->curr = -1;
207         atomic_set(&sdev->in_flight, 0);
208         atomic_set(&sdev->fixup_cnt, 0);
209         atomic_set(&sdev->cancel_req, 0);
210         sdev->csum_size = btrfs_super_csum_size(fs_info->super_copy);
211         INIT_LIST_HEAD(&sdev->csum_list);
212
213         spin_lock_init(&sdev->list_lock);
214         spin_lock_init(&sdev->stat_lock);
215         init_waitqueue_head(&sdev->list_wait);
216         return sdev;
217
218 nomem:
219         scrub_free_dev(sdev);
220         return ERR_PTR(-ENOMEM);
221 }
222
223 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx)
224 {
225         u64 isize;
226         u32 nlink;
227         int ret;
228         int i;
229         struct extent_buffer *eb;
230         struct btrfs_inode_item *inode_item;
231         struct scrub_warning *swarn = ctx;
232         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
233         struct inode_fs_paths *ipath = NULL;
234         struct btrfs_root *local_root;
235         struct btrfs_key root_key;
236
237         root_key.objectid = root;
238         root_key.type = BTRFS_ROOT_ITEM_KEY;
239         root_key.offset = (u64)-1;
240         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
241         if (IS_ERR(local_root)) {
242                 ret = PTR_ERR(local_root);
243                 goto err;
244         }
245
246         ret = inode_item_info(inum, 0, local_root, swarn->path);
247         if (ret) {
248                 btrfs_release_path(swarn->path);
249                 goto err;
250         }
251
252         eb = swarn->path->nodes[0];
253         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
254                                         struct btrfs_inode_item);
255         isize = btrfs_inode_size(eb, inode_item);
256         nlink = btrfs_inode_nlink(eb, inode_item);
257         btrfs_release_path(swarn->path);
258
259         ipath = init_ipath(4096, local_root, swarn->path);
260         if (IS_ERR(ipath)) {
261                 ret = PTR_ERR(ipath);
262                 ipath = NULL;
263                 goto err;
264         }
265         ret = paths_from_inode(inum, ipath);
266
267         if (ret < 0)
268                 goto err;
269
270         /*
271          * we deliberately ignore the bit ipath might have been too small to
272          * hold all of the paths here
273          */
274         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
275                 printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
276                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
277                         "length %llu, links %u (path: %s)\n", swarn->errstr,
278                         swarn->logical, swarn->dev->name,
279                         (unsigned long long)swarn->sector, root, inum, offset,
280                         min(isize - offset, (u64)PAGE_SIZE), nlink,
281                         (char *)(unsigned long)ipath->fspath->val[i]);
282
283         free_ipath(ipath);
284         return 0;
285
286 err:
287         printk(KERN_WARNING "btrfs: %s at logical %llu on dev "
288                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
289                 "resolving failed with ret=%d\n", swarn->errstr,
290                 swarn->logical, swarn->dev->name,
291                 (unsigned long long)swarn->sector, root, inum, offset, ret);
292
293         free_ipath(ipath);
294         return 0;
295 }
296
297 static void scrub_print_warning(const char *errstr, struct scrub_bio *sbio,
298                                 int ix)
299 {
300         struct btrfs_device *dev = sbio->sdev->dev;
301         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
302         struct btrfs_path *path;
303         struct btrfs_key found_key;
304         struct extent_buffer *eb;
305         struct btrfs_extent_item *ei;
306         struct scrub_warning swarn;
307         u32 item_size;
308         int ret;
309         u64 ref_root;
310         u8 ref_level;
311         unsigned long ptr = 0;
312         const int bufsize = 4096;
313         u64 extent_item_pos;
314
315         path = btrfs_alloc_path();
316
317         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
318         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
319         swarn.sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
320         swarn.logical = sbio->logical + ix * PAGE_SIZE;
321         swarn.errstr = errstr;
322         swarn.dev = dev;
323         swarn.msg_bufsize = bufsize;
324         swarn.scratch_bufsize = bufsize;
325
326         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
327                 goto out;
328
329         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key);
330         if (ret < 0)
331                 goto out;
332
333         extent_item_pos = swarn.logical - found_key.objectid;
334         swarn.extent_item_size = found_key.offset;
335
336         eb = path->nodes[0];
337         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
338         item_size = btrfs_item_size_nr(eb, path->slots[0]);
339         btrfs_release_path(path);
340
341         if (ret & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
342                 do {
343                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
344                                                         &ref_root, &ref_level);
345                         printk(KERN_WARNING "%s at logical %llu on dev %s, "
346                                 "sector %llu: metadata %s (level %d) in tree "
347                                 "%llu\n", errstr, swarn.logical, dev->name,
348                                 (unsigned long long)swarn.sector,
349                                 ref_level ? "node" : "leaf",
350                                 ret < 0 ? -1 : ref_level,
351                                 ret < 0 ? -1 : ref_root);
352                 } while (ret != 1);
353         } else {
354                 swarn.path = path;
355                 iterate_extent_inodes(fs_info, path, found_key.objectid,
356                                         extent_item_pos,
357                                         scrub_print_warning_inode, &swarn);
358         }
359
360 out:
361         btrfs_free_path(path);
362         kfree(swarn.scratch_buf);
363         kfree(swarn.msg_buf);
364 }
365
366 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx)
367 {
368         struct page *page = NULL;
369         unsigned long index;
370         struct scrub_fixup_nodatasum *fixup = ctx;
371         int ret;
372         int corrected = 0;
373         struct btrfs_key key;
374         struct inode *inode = NULL;
375         u64 end = offset + PAGE_SIZE - 1;
376         struct btrfs_root *local_root;
377
378         key.objectid = root;
379         key.type = BTRFS_ROOT_ITEM_KEY;
380         key.offset = (u64)-1;
381         local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
382         if (IS_ERR(local_root))
383                 return PTR_ERR(local_root);
384
385         key.type = BTRFS_INODE_ITEM_KEY;
386         key.objectid = inum;
387         key.offset = 0;
388         inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
389         if (IS_ERR(inode))
390                 return PTR_ERR(inode);
391
392         index = offset >> PAGE_CACHE_SHIFT;
393
394         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
395         if (!page) {
396                 ret = -ENOMEM;
397                 goto out;
398         }
399
400         if (PageUptodate(page)) {
401                 struct btrfs_mapping_tree *map_tree;
402                 if (PageDirty(page)) {
403                         /*
404                          * we need to write the data to the defect sector. the
405                          * data that was in that sector is not in memory,
406                          * because the page was modified. we must not write the
407                          * modified page to that sector.
408                          *
409                          * TODO: what could be done here: wait for the delalloc
410                          *       runner to write out that page (might involve
411                          *       COW) and see whether the sector is still
412                          *       referenced afterwards.
413                          *
414                          * For the meantime, we'll treat this error
415                          * incorrectable, although there is a chance that a
416                          * later scrub will find the bad sector again and that
417                          * there's no dirty page in memory, then.
418                          */
419                         ret = -EIO;
420                         goto out;
421                 }
422                 map_tree = &BTRFS_I(inode)->root->fs_info->mapping_tree;
423                 ret = repair_io_failure(map_tree, offset, PAGE_SIZE,
424                                         fixup->logical, page,
425                                         fixup->mirror_num);
426                 unlock_page(page);
427                 corrected = !ret;
428         } else {
429                 /*
430                  * we need to get good data first. the general readpage path
431                  * will call repair_io_failure for us, we just have to make
432                  * sure we read the bad mirror.
433                  */
434                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
435                                         EXTENT_DAMAGED, GFP_NOFS);
436                 if (ret) {
437                         /* set_extent_bits should give proper error */
438                         WARN_ON(ret > 0);
439                         if (ret > 0)
440                                 ret = -EFAULT;
441                         goto out;
442                 }
443
444                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
445                                                 btrfs_get_extent,
446                                                 fixup->mirror_num);
447                 wait_on_page_locked(page);
448
449                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
450                                                 end, EXTENT_DAMAGED, 0, NULL);
451                 if (!corrected)
452                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
453                                                 EXTENT_DAMAGED, GFP_NOFS);
454         }
455
456 out:
457         if (page)
458                 put_page(page);
459         if (inode)
460                 iput(inode);
461
462         if (ret < 0)
463                 return ret;
464
465         if (ret == 0 && corrected) {
466                 /*
467                  * we only need to call readpage for one of the inodes belonging
468                  * to this extent. so make iterate_extent_inodes stop
469                  */
470                 return 1;
471         }
472
473         return -EIO;
474 }
475
476 static void scrub_fixup_nodatasum(struct btrfs_work *work)
477 {
478         int ret;
479         struct scrub_fixup_nodatasum *fixup;
480         struct scrub_dev *sdev;
481         struct btrfs_trans_handle *trans = NULL;
482         struct btrfs_fs_info *fs_info;
483         struct btrfs_path *path;
484         int uncorrectable = 0;
485
486         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
487         sdev = fixup->sdev;
488         fs_info = fixup->root->fs_info;
489
490         path = btrfs_alloc_path();
491         if (!path) {
492                 spin_lock(&sdev->stat_lock);
493                 ++sdev->stat.malloc_errors;
494                 spin_unlock(&sdev->stat_lock);
495                 uncorrectable = 1;
496                 goto out;
497         }
498
499         trans = btrfs_join_transaction(fixup->root);
500         if (IS_ERR(trans)) {
501                 uncorrectable = 1;
502                 goto out;
503         }
504
505         /*
506          * the idea is to trigger a regular read through the standard path. we
507          * read a page from the (failed) logical address by specifying the
508          * corresponding copynum of the failed sector. thus, that readpage is
509          * expected to fail.
510          * that is the point where on-the-fly error correction will kick in
511          * (once it's finished) and rewrite the failed sector if a good copy
512          * can be found.
513          */
514         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
515                                                 path, scrub_fixup_readpage,
516                                                 fixup);
517         if (ret < 0) {
518                 uncorrectable = 1;
519                 goto out;
520         }
521         WARN_ON(ret != 1);
522
523         spin_lock(&sdev->stat_lock);
524         ++sdev->stat.corrected_errors;
525         spin_unlock(&sdev->stat_lock);
526
527 out:
528         if (trans && !IS_ERR(trans))
529                 btrfs_end_transaction(trans, fixup->root);
530         if (uncorrectable) {
531                 spin_lock(&sdev->stat_lock);
532                 ++sdev->stat.uncorrectable_errors;
533                 spin_unlock(&sdev->stat_lock);
534                 printk_ratelimited(KERN_ERR "btrfs: unable to fixup "
535                                         "(nodatasum) error at logical %llu\n",
536                                         fixup->logical);
537         }
538
539         btrfs_free_path(path);
540         kfree(fixup);
541
542         /* see caller why we're pretending to be paused in the scrub counters */
543         mutex_lock(&fs_info->scrub_lock);
544         atomic_dec(&fs_info->scrubs_running);
545         atomic_dec(&fs_info->scrubs_paused);
546         mutex_unlock(&fs_info->scrub_lock);
547         atomic_dec(&sdev->fixup_cnt);
548         wake_up(&fs_info->scrub_pause_wait);
549         wake_up(&sdev->list_wait);
550 }
551
552 /*
553  * scrub_recheck_error gets called when either verification of the page
554  * failed or the bio failed to read, e.g. with EIO. In the latter case,
555  * recheck_error gets called for every page in the bio, even though only
556  * one may be bad
557  */
558 static int scrub_recheck_error(struct scrub_bio *sbio, int ix)
559 {
560         struct scrub_dev *sdev = sbio->sdev;
561         u64 sector = (sbio->physical + ix * PAGE_SIZE) >> 9;
562         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
563                                         DEFAULT_RATELIMIT_BURST);
564
565         if (sbio->err) {
566                 if (scrub_fixup_io(READ, sbio->sdev->dev->bdev, sector,
567                                    sbio->bio->bi_io_vec[ix].bv_page) == 0) {
568                         if (scrub_fixup_check(sbio, ix) == 0)
569                                 return 0;
570                 }
571                 if (__ratelimit(&_rs))
572                         scrub_print_warning("i/o error", sbio, ix);
573         } else {
574                 if (__ratelimit(&_rs))
575                         scrub_print_warning("checksum error", sbio, ix);
576         }
577
578         spin_lock(&sdev->stat_lock);
579         ++sdev->stat.read_errors;
580         spin_unlock(&sdev->stat_lock);
581
582         scrub_fixup(sbio, ix);
583         return 1;
584 }
585
586 static int scrub_fixup_check(struct scrub_bio *sbio, int ix)
587 {
588         int ret = 1;
589         struct page *page;
590         void *buffer;
591         u64 flags = sbio->spag[ix].flags;
592
593         page = sbio->bio->bi_io_vec[ix].bv_page;
594         buffer = kmap_atomic(page);
595         if (flags & BTRFS_EXTENT_FLAG_DATA) {
596                 ret = scrub_checksum_data(sbio->sdev,
597                                           sbio->spag + ix, buffer);
598         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
599                 ret = scrub_checksum_tree_block(sbio->sdev,
600                                                 sbio->spag + ix,
601                                                 sbio->logical + ix * PAGE_SIZE,
602                                                 buffer);
603         } else {
604                 WARN_ON(1);
605         }
606         kunmap_atomic(buffer);
607
608         return ret;
609 }
610
611 static void scrub_fixup_end_io(struct bio *bio, int err)
612 {
613         complete((struct completion *)bio->bi_private);
614 }
615
616 static void scrub_fixup(struct scrub_bio *sbio, int ix)
617 {
618         struct scrub_dev *sdev = sbio->sdev;
619         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
620         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
621         struct btrfs_bio *bbio = NULL;
622         struct scrub_fixup_nodatasum *fixup;
623         u64 logical = sbio->logical + ix * PAGE_SIZE;
624         u64 length;
625         int i;
626         int ret;
627         DECLARE_COMPLETION_ONSTACK(complete);
628
629         if ((sbio->spag[ix].flags & BTRFS_EXTENT_FLAG_DATA) &&
630             (sbio->spag[ix].have_csum == 0)) {
631                 fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
632                 if (!fixup)
633                         goto uncorrectable;
634                 fixup->sdev = sdev;
635                 fixup->logical = logical;
636                 fixup->root = fs_info->extent_root;
637                 fixup->mirror_num = sbio->spag[ix].mirror_num;
638                 /*
639                  * increment scrubs_running to prevent cancel requests from
640                  * completing as long as a fixup worker is running. we must also
641                  * increment scrubs_paused to prevent deadlocking on pause
642                  * requests used for transactions commits (as the worker uses a
643                  * transaction context). it is safe to regard the fixup worker
644                  * as paused for all matters practical. effectively, we only
645                  * avoid cancellation requests from completing.
646                  */
647                 mutex_lock(&fs_info->scrub_lock);
648                 atomic_inc(&fs_info->scrubs_running);
649                 atomic_inc(&fs_info->scrubs_paused);
650                 mutex_unlock(&fs_info->scrub_lock);
651                 atomic_inc(&sdev->fixup_cnt);
652                 fixup->work.func = scrub_fixup_nodatasum;
653                 btrfs_queue_worker(&fs_info->scrub_workers, &fixup->work);
654                 return;
655         }
656
657         length = PAGE_SIZE;
658         ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length,
659                               &bbio, 0);
660         if (ret || !bbio || length < PAGE_SIZE) {
661                 printk(KERN_ERR
662                        "scrub_fixup: btrfs_map_block failed us for %llu\n",
663                        (unsigned long long)logical);
664                 WARN_ON(1);
665                 kfree(bbio);
666                 return;
667         }
668
669         if (bbio->num_stripes == 1)
670                 /* there aren't any replicas */
671                 goto uncorrectable;
672
673         /*
674          * first find a good copy
675          */
676         for (i = 0; i < bbio->num_stripes; ++i) {
677                 if (i + 1 == sbio->spag[ix].mirror_num)
678                         continue;
679
680                 if (scrub_fixup_io(READ, bbio->stripes[i].dev->bdev,
681                                    bbio->stripes[i].physical >> 9,
682                                    sbio->bio->bi_io_vec[ix].bv_page)) {
683                         /* I/O-error, this is not a good copy */
684                         continue;
685                 }
686
687                 if (scrub_fixup_check(sbio, ix) == 0)
688                         break;
689         }
690         if (i == bbio->num_stripes)
691                 goto uncorrectable;
692
693         if (!sdev->readonly) {
694                 /*
695                  * bi_io_vec[ix].bv_page now contains good data, write it back
696                  */
697                 if (scrub_fixup_io(WRITE, sdev->dev->bdev,
698                                    (sbio->physical + ix * PAGE_SIZE) >> 9,
699                                    sbio->bio->bi_io_vec[ix].bv_page)) {
700                         /* I/O-error, writeback failed, give up */
701                         goto uncorrectable;
702                 }
703         }
704
705         kfree(bbio);
706         spin_lock(&sdev->stat_lock);
707         ++sdev->stat.corrected_errors;
708         spin_unlock(&sdev->stat_lock);
709
710         printk_ratelimited(KERN_ERR "btrfs: fixed up error at logical %llu\n",
711                                (unsigned long long)logical);
712         return;
713
714 uncorrectable:
715         kfree(bbio);
716         spin_lock(&sdev->stat_lock);
717         ++sdev->stat.uncorrectable_errors;
718         spin_unlock(&sdev->stat_lock);
719
720         printk_ratelimited(KERN_ERR "btrfs: unable to fixup (regular) error at "
721                                 "logical %llu\n", (unsigned long long)logical);
722 }
723
724 static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
725                          struct page *page)
726 {
727         struct bio *bio = NULL;
728         int ret;
729         DECLARE_COMPLETION_ONSTACK(complete);
730
731         bio = bio_alloc(GFP_NOFS, 1);
732         bio->bi_bdev = bdev;
733         bio->bi_sector = sector;
734         bio_add_page(bio, page, PAGE_SIZE, 0);
735         bio->bi_end_io = scrub_fixup_end_io;
736         bio->bi_private = &complete;
737         btrfsic_submit_bio(rw, bio);
738
739         /* this will also unplug the queue */
740         wait_for_completion(&complete);
741
742         ret = !test_bit(BIO_UPTODATE, &bio->bi_flags);
743         bio_put(bio);
744         return ret;
745 }
746
747 static void scrub_bio_end_io(struct bio *bio, int err)
748 {
749         struct scrub_bio *sbio = bio->bi_private;
750         struct scrub_dev *sdev = sbio->sdev;
751         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
752
753         sbio->err = err;
754         sbio->bio = bio;
755
756         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
757 }
758
759 static void scrub_checksum(struct btrfs_work *work)
760 {
761         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
762         struct scrub_dev *sdev = sbio->sdev;
763         struct page *page;
764         void *buffer;
765         int i;
766         u64 flags;
767         u64 logical;
768         int ret;
769
770         if (sbio->err) {
771                 ret = 0;
772                 for (i = 0; i < sbio->count; ++i)
773                         ret |= scrub_recheck_error(sbio, i);
774                 if (!ret) {
775                         spin_lock(&sdev->stat_lock);
776                         ++sdev->stat.unverified_errors;
777                         spin_unlock(&sdev->stat_lock);
778                 }
779
780                 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
781                 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
782                 sbio->bio->bi_phys_segments = 0;
783                 sbio->bio->bi_idx = 0;
784
785                 for (i = 0; i < sbio->count; i++) {
786                         struct bio_vec *bi;
787                         bi = &sbio->bio->bi_io_vec[i];
788                         bi->bv_offset = 0;
789                         bi->bv_len = PAGE_SIZE;
790                 }
791                 goto out;
792         }
793         for (i = 0; i < sbio->count; ++i) {
794                 page = sbio->bio->bi_io_vec[i].bv_page;
795                 buffer = kmap_atomic(page);
796                 flags = sbio->spag[i].flags;
797                 logical = sbio->logical + i * PAGE_SIZE;
798                 ret = 0;
799                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
800                         ret = scrub_checksum_data(sdev, sbio->spag + i, buffer);
801                 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
802                         ret = scrub_checksum_tree_block(sdev, sbio->spag + i,
803                                                         logical, buffer);
804                 } else if (flags & BTRFS_EXTENT_FLAG_SUPER) {
805                         BUG_ON(i);
806                         (void)scrub_checksum_super(sbio, buffer);
807                 } else {
808                         WARN_ON(1);
809                 }
810                 kunmap_atomic(buffer);
811                 if (ret) {
812                         ret = scrub_recheck_error(sbio, i);
813                         if (!ret) {
814                                 spin_lock(&sdev->stat_lock);
815                                 ++sdev->stat.unverified_errors;
816                                 spin_unlock(&sdev->stat_lock);
817                         }
818                 }
819         }
820
821 out:
822         scrub_free_bio(sbio->bio);
823         sbio->bio = NULL;
824         spin_lock(&sdev->list_lock);
825         sbio->next_free = sdev->first_free;
826         sdev->first_free = sbio->index;
827         spin_unlock(&sdev->list_lock);
828         atomic_dec(&sdev->in_flight);
829         wake_up(&sdev->list_wait);
830 }
831
832 static int scrub_checksum_data(struct scrub_dev *sdev,
833                                struct scrub_page *spag, void *buffer)
834 {
835         u8 csum[BTRFS_CSUM_SIZE];
836         u32 crc = ~(u32)0;
837         int fail = 0;
838         struct btrfs_root *root = sdev->dev->dev_root;
839
840         if (!spag->have_csum)
841                 return 0;
842
843         crc = btrfs_csum_data(root, buffer, crc, PAGE_SIZE);
844         btrfs_csum_final(crc, csum);
845         if (memcmp(csum, spag->csum, sdev->csum_size))
846                 fail = 1;
847
848         spin_lock(&sdev->stat_lock);
849         ++sdev->stat.data_extents_scrubbed;
850         sdev->stat.data_bytes_scrubbed += PAGE_SIZE;
851         if (fail)
852                 ++sdev->stat.csum_errors;
853         spin_unlock(&sdev->stat_lock);
854
855         return fail;
856 }
857
858 static int scrub_checksum_tree_block(struct scrub_dev *sdev,
859                                      struct scrub_page *spag, u64 logical,
860                                      void *buffer)
861 {
862         struct btrfs_header *h;
863         struct btrfs_root *root = sdev->dev->dev_root;
864         struct btrfs_fs_info *fs_info = root->fs_info;
865         u8 csum[BTRFS_CSUM_SIZE];
866         u32 crc = ~(u32)0;
867         int fail = 0;
868         int crc_fail = 0;
869
870         /*
871          * we don't use the getter functions here, as we
872          * a) don't have an extent buffer and
873          * b) the page is already kmapped
874          */
875         h = (struct btrfs_header *)buffer;
876
877         if (logical != le64_to_cpu(h->bytenr))
878                 ++fail;
879
880         if (spag->generation != le64_to_cpu(h->generation))
881                 ++fail;
882
883         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
884                 ++fail;
885
886         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
887                    BTRFS_UUID_SIZE))
888                 ++fail;
889
890         crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
891                               PAGE_SIZE - BTRFS_CSUM_SIZE);
892         btrfs_csum_final(crc, csum);
893         if (memcmp(csum, h->csum, sdev->csum_size))
894                 ++crc_fail;
895
896         spin_lock(&sdev->stat_lock);
897         ++sdev->stat.tree_extents_scrubbed;
898         sdev->stat.tree_bytes_scrubbed += PAGE_SIZE;
899         if (crc_fail)
900                 ++sdev->stat.csum_errors;
901         if (fail)
902                 ++sdev->stat.verify_errors;
903         spin_unlock(&sdev->stat_lock);
904
905         return fail || crc_fail;
906 }
907
908 static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer)
909 {
910         struct btrfs_super_block *s;
911         u64 logical;
912         struct scrub_dev *sdev = sbio->sdev;
913         struct btrfs_root *root = sdev->dev->dev_root;
914         struct btrfs_fs_info *fs_info = root->fs_info;
915         u8 csum[BTRFS_CSUM_SIZE];
916         u32 crc = ~(u32)0;
917         int fail = 0;
918
919         s = (struct btrfs_super_block *)buffer;
920         logical = sbio->logical;
921
922         if (logical != le64_to_cpu(s->bytenr))
923                 ++fail;
924
925         if (sbio->spag[0].generation != le64_to_cpu(s->generation))
926                 ++fail;
927
928         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
929                 ++fail;
930
931         crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
932                               PAGE_SIZE - BTRFS_CSUM_SIZE);
933         btrfs_csum_final(crc, csum);
934         if (memcmp(csum, s->csum, sbio->sdev->csum_size))
935                 ++fail;
936
937         if (fail) {
938                 /*
939                  * if we find an error in a super block, we just report it.
940                  * They will get written with the next transaction commit
941                  * anyway
942                  */
943                 spin_lock(&sdev->stat_lock);
944                 ++sdev->stat.super_errors;
945                 spin_unlock(&sdev->stat_lock);
946         }
947
948         return fail;
949 }
950
951 static int scrub_submit(struct scrub_dev *sdev)
952 {
953         struct scrub_bio *sbio;
954
955         if (sdev->curr == -1)
956                 return 0;
957
958         sbio = sdev->bios[sdev->curr];
959         sbio->err = 0;
960         sdev->curr = -1;
961         atomic_inc(&sdev->in_flight);
962
963         btrfsic_submit_bio(READ, sbio->bio);
964
965         return 0;
966 }
967
968 static int scrub_page(struct scrub_dev *sdev, u64 logical, u64 len,
969                       u64 physical, u64 flags, u64 gen, int mirror_num,
970                       u8 *csum, int force)
971 {
972         struct scrub_bio *sbio;
973         struct page *page;
974         int ret;
975
976 again:
977         /*
978          * grab a fresh bio or wait for one to become available
979          */
980         while (sdev->curr == -1) {
981                 spin_lock(&sdev->list_lock);
982                 sdev->curr = sdev->first_free;
983                 if (sdev->curr != -1) {
984                         sdev->first_free = sdev->bios[sdev->curr]->next_free;
985                         sdev->bios[sdev->curr]->next_free = -1;
986                         sdev->bios[sdev->curr]->count = 0;
987                         spin_unlock(&sdev->list_lock);
988                 } else {
989                         spin_unlock(&sdev->list_lock);
990                         wait_event(sdev->list_wait, sdev->first_free != -1);
991                 }
992         }
993         sbio = sdev->bios[sdev->curr];
994         if (sbio->count == 0) {
995                 struct bio *bio;
996
997                 sbio->physical = physical;
998                 sbio->logical = logical;
999                 bio = bio_alloc(GFP_NOFS, SCRUB_PAGES_PER_BIO);
1000                 if (!bio)
1001                         return -ENOMEM;
1002
1003                 bio->bi_private = sbio;
1004                 bio->bi_end_io = scrub_bio_end_io;
1005                 bio->bi_bdev = sdev->dev->bdev;
1006                 bio->bi_sector = sbio->physical >> 9;
1007                 sbio->err = 0;
1008                 sbio->bio = bio;
1009         } else if (sbio->physical + sbio->count * PAGE_SIZE != physical ||
1010                    sbio->logical + sbio->count * PAGE_SIZE != logical) {
1011                 ret = scrub_submit(sdev);
1012                 if (ret)
1013                         return ret;
1014                 goto again;
1015         }
1016         sbio->spag[sbio->count].flags = flags;
1017         sbio->spag[sbio->count].generation = gen;
1018         sbio->spag[sbio->count].have_csum = 0;
1019         sbio->spag[sbio->count].mirror_num = mirror_num;
1020
1021         page = alloc_page(GFP_NOFS);
1022         if (!page)
1023                 return -ENOMEM;
1024
1025         ret = bio_add_page(sbio->bio, page, PAGE_SIZE, 0);
1026         if (!ret) {
1027                 __free_page(page);
1028                 ret = scrub_submit(sdev);
1029                 if (ret)
1030                         return ret;
1031                 goto again;
1032         }
1033
1034         if (csum) {
1035                 sbio->spag[sbio->count].have_csum = 1;
1036                 memcpy(sbio->spag[sbio->count].csum, csum, sdev->csum_size);
1037         }
1038         ++sbio->count;
1039         if (sbio->count == SCRUB_PAGES_PER_BIO || force) {
1040                 int ret;
1041
1042                 ret = scrub_submit(sdev);
1043                 if (ret)
1044                         return ret;
1045         }
1046
1047         return 0;
1048 }
1049
1050 static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
1051                            u8 *csum)
1052 {
1053         struct btrfs_ordered_sum *sum = NULL;
1054         int ret = 0;
1055         unsigned long i;
1056         unsigned long num_sectors;
1057         u32 sectorsize = sdev->dev->dev_root->sectorsize;
1058
1059         while (!list_empty(&sdev->csum_list)) {
1060                 sum = list_first_entry(&sdev->csum_list,
1061                                        struct btrfs_ordered_sum, list);
1062                 if (sum->bytenr > logical)
1063                         return 0;
1064                 if (sum->bytenr + sum->len > logical)
1065                         break;
1066
1067                 ++sdev->stat.csum_discards;
1068                 list_del(&sum->list);
1069                 kfree(sum);
1070                 sum = NULL;
1071         }
1072         if (!sum)
1073                 return 0;
1074
1075         num_sectors = sum->len / sectorsize;
1076         for (i = 0; i < num_sectors; ++i) {
1077                 if (sum->sums[i].bytenr == logical) {
1078                         memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
1079                         ret = 1;
1080                         break;
1081                 }
1082         }
1083         if (ret && i == num_sectors - 1) {
1084                 list_del(&sum->list);
1085                 kfree(sum);
1086         }
1087         return ret;
1088 }
1089
1090 /* scrub extent tries to collect up to 64 kB for each bio */
1091 static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
1092                         u64 physical, u64 flags, u64 gen, int mirror_num)
1093 {
1094         int ret;
1095         u8 csum[BTRFS_CSUM_SIZE];
1096
1097         while (len) {
1098                 u64 l = min_t(u64, len, PAGE_SIZE);
1099                 int have_csum = 0;
1100
1101                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
1102                         /* push csums to sbio */
1103                         have_csum = scrub_find_csum(sdev, logical, l, csum);
1104                         if (have_csum == 0)
1105                                 ++sdev->stat.no_csum;
1106                 }
1107                 ret = scrub_page(sdev, logical, l, physical, flags, gen,
1108                                  mirror_num, have_csum ? csum : NULL, 0);
1109                 if (ret)
1110                         return ret;
1111                 len -= l;
1112                 logical += l;
1113                 physical += l;
1114         }
1115         return 0;
1116 }
1117
1118 static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
1119         struct map_lookup *map, int num, u64 base, u64 length)
1120 {
1121         struct btrfs_path *path;
1122         struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
1123         struct btrfs_root *root = fs_info->extent_root;
1124         struct btrfs_root *csum_root = fs_info->csum_root;
1125         struct btrfs_extent_item *extent;
1126         struct blk_plug plug;
1127         u64 flags;
1128         int ret;
1129         int slot;
1130         int i;
1131         u64 nstripes;
1132         struct extent_buffer *l;
1133         struct btrfs_key key;
1134         u64 physical;
1135         u64 logical;
1136         u64 generation;
1137         int mirror_num;
1138         struct reada_control *reada1;
1139         struct reada_control *reada2;
1140         struct btrfs_key key_start;
1141         struct btrfs_key key_end;
1142
1143         u64 increment = map->stripe_len;
1144         u64 offset;
1145
1146         nstripes = length;
1147         offset = 0;
1148         do_div(nstripes, map->stripe_len);
1149         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
1150                 offset = map->stripe_len * num;
1151                 increment = map->stripe_len * map->num_stripes;
1152                 mirror_num = 1;
1153         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
1154                 int factor = map->num_stripes / map->sub_stripes;
1155                 offset = map->stripe_len * (num / map->sub_stripes);
1156                 increment = map->stripe_len * factor;
1157                 mirror_num = num % map->sub_stripes + 1;
1158         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
1159                 increment = map->stripe_len;
1160                 mirror_num = num % map->num_stripes + 1;
1161         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
1162                 increment = map->stripe_len;
1163                 mirror_num = num % map->num_stripes + 1;
1164         } else {
1165                 increment = map->stripe_len;
1166                 mirror_num = 1;
1167         }
1168
1169         path = btrfs_alloc_path();
1170         if (!path)
1171                 return -ENOMEM;
1172
1173         path->search_commit_root = 1;
1174         path->skip_locking = 1;
1175
1176         /*
1177          * trigger the readahead for extent tree csum tree and wait for
1178          * completion. During readahead, the scrub is officially paused
1179          * to not hold off transaction commits
1180          */
1181         logical = base + offset;
1182
1183         wait_event(sdev->list_wait,
1184                    atomic_read(&sdev->in_flight) == 0);
1185         atomic_inc(&fs_info->scrubs_paused);
1186         wake_up(&fs_info->scrub_pause_wait);
1187
1188         /* FIXME it might be better to start readahead at commit root */
1189         key_start.objectid = logical;
1190         key_start.type = BTRFS_EXTENT_ITEM_KEY;
1191         key_start.offset = (u64)0;
1192         key_end.objectid = base + offset + nstripes * increment;
1193         key_end.type = BTRFS_EXTENT_ITEM_KEY;
1194         key_end.offset = (u64)0;
1195         reada1 = btrfs_reada_add(root, &key_start, &key_end);
1196
1197         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1198         key_start.type = BTRFS_EXTENT_CSUM_KEY;
1199         key_start.offset = logical;
1200         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
1201         key_end.type = BTRFS_EXTENT_CSUM_KEY;
1202         key_end.offset = base + offset + nstripes * increment;
1203         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
1204
1205         if (!IS_ERR(reada1))
1206                 btrfs_reada_wait(reada1);
1207         if (!IS_ERR(reada2))
1208                 btrfs_reada_wait(reada2);
1209
1210         mutex_lock(&fs_info->scrub_lock);
1211         while (atomic_read(&fs_info->scrub_pause_req)) {
1212                 mutex_unlock(&fs_info->scrub_lock);
1213                 wait_event(fs_info->scrub_pause_wait,
1214                    atomic_read(&fs_info->scrub_pause_req) == 0);
1215                 mutex_lock(&fs_info->scrub_lock);
1216         }
1217         atomic_dec(&fs_info->scrubs_paused);
1218         mutex_unlock(&fs_info->scrub_lock);
1219         wake_up(&fs_info->scrub_pause_wait);
1220
1221         /*
1222          * collect all data csums for the stripe to avoid seeking during
1223          * the scrub. This might currently (crc32) end up to be about 1MB
1224          */
1225         blk_start_plug(&plug);
1226
1227         /*
1228          * now find all extents for each stripe and scrub them
1229          */
1230         logical = base + offset;
1231         physical = map->stripes[num].physical;
1232         ret = 0;
1233         for (i = 0; i < nstripes; ++i) {
1234                 /*
1235                  * canceled?
1236                  */
1237                 if (atomic_read(&fs_info->scrub_cancel_req) ||
1238                     atomic_read(&sdev->cancel_req)) {
1239                         ret = -ECANCELED;
1240                         goto out;
1241                 }
1242                 /*
1243                  * check to see if we have to pause
1244                  */
1245                 if (atomic_read(&fs_info->scrub_pause_req)) {
1246                         /* push queued extents */
1247                         scrub_submit(sdev);
1248                         wait_event(sdev->list_wait,
1249                                    atomic_read(&sdev->in_flight) == 0);
1250                         atomic_inc(&fs_info->scrubs_paused);
1251                         wake_up(&fs_info->scrub_pause_wait);
1252                         mutex_lock(&fs_info->scrub_lock);
1253                         while (atomic_read(&fs_info->scrub_pause_req)) {
1254                                 mutex_unlock(&fs_info->scrub_lock);
1255                                 wait_event(fs_info->scrub_pause_wait,
1256                                    atomic_read(&fs_info->scrub_pause_req) == 0);
1257                                 mutex_lock(&fs_info->scrub_lock);
1258                         }
1259                         atomic_dec(&fs_info->scrubs_paused);
1260                         mutex_unlock(&fs_info->scrub_lock);
1261                         wake_up(&fs_info->scrub_pause_wait);
1262                 }
1263
1264                 ret = btrfs_lookup_csums_range(csum_root, logical,
1265                                                logical + map->stripe_len - 1,
1266                                                &sdev->csum_list, 1);
1267                 if (ret)
1268                         goto out;
1269
1270                 key.objectid = logical;
1271                 key.type = BTRFS_EXTENT_ITEM_KEY;
1272                 key.offset = (u64)0;
1273
1274                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1275                 if (ret < 0)
1276                         goto out;
1277                 if (ret > 0) {
1278                         ret = btrfs_previous_item(root, path, 0,
1279                                                   BTRFS_EXTENT_ITEM_KEY);
1280                         if (ret < 0)
1281                                 goto out;
1282                         if (ret > 0) {
1283                                 /* there's no smaller item, so stick with the
1284                                  * larger one */
1285                                 btrfs_release_path(path);
1286                                 ret = btrfs_search_slot(NULL, root, &key,
1287                                                         path, 0, 0);
1288                                 if (ret < 0)
1289                                         goto out;
1290                         }
1291                 }
1292
1293                 while (1) {
1294                         l = path->nodes[0];
1295                         slot = path->slots[0];
1296                         if (slot >= btrfs_header_nritems(l)) {
1297                                 ret = btrfs_next_leaf(root, path);
1298                                 if (ret == 0)
1299                                         continue;
1300                                 if (ret < 0)
1301                                         goto out;
1302
1303                                 break;
1304                         }
1305                         btrfs_item_key_to_cpu(l, &key, slot);
1306
1307                         if (key.objectid + key.offset <= logical)
1308                                 goto next;
1309
1310                         if (key.objectid >= logical + map->stripe_len)
1311                                 break;
1312
1313                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
1314                                 goto next;
1315
1316                         extent = btrfs_item_ptr(l, slot,
1317                                                 struct btrfs_extent_item);
1318                         flags = btrfs_extent_flags(l, extent);
1319                         generation = btrfs_extent_generation(l, extent);
1320
1321                         if (key.objectid < logical &&
1322                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
1323                                 printk(KERN_ERR
1324                                        "btrfs scrub: tree block %llu spanning "
1325                                        "stripes, ignored. logical=%llu\n",
1326                                        (unsigned long long)key.objectid,
1327                                        (unsigned long long)logical);
1328                                 goto next;
1329                         }
1330
1331                         /*
1332                          * trim extent to this stripe
1333                          */
1334                         if (key.objectid < logical) {
1335                                 key.offset -= logical - key.objectid;
1336                                 key.objectid = logical;
1337                         }
1338                         if (key.objectid + key.offset >
1339                             logical + map->stripe_len) {
1340                                 key.offset = logical + map->stripe_len -
1341                                              key.objectid;
1342                         }
1343
1344                         ret = scrub_extent(sdev, key.objectid, key.offset,
1345                                            key.objectid - logical + physical,
1346                                            flags, generation, mirror_num);
1347                         if (ret)
1348                                 goto out;
1349
1350 next:
1351                         path->slots[0]++;
1352                 }
1353                 btrfs_release_path(path);
1354                 logical += increment;
1355                 physical += map->stripe_len;
1356                 spin_lock(&sdev->stat_lock);
1357                 sdev->stat.last_physical = physical;
1358                 spin_unlock(&sdev->stat_lock);
1359         }
1360         /* push queued extents */
1361         scrub_submit(sdev);
1362
1363 out:
1364         blk_finish_plug(&plug);
1365         btrfs_free_path(path);
1366         return ret < 0 ? ret : 0;
1367 }
1368
1369 static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
1370         u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length,
1371         u64 dev_offset)
1372 {
1373         struct btrfs_mapping_tree *map_tree =
1374                 &sdev->dev->dev_root->fs_info->mapping_tree;
1375         struct map_lookup *map;
1376         struct extent_map *em;
1377         int i;
1378         int ret = -EINVAL;
1379
1380         read_lock(&map_tree->map_tree.lock);
1381         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
1382         read_unlock(&map_tree->map_tree.lock);
1383
1384         if (!em)
1385                 return -EINVAL;
1386
1387         map = (struct map_lookup *)em->bdev;
1388         if (em->start != chunk_offset)
1389                 goto out;
1390
1391         if (em->len < length)
1392                 goto out;
1393
1394         for (i = 0; i < map->num_stripes; ++i) {
1395                 if (map->stripes[i].dev == sdev->dev &&
1396                     map->stripes[i].physical == dev_offset) {
1397                         ret = scrub_stripe(sdev, map, i, chunk_offset, length);
1398                         if (ret)
1399                                 goto out;
1400                 }
1401         }
1402 out:
1403         free_extent_map(em);
1404
1405         return ret;
1406 }
1407
1408 static noinline_for_stack
1409 int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
1410 {
1411         struct btrfs_dev_extent *dev_extent = NULL;
1412         struct btrfs_path *path;
1413         struct btrfs_root *root = sdev->dev->dev_root;
1414         struct btrfs_fs_info *fs_info = root->fs_info;
1415         u64 length;
1416         u64 chunk_tree;
1417         u64 chunk_objectid;
1418         u64 chunk_offset;
1419         int ret;
1420         int slot;
1421         struct extent_buffer *l;
1422         struct btrfs_key key;
1423         struct btrfs_key found_key;
1424         struct btrfs_block_group_cache *cache;
1425
1426         path = btrfs_alloc_path();
1427         if (!path)
1428                 return -ENOMEM;
1429
1430         path->reada = 2;
1431         path->search_commit_root = 1;
1432         path->skip_locking = 1;
1433
1434         key.objectid = sdev->dev->devid;
1435         key.offset = 0ull;
1436         key.type = BTRFS_DEV_EXTENT_KEY;
1437
1438
1439         while (1) {
1440                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1441                 if (ret < 0)
1442                         break;
1443                 if (ret > 0) {
1444                         if (path->slots[0] >=
1445                             btrfs_header_nritems(path->nodes[0])) {
1446                                 ret = btrfs_next_leaf(root, path);
1447                                 if (ret)
1448                                         break;
1449                         }
1450                 }
1451
1452                 l = path->nodes[0];
1453                 slot = path->slots[0];
1454
1455                 btrfs_item_key_to_cpu(l, &found_key, slot);
1456
1457                 if (found_key.objectid != sdev->dev->devid)
1458                         break;
1459
1460                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
1461                         break;
1462
1463                 if (found_key.offset >= end)
1464                         break;
1465
1466                 if (found_key.offset < key.offset)
1467                         break;
1468
1469                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1470                 length = btrfs_dev_extent_length(l, dev_extent);
1471
1472                 if (found_key.offset + length <= start) {
1473                         key.offset = found_key.offset + length;
1474                         btrfs_release_path(path);
1475                         continue;
1476                 }
1477
1478                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
1479                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
1480                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
1481
1482                 /*
1483                  * get a reference on the corresponding block group to prevent
1484                  * the chunk from going away while we scrub it
1485                  */
1486                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
1487                 if (!cache) {
1488                         ret = -ENOENT;
1489                         break;
1490                 }
1491                 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
1492                                   chunk_offset, length, found_key.offset);
1493                 btrfs_put_block_group(cache);
1494                 if (ret)
1495                         break;
1496
1497                 key.offset = found_key.offset + length;
1498                 btrfs_release_path(path);
1499         }
1500
1501         btrfs_free_path(path);
1502
1503         /*
1504          * ret can still be 1 from search_slot or next_leaf,
1505          * that's not an error
1506          */
1507         return ret < 0 ? ret : 0;
1508 }
1509
1510 static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
1511 {
1512         int     i;
1513         u64     bytenr;
1514         u64     gen;
1515         int     ret;
1516         struct btrfs_device *device = sdev->dev;
1517         struct btrfs_root *root = device->dev_root;
1518
1519         gen = root->fs_info->last_trans_committed;
1520
1521         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1522                 bytenr = btrfs_sb_offset(i);
1523                 if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes)
1524                         break;
1525
1526                 ret = scrub_page(sdev, bytenr, PAGE_SIZE, bytenr,
1527                                  BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
1528                 if (ret)
1529                         return ret;
1530         }
1531         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1532
1533         return 0;
1534 }
1535
1536 /*
1537  * get a reference count on fs_info->scrub_workers. start worker if necessary
1538  */
1539 static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
1540 {
1541         struct btrfs_fs_info *fs_info = root->fs_info;
1542         int ret = 0;
1543
1544         mutex_lock(&fs_info->scrub_lock);
1545         if (fs_info->scrub_workers_refcnt == 0) {
1546                 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
1547                            fs_info->thread_pool_size, &fs_info->generic_worker);
1548                 fs_info->scrub_workers.idle_thresh = 4;
1549                 ret = btrfs_start_workers(&fs_info->scrub_workers);
1550                 if (ret)
1551                         goto out;
1552         }
1553         ++fs_info->scrub_workers_refcnt;
1554 out:
1555         mutex_unlock(&fs_info->scrub_lock);
1556
1557         return ret;
1558 }
1559
1560 static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
1561 {
1562         struct btrfs_fs_info *fs_info = root->fs_info;
1563
1564         mutex_lock(&fs_info->scrub_lock);
1565         if (--fs_info->scrub_workers_refcnt == 0)
1566                 btrfs_stop_workers(&fs_info->scrub_workers);
1567         WARN_ON(fs_info->scrub_workers_refcnt < 0);
1568         mutex_unlock(&fs_info->scrub_lock);
1569 }
1570
1571
1572 int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
1573                     struct btrfs_scrub_progress *progress, int readonly)
1574 {
1575         struct scrub_dev *sdev;
1576         struct btrfs_fs_info *fs_info = root->fs_info;
1577         int ret;
1578         struct btrfs_device *dev;
1579
1580         if (btrfs_fs_closing(root->fs_info))
1581                 return -EINVAL;
1582
1583         /*
1584          * check some assumptions
1585          */
1586         if (root->sectorsize != PAGE_SIZE ||
1587             root->sectorsize != root->leafsize ||
1588             root->sectorsize != root->nodesize) {
1589                 printk(KERN_ERR "btrfs_scrub: size assumptions fail\n");
1590                 return -EINVAL;
1591         }
1592
1593         ret = scrub_workers_get(root);
1594         if (ret)
1595                 return ret;
1596
1597         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1598         dev = btrfs_find_device(root, devid, NULL, NULL);
1599         if (!dev || dev->missing) {
1600                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1601                 scrub_workers_put(root);
1602                 return -ENODEV;
1603         }
1604         mutex_lock(&fs_info->scrub_lock);
1605
1606         if (!dev->in_fs_metadata) {
1607                 mutex_unlock(&fs_info->scrub_lock);
1608                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1609                 scrub_workers_put(root);
1610                 return -ENODEV;
1611         }
1612
1613         if (dev->scrub_device) {
1614                 mutex_unlock(&fs_info->scrub_lock);
1615                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1616                 scrub_workers_put(root);
1617                 return -EINPROGRESS;
1618         }
1619         sdev = scrub_setup_dev(dev);
1620         if (IS_ERR(sdev)) {
1621                 mutex_unlock(&fs_info->scrub_lock);
1622                 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1623                 scrub_workers_put(root);
1624                 return PTR_ERR(sdev);
1625         }
1626         sdev->readonly = readonly;
1627         dev->scrub_device = sdev;
1628
1629         atomic_inc(&fs_info->scrubs_running);
1630         mutex_unlock(&fs_info->scrub_lock);
1631         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1632
1633         down_read(&fs_info->scrub_super_lock);
1634         ret = scrub_supers(sdev);
1635         up_read(&fs_info->scrub_super_lock);
1636
1637         if (!ret)
1638                 ret = scrub_enumerate_chunks(sdev, start, end);
1639
1640         wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1641         atomic_dec(&fs_info->scrubs_running);
1642         wake_up(&fs_info->scrub_pause_wait);
1643
1644         wait_event(sdev->list_wait, atomic_read(&sdev->fixup_cnt) == 0);
1645
1646         if (progress)
1647                 memcpy(progress, &sdev->stat, sizeof(*progress));
1648
1649         mutex_lock(&fs_info->scrub_lock);
1650         dev->scrub_device = NULL;
1651         mutex_unlock(&fs_info->scrub_lock);
1652
1653         scrub_free_dev(sdev);
1654         scrub_workers_put(root);
1655
1656         return ret;
1657 }
1658
1659 int btrfs_scrub_pause(struct btrfs_root *root)
1660 {
1661         struct btrfs_fs_info *fs_info = root->fs_info;
1662
1663         mutex_lock(&fs_info->scrub_lock);
1664         atomic_inc(&fs_info->scrub_pause_req);
1665         while (atomic_read(&fs_info->scrubs_paused) !=
1666                atomic_read(&fs_info->scrubs_running)) {
1667                 mutex_unlock(&fs_info->scrub_lock);
1668                 wait_event(fs_info->scrub_pause_wait,
1669                            atomic_read(&fs_info->scrubs_paused) ==
1670                            atomic_read(&fs_info->scrubs_running));
1671                 mutex_lock(&fs_info->scrub_lock);
1672         }
1673         mutex_unlock(&fs_info->scrub_lock);
1674
1675         return 0;
1676 }
1677
1678 int btrfs_scrub_continue(struct btrfs_root *root)
1679 {
1680         struct btrfs_fs_info *fs_info = root->fs_info;
1681
1682         atomic_dec(&fs_info->scrub_pause_req);
1683         wake_up(&fs_info->scrub_pause_wait);
1684         return 0;
1685 }
1686
1687 int btrfs_scrub_pause_super(struct btrfs_root *root)
1688 {
1689         down_write(&root->fs_info->scrub_super_lock);
1690         return 0;
1691 }
1692
1693 int btrfs_scrub_continue_super(struct btrfs_root *root)
1694 {
1695         up_write(&root->fs_info->scrub_super_lock);
1696         return 0;
1697 }
1698
1699 int btrfs_scrub_cancel(struct btrfs_root *root)
1700 {
1701         struct btrfs_fs_info *fs_info = root->fs_info;
1702
1703         mutex_lock(&fs_info->scrub_lock);
1704         if (!atomic_read(&fs_info->scrubs_running)) {
1705                 mutex_unlock(&fs_info->scrub_lock);
1706                 return -ENOTCONN;
1707         }
1708
1709         atomic_inc(&fs_info->scrub_cancel_req);
1710         while (atomic_read(&fs_info->scrubs_running)) {
1711                 mutex_unlock(&fs_info->scrub_lock);
1712                 wait_event(fs_info->scrub_pause_wait,
1713                            atomic_read(&fs_info->scrubs_running) == 0);
1714                 mutex_lock(&fs_info->scrub_lock);
1715         }
1716         atomic_dec(&fs_info->scrub_cancel_req);
1717         mutex_unlock(&fs_info->scrub_lock);
1718
1719         return 0;
1720 }
1721
1722 int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
1723 {
1724         struct btrfs_fs_info *fs_info = root->fs_info;
1725         struct scrub_dev *sdev;
1726
1727         mutex_lock(&fs_info->scrub_lock);
1728         sdev = dev->scrub_device;
1729         if (!sdev) {
1730                 mutex_unlock(&fs_info->scrub_lock);
1731                 return -ENOTCONN;
1732         }
1733         atomic_inc(&sdev->cancel_req);
1734         while (dev->scrub_device) {
1735                 mutex_unlock(&fs_info->scrub_lock);
1736                 wait_event(fs_info->scrub_pause_wait,
1737                            dev->scrub_device == NULL);
1738                 mutex_lock(&fs_info->scrub_lock);
1739         }
1740         mutex_unlock(&fs_info->scrub_lock);
1741
1742         return 0;
1743 }
1744 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
1745 {
1746         struct btrfs_fs_info *fs_info = root->fs_info;
1747         struct btrfs_device *dev;
1748         int ret;
1749
1750         /*
1751          * we have to hold the device_list_mutex here so the device
1752          * does not go away in cancel_dev. FIXME: find a better solution
1753          */
1754         mutex_lock(&fs_info->fs_devices->device_list_mutex);
1755         dev = btrfs_find_device(root, devid, NULL, NULL);
1756         if (!dev) {
1757                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1758                 return -ENODEV;
1759         }
1760         ret = btrfs_scrub_cancel_dev(root, dev);
1761         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1762
1763         return ret;
1764 }
1765
1766 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
1767                          struct btrfs_scrub_progress *progress)
1768 {
1769         struct btrfs_device *dev;
1770         struct scrub_dev *sdev = NULL;
1771
1772         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1773         dev = btrfs_find_device(root, devid, NULL, NULL);
1774         if (dev)
1775                 sdev = dev->scrub_device;
1776         if (sdev)
1777                 memcpy(progress, &sdev->stat, sizeof(*progress));
1778         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1779
1780         return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
1781 }