629313732521f4b25ca0966442c4d85d466ed239
[sfrench/cifs-2.6.git] / fs / btrfs / scrub.c
1 /*
2  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/blkdev.h>
20 #include <linux/ratelimit.h>
21 #include <linux/sched/mm.h>
22 #include "ctree.h"
23 #include "volumes.h"
24 #include "disk-io.h"
25 #include "ordered-data.h"
26 #include "transaction.h"
27 #include "backref.h"
28 #include "extent_io.h"
29 #include "dev-replace.h"
30 #include "check-integrity.h"
31 #include "rcu-string.h"
32 #include "raid56.h"
33
34 /*
35  * This is only the first step towards a full-features scrub. It reads all
36  * extent and super block and verifies the checksums. In case a bad checksum
37  * is found or the extent cannot be read, good data will be written back if
38  * any can be found.
39  *
40  * Future enhancements:
41  *  - In case an unrepairable extent is encountered, track which files are
42  *    affected and report them
43  *  - track and record media errors, throw out bad devices
44  *  - add a mode to also read unallocated space
45  */
46
47 struct scrub_block;
48 struct scrub_ctx;
49
50 /*
51  * the following three values only influence the performance.
52  * The last one configures the number of parallel and outstanding I/O
53  * operations. The first two values configure an upper limit for the number
54  * of (dynamically allocated) pages that are added to a bio.
55  */
56 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
57 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
58 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
59
60 /*
61  * the following value times PAGE_SIZE needs to be large enough to match the
62  * largest node/leaf/sector size that shall be supported.
63  * Values larger than BTRFS_STRIPE_LEN are not supported.
64  */
65 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
66
67 struct scrub_recover {
68         refcount_t              refs;
69         struct btrfs_bio        *bbio;
70         u64                     map_length;
71 };
72
73 struct scrub_page {
74         struct scrub_block      *sblock;
75         struct page             *page;
76         struct btrfs_device     *dev;
77         struct list_head        list;
78         u64                     flags;  /* extent flags */
79         u64                     generation;
80         u64                     logical;
81         u64                     physical;
82         u64                     physical_for_dev_replace;
83         atomic_t                refs;
84         struct {
85                 unsigned int    mirror_num:8;
86                 unsigned int    have_csum:1;
87                 unsigned int    io_error:1;
88         };
89         u8                      csum[BTRFS_CSUM_SIZE];
90
91         struct scrub_recover    *recover;
92 };
93
94 struct scrub_bio {
95         int                     index;
96         struct scrub_ctx        *sctx;
97         struct btrfs_device     *dev;
98         struct bio              *bio;
99         blk_status_t            status;
100         u64                     logical;
101         u64                     physical;
102 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
103         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
104 #else
105         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
106 #endif
107         int                     page_count;
108         int                     next_free;
109         struct btrfs_work       work;
110 };
111
112 struct scrub_block {
113         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
114         int                     page_count;
115         atomic_t                outstanding_pages;
116         refcount_t              refs; /* free mem on transition to zero */
117         struct scrub_ctx        *sctx;
118         struct scrub_parity     *sparity;
119         struct {
120                 unsigned int    header_error:1;
121                 unsigned int    checksum_error:1;
122                 unsigned int    no_io_error_seen:1;
123                 unsigned int    generation_error:1; /* also sets header_error */
124
125                 /* The following is for the data used to check parity */
126                 /* It is for the data with checksum */
127                 unsigned int    data_corrected:1;
128         };
129         struct btrfs_work       work;
130 };
131
132 /* Used for the chunks with parity stripe such RAID5/6 */
133 struct scrub_parity {
134         struct scrub_ctx        *sctx;
135
136         struct btrfs_device     *scrub_dev;
137
138         u64                     logic_start;
139
140         u64                     logic_end;
141
142         int                     nsectors;
143
144         u64                     stripe_len;
145
146         refcount_t              refs;
147
148         struct list_head        spages;
149
150         /* Work of parity check and repair */
151         struct btrfs_work       work;
152
153         /* Mark the parity blocks which have data */
154         unsigned long           *dbitmap;
155
156         /*
157          * Mark the parity blocks which have data, but errors happen when
158          * read data or check data
159          */
160         unsigned long           *ebitmap;
161
162         unsigned long           bitmap[0];
163 };
164
165 struct scrub_ctx {
166         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
167         struct btrfs_fs_info    *fs_info;
168         int                     first_free;
169         int                     curr;
170         atomic_t                bios_in_flight;
171         atomic_t                workers_pending;
172         spinlock_t              list_lock;
173         wait_queue_head_t       list_wait;
174         u16                     csum_size;
175         struct list_head        csum_list;
176         atomic_t                cancel_req;
177         int                     readonly;
178         int                     pages_per_rd_bio;
179
180         int                     is_dev_replace;
181
182         struct scrub_bio        *wr_curr_bio;
183         struct mutex            wr_lock;
184         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
185         struct btrfs_device     *wr_tgtdev;
186         bool                    flush_all_writes;
187
188         /*
189          * statistics
190          */
191         struct btrfs_scrub_progress stat;
192         spinlock_t              stat_lock;
193
194         /*
195          * Use a ref counter to avoid use-after-free issues. Scrub workers
196          * decrement bios_in_flight and workers_pending and then do a wakeup
197          * on the list_wait wait queue. We must ensure the main scrub task
198          * doesn't free the scrub context before or while the workers are
199          * doing the wakeup() call.
200          */
201         refcount_t              refs;
202 };
203
204 struct scrub_fixup_nodatasum {
205         struct scrub_ctx        *sctx;
206         struct btrfs_device     *dev;
207         u64                     logical;
208         struct btrfs_root       *root;
209         struct btrfs_work       work;
210         int                     mirror_num;
211 };
212
213 struct scrub_nocow_inode {
214         u64                     inum;
215         u64                     offset;
216         u64                     root;
217         struct list_head        list;
218 };
219
220 struct scrub_copy_nocow_ctx {
221         struct scrub_ctx        *sctx;
222         u64                     logical;
223         u64                     len;
224         int                     mirror_num;
225         u64                     physical_for_dev_replace;
226         struct list_head        inodes;
227         struct btrfs_work       work;
228 };
229
230 struct scrub_warning {
231         struct btrfs_path       *path;
232         u64                     extent_item_size;
233         const char              *errstr;
234         u64                     physical;
235         u64                     logical;
236         struct btrfs_device     *dev;
237 };
238
239 struct full_stripe_lock {
240         struct rb_node node;
241         u64 logical;
242         u64 refs;
243         struct mutex mutex;
244 };
245
246 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
247 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
248 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
249 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
250 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
251 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
252                                      struct scrub_block *sblocks_for_recheck);
253 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
254                                 struct scrub_block *sblock,
255                                 int retry_failed_mirror);
256 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
257 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
258                                              struct scrub_block *sblock_good);
259 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
260                                             struct scrub_block *sblock_good,
261                                             int page_num, int force_write);
262 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
263 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
264                                            int page_num);
265 static int scrub_checksum_data(struct scrub_block *sblock);
266 static int scrub_checksum_tree_block(struct scrub_block *sblock);
267 static int scrub_checksum_super(struct scrub_block *sblock);
268 static void scrub_block_get(struct scrub_block *sblock);
269 static void scrub_block_put(struct scrub_block *sblock);
270 static void scrub_page_get(struct scrub_page *spage);
271 static void scrub_page_put(struct scrub_page *spage);
272 static void scrub_parity_get(struct scrub_parity *sparity);
273 static void scrub_parity_put(struct scrub_parity *sparity);
274 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
275                                     struct scrub_page *spage);
276 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
277                        u64 physical, struct btrfs_device *dev, u64 flags,
278                        u64 gen, int mirror_num, u8 *csum, int force,
279                        u64 physical_for_dev_replace);
280 static void scrub_bio_end_io(struct bio *bio);
281 static void scrub_bio_end_io_worker(struct btrfs_work *work);
282 static void scrub_block_complete(struct scrub_block *sblock);
283 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
284                                u64 extent_logical, u64 extent_len,
285                                u64 *extent_physical,
286                                struct btrfs_device **extent_dev,
287                                int *extent_mirror_num);
288 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
289                                     struct scrub_page *spage);
290 static void scrub_wr_submit(struct scrub_ctx *sctx);
291 static void scrub_wr_bio_end_io(struct bio *bio);
292 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
293 static int write_page_nocow(struct scrub_ctx *sctx,
294                             u64 physical_for_dev_replace, struct page *page);
295 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
296                                       struct scrub_copy_nocow_ctx *ctx);
297 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
298                             int mirror_num, u64 physical_for_dev_replace);
299 static void copy_nocow_pages_worker(struct btrfs_work *work);
300 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
301 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
302 static void scrub_put_ctx(struct scrub_ctx *sctx);
303
304 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
305 {
306         return page->recover &&
307                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
308 }
309
310 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
311 {
312         refcount_inc(&sctx->refs);
313         atomic_inc(&sctx->bios_in_flight);
314 }
315
316 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
317 {
318         atomic_dec(&sctx->bios_in_flight);
319         wake_up(&sctx->list_wait);
320         scrub_put_ctx(sctx);
321 }
322
323 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
324 {
325         while (atomic_read(&fs_info->scrub_pause_req)) {
326                 mutex_unlock(&fs_info->scrub_lock);
327                 wait_event(fs_info->scrub_pause_wait,
328                    atomic_read(&fs_info->scrub_pause_req) == 0);
329                 mutex_lock(&fs_info->scrub_lock);
330         }
331 }
332
333 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
334 {
335         atomic_inc(&fs_info->scrubs_paused);
336         wake_up(&fs_info->scrub_pause_wait);
337 }
338
339 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
340 {
341         mutex_lock(&fs_info->scrub_lock);
342         __scrub_blocked_if_needed(fs_info);
343         atomic_dec(&fs_info->scrubs_paused);
344         mutex_unlock(&fs_info->scrub_lock);
345
346         wake_up(&fs_info->scrub_pause_wait);
347 }
348
349 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
350 {
351         scrub_pause_on(fs_info);
352         scrub_pause_off(fs_info);
353 }
354
355 /*
356  * Insert new full stripe lock into full stripe locks tree
357  *
358  * Return pointer to existing or newly inserted full_stripe_lock structure if
359  * everything works well.
360  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
361  *
362  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
363  * function
364  */
365 static struct full_stripe_lock *insert_full_stripe_lock(
366                 struct btrfs_full_stripe_locks_tree *locks_root,
367                 u64 fstripe_logical)
368 {
369         struct rb_node **p;
370         struct rb_node *parent = NULL;
371         struct full_stripe_lock *entry;
372         struct full_stripe_lock *ret;
373
374         WARN_ON(!mutex_is_locked(&locks_root->lock));
375
376         p = &locks_root->root.rb_node;
377         while (*p) {
378                 parent = *p;
379                 entry = rb_entry(parent, struct full_stripe_lock, node);
380                 if (fstripe_logical < entry->logical) {
381                         p = &(*p)->rb_left;
382                 } else if (fstripe_logical > entry->logical) {
383                         p = &(*p)->rb_right;
384                 } else {
385                         entry->refs++;
386                         return entry;
387                 }
388         }
389
390         /* Insert new lock */
391         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
392         if (!ret)
393                 return ERR_PTR(-ENOMEM);
394         ret->logical = fstripe_logical;
395         ret->refs = 1;
396         mutex_init(&ret->mutex);
397
398         rb_link_node(&ret->node, parent, p);
399         rb_insert_color(&ret->node, &locks_root->root);
400         return ret;
401 }
402
403 /*
404  * Search for a full stripe lock of a block group
405  *
406  * Return pointer to existing full stripe lock if found
407  * Return NULL if not found
408  */
409 static struct full_stripe_lock *search_full_stripe_lock(
410                 struct btrfs_full_stripe_locks_tree *locks_root,
411                 u64 fstripe_logical)
412 {
413         struct rb_node *node;
414         struct full_stripe_lock *entry;
415
416         WARN_ON(!mutex_is_locked(&locks_root->lock));
417
418         node = locks_root->root.rb_node;
419         while (node) {
420                 entry = rb_entry(node, struct full_stripe_lock, node);
421                 if (fstripe_logical < entry->logical)
422                         node = node->rb_left;
423                 else if (fstripe_logical > entry->logical)
424                         node = node->rb_right;
425                 else
426                         return entry;
427         }
428         return NULL;
429 }
430
431 /*
432  * Helper to get full stripe logical from a normal bytenr.
433  *
434  * Caller must ensure @cache is a RAID56 block group.
435  */
436 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
437                                    u64 bytenr)
438 {
439         u64 ret;
440
441         /*
442          * Due to chunk item size limit, full stripe length should not be
443          * larger than U32_MAX. Just a sanity check here.
444          */
445         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
446
447         /*
448          * round_down() can only handle power of 2, while RAID56 full
449          * stripe length can be 64KiB * n, so we need to manually round down.
450          */
451         ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
452                 cache->full_stripe_len + cache->key.objectid;
453         return ret;
454 }
455
456 /*
457  * Lock a full stripe to avoid concurrency of recovery and read
458  *
459  * It's only used for profiles with parities (RAID5/6), for other profiles it
460  * does nothing.
461  *
462  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
463  * So caller must call unlock_full_stripe() at the same context.
464  *
465  * Return <0 if encounters error.
466  */
467 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
468                             bool *locked_ret)
469 {
470         struct btrfs_block_group_cache *bg_cache;
471         struct btrfs_full_stripe_locks_tree *locks_root;
472         struct full_stripe_lock *existing;
473         u64 fstripe_start;
474         int ret = 0;
475
476         *locked_ret = false;
477         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
478         if (!bg_cache) {
479                 ASSERT(0);
480                 return -ENOENT;
481         }
482
483         /* Profiles not based on parity don't need full stripe lock */
484         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
485                 goto out;
486         locks_root = &bg_cache->full_stripe_locks_root;
487
488         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
489
490         /* Now insert the full stripe lock */
491         mutex_lock(&locks_root->lock);
492         existing = insert_full_stripe_lock(locks_root, fstripe_start);
493         mutex_unlock(&locks_root->lock);
494         if (IS_ERR(existing)) {
495                 ret = PTR_ERR(existing);
496                 goto out;
497         }
498         mutex_lock(&existing->mutex);
499         *locked_ret = true;
500 out:
501         btrfs_put_block_group(bg_cache);
502         return ret;
503 }
504
505 /*
506  * Unlock a full stripe.
507  *
508  * NOTE: Caller must ensure it's the same context calling corresponding
509  * lock_full_stripe().
510  *
511  * Return 0 if we unlock full stripe without problem.
512  * Return <0 for error
513  */
514 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
515                               bool locked)
516 {
517         struct btrfs_block_group_cache *bg_cache;
518         struct btrfs_full_stripe_locks_tree *locks_root;
519         struct full_stripe_lock *fstripe_lock;
520         u64 fstripe_start;
521         bool freeit = false;
522         int ret = 0;
523
524         /* If we didn't acquire full stripe lock, no need to continue */
525         if (!locked)
526                 return 0;
527
528         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
529         if (!bg_cache) {
530                 ASSERT(0);
531                 return -ENOENT;
532         }
533         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
534                 goto out;
535
536         locks_root = &bg_cache->full_stripe_locks_root;
537         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
538
539         mutex_lock(&locks_root->lock);
540         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
541         /* Unpaired unlock_full_stripe() detected */
542         if (!fstripe_lock) {
543                 WARN_ON(1);
544                 ret = -ENOENT;
545                 mutex_unlock(&locks_root->lock);
546                 goto out;
547         }
548
549         if (fstripe_lock->refs == 0) {
550                 WARN_ON(1);
551                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
552                         fstripe_lock->logical);
553         } else {
554                 fstripe_lock->refs--;
555         }
556
557         if (fstripe_lock->refs == 0) {
558                 rb_erase(&fstripe_lock->node, &locks_root->root);
559                 freeit = true;
560         }
561         mutex_unlock(&locks_root->lock);
562
563         mutex_unlock(&fstripe_lock->mutex);
564         if (freeit)
565                 kfree(fstripe_lock);
566 out:
567         btrfs_put_block_group(bg_cache);
568         return ret;
569 }
570
571 /*
572  * used for workers that require transaction commits (i.e., for the
573  * NOCOW case)
574  */
575 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
576 {
577         struct btrfs_fs_info *fs_info = sctx->fs_info;
578
579         refcount_inc(&sctx->refs);
580         /*
581          * increment scrubs_running to prevent cancel requests from
582          * completing as long as a worker is running. we must also
583          * increment scrubs_paused to prevent deadlocking on pause
584          * requests used for transactions commits (as the worker uses a
585          * transaction context). it is safe to regard the worker
586          * as paused for all matters practical. effectively, we only
587          * avoid cancellation requests from completing.
588          */
589         mutex_lock(&fs_info->scrub_lock);
590         atomic_inc(&fs_info->scrubs_running);
591         atomic_inc(&fs_info->scrubs_paused);
592         mutex_unlock(&fs_info->scrub_lock);
593
594         /*
595          * check if @scrubs_running=@scrubs_paused condition
596          * inside wait_event() is not an atomic operation.
597          * which means we may inc/dec @scrub_running/paused
598          * at any time. Let's wake up @scrub_pause_wait as
599          * much as we can to let commit transaction blocked less.
600          */
601         wake_up(&fs_info->scrub_pause_wait);
602
603         atomic_inc(&sctx->workers_pending);
604 }
605
606 /* used for workers that require transaction commits */
607 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
608 {
609         struct btrfs_fs_info *fs_info = sctx->fs_info;
610
611         /*
612          * see scrub_pending_trans_workers_inc() why we're pretending
613          * to be paused in the scrub counters
614          */
615         mutex_lock(&fs_info->scrub_lock);
616         atomic_dec(&fs_info->scrubs_running);
617         atomic_dec(&fs_info->scrubs_paused);
618         mutex_unlock(&fs_info->scrub_lock);
619         atomic_dec(&sctx->workers_pending);
620         wake_up(&fs_info->scrub_pause_wait);
621         wake_up(&sctx->list_wait);
622         scrub_put_ctx(sctx);
623 }
624
625 static void scrub_free_csums(struct scrub_ctx *sctx)
626 {
627         while (!list_empty(&sctx->csum_list)) {
628                 struct btrfs_ordered_sum *sum;
629                 sum = list_first_entry(&sctx->csum_list,
630                                        struct btrfs_ordered_sum, list);
631                 list_del(&sum->list);
632                 kfree(sum);
633         }
634 }
635
636 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
637 {
638         int i;
639
640         if (!sctx)
641                 return;
642
643         /* this can happen when scrub is cancelled */
644         if (sctx->curr != -1) {
645                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
646
647                 for (i = 0; i < sbio->page_count; i++) {
648                         WARN_ON(!sbio->pagev[i]->page);
649                         scrub_block_put(sbio->pagev[i]->sblock);
650                 }
651                 bio_put(sbio->bio);
652         }
653
654         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
655                 struct scrub_bio *sbio = sctx->bios[i];
656
657                 if (!sbio)
658                         break;
659                 kfree(sbio);
660         }
661
662         kfree(sctx->wr_curr_bio);
663         scrub_free_csums(sctx);
664         kfree(sctx);
665 }
666
667 static void scrub_put_ctx(struct scrub_ctx *sctx)
668 {
669         if (refcount_dec_and_test(&sctx->refs))
670                 scrub_free_ctx(sctx);
671 }
672
673 static noinline_for_stack
674 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
675 {
676         struct scrub_ctx *sctx;
677         int             i;
678         struct btrfs_fs_info *fs_info = dev->fs_info;
679
680         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
681         if (!sctx)
682                 goto nomem;
683         refcount_set(&sctx->refs, 1);
684         sctx->is_dev_replace = is_dev_replace;
685         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
686         sctx->curr = -1;
687         sctx->fs_info = dev->fs_info;
688         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
689                 struct scrub_bio *sbio;
690
691                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
692                 if (!sbio)
693                         goto nomem;
694                 sctx->bios[i] = sbio;
695
696                 sbio->index = i;
697                 sbio->sctx = sctx;
698                 sbio->page_count = 0;
699                 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
700                                 scrub_bio_end_io_worker, NULL, NULL);
701
702                 if (i != SCRUB_BIOS_PER_SCTX - 1)
703                         sctx->bios[i]->next_free = i + 1;
704                 else
705                         sctx->bios[i]->next_free = -1;
706         }
707         sctx->first_free = 0;
708         atomic_set(&sctx->bios_in_flight, 0);
709         atomic_set(&sctx->workers_pending, 0);
710         atomic_set(&sctx->cancel_req, 0);
711         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
712         INIT_LIST_HEAD(&sctx->csum_list);
713
714         spin_lock_init(&sctx->list_lock);
715         spin_lock_init(&sctx->stat_lock);
716         init_waitqueue_head(&sctx->list_wait);
717
718         WARN_ON(sctx->wr_curr_bio != NULL);
719         mutex_init(&sctx->wr_lock);
720         sctx->wr_curr_bio = NULL;
721         if (is_dev_replace) {
722                 WARN_ON(!fs_info->dev_replace.tgtdev);
723                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
724                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
725                 sctx->flush_all_writes = false;
726         }
727
728         return sctx;
729
730 nomem:
731         scrub_free_ctx(sctx);
732         return ERR_PTR(-ENOMEM);
733 }
734
735 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
736                                      void *warn_ctx)
737 {
738         u64 isize;
739         u32 nlink;
740         int ret;
741         int i;
742         unsigned nofs_flag;
743         struct extent_buffer *eb;
744         struct btrfs_inode_item *inode_item;
745         struct scrub_warning *swarn = warn_ctx;
746         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
747         struct inode_fs_paths *ipath = NULL;
748         struct btrfs_root *local_root;
749         struct btrfs_key root_key;
750         struct btrfs_key key;
751
752         root_key.objectid = root;
753         root_key.type = BTRFS_ROOT_ITEM_KEY;
754         root_key.offset = (u64)-1;
755         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
756         if (IS_ERR(local_root)) {
757                 ret = PTR_ERR(local_root);
758                 goto err;
759         }
760
761         /*
762          * this makes the path point to (inum INODE_ITEM ioff)
763          */
764         key.objectid = inum;
765         key.type = BTRFS_INODE_ITEM_KEY;
766         key.offset = 0;
767
768         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
769         if (ret) {
770                 btrfs_release_path(swarn->path);
771                 goto err;
772         }
773
774         eb = swarn->path->nodes[0];
775         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
776                                         struct btrfs_inode_item);
777         isize = btrfs_inode_size(eb, inode_item);
778         nlink = btrfs_inode_nlink(eb, inode_item);
779         btrfs_release_path(swarn->path);
780
781         /*
782          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
783          * uses GFP_NOFS in this context, so we keep it consistent but it does
784          * not seem to be strictly necessary.
785          */
786         nofs_flag = memalloc_nofs_save();
787         ipath = init_ipath(4096, local_root, swarn->path);
788         memalloc_nofs_restore(nofs_flag);
789         if (IS_ERR(ipath)) {
790                 ret = PTR_ERR(ipath);
791                 ipath = NULL;
792                 goto err;
793         }
794         ret = paths_from_inode(inum, ipath);
795
796         if (ret < 0)
797                 goto err;
798
799         /*
800          * we deliberately ignore the bit ipath might have been too small to
801          * hold all of the paths here
802          */
803         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
804                 btrfs_warn_in_rcu(fs_info,
805 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
806                                   swarn->errstr, swarn->logical,
807                                   rcu_str_deref(swarn->dev->name),
808                                   swarn->physical,
809                                   root, inum, offset,
810                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
811                                   (char *)(unsigned long)ipath->fspath->val[i]);
812
813         free_ipath(ipath);
814         return 0;
815
816 err:
817         btrfs_warn_in_rcu(fs_info,
818                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
819                           swarn->errstr, swarn->logical,
820                           rcu_str_deref(swarn->dev->name),
821                           swarn->physical,
822                           root, inum, offset, ret);
823
824         free_ipath(ipath);
825         return 0;
826 }
827
828 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
829 {
830         struct btrfs_device *dev;
831         struct btrfs_fs_info *fs_info;
832         struct btrfs_path *path;
833         struct btrfs_key found_key;
834         struct extent_buffer *eb;
835         struct btrfs_extent_item *ei;
836         struct scrub_warning swarn;
837         unsigned long ptr = 0;
838         u64 extent_item_pos;
839         u64 flags = 0;
840         u64 ref_root;
841         u32 item_size;
842         u8 ref_level = 0;
843         int ret;
844
845         WARN_ON(sblock->page_count < 1);
846         dev = sblock->pagev[0]->dev;
847         fs_info = sblock->sctx->fs_info;
848
849         path = btrfs_alloc_path();
850         if (!path)
851                 return;
852
853         swarn.physical = sblock->pagev[0]->physical;
854         swarn.logical = sblock->pagev[0]->logical;
855         swarn.errstr = errstr;
856         swarn.dev = NULL;
857
858         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
859                                   &flags);
860         if (ret < 0)
861                 goto out;
862
863         extent_item_pos = swarn.logical - found_key.objectid;
864         swarn.extent_item_size = found_key.offset;
865
866         eb = path->nodes[0];
867         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
868         item_size = btrfs_item_size_nr(eb, path->slots[0]);
869
870         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
871                 do {
872                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
873                                                       item_size, &ref_root,
874                                                       &ref_level);
875                         btrfs_warn_in_rcu(fs_info,
876 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
877                                 errstr, swarn.logical,
878                                 rcu_str_deref(dev->name),
879                                 swarn.physical,
880                                 ref_level ? "node" : "leaf",
881                                 ret < 0 ? -1 : ref_level,
882                                 ret < 0 ? -1 : ref_root);
883                 } while (ret != 1);
884                 btrfs_release_path(path);
885         } else {
886                 btrfs_release_path(path);
887                 swarn.path = path;
888                 swarn.dev = dev;
889                 iterate_extent_inodes(fs_info, found_key.objectid,
890                                         extent_item_pos, 1,
891                                         scrub_print_warning_inode, &swarn, false);
892         }
893
894 out:
895         btrfs_free_path(path);
896 }
897
898 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
899 {
900         struct page *page = NULL;
901         unsigned long index;
902         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
903         int ret;
904         int corrected = 0;
905         struct btrfs_key key;
906         struct inode *inode = NULL;
907         struct btrfs_fs_info *fs_info;
908         u64 end = offset + PAGE_SIZE - 1;
909         struct btrfs_root *local_root;
910         int srcu_index;
911
912         key.objectid = root;
913         key.type = BTRFS_ROOT_ITEM_KEY;
914         key.offset = (u64)-1;
915
916         fs_info = fixup->root->fs_info;
917         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
918
919         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
920         if (IS_ERR(local_root)) {
921                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
922                 return PTR_ERR(local_root);
923         }
924
925         key.type = BTRFS_INODE_ITEM_KEY;
926         key.objectid = inum;
927         key.offset = 0;
928         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
929         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
930         if (IS_ERR(inode))
931                 return PTR_ERR(inode);
932
933         index = offset >> PAGE_SHIFT;
934
935         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
936         if (!page) {
937                 ret = -ENOMEM;
938                 goto out;
939         }
940
941         if (PageUptodate(page)) {
942                 if (PageDirty(page)) {
943                         /*
944                          * we need to write the data to the defect sector. the
945                          * data that was in that sector is not in memory,
946                          * because the page was modified. we must not write the
947                          * modified page to that sector.
948                          *
949                          * TODO: what could be done here: wait for the delalloc
950                          *       runner to write out that page (might involve
951                          *       COW) and see whether the sector is still
952                          *       referenced afterwards.
953                          *
954                          * For the meantime, we'll treat this error
955                          * incorrectable, although there is a chance that a
956                          * later scrub will find the bad sector again and that
957                          * there's no dirty page in memory, then.
958                          */
959                         ret = -EIO;
960                         goto out;
961                 }
962                 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
963                                         fixup->logical, page,
964                                         offset - page_offset(page),
965                                         fixup->mirror_num);
966                 unlock_page(page);
967                 corrected = !ret;
968         } else {
969                 /*
970                  * we need to get good data first. the general readpage path
971                  * will call repair_io_failure for us, we just have to make
972                  * sure we read the bad mirror.
973                  */
974                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
975                                         EXTENT_DAMAGED);
976                 if (ret) {
977                         /* set_extent_bits should give proper error */
978                         WARN_ON(ret > 0);
979                         if (ret > 0)
980                                 ret = -EFAULT;
981                         goto out;
982                 }
983
984                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
985                                                 btrfs_get_extent,
986                                                 fixup->mirror_num);
987                 wait_on_page_locked(page);
988
989                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
990                                                 end, EXTENT_DAMAGED, 0, NULL);
991                 if (!corrected)
992                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
993                                                 EXTENT_DAMAGED);
994         }
995
996 out:
997         if (page)
998                 put_page(page);
999
1000         iput(inode);
1001
1002         if (ret < 0)
1003                 return ret;
1004
1005         if (ret == 0 && corrected) {
1006                 /*
1007                  * we only need to call readpage for one of the inodes belonging
1008                  * to this extent. so make iterate_extent_inodes stop
1009                  */
1010                 return 1;
1011         }
1012
1013         return -EIO;
1014 }
1015
1016 static void scrub_fixup_nodatasum(struct btrfs_work *work)
1017 {
1018         struct btrfs_fs_info *fs_info;
1019         int ret;
1020         struct scrub_fixup_nodatasum *fixup;
1021         struct scrub_ctx *sctx;
1022         struct btrfs_trans_handle *trans = NULL;
1023         struct btrfs_path *path;
1024         int uncorrectable = 0;
1025
1026         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1027         sctx = fixup->sctx;
1028         fs_info = fixup->root->fs_info;
1029
1030         path = btrfs_alloc_path();
1031         if (!path) {
1032                 spin_lock(&sctx->stat_lock);
1033                 ++sctx->stat.malloc_errors;
1034                 spin_unlock(&sctx->stat_lock);
1035                 uncorrectable = 1;
1036                 goto out;
1037         }
1038
1039         trans = btrfs_join_transaction(fixup->root);
1040         if (IS_ERR(trans)) {
1041                 uncorrectable = 1;
1042                 goto out;
1043         }
1044
1045         /*
1046          * the idea is to trigger a regular read through the standard path. we
1047          * read a page from the (failed) logical address by specifying the
1048          * corresponding copynum of the failed sector. thus, that readpage is
1049          * expected to fail.
1050          * that is the point where on-the-fly error correction will kick in
1051          * (once it's finished) and rewrite the failed sector if a good copy
1052          * can be found.
1053          */
1054         ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1055                                           scrub_fixup_readpage, fixup, false);
1056         if (ret < 0) {
1057                 uncorrectable = 1;
1058                 goto out;
1059         }
1060         WARN_ON(ret != 1);
1061
1062         spin_lock(&sctx->stat_lock);
1063         ++sctx->stat.corrected_errors;
1064         spin_unlock(&sctx->stat_lock);
1065
1066 out:
1067         if (trans && !IS_ERR(trans))
1068                 btrfs_end_transaction(trans);
1069         if (uncorrectable) {
1070                 spin_lock(&sctx->stat_lock);
1071                 ++sctx->stat.uncorrectable_errors;
1072                 spin_unlock(&sctx->stat_lock);
1073                 btrfs_dev_replace_stats_inc(
1074                         &fs_info->dev_replace.num_uncorrectable_read_errors);
1075                 btrfs_err_rl_in_rcu(fs_info,
1076                     "unable to fixup (nodatasum) error at logical %llu on dev %s",
1077                         fixup->logical, rcu_str_deref(fixup->dev->name));
1078         }
1079
1080         btrfs_free_path(path);
1081         kfree(fixup);
1082
1083         scrub_pending_trans_workers_dec(sctx);
1084 }
1085
1086 static inline void scrub_get_recover(struct scrub_recover *recover)
1087 {
1088         refcount_inc(&recover->refs);
1089 }
1090
1091 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1092                                      struct scrub_recover *recover)
1093 {
1094         if (refcount_dec_and_test(&recover->refs)) {
1095                 btrfs_bio_counter_dec(fs_info);
1096                 btrfs_put_bbio(recover->bbio);
1097                 kfree(recover);
1098         }
1099 }
1100
1101 /*
1102  * scrub_handle_errored_block gets called when either verification of the
1103  * pages failed or the bio failed to read, e.g. with EIO. In the latter
1104  * case, this function handles all pages in the bio, even though only one
1105  * may be bad.
1106  * The goal of this function is to repair the errored block by using the
1107  * contents of one of the mirrors.
1108  */
1109 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1110 {
1111         struct scrub_ctx *sctx = sblock_to_check->sctx;
1112         struct btrfs_device *dev;
1113         struct btrfs_fs_info *fs_info;
1114         u64 length;
1115         u64 logical;
1116         unsigned int failed_mirror_index;
1117         unsigned int is_metadata;
1118         unsigned int have_csum;
1119         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1120         struct scrub_block *sblock_bad;
1121         int ret;
1122         int mirror_index;
1123         int page_num;
1124         int success;
1125         bool full_stripe_locked;
1126         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1127                                       DEFAULT_RATELIMIT_BURST);
1128
1129         BUG_ON(sblock_to_check->page_count < 1);
1130         fs_info = sctx->fs_info;
1131         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1132                 /*
1133                  * if we find an error in a super block, we just report it.
1134                  * They will get written with the next transaction commit
1135                  * anyway
1136                  */
1137                 spin_lock(&sctx->stat_lock);
1138                 ++sctx->stat.super_errors;
1139                 spin_unlock(&sctx->stat_lock);
1140                 return 0;
1141         }
1142         length = sblock_to_check->page_count * PAGE_SIZE;
1143         logical = sblock_to_check->pagev[0]->logical;
1144         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1145         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1146         is_metadata = !(sblock_to_check->pagev[0]->flags &
1147                         BTRFS_EXTENT_FLAG_DATA);
1148         have_csum = sblock_to_check->pagev[0]->have_csum;
1149         dev = sblock_to_check->pagev[0]->dev;
1150
1151         /*
1152          * For RAID5/6, race can happen for a different device scrub thread.
1153          * For data corruption, Parity and Data threads will both try
1154          * to recovery the data.
1155          * Race can lead to doubly added csum error, or even unrecoverable
1156          * error.
1157          */
1158         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1159         if (ret < 0) {
1160                 spin_lock(&sctx->stat_lock);
1161                 if (ret == -ENOMEM)
1162                         sctx->stat.malloc_errors++;
1163                 sctx->stat.read_errors++;
1164                 sctx->stat.uncorrectable_errors++;
1165                 spin_unlock(&sctx->stat_lock);
1166                 return ret;
1167         }
1168
1169         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
1170                 sblocks_for_recheck = NULL;
1171                 goto nodatasum_case;
1172         }
1173
1174         /*
1175          * read all mirrors one after the other. This includes to
1176          * re-read the extent or metadata block that failed (that was
1177          * the cause that this fixup code is called) another time,
1178          * page by page this time in order to know which pages
1179          * caused I/O errors and which ones are good (for all mirrors).
1180          * It is the goal to handle the situation when more than one
1181          * mirror contains I/O errors, but the errors do not
1182          * overlap, i.e. the data can be repaired by selecting the
1183          * pages from those mirrors without I/O error on the
1184          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1185          * would be that mirror #1 has an I/O error on the first page,
1186          * the second page is good, and mirror #2 has an I/O error on
1187          * the second page, but the first page is good.
1188          * Then the first page of the first mirror can be repaired by
1189          * taking the first page of the second mirror, and the
1190          * second page of the second mirror can be repaired by
1191          * copying the contents of the 2nd page of the 1st mirror.
1192          * One more note: if the pages of one mirror contain I/O
1193          * errors, the checksum cannot be verified. In order to get
1194          * the best data for repairing, the first attempt is to find
1195          * a mirror without I/O errors and with a validated checksum.
1196          * Only if this is not possible, the pages are picked from
1197          * mirrors with I/O errors without considering the checksum.
1198          * If the latter is the case, at the end, the checksum of the
1199          * repaired area is verified in order to correctly maintain
1200          * the statistics.
1201          */
1202
1203         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1204                                       sizeof(*sblocks_for_recheck), GFP_NOFS);
1205         if (!sblocks_for_recheck) {
1206                 spin_lock(&sctx->stat_lock);
1207                 sctx->stat.malloc_errors++;
1208                 sctx->stat.read_errors++;
1209                 sctx->stat.uncorrectable_errors++;
1210                 spin_unlock(&sctx->stat_lock);
1211                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1212                 goto out;
1213         }
1214
1215         /* setup the context, map the logical blocks and alloc the pages */
1216         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1217         if (ret) {
1218                 spin_lock(&sctx->stat_lock);
1219                 sctx->stat.read_errors++;
1220                 sctx->stat.uncorrectable_errors++;
1221                 spin_unlock(&sctx->stat_lock);
1222                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1223                 goto out;
1224         }
1225         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1226         sblock_bad = sblocks_for_recheck + failed_mirror_index;
1227
1228         /* build and submit the bios for the failed mirror, check checksums */
1229         scrub_recheck_block(fs_info, sblock_bad, 1);
1230
1231         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1232             sblock_bad->no_io_error_seen) {
1233                 /*
1234                  * the error disappeared after reading page by page, or
1235                  * the area was part of a huge bio and other parts of the
1236                  * bio caused I/O errors, or the block layer merged several
1237                  * read requests into one and the error is caused by a
1238                  * different bio (usually one of the two latter cases is
1239                  * the cause)
1240                  */
1241                 spin_lock(&sctx->stat_lock);
1242                 sctx->stat.unverified_errors++;
1243                 sblock_to_check->data_corrected = 1;
1244                 spin_unlock(&sctx->stat_lock);
1245
1246                 if (sctx->is_dev_replace)
1247                         scrub_write_block_to_dev_replace(sblock_bad);
1248                 goto out;
1249         }
1250
1251         if (!sblock_bad->no_io_error_seen) {
1252                 spin_lock(&sctx->stat_lock);
1253                 sctx->stat.read_errors++;
1254                 spin_unlock(&sctx->stat_lock);
1255                 if (__ratelimit(&_rs))
1256                         scrub_print_warning("i/o error", sblock_to_check);
1257                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1258         } else if (sblock_bad->checksum_error) {
1259                 spin_lock(&sctx->stat_lock);
1260                 sctx->stat.csum_errors++;
1261                 spin_unlock(&sctx->stat_lock);
1262                 if (__ratelimit(&_rs))
1263                         scrub_print_warning("checksum error", sblock_to_check);
1264                 btrfs_dev_stat_inc_and_print(dev,
1265                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
1266         } else if (sblock_bad->header_error) {
1267                 spin_lock(&sctx->stat_lock);
1268                 sctx->stat.verify_errors++;
1269                 spin_unlock(&sctx->stat_lock);
1270                 if (__ratelimit(&_rs))
1271                         scrub_print_warning("checksum/header error",
1272                                             sblock_to_check);
1273                 if (sblock_bad->generation_error)
1274                         btrfs_dev_stat_inc_and_print(dev,
1275                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1276                 else
1277                         btrfs_dev_stat_inc_and_print(dev,
1278                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1279         }
1280
1281         if (sctx->readonly) {
1282                 ASSERT(!sctx->is_dev_replace);
1283                 goto out;
1284         }
1285
1286         if (!is_metadata && !have_csum) {
1287                 struct scrub_fixup_nodatasum *fixup_nodatasum;
1288
1289                 WARN_ON(sctx->is_dev_replace);
1290
1291 nodatasum_case:
1292
1293                 /*
1294                  * !is_metadata and !have_csum, this means that the data
1295                  * might not be COWed, that it might be modified
1296                  * concurrently. The general strategy to work on the
1297                  * commit root does not help in the case when COW is not
1298                  * used.
1299                  */
1300                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1301                 if (!fixup_nodatasum)
1302                         goto did_not_correct_error;
1303                 fixup_nodatasum->sctx = sctx;
1304                 fixup_nodatasum->dev = dev;
1305                 fixup_nodatasum->logical = logical;
1306                 fixup_nodatasum->root = fs_info->extent_root;
1307                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1308                 scrub_pending_trans_workers_inc(sctx);
1309                 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1310                                 scrub_fixup_nodatasum, NULL, NULL);
1311                 btrfs_queue_work(fs_info->scrub_workers,
1312                                  &fixup_nodatasum->work);
1313                 goto out;
1314         }
1315
1316         /*
1317          * now build and submit the bios for the other mirrors, check
1318          * checksums.
1319          * First try to pick the mirror which is completely without I/O
1320          * errors and also does not have a checksum error.
1321          * If one is found, and if a checksum is present, the full block
1322          * that is known to contain an error is rewritten. Afterwards
1323          * the block is known to be corrected.
1324          * If a mirror is found which is completely correct, and no
1325          * checksum is present, only those pages are rewritten that had
1326          * an I/O error in the block to be repaired, since it cannot be
1327          * determined, which copy of the other pages is better (and it
1328          * could happen otherwise that a correct page would be
1329          * overwritten by a bad one).
1330          */
1331         for (mirror_index = 0; ;mirror_index++) {
1332                 struct scrub_block *sblock_other;
1333
1334                 if (mirror_index == failed_mirror_index)
1335                         continue;
1336
1337                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1338                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1339                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1340                                 break;
1341                         if (!sblocks_for_recheck[mirror_index].page_count)
1342                                 break;
1343
1344                         sblock_other = sblocks_for_recheck + mirror_index;
1345                 } else {
1346                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1347                         int max_allowed = r->bbio->num_stripes -
1348                                                 r->bbio->num_tgtdevs;
1349
1350                         if (mirror_index >= max_allowed)
1351                                 break;
1352                         if (!sblocks_for_recheck[1].page_count)
1353                                 break;
1354
1355                         ASSERT(failed_mirror_index == 0);
1356                         sblock_other = sblocks_for_recheck + 1;
1357                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1358                 }
1359
1360                 /* build and submit the bios, check checksums */
1361                 scrub_recheck_block(fs_info, sblock_other, 0);
1362
1363                 if (!sblock_other->header_error &&
1364                     !sblock_other->checksum_error &&
1365                     sblock_other->no_io_error_seen) {
1366                         if (sctx->is_dev_replace) {
1367                                 scrub_write_block_to_dev_replace(sblock_other);
1368                                 goto corrected_error;
1369                         } else {
1370                                 ret = scrub_repair_block_from_good_copy(
1371                                                 sblock_bad, sblock_other);
1372                                 if (!ret)
1373                                         goto corrected_error;
1374                         }
1375                 }
1376         }
1377
1378         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1379                 goto did_not_correct_error;
1380
1381         /*
1382          * In case of I/O errors in the area that is supposed to be
1383          * repaired, continue by picking good copies of those pages.
1384          * Select the good pages from mirrors to rewrite bad pages from
1385          * the area to fix. Afterwards verify the checksum of the block
1386          * that is supposed to be repaired. This verification step is
1387          * only done for the purpose of statistic counting and for the
1388          * final scrub report, whether errors remain.
1389          * A perfect algorithm could make use of the checksum and try
1390          * all possible combinations of pages from the different mirrors
1391          * until the checksum verification succeeds. For example, when
1392          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1393          * of mirror #2 is readable but the final checksum test fails,
1394          * then the 2nd page of mirror #3 could be tried, whether now
1395          * the final checksum succeeds. But this would be a rare
1396          * exception and is therefore not implemented. At least it is
1397          * avoided that the good copy is overwritten.
1398          * A more useful improvement would be to pick the sectors
1399          * without I/O error based on sector sizes (512 bytes on legacy
1400          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1401          * mirror could be repaired by taking 512 byte of a different
1402          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1403          * area are unreadable.
1404          */
1405         success = 1;
1406         for (page_num = 0; page_num < sblock_bad->page_count;
1407              page_num++) {
1408                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1409                 struct scrub_block *sblock_other = NULL;
1410
1411                 /* skip no-io-error page in scrub */
1412                 if (!page_bad->io_error && !sctx->is_dev_replace)
1413                         continue;
1414
1415                 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1416                         /*
1417                          * In case of dev replace, if raid56 rebuild process
1418                          * didn't work out correct data, then copy the content
1419                          * in sblock_bad to make sure target device is identical
1420                          * to source device, instead of writing garbage data in
1421                          * sblock_for_recheck array to target device.
1422                          */
1423                         sblock_other = NULL;
1424                 } else if (page_bad->io_error) {
1425                         /* try to find no-io-error page in mirrors */
1426                         for (mirror_index = 0;
1427                              mirror_index < BTRFS_MAX_MIRRORS &&
1428                              sblocks_for_recheck[mirror_index].page_count > 0;
1429                              mirror_index++) {
1430                                 if (!sblocks_for_recheck[mirror_index].
1431                                     pagev[page_num]->io_error) {
1432                                         sblock_other = sblocks_for_recheck +
1433                                                        mirror_index;
1434                                         break;
1435                                 }
1436                         }
1437                         if (!sblock_other)
1438                                 success = 0;
1439                 }
1440
1441                 if (sctx->is_dev_replace) {
1442                         /*
1443                          * did not find a mirror to fetch the page
1444                          * from. scrub_write_page_to_dev_replace()
1445                          * handles this case (page->io_error), by
1446                          * filling the block with zeros before
1447                          * submitting the write request
1448                          */
1449                         if (!sblock_other)
1450                                 sblock_other = sblock_bad;
1451
1452                         if (scrub_write_page_to_dev_replace(sblock_other,
1453                                                             page_num) != 0) {
1454                                 btrfs_dev_replace_stats_inc(
1455                                         &fs_info->dev_replace.num_write_errors);
1456                                 success = 0;
1457                         }
1458                 } else if (sblock_other) {
1459                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1460                                                                sblock_other,
1461                                                                page_num, 0);
1462                         if (0 == ret)
1463                                 page_bad->io_error = 0;
1464                         else
1465                                 success = 0;
1466                 }
1467         }
1468
1469         if (success && !sctx->is_dev_replace) {
1470                 if (is_metadata || have_csum) {
1471                         /*
1472                          * need to verify the checksum now that all
1473                          * sectors on disk are repaired (the write
1474                          * request for data to be repaired is on its way).
1475                          * Just be lazy and use scrub_recheck_block()
1476                          * which re-reads the data before the checksum
1477                          * is verified, but most likely the data comes out
1478                          * of the page cache.
1479                          */
1480                         scrub_recheck_block(fs_info, sblock_bad, 1);
1481                         if (!sblock_bad->header_error &&
1482                             !sblock_bad->checksum_error &&
1483                             sblock_bad->no_io_error_seen)
1484                                 goto corrected_error;
1485                         else
1486                                 goto did_not_correct_error;
1487                 } else {
1488 corrected_error:
1489                         spin_lock(&sctx->stat_lock);
1490                         sctx->stat.corrected_errors++;
1491                         sblock_to_check->data_corrected = 1;
1492                         spin_unlock(&sctx->stat_lock);
1493                         btrfs_err_rl_in_rcu(fs_info,
1494                                 "fixed up error at logical %llu on dev %s",
1495                                 logical, rcu_str_deref(dev->name));
1496                 }
1497         } else {
1498 did_not_correct_error:
1499                 spin_lock(&sctx->stat_lock);
1500                 sctx->stat.uncorrectable_errors++;
1501                 spin_unlock(&sctx->stat_lock);
1502                 btrfs_err_rl_in_rcu(fs_info,
1503                         "unable to fixup (regular) error at logical %llu on dev %s",
1504                         logical, rcu_str_deref(dev->name));
1505         }
1506
1507 out:
1508         if (sblocks_for_recheck) {
1509                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1510                      mirror_index++) {
1511                         struct scrub_block *sblock = sblocks_for_recheck +
1512                                                      mirror_index;
1513                         struct scrub_recover *recover;
1514                         int page_index;
1515
1516                         for (page_index = 0; page_index < sblock->page_count;
1517                              page_index++) {
1518                                 sblock->pagev[page_index]->sblock = NULL;
1519                                 recover = sblock->pagev[page_index]->recover;
1520                                 if (recover) {
1521                                         scrub_put_recover(fs_info, recover);
1522                                         sblock->pagev[page_index]->recover =
1523                                                                         NULL;
1524                                 }
1525                                 scrub_page_put(sblock->pagev[page_index]);
1526                         }
1527                 }
1528                 kfree(sblocks_for_recheck);
1529         }
1530
1531         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1532         if (ret < 0)
1533                 return ret;
1534         return 0;
1535 }
1536
1537 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1538 {
1539         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1540                 return 2;
1541         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1542                 return 3;
1543         else
1544                 return (int)bbio->num_stripes;
1545 }
1546
1547 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1548                                                  u64 *raid_map,
1549                                                  u64 mapped_length,
1550                                                  int nstripes, int mirror,
1551                                                  int *stripe_index,
1552                                                  u64 *stripe_offset)
1553 {
1554         int i;
1555
1556         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1557                 /* RAID5/6 */
1558                 for (i = 0; i < nstripes; i++) {
1559                         if (raid_map[i] == RAID6_Q_STRIPE ||
1560                             raid_map[i] == RAID5_P_STRIPE)
1561                                 continue;
1562
1563                         if (logical >= raid_map[i] &&
1564                             logical < raid_map[i] + mapped_length)
1565                                 break;
1566                 }
1567
1568                 *stripe_index = i;
1569                 *stripe_offset = logical - raid_map[i];
1570         } else {
1571                 /* The other RAID type */
1572                 *stripe_index = mirror;
1573                 *stripe_offset = 0;
1574         }
1575 }
1576
1577 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1578                                      struct scrub_block *sblocks_for_recheck)
1579 {
1580         struct scrub_ctx *sctx = original_sblock->sctx;
1581         struct btrfs_fs_info *fs_info = sctx->fs_info;
1582         u64 length = original_sblock->page_count * PAGE_SIZE;
1583         u64 logical = original_sblock->pagev[0]->logical;
1584         u64 generation = original_sblock->pagev[0]->generation;
1585         u64 flags = original_sblock->pagev[0]->flags;
1586         u64 have_csum = original_sblock->pagev[0]->have_csum;
1587         struct scrub_recover *recover;
1588         struct btrfs_bio *bbio;
1589         u64 sublen;
1590         u64 mapped_length;
1591         u64 stripe_offset;
1592         int stripe_index;
1593         int page_index = 0;
1594         int mirror_index;
1595         int nmirrors;
1596         int ret;
1597
1598         /*
1599          * note: the two members refs and outstanding_pages
1600          * are not used (and not set) in the blocks that are used for
1601          * the recheck procedure
1602          */
1603
1604         while (length > 0) {
1605                 sublen = min_t(u64, length, PAGE_SIZE);
1606                 mapped_length = sublen;
1607                 bbio = NULL;
1608
1609                 /*
1610                  * with a length of PAGE_SIZE, each returned stripe
1611                  * represents one mirror
1612                  */
1613                 btrfs_bio_counter_inc_blocked(fs_info);
1614                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1615                                 logical, &mapped_length, &bbio);
1616                 if (ret || !bbio || mapped_length < sublen) {
1617                         btrfs_put_bbio(bbio);
1618                         btrfs_bio_counter_dec(fs_info);
1619                         return -EIO;
1620                 }
1621
1622                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1623                 if (!recover) {
1624                         btrfs_put_bbio(bbio);
1625                         btrfs_bio_counter_dec(fs_info);
1626                         return -ENOMEM;
1627                 }
1628
1629                 refcount_set(&recover->refs, 1);
1630                 recover->bbio = bbio;
1631                 recover->map_length = mapped_length;
1632
1633                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1634
1635                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1636
1637                 for (mirror_index = 0; mirror_index < nmirrors;
1638                      mirror_index++) {
1639                         struct scrub_block *sblock;
1640                         struct scrub_page *page;
1641
1642                         sblock = sblocks_for_recheck + mirror_index;
1643                         sblock->sctx = sctx;
1644
1645                         page = kzalloc(sizeof(*page), GFP_NOFS);
1646                         if (!page) {
1647 leave_nomem:
1648                                 spin_lock(&sctx->stat_lock);
1649                                 sctx->stat.malloc_errors++;
1650                                 spin_unlock(&sctx->stat_lock);
1651                                 scrub_put_recover(fs_info, recover);
1652                                 return -ENOMEM;
1653                         }
1654                         scrub_page_get(page);
1655                         sblock->pagev[page_index] = page;
1656                         page->sblock = sblock;
1657                         page->flags = flags;
1658                         page->generation = generation;
1659                         page->logical = logical;
1660                         page->have_csum = have_csum;
1661                         if (have_csum)
1662                                 memcpy(page->csum,
1663                                        original_sblock->pagev[0]->csum,
1664                                        sctx->csum_size);
1665
1666                         scrub_stripe_index_and_offset(logical,
1667                                                       bbio->map_type,
1668                                                       bbio->raid_map,
1669                                                       mapped_length,
1670                                                       bbio->num_stripes -
1671                                                       bbio->num_tgtdevs,
1672                                                       mirror_index,
1673                                                       &stripe_index,
1674                                                       &stripe_offset);
1675                         page->physical = bbio->stripes[stripe_index].physical +
1676                                          stripe_offset;
1677                         page->dev = bbio->stripes[stripe_index].dev;
1678
1679                         BUG_ON(page_index >= original_sblock->page_count);
1680                         page->physical_for_dev_replace =
1681                                 original_sblock->pagev[page_index]->
1682                                 physical_for_dev_replace;
1683                         /* for missing devices, dev->bdev is NULL */
1684                         page->mirror_num = mirror_index + 1;
1685                         sblock->page_count++;
1686                         page->page = alloc_page(GFP_NOFS);
1687                         if (!page->page)
1688                                 goto leave_nomem;
1689
1690                         scrub_get_recover(recover);
1691                         page->recover = recover;
1692                 }
1693                 scrub_put_recover(fs_info, recover);
1694                 length -= sublen;
1695                 logical += sublen;
1696                 page_index++;
1697         }
1698
1699         return 0;
1700 }
1701
1702 static void scrub_bio_wait_endio(struct bio *bio)
1703 {
1704         complete(bio->bi_private);
1705 }
1706
1707 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1708                                         struct bio *bio,
1709                                         struct scrub_page *page)
1710 {
1711         DECLARE_COMPLETION_ONSTACK(done);
1712         int ret;
1713         int mirror_num;
1714
1715         bio->bi_iter.bi_sector = page->logical >> 9;
1716         bio->bi_private = &done;
1717         bio->bi_end_io = scrub_bio_wait_endio;
1718
1719         mirror_num = page->sblock->pagev[0]->mirror_num;
1720         ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1721                                     page->recover->map_length,
1722                                     mirror_num, 0);
1723         if (ret)
1724                 return ret;
1725
1726         wait_for_completion_io(&done);
1727         return blk_status_to_errno(bio->bi_status);
1728 }
1729
1730 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1731                                           struct scrub_block *sblock)
1732 {
1733         struct scrub_page *first_page = sblock->pagev[0];
1734         struct bio *bio;
1735         int page_num;
1736
1737         /* All pages in sblock belong to the same stripe on the same device. */
1738         ASSERT(first_page->dev);
1739         if (!first_page->dev->bdev)
1740                 goto out;
1741
1742         bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1743         bio_set_dev(bio, first_page->dev->bdev);
1744
1745         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1746                 struct scrub_page *page = sblock->pagev[page_num];
1747
1748                 WARN_ON(!page->page);
1749                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1750         }
1751
1752         if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1753                 bio_put(bio);
1754                 goto out;
1755         }
1756
1757         bio_put(bio);
1758
1759         scrub_recheck_block_checksum(sblock);
1760
1761         return;
1762 out:
1763         for (page_num = 0; page_num < sblock->page_count; page_num++)
1764                 sblock->pagev[page_num]->io_error = 1;
1765
1766         sblock->no_io_error_seen = 0;
1767 }
1768
1769 /*
1770  * this function will check the on disk data for checksum errors, header
1771  * errors and read I/O errors. If any I/O errors happen, the exact pages
1772  * which are errored are marked as being bad. The goal is to enable scrub
1773  * to take those pages that are not errored from all the mirrors so that
1774  * the pages that are errored in the just handled mirror can be repaired.
1775  */
1776 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1777                                 struct scrub_block *sblock,
1778                                 int retry_failed_mirror)
1779 {
1780         int page_num;
1781
1782         sblock->no_io_error_seen = 1;
1783
1784         /* short cut for raid56 */
1785         if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1786                 return scrub_recheck_block_on_raid56(fs_info, sblock);
1787
1788         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1789                 struct bio *bio;
1790                 struct scrub_page *page = sblock->pagev[page_num];
1791
1792                 if (page->dev->bdev == NULL) {
1793                         page->io_error = 1;
1794                         sblock->no_io_error_seen = 0;
1795                         continue;
1796                 }
1797
1798                 WARN_ON(!page->page);
1799                 bio = btrfs_io_bio_alloc(1);
1800                 bio_set_dev(bio, page->dev->bdev);
1801
1802                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1803                 bio->bi_iter.bi_sector = page->physical >> 9;
1804                 bio->bi_opf = REQ_OP_READ;
1805
1806                 if (btrfsic_submit_bio_wait(bio)) {
1807                         page->io_error = 1;
1808                         sblock->no_io_error_seen = 0;
1809                 }
1810
1811                 bio_put(bio);
1812         }
1813
1814         if (sblock->no_io_error_seen)
1815                 scrub_recheck_block_checksum(sblock);
1816 }
1817
1818 static inline int scrub_check_fsid(u8 fsid[],
1819                                    struct scrub_page *spage)
1820 {
1821         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1822         int ret;
1823
1824         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1825         return !ret;
1826 }
1827
1828 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1829 {
1830         sblock->header_error = 0;
1831         sblock->checksum_error = 0;
1832         sblock->generation_error = 0;
1833
1834         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1835                 scrub_checksum_data(sblock);
1836         else
1837                 scrub_checksum_tree_block(sblock);
1838 }
1839
1840 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1841                                              struct scrub_block *sblock_good)
1842 {
1843         int page_num;
1844         int ret = 0;
1845
1846         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1847                 int ret_sub;
1848
1849                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1850                                                            sblock_good,
1851                                                            page_num, 1);
1852                 if (ret_sub)
1853                         ret = ret_sub;
1854         }
1855
1856         return ret;
1857 }
1858
1859 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1860                                             struct scrub_block *sblock_good,
1861                                             int page_num, int force_write)
1862 {
1863         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1864         struct scrub_page *page_good = sblock_good->pagev[page_num];
1865         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1866
1867         BUG_ON(page_bad->page == NULL);
1868         BUG_ON(page_good->page == NULL);
1869         if (force_write || sblock_bad->header_error ||
1870             sblock_bad->checksum_error || page_bad->io_error) {
1871                 struct bio *bio;
1872                 int ret;
1873
1874                 if (!page_bad->dev->bdev) {
1875                         btrfs_warn_rl(fs_info,
1876                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1877                         return -EIO;
1878                 }
1879
1880                 bio = btrfs_io_bio_alloc(1);
1881                 bio_set_dev(bio, page_bad->dev->bdev);
1882                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1883                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1884
1885                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1886                 if (PAGE_SIZE != ret) {
1887                         bio_put(bio);
1888                         return -EIO;
1889                 }
1890
1891                 if (btrfsic_submit_bio_wait(bio)) {
1892                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1893                                 BTRFS_DEV_STAT_WRITE_ERRS);
1894                         btrfs_dev_replace_stats_inc(
1895                                 &fs_info->dev_replace.num_write_errors);
1896                         bio_put(bio);
1897                         return -EIO;
1898                 }
1899                 bio_put(bio);
1900         }
1901
1902         return 0;
1903 }
1904
1905 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1906 {
1907         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1908         int page_num;
1909
1910         /*
1911          * This block is used for the check of the parity on the source device,
1912          * so the data needn't be written into the destination device.
1913          */
1914         if (sblock->sparity)
1915                 return;
1916
1917         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1918                 int ret;
1919
1920                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1921                 if (ret)
1922                         btrfs_dev_replace_stats_inc(
1923                                 &fs_info->dev_replace.num_write_errors);
1924         }
1925 }
1926
1927 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1928                                            int page_num)
1929 {
1930         struct scrub_page *spage = sblock->pagev[page_num];
1931
1932         BUG_ON(spage->page == NULL);
1933         if (spage->io_error) {
1934                 void *mapped_buffer = kmap_atomic(spage->page);
1935
1936                 clear_page(mapped_buffer);
1937                 flush_dcache_page(spage->page);
1938                 kunmap_atomic(mapped_buffer);
1939         }
1940         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1941 }
1942
1943 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1944                                     struct scrub_page *spage)
1945 {
1946         struct scrub_bio *sbio;
1947         int ret;
1948
1949         mutex_lock(&sctx->wr_lock);
1950 again:
1951         if (!sctx->wr_curr_bio) {
1952                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1953                                               GFP_KERNEL);
1954                 if (!sctx->wr_curr_bio) {
1955                         mutex_unlock(&sctx->wr_lock);
1956                         return -ENOMEM;
1957                 }
1958                 sctx->wr_curr_bio->sctx = sctx;
1959                 sctx->wr_curr_bio->page_count = 0;
1960         }
1961         sbio = sctx->wr_curr_bio;
1962         if (sbio->page_count == 0) {
1963                 struct bio *bio;
1964
1965                 sbio->physical = spage->physical_for_dev_replace;
1966                 sbio->logical = spage->logical;
1967                 sbio->dev = sctx->wr_tgtdev;
1968                 bio = sbio->bio;
1969                 if (!bio) {
1970                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1971                         sbio->bio = bio;
1972                 }
1973
1974                 bio->bi_private = sbio;
1975                 bio->bi_end_io = scrub_wr_bio_end_io;
1976                 bio_set_dev(bio, sbio->dev->bdev);
1977                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1978                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1979                 sbio->status = 0;
1980         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1981                    spage->physical_for_dev_replace ||
1982                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1983                    spage->logical) {
1984                 scrub_wr_submit(sctx);
1985                 goto again;
1986         }
1987
1988         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1989         if (ret != PAGE_SIZE) {
1990                 if (sbio->page_count < 1) {
1991                         bio_put(sbio->bio);
1992                         sbio->bio = NULL;
1993                         mutex_unlock(&sctx->wr_lock);
1994                         return -EIO;
1995                 }
1996                 scrub_wr_submit(sctx);
1997                 goto again;
1998         }
1999
2000         sbio->pagev[sbio->page_count] = spage;
2001         scrub_page_get(spage);
2002         sbio->page_count++;
2003         if (sbio->page_count == sctx->pages_per_wr_bio)
2004                 scrub_wr_submit(sctx);
2005         mutex_unlock(&sctx->wr_lock);
2006
2007         return 0;
2008 }
2009
2010 static void scrub_wr_submit(struct scrub_ctx *sctx)
2011 {
2012         struct scrub_bio *sbio;
2013
2014         if (!sctx->wr_curr_bio)
2015                 return;
2016
2017         sbio = sctx->wr_curr_bio;
2018         sctx->wr_curr_bio = NULL;
2019         WARN_ON(!sbio->bio->bi_disk);
2020         scrub_pending_bio_inc(sctx);
2021         /* process all writes in a single worker thread. Then the block layer
2022          * orders the requests before sending them to the driver which
2023          * doubled the write performance on spinning disks when measured
2024          * with Linux 3.5 */
2025         btrfsic_submit_bio(sbio->bio);
2026 }
2027
2028 static void scrub_wr_bio_end_io(struct bio *bio)
2029 {
2030         struct scrub_bio *sbio = bio->bi_private;
2031         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2032
2033         sbio->status = bio->bi_status;
2034         sbio->bio = bio;
2035
2036         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2037                          scrub_wr_bio_end_io_worker, NULL, NULL);
2038         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2039 }
2040
2041 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2042 {
2043         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2044         struct scrub_ctx *sctx = sbio->sctx;
2045         int i;
2046
2047         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2048         if (sbio->status) {
2049                 struct btrfs_dev_replace *dev_replace =
2050                         &sbio->sctx->fs_info->dev_replace;
2051
2052                 for (i = 0; i < sbio->page_count; i++) {
2053                         struct scrub_page *spage = sbio->pagev[i];
2054
2055                         spage->io_error = 1;
2056                         btrfs_dev_replace_stats_inc(&dev_replace->
2057                                                     num_write_errors);
2058                 }
2059         }
2060
2061         for (i = 0; i < sbio->page_count; i++)
2062                 scrub_page_put(sbio->pagev[i]);
2063
2064         bio_put(sbio->bio);
2065         kfree(sbio);
2066         scrub_pending_bio_dec(sctx);
2067 }
2068
2069 static int scrub_checksum(struct scrub_block *sblock)
2070 {
2071         u64 flags;
2072         int ret;
2073
2074         /*
2075          * No need to initialize these stats currently,
2076          * because this function only use return value
2077          * instead of these stats value.
2078          *
2079          * Todo:
2080          * always use stats
2081          */
2082         sblock->header_error = 0;
2083         sblock->generation_error = 0;
2084         sblock->checksum_error = 0;
2085
2086         WARN_ON(sblock->page_count < 1);
2087         flags = sblock->pagev[0]->flags;
2088         ret = 0;
2089         if (flags & BTRFS_EXTENT_FLAG_DATA)
2090                 ret = scrub_checksum_data(sblock);
2091         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2092                 ret = scrub_checksum_tree_block(sblock);
2093         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2094                 (void)scrub_checksum_super(sblock);
2095         else
2096                 WARN_ON(1);
2097         if (ret)
2098                 scrub_handle_errored_block(sblock);
2099
2100         return ret;
2101 }
2102
2103 static int scrub_checksum_data(struct scrub_block *sblock)
2104 {
2105         struct scrub_ctx *sctx = sblock->sctx;
2106         u8 csum[BTRFS_CSUM_SIZE];
2107         u8 *on_disk_csum;
2108         struct page *page;
2109         void *buffer;
2110         u32 crc = ~(u32)0;
2111         u64 len;
2112         int index;
2113
2114         BUG_ON(sblock->page_count < 1);
2115         if (!sblock->pagev[0]->have_csum)
2116                 return 0;
2117
2118         on_disk_csum = sblock->pagev[0]->csum;
2119         page = sblock->pagev[0]->page;
2120         buffer = kmap_atomic(page);
2121
2122         len = sctx->fs_info->sectorsize;
2123         index = 0;
2124         for (;;) {
2125                 u64 l = min_t(u64, len, PAGE_SIZE);
2126
2127                 crc = btrfs_csum_data(buffer, crc, l);
2128                 kunmap_atomic(buffer);
2129                 len -= l;
2130                 if (len == 0)
2131                         break;
2132                 index++;
2133                 BUG_ON(index >= sblock->page_count);
2134                 BUG_ON(!sblock->pagev[index]->page);
2135                 page = sblock->pagev[index]->page;
2136                 buffer = kmap_atomic(page);
2137         }
2138
2139         btrfs_csum_final(crc, csum);
2140         if (memcmp(csum, on_disk_csum, sctx->csum_size))
2141                 sblock->checksum_error = 1;
2142
2143         return sblock->checksum_error;
2144 }
2145
2146 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2147 {
2148         struct scrub_ctx *sctx = sblock->sctx;
2149         struct btrfs_header *h;
2150         struct btrfs_fs_info *fs_info = sctx->fs_info;
2151         u8 calculated_csum[BTRFS_CSUM_SIZE];
2152         u8 on_disk_csum[BTRFS_CSUM_SIZE];
2153         struct page *page;
2154         void *mapped_buffer;
2155         u64 mapped_size;
2156         void *p;
2157         u32 crc = ~(u32)0;
2158         u64 len;
2159         int index;
2160
2161         BUG_ON(sblock->page_count < 1);
2162         page = sblock->pagev[0]->page;
2163         mapped_buffer = kmap_atomic(page);
2164         h = (struct btrfs_header *)mapped_buffer;
2165         memcpy(on_disk_csum, h->csum, sctx->csum_size);
2166
2167         /*
2168          * we don't use the getter functions here, as we
2169          * a) don't have an extent buffer and
2170          * b) the page is already kmapped
2171          */
2172         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2173                 sblock->header_error = 1;
2174
2175         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2176                 sblock->header_error = 1;
2177                 sblock->generation_error = 1;
2178         }
2179
2180         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2181                 sblock->header_error = 1;
2182
2183         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2184                    BTRFS_UUID_SIZE))
2185                 sblock->header_error = 1;
2186
2187         len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2188         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2189         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2190         index = 0;
2191         for (;;) {
2192                 u64 l = min_t(u64, len, mapped_size);
2193
2194                 crc = btrfs_csum_data(p, crc, l);
2195                 kunmap_atomic(mapped_buffer);
2196                 len -= l;
2197                 if (len == 0)
2198                         break;
2199                 index++;
2200                 BUG_ON(index >= sblock->page_count);
2201                 BUG_ON(!sblock->pagev[index]->page);
2202                 page = sblock->pagev[index]->page;
2203                 mapped_buffer = kmap_atomic(page);
2204                 mapped_size = PAGE_SIZE;
2205                 p = mapped_buffer;
2206         }
2207
2208         btrfs_csum_final(crc, calculated_csum);
2209         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2210                 sblock->checksum_error = 1;
2211
2212         return sblock->header_error || sblock->checksum_error;
2213 }
2214
2215 static int scrub_checksum_super(struct scrub_block *sblock)
2216 {
2217         struct btrfs_super_block *s;
2218         struct scrub_ctx *sctx = sblock->sctx;
2219         u8 calculated_csum[BTRFS_CSUM_SIZE];
2220         u8 on_disk_csum[BTRFS_CSUM_SIZE];
2221         struct page *page;
2222         void *mapped_buffer;
2223         u64 mapped_size;
2224         void *p;
2225         u32 crc = ~(u32)0;
2226         int fail_gen = 0;
2227         int fail_cor = 0;
2228         u64 len;
2229         int index;
2230
2231         BUG_ON(sblock->page_count < 1);
2232         page = sblock->pagev[0]->page;
2233         mapped_buffer = kmap_atomic(page);
2234         s = (struct btrfs_super_block *)mapped_buffer;
2235         memcpy(on_disk_csum, s->csum, sctx->csum_size);
2236
2237         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2238                 ++fail_cor;
2239
2240         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2241                 ++fail_gen;
2242
2243         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2244                 ++fail_cor;
2245
2246         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2247         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2248         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2249         index = 0;
2250         for (;;) {
2251                 u64 l = min_t(u64, len, mapped_size);
2252
2253                 crc = btrfs_csum_data(p, crc, l);
2254                 kunmap_atomic(mapped_buffer);
2255                 len -= l;
2256                 if (len == 0)
2257                         break;
2258                 index++;
2259                 BUG_ON(index >= sblock->page_count);
2260                 BUG_ON(!sblock->pagev[index]->page);
2261                 page = sblock->pagev[index]->page;
2262                 mapped_buffer = kmap_atomic(page);
2263                 mapped_size = PAGE_SIZE;
2264                 p = mapped_buffer;
2265         }
2266
2267         btrfs_csum_final(crc, calculated_csum);
2268         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2269                 ++fail_cor;
2270
2271         if (fail_cor + fail_gen) {
2272                 /*
2273                  * if we find an error in a super block, we just report it.
2274                  * They will get written with the next transaction commit
2275                  * anyway
2276                  */
2277                 spin_lock(&sctx->stat_lock);
2278                 ++sctx->stat.super_errors;
2279                 spin_unlock(&sctx->stat_lock);
2280                 if (fail_cor)
2281                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2282                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2283                 else
2284                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2285                                 BTRFS_DEV_STAT_GENERATION_ERRS);
2286         }
2287
2288         return fail_cor + fail_gen;
2289 }
2290
2291 static void scrub_block_get(struct scrub_block *sblock)
2292 {
2293         refcount_inc(&sblock->refs);
2294 }
2295
2296 static void scrub_block_put(struct scrub_block *sblock)
2297 {
2298         if (refcount_dec_and_test(&sblock->refs)) {
2299                 int i;
2300
2301                 if (sblock->sparity)
2302                         scrub_parity_put(sblock->sparity);
2303
2304                 for (i = 0; i < sblock->page_count; i++)
2305                         scrub_page_put(sblock->pagev[i]);
2306                 kfree(sblock);
2307         }
2308 }
2309
2310 static void scrub_page_get(struct scrub_page *spage)
2311 {
2312         atomic_inc(&spage->refs);
2313 }
2314
2315 static void scrub_page_put(struct scrub_page *spage)
2316 {
2317         if (atomic_dec_and_test(&spage->refs)) {
2318                 if (spage->page)
2319                         __free_page(spage->page);
2320                 kfree(spage);
2321         }
2322 }
2323
2324 static void scrub_submit(struct scrub_ctx *sctx)
2325 {
2326         struct scrub_bio *sbio;
2327
2328         if (sctx->curr == -1)
2329                 return;
2330
2331         sbio = sctx->bios[sctx->curr];
2332         sctx->curr = -1;
2333         scrub_pending_bio_inc(sctx);
2334         btrfsic_submit_bio(sbio->bio);
2335 }
2336
2337 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2338                                     struct scrub_page *spage)
2339 {
2340         struct scrub_block *sblock = spage->sblock;
2341         struct scrub_bio *sbio;
2342         int ret;
2343
2344 again:
2345         /*
2346          * grab a fresh bio or wait for one to become available
2347          */
2348         while (sctx->curr == -1) {
2349                 spin_lock(&sctx->list_lock);
2350                 sctx->curr = sctx->first_free;
2351                 if (sctx->curr != -1) {
2352                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2353                         sctx->bios[sctx->curr]->next_free = -1;
2354                         sctx->bios[sctx->curr]->page_count = 0;
2355                         spin_unlock(&sctx->list_lock);
2356                 } else {
2357                         spin_unlock(&sctx->list_lock);
2358                         wait_event(sctx->list_wait, sctx->first_free != -1);
2359                 }
2360         }
2361         sbio = sctx->bios[sctx->curr];
2362         if (sbio->page_count == 0) {
2363                 struct bio *bio;
2364
2365                 sbio->physical = spage->physical;
2366                 sbio->logical = spage->logical;
2367                 sbio->dev = spage->dev;
2368                 bio = sbio->bio;
2369                 if (!bio) {
2370                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2371                         sbio->bio = bio;
2372                 }
2373
2374                 bio->bi_private = sbio;
2375                 bio->bi_end_io = scrub_bio_end_io;
2376                 bio_set_dev(bio, sbio->dev->bdev);
2377                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2378                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2379                 sbio->status = 0;
2380         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2381                    spage->physical ||
2382                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2383                    spage->logical ||
2384                    sbio->dev != spage->dev) {
2385                 scrub_submit(sctx);
2386                 goto again;
2387         }
2388
2389         sbio->pagev[sbio->page_count] = spage;
2390         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2391         if (ret != PAGE_SIZE) {
2392                 if (sbio->page_count < 1) {
2393                         bio_put(sbio->bio);
2394                         sbio->bio = NULL;
2395                         return -EIO;
2396                 }
2397                 scrub_submit(sctx);
2398                 goto again;
2399         }
2400
2401         scrub_block_get(sblock); /* one for the page added to the bio */
2402         atomic_inc(&sblock->outstanding_pages);
2403         sbio->page_count++;
2404         if (sbio->page_count == sctx->pages_per_rd_bio)
2405                 scrub_submit(sctx);
2406
2407         return 0;
2408 }
2409
2410 static void scrub_missing_raid56_end_io(struct bio *bio)
2411 {
2412         struct scrub_block *sblock = bio->bi_private;
2413         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2414
2415         if (bio->bi_status)
2416                 sblock->no_io_error_seen = 0;
2417
2418         bio_put(bio);
2419
2420         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2421 }
2422
2423 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2424 {
2425         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2426         struct scrub_ctx *sctx = sblock->sctx;
2427         struct btrfs_fs_info *fs_info = sctx->fs_info;
2428         u64 logical;
2429         struct btrfs_device *dev;
2430
2431         logical = sblock->pagev[0]->logical;
2432         dev = sblock->pagev[0]->dev;
2433
2434         if (sblock->no_io_error_seen)
2435                 scrub_recheck_block_checksum(sblock);
2436
2437         if (!sblock->no_io_error_seen) {
2438                 spin_lock(&sctx->stat_lock);
2439                 sctx->stat.read_errors++;
2440                 spin_unlock(&sctx->stat_lock);
2441                 btrfs_err_rl_in_rcu(fs_info,
2442                         "IO error rebuilding logical %llu for dev %s",
2443                         logical, rcu_str_deref(dev->name));
2444         } else if (sblock->header_error || sblock->checksum_error) {
2445                 spin_lock(&sctx->stat_lock);
2446                 sctx->stat.uncorrectable_errors++;
2447                 spin_unlock(&sctx->stat_lock);
2448                 btrfs_err_rl_in_rcu(fs_info,
2449                         "failed to rebuild valid logical %llu for dev %s",
2450                         logical, rcu_str_deref(dev->name));
2451         } else {
2452                 scrub_write_block_to_dev_replace(sblock);
2453         }
2454
2455         scrub_block_put(sblock);
2456
2457         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2458                 mutex_lock(&sctx->wr_lock);
2459                 scrub_wr_submit(sctx);
2460                 mutex_unlock(&sctx->wr_lock);
2461         }
2462
2463         scrub_pending_bio_dec(sctx);
2464 }
2465
2466 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2467 {
2468         struct scrub_ctx *sctx = sblock->sctx;
2469         struct btrfs_fs_info *fs_info = sctx->fs_info;
2470         u64 length = sblock->page_count * PAGE_SIZE;
2471         u64 logical = sblock->pagev[0]->logical;
2472         struct btrfs_bio *bbio = NULL;
2473         struct bio *bio;
2474         struct btrfs_raid_bio *rbio;
2475         int ret;
2476         int i;
2477
2478         btrfs_bio_counter_inc_blocked(fs_info);
2479         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2480                         &length, &bbio);
2481         if (ret || !bbio || !bbio->raid_map)
2482                 goto bbio_out;
2483
2484         if (WARN_ON(!sctx->is_dev_replace ||
2485                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2486                 /*
2487                  * We shouldn't be scrubbing a missing device. Even for dev
2488                  * replace, we should only get here for RAID 5/6. We either
2489                  * managed to mount something with no mirrors remaining or
2490                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2491                  */
2492                 goto bbio_out;
2493         }
2494
2495         bio = btrfs_io_bio_alloc(0);
2496         bio->bi_iter.bi_sector = logical >> 9;
2497         bio->bi_private = sblock;
2498         bio->bi_end_io = scrub_missing_raid56_end_io;
2499
2500         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2501         if (!rbio)
2502                 goto rbio_out;
2503
2504         for (i = 0; i < sblock->page_count; i++) {
2505                 struct scrub_page *spage = sblock->pagev[i];
2506
2507                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2508         }
2509
2510         btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2511                         scrub_missing_raid56_worker, NULL, NULL);
2512         scrub_block_get(sblock);
2513         scrub_pending_bio_inc(sctx);
2514         raid56_submit_missing_rbio(rbio);
2515         return;
2516
2517 rbio_out:
2518         bio_put(bio);
2519 bbio_out:
2520         btrfs_bio_counter_dec(fs_info);
2521         btrfs_put_bbio(bbio);
2522         spin_lock(&sctx->stat_lock);
2523         sctx->stat.malloc_errors++;
2524         spin_unlock(&sctx->stat_lock);
2525 }
2526
2527 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2528                        u64 physical, struct btrfs_device *dev, u64 flags,
2529                        u64 gen, int mirror_num, u8 *csum, int force,
2530                        u64 physical_for_dev_replace)
2531 {
2532         struct scrub_block *sblock;
2533         int index;
2534
2535         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2536         if (!sblock) {
2537                 spin_lock(&sctx->stat_lock);
2538                 sctx->stat.malloc_errors++;
2539                 spin_unlock(&sctx->stat_lock);
2540                 return -ENOMEM;
2541         }
2542
2543         /* one ref inside this function, plus one for each page added to
2544          * a bio later on */
2545         refcount_set(&sblock->refs, 1);
2546         sblock->sctx = sctx;
2547         sblock->no_io_error_seen = 1;
2548
2549         for (index = 0; len > 0; index++) {
2550                 struct scrub_page *spage;
2551                 u64 l = min_t(u64, len, PAGE_SIZE);
2552
2553                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2554                 if (!spage) {
2555 leave_nomem:
2556                         spin_lock(&sctx->stat_lock);
2557                         sctx->stat.malloc_errors++;
2558                         spin_unlock(&sctx->stat_lock);
2559                         scrub_block_put(sblock);
2560                         return -ENOMEM;
2561                 }
2562                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2563                 scrub_page_get(spage);
2564                 sblock->pagev[index] = spage;
2565                 spage->sblock = sblock;
2566                 spage->dev = dev;
2567                 spage->flags = flags;
2568                 spage->generation = gen;
2569                 spage->logical = logical;
2570                 spage->physical = physical;
2571                 spage->physical_for_dev_replace = physical_for_dev_replace;
2572                 spage->mirror_num = mirror_num;
2573                 if (csum) {
2574                         spage->have_csum = 1;
2575                         memcpy(spage->csum, csum, sctx->csum_size);
2576                 } else {
2577                         spage->have_csum = 0;
2578                 }
2579                 sblock->page_count++;
2580                 spage->page = alloc_page(GFP_KERNEL);
2581                 if (!spage->page)
2582                         goto leave_nomem;
2583                 len -= l;
2584                 logical += l;
2585                 physical += l;
2586                 physical_for_dev_replace += l;
2587         }
2588
2589         WARN_ON(sblock->page_count == 0);
2590         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2591                 /*
2592                  * This case should only be hit for RAID 5/6 device replace. See
2593                  * the comment in scrub_missing_raid56_pages() for details.
2594                  */
2595                 scrub_missing_raid56_pages(sblock);
2596         } else {
2597                 for (index = 0; index < sblock->page_count; index++) {
2598                         struct scrub_page *spage = sblock->pagev[index];
2599                         int ret;
2600
2601                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2602                         if (ret) {
2603                                 scrub_block_put(sblock);
2604                                 return ret;
2605                         }
2606                 }
2607
2608                 if (force)
2609                         scrub_submit(sctx);
2610         }
2611
2612         /* last one frees, either here or in bio completion for last page */
2613         scrub_block_put(sblock);
2614         return 0;
2615 }
2616
2617 static void scrub_bio_end_io(struct bio *bio)
2618 {
2619         struct scrub_bio *sbio = bio->bi_private;
2620         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2621
2622         sbio->status = bio->bi_status;
2623         sbio->bio = bio;
2624
2625         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2626 }
2627
2628 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2629 {
2630         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2631         struct scrub_ctx *sctx = sbio->sctx;
2632         int i;
2633
2634         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2635         if (sbio->status) {
2636                 for (i = 0; i < sbio->page_count; i++) {
2637                         struct scrub_page *spage = sbio->pagev[i];
2638
2639                         spage->io_error = 1;
2640                         spage->sblock->no_io_error_seen = 0;
2641                 }
2642         }
2643
2644         /* now complete the scrub_block items that have all pages completed */
2645         for (i = 0; i < sbio->page_count; i++) {
2646                 struct scrub_page *spage = sbio->pagev[i];
2647                 struct scrub_block *sblock = spage->sblock;
2648
2649                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2650                         scrub_block_complete(sblock);
2651                 scrub_block_put(sblock);
2652         }
2653
2654         bio_put(sbio->bio);
2655         sbio->bio = NULL;
2656         spin_lock(&sctx->list_lock);
2657         sbio->next_free = sctx->first_free;
2658         sctx->first_free = sbio->index;
2659         spin_unlock(&sctx->list_lock);
2660
2661         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2662                 mutex_lock(&sctx->wr_lock);
2663                 scrub_wr_submit(sctx);
2664                 mutex_unlock(&sctx->wr_lock);
2665         }
2666
2667         scrub_pending_bio_dec(sctx);
2668 }
2669
2670 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2671                                        unsigned long *bitmap,
2672                                        u64 start, u64 len)
2673 {
2674         u64 offset;
2675         u64 nsectors64;
2676         u32 nsectors;
2677         int sectorsize = sparity->sctx->fs_info->sectorsize;
2678
2679         if (len >= sparity->stripe_len) {
2680                 bitmap_set(bitmap, 0, sparity->nsectors);
2681                 return;
2682         }
2683
2684         start -= sparity->logic_start;
2685         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2686         offset = div_u64(offset, sectorsize);
2687         nsectors64 = div_u64(len, sectorsize);
2688
2689         ASSERT(nsectors64 < UINT_MAX);
2690         nsectors = (u32)nsectors64;
2691
2692         if (offset + nsectors <= sparity->nsectors) {
2693                 bitmap_set(bitmap, offset, nsectors);
2694                 return;
2695         }
2696
2697         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2698         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2699 }
2700
2701 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2702                                                    u64 start, u64 len)
2703 {
2704         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2705 }
2706
2707 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2708                                                   u64 start, u64 len)
2709 {
2710         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2711 }
2712
2713 static void scrub_block_complete(struct scrub_block *sblock)
2714 {
2715         int corrupted = 0;
2716
2717         if (!sblock->no_io_error_seen) {
2718                 corrupted = 1;
2719                 scrub_handle_errored_block(sblock);
2720         } else {
2721                 /*
2722                  * if has checksum error, write via repair mechanism in
2723                  * dev replace case, otherwise write here in dev replace
2724                  * case.
2725                  */
2726                 corrupted = scrub_checksum(sblock);
2727                 if (!corrupted && sblock->sctx->is_dev_replace)
2728                         scrub_write_block_to_dev_replace(sblock);
2729         }
2730
2731         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2732                 u64 start = sblock->pagev[0]->logical;
2733                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2734                           PAGE_SIZE;
2735
2736                 scrub_parity_mark_sectors_error(sblock->sparity,
2737                                                 start, end - start);
2738         }
2739 }
2740
2741 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2742 {
2743         struct btrfs_ordered_sum *sum = NULL;
2744         unsigned long index;
2745         unsigned long num_sectors;
2746
2747         while (!list_empty(&sctx->csum_list)) {
2748                 sum = list_first_entry(&sctx->csum_list,
2749                                        struct btrfs_ordered_sum, list);
2750                 if (sum->bytenr > logical)
2751                         return 0;
2752                 if (sum->bytenr + sum->len > logical)
2753                         break;
2754
2755                 ++sctx->stat.csum_discards;
2756                 list_del(&sum->list);
2757                 kfree(sum);
2758                 sum = NULL;
2759         }
2760         if (!sum)
2761                 return 0;
2762
2763         index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2764         ASSERT(index < UINT_MAX);
2765
2766         num_sectors = sum->len / sctx->fs_info->sectorsize;
2767         memcpy(csum, sum->sums + index, sctx->csum_size);
2768         if (index == num_sectors - 1) {
2769                 list_del(&sum->list);
2770                 kfree(sum);
2771         }
2772         return 1;
2773 }
2774
2775 /* scrub extent tries to collect up to 64 kB for each bio */
2776 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2777                         u64 logical, u64 len,
2778                         u64 physical, struct btrfs_device *dev, u64 flags,
2779                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2780 {
2781         int ret;
2782         u8 csum[BTRFS_CSUM_SIZE];
2783         u32 blocksize;
2784
2785         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2786                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2787                         blocksize = map->stripe_len;
2788                 else
2789                         blocksize = sctx->fs_info->sectorsize;
2790                 spin_lock(&sctx->stat_lock);
2791                 sctx->stat.data_extents_scrubbed++;
2792                 sctx->stat.data_bytes_scrubbed += len;
2793                 spin_unlock(&sctx->stat_lock);
2794         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2795                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2796                         blocksize = map->stripe_len;
2797                 else
2798                         blocksize = sctx->fs_info->nodesize;
2799                 spin_lock(&sctx->stat_lock);
2800                 sctx->stat.tree_extents_scrubbed++;
2801                 sctx->stat.tree_bytes_scrubbed += len;
2802                 spin_unlock(&sctx->stat_lock);
2803         } else {
2804                 blocksize = sctx->fs_info->sectorsize;
2805                 WARN_ON(1);
2806         }
2807
2808         while (len) {
2809                 u64 l = min_t(u64, len, blocksize);
2810                 int have_csum = 0;
2811
2812                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2813                         /* push csums to sbio */
2814                         have_csum = scrub_find_csum(sctx, logical, csum);
2815                         if (have_csum == 0)
2816                                 ++sctx->stat.no_csum;
2817                         if (sctx->is_dev_replace && !have_csum) {
2818                                 ret = copy_nocow_pages(sctx, logical, l,
2819                                                        mirror_num,
2820                                                       physical_for_dev_replace);
2821                                 goto behind_scrub_pages;
2822                         }
2823                 }
2824                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2825                                   mirror_num, have_csum ? csum : NULL, 0,
2826                                   physical_for_dev_replace);
2827 behind_scrub_pages:
2828                 if (ret)
2829                         return ret;
2830                 len -= l;
2831                 logical += l;
2832                 physical += l;
2833                 physical_for_dev_replace += l;
2834         }
2835         return 0;
2836 }
2837
2838 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2839                                   u64 logical, u64 len,
2840                                   u64 physical, struct btrfs_device *dev,
2841                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2842 {
2843         struct scrub_ctx *sctx = sparity->sctx;
2844         struct scrub_block *sblock;
2845         int index;
2846
2847         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2848         if (!sblock) {
2849                 spin_lock(&sctx->stat_lock);
2850                 sctx->stat.malloc_errors++;
2851                 spin_unlock(&sctx->stat_lock);
2852                 return -ENOMEM;
2853         }
2854
2855         /* one ref inside this function, plus one for each page added to
2856          * a bio later on */
2857         refcount_set(&sblock->refs, 1);
2858         sblock->sctx = sctx;
2859         sblock->no_io_error_seen = 1;
2860         sblock->sparity = sparity;
2861         scrub_parity_get(sparity);
2862
2863         for (index = 0; len > 0; index++) {
2864                 struct scrub_page *spage;
2865                 u64 l = min_t(u64, len, PAGE_SIZE);
2866
2867                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2868                 if (!spage) {
2869 leave_nomem:
2870                         spin_lock(&sctx->stat_lock);
2871                         sctx->stat.malloc_errors++;
2872                         spin_unlock(&sctx->stat_lock);
2873                         scrub_block_put(sblock);
2874                         return -ENOMEM;
2875                 }
2876                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2877                 /* For scrub block */
2878                 scrub_page_get(spage);
2879                 sblock->pagev[index] = spage;
2880                 /* For scrub parity */
2881                 scrub_page_get(spage);
2882                 list_add_tail(&spage->list, &sparity->spages);
2883                 spage->sblock = sblock;
2884                 spage->dev = dev;
2885                 spage->flags = flags;
2886                 spage->generation = gen;
2887                 spage->logical = logical;
2888                 spage->physical = physical;
2889                 spage->mirror_num = mirror_num;
2890                 if (csum) {
2891                         spage->have_csum = 1;
2892                         memcpy(spage->csum, csum, sctx->csum_size);
2893                 } else {
2894                         spage->have_csum = 0;
2895                 }
2896                 sblock->page_count++;
2897                 spage->page = alloc_page(GFP_KERNEL);
2898                 if (!spage->page)
2899                         goto leave_nomem;
2900                 len -= l;
2901                 logical += l;
2902                 physical += l;
2903         }
2904
2905         WARN_ON(sblock->page_count == 0);
2906         for (index = 0; index < sblock->page_count; index++) {
2907                 struct scrub_page *spage = sblock->pagev[index];
2908                 int ret;
2909
2910                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2911                 if (ret) {
2912                         scrub_block_put(sblock);
2913                         return ret;
2914                 }
2915         }
2916
2917         /* last one frees, either here or in bio completion for last page */
2918         scrub_block_put(sblock);
2919         return 0;
2920 }
2921
2922 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2923                                    u64 logical, u64 len,
2924                                    u64 physical, struct btrfs_device *dev,
2925                                    u64 flags, u64 gen, int mirror_num)
2926 {
2927         struct scrub_ctx *sctx = sparity->sctx;
2928         int ret;
2929         u8 csum[BTRFS_CSUM_SIZE];
2930         u32 blocksize;
2931
2932         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2933                 scrub_parity_mark_sectors_error(sparity, logical, len);
2934                 return 0;
2935         }
2936
2937         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2938                 blocksize = sparity->stripe_len;
2939         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2940                 blocksize = sparity->stripe_len;
2941         } else {
2942                 blocksize = sctx->fs_info->sectorsize;
2943                 WARN_ON(1);
2944         }
2945
2946         while (len) {
2947                 u64 l = min_t(u64, len, blocksize);
2948                 int have_csum = 0;
2949
2950                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2951                         /* push csums to sbio */
2952                         have_csum = scrub_find_csum(sctx, logical, csum);
2953                         if (have_csum == 0)
2954                                 goto skip;
2955                 }
2956                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2957                                              flags, gen, mirror_num,
2958                                              have_csum ? csum : NULL);
2959                 if (ret)
2960                         return ret;
2961 skip:
2962                 len -= l;
2963                 logical += l;
2964                 physical += l;
2965         }
2966         return 0;
2967 }
2968
2969 /*
2970  * Given a physical address, this will calculate it's
2971  * logical offset. if this is a parity stripe, it will return
2972  * the most left data stripe's logical offset.
2973  *
2974  * return 0 if it is a data stripe, 1 means parity stripe.
2975  */
2976 static int get_raid56_logic_offset(u64 physical, int num,
2977                                    struct map_lookup *map, u64 *offset,
2978                                    u64 *stripe_start)
2979 {
2980         int i;
2981         int j = 0;
2982         u64 stripe_nr;
2983         u64 last_offset;
2984         u32 stripe_index;
2985         u32 rot;
2986
2987         last_offset = (physical - map->stripes[num].physical) *
2988                       nr_data_stripes(map);
2989         if (stripe_start)
2990                 *stripe_start = last_offset;
2991
2992         *offset = last_offset;
2993         for (i = 0; i < nr_data_stripes(map); i++) {
2994                 *offset = last_offset + i * map->stripe_len;
2995
2996                 stripe_nr = div64_u64(*offset, map->stripe_len);
2997                 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2998
2999                 /* Work out the disk rotation on this stripe-set */
3000                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
3001                 /* calculate which stripe this data locates */
3002                 rot += i;
3003                 stripe_index = rot % map->num_stripes;
3004                 if (stripe_index == num)
3005                         return 0;
3006                 if (stripe_index < num)
3007                         j++;
3008         }
3009         *offset = last_offset + j * map->stripe_len;
3010         return 1;
3011 }
3012
3013 static void scrub_free_parity(struct scrub_parity *sparity)
3014 {
3015         struct scrub_ctx *sctx = sparity->sctx;
3016         struct scrub_page *curr, *next;
3017         int nbits;
3018
3019         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
3020         if (nbits) {
3021                 spin_lock(&sctx->stat_lock);
3022                 sctx->stat.read_errors += nbits;
3023                 sctx->stat.uncorrectable_errors += nbits;
3024                 spin_unlock(&sctx->stat_lock);
3025         }
3026
3027         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
3028                 list_del_init(&curr->list);
3029                 scrub_page_put(curr);
3030         }
3031
3032         kfree(sparity);
3033 }
3034
3035 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
3036 {
3037         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3038                                                     work);
3039         struct scrub_ctx *sctx = sparity->sctx;
3040
3041         scrub_free_parity(sparity);
3042         scrub_pending_bio_dec(sctx);
3043 }
3044
3045 static void scrub_parity_bio_endio(struct bio *bio)
3046 {
3047         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3048         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3049
3050         if (bio->bi_status)
3051                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3052                           sparity->nsectors);
3053
3054         bio_put(bio);
3055
3056         btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3057                         scrub_parity_bio_endio_worker, NULL, NULL);
3058         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3059 }
3060
3061 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3062 {
3063         struct scrub_ctx *sctx = sparity->sctx;
3064         struct btrfs_fs_info *fs_info = sctx->fs_info;
3065         struct bio *bio;
3066         struct btrfs_raid_bio *rbio;
3067         struct btrfs_bio *bbio = NULL;
3068         u64 length;
3069         int ret;
3070
3071         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3072                            sparity->nsectors))
3073                 goto out;
3074
3075         length = sparity->logic_end - sparity->logic_start;
3076
3077         btrfs_bio_counter_inc_blocked(fs_info);
3078         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3079                                &length, &bbio);
3080         if (ret || !bbio || !bbio->raid_map)
3081                 goto bbio_out;
3082
3083         bio = btrfs_io_bio_alloc(0);
3084         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3085         bio->bi_private = sparity;
3086         bio->bi_end_io = scrub_parity_bio_endio;
3087
3088         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3089                                               length, sparity->scrub_dev,
3090                                               sparity->dbitmap,
3091                                               sparity->nsectors);
3092         if (!rbio)
3093                 goto rbio_out;
3094
3095         scrub_pending_bio_inc(sctx);
3096         raid56_parity_submit_scrub_rbio(rbio);
3097         return;
3098
3099 rbio_out:
3100         bio_put(bio);
3101 bbio_out:
3102         btrfs_bio_counter_dec(fs_info);
3103         btrfs_put_bbio(bbio);
3104         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3105                   sparity->nsectors);
3106         spin_lock(&sctx->stat_lock);
3107         sctx->stat.malloc_errors++;
3108         spin_unlock(&sctx->stat_lock);
3109 out:
3110         scrub_free_parity(sparity);
3111 }
3112
3113 static inline int scrub_calc_parity_bitmap_len(int nsectors)
3114 {
3115         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3116 }
3117
3118 static void scrub_parity_get(struct scrub_parity *sparity)
3119 {
3120         refcount_inc(&sparity->refs);
3121 }
3122
3123 static void scrub_parity_put(struct scrub_parity *sparity)
3124 {
3125         if (!refcount_dec_and_test(&sparity->refs))
3126                 return;
3127
3128         scrub_parity_check_and_repair(sparity);
3129 }
3130
3131 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3132                                                   struct map_lookup *map,
3133                                                   struct btrfs_device *sdev,
3134                                                   struct btrfs_path *path,
3135                                                   u64 logic_start,
3136                                                   u64 logic_end)
3137 {
3138         struct btrfs_fs_info *fs_info = sctx->fs_info;
3139         struct btrfs_root *root = fs_info->extent_root;
3140         struct btrfs_root *csum_root = fs_info->csum_root;
3141         struct btrfs_extent_item *extent;
3142         struct btrfs_bio *bbio = NULL;
3143         u64 flags;
3144         int ret;
3145         int slot;
3146         struct extent_buffer *l;
3147         struct btrfs_key key;
3148         u64 generation;
3149         u64 extent_logical;
3150         u64 extent_physical;
3151         u64 extent_len;
3152         u64 mapped_length;
3153         struct btrfs_device *extent_dev;
3154         struct scrub_parity *sparity;
3155         int nsectors;
3156         int bitmap_len;
3157         int extent_mirror_num;
3158         int stop_loop = 0;
3159
3160         nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3161         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3162         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3163                           GFP_NOFS);
3164         if (!sparity) {
3165                 spin_lock(&sctx->stat_lock);
3166                 sctx->stat.malloc_errors++;
3167                 spin_unlock(&sctx->stat_lock);
3168                 return -ENOMEM;
3169         }
3170
3171         sparity->stripe_len = map->stripe_len;
3172         sparity->nsectors = nsectors;
3173         sparity->sctx = sctx;
3174         sparity->scrub_dev = sdev;
3175         sparity->logic_start = logic_start;
3176         sparity->logic_end = logic_end;
3177         refcount_set(&sparity->refs, 1);
3178         INIT_LIST_HEAD(&sparity->spages);
3179         sparity->dbitmap = sparity->bitmap;
3180         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3181
3182         ret = 0;
3183         while (logic_start < logic_end) {
3184                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3185                         key.type = BTRFS_METADATA_ITEM_KEY;
3186                 else
3187                         key.type = BTRFS_EXTENT_ITEM_KEY;
3188                 key.objectid = logic_start;
3189                 key.offset = (u64)-1;
3190
3191                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3192                 if (ret < 0)
3193                         goto out;
3194
3195                 if (ret > 0) {
3196                         ret = btrfs_previous_extent_item(root, path, 0);
3197                         if (ret < 0)
3198                                 goto out;
3199                         if (ret > 0) {
3200                                 btrfs_release_path(path);
3201                                 ret = btrfs_search_slot(NULL, root, &key,
3202                                                         path, 0, 0);
3203                                 if (ret < 0)
3204                                         goto out;
3205                         }
3206                 }
3207
3208                 stop_loop = 0;
3209                 while (1) {
3210                         u64 bytes;
3211
3212                         l = path->nodes[0];
3213                         slot = path->slots[0];
3214                         if (slot >= btrfs_header_nritems(l)) {
3215                                 ret = btrfs_next_leaf(root, path);
3216                                 if (ret == 0)
3217                                         continue;
3218                                 if (ret < 0)
3219                                         goto out;
3220
3221                                 stop_loop = 1;
3222                                 break;
3223                         }
3224                         btrfs_item_key_to_cpu(l, &key, slot);
3225
3226                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3227                             key.type != BTRFS_METADATA_ITEM_KEY)
3228                                 goto next;
3229
3230                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3231                                 bytes = fs_info->nodesize;
3232                         else
3233                                 bytes = key.offset;
3234
3235                         if (key.objectid + bytes <= logic_start)
3236                                 goto next;
3237
3238                         if (key.objectid >= logic_end) {
3239                                 stop_loop = 1;
3240                                 break;
3241                         }
3242
3243                         while (key.objectid >= logic_start + map->stripe_len)
3244                                 logic_start += map->stripe_len;
3245
3246                         extent = btrfs_item_ptr(l, slot,
3247                                                 struct btrfs_extent_item);
3248                         flags = btrfs_extent_flags(l, extent);
3249                         generation = btrfs_extent_generation(l, extent);
3250
3251                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3252                             (key.objectid < logic_start ||
3253                              key.objectid + bytes >
3254                              logic_start + map->stripe_len)) {
3255                                 btrfs_err(fs_info,
3256                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3257                                           key.objectid, logic_start);
3258                                 spin_lock(&sctx->stat_lock);
3259                                 sctx->stat.uncorrectable_errors++;
3260                                 spin_unlock(&sctx->stat_lock);
3261                                 goto next;
3262                         }
3263 again:
3264                         extent_logical = key.objectid;
3265                         extent_len = bytes;
3266
3267                         if (extent_logical < logic_start) {
3268                                 extent_len -= logic_start - extent_logical;
3269                                 extent_logical = logic_start;
3270                         }
3271
3272                         if (extent_logical + extent_len >
3273                             logic_start + map->stripe_len)
3274                                 extent_len = logic_start + map->stripe_len -
3275                                              extent_logical;
3276
3277                         scrub_parity_mark_sectors_data(sparity, extent_logical,
3278                                                        extent_len);
3279
3280                         mapped_length = extent_len;
3281                         bbio = NULL;
3282                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3283                                         extent_logical, &mapped_length, &bbio,
3284                                         0);
3285                         if (!ret) {
3286                                 if (!bbio || mapped_length < extent_len)
3287                                         ret = -EIO;
3288                         }
3289                         if (ret) {
3290                                 btrfs_put_bbio(bbio);
3291                                 goto out;
3292                         }
3293                         extent_physical = bbio->stripes[0].physical;
3294                         extent_mirror_num = bbio->mirror_num;
3295                         extent_dev = bbio->stripes[0].dev;
3296                         btrfs_put_bbio(bbio);
3297
3298                         ret = btrfs_lookup_csums_range(csum_root,
3299                                                 extent_logical,
3300                                                 extent_logical + extent_len - 1,
3301                                                 &sctx->csum_list, 1);
3302                         if (ret)
3303                                 goto out;
3304
3305                         ret = scrub_extent_for_parity(sparity, extent_logical,
3306                                                       extent_len,
3307                                                       extent_physical,
3308                                                       extent_dev, flags,
3309                                                       generation,
3310                                                       extent_mirror_num);
3311
3312                         scrub_free_csums(sctx);
3313
3314                         if (ret)
3315                                 goto out;
3316
3317                         if (extent_logical + extent_len <
3318                             key.objectid + bytes) {
3319                                 logic_start += map->stripe_len;
3320
3321                                 if (logic_start >= logic_end) {
3322                                         stop_loop = 1;
3323                                         break;
3324                                 }
3325
3326                                 if (logic_start < key.objectid + bytes) {
3327                                         cond_resched();
3328                                         goto again;
3329                                 }
3330                         }
3331 next:
3332                         path->slots[0]++;
3333                 }
3334
3335                 btrfs_release_path(path);
3336
3337                 if (stop_loop)
3338                         break;
3339
3340                 logic_start += map->stripe_len;
3341         }
3342 out:
3343         if (ret < 0)
3344                 scrub_parity_mark_sectors_error(sparity, logic_start,
3345                                                 logic_end - logic_start);
3346         scrub_parity_put(sparity);
3347         scrub_submit(sctx);
3348         mutex_lock(&sctx->wr_lock);
3349         scrub_wr_submit(sctx);
3350         mutex_unlock(&sctx->wr_lock);
3351
3352         btrfs_release_path(path);
3353         return ret < 0 ? ret : 0;
3354 }
3355
3356 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3357                                            struct map_lookup *map,
3358                                            struct btrfs_device *scrub_dev,
3359                                            int num, u64 base, u64 length,
3360                                            int is_dev_replace)
3361 {
3362         struct btrfs_path *path, *ppath;
3363         struct btrfs_fs_info *fs_info = sctx->fs_info;
3364         struct btrfs_root *root = fs_info->extent_root;
3365         struct btrfs_root *csum_root = fs_info->csum_root;
3366         struct btrfs_extent_item *extent;
3367         struct blk_plug plug;
3368         u64 flags;
3369         int ret;
3370         int slot;
3371         u64 nstripes;
3372         struct extent_buffer *l;
3373         u64 physical;
3374         u64 logical;
3375         u64 logic_end;
3376         u64 physical_end;
3377         u64 generation;
3378         int mirror_num;
3379         struct reada_control *reada1;
3380         struct reada_control *reada2;
3381         struct btrfs_key key;
3382         struct btrfs_key key_end;
3383         u64 increment = map->stripe_len;
3384         u64 offset;
3385         u64 extent_logical;
3386         u64 extent_physical;
3387         u64 extent_len;
3388         u64 stripe_logical;
3389         u64 stripe_end;
3390         struct btrfs_device *extent_dev;
3391         int extent_mirror_num;
3392         int stop_loop = 0;
3393
3394         physical = map->stripes[num].physical;
3395         offset = 0;
3396         nstripes = div64_u64(length, map->stripe_len);
3397         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3398                 offset = map->stripe_len * num;
3399                 increment = map->stripe_len * map->num_stripes;
3400                 mirror_num = 1;
3401         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3402                 int factor = map->num_stripes / map->sub_stripes;
3403                 offset = map->stripe_len * (num / map->sub_stripes);
3404                 increment = map->stripe_len * factor;
3405                 mirror_num = num % map->sub_stripes + 1;
3406         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3407                 increment = map->stripe_len;
3408                 mirror_num = num % map->num_stripes + 1;
3409         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3410                 increment = map->stripe_len;
3411                 mirror_num = num % map->num_stripes + 1;
3412         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3413                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3414                 increment = map->stripe_len * nr_data_stripes(map);
3415                 mirror_num = 1;
3416         } else {
3417                 increment = map->stripe_len;
3418                 mirror_num = 1;
3419         }
3420
3421         path = btrfs_alloc_path();
3422         if (!path)
3423                 return -ENOMEM;
3424
3425         ppath = btrfs_alloc_path();
3426         if (!ppath) {
3427                 btrfs_free_path(path);
3428                 return -ENOMEM;
3429         }
3430
3431         /*
3432          * work on commit root. The related disk blocks are static as
3433          * long as COW is applied. This means, it is save to rewrite
3434          * them to repair disk errors without any race conditions
3435          */
3436         path->search_commit_root = 1;
3437         path->skip_locking = 1;
3438
3439         ppath->search_commit_root = 1;
3440         ppath->skip_locking = 1;
3441         /*
3442          * trigger the readahead for extent tree csum tree and wait for
3443          * completion. During readahead, the scrub is officially paused
3444          * to not hold off transaction commits
3445          */
3446         logical = base + offset;
3447         physical_end = physical + nstripes * map->stripe_len;
3448         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3449                 get_raid56_logic_offset(physical_end, num,
3450                                         map, &logic_end, NULL);
3451                 logic_end += base;
3452         } else {
3453                 logic_end = logical + increment * nstripes;
3454         }
3455         wait_event(sctx->list_wait,
3456                    atomic_read(&sctx->bios_in_flight) == 0);
3457         scrub_blocked_if_needed(fs_info);
3458
3459         /* FIXME it might be better to start readahead at commit root */
3460         key.objectid = logical;
3461         key.type = BTRFS_EXTENT_ITEM_KEY;
3462         key.offset = (u64)0;
3463         key_end.objectid = logic_end;
3464         key_end.type = BTRFS_METADATA_ITEM_KEY;
3465         key_end.offset = (u64)-1;
3466         reada1 = btrfs_reada_add(root, &key, &key_end);
3467
3468         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3469         key.type = BTRFS_EXTENT_CSUM_KEY;
3470         key.offset = logical;
3471         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3472         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3473         key_end.offset = logic_end;
3474         reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3475
3476         if (!IS_ERR(reada1))
3477                 btrfs_reada_wait(reada1);
3478         if (!IS_ERR(reada2))
3479                 btrfs_reada_wait(reada2);
3480
3481
3482         /*
3483          * collect all data csums for the stripe to avoid seeking during
3484          * the scrub. This might currently (crc32) end up to be about 1MB
3485          */
3486         blk_start_plug(&plug);
3487
3488         /*
3489          * now find all extents for each stripe and scrub them
3490          */
3491         ret = 0;
3492         while (physical < physical_end) {
3493                 /*
3494                  * canceled?
3495                  */
3496                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3497                     atomic_read(&sctx->cancel_req)) {
3498                         ret = -ECANCELED;
3499                         goto out;
3500                 }
3501                 /*
3502                  * check to see if we have to pause
3503                  */
3504                 if (atomic_read(&fs_info->scrub_pause_req)) {
3505                         /* push queued extents */
3506                         sctx->flush_all_writes = true;
3507                         scrub_submit(sctx);
3508                         mutex_lock(&sctx->wr_lock);
3509                         scrub_wr_submit(sctx);
3510                         mutex_unlock(&sctx->wr_lock);
3511                         wait_event(sctx->list_wait,
3512                                    atomic_read(&sctx->bios_in_flight) == 0);
3513                         sctx->flush_all_writes = false;
3514                         scrub_blocked_if_needed(fs_info);
3515                 }
3516
3517                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3518                         ret = get_raid56_logic_offset(physical, num, map,
3519                                                       &logical,
3520                                                       &stripe_logical);
3521                         logical += base;
3522                         if (ret) {
3523                                 /* it is parity strip */
3524                                 stripe_logical += base;
3525                                 stripe_end = stripe_logical + increment;
3526                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3527                                                           ppath, stripe_logical,
3528                                                           stripe_end);
3529                                 if (ret)
3530                                         goto out;
3531                                 goto skip;
3532                         }
3533                 }
3534
3535                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3536                         key.type = BTRFS_METADATA_ITEM_KEY;
3537                 else
3538                         key.type = BTRFS_EXTENT_ITEM_KEY;
3539                 key.objectid = logical;
3540                 key.offset = (u64)-1;
3541
3542                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3543                 if (ret < 0)
3544                         goto out;
3545
3546                 if (ret > 0) {
3547                         ret = btrfs_previous_extent_item(root, path, 0);
3548                         if (ret < 0)
3549                                 goto out;
3550                         if (ret > 0) {
3551                                 /* there's no smaller item, so stick with the
3552                                  * larger one */
3553                                 btrfs_release_path(path);
3554                                 ret = btrfs_search_slot(NULL, root, &key,
3555                                                         path, 0, 0);
3556                                 if (ret < 0)
3557                                         goto out;
3558                         }
3559                 }
3560
3561                 stop_loop = 0;
3562                 while (1) {
3563                         u64 bytes;
3564
3565                         l = path->nodes[0];
3566                         slot = path->slots[0];
3567                         if (slot >= btrfs_header_nritems(l)) {
3568                                 ret = btrfs_next_leaf(root, path);
3569                                 if (ret == 0)
3570                                         continue;
3571                                 if (ret < 0)
3572                                         goto out;
3573
3574                                 stop_loop = 1;
3575                                 break;
3576                         }
3577                         btrfs_item_key_to_cpu(l, &key, slot);
3578
3579                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3580                             key.type != BTRFS_METADATA_ITEM_KEY)
3581                                 goto next;
3582
3583                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3584                                 bytes = fs_info->nodesize;
3585                         else
3586                                 bytes = key.offset;
3587
3588                         if (key.objectid + bytes <= logical)
3589                                 goto next;
3590
3591                         if (key.objectid >= logical + map->stripe_len) {
3592                                 /* out of this device extent */
3593                                 if (key.objectid >= logic_end)
3594                                         stop_loop = 1;
3595                                 break;
3596                         }
3597
3598                         extent = btrfs_item_ptr(l, slot,
3599                                                 struct btrfs_extent_item);
3600                         flags = btrfs_extent_flags(l, extent);
3601                         generation = btrfs_extent_generation(l, extent);
3602
3603                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3604                             (key.objectid < logical ||
3605                              key.objectid + bytes >
3606                              logical + map->stripe_len)) {
3607                                 btrfs_err(fs_info,
3608                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3609                                        key.objectid, logical);
3610                                 spin_lock(&sctx->stat_lock);
3611                                 sctx->stat.uncorrectable_errors++;
3612                                 spin_unlock(&sctx->stat_lock);
3613                                 goto next;
3614                         }
3615
3616 again:
3617                         extent_logical = key.objectid;
3618                         extent_len = bytes;
3619
3620                         /*
3621                          * trim extent to this stripe
3622                          */
3623                         if (extent_logical < logical) {
3624                                 extent_len -= logical - extent_logical;
3625                                 extent_logical = logical;
3626                         }
3627                         if (extent_logical + extent_len >
3628                             logical + map->stripe_len) {
3629