Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 */
5
6#include <linux/blkdev.h>
7#include <linux/ratelimit.h>
8#include <linux/sched/mm.h>
9#include <crypto/hash.h>
10#include "ctree.h"
11#include "discard.h"
12#include "volumes.h"
13#include "disk-io.h"
14#include "ordered-data.h"
15#include "transaction.h"
16#include "backref.h"
17#include "extent_io.h"
18#include "dev-replace.h"
19#include "check-integrity.h"
20#include "raid56.h"
21#include "block-group.h"
22#include "zoned.h"
23#include "fs.h"
24#include "accessors.h"
25#include "file-item.h"
26#include "scrub.h"
27
28/*
29 * This is only the first step towards a full-features scrub. It reads all
30 * extent and super block and verifies the checksums. In case a bad checksum
31 * is found or the extent cannot be read, good data will be written back if
32 * any can be found.
33 *
34 * Future enhancements:
35 * - In case an unrepairable extent is encountered, track which files are
36 * affected and report them
37 * - track and record media errors, throw out bad devices
38 * - add a mode to also read unallocated space
39 */
40
41struct scrub_block;
42struct scrub_ctx;
43
44/*
45 * The following three values only influence the performance.
46 *
47 * The last one configures the number of parallel and outstanding I/O
48 * operations. The first one configures an upper limit for the number
49 * of (dynamically allocated) pages that are added to a bio.
50 */
51#define SCRUB_SECTORS_PER_BIO 32 /* 128KiB per bio for 4KiB pages */
52#define SCRUB_BIOS_PER_SCTX 64 /* 8MiB per device in flight for 4KiB pages */
53
54/*
55 * The following value times PAGE_SIZE needs to be large enough to match the
56 * largest node/leaf/sector size that shall be supported.
57 */
58#define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
59
60#define SCRUB_MAX_PAGES (DIV_ROUND_UP(BTRFS_MAX_METADATA_BLOCKSIZE, PAGE_SIZE))
61
62/*
63 * Maximum number of mirrors that can be available for all profiles counting
64 * the target device of dev-replace as one. During an active device replace
65 * procedure, the target device of the copy operation is a mirror for the
66 * filesystem data as well that can be used to read data in order to repair
67 * read errors on other disks.
68 *
69 * Current value is derived from RAID1C4 with 4 copies.
70 */
71#define BTRFS_MAX_MIRRORS (4 + 1)
72
73struct scrub_recover {
74 refcount_t refs;
75 struct btrfs_io_context *bioc;
76 u64 map_length;
77};
78
79struct scrub_sector {
80 struct scrub_block *sblock;
81 struct list_head list;
82 u64 flags; /* extent flags */
83 u64 generation;
84 /* Offset in bytes to @sblock. */
85 u32 offset;
86 atomic_t refs;
87 unsigned int have_csum:1;
88 unsigned int io_error:1;
89 u8 csum[BTRFS_CSUM_SIZE];
90
91 struct scrub_recover *recover;
92};
93
94struct 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 struct scrub_sector *sectors[SCRUB_SECTORS_PER_BIO];
103 int sector_count;
104 int next_free;
105 struct work_struct work;
106};
107
108struct scrub_block {
109 /*
110 * Each page will have its page::private used to record the logical
111 * bytenr.
112 */
113 struct page *pages[SCRUB_MAX_PAGES];
114 struct scrub_sector *sectors[SCRUB_MAX_SECTORS_PER_BLOCK];
115 struct btrfs_device *dev;
116 /* Logical bytenr of the sblock */
117 u64 logical;
118 u64 physical;
119 u64 physical_for_dev_replace;
120 /* Length of sblock in bytes */
121 u32 len;
122 int sector_count;
123 int mirror_num;
124
125 atomic_t outstanding_sectors;
126 refcount_t refs; /* free mem on transition to zero */
127 struct scrub_ctx *sctx;
128 struct scrub_parity *sparity;
129 struct {
130 unsigned int header_error:1;
131 unsigned int checksum_error:1;
132 unsigned int no_io_error_seen:1;
133 unsigned int generation_error:1; /* also sets header_error */
134
135 /* The following is for the data used to check parity */
136 /* It is for the data with checksum */
137 unsigned int data_corrected:1;
138 };
139 struct work_struct work;
140};
141
142/* Used for the chunks with parity stripe such RAID5/6 */
143struct scrub_parity {
144 struct scrub_ctx *sctx;
145
146 struct btrfs_device *scrub_dev;
147
148 u64 logic_start;
149
150 u64 logic_end;
151
152 int nsectors;
153
154 u32 stripe_len;
155
156 refcount_t refs;
157
158 struct list_head sectors_list;
159
160 /* Work of parity check and repair */
161 struct work_struct work;
162
163 /* Mark the parity blocks which have data */
164 unsigned long dbitmap;
165
166 /*
167 * Mark the parity blocks which have data, but errors happen when
168 * read data or check data
169 */
170 unsigned long ebitmap;
171};
172
173struct scrub_ctx {
174 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
175 struct btrfs_fs_info *fs_info;
176 int first_free;
177 int curr;
178 atomic_t bios_in_flight;
179 atomic_t workers_pending;
180 spinlock_t list_lock;
181 wait_queue_head_t list_wait;
182 struct list_head csum_list;
183 atomic_t cancel_req;
184 int readonly;
185 int sectors_per_bio;
186
187 /* State of IO submission throttling affecting the associated device */
188 ktime_t throttle_deadline;
189 u64 throttle_sent;
190
191 int is_dev_replace;
192 u64 write_pointer;
193
194 struct scrub_bio *wr_curr_bio;
195 struct mutex wr_lock;
196 struct btrfs_device *wr_tgtdev;
197 bool flush_all_writes;
198
199 /*
200 * statistics
201 */
202 struct btrfs_scrub_progress stat;
203 spinlock_t stat_lock;
204
205 /*
206 * Use a ref counter to avoid use-after-free issues. Scrub workers
207 * decrement bios_in_flight and workers_pending and then do a wakeup
208 * on the list_wait wait queue. We must ensure the main scrub task
209 * doesn't free the scrub context before or while the workers are
210 * doing the wakeup() call.
211 */
212 refcount_t refs;
213};
214
215struct scrub_warning {
216 struct btrfs_path *path;
217 u64 extent_item_size;
218 const char *errstr;
219 u64 physical;
220 u64 logical;
221 struct btrfs_device *dev;
222};
223
224struct full_stripe_lock {
225 struct rb_node node;
226 u64 logical;
227 u64 refs;
228 struct mutex mutex;
229};
230
231#ifndef CONFIG_64BIT
232/* This structure is for archtectures whose (void *) is smaller than u64 */
233struct scrub_page_private {
234 u64 logical;
235};
236#endif
237
238static int attach_scrub_page_private(struct page *page, u64 logical)
239{
240#ifdef CONFIG_64BIT
241 attach_page_private(page, (void *)logical);
242 return 0;
243#else
244 struct scrub_page_private *spp;
245
246 spp = kmalloc(sizeof(*spp), GFP_KERNEL);
247 if (!spp)
248 return -ENOMEM;
249 spp->logical = logical;
250 attach_page_private(page, (void *)spp);
251 return 0;
252#endif
253}
254
255static void detach_scrub_page_private(struct page *page)
256{
257#ifdef CONFIG_64BIT
258 detach_page_private(page);
259 return;
260#else
261 struct scrub_page_private *spp;
262
263 spp = detach_page_private(page);
264 kfree(spp);
265 return;
266#endif
267}
268
269static struct scrub_block *alloc_scrub_block(struct scrub_ctx *sctx,
270 struct btrfs_device *dev,
271 u64 logical, u64 physical,
272 u64 physical_for_dev_replace,
273 int mirror_num)
274{
275 struct scrub_block *sblock;
276
277 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
278 if (!sblock)
279 return NULL;
280 refcount_set(&sblock->refs, 1);
281 sblock->sctx = sctx;
282 sblock->logical = logical;
283 sblock->physical = physical;
284 sblock->physical_for_dev_replace = physical_for_dev_replace;
285 sblock->dev = dev;
286 sblock->mirror_num = mirror_num;
287 sblock->no_io_error_seen = 1;
288 /*
289 * Scrub_block::pages will be allocated at alloc_scrub_sector() when
290 * the corresponding page is not allocated.
291 */
292 return sblock;
293}
294
295/*
296 * Allocate a new scrub sector and attach it to @sblock.
297 *
298 * Will also allocate new pages for @sblock if needed.
299 */
300static struct scrub_sector *alloc_scrub_sector(struct scrub_block *sblock,
301 u64 logical)
302{
303 const pgoff_t page_index = (logical - sblock->logical) >> PAGE_SHIFT;
304 struct scrub_sector *ssector;
305
306 /* We must never have scrub_block exceed U32_MAX in size. */
307 ASSERT(logical - sblock->logical < U32_MAX);
308
309 ssector = kzalloc(sizeof(*ssector), GFP_KERNEL);
310 if (!ssector)
311 return NULL;
312
313 /* Allocate a new page if the slot is not allocated */
314 if (!sblock->pages[page_index]) {
315 int ret;
316
317 sblock->pages[page_index] = alloc_page(GFP_KERNEL);
318 if (!sblock->pages[page_index]) {
319 kfree(ssector);
320 return NULL;
321 }
322 ret = attach_scrub_page_private(sblock->pages[page_index],
323 sblock->logical + (page_index << PAGE_SHIFT));
324 if (ret < 0) {
325 kfree(ssector);
326 __free_page(sblock->pages[page_index]);
327 sblock->pages[page_index] = NULL;
328 return NULL;
329 }
330 }
331
332 atomic_set(&ssector->refs, 1);
333 ssector->sblock = sblock;
334 /* The sector to be added should not be used */
335 ASSERT(sblock->sectors[sblock->sector_count] == NULL);
336 ssector->offset = logical - sblock->logical;
337
338 /* The sector count must be smaller than the limit */
339 ASSERT(sblock->sector_count < SCRUB_MAX_SECTORS_PER_BLOCK);
340
341 sblock->sectors[sblock->sector_count] = ssector;
342 sblock->sector_count++;
343 sblock->len += sblock->sctx->fs_info->sectorsize;
344
345 return ssector;
346}
347
348static struct page *scrub_sector_get_page(struct scrub_sector *ssector)
349{
350 struct scrub_block *sblock = ssector->sblock;
351 pgoff_t index;
352 /*
353 * When calling this function, ssector must be alreaday attached to the
354 * parent sblock.
355 */
356 ASSERT(sblock);
357
358 /* The range should be inside the sblock range */
359 ASSERT(ssector->offset < sblock->len);
360
361 index = ssector->offset >> PAGE_SHIFT;
362 ASSERT(index < SCRUB_MAX_PAGES);
363 ASSERT(sblock->pages[index]);
364 ASSERT(PagePrivate(sblock->pages[index]));
365 return sblock->pages[index];
366}
367
368static unsigned int scrub_sector_get_page_offset(struct scrub_sector *ssector)
369{
370 struct scrub_block *sblock = ssector->sblock;
371
372 /*
373 * When calling this function, ssector must be already attached to the
374 * parent sblock.
375 */
376 ASSERT(sblock);
377
378 /* The range should be inside the sblock range */
379 ASSERT(ssector->offset < sblock->len);
380
381 return offset_in_page(ssector->offset);
382}
383
384static char *scrub_sector_get_kaddr(struct scrub_sector *ssector)
385{
386 return page_address(scrub_sector_get_page(ssector)) +
387 scrub_sector_get_page_offset(ssector);
388}
389
390static int bio_add_scrub_sector(struct bio *bio, struct scrub_sector *ssector,
391 unsigned int len)
392{
393 return bio_add_page(bio, scrub_sector_get_page(ssector), len,
394 scrub_sector_get_page_offset(ssector));
395}
396
397static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
398 struct scrub_block *sblocks_for_recheck[]);
399static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
400 struct scrub_block *sblock,
401 int retry_failed_mirror);
402static void scrub_recheck_block_checksum(struct scrub_block *sblock);
403static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
404 struct scrub_block *sblock_good);
405static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
406 struct scrub_block *sblock_good,
407 int sector_num, int force_write);
408static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
409static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock,
410 int sector_num);
411static int scrub_checksum_data(struct scrub_block *sblock);
412static int scrub_checksum_tree_block(struct scrub_block *sblock);
413static int scrub_checksum_super(struct scrub_block *sblock);
414static void scrub_block_put(struct scrub_block *sblock);
415static void scrub_sector_get(struct scrub_sector *sector);
416static void scrub_sector_put(struct scrub_sector *sector);
417static void scrub_parity_get(struct scrub_parity *sparity);
418static void scrub_parity_put(struct scrub_parity *sparity);
419static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
420 u64 physical, struct btrfs_device *dev, u64 flags,
421 u64 gen, int mirror_num, u8 *csum,
422 u64 physical_for_dev_replace);
423static void scrub_bio_end_io(struct bio *bio);
424static void scrub_bio_end_io_worker(struct work_struct *work);
425static void scrub_block_complete(struct scrub_block *sblock);
426static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
427 u64 extent_logical, u32 extent_len,
428 u64 *extent_physical,
429 struct btrfs_device **extent_dev,
430 int *extent_mirror_num);
431static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
432 struct scrub_sector *sector);
433static void scrub_wr_submit(struct scrub_ctx *sctx);
434static void scrub_wr_bio_end_io(struct bio *bio);
435static void scrub_wr_bio_end_io_worker(struct work_struct *work);
436static void scrub_put_ctx(struct scrub_ctx *sctx);
437
438static inline int scrub_is_page_on_raid56(struct scrub_sector *sector)
439{
440 return sector->recover &&
441 (sector->recover->bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
442}
443
444static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
445{
446 refcount_inc(&sctx->refs);
447 atomic_inc(&sctx->bios_in_flight);
448}
449
450static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
451{
452 atomic_dec(&sctx->bios_in_flight);
453 wake_up(&sctx->list_wait);
454 scrub_put_ctx(sctx);
455}
456
457static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
458{
459 while (atomic_read(&fs_info->scrub_pause_req)) {
460 mutex_unlock(&fs_info->scrub_lock);
461 wait_event(fs_info->scrub_pause_wait,
462 atomic_read(&fs_info->scrub_pause_req) == 0);
463 mutex_lock(&fs_info->scrub_lock);
464 }
465}
466
467static void scrub_pause_on(struct btrfs_fs_info *fs_info)
468{
469 atomic_inc(&fs_info->scrubs_paused);
470 wake_up(&fs_info->scrub_pause_wait);
471}
472
473static void scrub_pause_off(struct btrfs_fs_info *fs_info)
474{
475 mutex_lock(&fs_info->scrub_lock);
476 __scrub_blocked_if_needed(fs_info);
477 atomic_dec(&fs_info->scrubs_paused);
478 mutex_unlock(&fs_info->scrub_lock);
479
480 wake_up(&fs_info->scrub_pause_wait);
481}
482
483static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
484{
485 scrub_pause_on(fs_info);
486 scrub_pause_off(fs_info);
487}
488
489/*
490 * Insert new full stripe lock into full stripe locks tree
491 *
492 * Return pointer to existing or newly inserted full_stripe_lock structure if
493 * everything works well.
494 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
495 *
496 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
497 * function
498 */
499static struct full_stripe_lock *insert_full_stripe_lock(
500 struct btrfs_full_stripe_locks_tree *locks_root,
501 u64 fstripe_logical)
502{
503 struct rb_node **p;
504 struct rb_node *parent = NULL;
505 struct full_stripe_lock *entry;
506 struct full_stripe_lock *ret;
507
508 lockdep_assert_held(&locks_root->lock);
509
510 p = &locks_root->root.rb_node;
511 while (*p) {
512 parent = *p;
513 entry = rb_entry(parent, struct full_stripe_lock, node);
514 if (fstripe_logical < entry->logical) {
515 p = &(*p)->rb_left;
516 } else if (fstripe_logical > entry->logical) {
517 p = &(*p)->rb_right;
518 } else {
519 entry->refs++;
520 return entry;
521 }
522 }
523
524 /*
525 * Insert new lock.
526 */
527 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
528 if (!ret)
529 return ERR_PTR(-ENOMEM);
530 ret->logical = fstripe_logical;
531 ret->refs = 1;
532 mutex_init(&ret->mutex);
533
534 rb_link_node(&ret->node, parent, p);
535 rb_insert_color(&ret->node, &locks_root->root);
536 return ret;
537}
538
539/*
540 * Search for a full stripe lock of a block group
541 *
542 * Return pointer to existing full stripe lock if found
543 * Return NULL if not found
544 */
545static struct full_stripe_lock *search_full_stripe_lock(
546 struct btrfs_full_stripe_locks_tree *locks_root,
547 u64 fstripe_logical)
548{
549 struct rb_node *node;
550 struct full_stripe_lock *entry;
551
552 lockdep_assert_held(&locks_root->lock);
553
554 node = locks_root->root.rb_node;
555 while (node) {
556 entry = rb_entry(node, struct full_stripe_lock, node);
557 if (fstripe_logical < entry->logical)
558 node = node->rb_left;
559 else if (fstripe_logical > entry->logical)
560 node = node->rb_right;
561 else
562 return entry;
563 }
564 return NULL;
565}
566
567/*
568 * Helper to get full stripe logical from a normal bytenr.
569 *
570 * Caller must ensure @cache is a RAID56 block group.
571 */
572static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
573{
574 u64 ret;
575
576 /*
577 * Due to chunk item size limit, full stripe length should not be
578 * larger than U32_MAX. Just a sanity check here.
579 */
580 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
581
582 /*
583 * round_down() can only handle power of 2, while RAID56 full
584 * stripe length can be 64KiB * n, so we need to manually round down.
585 */
586 ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
587 cache->full_stripe_len + cache->start;
588 return ret;
589}
590
591/*
592 * Lock a full stripe to avoid concurrency of recovery and read
593 *
594 * It's only used for profiles with parities (RAID5/6), for other profiles it
595 * does nothing.
596 *
597 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
598 * So caller must call unlock_full_stripe() at the same context.
599 *
600 * Return <0 if encounters error.
601 */
602static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
603 bool *locked_ret)
604{
605 struct btrfs_block_group *bg_cache;
606 struct btrfs_full_stripe_locks_tree *locks_root;
607 struct full_stripe_lock *existing;
608 u64 fstripe_start;
609 int ret = 0;
610
611 *locked_ret = false;
612 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
613 if (!bg_cache) {
614 ASSERT(0);
615 return -ENOENT;
616 }
617
618 /* Profiles not based on parity don't need full stripe lock */
619 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
620 goto out;
621 locks_root = &bg_cache->full_stripe_locks_root;
622
623 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
624
625 /* Now insert the full stripe lock */
626 mutex_lock(&locks_root->lock);
627 existing = insert_full_stripe_lock(locks_root, fstripe_start);
628 mutex_unlock(&locks_root->lock);
629 if (IS_ERR(existing)) {
630 ret = PTR_ERR(existing);
631 goto out;
632 }
633 mutex_lock(&existing->mutex);
634 *locked_ret = true;
635out:
636 btrfs_put_block_group(bg_cache);
637 return ret;
638}
639
640/*
641 * Unlock a full stripe.
642 *
643 * NOTE: Caller must ensure it's the same context calling corresponding
644 * lock_full_stripe().
645 *
646 * Return 0 if we unlock full stripe without problem.
647 * Return <0 for error
648 */
649static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
650 bool locked)
651{
652 struct btrfs_block_group *bg_cache;
653 struct btrfs_full_stripe_locks_tree *locks_root;
654 struct full_stripe_lock *fstripe_lock;
655 u64 fstripe_start;
656 bool freeit = false;
657 int ret = 0;
658
659 /* If we didn't acquire full stripe lock, no need to continue */
660 if (!locked)
661 return 0;
662
663 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
664 if (!bg_cache) {
665 ASSERT(0);
666 return -ENOENT;
667 }
668 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
669 goto out;
670
671 locks_root = &bg_cache->full_stripe_locks_root;
672 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
673
674 mutex_lock(&locks_root->lock);
675 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
676 /* Unpaired unlock_full_stripe() detected */
677 if (!fstripe_lock) {
678 WARN_ON(1);
679 ret = -ENOENT;
680 mutex_unlock(&locks_root->lock);
681 goto out;
682 }
683
684 if (fstripe_lock->refs == 0) {
685 WARN_ON(1);
686 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
687 fstripe_lock->logical);
688 } else {
689 fstripe_lock->refs--;
690 }
691
692 if (fstripe_lock->refs == 0) {
693 rb_erase(&fstripe_lock->node, &locks_root->root);
694 freeit = true;
695 }
696 mutex_unlock(&locks_root->lock);
697
698 mutex_unlock(&fstripe_lock->mutex);
699 if (freeit)
700 kfree(fstripe_lock);
701out:
702 btrfs_put_block_group(bg_cache);
703 return ret;
704}
705
706static void scrub_free_csums(struct scrub_ctx *sctx)
707{
708 while (!list_empty(&sctx->csum_list)) {
709 struct btrfs_ordered_sum *sum;
710 sum = list_first_entry(&sctx->csum_list,
711 struct btrfs_ordered_sum, list);
712 list_del(&sum->list);
713 kfree(sum);
714 }
715}
716
717static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
718{
719 int i;
720
721 if (!sctx)
722 return;
723
724 /* this can happen when scrub is cancelled */
725 if (sctx->curr != -1) {
726 struct scrub_bio *sbio = sctx->bios[sctx->curr];
727
728 for (i = 0; i < sbio->sector_count; i++)
729 scrub_block_put(sbio->sectors[i]->sblock);
730 bio_put(sbio->bio);
731 }
732
733 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
734 struct scrub_bio *sbio = sctx->bios[i];
735
736 if (!sbio)
737 break;
738 kfree(sbio);
739 }
740
741 kfree(sctx->wr_curr_bio);
742 scrub_free_csums(sctx);
743 kfree(sctx);
744}
745
746static void scrub_put_ctx(struct scrub_ctx *sctx)
747{
748 if (refcount_dec_and_test(&sctx->refs))
749 scrub_free_ctx(sctx);
750}
751
752static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
753 struct btrfs_fs_info *fs_info, int is_dev_replace)
754{
755 struct scrub_ctx *sctx;
756 int i;
757
758 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
759 if (!sctx)
760 goto nomem;
761 refcount_set(&sctx->refs, 1);
762 sctx->is_dev_replace = is_dev_replace;
763 sctx->sectors_per_bio = SCRUB_SECTORS_PER_BIO;
764 sctx->curr = -1;
765 sctx->fs_info = fs_info;
766 INIT_LIST_HEAD(&sctx->csum_list);
767 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
768 struct scrub_bio *sbio;
769
770 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
771 if (!sbio)
772 goto nomem;
773 sctx->bios[i] = sbio;
774
775 sbio->index = i;
776 sbio->sctx = sctx;
777 sbio->sector_count = 0;
778 INIT_WORK(&sbio->work, scrub_bio_end_io_worker);
779
780 if (i != SCRUB_BIOS_PER_SCTX - 1)
781 sctx->bios[i]->next_free = i + 1;
782 else
783 sctx->bios[i]->next_free = -1;
784 }
785 sctx->first_free = 0;
786 atomic_set(&sctx->bios_in_flight, 0);
787 atomic_set(&sctx->workers_pending, 0);
788 atomic_set(&sctx->cancel_req, 0);
789
790 spin_lock_init(&sctx->list_lock);
791 spin_lock_init(&sctx->stat_lock);
792 init_waitqueue_head(&sctx->list_wait);
793 sctx->throttle_deadline = 0;
794
795 WARN_ON(sctx->wr_curr_bio != NULL);
796 mutex_init(&sctx->wr_lock);
797 sctx->wr_curr_bio = NULL;
798 if (is_dev_replace) {
799 WARN_ON(!fs_info->dev_replace.tgtdev);
800 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
801 sctx->flush_all_writes = false;
802 }
803
804 return sctx;
805
806nomem:
807 scrub_free_ctx(sctx);
808 return ERR_PTR(-ENOMEM);
809}
810
811static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
812 u64 root, void *warn_ctx)
813{
814 u32 nlink;
815 int ret;
816 int i;
817 unsigned nofs_flag;
818 struct extent_buffer *eb;
819 struct btrfs_inode_item *inode_item;
820 struct scrub_warning *swarn = warn_ctx;
821 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
822 struct inode_fs_paths *ipath = NULL;
823 struct btrfs_root *local_root;
824 struct btrfs_key key;
825
826 local_root = btrfs_get_fs_root(fs_info, root, true);
827 if (IS_ERR(local_root)) {
828 ret = PTR_ERR(local_root);
829 goto err;
830 }
831
832 /*
833 * this makes the path point to (inum INODE_ITEM ioff)
834 */
835 key.objectid = inum;
836 key.type = BTRFS_INODE_ITEM_KEY;
837 key.offset = 0;
838
839 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
840 if (ret) {
841 btrfs_put_root(local_root);
842 btrfs_release_path(swarn->path);
843 goto err;
844 }
845
846 eb = swarn->path->nodes[0];
847 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
848 struct btrfs_inode_item);
849 nlink = btrfs_inode_nlink(eb, inode_item);
850 btrfs_release_path(swarn->path);
851
852 /*
853 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
854 * uses GFP_NOFS in this context, so we keep it consistent but it does
855 * not seem to be strictly necessary.
856 */
857 nofs_flag = memalloc_nofs_save();
858 ipath = init_ipath(4096, local_root, swarn->path);
859 memalloc_nofs_restore(nofs_flag);
860 if (IS_ERR(ipath)) {
861 btrfs_put_root(local_root);
862 ret = PTR_ERR(ipath);
863 ipath = NULL;
864 goto err;
865 }
866 ret = paths_from_inode(inum, ipath);
867
868 if (ret < 0)
869 goto err;
870
871 /*
872 * we deliberately ignore the bit ipath might have been too small to
873 * hold all of the paths here
874 */
875 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
876 btrfs_warn_in_rcu(fs_info,
877"%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
878 swarn->errstr, swarn->logical,
879 btrfs_dev_name(swarn->dev),
880 swarn->physical,
881 root, inum, offset,
882 fs_info->sectorsize, nlink,
883 (char *)(unsigned long)ipath->fspath->val[i]);
884
885 btrfs_put_root(local_root);
886 free_ipath(ipath);
887 return 0;
888
889err:
890 btrfs_warn_in_rcu(fs_info,
891 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
892 swarn->errstr, swarn->logical,
893 btrfs_dev_name(swarn->dev),
894 swarn->physical,
895 root, inum, offset, ret);
896
897 free_ipath(ipath);
898 return 0;
899}
900
901static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
902{
903 struct btrfs_device *dev;
904 struct btrfs_fs_info *fs_info;
905 struct btrfs_path *path;
906 struct btrfs_key found_key;
907 struct extent_buffer *eb;
908 struct btrfs_extent_item *ei;
909 struct scrub_warning swarn;
910 unsigned long ptr = 0;
911 u64 flags = 0;
912 u64 ref_root;
913 u32 item_size;
914 u8 ref_level = 0;
915 int ret;
916
917 WARN_ON(sblock->sector_count < 1);
918 dev = sblock->dev;
919 fs_info = sblock->sctx->fs_info;
920
921 /* Super block error, no need to search extent tree. */
922 if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
923 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
924 errstr, btrfs_dev_name(dev), sblock->physical);
925 return;
926 }
927 path = btrfs_alloc_path();
928 if (!path)
929 return;
930
931 swarn.physical = sblock->physical;
932 swarn.logical = sblock->logical;
933 swarn.errstr = errstr;
934 swarn.dev = NULL;
935
936 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
937 &flags);
938 if (ret < 0)
939 goto out;
940
941 swarn.extent_item_size = found_key.offset;
942
943 eb = path->nodes[0];
944 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
945 item_size = btrfs_item_size(eb, path->slots[0]);
946
947 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
948 do {
949 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
950 item_size, &ref_root,
951 &ref_level);
952 btrfs_warn_in_rcu(fs_info,
953"%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
954 errstr, swarn.logical,
955 btrfs_dev_name(dev),
956 swarn.physical,
957 ref_level ? "node" : "leaf",
958 ret < 0 ? -1 : ref_level,
959 ret < 0 ? -1 : ref_root);
960 } while (ret != 1);
961 btrfs_release_path(path);
962 } else {
963 struct btrfs_backref_walk_ctx ctx = { 0 };
964
965 btrfs_release_path(path);
966
967 ctx.bytenr = found_key.objectid;
968 ctx.extent_item_pos = swarn.logical - found_key.objectid;
969 ctx.fs_info = fs_info;
970
971 swarn.path = path;
972 swarn.dev = dev;
973
974 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
975 }
976
977out:
978 btrfs_free_path(path);
979}
980
981static inline void scrub_get_recover(struct scrub_recover *recover)
982{
983 refcount_inc(&recover->refs);
984}
985
986static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
987 struct scrub_recover *recover)
988{
989 if (refcount_dec_and_test(&recover->refs)) {
990 btrfs_bio_counter_dec(fs_info);
991 btrfs_put_bioc(recover->bioc);
992 kfree(recover);
993 }
994}
995
996/*
997 * scrub_handle_errored_block gets called when either verification of the
998 * sectors failed or the bio failed to read, e.g. with EIO. In the latter
999 * case, this function handles all sectors in the bio, even though only one
1000 * may be bad.
1001 * The goal of this function is to repair the errored block by using the
1002 * contents of one of the mirrors.
1003 */
1004static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1005{
1006 struct scrub_ctx *sctx = sblock_to_check->sctx;
1007 struct btrfs_device *dev = sblock_to_check->dev;
1008 struct btrfs_fs_info *fs_info;
1009 u64 logical;
1010 unsigned int failed_mirror_index;
1011 unsigned int is_metadata;
1012 unsigned int have_csum;
1013 /* One scrub_block for each mirror */
1014 struct scrub_block *sblocks_for_recheck[BTRFS_MAX_MIRRORS] = { 0 };
1015 struct scrub_block *sblock_bad;
1016 int ret;
1017 int mirror_index;
1018 int sector_num;
1019 int success;
1020 bool full_stripe_locked;
1021 unsigned int nofs_flag;
1022 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
1023 DEFAULT_RATELIMIT_BURST);
1024
1025 BUG_ON(sblock_to_check->sector_count < 1);
1026 fs_info = sctx->fs_info;
1027 if (sblock_to_check->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1028 /*
1029 * If we find an error in a super block, we just report it.
1030 * They will get written with the next transaction commit
1031 * anyway
1032 */
1033 scrub_print_warning("super block error", sblock_to_check);
1034 spin_lock(&sctx->stat_lock);
1035 ++sctx->stat.super_errors;
1036 spin_unlock(&sctx->stat_lock);
1037 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_CORRUPTION_ERRS);
1038 return 0;
1039 }
1040 logical = sblock_to_check->logical;
1041 ASSERT(sblock_to_check->mirror_num);
1042 failed_mirror_index = sblock_to_check->mirror_num - 1;
1043 is_metadata = !(sblock_to_check->sectors[0]->flags &
1044 BTRFS_EXTENT_FLAG_DATA);
1045 have_csum = sblock_to_check->sectors[0]->have_csum;
1046
1047 if (!sctx->is_dev_replace && btrfs_repair_one_zone(fs_info, logical))
1048 return 0;
1049
1050 /*
1051 * We must use GFP_NOFS because the scrub task might be waiting for a
1052 * worker task executing this function and in turn a transaction commit
1053 * might be waiting the scrub task to pause (which needs to wait for all
1054 * the worker tasks to complete before pausing).
1055 * We do allocations in the workers through insert_full_stripe_lock()
1056 * and scrub_add_sector_to_wr_bio(), which happens down the call chain of
1057 * this function.
1058 */
1059 nofs_flag = memalloc_nofs_save();
1060 /*
1061 * For RAID5/6, race can happen for a different device scrub thread.
1062 * For data corruption, Parity and Data threads will both try
1063 * to recovery the data.
1064 * Race can lead to doubly added csum error, or even unrecoverable
1065 * error.
1066 */
1067 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1068 if (ret < 0) {
1069 memalloc_nofs_restore(nofs_flag);
1070 spin_lock(&sctx->stat_lock);
1071 if (ret == -ENOMEM)
1072 sctx->stat.malloc_errors++;
1073 sctx->stat.read_errors++;
1074 sctx->stat.uncorrectable_errors++;
1075 spin_unlock(&sctx->stat_lock);
1076 return ret;
1077 }
1078
1079 /*
1080 * read all mirrors one after the other. This includes to
1081 * re-read the extent or metadata block that failed (that was
1082 * the cause that this fixup code is called) another time,
1083 * sector by sector this time in order to know which sectors
1084 * caused I/O errors and which ones are good (for all mirrors).
1085 * It is the goal to handle the situation when more than one
1086 * mirror contains I/O errors, but the errors do not
1087 * overlap, i.e. the data can be repaired by selecting the
1088 * sectors from those mirrors without I/O error on the
1089 * particular sectors. One example (with blocks >= 2 * sectorsize)
1090 * would be that mirror #1 has an I/O error on the first sector,
1091 * the second sector is good, and mirror #2 has an I/O error on
1092 * the second sector, but the first sector is good.
1093 * Then the first sector of the first mirror can be repaired by
1094 * taking the first sector of the second mirror, and the
1095 * second sector of the second mirror can be repaired by
1096 * copying the contents of the 2nd sector of the 1st mirror.
1097 * One more note: if the sectors of one mirror contain I/O
1098 * errors, the checksum cannot be verified. In order to get
1099 * the best data for repairing, the first attempt is to find
1100 * a mirror without I/O errors and with a validated checksum.
1101 * Only if this is not possible, the sectors are picked from
1102 * mirrors with I/O errors without considering the checksum.
1103 * If the latter is the case, at the end, the checksum of the
1104 * repaired area is verified in order to correctly maintain
1105 * the statistics.
1106 */
1107 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1108 /*
1109 * Note: the two members refs and outstanding_sectors are not
1110 * used in the blocks that are used for the recheck procedure.
1111 *
1112 * But alloc_scrub_block() will initialize sblock::ref anyway,
1113 * so we can use scrub_block_put() to clean them up.
1114 *
1115 * And here we don't setup the physical/dev for the sblock yet,
1116 * they will be correctly initialized in scrub_setup_recheck_block().
1117 */
1118 sblocks_for_recheck[mirror_index] = alloc_scrub_block(sctx, NULL,
1119 logical, 0, 0, mirror_index);
1120 if (!sblocks_for_recheck[mirror_index]) {
1121 spin_lock(&sctx->stat_lock);
1122 sctx->stat.malloc_errors++;
1123 sctx->stat.read_errors++;
1124 sctx->stat.uncorrectable_errors++;
1125 spin_unlock(&sctx->stat_lock);
1126 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1127 goto out;
1128 }
1129 }
1130
1131 /* Setup the context, map the logical blocks and alloc the sectors */
1132 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1133 if (ret) {
1134 spin_lock(&sctx->stat_lock);
1135 sctx->stat.read_errors++;
1136 sctx->stat.uncorrectable_errors++;
1137 spin_unlock(&sctx->stat_lock);
1138 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1139 goto out;
1140 }
1141 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1142 sblock_bad = sblocks_for_recheck[failed_mirror_index];
1143
1144 /* build and submit the bios for the failed mirror, check checksums */
1145 scrub_recheck_block(fs_info, sblock_bad, 1);
1146
1147 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1148 sblock_bad->no_io_error_seen) {
1149 /*
1150 * The error disappeared after reading sector by sector, or
1151 * the area was part of a huge bio and other parts of the
1152 * bio caused I/O errors, or the block layer merged several
1153 * read requests into one and the error is caused by a
1154 * different bio (usually one of the two latter cases is
1155 * the cause)
1156 */
1157 spin_lock(&sctx->stat_lock);
1158 sctx->stat.unverified_errors++;
1159 sblock_to_check->data_corrected = 1;
1160 spin_unlock(&sctx->stat_lock);
1161
1162 if (sctx->is_dev_replace)
1163 scrub_write_block_to_dev_replace(sblock_bad);
1164 goto out;
1165 }
1166
1167 if (!sblock_bad->no_io_error_seen) {
1168 spin_lock(&sctx->stat_lock);
1169 sctx->stat.read_errors++;
1170 spin_unlock(&sctx->stat_lock);
1171 if (__ratelimit(&rs))
1172 scrub_print_warning("i/o error", sblock_to_check);
1173 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1174 } else if (sblock_bad->checksum_error) {
1175 spin_lock(&sctx->stat_lock);
1176 sctx->stat.csum_errors++;
1177 spin_unlock(&sctx->stat_lock);
1178 if (__ratelimit(&rs))
1179 scrub_print_warning("checksum error", sblock_to_check);
1180 btrfs_dev_stat_inc_and_print(dev,
1181 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1182 } else if (sblock_bad->header_error) {
1183 spin_lock(&sctx->stat_lock);
1184 sctx->stat.verify_errors++;
1185 spin_unlock(&sctx->stat_lock);
1186 if (__ratelimit(&rs))
1187 scrub_print_warning("checksum/header error",
1188 sblock_to_check);
1189 if (sblock_bad->generation_error)
1190 btrfs_dev_stat_inc_and_print(dev,
1191 BTRFS_DEV_STAT_GENERATION_ERRS);
1192 else
1193 btrfs_dev_stat_inc_and_print(dev,
1194 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1195 }
1196
1197 if (sctx->readonly) {
1198 ASSERT(!sctx->is_dev_replace);
1199 goto out;
1200 }
1201
1202 /*
1203 * now build and submit the bios for the other mirrors, check
1204 * checksums.
1205 * First try to pick the mirror which is completely without I/O
1206 * errors and also does not have a checksum error.
1207 * If one is found, and if a checksum is present, the full block
1208 * that is known to contain an error is rewritten. Afterwards
1209 * the block is known to be corrected.
1210 * If a mirror is found which is completely correct, and no
1211 * checksum is present, only those sectors are rewritten that had
1212 * an I/O error in the block to be repaired, since it cannot be
1213 * determined, which copy of the other sectors is better (and it
1214 * could happen otherwise that a correct sector would be
1215 * overwritten by a bad one).
1216 */
1217 for (mirror_index = 0; ;mirror_index++) {
1218 struct scrub_block *sblock_other;
1219
1220 if (mirror_index == failed_mirror_index)
1221 continue;
1222
1223 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1224 if (!scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1225 if (mirror_index >= BTRFS_MAX_MIRRORS)
1226 break;
1227 if (!sblocks_for_recheck[mirror_index]->sector_count)
1228 break;
1229
1230 sblock_other = sblocks_for_recheck[mirror_index];
1231 } else {
1232 struct scrub_recover *r = sblock_bad->sectors[0]->recover;
1233 int max_allowed = r->bioc->num_stripes - r->bioc->num_tgtdevs;
1234
1235 if (mirror_index >= max_allowed)
1236 break;
1237 if (!sblocks_for_recheck[1]->sector_count)
1238 break;
1239
1240 ASSERT(failed_mirror_index == 0);
1241 sblock_other = sblocks_for_recheck[1];
1242 sblock_other->mirror_num = 1 + mirror_index;
1243 }
1244
1245 /* build and submit the bios, check checksums */
1246 scrub_recheck_block(fs_info, sblock_other, 0);
1247
1248 if (!sblock_other->header_error &&
1249 !sblock_other->checksum_error &&
1250 sblock_other->no_io_error_seen) {
1251 if (sctx->is_dev_replace) {
1252 scrub_write_block_to_dev_replace(sblock_other);
1253 goto corrected_error;
1254 } else {
1255 ret = scrub_repair_block_from_good_copy(
1256 sblock_bad, sblock_other);
1257 if (!ret)
1258 goto corrected_error;
1259 }
1260 }
1261 }
1262
1263 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1264 goto did_not_correct_error;
1265
1266 /*
1267 * In case of I/O errors in the area that is supposed to be
1268 * repaired, continue by picking good copies of those sectors.
1269 * Select the good sectors from mirrors to rewrite bad sectors from
1270 * the area to fix. Afterwards verify the checksum of the block
1271 * that is supposed to be repaired. This verification step is
1272 * only done for the purpose of statistic counting and for the
1273 * final scrub report, whether errors remain.
1274 * A perfect algorithm could make use of the checksum and try
1275 * all possible combinations of sectors from the different mirrors
1276 * until the checksum verification succeeds. For example, when
1277 * the 2nd sector of mirror #1 faces I/O errors, and the 2nd sector
1278 * of mirror #2 is readable but the final checksum test fails,
1279 * then the 2nd sector of mirror #3 could be tried, whether now
1280 * the final checksum succeeds. But this would be a rare
1281 * exception and is therefore not implemented. At least it is
1282 * avoided that the good copy is overwritten.
1283 * A more useful improvement would be to pick the sectors
1284 * without I/O error based on sector sizes (512 bytes on legacy
1285 * disks) instead of on sectorsize. Then maybe 512 byte of one
1286 * mirror could be repaired by taking 512 byte of a different
1287 * mirror, even if other 512 byte sectors in the same sectorsize
1288 * area are unreadable.
1289 */
1290 success = 1;
1291 for (sector_num = 0; sector_num < sblock_bad->sector_count;
1292 sector_num++) {
1293 struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1294 struct scrub_block *sblock_other = NULL;
1295
1296 /* Skip no-io-error sectors in scrub */
1297 if (!sector_bad->io_error && !sctx->is_dev_replace)
1298 continue;
1299
1300 if (scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1301 /*
1302 * In case of dev replace, if raid56 rebuild process
1303 * didn't work out correct data, then copy the content
1304 * in sblock_bad to make sure target device is identical
1305 * to source device, instead of writing garbage data in
1306 * sblock_for_recheck array to target device.
1307 */
1308 sblock_other = NULL;
1309 } else if (sector_bad->io_error) {
1310 /* Try to find no-io-error sector in mirrors */
1311 for (mirror_index = 0;
1312 mirror_index < BTRFS_MAX_MIRRORS &&
1313 sblocks_for_recheck[mirror_index]->sector_count > 0;
1314 mirror_index++) {
1315 if (!sblocks_for_recheck[mirror_index]->
1316 sectors[sector_num]->io_error) {
1317 sblock_other = sblocks_for_recheck[mirror_index];
1318 break;
1319 }
1320 }
1321 if (!sblock_other)
1322 success = 0;
1323 }
1324
1325 if (sctx->is_dev_replace) {
1326 /*
1327 * Did not find a mirror to fetch the sector from.
1328 * scrub_write_sector_to_dev_replace() handles this
1329 * case (sector->io_error), by filling the block with
1330 * zeros before submitting the write request
1331 */
1332 if (!sblock_other)
1333 sblock_other = sblock_bad;
1334
1335 if (scrub_write_sector_to_dev_replace(sblock_other,
1336 sector_num) != 0) {
1337 atomic64_inc(
1338 &fs_info->dev_replace.num_write_errors);
1339 success = 0;
1340 }
1341 } else if (sblock_other) {
1342 ret = scrub_repair_sector_from_good_copy(sblock_bad,
1343 sblock_other,
1344 sector_num, 0);
1345 if (0 == ret)
1346 sector_bad->io_error = 0;
1347 else
1348 success = 0;
1349 }
1350 }
1351
1352 if (success && !sctx->is_dev_replace) {
1353 if (is_metadata || have_csum) {
1354 /*
1355 * need to verify the checksum now that all
1356 * sectors on disk are repaired (the write
1357 * request for data to be repaired is on its way).
1358 * Just be lazy and use scrub_recheck_block()
1359 * which re-reads the data before the checksum
1360 * is verified, but most likely the data comes out
1361 * of the page cache.
1362 */
1363 scrub_recheck_block(fs_info, sblock_bad, 1);
1364 if (!sblock_bad->header_error &&
1365 !sblock_bad->checksum_error &&
1366 sblock_bad->no_io_error_seen)
1367 goto corrected_error;
1368 else
1369 goto did_not_correct_error;
1370 } else {
1371corrected_error:
1372 spin_lock(&sctx->stat_lock);
1373 sctx->stat.corrected_errors++;
1374 sblock_to_check->data_corrected = 1;
1375 spin_unlock(&sctx->stat_lock);
1376 btrfs_err_rl_in_rcu(fs_info,
1377 "fixed up error at logical %llu on dev %s",
1378 logical, btrfs_dev_name(dev));
1379 }
1380 } else {
1381did_not_correct_error:
1382 spin_lock(&sctx->stat_lock);
1383 sctx->stat.uncorrectable_errors++;
1384 spin_unlock(&sctx->stat_lock);
1385 btrfs_err_rl_in_rcu(fs_info,
1386 "unable to fixup (regular) error at logical %llu on dev %s",
1387 logical, btrfs_dev_name(dev));
1388 }
1389
1390out:
1391 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1392 struct scrub_block *sblock = sblocks_for_recheck[mirror_index];
1393 struct scrub_recover *recover;
1394 int sector_index;
1395
1396 /* Not allocated, continue checking the next mirror */
1397 if (!sblock)
1398 continue;
1399
1400 for (sector_index = 0; sector_index < sblock->sector_count;
1401 sector_index++) {
1402 /*
1403 * Here we just cleanup the recover, each sector will be
1404 * properly cleaned up by later scrub_block_put()
1405 */
1406 recover = sblock->sectors[sector_index]->recover;
1407 if (recover) {
1408 scrub_put_recover(fs_info, recover);
1409 sblock->sectors[sector_index]->recover = NULL;
1410 }
1411 }
1412 scrub_block_put(sblock);
1413 }
1414
1415 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1416 memalloc_nofs_restore(nofs_flag);
1417 if (ret < 0)
1418 return ret;
1419 return 0;
1420}
1421
1422static inline int scrub_nr_raid_mirrors(struct btrfs_io_context *bioc)
1423{
1424 if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID5)
1425 return 2;
1426 else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID6)
1427 return 3;
1428 else
1429 return (int)bioc->num_stripes;
1430}
1431
1432static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1433 u64 *raid_map,
1434 int nstripes, int mirror,
1435 int *stripe_index,
1436 u64 *stripe_offset)
1437{
1438 int i;
1439
1440 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1441 /* RAID5/6 */
1442 for (i = 0; i < nstripes; i++) {
1443 if (raid_map[i] == RAID6_Q_STRIPE ||
1444 raid_map[i] == RAID5_P_STRIPE)
1445 continue;
1446
1447 if (logical >= raid_map[i] &&
1448 logical < raid_map[i] + BTRFS_STRIPE_LEN)
1449 break;
1450 }
1451
1452 *stripe_index = i;
1453 *stripe_offset = logical - raid_map[i];
1454 } else {
1455 /* The other RAID type */
1456 *stripe_index = mirror;
1457 *stripe_offset = 0;
1458 }
1459}
1460
1461static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1462 struct scrub_block *sblocks_for_recheck[])
1463{
1464 struct scrub_ctx *sctx = original_sblock->sctx;
1465 struct btrfs_fs_info *fs_info = sctx->fs_info;
1466 u64 logical = original_sblock->logical;
1467 u64 length = original_sblock->sector_count << fs_info->sectorsize_bits;
1468 u64 generation = original_sblock->sectors[0]->generation;
1469 u64 flags = original_sblock->sectors[0]->flags;
1470 u64 have_csum = original_sblock->sectors[0]->have_csum;
1471 struct scrub_recover *recover;
1472 struct btrfs_io_context *bioc;
1473 u64 sublen;
1474 u64 mapped_length;
1475 u64 stripe_offset;
1476 int stripe_index;
1477 int sector_index = 0;
1478 int mirror_index;
1479 int nmirrors;
1480 int ret;
1481
1482 while (length > 0) {
1483 sublen = min_t(u64, length, fs_info->sectorsize);
1484 mapped_length = sublen;
1485 bioc = NULL;
1486
1487 /*
1488 * With a length of sectorsize, each returned stripe represents
1489 * one mirror
1490 */
1491 btrfs_bio_counter_inc_blocked(fs_info);
1492 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1493 logical, &mapped_length, &bioc);
1494 if (ret || !bioc || mapped_length < sublen) {
1495 btrfs_put_bioc(bioc);
1496 btrfs_bio_counter_dec(fs_info);
1497 return -EIO;
1498 }
1499
1500 recover = kzalloc(sizeof(struct scrub_recover), GFP_KERNEL);
1501 if (!recover) {
1502 btrfs_put_bioc(bioc);
1503 btrfs_bio_counter_dec(fs_info);
1504 return -ENOMEM;
1505 }
1506
1507 refcount_set(&recover->refs, 1);
1508 recover->bioc = bioc;
1509 recover->map_length = mapped_length;
1510
1511 ASSERT(sector_index < SCRUB_MAX_SECTORS_PER_BLOCK);
1512
1513 nmirrors = min(scrub_nr_raid_mirrors(bioc), BTRFS_MAX_MIRRORS);
1514
1515 for (mirror_index = 0; mirror_index < nmirrors;
1516 mirror_index++) {
1517 struct scrub_block *sblock;
1518 struct scrub_sector *sector;
1519
1520 sblock = sblocks_for_recheck[mirror_index];
1521 sblock->sctx = sctx;
1522
1523 sector = alloc_scrub_sector(sblock, logical);
1524 if (!sector) {
1525 spin_lock(&sctx->stat_lock);
1526 sctx->stat.malloc_errors++;
1527 spin_unlock(&sctx->stat_lock);
1528 scrub_put_recover(fs_info, recover);
1529 return -ENOMEM;
1530 }
1531 sector->flags = flags;
1532 sector->generation = generation;
1533 sector->have_csum = have_csum;
1534 if (have_csum)
1535 memcpy(sector->csum,
1536 original_sblock->sectors[0]->csum,
1537 sctx->fs_info->csum_size);
1538
1539 scrub_stripe_index_and_offset(logical,
1540 bioc->map_type,
1541 bioc->raid_map,
1542 bioc->num_stripes -
1543 bioc->num_tgtdevs,
1544 mirror_index,
1545 &stripe_index,
1546 &stripe_offset);
1547 /*
1548 * We're at the first sector, also populate @sblock
1549 * physical and dev.
1550 */
1551 if (sector_index == 0) {
1552 sblock->physical =
1553 bioc->stripes[stripe_index].physical +
1554 stripe_offset;
1555 sblock->dev = bioc->stripes[stripe_index].dev;
1556 sblock->physical_for_dev_replace =
1557 original_sblock->physical_for_dev_replace;
1558 }
1559
1560 BUG_ON(sector_index >= original_sblock->sector_count);
1561 scrub_get_recover(recover);
1562 sector->recover = recover;
1563 }
1564 scrub_put_recover(fs_info, recover);
1565 length -= sublen;
1566 logical += sublen;
1567 sector_index++;
1568 }
1569
1570 return 0;
1571}
1572
1573static void scrub_bio_wait_endio(struct bio *bio)
1574{
1575 complete(bio->bi_private);
1576}
1577
1578static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1579 struct bio *bio,
1580 struct scrub_sector *sector)
1581{
1582 DECLARE_COMPLETION_ONSTACK(done);
1583
1584 bio->bi_iter.bi_sector = (sector->offset + sector->sblock->logical) >>
1585 SECTOR_SHIFT;
1586 bio->bi_private = &done;
1587 bio->bi_end_io = scrub_bio_wait_endio;
1588 raid56_parity_recover(bio, sector->recover->bioc, sector->sblock->mirror_num);
1589
1590 wait_for_completion_io(&done);
1591 return blk_status_to_errno(bio->bi_status);
1592}
1593
1594static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1595 struct scrub_block *sblock)
1596{
1597 struct scrub_sector *first_sector = sblock->sectors[0];
1598 struct bio *bio;
1599 int i;
1600
1601 /* All sectors in sblock belong to the same stripe on the same device. */
1602 ASSERT(sblock->dev);
1603 if (!sblock->dev->bdev)
1604 goto out;
1605
1606 bio = bio_alloc(sblock->dev->bdev, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
1607
1608 for (i = 0; i < sblock->sector_count; i++) {
1609 struct scrub_sector *sector = sblock->sectors[i];
1610
1611 bio_add_scrub_sector(bio, sector, fs_info->sectorsize);
1612 }
1613
1614 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_sector)) {
1615 bio_put(bio);
1616 goto out;
1617 }
1618
1619 bio_put(bio);
1620
1621 scrub_recheck_block_checksum(sblock);
1622
1623 return;
1624out:
1625 for (i = 0; i < sblock->sector_count; i++)
1626 sblock->sectors[i]->io_error = 1;
1627
1628 sblock->no_io_error_seen = 0;
1629}
1630
1631/*
1632 * This function will check the on disk data for checksum errors, header errors
1633 * and read I/O errors. If any I/O errors happen, the exact sectors which are
1634 * errored are marked as being bad. The goal is to enable scrub to take those
1635 * sectors that are not errored from all the mirrors so that the sectors that
1636 * are errored in the just handled mirror can be repaired.
1637 */
1638static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1639 struct scrub_block *sblock,
1640 int retry_failed_mirror)
1641{
1642 int i;
1643
1644 sblock->no_io_error_seen = 1;
1645
1646 /* short cut for raid56 */
1647 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->sectors[0]))
1648 return scrub_recheck_block_on_raid56(fs_info, sblock);
1649
1650 for (i = 0; i < sblock->sector_count; i++) {
1651 struct scrub_sector *sector = sblock->sectors[i];
1652 struct bio bio;
1653 struct bio_vec bvec;
1654
1655 if (sblock->dev->bdev == NULL) {
1656 sector->io_error = 1;
1657 sblock->no_io_error_seen = 0;
1658 continue;
1659 }
1660
1661 bio_init(&bio, sblock->dev->bdev, &bvec, 1, REQ_OP_READ);
1662 bio_add_scrub_sector(&bio, sector, fs_info->sectorsize);
1663 bio.bi_iter.bi_sector = (sblock->physical + sector->offset) >>
1664 SECTOR_SHIFT;
1665
1666 btrfsic_check_bio(&bio);
1667 if (submit_bio_wait(&bio)) {
1668 sector->io_error = 1;
1669 sblock->no_io_error_seen = 0;
1670 }
1671
1672 bio_uninit(&bio);
1673 }
1674
1675 if (sblock->no_io_error_seen)
1676 scrub_recheck_block_checksum(sblock);
1677}
1678
1679static inline int scrub_check_fsid(u8 fsid[], struct scrub_sector *sector)
1680{
1681 struct btrfs_fs_devices *fs_devices = sector->sblock->dev->fs_devices;
1682 int ret;
1683
1684 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1685 return !ret;
1686}
1687
1688static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1689{
1690 sblock->header_error = 0;
1691 sblock->checksum_error = 0;
1692 sblock->generation_error = 0;
1693
1694 if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1695 scrub_checksum_data(sblock);
1696 else
1697 scrub_checksum_tree_block(sblock);
1698}
1699
1700static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1701 struct scrub_block *sblock_good)
1702{
1703 int i;
1704 int ret = 0;
1705
1706 for (i = 0; i < sblock_bad->sector_count; i++) {
1707 int ret_sub;
1708
1709 ret_sub = scrub_repair_sector_from_good_copy(sblock_bad,
1710 sblock_good, i, 1);
1711 if (ret_sub)
1712 ret = ret_sub;
1713 }
1714
1715 return ret;
1716}
1717
1718static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
1719 struct scrub_block *sblock_good,
1720 int sector_num, int force_write)
1721{
1722 struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1723 struct scrub_sector *sector_good = sblock_good->sectors[sector_num];
1724 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1725 const u32 sectorsize = fs_info->sectorsize;
1726
1727 if (force_write || sblock_bad->header_error ||
1728 sblock_bad->checksum_error || sector_bad->io_error) {
1729 struct bio bio;
1730 struct bio_vec bvec;
1731 int ret;
1732
1733 if (!sblock_bad->dev->bdev) {
1734 btrfs_warn_rl(fs_info,
1735 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1736 return -EIO;
1737 }
1738
1739 bio_init(&bio, sblock_bad->dev->bdev, &bvec, 1, REQ_OP_WRITE);
1740 bio.bi_iter.bi_sector = (sblock_bad->physical +
1741 sector_bad->offset) >> SECTOR_SHIFT;
1742 ret = bio_add_scrub_sector(&bio, sector_good, sectorsize);
1743
1744 btrfsic_check_bio(&bio);
1745 ret = submit_bio_wait(&bio);
1746 bio_uninit(&bio);
1747
1748 if (ret) {
1749 btrfs_dev_stat_inc_and_print(sblock_bad->dev,
1750 BTRFS_DEV_STAT_WRITE_ERRS);
1751 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1752 return -EIO;
1753 }
1754 }
1755
1756 return 0;
1757}
1758
1759static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1760{
1761 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1762 int i;
1763
1764 /*
1765 * This block is used for the check of the parity on the source device,
1766 * so the data needn't be written into the destination device.
1767 */
1768 if (sblock->sparity)
1769 return;
1770
1771 for (i = 0; i < sblock->sector_count; i++) {
1772 int ret;
1773
1774 ret = scrub_write_sector_to_dev_replace(sblock, i);
1775 if (ret)
1776 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1777 }
1778}
1779
1780static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock, int sector_num)
1781{
1782 const u32 sectorsize = sblock->sctx->fs_info->sectorsize;
1783 struct scrub_sector *sector = sblock->sectors[sector_num];
1784
1785 if (sector->io_error)
1786 memset(scrub_sector_get_kaddr(sector), 0, sectorsize);
1787
1788 return scrub_add_sector_to_wr_bio(sblock->sctx, sector);
1789}
1790
1791static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
1792{
1793 int ret = 0;
1794 u64 length;
1795
1796 if (!btrfs_is_zoned(sctx->fs_info))
1797 return 0;
1798
1799 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1800 return 0;
1801
1802 if (sctx->write_pointer < physical) {
1803 length = physical - sctx->write_pointer;
1804
1805 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1806 sctx->write_pointer, length);
1807 if (!ret)
1808 sctx->write_pointer = physical;
1809 }
1810 return ret;
1811}
1812
1813static void scrub_block_get(struct scrub_block *sblock)
1814{
1815 refcount_inc(&sblock->refs);
1816}
1817
1818static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
1819 struct scrub_sector *sector)
1820{
1821 struct scrub_block *sblock = sector->sblock;
1822 struct scrub_bio *sbio;
1823 int ret;
1824 const u32 sectorsize = sctx->fs_info->sectorsize;
1825
1826 mutex_lock(&sctx->wr_lock);
1827again:
1828 if (!sctx->wr_curr_bio) {
1829 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1830 GFP_KERNEL);
1831 if (!sctx->wr_curr_bio) {
1832 mutex_unlock(&sctx->wr_lock);
1833 return -ENOMEM;
1834 }
1835 sctx->wr_curr_bio->sctx = sctx;
1836 sctx->wr_curr_bio->sector_count = 0;
1837 }
1838 sbio = sctx->wr_curr_bio;
1839 if (sbio->sector_count == 0) {
1840 ret = fill_writer_pointer_gap(sctx, sector->offset +
1841 sblock->physical_for_dev_replace);
1842 if (ret) {
1843 mutex_unlock(&sctx->wr_lock);
1844 return ret;
1845 }
1846
1847 sbio->physical = sblock->physical_for_dev_replace + sector->offset;
1848 sbio->logical = sblock->logical + sector->offset;
1849 sbio->dev = sctx->wr_tgtdev;
1850 if (!sbio->bio) {
1851 sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
1852 REQ_OP_WRITE, GFP_NOFS);
1853 }
1854 sbio->bio->bi_private = sbio;
1855 sbio->bio->bi_end_io = scrub_wr_bio_end_io;
1856 sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
1857 sbio->status = 0;
1858 } else if (sbio->physical + sbio->sector_count * sectorsize !=
1859 sblock->physical_for_dev_replace + sector->offset ||
1860 sbio->logical + sbio->sector_count * sectorsize !=
1861 sblock->logical + sector->offset) {
1862 scrub_wr_submit(sctx);
1863 goto again;
1864 }
1865
1866 ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
1867 if (ret != sectorsize) {
1868 if (sbio->sector_count < 1) {
1869 bio_put(sbio->bio);
1870 sbio->bio = NULL;
1871 mutex_unlock(&sctx->wr_lock);
1872 return -EIO;
1873 }
1874 scrub_wr_submit(sctx);
1875 goto again;
1876 }
1877
1878 sbio->sectors[sbio->sector_count] = sector;
1879 scrub_sector_get(sector);
1880 /*
1881 * Since ssector no longer holds a page, but uses sblock::pages, we
1882 * have to ensure the sblock had not been freed before our write bio
1883 * finished.
1884 */
1885 scrub_block_get(sector->sblock);
1886
1887 sbio->sector_count++;
1888 if (sbio->sector_count == sctx->sectors_per_bio)
1889 scrub_wr_submit(sctx);
1890 mutex_unlock(&sctx->wr_lock);
1891
1892 return 0;
1893}
1894
1895static void scrub_wr_submit(struct scrub_ctx *sctx)
1896{
1897 struct scrub_bio *sbio;
1898
1899 if (!sctx->wr_curr_bio)
1900 return;
1901
1902 sbio = sctx->wr_curr_bio;
1903 sctx->wr_curr_bio = NULL;
1904 scrub_pending_bio_inc(sctx);
1905 /* process all writes in a single worker thread. Then the block layer
1906 * orders the requests before sending them to the driver which
1907 * doubled the write performance on spinning disks when measured
1908 * with Linux 3.5 */
1909 btrfsic_check_bio(sbio->bio);
1910 submit_bio(sbio->bio);
1911
1912 if (btrfs_is_zoned(sctx->fs_info))
1913 sctx->write_pointer = sbio->physical + sbio->sector_count *
1914 sctx->fs_info->sectorsize;
1915}
1916
1917static void scrub_wr_bio_end_io(struct bio *bio)
1918{
1919 struct scrub_bio *sbio = bio->bi_private;
1920 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1921
1922 sbio->status = bio->bi_status;
1923 sbio->bio = bio;
1924
1925 INIT_WORK(&sbio->work, scrub_wr_bio_end_io_worker);
1926 queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1927}
1928
1929static void scrub_wr_bio_end_io_worker(struct work_struct *work)
1930{
1931 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1932 struct scrub_ctx *sctx = sbio->sctx;
1933 int i;
1934
1935 ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
1936 if (sbio->status) {
1937 struct btrfs_dev_replace *dev_replace =
1938 &sbio->sctx->fs_info->dev_replace;
1939
1940 for (i = 0; i < sbio->sector_count; i++) {
1941 struct scrub_sector *sector = sbio->sectors[i];
1942
1943 sector->io_error = 1;
1944 atomic64_inc(&dev_replace->num_write_errors);
1945 }
1946 }
1947
1948 /*
1949 * In scrub_add_sector_to_wr_bio() we grab extra ref for sblock, now in
1950 * endio we should put the sblock.
1951 */
1952 for (i = 0; i < sbio->sector_count; i++) {
1953 scrub_block_put(sbio->sectors[i]->sblock);
1954 scrub_sector_put(sbio->sectors[i]);
1955 }
1956
1957 bio_put(sbio->bio);
1958 kfree(sbio);
1959 scrub_pending_bio_dec(sctx);
1960}
1961
1962static int scrub_checksum(struct scrub_block *sblock)
1963{
1964 u64 flags;
1965 int ret;
1966
1967 /*
1968 * No need to initialize these stats currently,
1969 * because this function only use return value
1970 * instead of these stats value.
1971 *
1972 * Todo:
1973 * always use stats
1974 */
1975 sblock->header_error = 0;
1976 sblock->generation_error = 0;
1977 sblock->checksum_error = 0;
1978
1979 WARN_ON(sblock->sector_count < 1);
1980 flags = sblock->sectors[0]->flags;
1981 ret = 0;
1982 if (flags & BTRFS_EXTENT_FLAG_DATA)
1983 ret = scrub_checksum_data(sblock);
1984 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1985 ret = scrub_checksum_tree_block(sblock);
1986 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1987 ret = scrub_checksum_super(sblock);
1988 else
1989 WARN_ON(1);
1990 if (ret)
1991 scrub_handle_errored_block(sblock);
1992
1993 return ret;
1994}
1995
1996static int scrub_checksum_data(struct scrub_block *sblock)
1997{
1998 struct scrub_ctx *sctx = sblock->sctx;
1999 struct btrfs_fs_info *fs_info = sctx->fs_info;
2000 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2001 u8 csum[BTRFS_CSUM_SIZE];
2002 struct scrub_sector *sector;
2003 char *kaddr;
2004
2005 BUG_ON(sblock->sector_count < 1);
2006 sector = sblock->sectors[0];
2007 if (!sector->have_csum)
2008 return 0;
2009
2010 kaddr = scrub_sector_get_kaddr(sector);
2011
2012 shash->tfm = fs_info->csum_shash;
2013 crypto_shash_init(shash);
2014
2015 crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
2016
2017 if (memcmp(csum, sector->csum, fs_info->csum_size))
2018 sblock->checksum_error = 1;
2019 return sblock->checksum_error;
2020}
2021
2022static int scrub_checksum_tree_block(struct scrub_block *sblock)
2023{
2024 struct scrub_ctx *sctx = sblock->sctx;
2025 struct btrfs_header *h;
2026 struct btrfs_fs_info *fs_info = sctx->fs_info;
2027 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2028 u8 calculated_csum[BTRFS_CSUM_SIZE];
2029 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2030 /*
2031 * This is done in sectorsize steps even for metadata as there's a
2032 * constraint for nodesize to be aligned to sectorsize. This will need
2033 * to change so we don't misuse data and metadata units like that.
2034 */
2035 const u32 sectorsize = sctx->fs_info->sectorsize;
2036 const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
2037 int i;
2038 struct scrub_sector *sector;
2039 char *kaddr;
2040
2041 BUG_ON(sblock->sector_count < 1);
2042
2043 /* Each member in sectors is just one sector */
2044 ASSERT(sblock->sector_count == num_sectors);
2045
2046 sector = sblock->sectors[0];
2047 kaddr = scrub_sector_get_kaddr(sector);
2048 h = (struct btrfs_header *)kaddr;
2049 memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
2050
2051 /*
2052 * we don't use the getter functions here, as we
2053 * a) don't have an extent buffer and
2054 * b) the page is already kmapped
2055 */
2056 if (sblock->logical != btrfs_stack_header_bytenr(h))
2057 sblock->header_error = 1;
2058
2059 if (sector->generation != btrfs_stack_header_generation(h)) {
2060 sblock->header_error = 1;
2061 sblock->generation_error = 1;
2062 }
2063
2064 if (!scrub_check_fsid(h->fsid, sector))
2065 sblock->header_error = 1;
2066
2067 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2068 BTRFS_UUID_SIZE))
2069 sblock->header_error = 1;
2070
2071 shash->tfm = fs_info->csum_shash;
2072 crypto_shash_init(shash);
2073 crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
2074 sectorsize - BTRFS_CSUM_SIZE);
2075
2076 for (i = 1; i < num_sectors; i++) {
2077 kaddr = scrub_sector_get_kaddr(sblock->sectors[i]);
2078 crypto_shash_update(shash, kaddr, sectorsize);
2079 }
2080
2081 crypto_shash_final(shash, calculated_csum);
2082 if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size))
2083 sblock->checksum_error = 1;
2084
2085 return sblock->header_error || sblock->checksum_error;
2086}
2087
2088static int scrub_checksum_super(struct scrub_block *sblock)
2089{
2090 struct btrfs_super_block *s;
2091 struct scrub_ctx *sctx = sblock->sctx;
2092 struct btrfs_fs_info *fs_info = sctx->fs_info;
2093 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2094 u8 calculated_csum[BTRFS_CSUM_SIZE];
2095 struct scrub_sector *sector;
2096 char *kaddr;
2097 int fail_gen = 0;
2098 int fail_cor = 0;
2099
2100 BUG_ON(sblock->sector_count < 1);
2101 sector = sblock->sectors[0];
2102 kaddr = scrub_sector_get_kaddr(sector);
2103 s = (struct btrfs_super_block *)kaddr;
2104
2105 if (sblock->logical != btrfs_super_bytenr(s))
2106 ++fail_cor;
2107
2108 if (sector->generation != btrfs_super_generation(s))
2109 ++fail_gen;
2110
2111 if (!scrub_check_fsid(s->fsid, sector))
2112 ++fail_cor;
2113
2114 shash->tfm = fs_info->csum_shash;
2115 crypto_shash_init(shash);
2116 crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
2117 BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
2118
2119 if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
2120 ++fail_cor;
2121
2122 return fail_cor + fail_gen;
2123}
2124
2125static void scrub_block_put(struct scrub_block *sblock)
2126{
2127 if (refcount_dec_and_test(&sblock->refs)) {
2128 int i;
2129
2130 if (sblock->sparity)
2131 scrub_parity_put(sblock->sparity);
2132
2133 for (i = 0; i < sblock->sector_count; i++)
2134 scrub_sector_put(sblock->sectors[i]);
2135 for (i = 0; i < DIV_ROUND_UP(sblock->len, PAGE_SIZE); i++) {
2136 if (sblock->pages[i]) {
2137 detach_scrub_page_private(sblock->pages[i]);
2138 __free_page(sblock->pages[i]);
2139 }
2140 }
2141 kfree(sblock);
2142 }
2143}
2144
2145static void scrub_sector_get(struct scrub_sector *sector)
2146{
2147 atomic_inc(§or->refs);
2148}
2149
2150static void scrub_sector_put(struct scrub_sector *sector)
2151{
2152 if (atomic_dec_and_test(§or->refs))
2153 kfree(sector);
2154}
2155
2156/*
2157 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
2158 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
2159 */
2160static void scrub_throttle(struct scrub_ctx *sctx)
2161{
2162 const int time_slice = 1000;
2163 struct scrub_bio *sbio;
2164 struct btrfs_device *device;
2165 s64 delta;
2166 ktime_t now;
2167 u32 div;
2168 u64 bwlimit;
2169
2170 sbio = sctx->bios[sctx->curr];
2171 device = sbio->dev;
2172 bwlimit = READ_ONCE(device->scrub_speed_max);
2173 if (bwlimit == 0)
2174 return;
2175
2176 /*
2177 * Slice is divided into intervals when the IO is submitted, adjust by
2178 * bwlimit and maximum of 64 intervals.
2179 */
2180 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
2181 div = min_t(u32, 64, div);
2182
2183 /* Start new epoch, set deadline */
2184 now = ktime_get();
2185 if (sctx->throttle_deadline == 0) {
2186 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
2187 sctx->throttle_sent = 0;
2188 }
2189
2190 /* Still in the time to send? */
2191 if (ktime_before(now, sctx->throttle_deadline)) {
2192 /* If current bio is within the limit, send it */
2193 sctx->throttle_sent += sbio->bio->bi_iter.bi_size;
2194 if (sctx->throttle_sent <= div_u64(bwlimit, div))
2195 return;
2196
2197 /* We're over the limit, sleep until the rest of the slice */
2198 delta = ktime_ms_delta(sctx->throttle_deadline, now);
2199 } else {
2200 /* New request after deadline, start new epoch */
2201 delta = 0;
2202 }
2203
2204 if (delta) {
2205 long timeout;
2206
2207 timeout = div_u64(delta * HZ, 1000);
2208 schedule_timeout_interruptible(timeout);
2209 }
2210
2211 /* Next call will start the deadline period */
2212 sctx->throttle_deadline = 0;
2213}
2214
2215static void scrub_submit(struct scrub_ctx *sctx)
2216{
2217 struct scrub_bio *sbio;
2218
2219 if (sctx->curr == -1)
2220 return;
2221
2222 scrub_throttle(sctx);
2223
2224 sbio = sctx->bios[sctx->curr];
2225 sctx->curr = -1;
2226 scrub_pending_bio_inc(sctx);
2227 btrfsic_check_bio(sbio->bio);
2228 submit_bio(sbio->bio);
2229}
2230
2231static int scrub_add_sector_to_rd_bio(struct scrub_ctx *sctx,
2232 struct scrub_sector *sector)
2233{
2234 struct scrub_block *sblock = sector->sblock;
2235 struct scrub_bio *sbio;
2236 const u32 sectorsize = sctx->fs_info->sectorsize;
2237 int ret;
2238
2239again:
2240 /*
2241 * grab a fresh bio or wait for one to become available
2242 */
2243 while (sctx->curr == -1) {
2244 spin_lock(&sctx->list_lock);
2245 sctx->curr = sctx->first_free;
2246 if (sctx->curr != -1) {
2247 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2248 sctx->bios[sctx->curr]->next_free = -1;
2249 sctx->bios[sctx->curr]->sector_count = 0;
2250 spin_unlock(&sctx->list_lock);
2251 } else {
2252 spin_unlock(&sctx->list_lock);
2253 wait_event(sctx->list_wait, sctx->first_free != -1);
2254 }
2255 }
2256 sbio = sctx->bios[sctx->curr];
2257 if (sbio->sector_count == 0) {
2258 sbio->physical = sblock->physical + sector->offset;
2259 sbio->logical = sblock->logical + sector->offset;
2260 sbio->dev = sblock->dev;
2261 if (!sbio->bio) {
2262 sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
2263 REQ_OP_READ, GFP_NOFS);
2264 }
2265 sbio->bio->bi_private = sbio;
2266 sbio->bio->bi_end_io = scrub_bio_end_io;
2267 sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
2268 sbio->status = 0;
2269 } else if (sbio->physical + sbio->sector_count * sectorsize !=
2270 sblock->physical + sector->offset ||
2271 sbio->logical + sbio->sector_count * sectorsize !=
2272 sblock->logical + sector->offset ||
2273 sbio->dev != sblock->dev) {
2274 scrub_submit(sctx);
2275 goto again;
2276 }
2277
2278 sbio->sectors[sbio->sector_count] = sector;
2279 ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
2280 if (ret != sectorsize) {
2281 if (sbio->sector_count < 1) {
2282 bio_put(sbio->bio);
2283 sbio->bio = NULL;
2284 return -EIO;
2285 }
2286 scrub_submit(sctx);
2287 goto again;
2288 }
2289
2290 scrub_block_get(sblock); /* one for the page added to the bio */
2291 atomic_inc(&sblock->outstanding_sectors);
2292 sbio->sector_count++;
2293 if (sbio->sector_count == sctx->sectors_per_bio)
2294 scrub_submit(sctx);
2295
2296 return 0;
2297}
2298
2299static void scrub_missing_raid56_end_io(struct bio *bio)
2300{
2301 struct scrub_block *sblock = bio->bi_private;
2302 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2303
2304 btrfs_bio_counter_dec(fs_info);
2305 if (bio->bi_status)
2306 sblock->no_io_error_seen = 0;
2307
2308 bio_put(bio);
2309
2310 queue_work(fs_info->scrub_workers, &sblock->work);
2311}
2312
2313static void scrub_missing_raid56_worker(struct work_struct *work)
2314{
2315 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2316 struct scrub_ctx *sctx = sblock->sctx;
2317 struct btrfs_fs_info *fs_info = sctx->fs_info;
2318 u64 logical;
2319 struct btrfs_device *dev;
2320
2321 logical = sblock->logical;
2322 dev = sblock->dev;
2323
2324 if (sblock->no_io_error_seen)
2325 scrub_recheck_block_checksum(sblock);
2326
2327 if (!sblock->no_io_error_seen) {
2328 spin_lock(&sctx->stat_lock);
2329 sctx->stat.read_errors++;
2330 spin_unlock(&sctx->stat_lock);
2331 btrfs_err_rl_in_rcu(fs_info,
2332 "IO error rebuilding logical %llu for dev %s",
2333 logical, btrfs_dev_name(dev));
2334 } else if (sblock->header_error || sblock->checksum_error) {
2335 spin_lock(&sctx->stat_lock);
2336 sctx->stat.uncorrectable_errors++;
2337 spin_unlock(&sctx->stat_lock);
2338 btrfs_err_rl_in_rcu(fs_info,
2339 "failed to rebuild valid logical %llu for dev %s",
2340 logical, btrfs_dev_name(dev));
2341 } else {
2342 scrub_write_block_to_dev_replace(sblock);
2343 }
2344
2345 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2346 mutex_lock(&sctx->wr_lock);
2347 scrub_wr_submit(sctx);
2348 mutex_unlock(&sctx->wr_lock);
2349 }
2350
2351 scrub_block_put(sblock);
2352 scrub_pending_bio_dec(sctx);
2353}
2354
2355static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2356{
2357 struct scrub_ctx *sctx = sblock->sctx;
2358 struct btrfs_fs_info *fs_info = sctx->fs_info;
2359 u64 length = sblock->sector_count << fs_info->sectorsize_bits;
2360 u64 logical = sblock->logical;
2361 struct btrfs_io_context *bioc = NULL;
2362 struct bio *bio;
2363 struct btrfs_raid_bio *rbio;
2364 int ret;
2365 int i;
2366
2367 btrfs_bio_counter_inc_blocked(fs_info);
2368 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2369 &length, &bioc);
2370 if (ret || !bioc || !bioc->raid_map)
2371 goto bioc_out;
2372
2373 if (WARN_ON(!sctx->is_dev_replace ||
2374 !(bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2375 /*
2376 * We shouldn't be scrubbing a missing device. Even for dev
2377 * replace, we should only get here for RAID 5/6. We either
2378 * managed to mount something with no mirrors remaining or
2379 * there's a bug in scrub_find_good_copy()/btrfs_map_block().
2380 */
2381 goto bioc_out;
2382 }
2383
2384 bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2385 bio->bi_iter.bi_sector = logical >> 9;
2386 bio->bi_private = sblock;
2387 bio->bi_end_io = scrub_missing_raid56_end_io;
2388
2389 rbio = raid56_alloc_missing_rbio(bio, bioc);
2390 if (!rbio)
2391 goto rbio_out;
2392
2393 for (i = 0; i < sblock->sector_count; i++) {
2394 struct scrub_sector *sector = sblock->sectors[i];
2395
2396 raid56_add_scrub_pages(rbio, scrub_sector_get_page(sector),
2397 scrub_sector_get_page_offset(sector),
2398 sector->offset + sector->sblock->logical);
2399 }
2400
2401 INIT_WORK(&sblock->work, scrub_missing_raid56_worker);
2402 scrub_block_get(sblock);
2403 scrub_pending_bio_inc(sctx);
2404 raid56_submit_missing_rbio(rbio);
2405 btrfs_put_bioc(bioc);
2406 return;
2407
2408rbio_out:
2409 bio_put(bio);
2410bioc_out:
2411 btrfs_bio_counter_dec(fs_info);
2412 btrfs_put_bioc(bioc);
2413 spin_lock(&sctx->stat_lock);
2414 sctx->stat.malloc_errors++;
2415 spin_unlock(&sctx->stat_lock);
2416}
2417
2418static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
2419 u64 physical, struct btrfs_device *dev, u64 flags,
2420 u64 gen, int mirror_num, u8 *csum,
2421 u64 physical_for_dev_replace)
2422{
2423 struct scrub_block *sblock;
2424 const u32 sectorsize = sctx->fs_info->sectorsize;
2425 int index;
2426
2427 sblock = alloc_scrub_block(sctx, dev, logical, physical,
2428 physical_for_dev_replace, mirror_num);
2429 if (!sblock) {
2430 spin_lock(&sctx->stat_lock);
2431 sctx->stat.malloc_errors++;
2432 spin_unlock(&sctx->stat_lock);
2433 return -ENOMEM;
2434 }
2435
2436 for (index = 0; len > 0; index++) {
2437 struct scrub_sector *sector;
2438 /*
2439 * Here we will allocate one page for one sector to scrub.
2440 * This is fine if PAGE_SIZE == sectorsize, but will cost
2441 * more memory for PAGE_SIZE > sectorsize case.
2442 */
2443 u32 l = min(sectorsize, len);
2444
2445 sector = alloc_scrub_sector(sblock, logical);
2446 if (!sector) {
2447 spin_lock(&sctx->stat_lock);
2448 sctx->stat.malloc_errors++;
2449 spin_unlock(&sctx->stat_lock);
2450 scrub_block_put(sblock);
2451 return -ENOMEM;
2452 }
2453 sector->flags = flags;
2454 sector->generation = gen;
2455 if (csum) {
2456 sector->have_csum = 1;
2457 memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2458 } else {
2459 sector->have_csum = 0;
2460 }
2461 len -= l;
2462 logical += l;
2463 physical += l;
2464 physical_for_dev_replace += l;
2465 }
2466
2467 WARN_ON(sblock->sector_count == 0);
2468 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2469 /*
2470 * This case should only be hit for RAID 5/6 device replace. See
2471 * the comment in scrub_missing_raid56_pages() for details.
2472 */
2473 scrub_missing_raid56_pages(sblock);
2474 } else {
2475 for (index = 0; index < sblock->sector_count; index++) {
2476 struct scrub_sector *sector = sblock->sectors[index];
2477 int ret;
2478
2479 ret = scrub_add_sector_to_rd_bio(sctx, sector);
2480 if (ret) {
2481 scrub_block_put(sblock);
2482 return ret;
2483 }
2484 }
2485
2486 if (flags & BTRFS_EXTENT_FLAG_SUPER)
2487 scrub_submit(sctx);
2488 }
2489
2490 /* last one frees, either here or in bio completion for last page */
2491 scrub_block_put(sblock);
2492 return 0;
2493}
2494
2495static void scrub_bio_end_io(struct bio *bio)
2496{
2497 struct scrub_bio *sbio = bio->bi_private;
2498 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2499
2500 sbio->status = bio->bi_status;
2501 sbio->bio = bio;
2502
2503 queue_work(fs_info->scrub_workers, &sbio->work);
2504}
2505
2506static void scrub_bio_end_io_worker(struct work_struct *work)
2507{
2508 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2509 struct scrub_ctx *sctx = sbio->sctx;
2510 int i;
2511
2512 ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
2513 if (sbio->status) {
2514 for (i = 0; i < sbio->sector_count; i++) {
2515 struct scrub_sector *sector = sbio->sectors[i];
2516
2517 sector->io_error = 1;
2518 sector->sblock->no_io_error_seen = 0;
2519 }
2520 }
2521
2522 /* Now complete the scrub_block items that have all pages completed */
2523 for (i = 0; i < sbio->sector_count; i++) {
2524 struct scrub_sector *sector = sbio->sectors[i];
2525 struct scrub_block *sblock = sector->sblock;
2526
2527 if (atomic_dec_and_test(&sblock->outstanding_sectors))
2528 scrub_block_complete(sblock);
2529 scrub_block_put(sblock);
2530 }
2531
2532 bio_put(sbio->bio);
2533 sbio->bio = NULL;
2534 spin_lock(&sctx->list_lock);
2535 sbio->next_free = sctx->first_free;
2536 sctx->first_free = sbio->index;
2537 spin_unlock(&sctx->list_lock);
2538
2539 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2540 mutex_lock(&sctx->wr_lock);
2541 scrub_wr_submit(sctx);
2542 mutex_unlock(&sctx->wr_lock);
2543 }
2544
2545 scrub_pending_bio_dec(sctx);
2546}
2547
2548static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2549 unsigned long *bitmap,
2550 u64 start, u32 len)
2551{
2552 u64 offset;
2553 u32 nsectors;
2554 u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2555
2556 if (len >= sparity->stripe_len) {
2557 bitmap_set(bitmap, 0, sparity->nsectors);
2558 return;
2559 }
2560
2561 start -= sparity->logic_start;
2562 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2563 offset = offset >> sectorsize_bits;
2564 nsectors = len >> sectorsize_bits;
2565
2566 if (offset + nsectors <= sparity->nsectors) {
2567 bitmap_set(bitmap, offset, nsectors);
2568 return;
2569 }
2570
2571 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2572 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2573}
2574
2575static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2576 u64 start, u32 len)
2577{
2578 __scrub_mark_bitmap(sparity, &sparity->ebitmap, start, len);
2579}
2580
2581static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2582 u64 start, u32 len)
2583{
2584 __scrub_mark_bitmap(sparity, &sparity->dbitmap, start, len);
2585}
2586
2587static void scrub_block_complete(struct scrub_block *sblock)
2588{
2589 int corrupted = 0;
2590
2591 if (!sblock->no_io_error_seen) {
2592 corrupted = 1;
2593 scrub_handle_errored_block(sblock);
2594 } else {
2595 /*
2596 * if has checksum error, write via repair mechanism in
2597 * dev replace case, otherwise write here in dev replace
2598 * case.
2599 */
2600 corrupted = scrub_checksum(sblock);
2601 if (!corrupted && sblock->sctx->is_dev_replace)
2602 scrub_write_block_to_dev_replace(sblock);
2603 }
2604
2605 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2606 u64 start = sblock->logical;
2607 u64 end = sblock->logical +
2608 sblock->sectors[sblock->sector_count - 1]->offset +
2609 sblock->sctx->fs_info->sectorsize;
2610
2611 ASSERT(end - start <= U32_MAX);
2612 scrub_parity_mark_sectors_error(sblock->sparity,
2613 start, end - start);
2614 }
2615}
2616
2617static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2618{
2619 sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2620 list_del(&sum->list);
2621 kfree(sum);
2622}
2623
2624/*
2625 * Find the desired csum for range [logical, logical + sectorsize), and store
2626 * the csum into @csum.
2627 *
2628 * The search source is sctx->csum_list, which is a pre-populated list
2629 * storing bytenr ordered csum ranges. We're responsible to cleanup any range
2630 * that is before @logical.
2631 *
2632 * Return 0 if there is no csum for the range.
2633 * Return 1 if there is csum for the range and copied to @csum.
2634 */
2635static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2636{
2637 bool found = false;
2638
2639 while (!list_empty(&sctx->csum_list)) {
2640 struct btrfs_ordered_sum *sum = NULL;
2641 unsigned long index;
2642 unsigned long num_sectors;
2643
2644 sum = list_first_entry(&sctx->csum_list,
2645 struct btrfs_ordered_sum, list);
2646 /* The current csum range is beyond our range, no csum found */
2647 if (sum->bytenr > logical)
2648 break;
2649
2650 /*
2651 * The current sum is before our bytenr, since scrub is always
2652 * done in bytenr order, the csum will never be used anymore,
2653 * clean it up so that later calls won't bother with the range,
2654 * and continue search the next range.
2655 */
2656 if (sum->bytenr + sum->len <= logical) {
2657 drop_csum_range(sctx, sum);
2658 continue;
2659 }
2660
2661 /* Now the csum range covers our bytenr, copy the csum */
2662 found = true;
2663 index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2664 num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2665
2666 memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2667 sctx->fs_info->csum_size);
2668
2669 /* Cleanup the range if we're at the end of the csum range */
2670 if (index == num_sectors - 1)
2671 drop_csum_range(sctx, sum);
2672 break;
2673 }
2674 if (!found)
2675 return 0;
2676 return 1;
2677}
2678
2679/* scrub extent tries to collect up to 64 kB for each bio */
2680static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2681 u64 logical, u32 len,
2682 u64 physical, struct btrfs_device *dev, u64 flags,
2683 u64 gen, int mirror_num)
2684{
2685 struct btrfs_device *src_dev = dev;
2686 u64 src_physical = physical;
2687 int src_mirror = mirror_num;
2688 int ret;
2689 u8 csum[BTRFS_CSUM_SIZE];
2690 u32 blocksize;
2691
2692 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2693 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2694 blocksize = map->stripe_len;
2695 else
2696 blocksize = sctx->fs_info->sectorsize;
2697 spin_lock(&sctx->stat_lock);
2698 sctx->stat.data_extents_scrubbed++;
2699 sctx->stat.data_bytes_scrubbed += len;
2700 spin_unlock(&sctx->stat_lock);
2701 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2702 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2703 blocksize = map->stripe_len;
2704 else
2705 blocksize = sctx->fs_info->nodesize;
2706 spin_lock(&sctx->stat_lock);
2707 sctx->stat.tree_extents_scrubbed++;
2708 sctx->stat.tree_bytes_scrubbed += len;
2709 spin_unlock(&sctx->stat_lock);
2710 } else {
2711 blocksize = sctx->fs_info->sectorsize;
2712 WARN_ON(1);
2713 }
2714
2715 /*
2716 * For dev-replace case, we can have @dev being a missing device.
2717 * Regular scrub will avoid its execution on missing device at all,
2718 * as that would trigger tons of read error.
2719 *
2720 * Reading from missing device will cause read error counts to
2721 * increase unnecessarily.
2722 * So here we change the read source to a good mirror.
2723 */
2724 if (sctx->is_dev_replace && !dev->bdev)
2725 scrub_find_good_copy(sctx->fs_info, logical, len, &src_physical,
2726 &src_dev, &src_mirror);
2727 while (len) {
2728 u32 l = min(len, blocksize);
2729 int have_csum = 0;
2730
2731 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2732 /* push csums to sbio */
2733 have_csum = scrub_find_csum(sctx, logical, csum);
2734 if (have_csum == 0)
2735 ++sctx->stat.no_csum;
2736 }
2737 ret = scrub_sectors(sctx, logical, l, src_physical, src_dev,
2738 flags, gen, src_mirror,
2739 have_csum ? csum : NULL, physical);
2740 if (ret)
2741 return ret;
2742 len -= l;
2743 logical += l;
2744 physical += l;
2745 src_physical += l;
2746 }
2747 return 0;
2748}
2749
2750static int scrub_sectors_for_parity(struct scrub_parity *sparity,
2751 u64 logical, u32 len,
2752 u64 physical, struct btrfs_device *dev,
2753 u64 flags, u64 gen, int mirror_num, u8 *csum)
2754{
2755 struct scrub_ctx *sctx = sparity->sctx;
2756 struct scrub_block *sblock;
2757 const u32 sectorsize = sctx->fs_info->sectorsize;
2758 int index;
2759
2760 ASSERT(IS_ALIGNED(len, sectorsize));
2761
2762 sblock = alloc_scrub_block(sctx, dev, logical, physical, physical, mirror_num);
2763 if (!sblock) {
2764 spin_lock(&sctx->stat_lock);
2765 sctx->stat.malloc_errors++;
2766 spin_unlock(&sctx->stat_lock);
2767 return -ENOMEM;
2768 }
2769
2770 sblock->sparity = sparity;
2771 scrub_parity_get(sparity);
2772
2773 for (index = 0; len > 0; index++) {
2774 struct scrub_sector *sector;
2775
2776 sector = alloc_scrub_sector(sblock, logical);
2777 if (!sector) {
2778 spin_lock(&sctx->stat_lock);
2779 sctx->stat.malloc_errors++;
2780 spin_unlock(&sctx->stat_lock);
2781 scrub_block_put(sblock);
2782 return -ENOMEM;
2783 }
2784 sblock->sectors[index] = sector;
2785 /* For scrub parity */
2786 scrub_sector_get(sector);
2787 list_add_tail(§or->list, &sparity->sectors_list);
2788 sector->flags = flags;
2789 sector->generation = gen;
2790 if (csum) {
2791 sector->have_csum = 1;
2792 memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2793 } else {
2794 sector->have_csum = 0;
2795 }
2796
2797 /* Iterate over the stripe range in sectorsize steps */
2798 len -= sectorsize;
2799 logical += sectorsize;
2800 physical += sectorsize;
2801 }
2802
2803 WARN_ON(sblock->sector_count == 0);
2804 for (index = 0; index < sblock->sector_count; index++) {
2805 struct scrub_sector *sector = sblock->sectors[index];
2806 int ret;
2807
2808 ret = scrub_add_sector_to_rd_bio(sctx, sector);
2809 if (ret) {
2810 scrub_block_put(sblock);
2811 return ret;
2812 }
2813 }
2814
2815 /* Last one frees, either here or in bio completion for last sector */
2816 scrub_block_put(sblock);
2817 return 0;
2818}
2819
2820static int scrub_extent_for_parity(struct scrub_parity *sparity,
2821 u64 logical, u32 len,
2822 u64 physical, struct btrfs_device *dev,
2823 u64 flags, u64 gen, int mirror_num)
2824{
2825 struct scrub_ctx *sctx = sparity->sctx;
2826 int ret;
2827 u8 csum[BTRFS_CSUM_SIZE];
2828 u32 blocksize;
2829
2830 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2831 scrub_parity_mark_sectors_error(sparity, logical, len);
2832 return 0;
2833 }
2834
2835 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2836 blocksize = sparity->stripe_len;
2837 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2838 blocksize = sparity->stripe_len;
2839 } else {
2840 blocksize = sctx->fs_info->sectorsize;
2841 WARN_ON(1);
2842 }
2843
2844 while (len) {
2845 u32 l = min(len, blocksize);
2846 int have_csum = 0;
2847
2848 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2849 /* push csums to sbio */
2850 have_csum = scrub_find_csum(sctx, logical, csum);
2851 if (have_csum == 0)
2852 goto skip;
2853 }
2854 ret = scrub_sectors_for_parity(sparity, logical, l, physical, dev,
2855 flags, gen, mirror_num,
2856 have_csum ? csum : NULL);
2857 if (ret)
2858 return ret;
2859skip:
2860 len -= l;
2861 logical += l;
2862 physical += l;
2863 }
2864 return 0;
2865}
2866
2867/*
2868 * Given a physical address, this will calculate it's
2869 * logical offset. if this is a parity stripe, it will return
2870 * the most left data stripe's logical offset.
2871 *
2872 * return 0 if it is a data stripe, 1 means parity stripe.
2873 */
2874static int get_raid56_logic_offset(u64 physical, int num,
2875 struct map_lookup *map, u64 *offset,
2876 u64 *stripe_start)
2877{
2878 int i;
2879 int j = 0;
2880 u64 stripe_nr;
2881 u64 last_offset;
2882 u32 stripe_index;
2883 u32 rot;
2884 const int data_stripes = nr_data_stripes(map);
2885
2886 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2887 if (stripe_start)
2888 *stripe_start = last_offset;
2889
2890 *offset = last_offset;
2891 for (i = 0; i < data_stripes; i++) {
2892 *offset = last_offset + i * map->stripe_len;
2893
2894 stripe_nr = div64_u64(*offset, map->stripe_len);
2895 stripe_nr = div_u64(stripe_nr, data_stripes);
2896
2897 /* Work out the disk rotation on this stripe-set */
2898 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2899 /* calculate which stripe this data locates */
2900 rot += i;
2901 stripe_index = rot % map->num_stripes;
2902 if (stripe_index == num)
2903 return 0;
2904 if (stripe_index < num)
2905 j++;
2906 }
2907 *offset = last_offset + j * map->stripe_len;
2908 return 1;
2909}
2910
2911static void scrub_free_parity(struct scrub_parity *sparity)
2912{
2913 struct scrub_ctx *sctx = sparity->sctx;
2914 struct scrub_sector *curr, *next;
2915 int nbits;
2916
2917 nbits = bitmap_weight(&sparity->ebitmap, sparity->nsectors);
2918 if (nbits) {
2919 spin_lock(&sctx->stat_lock);
2920 sctx->stat.read_errors += nbits;
2921 sctx->stat.uncorrectable_errors += nbits;
2922 spin_unlock(&sctx->stat_lock);
2923 }
2924
2925 list_for_each_entry_safe(curr, next, &sparity->sectors_list, list) {
2926 list_del_init(&curr->list);
2927 scrub_sector_put(curr);
2928 }
2929
2930 kfree(sparity);
2931}
2932
2933static void scrub_parity_bio_endio_worker(struct work_struct *work)
2934{
2935 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2936 work);
2937 struct scrub_ctx *sctx = sparity->sctx;
2938
2939 btrfs_bio_counter_dec(sctx->fs_info);
2940 scrub_free_parity(sparity);
2941 scrub_pending_bio_dec(sctx);
2942}
2943
2944static void scrub_parity_bio_endio(struct bio *bio)
2945{
2946 struct scrub_parity *sparity = bio->bi_private;
2947 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2948
2949 if (bio->bi_status)
2950 bitmap_or(&sparity->ebitmap, &sparity->ebitmap,
2951 &sparity->dbitmap, sparity->nsectors);
2952
2953 bio_put(bio);
2954
2955 INIT_WORK(&sparity->work, scrub_parity_bio_endio_worker);
2956 queue_work(fs_info->scrub_parity_workers, &sparity->work);
2957}
2958
2959static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2960{
2961 struct scrub_ctx *sctx = sparity->sctx;
2962 struct btrfs_fs_info *fs_info = sctx->fs_info;
2963 struct bio *bio;
2964 struct btrfs_raid_bio *rbio;
2965 struct btrfs_io_context *bioc = NULL;
2966 u64 length;
2967 int ret;
2968
2969 if (!bitmap_andnot(&sparity->dbitmap, &sparity->dbitmap,
2970 &sparity->ebitmap, sparity->nsectors))
2971 goto out;
2972
2973 length = sparity->logic_end - sparity->logic_start;
2974
2975 btrfs_bio_counter_inc_blocked(fs_info);
2976 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2977 &length, &bioc);
2978 if (ret || !bioc || !bioc->raid_map)
2979 goto bioc_out;
2980
2981 bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2982 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2983 bio->bi_private = sparity;
2984 bio->bi_end_io = scrub_parity_bio_endio;
2985
2986 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc,
2987 sparity->scrub_dev,
2988 &sparity->dbitmap,
2989 sparity->nsectors);
2990 btrfs_put_bioc(bioc);
2991 if (!rbio)
2992 goto rbio_out;
2993
2994 scrub_pending_bio_inc(sctx);
2995 raid56_parity_submit_scrub_rbio(rbio);
2996 return;
2997
2998rbio_out:
2999 bio_put(bio);
3000bioc_out:
3001 btrfs_bio_counter_dec(fs_info);
3002 bitmap_or(&sparity->ebitmap, &sparity->ebitmap, &sparity->dbitmap,
3003 sparity->nsectors);
3004 spin_lock(&sctx->stat_lock);
3005 sctx->stat.malloc_errors++;
3006 spin_unlock(&sctx->stat_lock);
3007out:
3008 scrub_free_parity(sparity);
3009}
3010
3011static void scrub_parity_get(struct scrub_parity *sparity)
3012{
3013 refcount_inc(&sparity->refs);
3014}
3015
3016static void scrub_parity_put(struct scrub_parity *sparity)
3017{
3018 if (!refcount_dec_and_test(&sparity->refs))
3019 return;
3020
3021 scrub_parity_check_and_repair(sparity);
3022}
3023
3024/*
3025 * Return 0 if the extent item range covers any byte of the range.
3026 * Return <0 if the extent item is before @search_start.
3027 * Return >0 if the extent item is after @start_start + @search_len.
3028 */
3029static int compare_extent_item_range(struct btrfs_path *path,
3030 u64 search_start, u64 search_len)
3031{
3032 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
3033 u64 len;
3034 struct btrfs_key key;
3035
3036 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3037 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
3038 key.type == BTRFS_METADATA_ITEM_KEY);
3039 if (key.type == BTRFS_METADATA_ITEM_KEY)
3040 len = fs_info->nodesize;
3041 else
3042 len = key.offset;
3043
3044 if (key.objectid + len <= search_start)
3045 return -1;
3046 if (key.objectid >= search_start + search_len)
3047 return 1;
3048 return 0;
3049}
3050
3051/*
3052 * Locate one extent item which covers any byte in range
3053 * [@search_start, @search_start + @search_length)
3054 *
3055 * If the path is not initialized, we will initialize the search by doing
3056 * a btrfs_search_slot().
3057 * If the path is already initialized, we will use the path as the initial
3058 * slot, to avoid duplicated btrfs_search_slot() calls.
3059 *
3060 * NOTE: If an extent item starts before @search_start, we will still
3061 * return the extent item. This is for data extent crossing stripe boundary.
3062 *
3063 * Return 0 if we found such extent item, and @path will point to the extent item.
3064 * Return >0 if no such extent item can be found, and @path will be released.
3065 * Return <0 if hit fatal error, and @path will be released.
3066 */
3067static int find_first_extent_item(struct btrfs_root *extent_root,
3068 struct btrfs_path *path,
3069 u64 search_start, u64 search_len)
3070{
3071 struct btrfs_fs_info *fs_info = extent_root->fs_info;
3072 struct btrfs_key key;
3073 int ret;
3074
3075 /* Continue using the existing path */
3076 if (path->nodes[0])
3077 goto search_forward;
3078
3079 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3080 key.type = BTRFS_METADATA_ITEM_KEY;
3081 else
3082 key.type = BTRFS_EXTENT_ITEM_KEY;
3083 key.objectid = search_start;
3084 key.offset = (u64)-1;
3085
3086 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3087 if (ret < 0)
3088 return ret;
3089
3090 ASSERT(ret > 0);
3091 /*
3092 * Here we intentionally pass 0 as @min_objectid, as there could be
3093 * an extent item starting before @search_start.
3094 */
3095 ret = btrfs_previous_extent_item(extent_root, path, 0);
3096 if (ret < 0)
3097 return ret;
3098 /*
3099 * No matter whether we have found an extent item, the next loop will
3100 * properly do every check on the key.
3101 */
3102search_forward:
3103 while (true) {
3104 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3105 if (key.objectid >= search_start + search_len)
3106 break;
3107 if (key.type != BTRFS_METADATA_ITEM_KEY &&
3108 key.type != BTRFS_EXTENT_ITEM_KEY)
3109 goto next;
3110
3111 ret = compare_extent_item_range(path, search_start, search_len);
3112 if (ret == 0)
3113 return ret;
3114 if (ret > 0)
3115 break;
3116next:
3117 path->slots[0]++;
3118 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
3119 ret = btrfs_next_leaf(extent_root, path);
3120 if (ret) {
3121 /* Either no more item or fatal error */
3122 btrfs_release_path(path);
3123 return ret;
3124 }
3125 }
3126 }
3127 btrfs_release_path(path);
3128 return 1;
3129}
3130
3131static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
3132 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
3133{
3134 struct btrfs_key key;
3135 struct btrfs_extent_item *ei;
3136
3137 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3138 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
3139 key.type == BTRFS_EXTENT_ITEM_KEY);
3140 *extent_start_ret = key.objectid;
3141 if (key.type == BTRFS_METADATA_ITEM_KEY)
3142 *size_ret = path->nodes[0]->fs_info->nodesize;
3143 else
3144 *size_ret = key.offset;
3145 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
3146 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
3147 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
3148}
3149
3150static bool does_range_cross_boundary(u64 extent_start, u64 extent_len,
3151 u64 boundary_start, u64 boudary_len)
3152{
3153 return (extent_start < boundary_start &&
3154 extent_start + extent_len > boundary_start) ||
3155 (extent_start < boundary_start + boudary_len &&
3156 extent_start + extent_len > boundary_start + boudary_len);
3157}
3158
3159static int scrub_raid56_data_stripe_for_parity(struct scrub_ctx *sctx,
3160 struct scrub_parity *sparity,
3161 struct map_lookup *map,
3162 struct btrfs_device *sdev,
3163 struct btrfs_path *path,
3164 u64 logical)
3165{
3166 struct btrfs_fs_info *fs_info = sctx->fs_info;
3167 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
3168 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, logical);
3169 u64 cur_logical = logical;
3170 int ret;
3171
3172 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3173
3174 /* Path must not be populated */
3175 ASSERT(!path->nodes[0]);
3176
3177 while (cur_logical < logical + map->stripe_len) {
3178 struct btrfs_io_context *bioc = NULL;
3179 struct btrfs_device *extent_dev;
3180 u64 extent_start;
3181 u64 extent_size;
3182 u64 mapped_length;
3183 u64 extent_flags;
3184 u64 extent_gen;
3185 u64 extent_physical;
3186 u64 extent_mirror_num;
3187
3188 ret = find_first_extent_item(extent_root, path, cur_logical,
3189 logical + map->stripe_len - cur_logical);
3190 /* No more extent item in this data stripe */
3191 if (ret > 0) {
3192 ret = 0;
3193 break;
3194 }
3195 if (ret < 0)
3196 break;
3197 get_extent_info(path, &extent_start, &extent_size, &extent_flags,
3198 &extent_gen);
3199
3200 /* Metadata should not cross stripe boundaries */
3201 if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3202 does_range_cross_boundary(extent_start, extent_size,
3203 logical, map->stripe_len)) {
3204 btrfs_err(fs_info,
3205 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3206 extent_start, logical);
3207 spin_lock(&sctx->stat_lock);
3208 sctx->stat.uncorrectable_errors++;
3209 spin_unlock(&sctx->stat_lock);
3210 cur_logical += extent_size;
3211 continue;
3212 }
3213
3214 /* Skip hole range which doesn't have any extent */
3215 cur_logical = max(extent_start, cur_logical);
3216
3217 /* Truncate the range inside this data stripe */
3218 extent_size = min(extent_start + extent_size,
3219 logical + map->stripe_len) - cur_logical;
3220 extent_start = cur_logical;
3221 ASSERT(extent_size <= U32_MAX);
3222
3223 scrub_parity_mark_sectors_data(sparity, extent_start, extent_size);
3224
3225 mapped_length = extent_size;
3226 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_start,
3227 &mapped_length, &bioc, 0);
3228 if (!ret && (!bioc || mapped_length < extent_size))
3229 ret = -EIO;
3230 if (ret) {
3231 btrfs_put_bioc(bioc);
3232 scrub_parity_mark_sectors_error(sparity, extent_start,
3233 extent_size);
3234 break;
3235 }
3236 extent_physical = bioc->stripes[0].physical;
3237 extent_mirror_num = bioc->mirror_num;
3238 extent_dev = bioc->stripes[0].dev;
3239 btrfs_put_bioc(bioc);
3240
3241 ret = btrfs_lookup_csums_list(csum_root, extent_start,
3242 extent_start + extent_size - 1,
3243 &sctx->csum_list, 1, false);
3244 if (ret) {
3245 scrub_parity_mark_sectors_error(sparity, extent_start,
3246 extent_size);
3247 break;
3248 }
3249
3250 ret = scrub_extent_for_parity(sparity, extent_start,
3251 extent_size, extent_physical,
3252 extent_dev, extent_flags,
3253 extent_gen, extent_mirror_num);
3254 scrub_free_csums(sctx);
3255
3256 if (ret) {
3257 scrub_parity_mark_sectors_error(sparity, extent_start,
3258 extent_size);
3259 break;
3260 }
3261
3262 cond_resched();
3263 cur_logical += extent_size;
3264 }
3265 btrfs_release_path(path);
3266 return ret;
3267}
3268
3269static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3270 struct map_lookup *map,
3271 struct btrfs_device *sdev,
3272 u64 logic_start,
3273 u64 logic_end)
3274{
3275 struct btrfs_fs_info *fs_info = sctx->fs_info;
3276 struct btrfs_path *path;
3277 u64 cur_logical;
3278 int ret;
3279 struct scrub_parity *sparity;
3280 int nsectors;
3281
3282 path = btrfs_alloc_path();
3283 if (!path) {
3284 spin_lock(&sctx->stat_lock);
3285 sctx->stat.malloc_errors++;
3286 spin_unlock(&sctx->stat_lock);
3287 return -ENOMEM;
3288 }
3289 path->search_commit_root = 1;
3290 path->skip_locking = 1;
3291
3292 ASSERT(map->stripe_len <= U32_MAX);
3293 nsectors = map->stripe_len >> fs_info->sectorsize_bits;
3294 ASSERT(nsectors <= BITS_PER_LONG);
3295 sparity = kzalloc(sizeof(struct scrub_parity), GFP_NOFS);
3296 if (!sparity) {
3297 spin_lock(&sctx->stat_lock);
3298 sctx->stat.malloc_errors++;
3299 spin_unlock(&sctx->stat_lock);
3300 btrfs_free_path(path);
3301 return -ENOMEM;
3302 }
3303
3304 ASSERT(map->stripe_len <= U32_MAX);
3305 sparity->stripe_len = map->stripe_len;
3306 sparity->nsectors = nsectors;
3307 sparity->sctx = sctx;
3308 sparity->scrub_dev = sdev;
3309 sparity->logic_start = logic_start;
3310 sparity->logic_end = logic_end;
3311 refcount_set(&sparity->refs, 1);
3312 INIT_LIST_HEAD(&sparity->sectors_list);
3313
3314 ret = 0;
3315 for (cur_logical = logic_start; cur_logical < logic_end;
3316 cur_logical += map->stripe_len) {
3317 ret = scrub_raid56_data_stripe_for_parity(sctx, sparity, map,
3318 sdev, path, cur_logical);
3319 if (ret < 0)
3320 break;
3321 }
3322
3323 scrub_parity_put(sparity);
3324 scrub_submit(sctx);
3325 mutex_lock(&sctx->wr_lock);
3326 scrub_wr_submit(sctx);
3327 mutex_unlock(&sctx->wr_lock);
3328
3329 btrfs_free_path(path);
3330 return ret < 0 ? ret : 0;
3331}
3332
3333static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3334{
3335 if (!btrfs_is_zoned(sctx->fs_info))
3336 return;
3337
3338 sctx->flush_all_writes = true;
3339 scrub_submit(sctx);
3340 mutex_lock(&sctx->wr_lock);
3341 scrub_wr_submit(sctx);
3342 mutex_unlock(&sctx->wr_lock);
3343
3344 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3345}
3346
3347static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3348 u64 physical, u64 physical_end)
3349{
3350 struct btrfs_fs_info *fs_info = sctx->fs_info;
3351 int ret = 0;
3352
3353 if (!btrfs_is_zoned(fs_info))
3354 return 0;
3355
3356 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3357
3358 mutex_lock(&sctx->wr_lock);
3359 if (sctx->write_pointer < physical_end) {
3360 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3361 physical,
3362 sctx->write_pointer);
3363 if (ret)
3364 btrfs_err(fs_info,
3365 "zoned: failed to recover write pointer");
3366 }
3367 mutex_unlock(&sctx->wr_lock);
3368 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3369
3370 return ret;
3371}
3372
3373/*
3374 * Scrub one range which can only has simple mirror based profile.
3375 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
3376 * RAID0/RAID10).
3377 *
3378 * Since we may need to handle a subset of block group, we need @logical_start
3379 * and @logical_length parameter.
3380 */
3381static int scrub_simple_mirror(struct scrub_ctx *sctx,
3382 struct btrfs_root *extent_root,
3383 struct btrfs_root *csum_root,
3384 struct btrfs_block_group *bg,
3385 struct map_lookup *map,
3386 u64 logical_start, u64 logical_length,
3387 struct btrfs_device *device,
3388 u64 physical, int mirror_num)
3389{
3390 struct btrfs_fs_info *fs_info = sctx->fs_info;
3391 const u64 logical_end = logical_start + logical_length;
3392 /* An artificial limit, inherit from old scrub behavior */
3393 const u32 max_length = SZ_64K;
3394 struct btrfs_path path = { 0 };
3395 u64 cur_logical = logical_start;
3396 int ret;
3397
3398 /* The range must be inside the bg */
3399 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
3400
3401 path.search_commit_root = 1;
3402 path.skip_locking = 1;
3403 /* Go through each extent items inside the logical range */
3404 while (cur_logical < logical_end) {
3405 u64 extent_start;
3406 u64 extent_len;
3407 u64 extent_flags;
3408 u64 extent_gen;
3409 u64 scrub_len;
3410
3411 /* Canceled? */
3412 if (atomic_read(&fs_info->scrub_cancel_req) ||
3413 atomic_read(&sctx->cancel_req)) {
3414 ret = -ECANCELED;
3415 break;
3416 }
3417 /* Paused? */
3418 if (atomic_read(&fs_info->scrub_pause_req)) {
3419 /* Push queued extents */
3420 sctx->flush_all_writes = true;
3421 scrub_submit(sctx);
3422 mutex_lock(&sctx->wr_lock);
3423 scrub_wr_submit(sctx);
3424 mutex_unlock(&sctx->wr_lock);
3425 wait_event(sctx->list_wait,
3426 atomic_read(&sctx->bios_in_flight) == 0);
3427 sctx->flush_all_writes = false;
3428 scrub_blocked_if_needed(fs_info);
3429 }
3430 /* Block group removed? */
3431 spin_lock(&bg->lock);
3432 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
3433 spin_unlock(&bg->lock);
3434 ret = 0;
3435 break;
3436 }
3437 spin_unlock(&bg->lock);
3438
3439 ret = find_first_extent_item(extent_root, &path, cur_logical,
3440 logical_end - cur_logical);
3441 if (ret > 0) {
3442 /* No more extent, just update the accounting */
3443 sctx->stat.last_physical = physical + logical_length;
3444 ret = 0;
3445 break;
3446 }
3447 if (ret < 0)
3448 break;
3449 get_extent_info(&path, &extent_start, &extent_len,
3450 &extent_flags, &extent_gen);
3451 /* Skip hole range which doesn't have any extent */
3452 cur_logical = max(extent_start, cur_logical);
3453
3454 /*
3455 * Scrub len has three limits:
3456 * - Extent size limit
3457 * - Scrub range limit
3458 * This is especially imporatant for RAID0/RAID10 to reuse
3459 * this function
3460 * - Max scrub size limit
3461 */
3462 scrub_len = min(min(extent_start + extent_len,
3463 logical_end), cur_logical + max_length) -
3464 cur_logical;
3465
3466 if (extent_flags & BTRFS_EXTENT_FLAG_DATA) {
3467 ret = btrfs_lookup_csums_list(csum_root, cur_logical,
3468 cur_logical + scrub_len - 1,
3469 &sctx->csum_list, 1, false);
3470 if (ret)
3471 break;
3472 }
3473 if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3474 does_range_cross_boundary(extent_start, extent_len,
3475 logical_start, logical_length)) {
3476 btrfs_err(fs_info,
3477"scrub: tree block %llu spanning boundaries, ignored. boundary=[%llu, %llu)",
3478 extent_start, logical_start, logical_end);
3479 spin_lock(&sctx->stat_lock);
3480 sctx->stat.uncorrectable_errors++;
3481 spin_unlock(&sctx->stat_lock);
3482 cur_logical += scrub_len;
3483 continue;
3484 }
3485 ret = scrub_extent(sctx, map, cur_logical, scrub_len,
3486 cur_logical - logical_start + physical,
3487 device, extent_flags, extent_gen,
3488 mirror_num);
3489 scrub_free_csums(sctx);
3490 if (ret)
3491 break;
3492 if (sctx->is_dev_replace)
3493 sync_replace_for_zoned(sctx);
3494 cur_logical += scrub_len;
3495 /* Don't hold CPU for too long time */
3496 cond_resched();
3497 }
3498 btrfs_release_path(&path);
3499 return ret;
3500}
3501
3502/* Calculate the full stripe length for simple stripe based profiles */
3503static u64 simple_stripe_full_stripe_len(const struct map_lookup *map)
3504{
3505 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3506 BTRFS_BLOCK_GROUP_RAID10));
3507
3508 return map->num_stripes / map->sub_stripes * map->stripe_len;
3509}
3510
3511/* Get the logical bytenr for the stripe */
3512static u64 simple_stripe_get_logical(struct map_lookup *map,
3513 struct btrfs_block_group *bg,
3514 int stripe_index)
3515{
3516 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3517 BTRFS_BLOCK_GROUP_RAID10));
3518 ASSERT(stripe_index < map->num_stripes);
3519
3520 /*
3521 * (stripe_index / sub_stripes) gives how many data stripes we need to
3522 * skip.
3523 */
3524 return (stripe_index / map->sub_stripes) * map->stripe_len + bg->start;
3525}
3526
3527/* Get the mirror number for the stripe */
3528static int simple_stripe_mirror_num(struct map_lookup *map, int stripe_index)
3529{
3530 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3531 BTRFS_BLOCK_GROUP_RAID10));
3532 ASSERT(stripe_index < map->num_stripes);
3533
3534 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
3535 return stripe_index % map->sub_stripes + 1;
3536}
3537
3538static int scrub_simple_stripe(struct scrub_ctx *sctx,
3539 struct btrfs_root *extent_root,
3540 struct btrfs_root *csum_root,
3541 struct btrfs_block_group *bg,
3542 struct map_lookup *map,
3543 struct btrfs_device *device,
3544 int stripe_index)
3545{
3546 const u64 logical_increment = simple_stripe_full_stripe_len(map);
3547 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
3548 const u64 orig_physical = map->stripes[stripe_index].physical;
3549 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
3550 u64 cur_logical = orig_logical;
3551 u64 cur_physical = orig_physical;
3552 int ret = 0;
3553
3554 while (cur_logical < bg->start + bg->length) {
3555 /*
3556 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
3557 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
3558 * this stripe.
3559 */
3560 ret = scrub_simple_mirror(sctx, extent_root, csum_root, bg, map,
3561 cur_logical, map->stripe_len, device,
3562 cur_physical, mirror_num);
3563 if (ret)
3564 return ret;
3565 /* Skip to next stripe which belongs to the target device */
3566 cur_logical += logical_increment;
3567 /* For physical offset, we just go to next stripe */
3568 cur_physical += map->stripe_len;
3569 }
3570 return ret;
3571}
3572
3573static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3574 struct btrfs_block_group *bg,
3575 struct extent_map *em,
3576 struct btrfs_device *scrub_dev,
3577 int stripe_index)
3578{
3579 struct btrfs_path *path;
3580 struct btrfs_fs_info *fs_info = sctx->fs_info;
3581 struct btrfs_root *root;
3582 struct btrfs_root *csum_root;
3583 struct blk_plug plug;
3584 struct map_lookup *map = em->map_lookup;
3585 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3586 const u64 chunk_logical = bg->start;
3587 int ret;
3588 u64 physical = map->stripes[stripe_index].physical;
3589 const u64 dev_stripe_len = btrfs_calc_stripe_length(em);
3590 const u64 physical_end = physical + dev_stripe_len;
3591 u64 logical;
3592 u64 logic_end;
3593 /* The logical increment after finishing one stripe */
3594 u64 increment;
3595 /* Offset inside the chunk */
3596 u64 offset;
3597 u64 stripe_logical;
3598 u64 stripe_end;
3599 int stop_loop = 0;
3600
3601 path = btrfs_alloc_path();
3602 if (!path)
3603 return -ENOMEM;
3604
3605 /*
3606 * work on commit root. The related disk blocks are static as
3607 * long as COW is applied. This means, it is save to rewrite
3608 * them to repair disk errors without any race conditions
3609 */
3610 path->search_commit_root = 1;
3611 path->skip_locking = 1;
3612 path->reada = READA_FORWARD;
3613
3614 wait_event(sctx->list_wait,
3615 atomic_read(&sctx->bios_in_flight) == 0);
3616 scrub_blocked_if_needed(fs_info);
3617
3618 root = btrfs_extent_root(fs_info, bg->start);
3619 csum_root = btrfs_csum_root(fs_info, bg->start);
3620
3621 /*
3622 * collect all data csums for the stripe to avoid seeking during
3623 * the scrub. This might currently (crc32) end up to be about 1MB
3624 */
3625 blk_start_plug(&plug);
3626
3627 if (sctx->is_dev_replace &&
3628 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3629 mutex_lock(&sctx->wr_lock);
3630 sctx->write_pointer = physical;
3631 mutex_unlock(&sctx->wr_lock);
3632 sctx->flush_all_writes = true;
3633 }
3634
3635 /*
3636 * There used to be a big double loop to handle all profiles using the
3637 * same routine, which grows larger and more gross over time.
3638 *
3639 * So here we handle each profile differently, so simpler profiles
3640 * have simpler scrubbing function.
3641 */
3642 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
3643 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
3644 /*
3645 * Above check rules out all complex profile, the remaining
3646 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
3647 * mirrored duplication without stripe.
3648 *
3649 * Only @physical and @mirror_num needs to calculated using
3650 * @stripe_index.
3651 */
3652 ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3653 bg->start, bg->length, scrub_dev,
3654 map->stripes[stripe_index].physical,
3655 stripe_index + 1);
3656 offset = 0;
3657 goto out;
3658 }
3659 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
3660 ret = scrub_simple_stripe(sctx, root, csum_root, bg, map,
3661 scrub_dev, stripe_index);
3662 offset = map->stripe_len * (stripe_index / map->sub_stripes);
3663 goto out;
3664 }
3665
3666 /* Only RAID56 goes through the old code */
3667 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3668 ret = 0;
3669
3670 /* Calculate the logical end of the stripe */
3671 get_raid56_logic_offset(physical_end, stripe_index,
3672 map, &logic_end, NULL);
3673 logic_end += chunk_logical;
3674
3675 /* Initialize @offset in case we need to go to out: label */
3676 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
3677 increment = map->stripe_len * nr_data_stripes(map);
3678
3679 /*
3680 * Due to the rotation, for RAID56 it's better to iterate each stripe
3681 * using their physical offset.
3682 */
3683 while (physical < physical_end) {
3684 ret = get_raid56_logic_offset(physical, stripe_index, map,
3685 &logical, &stripe_logical);
3686 logical += chunk_logical;
3687 if (ret) {
3688 /* it is parity strip */
3689 stripe_logical += chunk_logical;
3690 stripe_end = stripe_logical + increment;
3691 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3692 stripe_logical,
3693 stripe_end);
3694 if (ret)
3695 goto out;
3696 goto next;
3697 }
3698
3699 /*
3700 * Now we're at a data stripe, scrub each extents in the range.
3701 *
3702 * At this stage, if we ignore the repair part, inside each data
3703 * stripe it is no different than SINGLE profile.
3704 * We can reuse scrub_simple_mirror() here, as the repair part
3705 * is still based on @mirror_num.
3706 */
3707 ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3708 logical, map->stripe_len,
3709 scrub_dev, physical, 1);
3710 if (ret < 0)
3711 goto out;
3712next:
3713 logical += increment;
3714 physical += map->stripe_len;
3715 spin_lock(&sctx->stat_lock);
3716 if (stop_loop)
3717 sctx->stat.last_physical =
3718 map->stripes[stripe_index].physical + dev_stripe_len;
3719 else
3720 sctx->stat.last_physical = physical;
3721 spin_unlock(&sctx->stat_lock);
3722 if (stop_loop)
3723 break;
3724 }
3725out:
3726 /* push queued extents */
3727 scrub_submit(sctx);
3728 mutex_lock(&sctx->wr_lock);
3729 scrub_wr_submit(sctx);
3730 mutex_unlock(&sctx->wr_lock);
3731
3732 blk_finish_plug(&plug);
3733 btrfs_free_path(path);
3734
3735 if (sctx->is_dev_replace && ret >= 0) {
3736 int ret2;
3737
3738 ret2 = sync_write_pointer_for_zoned(sctx,
3739 chunk_logical + offset,
3740 map->stripes[stripe_index].physical,
3741 physical_end);
3742 if (ret2)
3743 ret = ret2;
3744 }
3745
3746 return ret < 0 ? ret : 0;
3747}
3748
3749static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3750 struct btrfs_block_group *bg,
3751 struct btrfs_device *scrub_dev,
3752 u64 dev_offset,
3753 u64 dev_extent_len)
3754{
3755 struct btrfs_fs_info *fs_info = sctx->fs_info;
3756 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3757 struct map_lookup *map;
3758 struct extent_map *em;
3759 int i;
3760 int ret = 0;
3761
3762 read_lock(&map_tree->lock);
3763 em = lookup_extent_mapping(map_tree, bg->start, bg->length);
3764 read_unlock(&map_tree->lock);
3765
3766 if (!em) {
3767 /*
3768 * Might have been an unused block group deleted by the cleaner
3769 * kthread or relocation.
3770 */
3771 spin_lock(&bg->lock);
3772 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
3773 ret = -EINVAL;
3774 spin_unlock(&bg->lock);
3775
3776 return ret;
3777 }
3778 if (em->start != bg->start)
3779 goto out;
3780 if (em->len < dev_extent_len)
3781 goto out;
3782
3783 map = em->map_lookup;
3784 for (i = 0; i < map->num_stripes; ++i) {
3785 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3786 map->stripes[i].physical == dev_offset) {
3787 ret = scrub_stripe(sctx, bg, em, scrub_dev, i);
3788 if (ret)
3789 goto out;
3790 }
3791 }
3792out:
3793 free_extent_map(em);
3794
3795 return ret;
3796}
3797
3798static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3799 struct btrfs_block_group *cache)
3800{
3801 struct btrfs_fs_info *fs_info = cache->fs_info;
3802 struct btrfs_trans_handle *trans;
3803
3804 if (!btrfs_is_zoned(fs_info))
3805 return 0;
3806
3807 btrfs_wait_block_group_reservations(cache);
3808 btrfs_wait_nocow_writers(cache);
3809 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3810
3811 trans = btrfs_join_transaction(root);
3812 if (IS_ERR(trans))
3813 return PTR_ERR(trans);
3814 return btrfs_commit_transaction(trans);
3815}
3816
3817static noinline_for_stack
3818int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3819 struct btrfs_device *scrub_dev, u64 start, u64 end)
3820{
3821 struct btrfs_dev_extent *dev_extent = NULL;
3822 struct btrfs_path *path;
3823 struct btrfs_fs_info *fs_info = sctx->fs_info;
3824 struct btrfs_root *root = fs_info->dev_root;
3825 u64 chunk_offset;
3826 int ret = 0;
3827 int ro_set;
3828 int slot;
3829 struct extent_buffer *l;
3830 struct btrfs_key key;
3831 struct btrfs_key found_key;
3832 struct btrfs_block_group *cache;
3833 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3834
3835 path = btrfs_alloc_path();
3836 if (!path)
3837 return -ENOMEM;
3838
3839 path->reada = READA_FORWARD;
3840 path->search_commit_root = 1;
3841 path->skip_locking = 1;
3842
3843 key.objectid = scrub_dev->devid;
3844 key.offset = 0ull;
3845 key.type = BTRFS_DEV_EXTENT_KEY;
3846
3847 while (1) {
3848 u64 dev_extent_len;
3849
3850 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3851 if (ret < 0)
3852 break;
3853 if (ret > 0) {
3854 if (path->slots[0] >=
3855 btrfs_header_nritems(path->nodes[0])) {
3856 ret = btrfs_next_leaf(root, path);
3857 if (ret < 0)
3858 break;
3859 if (ret > 0) {
3860 ret = 0;
3861 break;
3862 }
3863 } else {
3864 ret = 0;
3865 }
3866 }
3867
3868 l = path->nodes[0];
3869 slot = path->slots[0];
3870
3871 btrfs_item_key_to_cpu(l, &found_key, slot);
3872
3873 if (found_key.objectid != scrub_dev->devid)
3874 break;
3875
3876 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3877 break;
3878
3879 if (found_key.offset >= end)
3880 break;
3881
3882 if (found_key.offset < key.offset)
3883 break;
3884
3885 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3886 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
3887
3888 if (found_key.offset + dev_extent_len <= start)
3889 goto skip;
3890
3891 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3892
3893 /*
3894 * get a reference on the corresponding block group to prevent
3895 * the chunk from going away while we scrub it
3896 */
3897 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3898
3899 /* some chunks are removed but not committed to disk yet,
3900 * continue scrubbing */
3901 if (!cache)
3902 goto skip;
3903
3904 ASSERT(cache->start <= chunk_offset);
3905 /*
3906 * We are using the commit root to search for device extents, so
3907 * that means we could have found a device extent item from a
3908 * block group that was deleted in the current transaction. The
3909 * logical start offset of the deleted block group, stored at
3910 * @chunk_offset, might be part of the logical address range of
3911 * a new block group (which uses different physical extents).
3912 * In this case btrfs_lookup_block_group() has returned the new
3913 * block group, and its start address is less than @chunk_offset.
3914 *
3915 * We skip such new block groups, because it's pointless to
3916 * process them, as we won't find their extents because we search
3917 * for them using the commit root of the extent tree. For a device
3918 * replace it's also fine to skip it, we won't miss copying them
3919 * to the target device because we have the write duplication
3920 * setup through the regular write path (by btrfs_map_block()),
3921 * and we have committed a transaction when we started the device
3922 * replace, right after setting up the device replace state.
3923 */
3924 if (cache->start < chunk_offset) {
3925 btrfs_put_block_group(cache);
3926 goto skip;
3927 }
3928
3929 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3930 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
3931 btrfs_put_block_group(cache);
3932 goto skip;
3933 }
3934 }
3935
3936 /*
3937 * Make sure that while we are scrubbing the corresponding block
3938 * group doesn't get its logical address and its device extents
3939 * reused for another block group, which can possibly be of a
3940 * different type and different profile. We do this to prevent
3941 * false error detections and crashes due to bogus attempts to
3942 * repair extents.
3943 */
3944 spin_lock(&cache->lock);
3945 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
3946 spin_unlock(&cache->lock);
3947 btrfs_put_block_group(cache);
3948 goto skip;
3949 }
3950 btrfs_freeze_block_group(cache);
3951 spin_unlock(&cache->lock);
3952
3953 /*
3954 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3955 * to avoid deadlock caused by:
3956 * btrfs_inc_block_group_ro()
3957 * -> btrfs_wait_for_commit()
3958 * -> btrfs_commit_transaction()
3959 * -> btrfs_scrub_pause()
3960 */
3961 scrub_pause_on(fs_info);
3962
3963 /*
3964 * Don't do chunk preallocation for scrub.
3965 *
3966 * This is especially important for SYSTEM bgs, or we can hit
3967 * -EFBIG from btrfs_finish_chunk_alloc() like:
3968 * 1. The only SYSTEM bg is marked RO.
3969 * Since SYSTEM bg is small, that's pretty common.
3970 * 2. New SYSTEM bg will be allocated
3971 * Due to regular version will allocate new chunk.
3972 * 3. New SYSTEM bg is empty and will get cleaned up
3973 * Before cleanup really happens, it's marked RO again.
3974 * 4. Empty SYSTEM bg get scrubbed
3975 * We go back to 2.
3976 *
3977 * This can easily boost the amount of SYSTEM chunks if cleaner
3978 * thread can't be triggered fast enough, and use up all space
3979 * of btrfs_super_block::sys_chunk_array
3980 *
3981 * While for dev replace, we need to try our best to mark block
3982 * group RO, to prevent race between:
3983 * - Write duplication
3984 * Contains latest data
3985 * - Scrub copy
3986 * Contains data from commit tree
3987 *
3988 * If target block group is not marked RO, nocow writes can
3989 * be overwritten by scrub copy, causing data corruption.
3990 * So for dev-replace, it's not allowed to continue if a block
3991 * group is not RO.
3992 */
3993 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3994 if (!ret && sctx->is_dev_replace) {
3995 ret = finish_extent_writes_for_zoned(root, cache);
3996 if (ret) {
3997 btrfs_dec_block_group_ro(cache);
3998 scrub_pause_off(fs_info);
3999 btrfs_put_block_group(cache);
4000 break;
4001 }
4002 }
4003
4004 if (ret == 0) {
4005 ro_set = 1;
4006 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
4007 /*
4008 * btrfs_inc_block_group_ro return -ENOSPC when it
4009 * failed in creating new chunk for metadata.
4010 * It is not a problem for scrub, because
4011 * metadata are always cowed, and our scrub paused
4012 * commit_transactions.
4013 */
4014 ro_set = 0;
4015 } else if (ret == -ETXTBSY) {
4016 btrfs_warn(fs_info,
4017 "skipping scrub of block group %llu due to active swapfile",
4018 cache->start);
4019 scrub_pause_off(fs_info);
4020 ret = 0;
4021 goto skip_unfreeze;
4022 } else {
4023 btrfs_warn(fs_info,
4024 "failed setting block group ro: %d", ret);
4025 btrfs_unfreeze_block_group(cache);
4026 btrfs_put_block_group(cache);
4027 scrub_pause_off(fs_info);
4028 break;
4029 }
4030
4031 /*
4032 * Now the target block is marked RO, wait for nocow writes to
4033 * finish before dev-replace.
4034 * COW is fine, as COW never overwrites extents in commit tree.
4035 */
4036 if (sctx->is_dev_replace) {
4037 btrfs_wait_nocow_writers(cache);
4038 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
4039 cache->length);
4040 }
4041
4042 scrub_pause_off(fs_info);
4043 down_write(&dev_replace->rwsem);
4044 dev_replace->cursor_right = found_key.offset + dev_extent_len;
4045 dev_replace->cursor_left = found_key.offset;
4046 dev_replace->item_needs_writeback = 1;
4047 up_write(&dev_replace->rwsem);
4048
4049 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
4050 dev_extent_len);
4051
4052 /*
4053 * flush, submit all pending read and write bios, afterwards
4054 * wait for them.
4055 * Note that in the dev replace case, a read request causes
4056 * write requests that are submitted in the read completion
4057 * worker. Therefore in the current situation, it is required
4058 * that all write requests are flushed, so that all read and
4059 * write requests are really completed when bios_in_flight
4060 * changes to 0.
4061 */
4062 sctx->flush_all_writes = true;
4063 scrub_submit(sctx);
4064 mutex_lock(&sctx->wr_lock);
4065 scrub_wr_submit(sctx);
4066 mutex_unlock(&sctx->wr_lock);
4067
4068 wait_event(sctx->list_wait,
4069 atomic_read(&sctx->bios_in_flight) == 0);
4070
4071 scrub_pause_on(fs_info);
4072
4073 /*
4074 * must be called before we decrease @scrub_paused.
4075 * make sure we don't block transaction commit while
4076 * we are waiting pending workers finished.
4077 */
4078 wait_event(sctx->list_wait,
4079 atomic_read(&sctx->workers_pending) == 0);
4080 sctx->flush_all_writes = false;
4081
4082 scrub_pause_off(fs_info);
4083
4084 if (sctx->is_dev_replace &&
4085 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
4086 cache, found_key.offset))
4087 ro_set = 0;
4088
4089 down_write(&dev_replace->rwsem);
4090 dev_replace->cursor_left = dev_replace->cursor_right;
4091 dev_replace->item_needs_writeback = 1;
4092 up_write(&dev_replace->rwsem);
4093
4094 if (ro_set)
4095 btrfs_dec_block_group_ro(cache);
4096
4097 /*
4098 * We might have prevented the cleaner kthread from deleting
4099 * this block group if it was already unused because we raced
4100 * and set it to RO mode first. So add it back to the unused
4101 * list, otherwise it might not ever be deleted unless a manual
4102 * balance is triggered or it becomes used and unused again.
4103 */
4104 spin_lock(&cache->lock);
4105 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
4106 !cache->ro && cache->reserved == 0 && cache->used == 0) {
4107 spin_unlock(&cache->lock);
4108 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
4109 btrfs_discard_queue_work(&fs_info->discard_ctl,
4110 cache);
4111 else
4112 btrfs_mark_bg_unused(cache);
4113 } else {
4114 spin_unlock(&cache->lock);
4115 }
4116skip_unfreeze:
4117 btrfs_unfreeze_block_group(cache);
4118 btrfs_put_block_group(cache);
4119 if (ret)
4120 break;
4121 if (sctx->is_dev_replace &&
4122 atomic64_read(&dev_replace->num_write_errors) > 0) {
4123 ret = -EIO;
4124 break;
4125 }
4126 if (sctx->stat.malloc_errors > 0) {
4127 ret = -ENOMEM;
4128 break;
4129 }
4130skip:
4131 key.offset = found_key.offset + dev_extent_len;
4132 btrfs_release_path(path);
4133 }
4134
4135 btrfs_free_path(path);
4136
4137 return ret;
4138}
4139
4140static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4141 struct btrfs_device *scrub_dev)
4142{
4143 int i;
4144 u64 bytenr;
4145 u64 gen;
4146 int ret;
4147 struct btrfs_fs_info *fs_info = sctx->fs_info;
4148
4149 if (BTRFS_FS_ERROR(fs_info))
4150 return -EROFS;
4151
4152 /* Seed devices of a new filesystem has their own generation. */
4153 if (scrub_dev->fs_devices != fs_info->fs_devices)
4154 gen = scrub_dev->generation;
4155 else
4156 gen = fs_info->last_trans_committed;
4157
4158 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4159 bytenr = btrfs_sb_offset(i);
4160 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4161 scrub_dev->commit_total_bytes)
4162 break;
4163 if (!btrfs_check_super_location(scrub_dev, bytenr))
4164 continue;
4165
4166 ret = scrub_sectors(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4167 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4168 NULL, bytenr);
4169 if (ret)
4170 return ret;
4171 }
4172 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4173
4174 return 0;
4175}
4176
4177static void scrub_workers_put(struct btrfs_fs_info *fs_info)
4178{
4179 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
4180 &fs_info->scrub_lock)) {
4181 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
4182 struct workqueue_struct *scrub_wr_comp =
4183 fs_info->scrub_wr_completion_workers;
4184 struct workqueue_struct *scrub_parity =
4185 fs_info->scrub_parity_workers;
4186
4187 fs_info->scrub_workers = NULL;
4188 fs_info->scrub_wr_completion_workers = NULL;
4189 fs_info->scrub_parity_workers = NULL;
4190 mutex_unlock(&fs_info->scrub_lock);
4191
4192 if (scrub_workers)
4193 destroy_workqueue(scrub_workers);
4194 if (scrub_wr_comp)
4195 destroy_workqueue(scrub_wr_comp);
4196 if (scrub_parity)
4197 destroy_workqueue(scrub_parity);
4198 }
4199}
4200
4201/*
4202 * get a reference count on fs_info->scrub_workers. start worker if necessary
4203 */
4204static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4205 int is_dev_replace)
4206{
4207 struct workqueue_struct *scrub_workers = NULL;
4208 struct workqueue_struct *scrub_wr_comp = NULL;
4209 struct workqueue_struct *scrub_parity = NULL;
4210 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4211 int max_active = fs_info->thread_pool_size;
4212 int ret = -ENOMEM;
4213
4214 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
4215 return 0;
4216
4217 scrub_workers = alloc_workqueue("btrfs-scrub", flags,
4218 is_dev_replace ? 1 : max_active);
4219 if (!scrub_workers)
4220 goto fail_scrub_workers;
4221
4222 scrub_wr_comp = alloc_workqueue("btrfs-scrubwrc", flags, max_active);
4223 if (!scrub_wr_comp)
4224 goto fail_scrub_wr_completion_workers;
4225
4226 scrub_parity = alloc_workqueue("btrfs-scrubparity", flags, max_active);
4227 if (!scrub_parity)
4228 goto fail_scrub_parity_workers;
4229
4230 mutex_lock(&fs_info->scrub_lock);
4231 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
4232 ASSERT(fs_info->scrub_workers == NULL &&
4233 fs_info->scrub_wr_completion_workers == NULL &&
4234 fs_info->scrub_parity_workers == NULL);
4235 fs_info->scrub_workers = scrub_workers;
4236 fs_info->scrub_wr_completion_workers = scrub_wr_comp;
4237 fs_info->scrub_parity_workers = scrub_parity;
4238 refcount_set(&fs_info->scrub_workers_refcnt, 1);
4239 mutex_unlock(&fs_info->scrub_lock);
4240 return 0;
4241 }
4242 /* Other thread raced in and created the workers for us */
4243 refcount_inc(&fs_info->scrub_workers_refcnt);
4244 mutex_unlock(&fs_info->scrub_lock);
4245
4246 ret = 0;
4247 destroy_workqueue(scrub_parity);
4248fail_scrub_parity_workers:
4249 destroy_workqueue(scrub_wr_comp);
4250fail_scrub_wr_completion_workers:
4251 destroy_workqueue(scrub_workers);
4252fail_scrub_workers:
4253 return ret;
4254}
4255
4256int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4257 u64 end, struct btrfs_scrub_progress *progress,
4258 int readonly, int is_dev_replace)
4259{
4260 struct btrfs_dev_lookup_args args = { .devid = devid };
4261 struct scrub_ctx *sctx;
4262 int ret;
4263 struct btrfs_device *dev;
4264 unsigned int nofs_flag;
4265 bool need_commit = false;
4266
4267 if (btrfs_fs_closing(fs_info))
4268 return -EAGAIN;
4269
4270 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
4271 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
4272
4273 /*
4274 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
4275 * value (max nodesize / min sectorsize), thus nodesize should always
4276 * be fine.
4277 */
4278 ASSERT(fs_info->nodesize <=
4279 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
4280
4281 /* Allocate outside of device_list_mutex */
4282 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4283 if (IS_ERR(sctx))
4284 return PTR_ERR(sctx);
4285
4286 ret = scrub_workers_get(fs_info, is_dev_replace);
4287 if (ret)
4288 goto out_free_ctx;
4289
4290 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4291 dev = btrfs_find_device(fs_info->fs_devices, &args);
4292 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4293 !is_dev_replace)) {
4294 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4295 ret = -ENODEV;
4296 goto out;
4297 }
4298
4299 if (!is_dev_replace && !readonly &&
4300 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4301 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4302 btrfs_err_in_rcu(fs_info,
4303 "scrub on devid %llu: filesystem on %s is not writable",
4304 devid, btrfs_dev_name(dev));
4305 ret = -EROFS;
4306 goto out;
4307 }
4308
4309 mutex_lock(&fs_info->scrub_lock);
4310 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4311 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4312 mutex_unlock(&fs_info->scrub_lock);
4313 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4314 ret = -EIO;
4315 goto out;
4316 }
4317
4318 down_read(&fs_info->dev_replace.rwsem);
4319 if (dev->scrub_ctx ||
4320 (!is_dev_replace &&
4321 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4322 up_read(&fs_info->dev_replace.rwsem);
4323 mutex_unlock(&fs_info->scrub_lock);
4324 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4325 ret = -EINPROGRESS;
4326 goto out;
4327 }
4328 up_read(&fs_info->dev_replace.rwsem);
4329
4330 sctx->readonly = readonly;
4331 dev->scrub_ctx = sctx;
4332 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4333
4334 /*
4335 * checking @scrub_pause_req here, we can avoid
4336 * race between committing transaction and scrubbing.
4337 */
4338 __scrub_blocked_if_needed(fs_info);
4339 atomic_inc(&fs_info->scrubs_running);
4340 mutex_unlock(&fs_info->scrub_lock);
4341
4342 /*
4343 * In order to avoid deadlock with reclaim when there is a transaction
4344 * trying to pause scrub, make sure we use GFP_NOFS for all the
4345 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
4346 * invoked by our callees. The pausing request is done when the
4347 * transaction commit starts, and it blocks the transaction until scrub
4348 * is paused (done at specific points at scrub_stripe() or right above
4349 * before incrementing fs_info->scrubs_running).
4350 */
4351 nofs_flag = memalloc_nofs_save();
4352 if (!is_dev_replace) {
4353 u64 old_super_errors;
4354
4355 spin_lock(&sctx->stat_lock);
4356 old_super_errors = sctx->stat.super_errors;
4357 spin_unlock(&sctx->stat_lock);
4358
4359 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4360 /*
4361 * by holding device list mutex, we can
4362 * kick off writing super in log tree sync.
4363 */
4364 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4365 ret = scrub_supers(sctx, dev);
4366 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4367
4368 spin_lock(&sctx->stat_lock);
4369 /*
4370 * Super block errors found, but we can not commit transaction
4371 * at current context, since btrfs_commit_transaction() needs
4372 * to pause the current running scrub (hold by ourselves).
4373 */
4374 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
4375 need_commit = true;
4376 spin_unlock(&sctx->stat_lock);
4377 }
4378
4379 if (!ret)
4380 ret = scrub_enumerate_chunks(sctx, dev, start, end);
4381 memalloc_nofs_restore(nofs_flag);
4382
4383 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4384 atomic_dec(&fs_info->scrubs_running);
4385 wake_up(&fs_info->scrub_pause_wait);
4386
4387 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4388
4389 if (progress)
4390 memcpy(progress, &sctx->stat, sizeof(*progress));
4391
4392 if (!is_dev_replace)
4393 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4394 ret ? "not finished" : "finished", devid, ret);
4395
4396 mutex_lock(&fs_info->scrub_lock);
4397 dev->scrub_ctx = NULL;
4398 mutex_unlock(&fs_info->scrub_lock);
4399
4400 scrub_workers_put(fs_info);
4401 scrub_put_ctx(sctx);
4402
4403 /*
4404 * We found some super block errors before, now try to force a
4405 * transaction commit, as scrub has finished.
4406 */
4407 if (need_commit) {
4408 struct btrfs_trans_handle *trans;
4409
4410 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4411 if (IS_ERR(trans)) {
4412 ret = PTR_ERR(trans);
4413 btrfs_err(fs_info,
4414 "scrub: failed to start transaction to fix super block errors: %d", ret);
4415 return ret;
4416 }
4417 ret = btrfs_commit_transaction(trans);
4418 if (ret < 0)
4419 btrfs_err(fs_info,
4420 "scrub: failed to commit transaction to fix super block errors: %d", ret);
4421 }
4422 return ret;
4423out:
4424 scrub_workers_put(fs_info);
4425out_free_ctx:
4426 scrub_free_ctx(sctx);
4427
4428 return ret;
4429}
4430
4431void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4432{
4433 mutex_lock(&fs_info->scrub_lock);
4434 atomic_inc(&fs_info->scrub_pause_req);
4435 while (atomic_read(&fs_info->scrubs_paused) !=
4436 atomic_read(&fs_info->scrubs_running)) {
4437 mutex_unlock(&fs_info->scrub_lock);
4438 wait_event(fs_info->scrub_pause_wait,
4439 atomic_read(&fs_info->scrubs_paused) ==
4440 atomic_read(&fs_info->scrubs_running));
4441 mutex_lock(&fs_info->scrub_lock);
4442 }
4443 mutex_unlock(&fs_info->scrub_lock);
4444}
4445
4446void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4447{
4448 atomic_dec(&fs_info->scrub_pause_req);
4449 wake_up(&fs_info->scrub_pause_wait);
4450}
4451
4452int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4453{
4454 mutex_lock(&fs_info->scrub_lock);
4455 if (!atomic_read(&fs_info->scrubs_running)) {
4456 mutex_unlock(&fs_info->scrub_lock);
4457 return -ENOTCONN;
4458 }
4459
4460 atomic_inc(&fs_info->scrub_cancel_req);
4461 while (atomic_read(&fs_info->scrubs_running)) {
4462 mutex_unlock(&fs_info->scrub_lock);
4463 wait_event(fs_info->scrub_pause_wait,
4464 atomic_read(&fs_info->scrubs_running) == 0);
4465 mutex_lock(&fs_info->scrub_lock);
4466 }
4467 atomic_dec(&fs_info->scrub_cancel_req);
4468 mutex_unlock(&fs_info->scrub_lock);
4469
4470 return 0;
4471}
4472
4473int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4474{
4475 struct btrfs_fs_info *fs_info = dev->fs_info;
4476 struct scrub_ctx *sctx;
4477
4478 mutex_lock(&fs_info->scrub_lock);
4479 sctx = dev->scrub_ctx;
4480 if (!sctx) {
4481 mutex_unlock(&fs_info->scrub_lock);
4482 return -ENOTCONN;
4483 }
4484 atomic_inc(&sctx->cancel_req);
4485 while (dev->scrub_ctx) {
4486 mutex_unlock(&fs_info->scrub_lock);
4487 wait_event(fs_info->scrub_pause_wait,
4488 dev->scrub_ctx == NULL);
4489 mutex_lock(&fs_info->scrub_lock);
4490 }
4491 mutex_unlock(&fs_info->scrub_lock);
4492
4493 return 0;
4494}
4495
4496int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4497 struct btrfs_scrub_progress *progress)
4498{
4499 struct btrfs_dev_lookup_args args = { .devid = devid };
4500 struct btrfs_device *dev;
4501 struct scrub_ctx *sctx = NULL;
4502
4503 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4504 dev = btrfs_find_device(fs_info->fs_devices, &args);
4505 if (dev)
4506 sctx = dev->scrub_ctx;
4507 if (sctx)
4508 memcpy(progress, &sctx->stat, sizeof(*progress));
4509 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4510
4511 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4512}
4513
4514static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
4515 u64 extent_logical, u32 extent_len,
4516 u64 *extent_physical,
4517 struct btrfs_device **extent_dev,
4518 int *extent_mirror_num)
4519{
4520 u64 mapped_length;
4521 struct btrfs_io_context *bioc = NULL;
4522 int ret;
4523
4524 mapped_length = extent_len;
4525 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4526 &mapped_length, &bioc, 0);
4527 if (ret || !bioc || mapped_length < extent_len ||
4528 !bioc->stripes[0].dev->bdev) {
4529 btrfs_put_bioc(bioc);
4530 return;
4531 }
4532
4533 *extent_physical = bioc->stripes[0].physical;
4534 *extent_mirror_num = bioc->mirror_num;
4535 *extent_dev = bioc->stripes[0].dev;
4536 btrfs_put_bioc(bioc);
4537}
1/*
2 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/blkdev.h>
20#include <linux/ratelimit.h>
21#include "ctree.h"
22#include "volumes.h"
23#include "disk-io.h"
24#include "ordered-data.h"
25#include "transaction.h"
26#include "backref.h"
27#include "extent_io.h"
28#include "dev-replace.h"
29#include "check-integrity.h"
30#include "rcu-string.h"
31#include "raid56.h"
32
33/*
34 * This is only the first step towards a full-features scrub. It reads all
35 * extent and super block and verifies the checksums. In case a bad checksum
36 * is found or the extent cannot be read, good data will be written back if
37 * any can be found.
38 *
39 * Future enhancements:
40 * - In case an unrepairable extent is encountered, track which files are
41 * affected and report them
42 * - track and record media errors, throw out bad devices
43 * - add a mode to also read unallocated space
44 */
45
46struct scrub_block;
47struct scrub_ctx;
48
49/*
50 * the following three values only influence the performance.
51 * The last one configures the number of parallel and outstanding I/O
52 * operations. The first two values configure an upper limit for the number
53 * of (dynamically allocated) pages that are added to a bio.
54 */
55#define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
56#define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
57#define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
58
59/*
60 * the following value times PAGE_SIZE needs to be large enough to match the
61 * largest node/leaf/sector size that shall be supported.
62 * Values larger than BTRFS_STRIPE_LEN are not supported.
63 */
64#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
65
66struct scrub_page {
67 struct scrub_block *sblock;
68 struct page *page;
69 struct btrfs_device *dev;
70 u64 flags; /* extent flags */
71 u64 generation;
72 u64 logical;
73 u64 physical;
74 u64 physical_for_dev_replace;
75 atomic_t ref_count;
76 struct {
77 unsigned int mirror_num:8;
78 unsigned int have_csum:1;
79 unsigned int io_error:1;
80 };
81 u8 csum[BTRFS_CSUM_SIZE];
82};
83
84struct scrub_bio {
85 int index;
86 struct scrub_ctx *sctx;
87 struct btrfs_device *dev;
88 struct bio *bio;
89 int err;
90 u64 logical;
91 u64 physical;
92#if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
94#else
95 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
96#endif
97 int page_count;
98 int next_free;
99 struct btrfs_work work;
100};
101
102struct scrub_block {
103 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 int page_count;
105 atomic_t outstanding_pages;
106 atomic_t ref_count; /* free mem on transition to zero */
107 struct scrub_ctx *sctx;
108 struct {
109 unsigned int header_error:1;
110 unsigned int checksum_error:1;
111 unsigned int no_io_error_seen:1;
112 unsigned int generation_error:1; /* also sets header_error */
113 };
114};
115
116struct scrub_wr_ctx {
117 struct scrub_bio *wr_curr_bio;
118 struct btrfs_device *tgtdev;
119 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
120 atomic_t flush_all_writes;
121 struct mutex wr_lock;
122};
123
124struct scrub_ctx {
125 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
126 struct btrfs_root *dev_root;
127 int first_free;
128 int curr;
129 atomic_t bios_in_flight;
130 atomic_t workers_pending;
131 spinlock_t list_lock;
132 wait_queue_head_t list_wait;
133 u16 csum_size;
134 struct list_head csum_list;
135 atomic_t cancel_req;
136 int readonly;
137 int pages_per_rd_bio;
138 u32 sectorsize;
139 u32 nodesize;
140 u32 leafsize;
141
142 int is_dev_replace;
143 struct scrub_wr_ctx wr_ctx;
144
145 /*
146 * statistics
147 */
148 struct btrfs_scrub_progress stat;
149 spinlock_t stat_lock;
150};
151
152struct scrub_fixup_nodatasum {
153 struct scrub_ctx *sctx;
154 struct btrfs_device *dev;
155 u64 logical;
156 struct btrfs_root *root;
157 struct btrfs_work work;
158 int mirror_num;
159};
160
161struct scrub_nocow_inode {
162 u64 inum;
163 u64 offset;
164 u64 root;
165 struct list_head list;
166};
167
168struct scrub_copy_nocow_ctx {
169 struct scrub_ctx *sctx;
170 u64 logical;
171 u64 len;
172 int mirror_num;
173 u64 physical_for_dev_replace;
174 struct list_head inodes;
175 struct btrfs_work work;
176};
177
178struct scrub_warning {
179 struct btrfs_path *path;
180 u64 extent_item_size;
181 char *scratch_buf;
182 char *msg_buf;
183 const char *errstr;
184 sector_t sector;
185 u64 logical;
186 struct btrfs_device *dev;
187 int msg_bufsize;
188 int scratch_bufsize;
189};
190
191
192static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
193static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
194static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
195static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
196static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
197static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
198 struct btrfs_fs_info *fs_info,
199 struct scrub_block *original_sblock,
200 u64 length, u64 logical,
201 struct scrub_block *sblocks_for_recheck);
202static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
203 struct scrub_block *sblock, int is_metadata,
204 int have_csum, u8 *csum, u64 generation,
205 u16 csum_size);
206static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
207 struct scrub_block *sblock,
208 int is_metadata, int have_csum,
209 const u8 *csum, u64 generation,
210 u16 csum_size);
211static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
212 struct scrub_block *sblock_good,
213 int force_write);
214static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
215 struct scrub_block *sblock_good,
216 int page_num, int force_write);
217static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
218static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
219 int page_num);
220static int scrub_checksum_data(struct scrub_block *sblock);
221static int scrub_checksum_tree_block(struct scrub_block *sblock);
222static int scrub_checksum_super(struct scrub_block *sblock);
223static void scrub_block_get(struct scrub_block *sblock);
224static void scrub_block_put(struct scrub_block *sblock);
225static void scrub_page_get(struct scrub_page *spage);
226static void scrub_page_put(struct scrub_page *spage);
227static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
228 struct scrub_page *spage);
229static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
230 u64 physical, struct btrfs_device *dev, u64 flags,
231 u64 gen, int mirror_num, u8 *csum, int force,
232 u64 physical_for_dev_replace);
233static void scrub_bio_end_io(struct bio *bio, int err);
234static void scrub_bio_end_io_worker(struct btrfs_work *work);
235static void scrub_block_complete(struct scrub_block *sblock);
236static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
237 u64 extent_logical, u64 extent_len,
238 u64 *extent_physical,
239 struct btrfs_device **extent_dev,
240 int *extent_mirror_num);
241static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
242 struct scrub_wr_ctx *wr_ctx,
243 struct btrfs_fs_info *fs_info,
244 struct btrfs_device *dev,
245 int is_dev_replace);
246static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
247static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
248 struct scrub_page *spage);
249static void scrub_wr_submit(struct scrub_ctx *sctx);
250static void scrub_wr_bio_end_io(struct bio *bio, int err);
251static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
252static int write_page_nocow(struct scrub_ctx *sctx,
253 u64 physical_for_dev_replace, struct page *page);
254static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
255 struct scrub_copy_nocow_ctx *ctx);
256static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
257 int mirror_num, u64 physical_for_dev_replace);
258static void copy_nocow_pages_worker(struct btrfs_work *work);
259static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
260static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
261
262
263static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
264{
265 atomic_inc(&sctx->bios_in_flight);
266}
267
268static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
269{
270 atomic_dec(&sctx->bios_in_flight);
271 wake_up(&sctx->list_wait);
272}
273
274static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
275{
276 while (atomic_read(&fs_info->scrub_pause_req)) {
277 mutex_unlock(&fs_info->scrub_lock);
278 wait_event(fs_info->scrub_pause_wait,
279 atomic_read(&fs_info->scrub_pause_req) == 0);
280 mutex_lock(&fs_info->scrub_lock);
281 }
282}
283
284static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
285{
286 atomic_inc(&fs_info->scrubs_paused);
287 wake_up(&fs_info->scrub_pause_wait);
288
289 mutex_lock(&fs_info->scrub_lock);
290 __scrub_blocked_if_needed(fs_info);
291 atomic_dec(&fs_info->scrubs_paused);
292 mutex_unlock(&fs_info->scrub_lock);
293
294 wake_up(&fs_info->scrub_pause_wait);
295}
296
297/*
298 * used for workers that require transaction commits (i.e., for the
299 * NOCOW case)
300 */
301static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
302{
303 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
304
305 /*
306 * increment scrubs_running to prevent cancel requests from
307 * completing as long as a worker is running. we must also
308 * increment scrubs_paused to prevent deadlocking on pause
309 * requests used for transactions commits (as the worker uses a
310 * transaction context). it is safe to regard the worker
311 * as paused for all matters practical. effectively, we only
312 * avoid cancellation requests from completing.
313 */
314 mutex_lock(&fs_info->scrub_lock);
315 atomic_inc(&fs_info->scrubs_running);
316 atomic_inc(&fs_info->scrubs_paused);
317 mutex_unlock(&fs_info->scrub_lock);
318
319 /*
320 * check if @scrubs_running=@scrubs_paused condition
321 * inside wait_event() is not an atomic operation.
322 * which means we may inc/dec @scrub_running/paused
323 * at any time. Let's wake up @scrub_pause_wait as
324 * much as we can to let commit transaction blocked less.
325 */
326 wake_up(&fs_info->scrub_pause_wait);
327
328 atomic_inc(&sctx->workers_pending);
329}
330
331/* used for workers that require transaction commits */
332static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
333{
334 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
335
336 /*
337 * see scrub_pending_trans_workers_inc() why we're pretending
338 * to be paused in the scrub counters
339 */
340 mutex_lock(&fs_info->scrub_lock);
341 atomic_dec(&fs_info->scrubs_running);
342 atomic_dec(&fs_info->scrubs_paused);
343 mutex_unlock(&fs_info->scrub_lock);
344 atomic_dec(&sctx->workers_pending);
345 wake_up(&fs_info->scrub_pause_wait);
346 wake_up(&sctx->list_wait);
347}
348
349static void scrub_free_csums(struct scrub_ctx *sctx)
350{
351 while (!list_empty(&sctx->csum_list)) {
352 struct btrfs_ordered_sum *sum;
353 sum = list_first_entry(&sctx->csum_list,
354 struct btrfs_ordered_sum, list);
355 list_del(&sum->list);
356 kfree(sum);
357 }
358}
359
360static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
361{
362 int i;
363
364 if (!sctx)
365 return;
366
367 scrub_free_wr_ctx(&sctx->wr_ctx);
368
369 /* this can happen when scrub is cancelled */
370 if (sctx->curr != -1) {
371 struct scrub_bio *sbio = sctx->bios[sctx->curr];
372
373 for (i = 0; i < sbio->page_count; i++) {
374 WARN_ON(!sbio->pagev[i]->page);
375 scrub_block_put(sbio->pagev[i]->sblock);
376 }
377 bio_put(sbio->bio);
378 }
379
380 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
381 struct scrub_bio *sbio = sctx->bios[i];
382
383 if (!sbio)
384 break;
385 kfree(sbio);
386 }
387
388 scrub_free_csums(sctx);
389 kfree(sctx);
390}
391
392static noinline_for_stack
393struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
394{
395 struct scrub_ctx *sctx;
396 int i;
397 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
398 int pages_per_rd_bio;
399 int ret;
400
401 /*
402 * the setting of pages_per_rd_bio is correct for scrub but might
403 * be wrong for the dev_replace code where we might read from
404 * different devices in the initial huge bios. However, that
405 * code is able to correctly handle the case when adding a page
406 * to a bio fails.
407 */
408 if (dev->bdev)
409 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
410 bio_get_nr_vecs(dev->bdev));
411 else
412 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
413 sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
414 if (!sctx)
415 goto nomem;
416 sctx->is_dev_replace = is_dev_replace;
417 sctx->pages_per_rd_bio = pages_per_rd_bio;
418 sctx->curr = -1;
419 sctx->dev_root = dev->dev_root;
420 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
421 struct scrub_bio *sbio;
422
423 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
424 if (!sbio)
425 goto nomem;
426 sctx->bios[i] = sbio;
427
428 sbio->index = i;
429 sbio->sctx = sctx;
430 sbio->page_count = 0;
431 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker,
432 NULL, NULL);
433
434 if (i != SCRUB_BIOS_PER_SCTX - 1)
435 sctx->bios[i]->next_free = i + 1;
436 else
437 sctx->bios[i]->next_free = -1;
438 }
439 sctx->first_free = 0;
440 sctx->nodesize = dev->dev_root->nodesize;
441 sctx->leafsize = dev->dev_root->leafsize;
442 sctx->sectorsize = dev->dev_root->sectorsize;
443 atomic_set(&sctx->bios_in_flight, 0);
444 atomic_set(&sctx->workers_pending, 0);
445 atomic_set(&sctx->cancel_req, 0);
446 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
447 INIT_LIST_HEAD(&sctx->csum_list);
448
449 spin_lock_init(&sctx->list_lock);
450 spin_lock_init(&sctx->stat_lock);
451 init_waitqueue_head(&sctx->list_wait);
452
453 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
454 fs_info->dev_replace.tgtdev, is_dev_replace);
455 if (ret) {
456 scrub_free_ctx(sctx);
457 return ERR_PTR(ret);
458 }
459 return sctx;
460
461nomem:
462 scrub_free_ctx(sctx);
463 return ERR_PTR(-ENOMEM);
464}
465
466static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
467 void *warn_ctx)
468{
469 u64 isize;
470 u32 nlink;
471 int ret;
472 int i;
473 struct extent_buffer *eb;
474 struct btrfs_inode_item *inode_item;
475 struct scrub_warning *swarn = warn_ctx;
476 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
477 struct inode_fs_paths *ipath = NULL;
478 struct btrfs_root *local_root;
479 struct btrfs_key root_key;
480
481 root_key.objectid = root;
482 root_key.type = BTRFS_ROOT_ITEM_KEY;
483 root_key.offset = (u64)-1;
484 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
485 if (IS_ERR(local_root)) {
486 ret = PTR_ERR(local_root);
487 goto err;
488 }
489
490 ret = inode_item_info(inum, 0, local_root, swarn->path);
491 if (ret) {
492 btrfs_release_path(swarn->path);
493 goto err;
494 }
495
496 eb = swarn->path->nodes[0];
497 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
498 struct btrfs_inode_item);
499 isize = btrfs_inode_size(eb, inode_item);
500 nlink = btrfs_inode_nlink(eb, inode_item);
501 btrfs_release_path(swarn->path);
502
503 ipath = init_ipath(4096, local_root, swarn->path);
504 if (IS_ERR(ipath)) {
505 ret = PTR_ERR(ipath);
506 ipath = NULL;
507 goto err;
508 }
509 ret = paths_from_inode(inum, ipath);
510
511 if (ret < 0)
512 goto err;
513
514 /*
515 * we deliberately ignore the bit ipath might have been too small to
516 * hold all of the paths here
517 */
518 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
519 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
520 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
521 "length %llu, links %u (path: %s)\n", swarn->errstr,
522 swarn->logical, rcu_str_deref(swarn->dev->name),
523 (unsigned long long)swarn->sector, root, inum, offset,
524 min(isize - offset, (u64)PAGE_SIZE), nlink,
525 (char *)(unsigned long)ipath->fspath->val[i]);
526
527 free_ipath(ipath);
528 return 0;
529
530err:
531 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
532 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
533 "resolving failed with ret=%d\n", swarn->errstr,
534 swarn->logical, rcu_str_deref(swarn->dev->name),
535 (unsigned long long)swarn->sector, root, inum, offset, ret);
536
537 free_ipath(ipath);
538 return 0;
539}
540
541static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
542{
543 struct btrfs_device *dev;
544 struct btrfs_fs_info *fs_info;
545 struct btrfs_path *path;
546 struct btrfs_key found_key;
547 struct extent_buffer *eb;
548 struct btrfs_extent_item *ei;
549 struct scrub_warning swarn;
550 unsigned long ptr = 0;
551 u64 extent_item_pos;
552 u64 flags = 0;
553 u64 ref_root;
554 u32 item_size;
555 u8 ref_level;
556 const int bufsize = 4096;
557 int ret;
558
559 WARN_ON(sblock->page_count < 1);
560 dev = sblock->pagev[0]->dev;
561 fs_info = sblock->sctx->dev_root->fs_info;
562
563 path = btrfs_alloc_path();
564
565 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
566 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
567 swarn.sector = (sblock->pagev[0]->physical) >> 9;
568 swarn.logical = sblock->pagev[0]->logical;
569 swarn.errstr = errstr;
570 swarn.dev = NULL;
571 swarn.msg_bufsize = bufsize;
572 swarn.scratch_bufsize = bufsize;
573
574 if (!path || !swarn.scratch_buf || !swarn.msg_buf)
575 goto out;
576
577 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
578 &flags);
579 if (ret < 0)
580 goto out;
581
582 extent_item_pos = swarn.logical - found_key.objectid;
583 swarn.extent_item_size = found_key.offset;
584
585 eb = path->nodes[0];
586 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
587 item_size = btrfs_item_size_nr(eb, path->slots[0]);
588
589 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
590 do {
591 ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
592 &ref_root, &ref_level);
593 printk_in_rcu(KERN_WARNING
594 "BTRFS: %s at logical %llu on dev %s, "
595 "sector %llu: metadata %s (level %d) in tree "
596 "%llu\n", errstr, swarn.logical,
597 rcu_str_deref(dev->name),
598 (unsigned long long)swarn.sector,
599 ref_level ? "node" : "leaf",
600 ret < 0 ? -1 : ref_level,
601 ret < 0 ? -1 : ref_root);
602 } while (ret != 1);
603 btrfs_release_path(path);
604 } else {
605 btrfs_release_path(path);
606 swarn.path = path;
607 swarn.dev = dev;
608 iterate_extent_inodes(fs_info, found_key.objectid,
609 extent_item_pos, 1,
610 scrub_print_warning_inode, &swarn);
611 }
612
613out:
614 btrfs_free_path(path);
615 kfree(swarn.scratch_buf);
616 kfree(swarn.msg_buf);
617}
618
619static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
620{
621 struct page *page = NULL;
622 unsigned long index;
623 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
624 int ret;
625 int corrected = 0;
626 struct btrfs_key key;
627 struct inode *inode = NULL;
628 struct btrfs_fs_info *fs_info;
629 u64 end = offset + PAGE_SIZE - 1;
630 struct btrfs_root *local_root;
631 int srcu_index;
632
633 key.objectid = root;
634 key.type = BTRFS_ROOT_ITEM_KEY;
635 key.offset = (u64)-1;
636
637 fs_info = fixup->root->fs_info;
638 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
639
640 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
641 if (IS_ERR(local_root)) {
642 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
643 return PTR_ERR(local_root);
644 }
645
646 key.type = BTRFS_INODE_ITEM_KEY;
647 key.objectid = inum;
648 key.offset = 0;
649 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
650 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
651 if (IS_ERR(inode))
652 return PTR_ERR(inode);
653
654 index = offset >> PAGE_CACHE_SHIFT;
655
656 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
657 if (!page) {
658 ret = -ENOMEM;
659 goto out;
660 }
661
662 if (PageUptodate(page)) {
663 if (PageDirty(page)) {
664 /*
665 * we need to write the data to the defect sector. the
666 * data that was in that sector is not in memory,
667 * because the page was modified. we must not write the
668 * modified page to that sector.
669 *
670 * TODO: what could be done here: wait for the delalloc
671 * runner to write out that page (might involve
672 * COW) and see whether the sector is still
673 * referenced afterwards.
674 *
675 * For the meantime, we'll treat this error
676 * incorrectable, although there is a chance that a
677 * later scrub will find the bad sector again and that
678 * there's no dirty page in memory, then.
679 */
680 ret = -EIO;
681 goto out;
682 }
683 fs_info = BTRFS_I(inode)->root->fs_info;
684 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
685 fixup->logical, page,
686 fixup->mirror_num);
687 unlock_page(page);
688 corrected = !ret;
689 } else {
690 /*
691 * we need to get good data first. the general readpage path
692 * will call repair_io_failure for us, we just have to make
693 * sure we read the bad mirror.
694 */
695 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
696 EXTENT_DAMAGED, GFP_NOFS);
697 if (ret) {
698 /* set_extent_bits should give proper error */
699 WARN_ON(ret > 0);
700 if (ret > 0)
701 ret = -EFAULT;
702 goto out;
703 }
704
705 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
706 btrfs_get_extent,
707 fixup->mirror_num);
708 wait_on_page_locked(page);
709
710 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
711 end, EXTENT_DAMAGED, 0, NULL);
712 if (!corrected)
713 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
714 EXTENT_DAMAGED, GFP_NOFS);
715 }
716
717out:
718 if (page)
719 put_page(page);
720 if (inode)
721 iput(inode);
722
723 if (ret < 0)
724 return ret;
725
726 if (ret == 0 && corrected) {
727 /*
728 * we only need to call readpage for one of the inodes belonging
729 * to this extent. so make iterate_extent_inodes stop
730 */
731 return 1;
732 }
733
734 return -EIO;
735}
736
737static void scrub_fixup_nodatasum(struct btrfs_work *work)
738{
739 int ret;
740 struct scrub_fixup_nodatasum *fixup;
741 struct scrub_ctx *sctx;
742 struct btrfs_trans_handle *trans = NULL;
743 struct btrfs_path *path;
744 int uncorrectable = 0;
745
746 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
747 sctx = fixup->sctx;
748
749 path = btrfs_alloc_path();
750 if (!path) {
751 spin_lock(&sctx->stat_lock);
752 ++sctx->stat.malloc_errors;
753 spin_unlock(&sctx->stat_lock);
754 uncorrectable = 1;
755 goto out;
756 }
757
758 trans = btrfs_join_transaction(fixup->root);
759 if (IS_ERR(trans)) {
760 uncorrectable = 1;
761 goto out;
762 }
763
764 /*
765 * the idea is to trigger a regular read through the standard path. we
766 * read a page from the (failed) logical address by specifying the
767 * corresponding copynum of the failed sector. thus, that readpage is
768 * expected to fail.
769 * that is the point where on-the-fly error correction will kick in
770 * (once it's finished) and rewrite the failed sector if a good copy
771 * can be found.
772 */
773 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
774 path, scrub_fixup_readpage,
775 fixup);
776 if (ret < 0) {
777 uncorrectable = 1;
778 goto out;
779 }
780 WARN_ON(ret != 1);
781
782 spin_lock(&sctx->stat_lock);
783 ++sctx->stat.corrected_errors;
784 spin_unlock(&sctx->stat_lock);
785
786out:
787 if (trans && !IS_ERR(trans))
788 btrfs_end_transaction(trans, fixup->root);
789 if (uncorrectable) {
790 spin_lock(&sctx->stat_lock);
791 ++sctx->stat.uncorrectable_errors;
792 spin_unlock(&sctx->stat_lock);
793 btrfs_dev_replace_stats_inc(
794 &sctx->dev_root->fs_info->dev_replace.
795 num_uncorrectable_read_errors);
796 printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
797 "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
798 fixup->logical, rcu_str_deref(fixup->dev->name));
799 }
800
801 btrfs_free_path(path);
802 kfree(fixup);
803
804 scrub_pending_trans_workers_dec(sctx);
805}
806
807/*
808 * scrub_handle_errored_block gets called when either verification of the
809 * pages failed or the bio failed to read, e.g. with EIO. In the latter
810 * case, this function handles all pages in the bio, even though only one
811 * may be bad.
812 * The goal of this function is to repair the errored block by using the
813 * contents of one of the mirrors.
814 */
815static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
816{
817 struct scrub_ctx *sctx = sblock_to_check->sctx;
818 struct btrfs_device *dev;
819 struct btrfs_fs_info *fs_info;
820 u64 length;
821 u64 logical;
822 u64 generation;
823 unsigned int failed_mirror_index;
824 unsigned int is_metadata;
825 unsigned int have_csum;
826 u8 *csum;
827 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
828 struct scrub_block *sblock_bad;
829 int ret;
830 int mirror_index;
831 int page_num;
832 int success;
833 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
834 DEFAULT_RATELIMIT_BURST);
835
836 BUG_ON(sblock_to_check->page_count < 1);
837 fs_info = sctx->dev_root->fs_info;
838 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
839 /*
840 * if we find an error in a super block, we just report it.
841 * They will get written with the next transaction commit
842 * anyway
843 */
844 spin_lock(&sctx->stat_lock);
845 ++sctx->stat.super_errors;
846 spin_unlock(&sctx->stat_lock);
847 return 0;
848 }
849 length = sblock_to_check->page_count * PAGE_SIZE;
850 logical = sblock_to_check->pagev[0]->logical;
851 generation = sblock_to_check->pagev[0]->generation;
852 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
853 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
854 is_metadata = !(sblock_to_check->pagev[0]->flags &
855 BTRFS_EXTENT_FLAG_DATA);
856 have_csum = sblock_to_check->pagev[0]->have_csum;
857 csum = sblock_to_check->pagev[0]->csum;
858 dev = sblock_to_check->pagev[0]->dev;
859
860 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
861 sblocks_for_recheck = NULL;
862 goto nodatasum_case;
863 }
864
865 /*
866 * read all mirrors one after the other. This includes to
867 * re-read the extent or metadata block that failed (that was
868 * the cause that this fixup code is called) another time,
869 * page by page this time in order to know which pages
870 * caused I/O errors and which ones are good (for all mirrors).
871 * It is the goal to handle the situation when more than one
872 * mirror contains I/O errors, but the errors do not
873 * overlap, i.e. the data can be repaired by selecting the
874 * pages from those mirrors without I/O error on the
875 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
876 * would be that mirror #1 has an I/O error on the first page,
877 * the second page is good, and mirror #2 has an I/O error on
878 * the second page, but the first page is good.
879 * Then the first page of the first mirror can be repaired by
880 * taking the first page of the second mirror, and the
881 * second page of the second mirror can be repaired by
882 * copying the contents of the 2nd page of the 1st mirror.
883 * One more note: if the pages of one mirror contain I/O
884 * errors, the checksum cannot be verified. In order to get
885 * the best data for repairing, the first attempt is to find
886 * a mirror without I/O errors and with a validated checksum.
887 * Only if this is not possible, the pages are picked from
888 * mirrors with I/O errors without considering the checksum.
889 * If the latter is the case, at the end, the checksum of the
890 * repaired area is verified in order to correctly maintain
891 * the statistics.
892 */
893
894 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
895 sizeof(*sblocks_for_recheck),
896 GFP_NOFS);
897 if (!sblocks_for_recheck) {
898 spin_lock(&sctx->stat_lock);
899 sctx->stat.malloc_errors++;
900 sctx->stat.read_errors++;
901 sctx->stat.uncorrectable_errors++;
902 spin_unlock(&sctx->stat_lock);
903 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
904 goto out;
905 }
906
907 /* setup the context, map the logical blocks and alloc the pages */
908 ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
909 logical, sblocks_for_recheck);
910 if (ret) {
911 spin_lock(&sctx->stat_lock);
912 sctx->stat.read_errors++;
913 sctx->stat.uncorrectable_errors++;
914 spin_unlock(&sctx->stat_lock);
915 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
916 goto out;
917 }
918 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
919 sblock_bad = sblocks_for_recheck + failed_mirror_index;
920
921 /* build and submit the bios for the failed mirror, check checksums */
922 scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
923 csum, generation, sctx->csum_size);
924
925 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
926 sblock_bad->no_io_error_seen) {
927 /*
928 * the error disappeared after reading page by page, or
929 * the area was part of a huge bio and other parts of the
930 * bio caused I/O errors, or the block layer merged several
931 * read requests into one and the error is caused by a
932 * different bio (usually one of the two latter cases is
933 * the cause)
934 */
935 spin_lock(&sctx->stat_lock);
936 sctx->stat.unverified_errors++;
937 spin_unlock(&sctx->stat_lock);
938
939 if (sctx->is_dev_replace)
940 scrub_write_block_to_dev_replace(sblock_bad);
941 goto out;
942 }
943
944 if (!sblock_bad->no_io_error_seen) {
945 spin_lock(&sctx->stat_lock);
946 sctx->stat.read_errors++;
947 spin_unlock(&sctx->stat_lock);
948 if (__ratelimit(&_rs))
949 scrub_print_warning("i/o error", sblock_to_check);
950 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
951 } else if (sblock_bad->checksum_error) {
952 spin_lock(&sctx->stat_lock);
953 sctx->stat.csum_errors++;
954 spin_unlock(&sctx->stat_lock);
955 if (__ratelimit(&_rs))
956 scrub_print_warning("checksum error", sblock_to_check);
957 btrfs_dev_stat_inc_and_print(dev,
958 BTRFS_DEV_STAT_CORRUPTION_ERRS);
959 } else if (sblock_bad->header_error) {
960 spin_lock(&sctx->stat_lock);
961 sctx->stat.verify_errors++;
962 spin_unlock(&sctx->stat_lock);
963 if (__ratelimit(&_rs))
964 scrub_print_warning("checksum/header error",
965 sblock_to_check);
966 if (sblock_bad->generation_error)
967 btrfs_dev_stat_inc_and_print(dev,
968 BTRFS_DEV_STAT_GENERATION_ERRS);
969 else
970 btrfs_dev_stat_inc_and_print(dev,
971 BTRFS_DEV_STAT_CORRUPTION_ERRS);
972 }
973
974 if (sctx->readonly) {
975 ASSERT(!sctx->is_dev_replace);
976 goto out;
977 }
978
979 if (!is_metadata && !have_csum) {
980 struct scrub_fixup_nodatasum *fixup_nodatasum;
981
982nodatasum_case:
983 WARN_ON(sctx->is_dev_replace);
984
985 /*
986 * !is_metadata and !have_csum, this means that the data
987 * might not be COW'ed, that it might be modified
988 * concurrently. The general strategy to work on the
989 * commit root does not help in the case when COW is not
990 * used.
991 */
992 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
993 if (!fixup_nodatasum)
994 goto did_not_correct_error;
995 fixup_nodatasum->sctx = sctx;
996 fixup_nodatasum->dev = dev;
997 fixup_nodatasum->logical = logical;
998 fixup_nodatasum->root = fs_info->extent_root;
999 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1000 scrub_pending_trans_workers_inc(sctx);
1001 btrfs_init_work(&fixup_nodatasum->work, scrub_fixup_nodatasum,
1002 NULL, NULL);
1003 btrfs_queue_work(fs_info->scrub_workers,
1004 &fixup_nodatasum->work);
1005 goto out;
1006 }
1007
1008 /*
1009 * now build and submit the bios for the other mirrors, check
1010 * checksums.
1011 * First try to pick the mirror which is completely without I/O
1012 * errors and also does not have a checksum error.
1013 * If one is found, and if a checksum is present, the full block
1014 * that is known to contain an error is rewritten. Afterwards
1015 * the block is known to be corrected.
1016 * If a mirror is found which is completely correct, and no
1017 * checksum is present, only those pages are rewritten that had
1018 * an I/O error in the block to be repaired, since it cannot be
1019 * determined, which copy of the other pages is better (and it
1020 * could happen otherwise that a correct page would be
1021 * overwritten by a bad one).
1022 */
1023 for (mirror_index = 0;
1024 mirror_index < BTRFS_MAX_MIRRORS &&
1025 sblocks_for_recheck[mirror_index].page_count > 0;
1026 mirror_index++) {
1027 struct scrub_block *sblock_other;
1028
1029 if (mirror_index == failed_mirror_index)
1030 continue;
1031 sblock_other = sblocks_for_recheck + mirror_index;
1032
1033 /* build and submit the bios, check checksums */
1034 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1035 have_csum, csum, generation,
1036 sctx->csum_size);
1037
1038 if (!sblock_other->header_error &&
1039 !sblock_other->checksum_error &&
1040 sblock_other->no_io_error_seen) {
1041 if (sctx->is_dev_replace) {
1042 scrub_write_block_to_dev_replace(sblock_other);
1043 } else {
1044 int force_write = is_metadata || have_csum;
1045
1046 ret = scrub_repair_block_from_good_copy(
1047 sblock_bad, sblock_other,
1048 force_write);
1049 }
1050 if (0 == ret)
1051 goto corrected_error;
1052 }
1053 }
1054
1055 /*
1056 * for dev_replace, pick good pages and write to the target device.
1057 */
1058 if (sctx->is_dev_replace) {
1059 success = 1;
1060 for (page_num = 0; page_num < sblock_bad->page_count;
1061 page_num++) {
1062 int sub_success;
1063
1064 sub_success = 0;
1065 for (mirror_index = 0;
1066 mirror_index < BTRFS_MAX_MIRRORS &&
1067 sblocks_for_recheck[mirror_index].page_count > 0;
1068 mirror_index++) {
1069 struct scrub_block *sblock_other =
1070 sblocks_for_recheck + mirror_index;
1071 struct scrub_page *page_other =
1072 sblock_other->pagev[page_num];
1073
1074 if (!page_other->io_error) {
1075 ret = scrub_write_page_to_dev_replace(
1076 sblock_other, page_num);
1077 if (ret == 0) {
1078 /* succeeded for this page */
1079 sub_success = 1;
1080 break;
1081 } else {
1082 btrfs_dev_replace_stats_inc(
1083 &sctx->dev_root->
1084 fs_info->dev_replace.
1085 num_write_errors);
1086 }
1087 }
1088 }
1089
1090 if (!sub_success) {
1091 /*
1092 * did not find a mirror to fetch the page
1093 * from. scrub_write_page_to_dev_replace()
1094 * handles this case (page->io_error), by
1095 * filling the block with zeros before
1096 * submitting the write request
1097 */
1098 success = 0;
1099 ret = scrub_write_page_to_dev_replace(
1100 sblock_bad, page_num);
1101 if (ret)
1102 btrfs_dev_replace_stats_inc(
1103 &sctx->dev_root->fs_info->
1104 dev_replace.num_write_errors);
1105 }
1106 }
1107
1108 goto out;
1109 }
1110
1111 /*
1112 * for regular scrub, repair those pages that are errored.
1113 * In case of I/O errors in the area that is supposed to be
1114 * repaired, continue by picking good copies of those pages.
1115 * Select the good pages from mirrors to rewrite bad pages from
1116 * the area to fix. Afterwards verify the checksum of the block
1117 * that is supposed to be repaired. This verification step is
1118 * only done for the purpose of statistic counting and for the
1119 * final scrub report, whether errors remain.
1120 * A perfect algorithm could make use of the checksum and try
1121 * all possible combinations of pages from the different mirrors
1122 * until the checksum verification succeeds. For example, when
1123 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1124 * of mirror #2 is readable but the final checksum test fails,
1125 * then the 2nd page of mirror #3 could be tried, whether now
1126 * the final checksum succeedes. But this would be a rare
1127 * exception and is therefore not implemented. At least it is
1128 * avoided that the good copy is overwritten.
1129 * A more useful improvement would be to pick the sectors
1130 * without I/O error based on sector sizes (512 bytes on legacy
1131 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1132 * mirror could be repaired by taking 512 byte of a different
1133 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1134 * area are unreadable.
1135 */
1136
1137 /* can only fix I/O errors from here on */
1138 if (sblock_bad->no_io_error_seen)
1139 goto did_not_correct_error;
1140
1141 success = 1;
1142 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1143 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1144
1145 if (!page_bad->io_error)
1146 continue;
1147
1148 for (mirror_index = 0;
1149 mirror_index < BTRFS_MAX_MIRRORS &&
1150 sblocks_for_recheck[mirror_index].page_count > 0;
1151 mirror_index++) {
1152 struct scrub_block *sblock_other = sblocks_for_recheck +
1153 mirror_index;
1154 struct scrub_page *page_other = sblock_other->pagev[
1155 page_num];
1156
1157 if (!page_other->io_error) {
1158 ret = scrub_repair_page_from_good_copy(
1159 sblock_bad, sblock_other, page_num, 0);
1160 if (0 == ret) {
1161 page_bad->io_error = 0;
1162 break; /* succeeded for this page */
1163 }
1164 }
1165 }
1166
1167 if (page_bad->io_error) {
1168 /* did not find a mirror to copy the page from */
1169 success = 0;
1170 }
1171 }
1172
1173 if (success) {
1174 if (is_metadata || have_csum) {
1175 /*
1176 * need to verify the checksum now that all
1177 * sectors on disk are repaired (the write
1178 * request for data to be repaired is on its way).
1179 * Just be lazy and use scrub_recheck_block()
1180 * which re-reads the data before the checksum
1181 * is verified, but most likely the data comes out
1182 * of the page cache.
1183 */
1184 scrub_recheck_block(fs_info, sblock_bad,
1185 is_metadata, have_csum, csum,
1186 generation, sctx->csum_size);
1187 if (!sblock_bad->header_error &&
1188 !sblock_bad->checksum_error &&
1189 sblock_bad->no_io_error_seen)
1190 goto corrected_error;
1191 else
1192 goto did_not_correct_error;
1193 } else {
1194corrected_error:
1195 spin_lock(&sctx->stat_lock);
1196 sctx->stat.corrected_errors++;
1197 spin_unlock(&sctx->stat_lock);
1198 printk_ratelimited_in_rcu(KERN_ERR
1199 "BTRFS: fixed up error at logical %llu on dev %s\n",
1200 logical, rcu_str_deref(dev->name));
1201 }
1202 } else {
1203did_not_correct_error:
1204 spin_lock(&sctx->stat_lock);
1205 sctx->stat.uncorrectable_errors++;
1206 spin_unlock(&sctx->stat_lock);
1207 printk_ratelimited_in_rcu(KERN_ERR
1208 "BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1209 logical, rcu_str_deref(dev->name));
1210 }
1211
1212out:
1213 if (sblocks_for_recheck) {
1214 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1215 mirror_index++) {
1216 struct scrub_block *sblock = sblocks_for_recheck +
1217 mirror_index;
1218 int page_index;
1219
1220 for (page_index = 0; page_index < sblock->page_count;
1221 page_index++) {
1222 sblock->pagev[page_index]->sblock = NULL;
1223 scrub_page_put(sblock->pagev[page_index]);
1224 }
1225 }
1226 kfree(sblocks_for_recheck);
1227 }
1228
1229 return 0;
1230}
1231
1232static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1233 struct btrfs_fs_info *fs_info,
1234 struct scrub_block *original_sblock,
1235 u64 length, u64 logical,
1236 struct scrub_block *sblocks_for_recheck)
1237{
1238 int page_index;
1239 int mirror_index;
1240 int ret;
1241
1242 /*
1243 * note: the two members ref_count and outstanding_pages
1244 * are not used (and not set) in the blocks that are used for
1245 * the recheck procedure
1246 */
1247
1248 page_index = 0;
1249 while (length > 0) {
1250 u64 sublen = min_t(u64, length, PAGE_SIZE);
1251 u64 mapped_length = sublen;
1252 struct btrfs_bio *bbio = NULL;
1253
1254 /*
1255 * with a length of PAGE_SIZE, each returned stripe
1256 * represents one mirror
1257 */
1258 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1259 &mapped_length, &bbio, 0);
1260 if (ret || !bbio || mapped_length < sublen) {
1261 kfree(bbio);
1262 return -EIO;
1263 }
1264
1265 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1266 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1267 mirror_index++) {
1268 struct scrub_block *sblock;
1269 struct scrub_page *page;
1270
1271 if (mirror_index >= BTRFS_MAX_MIRRORS)
1272 continue;
1273
1274 sblock = sblocks_for_recheck + mirror_index;
1275 sblock->sctx = sctx;
1276 page = kzalloc(sizeof(*page), GFP_NOFS);
1277 if (!page) {
1278leave_nomem:
1279 spin_lock(&sctx->stat_lock);
1280 sctx->stat.malloc_errors++;
1281 spin_unlock(&sctx->stat_lock);
1282 kfree(bbio);
1283 return -ENOMEM;
1284 }
1285 scrub_page_get(page);
1286 sblock->pagev[page_index] = page;
1287 page->logical = logical;
1288 page->physical = bbio->stripes[mirror_index].physical;
1289 BUG_ON(page_index >= original_sblock->page_count);
1290 page->physical_for_dev_replace =
1291 original_sblock->pagev[page_index]->
1292 physical_for_dev_replace;
1293 /* for missing devices, dev->bdev is NULL */
1294 page->dev = bbio->stripes[mirror_index].dev;
1295 page->mirror_num = mirror_index + 1;
1296 sblock->page_count++;
1297 page->page = alloc_page(GFP_NOFS);
1298 if (!page->page)
1299 goto leave_nomem;
1300 }
1301 kfree(bbio);
1302 length -= sublen;
1303 logical += sublen;
1304 page_index++;
1305 }
1306
1307 return 0;
1308}
1309
1310/*
1311 * this function will check the on disk data for checksum errors, header
1312 * errors and read I/O errors. If any I/O errors happen, the exact pages
1313 * which are errored are marked as being bad. The goal is to enable scrub
1314 * to take those pages that are not errored from all the mirrors so that
1315 * the pages that are errored in the just handled mirror can be repaired.
1316 */
1317static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1318 struct scrub_block *sblock, int is_metadata,
1319 int have_csum, u8 *csum, u64 generation,
1320 u16 csum_size)
1321{
1322 int page_num;
1323
1324 sblock->no_io_error_seen = 1;
1325 sblock->header_error = 0;
1326 sblock->checksum_error = 0;
1327
1328 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1329 struct bio *bio;
1330 struct scrub_page *page = sblock->pagev[page_num];
1331
1332 if (page->dev->bdev == NULL) {
1333 page->io_error = 1;
1334 sblock->no_io_error_seen = 0;
1335 continue;
1336 }
1337
1338 WARN_ON(!page->page);
1339 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1340 if (!bio) {
1341 page->io_error = 1;
1342 sblock->no_io_error_seen = 0;
1343 continue;
1344 }
1345 bio->bi_bdev = page->dev->bdev;
1346 bio->bi_iter.bi_sector = page->physical >> 9;
1347
1348 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1349 if (btrfsic_submit_bio_wait(READ, bio))
1350 sblock->no_io_error_seen = 0;
1351
1352 bio_put(bio);
1353 }
1354
1355 if (sblock->no_io_error_seen)
1356 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1357 have_csum, csum, generation,
1358 csum_size);
1359
1360 return;
1361}
1362
1363static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1364 struct scrub_block *sblock,
1365 int is_metadata, int have_csum,
1366 const u8 *csum, u64 generation,
1367 u16 csum_size)
1368{
1369 int page_num;
1370 u8 calculated_csum[BTRFS_CSUM_SIZE];
1371 u32 crc = ~(u32)0;
1372 void *mapped_buffer;
1373
1374 WARN_ON(!sblock->pagev[0]->page);
1375 if (is_metadata) {
1376 struct btrfs_header *h;
1377
1378 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1379 h = (struct btrfs_header *)mapped_buffer;
1380
1381 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1382 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1383 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1384 BTRFS_UUID_SIZE)) {
1385 sblock->header_error = 1;
1386 } else if (generation != btrfs_stack_header_generation(h)) {
1387 sblock->header_error = 1;
1388 sblock->generation_error = 1;
1389 }
1390 csum = h->csum;
1391 } else {
1392 if (!have_csum)
1393 return;
1394
1395 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1396 }
1397
1398 for (page_num = 0;;) {
1399 if (page_num == 0 && is_metadata)
1400 crc = btrfs_csum_data(
1401 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1402 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1403 else
1404 crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1405
1406 kunmap_atomic(mapped_buffer);
1407 page_num++;
1408 if (page_num >= sblock->page_count)
1409 break;
1410 WARN_ON(!sblock->pagev[page_num]->page);
1411
1412 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1413 }
1414
1415 btrfs_csum_final(crc, calculated_csum);
1416 if (memcmp(calculated_csum, csum, csum_size))
1417 sblock->checksum_error = 1;
1418}
1419
1420static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1421 struct scrub_block *sblock_good,
1422 int force_write)
1423{
1424 int page_num;
1425 int ret = 0;
1426
1427 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1428 int ret_sub;
1429
1430 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1431 sblock_good,
1432 page_num,
1433 force_write);
1434 if (ret_sub)
1435 ret = ret_sub;
1436 }
1437
1438 return ret;
1439}
1440
1441static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1442 struct scrub_block *sblock_good,
1443 int page_num, int force_write)
1444{
1445 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1446 struct scrub_page *page_good = sblock_good->pagev[page_num];
1447
1448 BUG_ON(page_bad->page == NULL);
1449 BUG_ON(page_good->page == NULL);
1450 if (force_write || sblock_bad->header_error ||
1451 sblock_bad->checksum_error || page_bad->io_error) {
1452 struct bio *bio;
1453 int ret;
1454
1455 if (!page_bad->dev->bdev) {
1456 printk_ratelimited(KERN_WARNING "BTRFS: "
1457 "scrub_repair_page_from_good_copy(bdev == NULL) "
1458 "is unexpected!\n");
1459 return -EIO;
1460 }
1461
1462 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1463 if (!bio)
1464 return -EIO;
1465 bio->bi_bdev = page_bad->dev->bdev;
1466 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1467
1468 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1469 if (PAGE_SIZE != ret) {
1470 bio_put(bio);
1471 return -EIO;
1472 }
1473
1474 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1475 btrfs_dev_stat_inc_and_print(page_bad->dev,
1476 BTRFS_DEV_STAT_WRITE_ERRS);
1477 btrfs_dev_replace_stats_inc(
1478 &sblock_bad->sctx->dev_root->fs_info->
1479 dev_replace.num_write_errors);
1480 bio_put(bio);
1481 return -EIO;
1482 }
1483 bio_put(bio);
1484 }
1485
1486 return 0;
1487}
1488
1489static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1490{
1491 int page_num;
1492
1493 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1494 int ret;
1495
1496 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1497 if (ret)
1498 btrfs_dev_replace_stats_inc(
1499 &sblock->sctx->dev_root->fs_info->dev_replace.
1500 num_write_errors);
1501 }
1502}
1503
1504static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1505 int page_num)
1506{
1507 struct scrub_page *spage = sblock->pagev[page_num];
1508
1509 BUG_ON(spage->page == NULL);
1510 if (spage->io_error) {
1511 void *mapped_buffer = kmap_atomic(spage->page);
1512
1513 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1514 flush_dcache_page(spage->page);
1515 kunmap_atomic(mapped_buffer);
1516 }
1517 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1518}
1519
1520static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1521 struct scrub_page *spage)
1522{
1523 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1524 struct scrub_bio *sbio;
1525 int ret;
1526
1527 mutex_lock(&wr_ctx->wr_lock);
1528again:
1529 if (!wr_ctx->wr_curr_bio) {
1530 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1531 GFP_NOFS);
1532 if (!wr_ctx->wr_curr_bio) {
1533 mutex_unlock(&wr_ctx->wr_lock);
1534 return -ENOMEM;
1535 }
1536 wr_ctx->wr_curr_bio->sctx = sctx;
1537 wr_ctx->wr_curr_bio->page_count = 0;
1538 }
1539 sbio = wr_ctx->wr_curr_bio;
1540 if (sbio->page_count == 0) {
1541 struct bio *bio;
1542
1543 sbio->physical = spage->physical_for_dev_replace;
1544 sbio->logical = spage->logical;
1545 sbio->dev = wr_ctx->tgtdev;
1546 bio = sbio->bio;
1547 if (!bio) {
1548 bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1549 if (!bio) {
1550 mutex_unlock(&wr_ctx->wr_lock);
1551 return -ENOMEM;
1552 }
1553 sbio->bio = bio;
1554 }
1555
1556 bio->bi_private = sbio;
1557 bio->bi_end_io = scrub_wr_bio_end_io;
1558 bio->bi_bdev = sbio->dev->bdev;
1559 bio->bi_iter.bi_sector = sbio->physical >> 9;
1560 sbio->err = 0;
1561 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1562 spage->physical_for_dev_replace ||
1563 sbio->logical + sbio->page_count * PAGE_SIZE !=
1564 spage->logical) {
1565 scrub_wr_submit(sctx);
1566 goto again;
1567 }
1568
1569 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1570 if (ret != PAGE_SIZE) {
1571 if (sbio->page_count < 1) {
1572 bio_put(sbio->bio);
1573 sbio->bio = NULL;
1574 mutex_unlock(&wr_ctx->wr_lock);
1575 return -EIO;
1576 }
1577 scrub_wr_submit(sctx);
1578 goto again;
1579 }
1580
1581 sbio->pagev[sbio->page_count] = spage;
1582 scrub_page_get(spage);
1583 sbio->page_count++;
1584 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1585 scrub_wr_submit(sctx);
1586 mutex_unlock(&wr_ctx->wr_lock);
1587
1588 return 0;
1589}
1590
1591static void scrub_wr_submit(struct scrub_ctx *sctx)
1592{
1593 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1594 struct scrub_bio *sbio;
1595
1596 if (!wr_ctx->wr_curr_bio)
1597 return;
1598
1599 sbio = wr_ctx->wr_curr_bio;
1600 wr_ctx->wr_curr_bio = NULL;
1601 WARN_ON(!sbio->bio->bi_bdev);
1602 scrub_pending_bio_inc(sctx);
1603 /* process all writes in a single worker thread. Then the block layer
1604 * orders the requests before sending them to the driver which
1605 * doubled the write performance on spinning disks when measured
1606 * with Linux 3.5 */
1607 btrfsic_submit_bio(WRITE, sbio->bio);
1608}
1609
1610static void scrub_wr_bio_end_io(struct bio *bio, int err)
1611{
1612 struct scrub_bio *sbio = bio->bi_private;
1613 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1614
1615 sbio->err = err;
1616 sbio->bio = bio;
1617
1618 btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1619 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1620}
1621
1622static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1623{
1624 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1625 struct scrub_ctx *sctx = sbio->sctx;
1626 int i;
1627
1628 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1629 if (sbio->err) {
1630 struct btrfs_dev_replace *dev_replace =
1631 &sbio->sctx->dev_root->fs_info->dev_replace;
1632
1633 for (i = 0; i < sbio->page_count; i++) {
1634 struct scrub_page *spage = sbio->pagev[i];
1635
1636 spage->io_error = 1;
1637 btrfs_dev_replace_stats_inc(&dev_replace->
1638 num_write_errors);
1639 }
1640 }
1641
1642 for (i = 0; i < sbio->page_count; i++)
1643 scrub_page_put(sbio->pagev[i]);
1644
1645 bio_put(sbio->bio);
1646 kfree(sbio);
1647 scrub_pending_bio_dec(sctx);
1648}
1649
1650static int scrub_checksum(struct scrub_block *sblock)
1651{
1652 u64 flags;
1653 int ret;
1654
1655 WARN_ON(sblock->page_count < 1);
1656 flags = sblock->pagev[0]->flags;
1657 ret = 0;
1658 if (flags & BTRFS_EXTENT_FLAG_DATA)
1659 ret = scrub_checksum_data(sblock);
1660 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1661 ret = scrub_checksum_tree_block(sblock);
1662 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1663 (void)scrub_checksum_super(sblock);
1664 else
1665 WARN_ON(1);
1666 if (ret)
1667 scrub_handle_errored_block(sblock);
1668
1669 return ret;
1670}
1671
1672static int scrub_checksum_data(struct scrub_block *sblock)
1673{
1674 struct scrub_ctx *sctx = sblock->sctx;
1675 u8 csum[BTRFS_CSUM_SIZE];
1676 u8 *on_disk_csum;
1677 struct page *page;
1678 void *buffer;
1679 u32 crc = ~(u32)0;
1680 int fail = 0;
1681 u64 len;
1682 int index;
1683
1684 BUG_ON(sblock->page_count < 1);
1685 if (!sblock->pagev[0]->have_csum)
1686 return 0;
1687
1688 on_disk_csum = sblock->pagev[0]->csum;
1689 page = sblock->pagev[0]->page;
1690 buffer = kmap_atomic(page);
1691
1692 len = sctx->sectorsize;
1693 index = 0;
1694 for (;;) {
1695 u64 l = min_t(u64, len, PAGE_SIZE);
1696
1697 crc = btrfs_csum_data(buffer, crc, l);
1698 kunmap_atomic(buffer);
1699 len -= l;
1700 if (len == 0)
1701 break;
1702 index++;
1703 BUG_ON(index >= sblock->page_count);
1704 BUG_ON(!sblock->pagev[index]->page);
1705 page = sblock->pagev[index]->page;
1706 buffer = kmap_atomic(page);
1707 }
1708
1709 btrfs_csum_final(crc, csum);
1710 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1711 fail = 1;
1712
1713 return fail;
1714}
1715
1716static int scrub_checksum_tree_block(struct scrub_block *sblock)
1717{
1718 struct scrub_ctx *sctx = sblock->sctx;
1719 struct btrfs_header *h;
1720 struct btrfs_root *root = sctx->dev_root;
1721 struct btrfs_fs_info *fs_info = root->fs_info;
1722 u8 calculated_csum[BTRFS_CSUM_SIZE];
1723 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1724 struct page *page;
1725 void *mapped_buffer;
1726 u64 mapped_size;
1727 void *p;
1728 u32 crc = ~(u32)0;
1729 int fail = 0;
1730 int crc_fail = 0;
1731 u64 len;
1732 int index;
1733
1734 BUG_ON(sblock->page_count < 1);
1735 page = sblock->pagev[0]->page;
1736 mapped_buffer = kmap_atomic(page);
1737 h = (struct btrfs_header *)mapped_buffer;
1738 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1739
1740 /*
1741 * we don't use the getter functions here, as we
1742 * a) don't have an extent buffer and
1743 * b) the page is already kmapped
1744 */
1745
1746 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1747 ++fail;
1748
1749 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1750 ++fail;
1751
1752 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1753 ++fail;
1754
1755 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1756 BTRFS_UUID_SIZE))
1757 ++fail;
1758
1759 WARN_ON(sctx->nodesize != sctx->leafsize);
1760 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1761 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1762 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1763 index = 0;
1764 for (;;) {
1765 u64 l = min_t(u64, len, mapped_size);
1766
1767 crc = btrfs_csum_data(p, crc, l);
1768 kunmap_atomic(mapped_buffer);
1769 len -= l;
1770 if (len == 0)
1771 break;
1772 index++;
1773 BUG_ON(index >= sblock->page_count);
1774 BUG_ON(!sblock->pagev[index]->page);
1775 page = sblock->pagev[index]->page;
1776 mapped_buffer = kmap_atomic(page);
1777 mapped_size = PAGE_SIZE;
1778 p = mapped_buffer;
1779 }
1780
1781 btrfs_csum_final(crc, calculated_csum);
1782 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1783 ++crc_fail;
1784
1785 return fail || crc_fail;
1786}
1787
1788static int scrub_checksum_super(struct scrub_block *sblock)
1789{
1790 struct btrfs_super_block *s;
1791 struct scrub_ctx *sctx = sblock->sctx;
1792 struct btrfs_root *root = sctx->dev_root;
1793 struct btrfs_fs_info *fs_info = root->fs_info;
1794 u8 calculated_csum[BTRFS_CSUM_SIZE];
1795 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1796 struct page *page;
1797 void *mapped_buffer;
1798 u64 mapped_size;
1799 void *p;
1800 u32 crc = ~(u32)0;
1801 int fail_gen = 0;
1802 int fail_cor = 0;
1803 u64 len;
1804 int index;
1805
1806 BUG_ON(sblock->page_count < 1);
1807 page = sblock->pagev[0]->page;
1808 mapped_buffer = kmap_atomic(page);
1809 s = (struct btrfs_super_block *)mapped_buffer;
1810 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1811
1812 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1813 ++fail_cor;
1814
1815 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1816 ++fail_gen;
1817
1818 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1819 ++fail_cor;
1820
1821 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1822 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1823 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1824 index = 0;
1825 for (;;) {
1826 u64 l = min_t(u64, len, mapped_size);
1827
1828 crc = btrfs_csum_data(p, crc, l);
1829 kunmap_atomic(mapped_buffer);
1830 len -= l;
1831 if (len == 0)
1832 break;
1833 index++;
1834 BUG_ON(index >= sblock->page_count);
1835 BUG_ON(!sblock->pagev[index]->page);
1836 page = sblock->pagev[index]->page;
1837 mapped_buffer = kmap_atomic(page);
1838 mapped_size = PAGE_SIZE;
1839 p = mapped_buffer;
1840 }
1841
1842 btrfs_csum_final(crc, calculated_csum);
1843 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1844 ++fail_cor;
1845
1846 if (fail_cor + fail_gen) {
1847 /*
1848 * if we find an error in a super block, we just report it.
1849 * They will get written with the next transaction commit
1850 * anyway
1851 */
1852 spin_lock(&sctx->stat_lock);
1853 ++sctx->stat.super_errors;
1854 spin_unlock(&sctx->stat_lock);
1855 if (fail_cor)
1856 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1857 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1858 else
1859 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1860 BTRFS_DEV_STAT_GENERATION_ERRS);
1861 }
1862
1863 return fail_cor + fail_gen;
1864}
1865
1866static void scrub_block_get(struct scrub_block *sblock)
1867{
1868 atomic_inc(&sblock->ref_count);
1869}
1870
1871static void scrub_block_put(struct scrub_block *sblock)
1872{
1873 if (atomic_dec_and_test(&sblock->ref_count)) {
1874 int i;
1875
1876 for (i = 0; i < sblock->page_count; i++)
1877 scrub_page_put(sblock->pagev[i]);
1878 kfree(sblock);
1879 }
1880}
1881
1882static void scrub_page_get(struct scrub_page *spage)
1883{
1884 atomic_inc(&spage->ref_count);
1885}
1886
1887static void scrub_page_put(struct scrub_page *spage)
1888{
1889 if (atomic_dec_and_test(&spage->ref_count)) {
1890 if (spage->page)
1891 __free_page(spage->page);
1892 kfree(spage);
1893 }
1894}
1895
1896static void scrub_submit(struct scrub_ctx *sctx)
1897{
1898 struct scrub_bio *sbio;
1899
1900 if (sctx->curr == -1)
1901 return;
1902
1903 sbio = sctx->bios[sctx->curr];
1904 sctx->curr = -1;
1905 scrub_pending_bio_inc(sctx);
1906
1907 if (!sbio->bio->bi_bdev) {
1908 /*
1909 * this case should not happen. If btrfs_map_block() is
1910 * wrong, it could happen for dev-replace operations on
1911 * missing devices when no mirrors are available, but in
1912 * this case it should already fail the mount.
1913 * This case is handled correctly (but _very_ slowly).
1914 */
1915 printk_ratelimited(KERN_WARNING
1916 "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1917 bio_endio(sbio->bio, -EIO);
1918 } else {
1919 btrfsic_submit_bio(READ, sbio->bio);
1920 }
1921}
1922
1923static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1924 struct scrub_page *spage)
1925{
1926 struct scrub_block *sblock = spage->sblock;
1927 struct scrub_bio *sbio;
1928 int ret;
1929
1930again:
1931 /*
1932 * grab a fresh bio or wait for one to become available
1933 */
1934 while (sctx->curr == -1) {
1935 spin_lock(&sctx->list_lock);
1936 sctx->curr = sctx->first_free;
1937 if (sctx->curr != -1) {
1938 sctx->first_free = sctx->bios[sctx->curr]->next_free;
1939 sctx->bios[sctx->curr]->next_free = -1;
1940 sctx->bios[sctx->curr]->page_count = 0;
1941 spin_unlock(&sctx->list_lock);
1942 } else {
1943 spin_unlock(&sctx->list_lock);
1944 wait_event(sctx->list_wait, sctx->first_free != -1);
1945 }
1946 }
1947 sbio = sctx->bios[sctx->curr];
1948 if (sbio->page_count == 0) {
1949 struct bio *bio;
1950
1951 sbio->physical = spage->physical;
1952 sbio->logical = spage->logical;
1953 sbio->dev = spage->dev;
1954 bio = sbio->bio;
1955 if (!bio) {
1956 bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1957 if (!bio)
1958 return -ENOMEM;
1959 sbio->bio = bio;
1960 }
1961
1962 bio->bi_private = sbio;
1963 bio->bi_end_io = scrub_bio_end_io;
1964 bio->bi_bdev = sbio->dev->bdev;
1965 bio->bi_iter.bi_sector = sbio->physical >> 9;
1966 sbio->err = 0;
1967 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1968 spage->physical ||
1969 sbio->logical + sbio->page_count * PAGE_SIZE !=
1970 spage->logical ||
1971 sbio->dev != spage->dev) {
1972 scrub_submit(sctx);
1973 goto again;
1974 }
1975
1976 sbio->pagev[sbio->page_count] = spage;
1977 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1978 if (ret != PAGE_SIZE) {
1979 if (sbio->page_count < 1) {
1980 bio_put(sbio->bio);
1981 sbio->bio = NULL;
1982 return -EIO;
1983 }
1984 scrub_submit(sctx);
1985 goto again;
1986 }
1987
1988 scrub_block_get(sblock); /* one for the page added to the bio */
1989 atomic_inc(&sblock->outstanding_pages);
1990 sbio->page_count++;
1991 if (sbio->page_count == sctx->pages_per_rd_bio)
1992 scrub_submit(sctx);
1993
1994 return 0;
1995}
1996
1997static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1998 u64 physical, struct btrfs_device *dev, u64 flags,
1999 u64 gen, int mirror_num, u8 *csum, int force,
2000 u64 physical_for_dev_replace)
2001{
2002 struct scrub_block *sblock;
2003 int index;
2004
2005 sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2006 if (!sblock) {
2007 spin_lock(&sctx->stat_lock);
2008 sctx->stat.malloc_errors++;
2009 spin_unlock(&sctx->stat_lock);
2010 return -ENOMEM;
2011 }
2012
2013 /* one ref inside this function, plus one for each page added to
2014 * a bio later on */
2015 atomic_set(&sblock->ref_count, 1);
2016 sblock->sctx = sctx;
2017 sblock->no_io_error_seen = 1;
2018
2019 for (index = 0; len > 0; index++) {
2020 struct scrub_page *spage;
2021 u64 l = min_t(u64, len, PAGE_SIZE);
2022
2023 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2024 if (!spage) {
2025leave_nomem:
2026 spin_lock(&sctx->stat_lock);
2027 sctx->stat.malloc_errors++;
2028 spin_unlock(&sctx->stat_lock);
2029 scrub_block_put(sblock);
2030 return -ENOMEM;
2031 }
2032 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2033 scrub_page_get(spage);
2034 sblock->pagev[index] = spage;
2035 spage->sblock = sblock;
2036 spage->dev = dev;
2037 spage->flags = flags;
2038 spage->generation = gen;
2039 spage->logical = logical;
2040 spage->physical = physical;
2041 spage->physical_for_dev_replace = physical_for_dev_replace;
2042 spage->mirror_num = mirror_num;
2043 if (csum) {
2044 spage->have_csum = 1;
2045 memcpy(spage->csum, csum, sctx->csum_size);
2046 } else {
2047 spage->have_csum = 0;
2048 }
2049 sblock->page_count++;
2050 spage->page = alloc_page(GFP_NOFS);
2051 if (!spage->page)
2052 goto leave_nomem;
2053 len -= l;
2054 logical += l;
2055 physical += l;
2056 physical_for_dev_replace += l;
2057 }
2058
2059 WARN_ON(sblock->page_count == 0);
2060 for (index = 0; index < sblock->page_count; index++) {
2061 struct scrub_page *spage = sblock->pagev[index];
2062 int ret;
2063
2064 ret = scrub_add_page_to_rd_bio(sctx, spage);
2065 if (ret) {
2066 scrub_block_put(sblock);
2067 return ret;
2068 }
2069 }
2070
2071 if (force)
2072 scrub_submit(sctx);
2073
2074 /* last one frees, either here or in bio completion for last page */
2075 scrub_block_put(sblock);
2076 return 0;
2077}
2078
2079static void scrub_bio_end_io(struct bio *bio, int err)
2080{
2081 struct scrub_bio *sbio = bio->bi_private;
2082 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2083
2084 sbio->err = err;
2085 sbio->bio = bio;
2086
2087 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2088}
2089
2090static void scrub_bio_end_io_worker(struct btrfs_work *work)
2091{
2092 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2093 struct scrub_ctx *sctx = sbio->sctx;
2094 int i;
2095
2096 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2097 if (sbio->err) {
2098 for (i = 0; i < sbio->page_count; i++) {
2099 struct scrub_page *spage = sbio->pagev[i];
2100
2101 spage->io_error = 1;
2102 spage->sblock->no_io_error_seen = 0;
2103 }
2104 }
2105
2106 /* now complete the scrub_block items that have all pages completed */
2107 for (i = 0; i < sbio->page_count; i++) {
2108 struct scrub_page *spage = sbio->pagev[i];
2109 struct scrub_block *sblock = spage->sblock;
2110
2111 if (atomic_dec_and_test(&sblock->outstanding_pages))
2112 scrub_block_complete(sblock);
2113 scrub_block_put(sblock);
2114 }
2115
2116 bio_put(sbio->bio);
2117 sbio->bio = NULL;
2118 spin_lock(&sctx->list_lock);
2119 sbio->next_free = sctx->first_free;
2120 sctx->first_free = sbio->index;
2121 spin_unlock(&sctx->list_lock);
2122
2123 if (sctx->is_dev_replace &&
2124 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2125 mutex_lock(&sctx->wr_ctx.wr_lock);
2126 scrub_wr_submit(sctx);
2127 mutex_unlock(&sctx->wr_ctx.wr_lock);
2128 }
2129
2130 scrub_pending_bio_dec(sctx);
2131}
2132
2133static void scrub_block_complete(struct scrub_block *sblock)
2134{
2135 if (!sblock->no_io_error_seen) {
2136 scrub_handle_errored_block(sblock);
2137 } else {
2138 /*
2139 * if has checksum error, write via repair mechanism in
2140 * dev replace case, otherwise write here in dev replace
2141 * case.
2142 */
2143 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2144 scrub_write_block_to_dev_replace(sblock);
2145 }
2146}
2147
2148static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2149 u8 *csum)
2150{
2151 struct btrfs_ordered_sum *sum = NULL;
2152 unsigned long index;
2153 unsigned long num_sectors;
2154
2155 while (!list_empty(&sctx->csum_list)) {
2156 sum = list_first_entry(&sctx->csum_list,
2157 struct btrfs_ordered_sum, list);
2158 if (sum->bytenr > logical)
2159 return 0;
2160 if (sum->bytenr + sum->len > logical)
2161 break;
2162
2163 ++sctx->stat.csum_discards;
2164 list_del(&sum->list);
2165 kfree(sum);
2166 sum = NULL;
2167 }
2168 if (!sum)
2169 return 0;
2170
2171 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2172 num_sectors = sum->len / sctx->sectorsize;
2173 memcpy(csum, sum->sums + index, sctx->csum_size);
2174 if (index == num_sectors - 1) {
2175 list_del(&sum->list);
2176 kfree(sum);
2177 }
2178 return 1;
2179}
2180
2181/* scrub extent tries to collect up to 64 kB for each bio */
2182static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2183 u64 physical, struct btrfs_device *dev, u64 flags,
2184 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2185{
2186 int ret;
2187 u8 csum[BTRFS_CSUM_SIZE];
2188 u32 blocksize;
2189
2190 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2191 blocksize = sctx->sectorsize;
2192 spin_lock(&sctx->stat_lock);
2193 sctx->stat.data_extents_scrubbed++;
2194 sctx->stat.data_bytes_scrubbed += len;
2195 spin_unlock(&sctx->stat_lock);
2196 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2197 WARN_ON(sctx->nodesize != sctx->leafsize);
2198 blocksize = sctx->nodesize;
2199 spin_lock(&sctx->stat_lock);
2200 sctx->stat.tree_extents_scrubbed++;
2201 sctx->stat.tree_bytes_scrubbed += len;
2202 spin_unlock(&sctx->stat_lock);
2203 } else {
2204 blocksize = sctx->sectorsize;
2205 WARN_ON(1);
2206 }
2207
2208 while (len) {
2209 u64 l = min_t(u64, len, blocksize);
2210 int have_csum = 0;
2211
2212 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2213 /* push csums to sbio */
2214 have_csum = scrub_find_csum(sctx, logical, l, csum);
2215 if (have_csum == 0)
2216 ++sctx->stat.no_csum;
2217 if (sctx->is_dev_replace && !have_csum) {
2218 ret = copy_nocow_pages(sctx, logical, l,
2219 mirror_num,
2220 physical_for_dev_replace);
2221 goto behind_scrub_pages;
2222 }
2223 }
2224 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2225 mirror_num, have_csum ? csum : NULL, 0,
2226 physical_for_dev_replace);
2227behind_scrub_pages:
2228 if (ret)
2229 return ret;
2230 len -= l;
2231 logical += l;
2232 physical += l;
2233 physical_for_dev_replace += l;
2234 }
2235 return 0;
2236}
2237
2238/*
2239 * Given a physical address, this will calculate it's
2240 * logical offset. if this is a parity stripe, it will return
2241 * the most left data stripe's logical offset.
2242 *
2243 * return 0 if it is a data stripe, 1 means parity stripe.
2244 */
2245static int get_raid56_logic_offset(u64 physical, int num,
2246 struct map_lookup *map, u64 *offset)
2247{
2248 int i;
2249 int j = 0;
2250 u64 stripe_nr;
2251 u64 last_offset;
2252 int stripe_index;
2253 int rot;
2254
2255 last_offset = (physical - map->stripes[num].physical) *
2256 nr_data_stripes(map);
2257 *offset = last_offset;
2258 for (i = 0; i < nr_data_stripes(map); i++) {
2259 *offset = last_offset + i * map->stripe_len;
2260
2261 stripe_nr = *offset;
2262 do_div(stripe_nr, map->stripe_len);
2263 do_div(stripe_nr, nr_data_stripes(map));
2264
2265 /* Work out the disk rotation on this stripe-set */
2266 rot = do_div(stripe_nr, map->num_stripes);
2267 /* calculate which stripe this data locates */
2268 rot += i;
2269 stripe_index = rot % map->num_stripes;
2270 if (stripe_index == num)
2271 return 0;
2272 if (stripe_index < num)
2273 j++;
2274 }
2275 *offset = last_offset + j * map->stripe_len;
2276 return 1;
2277}
2278
2279static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2280 struct map_lookup *map,
2281 struct btrfs_device *scrub_dev,
2282 int num, u64 base, u64 length,
2283 int is_dev_replace)
2284{
2285 struct btrfs_path *path;
2286 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2287 struct btrfs_root *root = fs_info->extent_root;
2288 struct btrfs_root *csum_root = fs_info->csum_root;
2289 struct btrfs_extent_item *extent;
2290 struct blk_plug plug;
2291 u64 flags;
2292 int ret;
2293 int slot;
2294 u64 nstripes;
2295 struct extent_buffer *l;
2296 struct btrfs_key key;
2297 u64 physical;
2298 u64 logical;
2299 u64 logic_end;
2300 u64 physical_end;
2301 u64 generation;
2302 int mirror_num;
2303 struct reada_control *reada1;
2304 struct reada_control *reada2;
2305 struct btrfs_key key_start;
2306 struct btrfs_key key_end;
2307 u64 increment = map->stripe_len;
2308 u64 offset;
2309 u64 extent_logical;
2310 u64 extent_physical;
2311 u64 extent_len;
2312 struct btrfs_device *extent_dev;
2313 int extent_mirror_num;
2314 int stop_loop = 0;
2315
2316 nstripes = length;
2317 physical = map->stripes[num].physical;
2318 offset = 0;
2319 do_div(nstripes, map->stripe_len);
2320 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2321 offset = map->stripe_len * num;
2322 increment = map->stripe_len * map->num_stripes;
2323 mirror_num = 1;
2324 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2325 int factor = map->num_stripes / map->sub_stripes;
2326 offset = map->stripe_len * (num / map->sub_stripes);
2327 increment = map->stripe_len * factor;
2328 mirror_num = num % map->sub_stripes + 1;
2329 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2330 increment = map->stripe_len;
2331 mirror_num = num % map->num_stripes + 1;
2332 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2333 increment = map->stripe_len;
2334 mirror_num = num % map->num_stripes + 1;
2335 } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2336 BTRFS_BLOCK_GROUP_RAID6)) {
2337 get_raid56_logic_offset(physical, num, map, &offset);
2338 increment = map->stripe_len * nr_data_stripes(map);
2339 mirror_num = 1;
2340 } else {
2341 increment = map->stripe_len;
2342 mirror_num = 1;
2343 }
2344
2345 path = btrfs_alloc_path();
2346 if (!path)
2347 return -ENOMEM;
2348
2349 /*
2350 * work on commit root. The related disk blocks are static as
2351 * long as COW is applied. This means, it is save to rewrite
2352 * them to repair disk errors without any race conditions
2353 */
2354 path->search_commit_root = 1;
2355 path->skip_locking = 1;
2356
2357 /*
2358 * trigger the readahead for extent tree csum tree and wait for
2359 * completion. During readahead, the scrub is officially paused
2360 * to not hold off transaction commits
2361 */
2362 logical = base + offset;
2363 physical_end = physical + nstripes * map->stripe_len;
2364 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2365 BTRFS_BLOCK_GROUP_RAID6)) {
2366 get_raid56_logic_offset(physical_end, num,
2367 map, &logic_end);
2368 logic_end += base;
2369 } else {
2370 logic_end = logical + increment * nstripes;
2371 }
2372 wait_event(sctx->list_wait,
2373 atomic_read(&sctx->bios_in_flight) == 0);
2374 scrub_blocked_if_needed(fs_info);
2375
2376 /* FIXME it might be better to start readahead at commit root */
2377 key_start.objectid = logical;
2378 key_start.type = BTRFS_EXTENT_ITEM_KEY;
2379 key_start.offset = (u64)0;
2380 key_end.objectid = logic_end;
2381 key_end.type = BTRFS_METADATA_ITEM_KEY;
2382 key_end.offset = (u64)-1;
2383 reada1 = btrfs_reada_add(root, &key_start, &key_end);
2384
2385 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2386 key_start.type = BTRFS_EXTENT_CSUM_KEY;
2387 key_start.offset = logical;
2388 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2389 key_end.type = BTRFS_EXTENT_CSUM_KEY;
2390 key_end.offset = logic_end;
2391 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2392
2393 if (!IS_ERR(reada1))
2394 btrfs_reada_wait(reada1);
2395 if (!IS_ERR(reada2))
2396 btrfs_reada_wait(reada2);
2397
2398
2399 /*
2400 * collect all data csums for the stripe to avoid seeking during
2401 * the scrub. This might currently (crc32) end up to be about 1MB
2402 */
2403 blk_start_plug(&plug);
2404
2405 /*
2406 * now find all extents for each stripe and scrub them
2407 */
2408 ret = 0;
2409 while (physical < physical_end) {
2410 /* for raid56, we skip parity stripe */
2411 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2412 BTRFS_BLOCK_GROUP_RAID6)) {
2413 ret = get_raid56_logic_offset(physical, num,
2414 map, &logical);
2415 logical += base;
2416 if (ret)
2417 goto skip;
2418 }
2419 /*
2420 * canceled?
2421 */
2422 if (atomic_read(&fs_info->scrub_cancel_req) ||
2423 atomic_read(&sctx->cancel_req)) {
2424 ret = -ECANCELED;
2425 goto out;
2426 }
2427 /*
2428 * check to see if we have to pause
2429 */
2430 if (atomic_read(&fs_info->scrub_pause_req)) {
2431 /* push queued extents */
2432 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2433 scrub_submit(sctx);
2434 mutex_lock(&sctx->wr_ctx.wr_lock);
2435 scrub_wr_submit(sctx);
2436 mutex_unlock(&sctx->wr_ctx.wr_lock);
2437 wait_event(sctx->list_wait,
2438 atomic_read(&sctx->bios_in_flight) == 0);
2439 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2440 scrub_blocked_if_needed(fs_info);
2441 }
2442
2443 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2444 key.type = BTRFS_METADATA_ITEM_KEY;
2445 else
2446 key.type = BTRFS_EXTENT_ITEM_KEY;
2447 key.objectid = logical;
2448 key.offset = (u64)-1;
2449
2450 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2451 if (ret < 0)
2452 goto out;
2453
2454 if (ret > 0) {
2455 ret = btrfs_previous_extent_item(root, path, 0);
2456 if (ret < 0)
2457 goto out;
2458 if (ret > 0) {
2459 /* there's no smaller item, so stick with the
2460 * larger one */
2461 btrfs_release_path(path);
2462 ret = btrfs_search_slot(NULL, root, &key,
2463 path, 0, 0);
2464 if (ret < 0)
2465 goto out;
2466 }
2467 }
2468
2469 stop_loop = 0;
2470 while (1) {
2471 u64 bytes;
2472
2473 l = path->nodes[0];
2474 slot = path->slots[0];
2475 if (slot >= btrfs_header_nritems(l)) {
2476 ret = btrfs_next_leaf(root, path);
2477 if (ret == 0)
2478 continue;
2479 if (ret < 0)
2480 goto out;
2481
2482 stop_loop = 1;
2483 break;
2484 }
2485 btrfs_item_key_to_cpu(l, &key, slot);
2486
2487 if (key.type == BTRFS_METADATA_ITEM_KEY)
2488 bytes = root->leafsize;
2489 else
2490 bytes = key.offset;
2491
2492 if (key.objectid + bytes <= logical)
2493 goto next;
2494
2495 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2496 key.type != BTRFS_METADATA_ITEM_KEY)
2497 goto next;
2498
2499 if (key.objectid >= logical + map->stripe_len) {
2500 /* out of this device extent */
2501 if (key.objectid >= logic_end)
2502 stop_loop = 1;
2503 break;
2504 }
2505
2506 extent = btrfs_item_ptr(l, slot,
2507 struct btrfs_extent_item);
2508 flags = btrfs_extent_flags(l, extent);
2509 generation = btrfs_extent_generation(l, extent);
2510
2511 if (key.objectid < logical &&
2512 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2513 btrfs_err(fs_info,
2514 "scrub: tree block %llu spanning "
2515 "stripes, ignored. logical=%llu",
2516 key.objectid, logical);
2517 goto next;
2518 }
2519
2520again:
2521 extent_logical = key.objectid;
2522 extent_len = bytes;
2523
2524 /*
2525 * trim extent to this stripe
2526 */
2527 if (extent_logical < logical) {
2528 extent_len -= logical - extent_logical;
2529 extent_logical = logical;
2530 }
2531 if (extent_logical + extent_len >
2532 logical + map->stripe_len) {
2533 extent_len = logical + map->stripe_len -
2534 extent_logical;
2535 }
2536
2537 extent_physical = extent_logical - logical + physical;
2538 extent_dev = scrub_dev;
2539 extent_mirror_num = mirror_num;
2540 if (is_dev_replace)
2541 scrub_remap_extent(fs_info, extent_logical,
2542 extent_len, &extent_physical,
2543 &extent_dev,
2544 &extent_mirror_num);
2545
2546 ret = btrfs_lookup_csums_range(csum_root, logical,
2547 logical + map->stripe_len - 1,
2548 &sctx->csum_list, 1);
2549 if (ret)
2550 goto out;
2551
2552 ret = scrub_extent(sctx, extent_logical, extent_len,
2553 extent_physical, extent_dev, flags,
2554 generation, extent_mirror_num,
2555 extent_logical - logical + physical);
2556 if (ret)
2557 goto out;
2558
2559 scrub_free_csums(sctx);
2560 if (extent_logical + extent_len <
2561 key.objectid + bytes) {
2562 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2563 BTRFS_BLOCK_GROUP_RAID6)) {
2564 /*
2565 * loop until we find next data stripe
2566 * or we have finished all stripes.
2567 */
2568 do {
2569 physical += map->stripe_len;
2570 ret = get_raid56_logic_offset(
2571 physical, num,
2572 map, &logical);
2573 logical += base;
2574 } while (physical < physical_end && ret);
2575 } else {
2576 physical += map->stripe_len;
2577 logical += increment;
2578 }
2579 if (logical < key.objectid + bytes) {
2580 cond_resched();
2581 goto again;
2582 }
2583
2584 if (physical >= physical_end) {
2585 stop_loop = 1;
2586 break;
2587 }
2588 }
2589next:
2590 path->slots[0]++;
2591 }
2592 btrfs_release_path(path);
2593skip:
2594 logical += increment;
2595 physical += map->stripe_len;
2596 spin_lock(&sctx->stat_lock);
2597 if (stop_loop)
2598 sctx->stat.last_physical = map->stripes[num].physical +
2599 length;
2600 else
2601 sctx->stat.last_physical = physical;
2602 spin_unlock(&sctx->stat_lock);
2603 if (stop_loop)
2604 break;
2605 }
2606out:
2607 /* push queued extents */
2608 scrub_submit(sctx);
2609 mutex_lock(&sctx->wr_ctx.wr_lock);
2610 scrub_wr_submit(sctx);
2611 mutex_unlock(&sctx->wr_ctx.wr_lock);
2612
2613 blk_finish_plug(&plug);
2614 btrfs_free_path(path);
2615 return ret < 0 ? ret : 0;
2616}
2617
2618static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2619 struct btrfs_device *scrub_dev,
2620 u64 chunk_tree, u64 chunk_objectid,
2621 u64 chunk_offset, u64 length,
2622 u64 dev_offset, int is_dev_replace)
2623{
2624 struct btrfs_mapping_tree *map_tree =
2625 &sctx->dev_root->fs_info->mapping_tree;
2626 struct map_lookup *map;
2627 struct extent_map *em;
2628 int i;
2629 int ret = 0;
2630
2631 read_lock(&map_tree->map_tree.lock);
2632 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2633 read_unlock(&map_tree->map_tree.lock);
2634
2635 if (!em)
2636 return -EINVAL;
2637
2638 map = (struct map_lookup *)em->bdev;
2639 if (em->start != chunk_offset)
2640 goto out;
2641
2642 if (em->len < length)
2643 goto out;
2644
2645 for (i = 0; i < map->num_stripes; ++i) {
2646 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2647 map->stripes[i].physical == dev_offset) {
2648 ret = scrub_stripe(sctx, map, scrub_dev, i,
2649 chunk_offset, length,
2650 is_dev_replace);
2651 if (ret)
2652 goto out;
2653 }
2654 }
2655out:
2656 free_extent_map(em);
2657
2658 return ret;
2659}
2660
2661static noinline_for_stack
2662int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2663 struct btrfs_device *scrub_dev, u64 start, u64 end,
2664 int is_dev_replace)
2665{
2666 struct btrfs_dev_extent *dev_extent = NULL;
2667 struct btrfs_path *path;
2668 struct btrfs_root *root = sctx->dev_root;
2669 struct btrfs_fs_info *fs_info = root->fs_info;
2670 u64 length;
2671 u64 chunk_tree;
2672 u64 chunk_objectid;
2673 u64 chunk_offset;
2674 int ret;
2675 int slot;
2676 struct extent_buffer *l;
2677 struct btrfs_key key;
2678 struct btrfs_key found_key;
2679 struct btrfs_block_group_cache *cache;
2680 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2681
2682 path = btrfs_alloc_path();
2683 if (!path)
2684 return -ENOMEM;
2685
2686 path->reada = 2;
2687 path->search_commit_root = 1;
2688 path->skip_locking = 1;
2689
2690 key.objectid = scrub_dev->devid;
2691 key.offset = 0ull;
2692 key.type = BTRFS_DEV_EXTENT_KEY;
2693
2694 while (1) {
2695 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2696 if (ret < 0)
2697 break;
2698 if (ret > 0) {
2699 if (path->slots[0] >=
2700 btrfs_header_nritems(path->nodes[0])) {
2701 ret = btrfs_next_leaf(root, path);
2702 if (ret)
2703 break;
2704 }
2705 }
2706
2707 l = path->nodes[0];
2708 slot = path->slots[0];
2709
2710 btrfs_item_key_to_cpu(l, &found_key, slot);
2711
2712 if (found_key.objectid != scrub_dev->devid)
2713 break;
2714
2715 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2716 break;
2717
2718 if (found_key.offset >= end)
2719 break;
2720
2721 if (found_key.offset < key.offset)
2722 break;
2723
2724 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2725 length = btrfs_dev_extent_length(l, dev_extent);
2726
2727 if (found_key.offset + length <= start) {
2728 key.offset = found_key.offset + length;
2729 btrfs_release_path(path);
2730 continue;
2731 }
2732
2733 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2734 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2735 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2736
2737 /*
2738 * get a reference on the corresponding block group to prevent
2739 * the chunk from going away while we scrub it
2740 */
2741 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2742 if (!cache) {
2743 ret = -ENOENT;
2744 break;
2745 }
2746 dev_replace->cursor_right = found_key.offset + length;
2747 dev_replace->cursor_left = found_key.offset;
2748 dev_replace->item_needs_writeback = 1;
2749 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2750 chunk_offset, length, found_key.offset,
2751 is_dev_replace);
2752
2753 /*
2754 * flush, submit all pending read and write bios, afterwards
2755 * wait for them.
2756 * Note that in the dev replace case, a read request causes
2757 * write requests that are submitted in the read completion
2758 * worker. Therefore in the current situation, it is required
2759 * that all write requests are flushed, so that all read and
2760 * write requests are really completed when bios_in_flight
2761 * changes to 0.
2762 */
2763 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2764 scrub_submit(sctx);
2765 mutex_lock(&sctx->wr_ctx.wr_lock);
2766 scrub_wr_submit(sctx);
2767 mutex_unlock(&sctx->wr_ctx.wr_lock);
2768
2769 wait_event(sctx->list_wait,
2770 atomic_read(&sctx->bios_in_flight) == 0);
2771 atomic_inc(&fs_info->scrubs_paused);
2772 wake_up(&fs_info->scrub_pause_wait);
2773
2774 /*
2775 * must be called before we decrease @scrub_paused.
2776 * make sure we don't block transaction commit while
2777 * we are waiting pending workers finished.
2778 */
2779 wait_event(sctx->list_wait,
2780 atomic_read(&sctx->workers_pending) == 0);
2781 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2782
2783 mutex_lock(&fs_info->scrub_lock);
2784 __scrub_blocked_if_needed(fs_info);
2785 atomic_dec(&fs_info->scrubs_paused);
2786 mutex_unlock(&fs_info->scrub_lock);
2787 wake_up(&fs_info->scrub_pause_wait);
2788
2789 btrfs_put_block_group(cache);
2790 if (ret)
2791 break;
2792 if (is_dev_replace &&
2793 atomic64_read(&dev_replace->num_write_errors) > 0) {
2794 ret = -EIO;
2795 break;
2796 }
2797 if (sctx->stat.malloc_errors > 0) {
2798 ret = -ENOMEM;
2799 break;
2800 }
2801
2802 dev_replace->cursor_left = dev_replace->cursor_right;
2803 dev_replace->item_needs_writeback = 1;
2804
2805 key.offset = found_key.offset + length;
2806 btrfs_release_path(path);
2807 }
2808
2809 btrfs_free_path(path);
2810
2811 /*
2812 * ret can still be 1 from search_slot or next_leaf,
2813 * that's not an error
2814 */
2815 return ret < 0 ? ret : 0;
2816}
2817
2818static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2819 struct btrfs_device *scrub_dev)
2820{
2821 int i;
2822 u64 bytenr;
2823 u64 gen;
2824 int ret;
2825 struct btrfs_root *root = sctx->dev_root;
2826
2827 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2828 return -EIO;
2829
2830 gen = root->fs_info->last_trans_committed;
2831
2832 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2833 bytenr = btrfs_sb_offset(i);
2834 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2835 break;
2836
2837 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2838 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2839 NULL, 1, bytenr);
2840 if (ret)
2841 return ret;
2842 }
2843 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2844
2845 return 0;
2846}
2847
2848/*
2849 * get a reference count on fs_info->scrub_workers. start worker if necessary
2850 */
2851static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2852 int is_dev_replace)
2853{
2854 int ret = 0;
2855 int flags = WQ_FREEZABLE | WQ_UNBOUND;
2856 int max_active = fs_info->thread_pool_size;
2857
2858 if (fs_info->scrub_workers_refcnt == 0) {
2859 if (is_dev_replace)
2860 fs_info->scrub_workers =
2861 btrfs_alloc_workqueue("btrfs-scrub", flags,
2862 1, 4);
2863 else
2864 fs_info->scrub_workers =
2865 btrfs_alloc_workqueue("btrfs-scrub", flags,
2866 max_active, 4);
2867 if (!fs_info->scrub_workers) {
2868 ret = -ENOMEM;
2869 goto out;
2870 }
2871 fs_info->scrub_wr_completion_workers =
2872 btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
2873 max_active, 2);
2874 if (!fs_info->scrub_wr_completion_workers) {
2875 ret = -ENOMEM;
2876 goto out;
2877 }
2878 fs_info->scrub_nocow_workers =
2879 btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
2880 if (!fs_info->scrub_nocow_workers) {
2881 ret = -ENOMEM;
2882 goto out;
2883 }
2884 }
2885 ++fs_info->scrub_workers_refcnt;
2886out:
2887 return ret;
2888}
2889
2890static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2891{
2892 if (--fs_info->scrub_workers_refcnt == 0) {
2893 btrfs_destroy_workqueue(fs_info->scrub_workers);
2894 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
2895 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
2896 }
2897 WARN_ON(fs_info->scrub_workers_refcnt < 0);
2898}
2899
2900int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2901 u64 end, struct btrfs_scrub_progress *progress,
2902 int readonly, int is_dev_replace)
2903{
2904 struct scrub_ctx *sctx;
2905 int ret;
2906 struct btrfs_device *dev;
2907
2908 if (btrfs_fs_closing(fs_info))
2909 return -EINVAL;
2910
2911 /*
2912 * check some assumptions
2913 */
2914 if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2915 btrfs_err(fs_info,
2916 "scrub: size assumption nodesize == leafsize (%d == %d) fails",
2917 fs_info->chunk_root->nodesize,
2918 fs_info->chunk_root->leafsize);
2919 return -EINVAL;
2920 }
2921
2922 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2923 /*
2924 * in this case scrub is unable to calculate the checksum
2925 * the way scrub is implemented. Do not handle this
2926 * situation at all because it won't ever happen.
2927 */
2928 btrfs_err(fs_info,
2929 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2930 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2931 return -EINVAL;
2932 }
2933
2934 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2935 /* not supported for data w/o checksums */
2936 btrfs_err(fs_info,
2937 "scrub: size assumption sectorsize != PAGE_SIZE "
2938 "(%d != %lu) fails",
2939 fs_info->chunk_root->sectorsize, PAGE_SIZE);
2940 return -EINVAL;
2941 }
2942
2943 if (fs_info->chunk_root->nodesize >
2944 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2945 fs_info->chunk_root->sectorsize >
2946 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2947 /*
2948 * would exhaust the array bounds of pagev member in
2949 * struct scrub_block
2950 */
2951 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2952 "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2953 fs_info->chunk_root->nodesize,
2954 SCRUB_MAX_PAGES_PER_BLOCK,
2955 fs_info->chunk_root->sectorsize,
2956 SCRUB_MAX_PAGES_PER_BLOCK);
2957 return -EINVAL;
2958 }
2959
2960
2961 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2962 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2963 if (!dev || (dev->missing && !is_dev_replace)) {
2964 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2965 return -ENODEV;
2966 }
2967
2968 mutex_lock(&fs_info->scrub_lock);
2969 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2970 mutex_unlock(&fs_info->scrub_lock);
2971 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2972 return -EIO;
2973 }
2974
2975 btrfs_dev_replace_lock(&fs_info->dev_replace);
2976 if (dev->scrub_device ||
2977 (!is_dev_replace &&
2978 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2979 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2980 mutex_unlock(&fs_info->scrub_lock);
2981 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2982 return -EINPROGRESS;
2983 }
2984 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2985
2986 ret = scrub_workers_get(fs_info, is_dev_replace);
2987 if (ret) {
2988 mutex_unlock(&fs_info->scrub_lock);
2989 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2990 return ret;
2991 }
2992
2993 sctx = scrub_setup_ctx(dev, is_dev_replace);
2994 if (IS_ERR(sctx)) {
2995 mutex_unlock(&fs_info->scrub_lock);
2996 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2997 scrub_workers_put(fs_info);
2998 return PTR_ERR(sctx);
2999 }
3000 sctx->readonly = readonly;
3001 dev->scrub_device = sctx;
3002 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3003
3004 /*
3005 * checking @scrub_pause_req here, we can avoid
3006 * race between committing transaction and scrubbing.
3007 */
3008 __scrub_blocked_if_needed(fs_info);
3009 atomic_inc(&fs_info->scrubs_running);
3010 mutex_unlock(&fs_info->scrub_lock);
3011
3012 if (!is_dev_replace) {
3013 /*
3014 * by holding device list mutex, we can
3015 * kick off writing super in log tree sync.
3016 */
3017 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3018 ret = scrub_supers(sctx, dev);
3019 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3020 }
3021
3022 if (!ret)
3023 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3024 is_dev_replace);
3025
3026 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3027 atomic_dec(&fs_info->scrubs_running);
3028 wake_up(&fs_info->scrub_pause_wait);
3029
3030 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3031
3032 if (progress)
3033 memcpy(progress, &sctx->stat, sizeof(*progress));
3034
3035 mutex_lock(&fs_info->scrub_lock);
3036 dev->scrub_device = NULL;
3037 scrub_workers_put(fs_info);
3038 mutex_unlock(&fs_info->scrub_lock);
3039
3040 scrub_free_ctx(sctx);
3041
3042 return ret;
3043}
3044
3045void btrfs_scrub_pause(struct btrfs_root *root)
3046{
3047 struct btrfs_fs_info *fs_info = root->fs_info;
3048
3049 mutex_lock(&fs_info->scrub_lock);
3050 atomic_inc(&fs_info->scrub_pause_req);
3051 while (atomic_read(&fs_info->scrubs_paused) !=
3052 atomic_read(&fs_info->scrubs_running)) {
3053 mutex_unlock(&fs_info->scrub_lock);
3054 wait_event(fs_info->scrub_pause_wait,
3055 atomic_read(&fs_info->scrubs_paused) ==
3056 atomic_read(&fs_info->scrubs_running));
3057 mutex_lock(&fs_info->scrub_lock);
3058 }
3059 mutex_unlock(&fs_info->scrub_lock);
3060}
3061
3062void btrfs_scrub_continue(struct btrfs_root *root)
3063{
3064 struct btrfs_fs_info *fs_info = root->fs_info;
3065
3066 atomic_dec(&fs_info->scrub_pause_req);
3067 wake_up(&fs_info->scrub_pause_wait);
3068}
3069
3070int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3071{
3072 mutex_lock(&fs_info->scrub_lock);
3073 if (!atomic_read(&fs_info->scrubs_running)) {
3074 mutex_unlock(&fs_info->scrub_lock);
3075 return -ENOTCONN;
3076 }
3077
3078 atomic_inc(&fs_info->scrub_cancel_req);
3079 while (atomic_read(&fs_info->scrubs_running)) {
3080 mutex_unlock(&fs_info->scrub_lock);
3081 wait_event(fs_info->scrub_pause_wait,
3082 atomic_read(&fs_info->scrubs_running) == 0);
3083 mutex_lock(&fs_info->scrub_lock);
3084 }
3085 atomic_dec(&fs_info->scrub_cancel_req);
3086 mutex_unlock(&fs_info->scrub_lock);
3087
3088 return 0;
3089}
3090
3091int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3092 struct btrfs_device *dev)
3093{
3094 struct scrub_ctx *sctx;
3095
3096 mutex_lock(&fs_info->scrub_lock);
3097 sctx = dev->scrub_device;
3098 if (!sctx) {
3099 mutex_unlock(&fs_info->scrub_lock);
3100 return -ENOTCONN;
3101 }
3102 atomic_inc(&sctx->cancel_req);
3103 while (dev->scrub_device) {
3104 mutex_unlock(&fs_info->scrub_lock);
3105 wait_event(fs_info->scrub_pause_wait,
3106 dev->scrub_device == NULL);
3107 mutex_lock(&fs_info->scrub_lock);
3108 }
3109 mutex_unlock(&fs_info->scrub_lock);
3110
3111 return 0;
3112}
3113
3114int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3115 struct btrfs_scrub_progress *progress)
3116{
3117 struct btrfs_device *dev;
3118 struct scrub_ctx *sctx = NULL;
3119
3120 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3121 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3122 if (dev)
3123 sctx = dev->scrub_device;
3124 if (sctx)
3125 memcpy(progress, &sctx->stat, sizeof(*progress));
3126 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3127
3128 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3129}
3130
3131static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3132 u64 extent_logical, u64 extent_len,
3133 u64 *extent_physical,
3134 struct btrfs_device **extent_dev,
3135 int *extent_mirror_num)
3136{
3137 u64 mapped_length;
3138 struct btrfs_bio *bbio = NULL;
3139 int ret;
3140
3141 mapped_length = extent_len;
3142 ret = btrfs_map_block(fs_info, READ, extent_logical,
3143 &mapped_length, &bbio, 0);
3144 if (ret || !bbio || mapped_length < extent_len ||
3145 !bbio->stripes[0].dev->bdev) {
3146 kfree(bbio);
3147 return;
3148 }
3149
3150 *extent_physical = bbio->stripes[0].physical;
3151 *extent_mirror_num = bbio->mirror_num;
3152 *extent_dev = bbio->stripes[0].dev;
3153 kfree(bbio);
3154}
3155
3156static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3157 struct scrub_wr_ctx *wr_ctx,
3158 struct btrfs_fs_info *fs_info,
3159 struct btrfs_device *dev,
3160 int is_dev_replace)
3161{
3162 WARN_ON(wr_ctx->wr_curr_bio != NULL);
3163
3164 mutex_init(&wr_ctx->wr_lock);
3165 wr_ctx->wr_curr_bio = NULL;
3166 if (!is_dev_replace)
3167 return 0;
3168
3169 WARN_ON(!dev->bdev);
3170 wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3171 bio_get_nr_vecs(dev->bdev));
3172 wr_ctx->tgtdev = dev;
3173 atomic_set(&wr_ctx->flush_all_writes, 0);
3174 return 0;
3175}
3176
3177static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3178{
3179 mutex_lock(&wr_ctx->wr_lock);
3180 kfree(wr_ctx->wr_curr_bio);
3181 wr_ctx->wr_curr_bio = NULL;
3182 mutex_unlock(&wr_ctx->wr_lock);
3183}
3184
3185static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3186 int mirror_num, u64 physical_for_dev_replace)
3187{
3188 struct scrub_copy_nocow_ctx *nocow_ctx;
3189 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3190
3191 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3192 if (!nocow_ctx) {
3193 spin_lock(&sctx->stat_lock);
3194 sctx->stat.malloc_errors++;
3195 spin_unlock(&sctx->stat_lock);
3196 return -ENOMEM;
3197 }
3198
3199 scrub_pending_trans_workers_inc(sctx);
3200
3201 nocow_ctx->sctx = sctx;
3202 nocow_ctx->logical = logical;
3203 nocow_ctx->len = len;
3204 nocow_ctx->mirror_num = mirror_num;
3205 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3206 btrfs_init_work(&nocow_ctx->work, copy_nocow_pages_worker, NULL, NULL);
3207 INIT_LIST_HEAD(&nocow_ctx->inodes);
3208 btrfs_queue_work(fs_info->scrub_nocow_workers,
3209 &nocow_ctx->work);
3210
3211 return 0;
3212}
3213
3214static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3215{
3216 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3217 struct scrub_nocow_inode *nocow_inode;
3218
3219 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3220 if (!nocow_inode)
3221 return -ENOMEM;
3222 nocow_inode->inum = inum;
3223 nocow_inode->offset = offset;
3224 nocow_inode->root = root;
3225 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3226 return 0;
3227}
3228
3229#define COPY_COMPLETE 1
3230
3231static void copy_nocow_pages_worker(struct btrfs_work *work)
3232{
3233 struct scrub_copy_nocow_ctx *nocow_ctx =
3234 container_of(work, struct scrub_copy_nocow_ctx, work);
3235 struct scrub_ctx *sctx = nocow_ctx->sctx;
3236 u64 logical = nocow_ctx->logical;
3237 u64 len = nocow_ctx->len;
3238 int mirror_num = nocow_ctx->mirror_num;
3239 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3240 int ret;
3241 struct btrfs_trans_handle *trans = NULL;
3242 struct btrfs_fs_info *fs_info;
3243 struct btrfs_path *path;
3244 struct btrfs_root *root;
3245 int not_written = 0;
3246
3247 fs_info = sctx->dev_root->fs_info;
3248 root = fs_info->extent_root;
3249
3250 path = btrfs_alloc_path();
3251 if (!path) {
3252 spin_lock(&sctx->stat_lock);
3253 sctx->stat.malloc_errors++;
3254 spin_unlock(&sctx->stat_lock);
3255 not_written = 1;
3256 goto out;
3257 }
3258
3259 trans = btrfs_join_transaction(root);
3260 if (IS_ERR(trans)) {
3261 not_written = 1;
3262 goto out;
3263 }
3264
3265 ret = iterate_inodes_from_logical(logical, fs_info, path,
3266 record_inode_for_nocow, nocow_ctx);
3267 if (ret != 0 && ret != -ENOENT) {
3268 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3269 "phys %llu, len %llu, mir %u, ret %d",
3270 logical, physical_for_dev_replace, len, mirror_num,
3271 ret);
3272 not_written = 1;
3273 goto out;
3274 }
3275
3276 btrfs_end_transaction(trans, root);
3277 trans = NULL;
3278 while (!list_empty(&nocow_ctx->inodes)) {
3279 struct scrub_nocow_inode *entry;
3280 entry = list_first_entry(&nocow_ctx->inodes,
3281 struct scrub_nocow_inode,
3282 list);
3283 list_del_init(&entry->list);
3284 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3285 entry->root, nocow_ctx);
3286 kfree(entry);
3287 if (ret == COPY_COMPLETE) {
3288 ret = 0;
3289 break;
3290 } else if (ret) {
3291 break;
3292 }
3293 }
3294out:
3295 while (!list_empty(&nocow_ctx->inodes)) {
3296 struct scrub_nocow_inode *entry;
3297 entry = list_first_entry(&nocow_ctx->inodes,
3298 struct scrub_nocow_inode,
3299 list);
3300 list_del_init(&entry->list);
3301 kfree(entry);
3302 }
3303 if (trans && !IS_ERR(trans))
3304 btrfs_end_transaction(trans, root);
3305 if (not_written)
3306 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3307 num_uncorrectable_read_errors);
3308
3309 btrfs_free_path(path);
3310 kfree(nocow_ctx);
3311
3312 scrub_pending_trans_workers_dec(sctx);
3313}
3314
3315static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3316 struct scrub_copy_nocow_ctx *nocow_ctx)
3317{
3318 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3319 struct btrfs_key key;
3320 struct inode *inode;
3321 struct page *page;
3322 struct btrfs_root *local_root;
3323 struct btrfs_ordered_extent *ordered;
3324 struct extent_map *em;
3325 struct extent_state *cached_state = NULL;
3326 struct extent_io_tree *io_tree;
3327 u64 physical_for_dev_replace;
3328 u64 len = nocow_ctx->len;
3329 u64 lockstart = offset, lockend = offset + len - 1;
3330 unsigned long index;
3331 int srcu_index;
3332 int ret = 0;
3333 int err = 0;
3334
3335 key.objectid = root;
3336 key.type = BTRFS_ROOT_ITEM_KEY;
3337 key.offset = (u64)-1;
3338
3339 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3340
3341 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3342 if (IS_ERR(local_root)) {
3343 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3344 return PTR_ERR(local_root);
3345 }
3346
3347 key.type = BTRFS_INODE_ITEM_KEY;
3348 key.objectid = inum;
3349 key.offset = 0;
3350 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3351 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3352 if (IS_ERR(inode))
3353 return PTR_ERR(inode);
3354
3355 /* Avoid truncate/dio/punch hole.. */
3356 mutex_lock(&inode->i_mutex);
3357 inode_dio_wait(inode);
3358
3359 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3360 io_tree = &BTRFS_I(inode)->io_tree;
3361
3362 lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3363 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3364 if (ordered) {
3365 btrfs_put_ordered_extent(ordered);
3366 goto out_unlock;
3367 }
3368
3369 em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3370 if (IS_ERR(em)) {
3371 ret = PTR_ERR(em);
3372 goto out_unlock;
3373 }
3374
3375 /*
3376 * This extent does not actually cover the logical extent anymore,
3377 * move on to the next inode.
3378 */
3379 if (em->block_start > nocow_ctx->logical ||
3380 em->block_start + em->block_len < nocow_ctx->logical + len) {
3381 free_extent_map(em);
3382 goto out_unlock;
3383 }
3384 free_extent_map(em);
3385
3386 while (len >= PAGE_CACHE_SIZE) {
3387 index = offset >> PAGE_CACHE_SHIFT;
3388again:
3389 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3390 if (!page) {
3391 btrfs_err(fs_info, "find_or_create_page() failed");
3392 ret = -ENOMEM;
3393 goto out;
3394 }
3395
3396 if (PageUptodate(page)) {
3397 if (PageDirty(page))
3398 goto next_page;
3399 } else {
3400 ClearPageError(page);
3401 err = extent_read_full_page_nolock(io_tree, page,
3402 btrfs_get_extent,
3403 nocow_ctx->mirror_num);
3404 if (err) {
3405 ret = err;
3406 goto next_page;
3407 }
3408
3409 lock_page(page);
3410 /*
3411 * If the page has been remove from the page cache,
3412 * the data on it is meaningless, because it may be
3413 * old one, the new data may be written into the new
3414 * page in the page cache.
3415 */
3416 if (page->mapping != inode->i_mapping) {
3417 unlock_page(page);
3418 page_cache_release(page);
3419 goto again;
3420 }
3421 if (!PageUptodate(page)) {
3422 ret = -EIO;
3423 goto next_page;
3424 }
3425 }
3426 err = write_page_nocow(nocow_ctx->sctx,
3427 physical_for_dev_replace, page);
3428 if (err)
3429 ret = err;
3430next_page:
3431 unlock_page(page);
3432 page_cache_release(page);
3433
3434 if (ret)
3435 break;
3436
3437 offset += PAGE_CACHE_SIZE;
3438 physical_for_dev_replace += PAGE_CACHE_SIZE;
3439 len -= PAGE_CACHE_SIZE;
3440 }
3441 ret = COPY_COMPLETE;
3442out_unlock:
3443 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3444 GFP_NOFS);
3445out:
3446 mutex_unlock(&inode->i_mutex);
3447 iput(inode);
3448 return ret;
3449}
3450
3451static int write_page_nocow(struct scrub_ctx *sctx,
3452 u64 physical_for_dev_replace, struct page *page)
3453{
3454 struct bio *bio;
3455 struct btrfs_device *dev;
3456 int ret;
3457
3458 dev = sctx->wr_ctx.tgtdev;
3459 if (!dev)
3460 return -EIO;
3461 if (!dev->bdev) {
3462 printk_ratelimited(KERN_WARNING
3463 "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3464 return -EIO;
3465 }
3466 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3467 if (!bio) {
3468 spin_lock(&sctx->stat_lock);
3469 sctx->stat.malloc_errors++;
3470 spin_unlock(&sctx->stat_lock);
3471 return -ENOMEM;
3472 }
3473 bio->bi_iter.bi_size = 0;
3474 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3475 bio->bi_bdev = dev->bdev;
3476 ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3477 if (ret != PAGE_CACHE_SIZE) {
3478leave_with_eio:
3479 bio_put(bio);
3480 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3481 return -EIO;
3482 }
3483
3484 if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3485 goto leave_with_eio;
3486
3487 bio_put(bio);
3488 return 0;
3489}