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 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 "ctree.h"
21#include "volumes.h"
22#include "disk-io.h"
23#include "ordered-data.h"
24
25/*
26 * This is only the first step towards a full-features scrub. It reads all
27 * extent and super block and verifies the checksums. In case a bad checksum
28 * is found or the extent cannot be read, good data will be written back if
29 * any can be found.
30 *
31 * Future enhancements:
32 * - To enhance the performance, better read-ahead strategies for the
33 * extent-tree can be employed.
34 * - In case an unrepairable extent is encountered, track which files are
35 * affected and report them
36 * - In case of a read error on files with nodatasum, map the file and read
37 * the extent to trigger a writeback of the good copy
38 * - track and record media errors, throw out bad devices
39 * - add a mode to also read unallocated space
40 * - make the prefetch cancellable
41 */
42
43struct scrub_bio;
44struct scrub_page;
45struct scrub_dev;
46static void scrub_bio_end_io(struct bio *bio, int err);
47static void scrub_checksum(struct btrfs_work *work);
48static int scrub_checksum_data(struct scrub_dev *sdev,
49 struct scrub_page *spag, void *buffer);
50static int scrub_checksum_tree_block(struct scrub_dev *sdev,
51 struct scrub_page *spag, u64 logical,
52 void *buffer);
53static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer);
54static int scrub_fixup_check(struct scrub_bio *sbio, int ix);
55static void scrub_fixup_end_io(struct bio *bio, int err);
56static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
57 struct page *page);
58static void scrub_fixup(struct scrub_bio *sbio, int ix);
59
60#define SCRUB_PAGES_PER_BIO 16 /* 64k per bio */
61#define SCRUB_BIOS_PER_DEV 16 /* 1 MB per device in flight */
62
63struct scrub_page {
64 u64 flags; /* extent flags */
65 u64 generation;
66 u64 mirror_num;
67 int have_csum;
68 u8 csum[BTRFS_CSUM_SIZE];
69};
70
71struct scrub_bio {
72 int index;
73 struct scrub_dev *sdev;
74 struct bio *bio;
75 int err;
76 u64 logical;
77 u64 physical;
78 struct scrub_page spag[SCRUB_PAGES_PER_BIO];
79 u64 count;
80 int next_free;
81 struct btrfs_work work;
82};
83
84struct scrub_dev {
85 struct scrub_bio *bios[SCRUB_BIOS_PER_DEV];
86 struct btrfs_device *dev;
87 int first_free;
88 int curr;
89 atomic_t in_flight;
90 spinlock_t list_lock;
91 wait_queue_head_t list_wait;
92 u16 csum_size;
93 struct list_head csum_list;
94 atomic_t cancel_req;
95 int readonly;
96 /*
97 * statistics
98 */
99 struct btrfs_scrub_progress stat;
100 spinlock_t stat_lock;
101};
102
103static void scrub_free_csums(struct scrub_dev *sdev)
104{
105 while (!list_empty(&sdev->csum_list)) {
106 struct btrfs_ordered_sum *sum;
107 sum = list_first_entry(&sdev->csum_list,
108 struct btrfs_ordered_sum, list);
109 list_del(&sum->list);
110 kfree(sum);
111 }
112}
113
114static void scrub_free_bio(struct bio *bio)
115{
116 int i;
117 struct page *last_page = NULL;
118
119 if (!bio)
120 return;
121
122 for (i = 0; i < bio->bi_vcnt; ++i) {
123 if (bio->bi_io_vec[i].bv_page == last_page)
124 continue;
125 last_page = bio->bi_io_vec[i].bv_page;
126 __free_page(last_page);
127 }
128 bio_put(bio);
129}
130
131static noinline_for_stack void scrub_free_dev(struct scrub_dev *sdev)
132{
133 int i;
134
135 if (!sdev)
136 return;
137
138 for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
139 struct scrub_bio *sbio = sdev->bios[i];
140
141 if (!sbio)
142 break;
143
144 scrub_free_bio(sbio->bio);
145 kfree(sbio);
146 }
147
148 scrub_free_csums(sdev);
149 kfree(sdev);
150}
151
152static noinline_for_stack
153struct scrub_dev *scrub_setup_dev(struct btrfs_device *dev)
154{
155 struct scrub_dev *sdev;
156 int i;
157 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
158
159 sdev = kzalloc(sizeof(*sdev), GFP_NOFS);
160 if (!sdev)
161 goto nomem;
162 sdev->dev = dev;
163 for (i = 0; i < SCRUB_BIOS_PER_DEV; ++i) {
164 struct scrub_bio *sbio;
165
166 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
167 if (!sbio)
168 goto nomem;
169 sdev->bios[i] = sbio;
170
171 sbio->index = i;
172 sbio->sdev = sdev;
173 sbio->count = 0;
174 sbio->work.func = scrub_checksum;
175
176 if (i != SCRUB_BIOS_PER_DEV-1)
177 sdev->bios[i]->next_free = i + 1;
178 else
179 sdev->bios[i]->next_free = -1;
180 }
181 sdev->first_free = 0;
182 sdev->curr = -1;
183 atomic_set(&sdev->in_flight, 0);
184 atomic_set(&sdev->cancel_req, 0);
185 sdev->csum_size = btrfs_super_csum_size(&fs_info->super_copy);
186 INIT_LIST_HEAD(&sdev->csum_list);
187
188 spin_lock_init(&sdev->list_lock);
189 spin_lock_init(&sdev->stat_lock);
190 init_waitqueue_head(&sdev->list_wait);
191 return sdev;
192
193nomem:
194 scrub_free_dev(sdev);
195 return ERR_PTR(-ENOMEM);
196}
197
198/*
199 * scrub_recheck_error gets called when either verification of the page
200 * failed or the bio failed to read, e.g. with EIO. In the latter case,
201 * recheck_error gets called for every page in the bio, even though only
202 * one may be bad
203 */
204static void scrub_recheck_error(struct scrub_bio *sbio, int ix)
205{
206 if (sbio->err) {
207 if (scrub_fixup_io(READ, sbio->sdev->dev->bdev,
208 (sbio->physical + ix * PAGE_SIZE) >> 9,
209 sbio->bio->bi_io_vec[ix].bv_page) == 0) {
210 if (scrub_fixup_check(sbio, ix) == 0)
211 return;
212 }
213 }
214
215 scrub_fixup(sbio, ix);
216}
217
218static int scrub_fixup_check(struct scrub_bio *sbio, int ix)
219{
220 int ret = 1;
221 struct page *page;
222 void *buffer;
223 u64 flags = sbio->spag[ix].flags;
224
225 page = sbio->bio->bi_io_vec[ix].bv_page;
226 buffer = kmap_atomic(page, KM_USER0);
227 if (flags & BTRFS_EXTENT_FLAG_DATA) {
228 ret = scrub_checksum_data(sbio->sdev,
229 sbio->spag + ix, buffer);
230 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
231 ret = scrub_checksum_tree_block(sbio->sdev,
232 sbio->spag + ix,
233 sbio->logical + ix * PAGE_SIZE,
234 buffer);
235 } else {
236 WARN_ON(1);
237 }
238 kunmap_atomic(buffer, KM_USER0);
239
240 return ret;
241}
242
243static void scrub_fixup_end_io(struct bio *bio, int err)
244{
245 complete((struct completion *)bio->bi_private);
246}
247
248static void scrub_fixup(struct scrub_bio *sbio, int ix)
249{
250 struct scrub_dev *sdev = sbio->sdev;
251 struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
252 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
253 struct btrfs_multi_bio *multi = NULL;
254 u64 logical = sbio->logical + ix * PAGE_SIZE;
255 u64 length;
256 int i;
257 int ret;
258 DECLARE_COMPLETION_ONSTACK(complete);
259
260 if ((sbio->spag[ix].flags & BTRFS_EXTENT_FLAG_DATA) &&
261 (sbio->spag[ix].have_csum == 0)) {
262 /*
263 * nodatasum, don't try to fix anything
264 * FIXME: we can do better, open the inode and trigger a
265 * writeback
266 */
267 goto uncorrectable;
268 }
269
270 length = PAGE_SIZE;
271 ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length,
272 &multi, 0);
273 if (ret || !multi || length < PAGE_SIZE) {
274 printk(KERN_ERR
275 "scrub_fixup: btrfs_map_block failed us for %llu\n",
276 (unsigned long long)logical);
277 WARN_ON(1);
278 return;
279 }
280
281 if (multi->num_stripes == 1)
282 /* there aren't any replicas */
283 goto uncorrectable;
284
285 /*
286 * first find a good copy
287 */
288 for (i = 0; i < multi->num_stripes; ++i) {
289 if (i == sbio->spag[ix].mirror_num)
290 continue;
291
292 if (scrub_fixup_io(READ, multi->stripes[i].dev->bdev,
293 multi->stripes[i].physical >> 9,
294 sbio->bio->bi_io_vec[ix].bv_page)) {
295 /* I/O-error, this is not a good copy */
296 continue;
297 }
298
299 if (scrub_fixup_check(sbio, ix) == 0)
300 break;
301 }
302 if (i == multi->num_stripes)
303 goto uncorrectable;
304
305 if (!sdev->readonly) {
306 /*
307 * bi_io_vec[ix].bv_page now contains good data, write it back
308 */
309 if (scrub_fixup_io(WRITE, sdev->dev->bdev,
310 (sbio->physical + ix * PAGE_SIZE) >> 9,
311 sbio->bio->bi_io_vec[ix].bv_page)) {
312 /* I/O-error, writeback failed, give up */
313 goto uncorrectable;
314 }
315 }
316
317 kfree(multi);
318 spin_lock(&sdev->stat_lock);
319 ++sdev->stat.corrected_errors;
320 spin_unlock(&sdev->stat_lock);
321
322 if (printk_ratelimit())
323 printk(KERN_ERR "btrfs: fixed up at %llu\n",
324 (unsigned long long)logical);
325 return;
326
327uncorrectable:
328 kfree(multi);
329 spin_lock(&sdev->stat_lock);
330 ++sdev->stat.uncorrectable_errors;
331 spin_unlock(&sdev->stat_lock);
332
333 if (printk_ratelimit())
334 printk(KERN_ERR "btrfs: unable to fixup at %llu\n",
335 (unsigned long long)logical);
336}
337
338static int scrub_fixup_io(int rw, struct block_device *bdev, sector_t sector,
339 struct page *page)
340{
341 struct bio *bio = NULL;
342 int ret;
343 DECLARE_COMPLETION_ONSTACK(complete);
344
345 bio = bio_alloc(GFP_NOFS, 1);
346 bio->bi_bdev = bdev;
347 bio->bi_sector = sector;
348 bio_add_page(bio, page, PAGE_SIZE, 0);
349 bio->bi_end_io = scrub_fixup_end_io;
350 bio->bi_private = &complete;
351 submit_bio(rw, bio);
352
353 /* this will also unplug the queue */
354 wait_for_completion(&complete);
355
356 ret = !test_bit(BIO_UPTODATE, &bio->bi_flags);
357 bio_put(bio);
358 return ret;
359}
360
361static void scrub_bio_end_io(struct bio *bio, int err)
362{
363 struct scrub_bio *sbio = bio->bi_private;
364 struct scrub_dev *sdev = sbio->sdev;
365 struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
366
367 sbio->err = err;
368 sbio->bio = bio;
369
370 btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
371}
372
373static void scrub_checksum(struct btrfs_work *work)
374{
375 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
376 struct scrub_dev *sdev = sbio->sdev;
377 struct page *page;
378 void *buffer;
379 int i;
380 u64 flags;
381 u64 logical;
382 int ret;
383
384 if (sbio->err) {
385 for (i = 0; i < sbio->count; ++i)
386 scrub_recheck_error(sbio, i);
387
388 sbio->bio->bi_flags &= ~(BIO_POOL_MASK - 1);
389 sbio->bio->bi_flags |= 1 << BIO_UPTODATE;
390 sbio->bio->bi_phys_segments = 0;
391 sbio->bio->bi_idx = 0;
392
393 for (i = 0; i < sbio->count; i++) {
394 struct bio_vec *bi;
395 bi = &sbio->bio->bi_io_vec[i];
396 bi->bv_offset = 0;
397 bi->bv_len = PAGE_SIZE;
398 }
399
400 spin_lock(&sdev->stat_lock);
401 ++sdev->stat.read_errors;
402 spin_unlock(&sdev->stat_lock);
403 goto out;
404 }
405 for (i = 0; i < sbio->count; ++i) {
406 page = sbio->bio->bi_io_vec[i].bv_page;
407 buffer = kmap_atomic(page, KM_USER0);
408 flags = sbio->spag[i].flags;
409 logical = sbio->logical + i * PAGE_SIZE;
410 ret = 0;
411 if (flags & BTRFS_EXTENT_FLAG_DATA) {
412 ret = scrub_checksum_data(sdev, sbio->spag + i, buffer);
413 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
414 ret = scrub_checksum_tree_block(sdev, sbio->spag + i,
415 logical, buffer);
416 } else if (flags & BTRFS_EXTENT_FLAG_SUPER) {
417 BUG_ON(i);
418 (void)scrub_checksum_super(sbio, buffer);
419 } else {
420 WARN_ON(1);
421 }
422 kunmap_atomic(buffer, KM_USER0);
423 if (ret)
424 scrub_recheck_error(sbio, i);
425 }
426
427out:
428 scrub_free_bio(sbio->bio);
429 sbio->bio = NULL;
430 spin_lock(&sdev->list_lock);
431 sbio->next_free = sdev->first_free;
432 sdev->first_free = sbio->index;
433 spin_unlock(&sdev->list_lock);
434 atomic_dec(&sdev->in_flight);
435 wake_up(&sdev->list_wait);
436}
437
438static int scrub_checksum_data(struct scrub_dev *sdev,
439 struct scrub_page *spag, void *buffer)
440{
441 u8 csum[BTRFS_CSUM_SIZE];
442 u32 crc = ~(u32)0;
443 int fail = 0;
444 struct btrfs_root *root = sdev->dev->dev_root;
445
446 if (!spag->have_csum)
447 return 0;
448
449 crc = btrfs_csum_data(root, buffer, crc, PAGE_SIZE);
450 btrfs_csum_final(crc, csum);
451 if (memcmp(csum, spag->csum, sdev->csum_size))
452 fail = 1;
453
454 spin_lock(&sdev->stat_lock);
455 ++sdev->stat.data_extents_scrubbed;
456 sdev->stat.data_bytes_scrubbed += PAGE_SIZE;
457 if (fail)
458 ++sdev->stat.csum_errors;
459 spin_unlock(&sdev->stat_lock);
460
461 return fail;
462}
463
464static int scrub_checksum_tree_block(struct scrub_dev *sdev,
465 struct scrub_page *spag, u64 logical,
466 void *buffer)
467{
468 struct btrfs_header *h;
469 struct btrfs_root *root = sdev->dev->dev_root;
470 struct btrfs_fs_info *fs_info = root->fs_info;
471 u8 csum[BTRFS_CSUM_SIZE];
472 u32 crc = ~(u32)0;
473 int fail = 0;
474 int crc_fail = 0;
475
476 /*
477 * we don't use the getter functions here, as we
478 * a) don't have an extent buffer and
479 * b) the page is already kmapped
480 */
481 h = (struct btrfs_header *)buffer;
482
483 if (logical != le64_to_cpu(h->bytenr))
484 ++fail;
485
486 if (spag->generation != le64_to_cpu(h->generation))
487 ++fail;
488
489 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
490 ++fail;
491
492 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
493 BTRFS_UUID_SIZE))
494 ++fail;
495
496 crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
497 PAGE_SIZE - BTRFS_CSUM_SIZE);
498 btrfs_csum_final(crc, csum);
499 if (memcmp(csum, h->csum, sdev->csum_size))
500 ++crc_fail;
501
502 spin_lock(&sdev->stat_lock);
503 ++sdev->stat.tree_extents_scrubbed;
504 sdev->stat.tree_bytes_scrubbed += PAGE_SIZE;
505 if (crc_fail)
506 ++sdev->stat.csum_errors;
507 if (fail)
508 ++sdev->stat.verify_errors;
509 spin_unlock(&sdev->stat_lock);
510
511 return fail || crc_fail;
512}
513
514static int scrub_checksum_super(struct scrub_bio *sbio, void *buffer)
515{
516 struct btrfs_super_block *s;
517 u64 logical;
518 struct scrub_dev *sdev = sbio->sdev;
519 struct btrfs_root *root = sdev->dev->dev_root;
520 struct btrfs_fs_info *fs_info = root->fs_info;
521 u8 csum[BTRFS_CSUM_SIZE];
522 u32 crc = ~(u32)0;
523 int fail = 0;
524
525 s = (struct btrfs_super_block *)buffer;
526 logical = sbio->logical;
527
528 if (logical != le64_to_cpu(s->bytenr))
529 ++fail;
530
531 if (sbio->spag[0].generation != le64_to_cpu(s->generation))
532 ++fail;
533
534 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
535 ++fail;
536
537 crc = btrfs_csum_data(root, buffer + BTRFS_CSUM_SIZE, crc,
538 PAGE_SIZE - BTRFS_CSUM_SIZE);
539 btrfs_csum_final(crc, csum);
540 if (memcmp(csum, s->csum, sbio->sdev->csum_size))
541 ++fail;
542
543 if (fail) {
544 /*
545 * if we find an error in a super block, we just report it.
546 * They will get written with the next transaction commit
547 * anyway
548 */
549 spin_lock(&sdev->stat_lock);
550 ++sdev->stat.super_errors;
551 spin_unlock(&sdev->stat_lock);
552 }
553
554 return fail;
555}
556
557static int scrub_submit(struct scrub_dev *sdev)
558{
559 struct scrub_bio *sbio;
560 struct bio *bio;
561 int i;
562
563 if (sdev->curr == -1)
564 return 0;
565
566 sbio = sdev->bios[sdev->curr];
567
568 bio = bio_alloc(GFP_NOFS, sbio->count);
569 if (!bio)
570 goto nomem;
571
572 bio->bi_private = sbio;
573 bio->bi_end_io = scrub_bio_end_io;
574 bio->bi_bdev = sdev->dev->bdev;
575 bio->bi_sector = sbio->physical >> 9;
576
577 for (i = 0; i < sbio->count; ++i) {
578 struct page *page;
579 int ret;
580
581 page = alloc_page(GFP_NOFS);
582 if (!page)
583 goto nomem;
584
585 ret = bio_add_page(bio, page, PAGE_SIZE, 0);
586 if (!ret) {
587 __free_page(page);
588 goto nomem;
589 }
590 }
591
592 sbio->err = 0;
593 sdev->curr = -1;
594 atomic_inc(&sdev->in_flight);
595
596 submit_bio(READ, bio);
597
598 return 0;
599
600nomem:
601 scrub_free_bio(bio);
602
603 return -ENOMEM;
604}
605
606static int scrub_page(struct scrub_dev *sdev, u64 logical, u64 len,
607 u64 physical, u64 flags, u64 gen, u64 mirror_num,
608 u8 *csum, int force)
609{
610 struct scrub_bio *sbio;
611
612again:
613 /*
614 * grab a fresh bio or wait for one to become available
615 */
616 while (sdev->curr == -1) {
617 spin_lock(&sdev->list_lock);
618 sdev->curr = sdev->first_free;
619 if (sdev->curr != -1) {
620 sdev->first_free = sdev->bios[sdev->curr]->next_free;
621 sdev->bios[sdev->curr]->next_free = -1;
622 sdev->bios[sdev->curr]->count = 0;
623 spin_unlock(&sdev->list_lock);
624 } else {
625 spin_unlock(&sdev->list_lock);
626 wait_event(sdev->list_wait, sdev->first_free != -1);
627 }
628 }
629 sbio = sdev->bios[sdev->curr];
630 if (sbio->count == 0) {
631 sbio->physical = physical;
632 sbio->logical = logical;
633 } else if (sbio->physical + sbio->count * PAGE_SIZE != physical ||
634 sbio->logical + sbio->count * PAGE_SIZE != logical) {
635 int ret;
636
637 ret = scrub_submit(sdev);
638 if (ret)
639 return ret;
640 goto again;
641 }
642 sbio->spag[sbio->count].flags = flags;
643 sbio->spag[sbio->count].generation = gen;
644 sbio->spag[sbio->count].have_csum = 0;
645 sbio->spag[sbio->count].mirror_num = mirror_num;
646 if (csum) {
647 sbio->spag[sbio->count].have_csum = 1;
648 memcpy(sbio->spag[sbio->count].csum, csum, sdev->csum_size);
649 }
650 ++sbio->count;
651 if (sbio->count == SCRUB_PAGES_PER_BIO || force) {
652 int ret;
653
654 ret = scrub_submit(sdev);
655 if (ret)
656 return ret;
657 }
658
659 return 0;
660}
661
662static int scrub_find_csum(struct scrub_dev *sdev, u64 logical, u64 len,
663 u8 *csum)
664{
665 struct btrfs_ordered_sum *sum = NULL;
666 int ret = 0;
667 unsigned long i;
668 unsigned long num_sectors;
669 u32 sectorsize = sdev->dev->dev_root->sectorsize;
670
671 while (!list_empty(&sdev->csum_list)) {
672 sum = list_first_entry(&sdev->csum_list,
673 struct btrfs_ordered_sum, list);
674 if (sum->bytenr > logical)
675 return 0;
676 if (sum->bytenr + sum->len > logical)
677 break;
678
679 ++sdev->stat.csum_discards;
680 list_del(&sum->list);
681 kfree(sum);
682 sum = NULL;
683 }
684 if (!sum)
685 return 0;
686
687 num_sectors = sum->len / sectorsize;
688 for (i = 0; i < num_sectors; ++i) {
689 if (sum->sums[i].bytenr == logical) {
690 memcpy(csum, &sum->sums[i].sum, sdev->csum_size);
691 ret = 1;
692 break;
693 }
694 }
695 if (ret && i == num_sectors - 1) {
696 list_del(&sum->list);
697 kfree(sum);
698 }
699 return ret;
700}
701
702/* scrub extent tries to collect up to 64 kB for each bio */
703static int scrub_extent(struct scrub_dev *sdev, u64 logical, u64 len,
704 u64 physical, u64 flags, u64 gen, u64 mirror_num)
705{
706 int ret;
707 u8 csum[BTRFS_CSUM_SIZE];
708
709 while (len) {
710 u64 l = min_t(u64, len, PAGE_SIZE);
711 int have_csum = 0;
712
713 if (flags & BTRFS_EXTENT_FLAG_DATA) {
714 /* push csums to sbio */
715 have_csum = scrub_find_csum(sdev, logical, l, csum);
716 if (have_csum == 0)
717 ++sdev->stat.no_csum;
718 }
719 ret = scrub_page(sdev, logical, l, physical, flags, gen,
720 mirror_num, have_csum ? csum : NULL, 0);
721 if (ret)
722 return ret;
723 len -= l;
724 logical += l;
725 physical += l;
726 }
727 return 0;
728}
729
730static noinline_for_stack int scrub_stripe(struct scrub_dev *sdev,
731 struct map_lookup *map, int num, u64 base, u64 length)
732{
733 struct btrfs_path *path;
734 struct btrfs_fs_info *fs_info = sdev->dev->dev_root->fs_info;
735 struct btrfs_root *root = fs_info->extent_root;
736 struct btrfs_root *csum_root = fs_info->csum_root;
737 struct btrfs_extent_item *extent;
738 struct blk_plug plug;
739 u64 flags;
740 int ret;
741 int slot;
742 int i;
743 u64 nstripes;
744 int start_stripe;
745 struct extent_buffer *l;
746 struct btrfs_key key;
747 u64 physical;
748 u64 logical;
749 u64 generation;
750 u64 mirror_num;
751
752 u64 increment = map->stripe_len;
753 u64 offset;
754
755 nstripes = length;
756 offset = 0;
757 do_div(nstripes, map->stripe_len);
758 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
759 offset = map->stripe_len * num;
760 increment = map->stripe_len * map->num_stripes;
761 mirror_num = 0;
762 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
763 int factor = map->num_stripes / map->sub_stripes;
764 offset = map->stripe_len * (num / map->sub_stripes);
765 increment = map->stripe_len * factor;
766 mirror_num = num % map->sub_stripes;
767 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
768 increment = map->stripe_len;
769 mirror_num = num % map->num_stripes;
770 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
771 increment = map->stripe_len;
772 mirror_num = num % map->num_stripes;
773 } else {
774 increment = map->stripe_len;
775 mirror_num = 0;
776 }
777
778 path = btrfs_alloc_path();
779 if (!path)
780 return -ENOMEM;
781
782 path->reada = 2;
783 path->search_commit_root = 1;
784 path->skip_locking = 1;
785
786 /*
787 * find all extents for each stripe and just read them to get
788 * them into the page cache
789 * FIXME: we can do better. build a more intelligent prefetching
790 */
791 logical = base + offset;
792 physical = map->stripes[num].physical;
793 ret = 0;
794 for (i = 0; i < nstripes; ++i) {
795 key.objectid = logical;
796 key.type = BTRFS_EXTENT_ITEM_KEY;
797 key.offset = (u64)0;
798
799 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
800 if (ret < 0)
801 goto out_noplug;
802
803 /*
804 * we might miss half an extent here, but that doesn't matter,
805 * as it's only the prefetch
806 */
807 while (1) {
808 l = path->nodes[0];
809 slot = path->slots[0];
810 if (slot >= btrfs_header_nritems(l)) {
811 ret = btrfs_next_leaf(root, path);
812 if (ret == 0)
813 continue;
814 if (ret < 0)
815 goto out_noplug;
816
817 break;
818 }
819 btrfs_item_key_to_cpu(l, &key, slot);
820
821 if (key.objectid >= logical + map->stripe_len)
822 break;
823
824 path->slots[0]++;
825 }
826 btrfs_release_path(path);
827 logical += increment;
828 physical += map->stripe_len;
829 cond_resched();
830 }
831
832 /*
833 * collect all data csums for the stripe to avoid seeking during
834 * the scrub. This might currently (crc32) end up to be about 1MB
835 */
836 start_stripe = 0;
837 blk_start_plug(&plug);
838again:
839 logical = base + offset + start_stripe * increment;
840 for (i = start_stripe; i < nstripes; ++i) {
841 ret = btrfs_lookup_csums_range(csum_root, logical,
842 logical + map->stripe_len - 1,
843 &sdev->csum_list, 1);
844 if (ret)
845 goto out;
846
847 logical += increment;
848 cond_resched();
849 }
850 /*
851 * now find all extents for each stripe and scrub them
852 */
853 logical = base + offset + start_stripe * increment;
854 physical = map->stripes[num].physical + start_stripe * map->stripe_len;
855 ret = 0;
856 for (i = start_stripe; i < nstripes; ++i) {
857 /*
858 * canceled?
859 */
860 if (atomic_read(&fs_info->scrub_cancel_req) ||
861 atomic_read(&sdev->cancel_req)) {
862 ret = -ECANCELED;
863 goto out;
864 }
865 /*
866 * check to see if we have to pause
867 */
868 if (atomic_read(&fs_info->scrub_pause_req)) {
869 /* push queued extents */
870 scrub_submit(sdev);
871 wait_event(sdev->list_wait,
872 atomic_read(&sdev->in_flight) == 0);
873 atomic_inc(&fs_info->scrubs_paused);
874 wake_up(&fs_info->scrub_pause_wait);
875 mutex_lock(&fs_info->scrub_lock);
876 while (atomic_read(&fs_info->scrub_pause_req)) {
877 mutex_unlock(&fs_info->scrub_lock);
878 wait_event(fs_info->scrub_pause_wait,
879 atomic_read(&fs_info->scrub_pause_req) == 0);
880 mutex_lock(&fs_info->scrub_lock);
881 }
882 atomic_dec(&fs_info->scrubs_paused);
883 mutex_unlock(&fs_info->scrub_lock);
884 wake_up(&fs_info->scrub_pause_wait);
885 scrub_free_csums(sdev);
886 start_stripe = i;
887 goto again;
888 }
889
890 key.objectid = logical;
891 key.type = BTRFS_EXTENT_ITEM_KEY;
892 key.offset = (u64)0;
893
894 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
895 if (ret < 0)
896 goto out;
897 if (ret > 0) {
898 ret = btrfs_previous_item(root, path, 0,
899 BTRFS_EXTENT_ITEM_KEY);
900 if (ret < 0)
901 goto out;
902 if (ret > 0) {
903 /* there's no smaller item, so stick with the
904 * larger one */
905 btrfs_release_path(path);
906 ret = btrfs_search_slot(NULL, root, &key,
907 path, 0, 0);
908 if (ret < 0)
909 goto out;
910 }
911 }
912
913 while (1) {
914 l = path->nodes[0];
915 slot = path->slots[0];
916 if (slot >= btrfs_header_nritems(l)) {
917 ret = btrfs_next_leaf(root, path);
918 if (ret == 0)
919 continue;
920 if (ret < 0)
921 goto out;
922
923 break;
924 }
925 btrfs_item_key_to_cpu(l, &key, slot);
926
927 if (key.objectid + key.offset <= logical)
928 goto next;
929
930 if (key.objectid >= logical + map->stripe_len)
931 break;
932
933 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
934 goto next;
935
936 extent = btrfs_item_ptr(l, slot,
937 struct btrfs_extent_item);
938 flags = btrfs_extent_flags(l, extent);
939 generation = btrfs_extent_generation(l, extent);
940
941 if (key.objectid < logical &&
942 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
943 printk(KERN_ERR
944 "btrfs scrub: tree block %llu spanning "
945 "stripes, ignored. logical=%llu\n",
946 (unsigned long long)key.objectid,
947 (unsigned long long)logical);
948 goto next;
949 }
950
951 /*
952 * trim extent to this stripe
953 */
954 if (key.objectid < logical) {
955 key.offset -= logical - key.objectid;
956 key.objectid = logical;
957 }
958 if (key.objectid + key.offset >
959 logical + map->stripe_len) {
960 key.offset = logical + map->stripe_len -
961 key.objectid;
962 }
963
964 ret = scrub_extent(sdev, key.objectid, key.offset,
965 key.objectid - logical + physical,
966 flags, generation, mirror_num);
967 if (ret)
968 goto out;
969
970next:
971 path->slots[0]++;
972 }
973 btrfs_release_path(path);
974 logical += increment;
975 physical += map->stripe_len;
976 spin_lock(&sdev->stat_lock);
977 sdev->stat.last_physical = physical;
978 spin_unlock(&sdev->stat_lock);
979 }
980 /* push queued extents */
981 scrub_submit(sdev);
982
983out:
984 blk_finish_plug(&plug);
985out_noplug:
986 btrfs_free_path(path);
987 return ret < 0 ? ret : 0;
988}
989
990static noinline_for_stack int scrub_chunk(struct scrub_dev *sdev,
991 u64 chunk_tree, u64 chunk_objectid, u64 chunk_offset, u64 length)
992{
993 struct btrfs_mapping_tree *map_tree =
994 &sdev->dev->dev_root->fs_info->mapping_tree;
995 struct map_lookup *map;
996 struct extent_map *em;
997 int i;
998 int ret = -EINVAL;
999
1000 read_lock(&map_tree->map_tree.lock);
1001 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
1002 read_unlock(&map_tree->map_tree.lock);
1003
1004 if (!em)
1005 return -EINVAL;
1006
1007 map = (struct map_lookup *)em->bdev;
1008 if (em->start != chunk_offset)
1009 goto out;
1010
1011 if (em->len < length)
1012 goto out;
1013
1014 for (i = 0; i < map->num_stripes; ++i) {
1015 if (map->stripes[i].dev == sdev->dev) {
1016 ret = scrub_stripe(sdev, map, i, chunk_offset, length);
1017 if (ret)
1018 goto out;
1019 }
1020 }
1021out:
1022 free_extent_map(em);
1023
1024 return ret;
1025}
1026
1027static noinline_for_stack
1028int scrub_enumerate_chunks(struct scrub_dev *sdev, u64 start, u64 end)
1029{
1030 struct btrfs_dev_extent *dev_extent = NULL;
1031 struct btrfs_path *path;
1032 struct btrfs_root *root = sdev->dev->dev_root;
1033 struct btrfs_fs_info *fs_info = root->fs_info;
1034 u64 length;
1035 u64 chunk_tree;
1036 u64 chunk_objectid;
1037 u64 chunk_offset;
1038 int ret;
1039 int slot;
1040 struct extent_buffer *l;
1041 struct btrfs_key key;
1042 struct btrfs_key found_key;
1043 struct btrfs_block_group_cache *cache;
1044
1045 path = btrfs_alloc_path();
1046 if (!path)
1047 return -ENOMEM;
1048
1049 path->reada = 2;
1050 path->search_commit_root = 1;
1051 path->skip_locking = 1;
1052
1053 key.objectid = sdev->dev->devid;
1054 key.offset = 0ull;
1055 key.type = BTRFS_DEV_EXTENT_KEY;
1056
1057
1058 while (1) {
1059 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1060 if (ret < 0)
1061 break;
1062 if (ret > 0) {
1063 if (path->slots[0] >=
1064 btrfs_header_nritems(path->nodes[0])) {
1065 ret = btrfs_next_leaf(root, path);
1066 if (ret)
1067 break;
1068 }
1069 }
1070
1071 l = path->nodes[0];
1072 slot = path->slots[0];
1073
1074 btrfs_item_key_to_cpu(l, &found_key, slot);
1075
1076 if (found_key.objectid != sdev->dev->devid)
1077 break;
1078
1079 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
1080 break;
1081
1082 if (found_key.offset >= end)
1083 break;
1084
1085 if (found_key.offset < key.offset)
1086 break;
1087
1088 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
1089 length = btrfs_dev_extent_length(l, dev_extent);
1090
1091 if (found_key.offset + length <= start) {
1092 key.offset = found_key.offset + length;
1093 btrfs_release_path(path);
1094 continue;
1095 }
1096
1097 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
1098 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
1099 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
1100
1101 /*
1102 * get a reference on the corresponding block group to prevent
1103 * the chunk from going away while we scrub it
1104 */
1105 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
1106 if (!cache) {
1107 ret = -ENOENT;
1108 break;
1109 }
1110 ret = scrub_chunk(sdev, chunk_tree, chunk_objectid,
1111 chunk_offset, length);
1112 btrfs_put_block_group(cache);
1113 if (ret)
1114 break;
1115
1116 key.offset = found_key.offset + length;
1117 btrfs_release_path(path);
1118 }
1119
1120 btrfs_free_path(path);
1121
1122 /*
1123 * ret can still be 1 from search_slot or next_leaf,
1124 * that's not an error
1125 */
1126 return ret < 0 ? ret : 0;
1127}
1128
1129static noinline_for_stack int scrub_supers(struct scrub_dev *sdev)
1130{
1131 int i;
1132 u64 bytenr;
1133 u64 gen;
1134 int ret;
1135 struct btrfs_device *device = sdev->dev;
1136 struct btrfs_root *root = device->dev_root;
1137
1138 gen = root->fs_info->last_trans_committed;
1139
1140 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1141 bytenr = btrfs_sb_offset(i);
1142 if (bytenr + BTRFS_SUPER_INFO_SIZE >= device->total_bytes)
1143 break;
1144
1145 ret = scrub_page(sdev, bytenr, PAGE_SIZE, bytenr,
1146 BTRFS_EXTENT_FLAG_SUPER, gen, i, NULL, 1);
1147 if (ret)
1148 return ret;
1149 }
1150 wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1151
1152 return 0;
1153}
1154
1155/*
1156 * get a reference count on fs_info->scrub_workers. start worker if necessary
1157 */
1158static noinline_for_stack int scrub_workers_get(struct btrfs_root *root)
1159{
1160 struct btrfs_fs_info *fs_info = root->fs_info;
1161
1162 mutex_lock(&fs_info->scrub_lock);
1163 if (fs_info->scrub_workers_refcnt == 0) {
1164 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
1165 fs_info->thread_pool_size, &fs_info->generic_worker);
1166 fs_info->scrub_workers.idle_thresh = 4;
1167 btrfs_start_workers(&fs_info->scrub_workers, 1);
1168 }
1169 ++fs_info->scrub_workers_refcnt;
1170 mutex_unlock(&fs_info->scrub_lock);
1171
1172 return 0;
1173}
1174
1175static noinline_for_stack void scrub_workers_put(struct btrfs_root *root)
1176{
1177 struct btrfs_fs_info *fs_info = root->fs_info;
1178
1179 mutex_lock(&fs_info->scrub_lock);
1180 if (--fs_info->scrub_workers_refcnt == 0)
1181 btrfs_stop_workers(&fs_info->scrub_workers);
1182 WARN_ON(fs_info->scrub_workers_refcnt < 0);
1183 mutex_unlock(&fs_info->scrub_lock);
1184}
1185
1186
1187int btrfs_scrub_dev(struct btrfs_root *root, u64 devid, u64 start, u64 end,
1188 struct btrfs_scrub_progress *progress, int readonly)
1189{
1190 struct scrub_dev *sdev;
1191 struct btrfs_fs_info *fs_info = root->fs_info;
1192 int ret;
1193 struct btrfs_device *dev;
1194
1195 if (btrfs_fs_closing(root->fs_info))
1196 return -EINVAL;
1197
1198 /*
1199 * check some assumptions
1200 */
1201 if (root->sectorsize != PAGE_SIZE ||
1202 root->sectorsize != root->leafsize ||
1203 root->sectorsize != root->nodesize) {
1204 printk(KERN_ERR "btrfs_scrub: size assumptions fail\n");
1205 return -EINVAL;
1206 }
1207
1208 ret = scrub_workers_get(root);
1209 if (ret)
1210 return ret;
1211
1212 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1213 dev = btrfs_find_device(root, devid, NULL, NULL);
1214 if (!dev || dev->missing) {
1215 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1216 scrub_workers_put(root);
1217 return -ENODEV;
1218 }
1219 mutex_lock(&fs_info->scrub_lock);
1220
1221 if (!dev->in_fs_metadata) {
1222 mutex_unlock(&fs_info->scrub_lock);
1223 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1224 scrub_workers_put(root);
1225 return -ENODEV;
1226 }
1227
1228 if (dev->scrub_device) {
1229 mutex_unlock(&fs_info->scrub_lock);
1230 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1231 scrub_workers_put(root);
1232 return -EINPROGRESS;
1233 }
1234 sdev = scrub_setup_dev(dev);
1235 if (IS_ERR(sdev)) {
1236 mutex_unlock(&fs_info->scrub_lock);
1237 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1238 scrub_workers_put(root);
1239 return PTR_ERR(sdev);
1240 }
1241 sdev->readonly = readonly;
1242 dev->scrub_device = sdev;
1243
1244 atomic_inc(&fs_info->scrubs_running);
1245 mutex_unlock(&fs_info->scrub_lock);
1246 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1247
1248 down_read(&fs_info->scrub_super_lock);
1249 ret = scrub_supers(sdev);
1250 up_read(&fs_info->scrub_super_lock);
1251
1252 if (!ret)
1253 ret = scrub_enumerate_chunks(sdev, start, end);
1254
1255 wait_event(sdev->list_wait, atomic_read(&sdev->in_flight) == 0);
1256
1257 atomic_dec(&fs_info->scrubs_running);
1258 wake_up(&fs_info->scrub_pause_wait);
1259
1260 if (progress)
1261 memcpy(progress, &sdev->stat, sizeof(*progress));
1262
1263 mutex_lock(&fs_info->scrub_lock);
1264 dev->scrub_device = NULL;
1265 mutex_unlock(&fs_info->scrub_lock);
1266
1267 scrub_free_dev(sdev);
1268 scrub_workers_put(root);
1269
1270 return ret;
1271}
1272
1273int btrfs_scrub_pause(struct btrfs_root *root)
1274{
1275 struct btrfs_fs_info *fs_info = root->fs_info;
1276
1277 mutex_lock(&fs_info->scrub_lock);
1278 atomic_inc(&fs_info->scrub_pause_req);
1279 while (atomic_read(&fs_info->scrubs_paused) !=
1280 atomic_read(&fs_info->scrubs_running)) {
1281 mutex_unlock(&fs_info->scrub_lock);
1282 wait_event(fs_info->scrub_pause_wait,
1283 atomic_read(&fs_info->scrubs_paused) ==
1284 atomic_read(&fs_info->scrubs_running));
1285 mutex_lock(&fs_info->scrub_lock);
1286 }
1287 mutex_unlock(&fs_info->scrub_lock);
1288
1289 return 0;
1290}
1291
1292int btrfs_scrub_continue(struct btrfs_root *root)
1293{
1294 struct btrfs_fs_info *fs_info = root->fs_info;
1295
1296 atomic_dec(&fs_info->scrub_pause_req);
1297 wake_up(&fs_info->scrub_pause_wait);
1298 return 0;
1299}
1300
1301int btrfs_scrub_pause_super(struct btrfs_root *root)
1302{
1303 down_write(&root->fs_info->scrub_super_lock);
1304 return 0;
1305}
1306
1307int btrfs_scrub_continue_super(struct btrfs_root *root)
1308{
1309 up_write(&root->fs_info->scrub_super_lock);
1310 return 0;
1311}
1312
1313int btrfs_scrub_cancel(struct btrfs_root *root)
1314{
1315 struct btrfs_fs_info *fs_info = root->fs_info;
1316
1317 mutex_lock(&fs_info->scrub_lock);
1318 if (!atomic_read(&fs_info->scrubs_running)) {
1319 mutex_unlock(&fs_info->scrub_lock);
1320 return -ENOTCONN;
1321 }
1322
1323 atomic_inc(&fs_info->scrub_cancel_req);
1324 while (atomic_read(&fs_info->scrubs_running)) {
1325 mutex_unlock(&fs_info->scrub_lock);
1326 wait_event(fs_info->scrub_pause_wait,
1327 atomic_read(&fs_info->scrubs_running) == 0);
1328 mutex_lock(&fs_info->scrub_lock);
1329 }
1330 atomic_dec(&fs_info->scrub_cancel_req);
1331 mutex_unlock(&fs_info->scrub_lock);
1332
1333 return 0;
1334}
1335
1336int btrfs_scrub_cancel_dev(struct btrfs_root *root, struct btrfs_device *dev)
1337{
1338 struct btrfs_fs_info *fs_info = root->fs_info;
1339 struct scrub_dev *sdev;
1340
1341 mutex_lock(&fs_info->scrub_lock);
1342 sdev = dev->scrub_device;
1343 if (!sdev) {
1344 mutex_unlock(&fs_info->scrub_lock);
1345 return -ENOTCONN;
1346 }
1347 atomic_inc(&sdev->cancel_req);
1348 while (dev->scrub_device) {
1349 mutex_unlock(&fs_info->scrub_lock);
1350 wait_event(fs_info->scrub_pause_wait,
1351 dev->scrub_device == NULL);
1352 mutex_lock(&fs_info->scrub_lock);
1353 }
1354 mutex_unlock(&fs_info->scrub_lock);
1355
1356 return 0;
1357}
1358int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
1359{
1360 struct btrfs_fs_info *fs_info = root->fs_info;
1361 struct btrfs_device *dev;
1362 int ret;
1363
1364 /*
1365 * we have to hold the device_list_mutex here so the device
1366 * does not go away in cancel_dev. FIXME: find a better solution
1367 */
1368 mutex_lock(&fs_info->fs_devices->device_list_mutex);
1369 dev = btrfs_find_device(root, devid, NULL, NULL);
1370 if (!dev) {
1371 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1372 return -ENODEV;
1373 }
1374 ret = btrfs_scrub_cancel_dev(root, dev);
1375 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
1376
1377 return ret;
1378}
1379
1380int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
1381 struct btrfs_scrub_progress *progress)
1382{
1383 struct btrfs_device *dev;
1384 struct scrub_dev *sdev = NULL;
1385
1386 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
1387 dev = btrfs_find_device(root, devid, NULL, NULL);
1388 if (dev)
1389 sdev = dev->scrub_device;
1390 if (sdev)
1391 memcpy(progress, &sdev->stat, sizeof(*progress));
1392 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
1393
1394 return dev ? (sdev ? 0 : -ENOTCONN) : -ENODEV;
1395}