1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
9 #include <crypto/hash.h>
14 #include "ordered-data.h"
15 #include "transaction.h"
17 #include "extent_io.h"
18 #include "dev-replace.h"
19 #include "check-integrity.h"
20 #include "rcu-string.h"
22 #include "block-group.h"
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
31 * Future enhancements:
32 * - In case an unrepairable extent is encountered, track which files are
33 * affected and report them
34 * - track and record media errors, throw out bad devices
35 * - add a mode to also read unallocated space
42 * The following three values only influence the performance.
44 * The last one configures the number of parallel and outstanding I/O
45 * operations. The first one configures an upper limit for the number
46 * of (dynamically allocated) pages that are added to a bio.
48 #define SCRUB_SECTORS_PER_BIO 32 /* 128KiB per bio for 4KiB pages */
49 #define SCRUB_BIOS_PER_SCTX 64 /* 8MiB per device in flight for 4KiB pages */
52 * The following value times PAGE_SIZE needs to be large enough to match the
53 * largest node/leaf/sector size that shall be supported.
55 #define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
57 #define SCRUB_MAX_PAGES (DIV_ROUND_UP(BTRFS_MAX_METADATA_BLOCKSIZE, PAGE_SIZE))
59 struct scrub_recover {
61 struct btrfs_io_context *bioc;
66 struct scrub_block *sblock;
67 struct list_head list;
68 u64 flags; /* extent flags */
70 /* Offset in bytes to @sblock. */
73 unsigned int have_csum:1;
74 unsigned int io_error:1;
75 u8 csum[BTRFS_CSUM_SIZE];
77 struct scrub_recover *recover;
82 struct scrub_ctx *sctx;
83 struct btrfs_device *dev;
88 struct scrub_sector *sectors[SCRUB_SECTORS_PER_BIO];
91 struct work_struct work;
96 * Each page will have its page::private used to record the logical
99 struct page *pages[SCRUB_MAX_PAGES];
100 struct scrub_sector *sectors[SCRUB_MAX_SECTORS_PER_BLOCK];
101 struct btrfs_device *dev;
102 /* Logical bytenr of the sblock */
105 u64 physical_for_dev_replace;
106 /* Length of sblock in bytes */
111 atomic_t outstanding_sectors;
112 refcount_t refs; /* free mem on transition to zero */
113 struct scrub_ctx *sctx;
114 struct scrub_parity *sparity;
116 unsigned int header_error:1;
117 unsigned int checksum_error:1;
118 unsigned int no_io_error_seen:1;
119 unsigned int generation_error:1; /* also sets header_error */
121 /* The following is for the data used to check parity */
122 /* It is for the data with checksum */
123 unsigned int data_corrected:1;
125 struct work_struct work;
128 /* Used for the chunks with parity stripe such RAID5/6 */
129 struct scrub_parity {
130 struct scrub_ctx *sctx;
132 struct btrfs_device *scrub_dev;
144 struct list_head sectors_list;
146 /* Work of parity check and repair */
147 struct work_struct work;
149 /* Mark the parity blocks which have data */
150 unsigned long dbitmap;
153 * Mark the parity blocks which have data, but errors happen when
154 * read data or check data
156 unsigned long ebitmap;
160 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
161 struct btrfs_fs_info *fs_info;
164 atomic_t bios_in_flight;
165 atomic_t workers_pending;
166 spinlock_t list_lock;
167 wait_queue_head_t list_wait;
168 struct list_head csum_list;
173 /* State of IO submission throttling affecting the associated device */
174 ktime_t throttle_deadline;
180 struct scrub_bio *wr_curr_bio;
181 struct mutex wr_lock;
182 struct btrfs_device *wr_tgtdev;
183 bool flush_all_writes;
188 struct btrfs_scrub_progress stat;
189 spinlock_t stat_lock;
192 * Use a ref counter to avoid use-after-free issues. Scrub workers
193 * decrement bios_in_flight and workers_pending and then do a wakeup
194 * on the list_wait wait queue. We must ensure the main scrub task
195 * doesn't free the scrub context before or while the workers are
196 * doing the wakeup() call.
201 struct scrub_warning {
202 struct btrfs_path *path;
203 u64 extent_item_size;
207 struct btrfs_device *dev;
210 struct full_stripe_lock {
218 /* This structure is for archtectures whose (void *) is smaller than u64 */
219 struct scrub_page_private {
224 static int attach_scrub_page_private(struct page *page, u64 logical)
227 attach_page_private(page, (void *)logical);
230 struct scrub_page_private *spp;
232 spp = kmalloc(sizeof(*spp), GFP_KERNEL);
235 spp->logical = logical;
236 attach_page_private(page, (void *)spp);
241 static void detach_scrub_page_private(struct page *page)
244 detach_page_private(page);
247 struct scrub_page_private *spp;
249 spp = detach_page_private(page);
255 static struct scrub_block *alloc_scrub_block(struct scrub_ctx *sctx,
256 struct btrfs_device *dev,
257 u64 logical, u64 physical,
258 u64 physical_for_dev_replace,
261 struct scrub_block *sblock;
263 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
266 refcount_set(&sblock->refs, 1);
268 sblock->logical = logical;
269 sblock->physical = physical;
270 sblock->physical_for_dev_replace = physical_for_dev_replace;
272 sblock->mirror_num = mirror_num;
273 sblock->no_io_error_seen = 1;
275 * Scrub_block::pages will be allocated at alloc_scrub_sector() when
276 * the corresponding page is not allocated.
282 * Allocate a new scrub sector and attach it to @sblock.
284 * Will also allocate new pages for @sblock if needed.
286 static struct scrub_sector *alloc_scrub_sector(struct scrub_block *sblock,
287 u64 logical, gfp_t gfp)
289 const pgoff_t page_index = (logical - sblock->logical) >> PAGE_SHIFT;
290 struct scrub_sector *ssector;
292 /* We must never have scrub_block exceed U32_MAX in size. */
293 ASSERT(logical - sblock->logical < U32_MAX);
295 ssector = kzalloc(sizeof(*ssector), gfp);
299 /* Allocate a new page if the slot is not allocated */
300 if (!sblock->pages[page_index]) {
303 sblock->pages[page_index] = alloc_page(gfp);
304 if (!sblock->pages[page_index]) {
308 ret = attach_scrub_page_private(sblock->pages[page_index],
309 sblock->logical + (page_index << PAGE_SHIFT));
312 __free_page(sblock->pages[page_index]);
313 sblock->pages[page_index] = NULL;
318 atomic_set(&ssector->refs, 1);
319 ssector->sblock = sblock;
320 /* The sector to be added should not be used */
321 ASSERT(sblock->sectors[sblock->sector_count] == NULL);
322 ssector->offset = logical - sblock->logical;
324 /* The sector count must be smaller than the limit */
325 ASSERT(sblock->sector_count < SCRUB_MAX_SECTORS_PER_BLOCK);
327 sblock->sectors[sblock->sector_count] = ssector;
328 sblock->sector_count++;
329 sblock->len += sblock->sctx->fs_info->sectorsize;
334 static struct page *scrub_sector_get_page(struct scrub_sector *ssector)
336 struct scrub_block *sblock = ssector->sblock;
339 * When calling this function, ssector must be alreaday attached to the
344 /* The range should be inside the sblock range */
345 ASSERT(ssector->offset < sblock->len);
347 index = ssector->offset >> PAGE_SHIFT;
348 ASSERT(index < SCRUB_MAX_PAGES);
349 ASSERT(sblock->pages[index]);
350 ASSERT(PagePrivate(sblock->pages[index]));
351 return sblock->pages[index];
354 static unsigned int scrub_sector_get_page_offset(struct scrub_sector *ssector)
356 struct scrub_block *sblock = ssector->sblock;
359 * When calling this function, ssector must be already attached to the
364 /* The range should be inside the sblock range */
365 ASSERT(ssector->offset < sblock->len);
367 return offset_in_page(ssector->offset);
370 static char *scrub_sector_get_kaddr(struct scrub_sector *ssector)
372 return page_address(scrub_sector_get_page(ssector)) +
373 scrub_sector_get_page_offset(ssector);
376 static int bio_add_scrub_sector(struct bio *bio, struct scrub_sector *ssector,
379 return bio_add_page(bio, scrub_sector_get_page(ssector), len,
380 scrub_sector_get_page_offset(ssector));
383 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
384 struct scrub_block *sblocks_for_recheck[]);
385 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
386 struct scrub_block *sblock,
387 int retry_failed_mirror);
388 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
389 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
390 struct scrub_block *sblock_good);
391 static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
392 struct scrub_block *sblock_good,
393 int sector_num, int force_write);
394 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
395 static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock,
397 static int scrub_checksum_data(struct scrub_block *sblock);
398 static int scrub_checksum_tree_block(struct scrub_block *sblock);
399 static int scrub_checksum_super(struct scrub_block *sblock);
400 static void scrub_block_put(struct scrub_block *sblock);
401 static void scrub_sector_get(struct scrub_sector *sector);
402 static void scrub_sector_put(struct scrub_sector *sector);
403 static void scrub_parity_get(struct scrub_parity *sparity);
404 static void scrub_parity_put(struct scrub_parity *sparity);
405 static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
406 u64 physical, struct btrfs_device *dev, u64 flags,
407 u64 gen, int mirror_num, u8 *csum,
408 u64 physical_for_dev_replace);
409 static void scrub_bio_end_io(struct bio *bio);
410 static void scrub_bio_end_io_worker(struct work_struct *work);
411 static void scrub_block_complete(struct scrub_block *sblock);
412 static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
413 u64 extent_logical, u32 extent_len,
414 u64 *extent_physical,
415 struct btrfs_device **extent_dev,
416 int *extent_mirror_num);
417 static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
418 struct scrub_sector *sector);
419 static void scrub_wr_submit(struct scrub_ctx *sctx);
420 static void scrub_wr_bio_end_io(struct bio *bio);
421 static void scrub_wr_bio_end_io_worker(struct work_struct *work);
422 static void scrub_put_ctx(struct scrub_ctx *sctx);
424 static inline int scrub_is_page_on_raid56(struct scrub_sector *sector)
426 return sector->recover &&
427 (sector->recover->bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
430 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
432 refcount_inc(&sctx->refs);
433 atomic_inc(&sctx->bios_in_flight);
436 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
438 atomic_dec(&sctx->bios_in_flight);
439 wake_up(&sctx->list_wait);
443 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
445 while (atomic_read(&fs_info->scrub_pause_req)) {
446 mutex_unlock(&fs_info->scrub_lock);
447 wait_event(fs_info->scrub_pause_wait,
448 atomic_read(&fs_info->scrub_pause_req) == 0);
449 mutex_lock(&fs_info->scrub_lock);
453 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
455 atomic_inc(&fs_info->scrubs_paused);
456 wake_up(&fs_info->scrub_pause_wait);
459 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
461 mutex_lock(&fs_info->scrub_lock);
462 __scrub_blocked_if_needed(fs_info);
463 atomic_dec(&fs_info->scrubs_paused);
464 mutex_unlock(&fs_info->scrub_lock);
466 wake_up(&fs_info->scrub_pause_wait);
469 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
471 scrub_pause_on(fs_info);
472 scrub_pause_off(fs_info);
476 * Insert new full stripe lock into full stripe locks tree
478 * Return pointer to existing or newly inserted full_stripe_lock structure if
479 * everything works well.
480 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
482 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
485 static struct full_stripe_lock *insert_full_stripe_lock(
486 struct btrfs_full_stripe_locks_tree *locks_root,
490 struct rb_node *parent = NULL;
491 struct full_stripe_lock *entry;
492 struct full_stripe_lock *ret;
494 lockdep_assert_held(&locks_root->lock);
496 p = &locks_root->root.rb_node;
499 entry = rb_entry(parent, struct full_stripe_lock, node);
500 if (fstripe_logical < entry->logical) {
502 } else if (fstripe_logical > entry->logical) {
513 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
515 return ERR_PTR(-ENOMEM);
516 ret->logical = fstripe_logical;
518 mutex_init(&ret->mutex);
520 rb_link_node(&ret->node, parent, p);
521 rb_insert_color(&ret->node, &locks_root->root);
526 * Search for a full stripe lock of a block group
528 * Return pointer to existing full stripe lock if found
529 * Return NULL if not found
531 static struct full_stripe_lock *search_full_stripe_lock(
532 struct btrfs_full_stripe_locks_tree *locks_root,
535 struct rb_node *node;
536 struct full_stripe_lock *entry;
538 lockdep_assert_held(&locks_root->lock);
540 node = locks_root->root.rb_node;
542 entry = rb_entry(node, struct full_stripe_lock, node);
543 if (fstripe_logical < entry->logical)
544 node = node->rb_left;
545 else if (fstripe_logical > entry->logical)
546 node = node->rb_right;
554 * Helper to get full stripe logical from a normal bytenr.
556 * Caller must ensure @cache is a RAID56 block group.
558 static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
563 * Due to chunk item size limit, full stripe length should not be
564 * larger than U32_MAX. Just a sanity check here.
566 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
569 * round_down() can only handle power of 2, while RAID56 full
570 * stripe length can be 64KiB * n, so we need to manually round down.
572 ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
573 cache->full_stripe_len + cache->start;
578 * Lock a full stripe to avoid concurrency of recovery and read
580 * It's only used for profiles with parities (RAID5/6), for other profiles it
583 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
584 * So caller must call unlock_full_stripe() at the same context.
586 * Return <0 if encounters error.
588 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
591 struct btrfs_block_group *bg_cache;
592 struct btrfs_full_stripe_locks_tree *locks_root;
593 struct full_stripe_lock *existing;
598 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
604 /* Profiles not based on parity don't need full stripe lock */
605 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
607 locks_root = &bg_cache->full_stripe_locks_root;
609 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
611 /* Now insert the full stripe lock */
612 mutex_lock(&locks_root->lock);
613 existing = insert_full_stripe_lock(locks_root, fstripe_start);
614 mutex_unlock(&locks_root->lock);
615 if (IS_ERR(existing)) {
616 ret = PTR_ERR(existing);
619 mutex_lock(&existing->mutex);
622 btrfs_put_block_group(bg_cache);
627 * Unlock a full stripe.
629 * NOTE: Caller must ensure it's the same context calling corresponding
630 * lock_full_stripe().
632 * Return 0 if we unlock full stripe without problem.
633 * Return <0 for error
635 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
638 struct btrfs_block_group *bg_cache;
639 struct btrfs_full_stripe_locks_tree *locks_root;
640 struct full_stripe_lock *fstripe_lock;
645 /* If we didn't acquire full stripe lock, no need to continue */
649 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
654 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
657 locks_root = &bg_cache->full_stripe_locks_root;
658 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
660 mutex_lock(&locks_root->lock);
661 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
662 /* Unpaired unlock_full_stripe() detected */
666 mutex_unlock(&locks_root->lock);
670 if (fstripe_lock->refs == 0) {
672 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
673 fstripe_lock->logical);
675 fstripe_lock->refs--;
678 if (fstripe_lock->refs == 0) {
679 rb_erase(&fstripe_lock->node, &locks_root->root);
682 mutex_unlock(&locks_root->lock);
684 mutex_unlock(&fstripe_lock->mutex);
688 btrfs_put_block_group(bg_cache);
692 static void scrub_free_csums(struct scrub_ctx *sctx)
694 while (!list_empty(&sctx->csum_list)) {
695 struct btrfs_ordered_sum *sum;
696 sum = list_first_entry(&sctx->csum_list,
697 struct btrfs_ordered_sum, list);
698 list_del(&sum->list);
703 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
710 /* this can happen when scrub is cancelled */
711 if (sctx->curr != -1) {
712 struct scrub_bio *sbio = sctx->bios[sctx->curr];
714 for (i = 0; i < sbio->sector_count; i++)
715 scrub_block_put(sbio->sectors[i]->sblock);
719 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
720 struct scrub_bio *sbio = sctx->bios[i];
727 kfree(sctx->wr_curr_bio);
728 scrub_free_csums(sctx);
732 static void scrub_put_ctx(struct scrub_ctx *sctx)
734 if (refcount_dec_and_test(&sctx->refs))
735 scrub_free_ctx(sctx);
738 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
739 struct btrfs_fs_info *fs_info, int is_dev_replace)
741 struct scrub_ctx *sctx;
744 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
747 refcount_set(&sctx->refs, 1);
748 sctx->is_dev_replace = is_dev_replace;
749 sctx->sectors_per_bio = SCRUB_SECTORS_PER_BIO;
751 sctx->fs_info = fs_info;
752 INIT_LIST_HEAD(&sctx->csum_list);
753 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
754 struct scrub_bio *sbio;
756 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
759 sctx->bios[i] = sbio;
763 sbio->sector_count = 0;
764 INIT_WORK(&sbio->work, scrub_bio_end_io_worker);
766 if (i != SCRUB_BIOS_PER_SCTX - 1)
767 sctx->bios[i]->next_free = i + 1;
769 sctx->bios[i]->next_free = -1;
771 sctx->first_free = 0;
772 atomic_set(&sctx->bios_in_flight, 0);
773 atomic_set(&sctx->workers_pending, 0);
774 atomic_set(&sctx->cancel_req, 0);
776 spin_lock_init(&sctx->list_lock);
777 spin_lock_init(&sctx->stat_lock);
778 init_waitqueue_head(&sctx->list_wait);
779 sctx->throttle_deadline = 0;
781 WARN_ON(sctx->wr_curr_bio != NULL);
782 mutex_init(&sctx->wr_lock);
783 sctx->wr_curr_bio = NULL;
784 if (is_dev_replace) {
785 WARN_ON(!fs_info->dev_replace.tgtdev);
786 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
787 sctx->flush_all_writes = false;
793 scrub_free_ctx(sctx);
794 return ERR_PTR(-ENOMEM);
797 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
804 struct extent_buffer *eb;
805 struct btrfs_inode_item *inode_item;
806 struct scrub_warning *swarn = warn_ctx;
807 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
808 struct inode_fs_paths *ipath = NULL;
809 struct btrfs_root *local_root;
810 struct btrfs_key key;
812 local_root = btrfs_get_fs_root(fs_info, root, true);
813 if (IS_ERR(local_root)) {
814 ret = PTR_ERR(local_root);
819 * this makes the path point to (inum INODE_ITEM ioff)
822 key.type = BTRFS_INODE_ITEM_KEY;
825 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
827 btrfs_put_root(local_root);
828 btrfs_release_path(swarn->path);
832 eb = swarn->path->nodes[0];
833 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
834 struct btrfs_inode_item);
835 nlink = btrfs_inode_nlink(eb, inode_item);
836 btrfs_release_path(swarn->path);
839 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
840 * uses GFP_NOFS in this context, so we keep it consistent but it does
841 * not seem to be strictly necessary.
843 nofs_flag = memalloc_nofs_save();
844 ipath = init_ipath(4096, local_root, swarn->path);
845 memalloc_nofs_restore(nofs_flag);
847 btrfs_put_root(local_root);
848 ret = PTR_ERR(ipath);
852 ret = paths_from_inode(inum, ipath);
858 * we deliberately ignore the bit ipath might have been too small to
859 * hold all of the paths here
861 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
862 btrfs_warn_in_rcu(fs_info,
863 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
864 swarn->errstr, swarn->logical,
865 rcu_str_deref(swarn->dev->name),
868 fs_info->sectorsize, nlink,
869 (char *)(unsigned long)ipath->fspath->val[i]);
871 btrfs_put_root(local_root);
876 btrfs_warn_in_rcu(fs_info,
877 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
878 swarn->errstr, swarn->logical,
879 rcu_str_deref(swarn->dev->name),
881 root, inum, offset, ret);
887 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
889 struct btrfs_device *dev;
890 struct btrfs_fs_info *fs_info;
891 struct btrfs_path *path;
892 struct btrfs_key found_key;
893 struct extent_buffer *eb;
894 struct btrfs_extent_item *ei;
895 struct scrub_warning swarn;
896 unsigned long ptr = 0;
904 WARN_ON(sblock->sector_count < 1);
906 fs_info = sblock->sctx->fs_info;
908 /* Super block error, no need to search extent tree. */
909 if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
910 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
911 errstr, rcu_str_deref(dev->name),
915 path = btrfs_alloc_path();
919 swarn.physical = sblock->physical;
920 swarn.logical = sblock->logical;
921 swarn.errstr = errstr;
924 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
929 extent_item_pos = swarn.logical - found_key.objectid;
930 swarn.extent_item_size = found_key.offset;
933 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
934 item_size = btrfs_item_size(eb, path->slots[0]);
936 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
938 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
939 item_size, &ref_root,
941 btrfs_warn_in_rcu(fs_info,
942 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
943 errstr, swarn.logical,
944 rcu_str_deref(dev->name),
946 ref_level ? "node" : "leaf",
947 ret < 0 ? -1 : ref_level,
948 ret < 0 ? -1 : ref_root);
950 btrfs_release_path(path);
952 btrfs_release_path(path);
955 iterate_extent_inodes(fs_info, found_key.objectid,
957 scrub_print_warning_inode, &swarn, false);
961 btrfs_free_path(path);
964 static inline void scrub_get_recover(struct scrub_recover *recover)
966 refcount_inc(&recover->refs);
969 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
970 struct scrub_recover *recover)
972 if (refcount_dec_and_test(&recover->refs)) {
973 btrfs_bio_counter_dec(fs_info);
974 btrfs_put_bioc(recover->bioc);
980 * scrub_handle_errored_block gets called when either verification of the
981 * sectors failed or the bio failed to read, e.g. with EIO. In the latter
982 * case, this function handles all sectors in the bio, even though only one
984 * The goal of this function is to repair the errored block by using the
985 * contents of one of the mirrors.
987 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
989 struct scrub_ctx *sctx = sblock_to_check->sctx;
990 struct btrfs_device *dev = sblock_to_check->dev;
991 struct btrfs_fs_info *fs_info;
993 unsigned int failed_mirror_index;
994 unsigned int is_metadata;
995 unsigned int have_csum;
996 /* One scrub_block for each mirror */
997 struct scrub_block *sblocks_for_recheck[BTRFS_MAX_MIRRORS] = { 0 };
998 struct scrub_block *sblock_bad;
1003 bool full_stripe_locked;
1004 unsigned int nofs_flag;
1005 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
1006 DEFAULT_RATELIMIT_BURST);
1008 BUG_ON(sblock_to_check->sector_count < 1);
1009 fs_info = sctx->fs_info;
1010 if (sblock_to_check->sectors[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1012 * If we find an error in a super block, we just report it.
1013 * They will get written with the next transaction commit
1016 scrub_print_warning("super block error", sblock_to_check);
1017 spin_lock(&sctx->stat_lock);
1018 ++sctx->stat.super_errors;
1019 spin_unlock(&sctx->stat_lock);
1020 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_CORRUPTION_ERRS);
1023 logical = sblock_to_check->logical;
1024 ASSERT(sblock_to_check->mirror_num);
1025 failed_mirror_index = sblock_to_check->mirror_num - 1;
1026 is_metadata = !(sblock_to_check->sectors[0]->flags &
1027 BTRFS_EXTENT_FLAG_DATA);
1028 have_csum = sblock_to_check->sectors[0]->have_csum;
1030 if (!sctx->is_dev_replace && btrfs_repair_one_zone(fs_info, logical))
1034 * We must use GFP_NOFS because the scrub task might be waiting for a
1035 * worker task executing this function and in turn a transaction commit
1036 * might be waiting the scrub task to pause (which needs to wait for all
1037 * the worker tasks to complete before pausing).
1038 * We do allocations in the workers through insert_full_stripe_lock()
1039 * and scrub_add_sector_to_wr_bio(), which happens down the call chain of
1042 nofs_flag = memalloc_nofs_save();
1044 * For RAID5/6, race can happen for a different device scrub thread.
1045 * For data corruption, Parity and Data threads will both try
1046 * to recovery the data.
1047 * Race can lead to doubly added csum error, or even unrecoverable
1050 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1052 memalloc_nofs_restore(nofs_flag);
1053 spin_lock(&sctx->stat_lock);
1055 sctx->stat.malloc_errors++;
1056 sctx->stat.read_errors++;
1057 sctx->stat.uncorrectable_errors++;
1058 spin_unlock(&sctx->stat_lock);
1063 * read all mirrors one after the other. This includes to
1064 * re-read the extent or metadata block that failed (that was
1065 * the cause that this fixup code is called) another time,
1066 * sector by sector this time in order to know which sectors
1067 * caused I/O errors and which ones are good (for all mirrors).
1068 * It is the goal to handle the situation when more than one
1069 * mirror contains I/O errors, but the errors do not
1070 * overlap, i.e. the data can be repaired by selecting the
1071 * sectors from those mirrors without I/O error on the
1072 * particular sectors. One example (with blocks >= 2 * sectorsize)
1073 * would be that mirror #1 has an I/O error on the first sector,
1074 * the second sector is good, and mirror #2 has an I/O error on
1075 * the second sector, but the first sector is good.
1076 * Then the first sector of the first mirror can be repaired by
1077 * taking the first sector of the second mirror, and the
1078 * second sector of the second mirror can be repaired by
1079 * copying the contents of the 2nd sector of the 1st mirror.
1080 * One more note: if the sectors of one mirror contain I/O
1081 * errors, the checksum cannot be verified. In order to get
1082 * the best data for repairing, the first attempt is to find
1083 * a mirror without I/O errors and with a validated checksum.
1084 * Only if this is not possible, the sectors are picked from
1085 * mirrors with I/O errors without considering the checksum.
1086 * If the latter is the case, at the end, the checksum of the
1087 * repaired area is verified in order to correctly maintain
1090 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1092 * Note: the two members refs and outstanding_sectors are not
1093 * used in the blocks that are used for the recheck procedure.
1095 * But alloc_scrub_block() will initialize sblock::ref anyway,
1096 * so we can use scrub_block_put() to clean them up.
1098 * And here we don't setup the physical/dev for the sblock yet,
1099 * they will be correctly initialized in scrub_setup_recheck_block().
1101 sblocks_for_recheck[mirror_index] = alloc_scrub_block(sctx, NULL,
1102 logical, 0, 0, mirror_index);
1103 if (!sblocks_for_recheck[mirror_index]) {
1104 spin_lock(&sctx->stat_lock);
1105 sctx->stat.malloc_errors++;
1106 sctx->stat.read_errors++;
1107 sctx->stat.uncorrectable_errors++;
1108 spin_unlock(&sctx->stat_lock);
1109 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1114 /* Setup the context, map the logical blocks and alloc the sectors */
1115 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1117 spin_lock(&sctx->stat_lock);
1118 sctx->stat.read_errors++;
1119 sctx->stat.uncorrectable_errors++;
1120 spin_unlock(&sctx->stat_lock);
1121 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1124 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1125 sblock_bad = sblocks_for_recheck[failed_mirror_index];
1127 /* build and submit the bios for the failed mirror, check checksums */
1128 scrub_recheck_block(fs_info, sblock_bad, 1);
1130 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1131 sblock_bad->no_io_error_seen) {
1133 * The error disappeared after reading sector by sector, or
1134 * the area was part of a huge bio and other parts of the
1135 * bio caused I/O errors, or the block layer merged several
1136 * read requests into one and the error is caused by a
1137 * different bio (usually one of the two latter cases is
1140 spin_lock(&sctx->stat_lock);
1141 sctx->stat.unverified_errors++;
1142 sblock_to_check->data_corrected = 1;
1143 spin_unlock(&sctx->stat_lock);
1145 if (sctx->is_dev_replace)
1146 scrub_write_block_to_dev_replace(sblock_bad);
1150 if (!sblock_bad->no_io_error_seen) {
1151 spin_lock(&sctx->stat_lock);
1152 sctx->stat.read_errors++;
1153 spin_unlock(&sctx->stat_lock);
1154 if (__ratelimit(&rs))
1155 scrub_print_warning("i/o error", sblock_to_check);
1156 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1157 } else if (sblock_bad->checksum_error) {
1158 spin_lock(&sctx->stat_lock);
1159 sctx->stat.csum_errors++;
1160 spin_unlock(&sctx->stat_lock);
1161 if (__ratelimit(&rs))
1162 scrub_print_warning("checksum error", sblock_to_check);
1163 btrfs_dev_stat_inc_and_print(dev,
1164 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1165 } else if (sblock_bad->header_error) {
1166 spin_lock(&sctx->stat_lock);
1167 sctx->stat.verify_errors++;
1168 spin_unlock(&sctx->stat_lock);
1169 if (__ratelimit(&rs))
1170 scrub_print_warning("checksum/header error",
1172 if (sblock_bad->generation_error)
1173 btrfs_dev_stat_inc_and_print(dev,
1174 BTRFS_DEV_STAT_GENERATION_ERRS);
1176 btrfs_dev_stat_inc_and_print(dev,
1177 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1180 if (sctx->readonly) {
1181 ASSERT(!sctx->is_dev_replace);
1186 * now build and submit the bios for the other mirrors, check
1188 * First try to pick the mirror which is completely without I/O
1189 * errors and also does not have a checksum error.
1190 * If one is found, and if a checksum is present, the full block
1191 * that is known to contain an error is rewritten. Afterwards
1192 * the block is known to be corrected.
1193 * If a mirror is found which is completely correct, and no
1194 * checksum is present, only those sectors are rewritten that had
1195 * an I/O error in the block to be repaired, since it cannot be
1196 * determined, which copy of the other sectors is better (and it
1197 * could happen otherwise that a correct sector would be
1198 * overwritten by a bad one).
1200 for (mirror_index = 0; ;mirror_index++) {
1201 struct scrub_block *sblock_other;
1203 if (mirror_index == failed_mirror_index)
1206 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1207 if (!scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1208 if (mirror_index >= BTRFS_MAX_MIRRORS)
1210 if (!sblocks_for_recheck[mirror_index]->sector_count)
1213 sblock_other = sblocks_for_recheck[mirror_index];
1215 struct scrub_recover *r = sblock_bad->sectors[0]->recover;
1216 int max_allowed = r->bioc->num_stripes - r->bioc->num_tgtdevs;
1218 if (mirror_index >= max_allowed)
1220 if (!sblocks_for_recheck[1]->sector_count)
1223 ASSERT(failed_mirror_index == 0);
1224 sblock_other = sblocks_for_recheck[1];
1225 sblock_other->mirror_num = 1 + mirror_index;
1228 /* build and submit the bios, check checksums */
1229 scrub_recheck_block(fs_info, sblock_other, 0);
1231 if (!sblock_other->header_error &&
1232 !sblock_other->checksum_error &&
1233 sblock_other->no_io_error_seen) {
1234 if (sctx->is_dev_replace) {
1235 scrub_write_block_to_dev_replace(sblock_other);
1236 goto corrected_error;
1238 ret = scrub_repair_block_from_good_copy(
1239 sblock_bad, sblock_other);
1241 goto corrected_error;
1246 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1247 goto did_not_correct_error;
1250 * In case of I/O errors in the area that is supposed to be
1251 * repaired, continue by picking good copies of those sectors.
1252 * Select the good sectors from mirrors to rewrite bad sectors from
1253 * the area to fix. Afterwards verify the checksum of the block
1254 * that is supposed to be repaired. This verification step is
1255 * only done for the purpose of statistic counting and for the
1256 * final scrub report, whether errors remain.
1257 * A perfect algorithm could make use of the checksum and try
1258 * all possible combinations of sectors from the different mirrors
1259 * until the checksum verification succeeds. For example, when
1260 * the 2nd sector of mirror #1 faces I/O errors, and the 2nd sector
1261 * of mirror #2 is readable but the final checksum test fails,
1262 * then the 2nd sector of mirror #3 could be tried, whether now
1263 * the final checksum succeeds. But this would be a rare
1264 * exception and is therefore not implemented. At least it is
1265 * avoided that the good copy is overwritten.
1266 * A more useful improvement would be to pick the sectors
1267 * without I/O error based on sector sizes (512 bytes on legacy
1268 * disks) instead of on sectorsize. Then maybe 512 byte of one
1269 * mirror could be repaired by taking 512 byte of a different
1270 * mirror, even if other 512 byte sectors in the same sectorsize
1271 * area are unreadable.
1274 for (sector_num = 0; sector_num < sblock_bad->sector_count;
1276 struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1277 struct scrub_block *sblock_other = NULL;
1279 /* Skip no-io-error sectors in scrub */
1280 if (!sector_bad->io_error && !sctx->is_dev_replace)
1283 if (scrub_is_page_on_raid56(sblock_bad->sectors[0])) {
1285 * In case of dev replace, if raid56 rebuild process
1286 * didn't work out correct data, then copy the content
1287 * in sblock_bad to make sure target device is identical
1288 * to source device, instead of writing garbage data in
1289 * sblock_for_recheck array to target device.
1291 sblock_other = NULL;
1292 } else if (sector_bad->io_error) {
1293 /* Try to find no-io-error sector in mirrors */
1294 for (mirror_index = 0;
1295 mirror_index < BTRFS_MAX_MIRRORS &&
1296 sblocks_for_recheck[mirror_index]->sector_count > 0;
1298 if (!sblocks_for_recheck[mirror_index]->
1299 sectors[sector_num]->io_error) {
1300 sblock_other = sblocks_for_recheck[mirror_index];
1308 if (sctx->is_dev_replace) {
1310 * Did not find a mirror to fetch the sector from.
1311 * scrub_write_sector_to_dev_replace() handles this
1312 * case (sector->io_error), by filling the block with
1313 * zeros before submitting the write request
1316 sblock_other = sblock_bad;
1318 if (scrub_write_sector_to_dev_replace(sblock_other,
1321 &fs_info->dev_replace.num_write_errors);
1324 } else if (sblock_other) {
1325 ret = scrub_repair_sector_from_good_copy(sblock_bad,
1329 sector_bad->io_error = 0;
1335 if (success && !sctx->is_dev_replace) {
1336 if (is_metadata || have_csum) {
1338 * need to verify the checksum now that all
1339 * sectors on disk are repaired (the write
1340 * request for data to be repaired is on its way).
1341 * Just be lazy and use scrub_recheck_block()
1342 * which re-reads the data before the checksum
1343 * is verified, but most likely the data comes out
1344 * of the page cache.
1346 scrub_recheck_block(fs_info, sblock_bad, 1);
1347 if (!sblock_bad->header_error &&
1348 !sblock_bad->checksum_error &&
1349 sblock_bad->no_io_error_seen)
1350 goto corrected_error;
1352 goto did_not_correct_error;
1355 spin_lock(&sctx->stat_lock);
1356 sctx->stat.corrected_errors++;
1357 sblock_to_check->data_corrected = 1;
1358 spin_unlock(&sctx->stat_lock);
1359 btrfs_err_rl_in_rcu(fs_info,
1360 "fixed up error at logical %llu on dev %s",
1361 logical, rcu_str_deref(dev->name));
1364 did_not_correct_error:
1365 spin_lock(&sctx->stat_lock);
1366 sctx->stat.uncorrectable_errors++;
1367 spin_unlock(&sctx->stat_lock);
1368 btrfs_err_rl_in_rcu(fs_info,
1369 "unable to fixup (regular) error at logical %llu on dev %s",
1370 logical, rcu_str_deref(dev->name));
1374 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; mirror_index++) {
1375 struct scrub_block *sblock = sblocks_for_recheck[mirror_index];
1376 struct scrub_recover *recover;
1379 /* Not allocated, continue checking the next mirror */
1383 for (sector_index = 0; sector_index < sblock->sector_count;
1386 * Here we just cleanup the recover, each sector will be
1387 * properly cleaned up by later scrub_block_put()
1389 recover = sblock->sectors[sector_index]->recover;
1391 scrub_put_recover(fs_info, recover);
1392 sblock->sectors[sector_index]->recover = NULL;
1395 scrub_block_put(sblock);
1398 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1399 memalloc_nofs_restore(nofs_flag);
1405 static inline int scrub_nr_raid_mirrors(struct btrfs_io_context *bioc)
1407 if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID5)
1409 else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID6)
1412 return (int)bioc->num_stripes;
1415 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1417 int nstripes, int mirror,
1423 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1425 for (i = 0; i < nstripes; i++) {
1426 if (raid_map[i] == RAID6_Q_STRIPE ||
1427 raid_map[i] == RAID5_P_STRIPE)
1430 if (logical >= raid_map[i] &&
1431 logical < raid_map[i] + BTRFS_STRIPE_LEN)
1436 *stripe_offset = logical - raid_map[i];
1438 /* The other RAID type */
1439 *stripe_index = mirror;
1444 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1445 struct scrub_block *sblocks_for_recheck[])
1447 struct scrub_ctx *sctx = original_sblock->sctx;
1448 struct btrfs_fs_info *fs_info = sctx->fs_info;
1449 u64 logical = original_sblock->logical;
1450 u64 length = original_sblock->sector_count << fs_info->sectorsize_bits;
1451 u64 generation = original_sblock->sectors[0]->generation;
1452 u64 flags = original_sblock->sectors[0]->flags;
1453 u64 have_csum = original_sblock->sectors[0]->have_csum;
1454 struct scrub_recover *recover;
1455 struct btrfs_io_context *bioc;
1460 int sector_index = 0;
1465 while (length > 0) {
1466 sublen = min_t(u64, length, fs_info->sectorsize);
1467 mapped_length = sublen;
1471 * With a length of sectorsize, each returned stripe represents
1474 btrfs_bio_counter_inc_blocked(fs_info);
1475 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1476 logical, &mapped_length, &bioc);
1477 if (ret || !bioc || mapped_length < sublen) {
1478 btrfs_put_bioc(bioc);
1479 btrfs_bio_counter_dec(fs_info);
1483 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1485 btrfs_put_bioc(bioc);
1486 btrfs_bio_counter_dec(fs_info);
1490 refcount_set(&recover->refs, 1);
1491 recover->bioc = bioc;
1492 recover->map_length = mapped_length;
1494 ASSERT(sector_index < SCRUB_MAX_SECTORS_PER_BLOCK);
1496 nmirrors = min(scrub_nr_raid_mirrors(bioc), BTRFS_MAX_MIRRORS);
1498 for (mirror_index = 0; mirror_index < nmirrors;
1500 struct scrub_block *sblock;
1501 struct scrub_sector *sector;
1503 sblock = sblocks_for_recheck[mirror_index];
1504 sblock->sctx = sctx;
1506 sector = alloc_scrub_sector(sblock, logical, GFP_NOFS);
1508 spin_lock(&sctx->stat_lock);
1509 sctx->stat.malloc_errors++;
1510 spin_unlock(&sctx->stat_lock);
1511 scrub_put_recover(fs_info, recover);
1514 sector->flags = flags;
1515 sector->generation = generation;
1516 sector->have_csum = have_csum;
1518 memcpy(sector->csum,
1519 original_sblock->sectors[0]->csum,
1520 sctx->fs_info->csum_size);
1522 scrub_stripe_index_and_offset(logical,
1531 * We're at the first sector, also populate @sblock
1534 if (sector_index == 0) {
1536 bioc->stripes[stripe_index].physical +
1538 sblock->dev = bioc->stripes[stripe_index].dev;
1539 sblock->physical_for_dev_replace =
1540 original_sblock->physical_for_dev_replace;
1543 BUG_ON(sector_index >= original_sblock->sector_count);
1544 scrub_get_recover(recover);
1545 sector->recover = recover;
1547 scrub_put_recover(fs_info, recover);
1556 static void scrub_bio_wait_endio(struct bio *bio)
1558 complete(bio->bi_private);
1561 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1563 struct scrub_sector *sector)
1565 DECLARE_COMPLETION_ONSTACK(done);
1567 bio->bi_iter.bi_sector = (sector->offset + sector->sblock->logical) >>
1569 bio->bi_private = &done;
1570 bio->bi_end_io = scrub_bio_wait_endio;
1571 raid56_parity_recover(bio, sector->recover->bioc, sector->sblock->mirror_num);
1573 wait_for_completion_io(&done);
1574 return blk_status_to_errno(bio->bi_status);
1577 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1578 struct scrub_block *sblock)
1580 struct scrub_sector *first_sector = sblock->sectors[0];
1584 /* All sectors in sblock belong to the same stripe on the same device. */
1585 ASSERT(sblock->dev);
1586 if (!sblock->dev->bdev)
1589 bio = bio_alloc(sblock->dev->bdev, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
1591 for (i = 0; i < sblock->sector_count; i++) {
1592 struct scrub_sector *sector = sblock->sectors[i];
1594 bio_add_scrub_sector(bio, sector, fs_info->sectorsize);
1597 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_sector)) {
1604 scrub_recheck_block_checksum(sblock);
1608 for (i = 0; i < sblock->sector_count; i++)
1609 sblock->sectors[i]->io_error = 1;
1611 sblock->no_io_error_seen = 0;
1615 * This function will check the on disk data for checksum errors, header errors
1616 * and read I/O errors. If any I/O errors happen, the exact sectors which are
1617 * errored are marked as being bad. The goal is to enable scrub to take those
1618 * sectors that are not errored from all the mirrors so that the sectors that
1619 * are errored in the just handled mirror can be repaired.
1621 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1622 struct scrub_block *sblock,
1623 int retry_failed_mirror)
1627 sblock->no_io_error_seen = 1;
1629 /* short cut for raid56 */
1630 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->sectors[0]))
1631 return scrub_recheck_block_on_raid56(fs_info, sblock);
1633 for (i = 0; i < sblock->sector_count; i++) {
1634 struct scrub_sector *sector = sblock->sectors[i];
1636 struct bio_vec bvec;
1638 if (sblock->dev->bdev == NULL) {
1639 sector->io_error = 1;
1640 sblock->no_io_error_seen = 0;
1644 bio_init(&bio, sblock->dev->bdev, &bvec, 1, REQ_OP_READ);
1645 bio_add_scrub_sector(&bio, sector, fs_info->sectorsize);
1646 bio.bi_iter.bi_sector = (sblock->physical + sector->offset) >>
1649 btrfsic_check_bio(&bio);
1650 if (submit_bio_wait(&bio)) {
1651 sector->io_error = 1;
1652 sblock->no_io_error_seen = 0;
1658 if (sblock->no_io_error_seen)
1659 scrub_recheck_block_checksum(sblock);
1662 static inline int scrub_check_fsid(u8 fsid[], struct scrub_sector *sector)
1664 struct btrfs_fs_devices *fs_devices = sector->sblock->dev->fs_devices;
1667 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1671 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1673 sblock->header_error = 0;
1674 sblock->checksum_error = 0;
1675 sblock->generation_error = 0;
1677 if (sblock->sectors[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1678 scrub_checksum_data(sblock);
1680 scrub_checksum_tree_block(sblock);
1683 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1684 struct scrub_block *sblock_good)
1689 for (i = 0; i < sblock_bad->sector_count; i++) {
1692 ret_sub = scrub_repair_sector_from_good_copy(sblock_bad,
1701 static int scrub_repair_sector_from_good_copy(struct scrub_block *sblock_bad,
1702 struct scrub_block *sblock_good,
1703 int sector_num, int force_write)
1705 struct scrub_sector *sector_bad = sblock_bad->sectors[sector_num];
1706 struct scrub_sector *sector_good = sblock_good->sectors[sector_num];
1707 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1708 const u32 sectorsize = fs_info->sectorsize;
1710 if (force_write || sblock_bad->header_error ||
1711 sblock_bad->checksum_error || sector_bad->io_error) {
1713 struct bio_vec bvec;
1716 if (!sblock_bad->dev->bdev) {
1717 btrfs_warn_rl(fs_info,
1718 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1722 bio_init(&bio, sblock_bad->dev->bdev, &bvec, 1, REQ_OP_WRITE);
1723 bio.bi_iter.bi_sector = (sblock_bad->physical +
1724 sector_bad->offset) >> SECTOR_SHIFT;
1725 ret = bio_add_scrub_sector(&bio, sector_good, sectorsize);
1727 btrfsic_check_bio(&bio);
1728 ret = submit_bio_wait(&bio);
1732 btrfs_dev_stat_inc_and_print(sblock_bad->dev,
1733 BTRFS_DEV_STAT_WRITE_ERRS);
1734 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1742 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1744 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1748 * This block is used for the check of the parity on the source device,
1749 * so the data needn't be written into the destination device.
1751 if (sblock->sparity)
1754 for (i = 0; i < sblock->sector_count; i++) {
1757 ret = scrub_write_sector_to_dev_replace(sblock, i);
1759 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1763 static int scrub_write_sector_to_dev_replace(struct scrub_block *sblock, int sector_num)
1765 const u32 sectorsize = sblock->sctx->fs_info->sectorsize;
1766 struct scrub_sector *sector = sblock->sectors[sector_num];
1768 if (sector->io_error)
1769 memset(scrub_sector_get_kaddr(sector), 0, sectorsize);
1771 return scrub_add_sector_to_wr_bio(sblock->sctx, sector);
1774 static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
1779 if (!btrfs_is_zoned(sctx->fs_info))
1782 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1785 if (sctx->write_pointer < physical) {
1786 length = physical - sctx->write_pointer;
1788 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1789 sctx->write_pointer, length);
1791 sctx->write_pointer = physical;
1796 static void scrub_block_get(struct scrub_block *sblock)
1798 refcount_inc(&sblock->refs);
1801 static int scrub_add_sector_to_wr_bio(struct scrub_ctx *sctx,
1802 struct scrub_sector *sector)
1804 struct scrub_block *sblock = sector->sblock;
1805 struct scrub_bio *sbio;
1807 const u32 sectorsize = sctx->fs_info->sectorsize;
1809 mutex_lock(&sctx->wr_lock);
1811 if (!sctx->wr_curr_bio) {
1812 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1814 if (!sctx->wr_curr_bio) {
1815 mutex_unlock(&sctx->wr_lock);
1818 sctx->wr_curr_bio->sctx = sctx;
1819 sctx->wr_curr_bio->sector_count = 0;
1821 sbio = sctx->wr_curr_bio;
1822 if (sbio->sector_count == 0) {
1823 ret = fill_writer_pointer_gap(sctx, sector->offset +
1824 sblock->physical_for_dev_replace);
1826 mutex_unlock(&sctx->wr_lock);
1830 sbio->physical = sblock->physical_for_dev_replace + sector->offset;
1831 sbio->logical = sblock->logical + sector->offset;
1832 sbio->dev = sctx->wr_tgtdev;
1834 sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
1835 REQ_OP_WRITE, GFP_NOFS);
1837 sbio->bio->bi_private = sbio;
1838 sbio->bio->bi_end_io = scrub_wr_bio_end_io;
1839 sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
1841 } else if (sbio->physical + sbio->sector_count * sectorsize !=
1842 sblock->physical_for_dev_replace + sector->offset ||
1843 sbio->logical + sbio->sector_count * sectorsize !=
1844 sblock->logical + sector->offset) {
1845 scrub_wr_submit(sctx);
1849 ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
1850 if (ret != sectorsize) {
1851 if (sbio->sector_count < 1) {
1854 mutex_unlock(&sctx->wr_lock);
1857 scrub_wr_submit(sctx);
1861 sbio->sectors[sbio->sector_count] = sector;
1862 scrub_sector_get(sector);
1864 * Since ssector no longer holds a page, but uses sblock::pages, we
1865 * have to ensure the sblock had not been freed before our write bio
1868 scrub_block_get(sector->sblock);
1870 sbio->sector_count++;
1871 if (sbio->sector_count == sctx->sectors_per_bio)
1872 scrub_wr_submit(sctx);
1873 mutex_unlock(&sctx->wr_lock);
1878 static void scrub_wr_submit(struct scrub_ctx *sctx)
1880 struct scrub_bio *sbio;
1882 if (!sctx->wr_curr_bio)
1885 sbio = sctx->wr_curr_bio;
1886 sctx->wr_curr_bio = NULL;
1887 scrub_pending_bio_inc(sctx);
1888 /* process all writes in a single worker thread. Then the block layer
1889 * orders the requests before sending them to the driver which
1890 * doubled the write performance on spinning disks when measured
1892 btrfsic_check_bio(sbio->bio);
1893 submit_bio(sbio->bio);
1895 if (btrfs_is_zoned(sctx->fs_info))
1896 sctx->write_pointer = sbio->physical + sbio->sector_count *
1897 sctx->fs_info->sectorsize;
1900 static void scrub_wr_bio_end_io(struct bio *bio)
1902 struct scrub_bio *sbio = bio->bi_private;
1903 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1905 sbio->status = bio->bi_status;
1908 INIT_WORK(&sbio->work, scrub_wr_bio_end_io_worker);
1909 queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1912 static void scrub_wr_bio_end_io_worker(struct work_struct *work)
1914 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1915 struct scrub_ctx *sctx = sbio->sctx;
1918 ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
1920 struct btrfs_dev_replace *dev_replace =
1921 &sbio->sctx->fs_info->dev_replace;
1923 for (i = 0; i < sbio->sector_count; i++) {
1924 struct scrub_sector *sector = sbio->sectors[i];
1926 sector->io_error = 1;
1927 atomic64_inc(&dev_replace->num_write_errors);
1932 * In scrub_add_sector_to_wr_bio() we grab extra ref for sblock, now in
1933 * endio we should put the sblock.
1935 for (i = 0; i < sbio->sector_count; i++) {
1936 scrub_block_put(sbio->sectors[i]->sblock);
1937 scrub_sector_put(sbio->sectors[i]);
1942 scrub_pending_bio_dec(sctx);
1945 static int scrub_checksum(struct scrub_block *sblock)
1951 * No need to initialize these stats currently,
1952 * because this function only use return value
1953 * instead of these stats value.
1958 sblock->header_error = 0;
1959 sblock->generation_error = 0;
1960 sblock->checksum_error = 0;
1962 WARN_ON(sblock->sector_count < 1);
1963 flags = sblock->sectors[0]->flags;
1965 if (flags & BTRFS_EXTENT_FLAG_DATA)
1966 ret = scrub_checksum_data(sblock);
1967 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1968 ret = scrub_checksum_tree_block(sblock);
1969 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1970 ret = scrub_checksum_super(sblock);
1974 scrub_handle_errored_block(sblock);
1979 static int scrub_checksum_data(struct scrub_block *sblock)
1981 struct scrub_ctx *sctx = sblock->sctx;
1982 struct btrfs_fs_info *fs_info = sctx->fs_info;
1983 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1984 u8 csum[BTRFS_CSUM_SIZE];
1985 struct scrub_sector *sector;
1988 BUG_ON(sblock->sector_count < 1);
1989 sector = sblock->sectors[0];
1990 if (!sector->have_csum)
1993 kaddr = scrub_sector_get_kaddr(sector);
1995 shash->tfm = fs_info->csum_shash;
1996 crypto_shash_init(shash);
1998 crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
2000 if (memcmp(csum, sector->csum, fs_info->csum_size))
2001 sblock->checksum_error = 1;
2002 return sblock->checksum_error;
2005 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2007 struct scrub_ctx *sctx = sblock->sctx;
2008 struct btrfs_header *h;
2009 struct btrfs_fs_info *fs_info = sctx->fs_info;
2010 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2011 u8 calculated_csum[BTRFS_CSUM_SIZE];
2012 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2014 * This is done in sectorsize steps even for metadata as there's a
2015 * constraint for nodesize to be aligned to sectorsize. This will need
2016 * to change so we don't misuse data and metadata units like that.
2018 const u32 sectorsize = sctx->fs_info->sectorsize;
2019 const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
2021 struct scrub_sector *sector;
2024 BUG_ON(sblock->sector_count < 1);
2026 /* Each member in sectors is just one sector */
2027 ASSERT(sblock->sector_count == num_sectors);
2029 sector = sblock->sectors[0];
2030 kaddr = scrub_sector_get_kaddr(sector);
2031 h = (struct btrfs_header *)kaddr;
2032 memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
2035 * we don't use the getter functions here, as we
2036 * a) don't have an extent buffer and
2037 * b) the page is already kmapped
2039 if (sblock->logical != btrfs_stack_header_bytenr(h)) {
2040 sblock->header_error = 1;
2041 btrfs_warn_rl(fs_info,
2042 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
2043 sblock->logical, sblock->mirror_num,
2044 btrfs_stack_header_bytenr(h),
2049 if (!scrub_check_fsid(h->fsid, sector)) {
2050 sblock->header_error = 1;
2051 btrfs_warn_rl(fs_info,
2052 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
2053 sblock->logical, sblock->mirror_num,
2054 h->fsid, sblock->dev->fs_devices->fsid);
2058 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid, BTRFS_UUID_SIZE)) {
2059 sblock->header_error = 1;
2060 btrfs_warn_rl(fs_info,
2061 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
2062 sblock->logical, sblock->mirror_num,
2063 h->chunk_tree_uuid, fs_info->chunk_tree_uuid);
2067 shash->tfm = fs_info->csum_shash;
2068 crypto_shash_init(shash);
2069 crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
2070 sectorsize - BTRFS_CSUM_SIZE);
2072 for (i = 1; i < num_sectors; i++) {
2073 kaddr = scrub_sector_get_kaddr(sblock->sectors[i]);
2074 crypto_shash_update(shash, kaddr, sectorsize);
2077 crypto_shash_final(shash, calculated_csum);
2078 if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size)) {
2079 sblock->checksum_error = 1;
2080 btrfs_warn_rl(fs_info,
2081 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
2082 sblock->logical, sblock->mirror_num,
2083 CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
2084 CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
2088 if (sector->generation != btrfs_stack_header_generation(h)) {
2089 sblock->header_error = 1;
2090 sblock->generation_error = 1;
2091 btrfs_warn_rl(fs_info,
2092 "tree block %llu mirror %u has bad generation, has %llu want %llu",
2093 sblock->logical, sblock->mirror_num,
2094 btrfs_stack_header_generation(h),
2095 sector->generation);
2099 return sblock->header_error || sblock->checksum_error;
2102 static int scrub_checksum_super(struct scrub_block *sblock)
2104 struct btrfs_super_block *s;
2105 struct scrub_ctx *sctx = sblock->sctx;
2106 struct btrfs_fs_info *fs_info = sctx->fs_info;
2107 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
2108 u8 calculated_csum[BTRFS_CSUM_SIZE];
2109 struct scrub_sector *sector;
2114 BUG_ON(sblock->sector_count < 1);
2115 sector = sblock->sectors[0];
2116 kaddr = scrub_sector_get_kaddr(sector);
2117 s = (struct btrfs_super_block *)kaddr;
2119 if (sblock->logical != btrfs_super_bytenr(s))
2122 if (sector->generation != btrfs_super_generation(s))
2125 if (!scrub_check_fsid(s->fsid, sector))
2128 shash->tfm = fs_info->csum_shash;
2129 crypto_shash_init(shash);
2130 crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
2131 BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
2133 if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
2136 return fail_cor + fail_gen;
2139 static void scrub_block_put(struct scrub_block *sblock)
2141 if (refcount_dec_and_test(&sblock->refs)) {
2144 if (sblock->sparity)
2145 scrub_parity_put(sblock->sparity);
2147 for (i = 0; i < sblock->sector_count; i++)
2148 scrub_sector_put(sblock->sectors[i]);
2149 for (i = 0; i < DIV_ROUND_UP(sblock->len, PAGE_SIZE); i++) {
2150 if (sblock->pages[i]) {
2151 detach_scrub_page_private(sblock->pages[i]);
2152 __free_page(sblock->pages[i]);
2159 static void scrub_sector_get(struct scrub_sector *sector)
2161 atomic_inc(§or->refs);
2164 static void scrub_sector_put(struct scrub_sector *sector)
2166 if (atomic_dec_and_test(§or->refs))
2171 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
2172 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
2174 static void scrub_throttle(struct scrub_ctx *sctx)
2176 const int time_slice = 1000;
2177 struct scrub_bio *sbio;
2178 struct btrfs_device *device;
2184 sbio = sctx->bios[sctx->curr];
2186 bwlimit = READ_ONCE(device->scrub_speed_max);
2191 * Slice is divided into intervals when the IO is submitted, adjust by
2192 * bwlimit and maximum of 64 intervals.
2194 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
2195 div = min_t(u32, 64, div);
2197 /* Start new epoch, set deadline */
2199 if (sctx->throttle_deadline == 0) {
2200 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
2201 sctx->throttle_sent = 0;
2204 /* Still in the time to send? */
2205 if (ktime_before(now, sctx->throttle_deadline)) {
2206 /* If current bio is within the limit, send it */
2207 sctx->throttle_sent += sbio->bio->bi_iter.bi_size;
2208 if (sctx->throttle_sent <= div_u64(bwlimit, div))
2211 /* We're over the limit, sleep until the rest of the slice */
2212 delta = ktime_ms_delta(sctx->throttle_deadline, now);
2214 /* New request after deadline, start new epoch */
2221 timeout = div_u64(delta * HZ, 1000);
2222 schedule_timeout_interruptible(timeout);
2225 /* Next call will start the deadline period */
2226 sctx->throttle_deadline = 0;
2229 static void scrub_submit(struct scrub_ctx *sctx)
2231 struct scrub_bio *sbio;
2233 if (sctx->curr == -1)
2236 scrub_throttle(sctx);
2238 sbio = sctx->bios[sctx->curr];
2240 scrub_pending_bio_inc(sctx);
2241 btrfsic_check_bio(sbio->bio);
2242 submit_bio(sbio->bio);
2245 static int scrub_add_sector_to_rd_bio(struct scrub_ctx *sctx,
2246 struct scrub_sector *sector)
2248 struct scrub_block *sblock = sector->sblock;
2249 struct scrub_bio *sbio;
2250 const u32 sectorsize = sctx->fs_info->sectorsize;
2255 * grab a fresh bio or wait for one to become available
2257 while (sctx->curr == -1) {
2258 spin_lock(&sctx->list_lock);
2259 sctx->curr = sctx->first_free;
2260 if (sctx->curr != -1) {
2261 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2262 sctx->bios[sctx->curr]->next_free = -1;
2263 sctx->bios[sctx->curr]->sector_count = 0;
2264 spin_unlock(&sctx->list_lock);
2266 spin_unlock(&sctx->list_lock);
2267 wait_event(sctx->list_wait, sctx->first_free != -1);
2270 sbio = sctx->bios[sctx->curr];
2271 if (sbio->sector_count == 0) {
2272 sbio->physical = sblock->physical + sector->offset;
2273 sbio->logical = sblock->logical + sector->offset;
2274 sbio->dev = sblock->dev;
2276 sbio->bio = bio_alloc(sbio->dev->bdev, sctx->sectors_per_bio,
2277 REQ_OP_READ, GFP_NOFS);
2279 sbio->bio->bi_private = sbio;
2280 sbio->bio->bi_end_io = scrub_bio_end_io;
2281 sbio->bio->bi_iter.bi_sector = sbio->physical >> 9;
2283 } else if (sbio->physical + sbio->sector_count * sectorsize !=
2284 sblock->physical + sector->offset ||
2285 sbio->logical + sbio->sector_count * sectorsize !=
2286 sblock->logical + sector->offset ||
2287 sbio->dev != sblock->dev) {
2292 sbio->sectors[sbio->sector_count] = sector;
2293 ret = bio_add_scrub_sector(sbio->bio, sector, sectorsize);
2294 if (ret != sectorsize) {
2295 if (sbio->sector_count < 1) {
2304 scrub_block_get(sblock); /* one for the page added to the bio */
2305 atomic_inc(&sblock->outstanding_sectors);
2306 sbio->sector_count++;
2307 if (sbio->sector_count == sctx->sectors_per_bio)
2313 static void scrub_missing_raid56_end_io(struct bio *bio)
2315 struct scrub_block *sblock = bio->bi_private;
2316 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2318 btrfs_bio_counter_dec(fs_info);
2320 sblock->no_io_error_seen = 0;
2324 queue_work(fs_info->scrub_workers, &sblock->work);
2327 static void scrub_missing_raid56_worker(struct work_struct *work)
2329 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2330 struct scrub_ctx *sctx = sblock->sctx;
2331 struct btrfs_fs_info *fs_info = sctx->fs_info;
2333 struct btrfs_device *dev;
2335 logical = sblock->logical;
2338 if (sblock->no_io_error_seen)
2339 scrub_recheck_block_checksum(sblock);
2341 if (!sblock->no_io_error_seen) {
2342 spin_lock(&sctx->stat_lock);
2343 sctx->stat.read_errors++;
2344 spin_unlock(&sctx->stat_lock);
2345 btrfs_err_rl_in_rcu(fs_info,
2346 "IO error rebuilding logical %llu for dev %s",
2347 logical, rcu_str_deref(dev->name));
2348 } else if (sblock->header_error || sblock->checksum_error) {
2349 spin_lock(&sctx->stat_lock);
2350 sctx->stat.uncorrectable_errors++;
2351 spin_unlock(&sctx->stat_lock);
2352 btrfs_err_rl_in_rcu(fs_info,
2353 "failed to rebuild valid logical %llu for dev %s",
2354 logical, rcu_str_deref(dev->name));
2356 scrub_write_block_to_dev_replace(sblock);
2359 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2360 mutex_lock(&sctx->wr_lock);
2361 scrub_wr_submit(sctx);
2362 mutex_unlock(&sctx->wr_lock);
2365 scrub_block_put(sblock);
2366 scrub_pending_bio_dec(sctx);
2369 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2371 struct scrub_ctx *sctx = sblock->sctx;
2372 struct btrfs_fs_info *fs_info = sctx->fs_info;
2373 u64 length = sblock->sector_count << fs_info->sectorsize_bits;
2374 u64 logical = sblock->logical;
2375 struct btrfs_io_context *bioc = NULL;
2377 struct btrfs_raid_bio *rbio;
2381 btrfs_bio_counter_inc_blocked(fs_info);
2382 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2384 if (ret || !bioc || !bioc->raid_map)
2387 if (WARN_ON(!sctx->is_dev_replace ||
2388 !(bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2390 * We shouldn't be scrubbing a missing device. Even for dev
2391 * replace, we should only get here for RAID 5/6. We either
2392 * managed to mount something with no mirrors remaining or
2393 * there's a bug in scrub_find_good_copy()/btrfs_map_block().
2398 bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2399 bio->bi_iter.bi_sector = logical >> 9;
2400 bio->bi_private = sblock;
2401 bio->bi_end_io = scrub_missing_raid56_end_io;
2403 rbio = raid56_alloc_missing_rbio(bio, bioc);
2407 for (i = 0; i < sblock->sector_count; i++) {
2408 struct scrub_sector *sector = sblock->sectors[i];
2410 raid56_add_scrub_pages(rbio, scrub_sector_get_page(sector),
2411 scrub_sector_get_page_offset(sector),
2412 sector->offset + sector->sblock->logical);
2415 INIT_WORK(&sblock->work, scrub_missing_raid56_worker);
2416 scrub_block_get(sblock);
2417 scrub_pending_bio_inc(sctx);
2418 raid56_submit_missing_rbio(rbio);
2419 btrfs_put_bioc(bioc);
2425 btrfs_bio_counter_dec(fs_info);
2426 btrfs_put_bioc(bioc);
2427 spin_lock(&sctx->stat_lock);
2428 sctx->stat.malloc_errors++;
2429 spin_unlock(&sctx->stat_lock);
2432 static int scrub_sectors(struct scrub_ctx *sctx, u64 logical, u32 len,
2433 u64 physical, struct btrfs_device *dev, u64 flags,
2434 u64 gen, int mirror_num, u8 *csum,
2435 u64 physical_for_dev_replace)
2437 struct scrub_block *sblock;
2438 const u32 sectorsize = sctx->fs_info->sectorsize;
2441 sblock = alloc_scrub_block(sctx, dev, logical, physical,
2442 physical_for_dev_replace, mirror_num);
2444 spin_lock(&sctx->stat_lock);
2445 sctx->stat.malloc_errors++;
2446 spin_unlock(&sctx->stat_lock);
2450 for (index = 0; len > 0; index++) {
2451 struct scrub_sector *sector;
2453 * Here we will allocate one page for one sector to scrub.
2454 * This is fine if PAGE_SIZE == sectorsize, but will cost
2455 * more memory for PAGE_SIZE > sectorsize case.
2457 u32 l = min(sectorsize, len);
2459 sector = alloc_scrub_sector(sblock, logical, GFP_KERNEL);
2461 spin_lock(&sctx->stat_lock);
2462 sctx->stat.malloc_errors++;
2463 spin_unlock(&sctx->stat_lock);
2464 scrub_block_put(sblock);
2467 sector->flags = flags;
2468 sector->generation = gen;
2470 sector->have_csum = 1;
2471 memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2473 sector->have_csum = 0;
2478 physical_for_dev_replace += l;
2481 WARN_ON(sblock->sector_count == 0);
2482 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2484 * This case should only be hit for RAID 5/6 device replace. See
2485 * the comment in scrub_missing_raid56_pages() for details.
2487 scrub_missing_raid56_pages(sblock);
2489 for (index = 0; index < sblock->sector_count; index++) {
2490 struct scrub_sector *sector = sblock->sectors[index];
2493 ret = scrub_add_sector_to_rd_bio(sctx, sector);
2495 scrub_block_put(sblock);
2500 if (flags & BTRFS_EXTENT_FLAG_SUPER)
2504 /* last one frees, either here or in bio completion for last page */
2505 scrub_block_put(sblock);
2509 static void scrub_bio_end_io(struct bio *bio)
2511 struct scrub_bio *sbio = bio->bi_private;
2512 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2514 sbio->status = bio->bi_status;
2517 queue_work(fs_info->scrub_workers, &sbio->work);
2520 static void scrub_bio_end_io_worker(struct work_struct *work)
2522 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2523 struct scrub_ctx *sctx = sbio->sctx;
2526 ASSERT(sbio->sector_count <= SCRUB_SECTORS_PER_BIO);
2528 for (i = 0; i < sbio->sector_count; i++) {
2529 struct scrub_sector *sector = sbio->sectors[i];
2531 sector->io_error = 1;
2532 sector->sblock->no_io_error_seen = 0;
2536 /* Now complete the scrub_block items that have all pages completed */
2537 for (i = 0; i < sbio->sector_count; i++) {
2538 struct scrub_sector *sector = sbio->sectors[i];
2539 struct scrub_block *sblock = sector->sblock;
2541 if (atomic_dec_and_test(&sblock->outstanding_sectors))
2542 scrub_block_complete(sblock);
2543 scrub_block_put(sblock);
2548 spin_lock(&sctx->list_lock);
2549 sbio->next_free = sctx->first_free;
2550 sctx->first_free = sbio->index;
2551 spin_unlock(&sctx->list_lock);
2553 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2554 mutex_lock(&sctx->wr_lock);
2555 scrub_wr_submit(sctx);
2556 mutex_unlock(&sctx->wr_lock);
2559 scrub_pending_bio_dec(sctx);
2562 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2563 unsigned long *bitmap,
2568 u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2570 if (len >= sparity->stripe_len) {
2571 bitmap_set(bitmap, 0, sparity->nsectors);
2575 start -= sparity->logic_start;
2576 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2577 offset = offset >> sectorsize_bits;
2578 nsectors = len >> sectorsize_bits;
2580 if (offset + nsectors <= sparity->nsectors) {
2581 bitmap_set(bitmap, offset, nsectors);
2585 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2586 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2589 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2592 __scrub_mark_bitmap(sparity, &sparity->ebitmap, start, len);
2595 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2598 __scrub_mark_bitmap(sparity, &sparity->dbitmap, start, len);
2601 static void scrub_block_complete(struct scrub_block *sblock)
2605 if (!sblock->no_io_error_seen) {
2607 scrub_handle_errored_block(sblock);
2610 * if has checksum error, write via repair mechanism in
2611 * dev replace case, otherwise write here in dev replace
2614 corrupted = scrub_checksum(sblock);
2615 if (!corrupted && sblock->sctx->is_dev_replace)
2616 scrub_write_block_to_dev_replace(sblock);
2619 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2620 u64 start = sblock->logical;
2621 u64 end = sblock->logical +
2622 sblock->sectors[sblock->sector_count - 1]->offset +
2623 sblock->sctx->fs_info->sectorsize;
2625 ASSERT(end - start <= U32_MAX);
2626 scrub_parity_mark_sectors_error(sblock->sparity,
2627 start, end - start);
2631 static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2633 sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2634 list_del(&sum->list);
2639 * Find the desired csum for range [logical, logical + sectorsize), and store
2640 * the csum into @csum.
2642 * The search source is sctx->csum_list, which is a pre-populated list
2643 * storing bytenr ordered csum ranges. We're responsible to cleanup any range
2644 * that is before @logical.
2646 * Return 0 if there is no csum for the range.
2647 * Return 1 if there is csum for the range and copied to @csum.
2649 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2653 while (!list_empty(&sctx->csum_list)) {
2654 struct btrfs_ordered_sum *sum = NULL;
2655 unsigned long index;
2656 unsigned long num_sectors;
2658 sum = list_first_entry(&sctx->csum_list,
2659 struct btrfs_ordered_sum, list);
2660 /* The current csum range is beyond our range, no csum found */
2661 if (sum->bytenr > logical)
2665 * The current sum is before our bytenr, since scrub is always
2666 * done in bytenr order, the csum will never be used anymore,
2667 * clean it up so that later calls won't bother with the range,
2668 * and continue search the next range.
2670 if (sum->bytenr + sum->len <= logical) {
2671 drop_csum_range(sctx, sum);
2675 /* Now the csum range covers our bytenr, copy the csum */
2677 index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2678 num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2680 memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2681 sctx->fs_info->csum_size);
2683 /* Cleanup the range if we're at the end of the csum range */
2684 if (index == num_sectors - 1)
2685 drop_csum_range(sctx, sum);
2693 /* scrub extent tries to collect up to 64 kB for each bio */
2694 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2695 u64 logical, u32 len,
2696 u64 physical, struct btrfs_device *dev, u64 flags,
2697 u64 gen, int mirror_num)
2699 struct btrfs_device *src_dev = dev;
2700 u64 src_physical = physical;
2701 int src_mirror = mirror_num;
2703 u8 csum[BTRFS_CSUM_SIZE];
2706 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2707 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2708 blocksize = map->stripe_len;
2710 blocksize = sctx->fs_info->sectorsize;
2711 spin_lock(&sctx->stat_lock);
2712 sctx->stat.data_extents_scrubbed++;
2713 sctx->stat.data_bytes_scrubbed += len;
2714 spin_unlock(&sctx->stat_lock);
2715 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2716 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2717 blocksize = map->stripe_len;
2719 blocksize = sctx->fs_info->nodesize;
2720 spin_lock(&sctx->stat_lock);
2721 sctx->stat.tree_extents_scrubbed++;
2722 sctx->stat.tree_bytes_scrubbed += len;
2723 spin_unlock(&sctx->stat_lock);
2725 blocksize = sctx->fs_info->sectorsize;
2730 * For dev-replace case, we can have @dev being a missing device.
2731 * Regular scrub will avoid its execution on missing device at all,
2732 * as that would trigger tons of read error.
2734 * Reading from missing device will cause read error counts to
2735 * increase unnecessarily.
2736 * So here we change the read source to a good mirror.
2738 if (sctx->is_dev_replace && !dev->bdev)
2739 scrub_find_good_copy(sctx->fs_info, logical, len, &src_physical,
2740 &src_dev, &src_mirror);
2742 u32 l = min(len, blocksize);
2745 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2746 /* push csums to sbio */
2747 have_csum = scrub_find_csum(sctx, logical, csum);
2749 ++sctx->stat.no_csum;
2751 ret = scrub_sectors(sctx, logical, l, src_physical, src_dev,
2752 flags, gen, src_mirror,
2753 have_csum ? csum : NULL, physical);
2764 static int scrub_sectors_for_parity(struct scrub_parity *sparity,
2765 u64 logical, u32 len,
2766 u64 physical, struct btrfs_device *dev,
2767 u64 flags, u64 gen, int mirror_num, u8 *csum)
2769 struct scrub_ctx *sctx = sparity->sctx;
2770 struct scrub_block *sblock;
2771 const u32 sectorsize = sctx->fs_info->sectorsize;
2774 ASSERT(IS_ALIGNED(len, sectorsize));
2776 sblock = alloc_scrub_block(sctx, dev, logical, physical, physical, mirror_num);
2778 spin_lock(&sctx->stat_lock);
2779 sctx->stat.malloc_errors++;
2780 spin_unlock(&sctx->stat_lock);
2784 sblock->sparity = sparity;
2785 scrub_parity_get(sparity);
2787 for (index = 0; len > 0; index++) {
2788 struct scrub_sector *sector;
2790 sector = alloc_scrub_sector(sblock, logical, GFP_KERNEL);
2792 spin_lock(&sctx->stat_lock);
2793 sctx->stat.malloc_errors++;
2794 spin_unlock(&sctx->stat_lock);
2795 scrub_block_put(sblock);
2798 sblock->sectors[index] = sector;
2799 /* For scrub parity */
2800 scrub_sector_get(sector);
2801 list_add_tail(§or->list, &sparity->sectors_list);
2802 sector->flags = flags;
2803 sector->generation = gen;
2805 sector->have_csum = 1;
2806 memcpy(sector->csum, csum, sctx->fs_info->csum_size);
2808 sector->have_csum = 0;
2811 /* Iterate over the stripe range in sectorsize steps */
2813 logical += sectorsize;
2814 physical += sectorsize;
2817 WARN_ON(sblock->sector_count == 0);
2818 for (index = 0; index < sblock->sector_count; index++) {
2819 struct scrub_sector *sector = sblock->sectors[index];
2822 ret = scrub_add_sector_to_rd_bio(sctx, sector);
2824 scrub_block_put(sblock);
2829 /* Last one frees, either here or in bio completion for last sector */
2830 scrub_block_put(sblock);
2834 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2835 u64 logical, u32 len,
2836 u64 physical, struct btrfs_device *dev,
2837 u64 flags, u64 gen, int mirror_num)
2839 struct scrub_ctx *sctx = sparity->sctx;
2841 u8 csum[BTRFS_CSUM_SIZE];
2844 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2845 scrub_parity_mark_sectors_error(sparity, logical, len);
2849 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2850 blocksize = sparity->stripe_len;
2851 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2852 blocksize = sparity->stripe_len;
2854 blocksize = sctx->fs_info->sectorsize;
2859 u32 l = min(len, blocksize);
2862 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2863 /* push csums to sbio */
2864 have_csum = scrub_find_csum(sctx, logical, csum);
2868 ret = scrub_sectors_for_parity(sparity, logical, l, physical, dev,
2869 flags, gen, mirror_num,
2870 have_csum ? csum : NULL);
2882 * Given a physical address, this will calculate it's
2883 * logical offset. if this is a parity stripe, it will return
2884 * the most left data stripe's logical offset.
2886 * return 0 if it is a data stripe, 1 means parity stripe.
2888 static int get_raid56_logic_offset(u64 physical, int num,
2889 struct map_lookup *map, u64 *offset,
2898 const int data_stripes = nr_data_stripes(map);
2900 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2902 *stripe_start = last_offset;
2904 *offset = last_offset;
2905 for (i = 0; i < data_stripes; i++) {
2906 *offset = last_offset + i * map->stripe_len;
2908 stripe_nr = div64_u64(*offset, map->stripe_len);
2909 stripe_nr = div_u64(stripe_nr, data_stripes);
2911 /* Work out the disk rotation on this stripe-set */
2912 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2913 /* calculate which stripe this data locates */
2915 stripe_index = rot % map->num_stripes;
2916 if (stripe_index == num)
2918 if (stripe_index < num)
2921 *offset = last_offset + j * map->stripe_len;
2925 static void scrub_free_parity(struct scrub_parity *sparity)
2927 struct scrub_ctx *sctx = sparity->sctx;
2928 struct scrub_sector *curr, *next;
2931 nbits = bitmap_weight(&sparity->ebitmap, sparity->nsectors);
2933 spin_lock(&sctx->stat_lock);
2934 sctx->stat.read_errors += nbits;
2935 sctx->stat.uncorrectable_errors += nbits;
2936 spin_unlock(&sctx->stat_lock);
2939 list_for_each_entry_safe(curr, next, &sparity->sectors_list, list) {
2940 list_del_init(&curr->list);
2941 scrub_sector_put(curr);
2947 static void scrub_parity_bio_endio_worker(struct work_struct *work)
2949 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2951 struct scrub_ctx *sctx = sparity->sctx;
2953 btrfs_bio_counter_dec(sctx->fs_info);
2954 scrub_free_parity(sparity);
2955 scrub_pending_bio_dec(sctx);
2958 static void scrub_parity_bio_endio(struct bio *bio)
2960 struct scrub_parity *sparity = bio->bi_private;
2961 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2964 bitmap_or(&sparity->ebitmap, &sparity->ebitmap,
2965 &sparity->dbitmap, sparity->nsectors);
2969 INIT_WORK(&sparity->work, scrub_parity_bio_endio_worker);
2970 queue_work(fs_info->scrub_parity_workers, &sparity->work);
2973 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2975 struct scrub_ctx *sctx = sparity->sctx;
2976 struct btrfs_fs_info *fs_info = sctx->fs_info;
2978 struct btrfs_raid_bio *rbio;
2979 struct btrfs_io_context *bioc = NULL;
2983 if (!bitmap_andnot(&sparity->dbitmap, &sparity->dbitmap,
2984 &sparity->ebitmap, sparity->nsectors))
2987 length = sparity->logic_end - sparity->logic_start;
2989 btrfs_bio_counter_inc_blocked(fs_info);
2990 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2992 if (ret || !bioc || !bioc->raid_map)
2995 bio = bio_alloc(NULL, BIO_MAX_VECS, REQ_OP_READ, GFP_NOFS);
2996 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2997 bio->bi_private = sparity;
2998 bio->bi_end_io = scrub_parity_bio_endio;
3000 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc,
3004 btrfs_put_bioc(bioc);
3008 scrub_pending_bio_inc(sctx);
3009 raid56_parity_submit_scrub_rbio(rbio);
3015 btrfs_bio_counter_dec(fs_info);
3016 bitmap_or(&sparity->ebitmap, &sparity->ebitmap, &sparity->dbitmap,
3018 spin_lock(&sctx->stat_lock);
3019 sctx->stat.malloc_errors++;
3020 spin_unlock(&sctx->stat_lock);
3022 scrub_free_parity(sparity);
3025 static void scrub_parity_get(struct scrub_parity *sparity)
3027 refcount_inc(&sparity->refs);
3030 static void scrub_parity_put(struct scrub_parity *sparity)
3032 if (!refcount_dec_and_test(&sparity->refs))
3035 scrub_parity_check_and_repair(sparity);
3039 * Return 0 if the extent item range covers any byte of the range.
3040 * Return <0 if the extent item is before @search_start.
3041 * Return >0 if the extent item is after @start_start + @search_len.
3043 static int compare_extent_item_range(struct btrfs_path *path,
3044 u64 search_start, u64 search_len)
3046 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
3048 struct btrfs_key key;
3050 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3051 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
3052 key.type == BTRFS_METADATA_ITEM_KEY);
3053 if (key.type == BTRFS_METADATA_ITEM_KEY)
3054 len = fs_info->nodesize;
3058 if (key.objectid + len <= search_start)
3060 if (key.objectid >= search_start + search_len)
3066 * Locate one extent item which covers any byte in range
3067 * [@search_start, @search_start + @search_length)
3069 * If the path is not initialized, we will initialize the search by doing
3070 * a btrfs_search_slot().
3071 * If the path is already initialized, we will use the path as the initial
3072 * slot, to avoid duplicated btrfs_search_slot() calls.
3074 * NOTE: If an extent item starts before @search_start, we will still
3075 * return the extent item. This is for data extent crossing stripe boundary.
3077 * Return 0 if we found such extent item, and @path will point to the extent item.
3078 * Return >0 if no such extent item can be found, and @path will be released.
3079 * Return <0 if hit fatal error, and @path will be released.
3081 static int find_first_extent_item(struct btrfs_root *extent_root,
3082 struct btrfs_path *path,
3083 u64 search_start, u64 search_len)
3085 struct btrfs_fs_info *fs_info = extent_root->fs_info;
3086 struct btrfs_key key;
3089 /* Continue using the existing path */
3091 goto search_forward;
3093 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3094 key.type = BTRFS_METADATA_ITEM_KEY;
3096 key.type = BTRFS_EXTENT_ITEM_KEY;
3097 key.objectid = search_start;
3098 key.offset = (u64)-1;
3100 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3106 * Here we intentionally pass 0 as @min_objectid, as there could be
3107 * an extent item starting before @search_start.
3109 ret = btrfs_previous_extent_item(extent_root, path, 0);
3113 * No matter whether we have found an extent item, the next loop will
3114 * properly do every check on the key.
3118 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3119 if (key.objectid >= search_start + search_len)
3121 if (key.type != BTRFS_METADATA_ITEM_KEY &&
3122 key.type != BTRFS_EXTENT_ITEM_KEY)
3125 ret = compare_extent_item_range(path, search_start, search_len);
3132 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
3133 ret = btrfs_next_leaf(extent_root, path);
3135 /* Either no more item or fatal error */
3136 btrfs_release_path(path);
3141 btrfs_release_path(path);
3145 static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
3146 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
3148 struct btrfs_key key;
3149 struct btrfs_extent_item *ei;
3151 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
3152 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
3153 key.type == BTRFS_EXTENT_ITEM_KEY);
3154 *extent_start_ret = key.objectid;
3155 if (key.type == BTRFS_METADATA_ITEM_KEY)
3156 *size_ret = path->nodes[0]->fs_info->nodesize;
3158 *size_ret = key.offset;
3159 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
3160 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
3161 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
3164 static bool does_range_cross_boundary(u64 extent_start, u64 extent_len,
3165 u64 boundary_start, u64 boudary_len)
3167 return (extent_start < boundary_start &&
3168 extent_start + extent_len > boundary_start) ||
3169 (extent_start < boundary_start + boudary_len &&
3170 extent_start + extent_len > boundary_start + boudary_len);
3173 static int scrub_raid56_data_stripe_for_parity(struct scrub_ctx *sctx,
3174 struct scrub_parity *sparity,
3175 struct map_lookup *map,
3176 struct btrfs_device *sdev,
3177 struct btrfs_path *path,
3180 struct btrfs_fs_info *fs_info = sctx->fs_info;
3181 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, logical);
3182 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, logical);
3183 u64 cur_logical = logical;
3186 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3188 /* Path must not be populated */
3189 ASSERT(!path->nodes[0]);
3191 while (cur_logical < logical + map->stripe_len) {
3192 struct btrfs_io_context *bioc = NULL;
3193 struct btrfs_device *extent_dev;
3199 u64 extent_physical;
3200 u64 extent_mirror_num;
3202 ret = find_first_extent_item(extent_root, path, cur_logical,
3203 logical + map->stripe_len - cur_logical);
3204 /* No more extent item in this data stripe */
3211 get_extent_info(path, &extent_start, &extent_size, &extent_flags,
3214 /* Metadata should not cross stripe boundaries */
3215 if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3216 does_range_cross_boundary(extent_start, extent_size,
3217 logical, map->stripe_len)) {
3219 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3220 extent_start, logical);
3221 spin_lock(&sctx->stat_lock);
3222 sctx->stat.uncorrectable_errors++;
3223 spin_unlock(&sctx->stat_lock);
3224 cur_logical += extent_size;
3228 /* Skip hole range which doesn't have any extent */
3229 cur_logical = max(extent_start, cur_logical);
3231 /* Truncate the range inside this data stripe */
3232 extent_size = min(extent_start + extent_size,
3233 logical + map->stripe_len) - cur_logical;
3234 extent_start = cur_logical;
3235 ASSERT(extent_size <= U32_MAX);
3237 scrub_parity_mark_sectors_data(sparity, extent_start, extent_size);
3239 mapped_length = extent_size;
3240 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_start,
3241 &mapped_length, &bioc, 0);
3242 if (!ret && (!bioc || mapped_length < extent_size))
3245 btrfs_put_bioc(bioc);
3246 scrub_parity_mark_sectors_error(sparity, extent_start,
3250 extent_physical = bioc->stripes[0].physical;
3251 extent_mirror_num = bioc->mirror_num;
3252 extent_dev = bioc->stripes[0].dev;
3253 btrfs_put_bioc(bioc);
3255 ret = btrfs_lookup_csums_range(csum_root, extent_start,
3256 extent_start + extent_size - 1,
3257 &sctx->csum_list, 1, false);
3259 scrub_parity_mark_sectors_error(sparity, extent_start,
3264 ret = scrub_extent_for_parity(sparity, extent_start,
3265 extent_size, extent_physical,
3266 extent_dev, extent_flags,
3267 extent_gen, extent_mirror_num);
3268 scrub_free_csums(sctx);
3271 scrub_parity_mark_sectors_error(sparity, extent_start,
3277 cur_logical += extent_size;
3279 btrfs_release_path(path);
3283 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3284 struct map_lookup *map,
3285 struct btrfs_device *sdev,
3289 struct btrfs_fs_info *fs_info = sctx->fs_info;
3290 struct btrfs_path *path;
3293 struct scrub_parity *sparity;
3296 path = btrfs_alloc_path();
3298 spin_lock(&sctx->stat_lock);
3299 sctx->stat.malloc_errors++;
3300 spin_unlock(&sctx->stat_lock);
3303 path->search_commit_root = 1;
3304 path->skip_locking = 1;
3306 ASSERT(map->stripe_len <= U32_MAX);
3307 nsectors = map->stripe_len >> fs_info->sectorsize_bits;
3308 ASSERT(nsectors <= BITS_PER_LONG);
3309 sparity = kzalloc(sizeof(struct scrub_parity), GFP_NOFS);
3311 spin_lock(&sctx->stat_lock);
3312 sctx->stat.malloc_errors++;
3313 spin_unlock(&sctx->stat_lock);
3314 btrfs_free_path(path);
3318 ASSERT(map->stripe_len <= U32_MAX);
3319 sparity->stripe_len = map->stripe_len;
3320 sparity->nsectors = nsectors;
3321 sparity->sctx = sctx;
3322 sparity->scrub_dev = sdev;
3323 sparity->logic_start = logic_start;
3324 sparity->logic_end = logic_end;
3325 refcount_set(&sparity->refs, 1);
3326 INIT_LIST_HEAD(&sparity->sectors_list);
3329 for (cur_logical = logic_start; cur_logical < logic_end;
3330 cur_logical += map->stripe_len) {
3331 ret = scrub_raid56_data_stripe_for_parity(sctx, sparity, map,
3332 sdev, path, cur_logical);
3337 scrub_parity_put(sparity);
3339 mutex_lock(&sctx->wr_lock);
3340 scrub_wr_submit(sctx);
3341 mutex_unlock(&sctx->wr_lock);
3343 btrfs_free_path(path);
3344 return ret < 0 ? ret : 0;
3347 static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3349 if (!btrfs_is_zoned(sctx->fs_info))
3352 sctx->flush_all_writes = true;
3354 mutex_lock(&sctx->wr_lock);
3355 scrub_wr_submit(sctx);
3356 mutex_unlock(&sctx->wr_lock);
3358 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3361 static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3362 u64 physical, u64 physical_end)
3364 struct btrfs_fs_info *fs_info = sctx->fs_info;
3367 if (!btrfs_is_zoned(fs_info))
3370 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3372 mutex_lock(&sctx->wr_lock);
3373 if (sctx->write_pointer < physical_end) {
3374 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3376 sctx->write_pointer);
3379 "zoned: failed to recover write pointer");
3381 mutex_unlock(&sctx->wr_lock);
3382 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3388 * Scrub one range which can only has simple mirror based profile.
3389 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
3392 * Since we may need to handle a subset of block group, we need @logical_start
3393 * and @logical_length parameter.
3395 static int scrub_simple_mirror(struct scrub_ctx *sctx,
3396 struct btrfs_root *extent_root,
3397 struct btrfs_root *csum_root,
3398 struct btrfs_block_group *bg,
3399 struct map_lookup *map,
3400 u64 logical_start, u64 logical_length,
3401 struct btrfs_device *device,
3402 u64 physical, int mirror_num)
3404 struct btrfs_fs_info *fs_info = sctx->fs_info;
3405 const u64 logical_end = logical_start + logical_length;
3406 /* An artificial limit, inherit from old scrub behavior */
3407 const u32 max_length = SZ_64K;
3408 struct btrfs_path path = { 0 };
3409 u64 cur_logical = logical_start;
3412 /* The range must be inside the bg */
3413 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
3415 path.search_commit_root = 1;
3416 path.skip_locking = 1;
3417 /* Go through each extent items inside the logical range */
3418 while (cur_logical < logical_end) {
3426 if (atomic_read(&fs_info->scrub_cancel_req) ||
3427 atomic_read(&sctx->cancel_req)) {
3432 if (atomic_read(&fs_info->scrub_pause_req)) {
3433 /* Push queued extents */
3434 sctx->flush_all_writes = true;
3436 mutex_lock(&sctx->wr_lock);
3437 scrub_wr_submit(sctx);
3438 mutex_unlock(&sctx->wr_lock);
3439 wait_event(sctx->list_wait,
3440 atomic_read(&sctx->bios_in_flight) == 0);
3441 sctx->flush_all_writes = false;
3442 scrub_blocked_if_needed(fs_info);
3444 /* Block group removed? */
3445 spin_lock(&bg->lock);
3446 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
3447 spin_unlock(&bg->lock);
3451 spin_unlock(&bg->lock);
3453 ret = find_first_extent_item(extent_root, &path, cur_logical,
3454 logical_end - cur_logical);
3456 /* No more extent, just update the accounting */
3457 sctx->stat.last_physical = physical + logical_length;
3463 get_extent_info(&path, &extent_start, &extent_len,
3464 &extent_flags, &extent_gen);
3465 /* Skip hole range which doesn't have any extent */
3466 cur_logical = max(extent_start, cur_logical);
3469 * Scrub len has three limits:
3470 * - Extent size limit
3471 * - Scrub range limit
3472 * This is especially imporatant for RAID0/RAID10 to reuse
3474 * - Max scrub size limit
3476 scrub_len = min(min(extent_start + extent_len,
3477 logical_end), cur_logical + max_length) -
3480 if (extent_flags & BTRFS_EXTENT_FLAG_DATA) {
3481 ret = btrfs_lookup_csums_range(csum_root, cur_logical,
3482 cur_logical + scrub_len - 1,
3483 &sctx->csum_list, 1, false);
3487 if ((extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3488 does_range_cross_boundary(extent_start, extent_len,
3489 logical_start, logical_length)) {
3491 "scrub: tree block %llu spanning boundaries, ignored. boundary=[%llu, %llu)",
3492 extent_start, logical_start, logical_end);
3493 spin_lock(&sctx->stat_lock);
3494 sctx->stat.uncorrectable_errors++;
3495 spin_unlock(&sctx->stat_lock);
3496 cur_logical += scrub_len;
3499 ret = scrub_extent(sctx, map, cur_logical, scrub_len,
3500 cur_logical - logical_start + physical,
3501 device, extent_flags, extent_gen,
3503 scrub_free_csums(sctx);
3506 if (sctx->is_dev_replace)
3507 sync_replace_for_zoned(sctx);
3508 cur_logical += scrub_len;
3509 /* Don't hold CPU for too long time */
3512 btrfs_release_path(&path);
3516 /* Calculate the full stripe length for simple stripe based profiles */
3517 static u64 simple_stripe_full_stripe_len(const struct map_lookup *map)
3519 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3520 BTRFS_BLOCK_GROUP_RAID10));
3522 return map->num_stripes / map->sub_stripes * map->stripe_len;
3525 /* Get the logical bytenr for the stripe */
3526 static u64 simple_stripe_get_logical(struct map_lookup *map,
3527 struct btrfs_block_group *bg,
3530 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3531 BTRFS_BLOCK_GROUP_RAID10));
3532 ASSERT(stripe_index < map->num_stripes);
3535 * (stripe_index / sub_stripes) gives how many data stripes we need to
3538 return (stripe_index / map->sub_stripes) * map->stripe_len + bg->start;
3541 /* Get the mirror number for the stripe */
3542 static int simple_stripe_mirror_num(struct map_lookup *map, int stripe_index)
3544 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
3545 BTRFS_BLOCK_GROUP_RAID10));
3546 ASSERT(stripe_index < map->num_stripes);
3548 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
3549 return stripe_index % map->sub_stripes + 1;
3552 static int scrub_simple_stripe(struct scrub_ctx *sctx,
3553 struct btrfs_root *extent_root,
3554 struct btrfs_root *csum_root,
3555 struct btrfs_block_group *bg,
3556 struct map_lookup *map,
3557 struct btrfs_device *device,
3560 const u64 logical_increment = simple_stripe_full_stripe_len(map);
3561 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
3562 const u64 orig_physical = map->stripes[stripe_index].physical;
3563 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
3564 u64 cur_logical = orig_logical;
3565 u64 cur_physical = orig_physical;
3568 while (cur_logical < bg->start + bg->length) {
3570 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
3571 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
3574 ret = scrub_simple_mirror(sctx, extent_root, csum_root, bg, map,
3575 cur_logical, map->stripe_len, device,
3576 cur_physical, mirror_num);
3579 /* Skip to next stripe which belongs to the target device */
3580 cur_logical += logical_increment;
3581 /* For physical offset, we just go to next stripe */
3582 cur_physical += map->stripe_len;
3587 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3588 struct btrfs_block_group *bg,
3589 struct extent_map *em,
3590 struct btrfs_device *scrub_dev,
3593 struct btrfs_path *path;
3594 struct btrfs_fs_info *fs_info = sctx->fs_info;
3595 struct btrfs_root *root;
3596 struct btrfs_root *csum_root;
3597 struct blk_plug plug;
3598 struct map_lookup *map = em->map_lookup;
3599 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3600 const u64 chunk_logical = bg->start;
3602 u64 physical = map->stripes[stripe_index].physical;
3603 const u64 dev_stripe_len = btrfs_calc_stripe_length(em);
3604 const u64 physical_end = physical + dev_stripe_len;
3607 /* The logical increment after finishing one stripe */
3609 /* Offset inside the chunk */
3615 path = btrfs_alloc_path();
3620 * work on commit root. The related disk blocks are static as
3621 * long as COW is applied. This means, it is save to rewrite
3622 * them to repair disk errors without any race conditions
3624 path->search_commit_root = 1;
3625 path->skip_locking = 1;
3626 path->reada = READA_FORWARD;
3628 wait_event(sctx->list_wait,
3629 atomic_read(&sctx->bios_in_flight) == 0);
3630 scrub_blocked_if_needed(fs_info);
3632 root = btrfs_extent_root(fs_info, bg->start);
3633 csum_root = btrfs_csum_root(fs_info, bg->start);
3636 * collect all data csums for the stripe to avoid seeking during
3637 * the scrub. This might currently (crc32) end up to be about 1MB
3639 blk_start_plug(&plug);
3641 if (sctx->is_dev_replace &&
3642 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3643 mutex_lock(&sctx->wr_lock);
3644 sctx->write_pointer = physical;
3645 mutex_unlock(&sctx->wr_lock);
3646 sctx->flush_all_writes = true;
3650 * There used to be a big double loop to handle all profiles using the
3651 * same routine, which grows larger and more gross over time.
3653 * So here we handle each profile differently, so simpler profiles
3654 * have simpler scrubbing function.
3656 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
3657 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
3659 * Above check rules out all complex profile, the remaining
3660 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
3661 * mirrored duplication without stripe.
3663 * Only @physical and @mirror_num needs to calculated using
3666 ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3667 bg->start, bg->length, scrub_dev,
3668 map->stripes[stripe_index].physical,
3673 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
3674 ret = scrub_simple_stripe(sctx, root, csum_root, bg, map,
3675 scrub_dev, stripe_index);
3676 offset = map->stripe_len * (stripe_index / map->sub_stripes);
3680 /* Only RAID56 goes through the old code */
3681 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
3684 /* Calculate the logical end of the stripe */
3685 get_raid56_logic_offset(physical_end, stripe_index,
3686 map, &logic_end, NULL);
3687 logic_end += chunk_logical;
3689 /* Initialize @offset in case we need to go to out: label */
3690 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
3691 increment = map->stripe_len * nr_data_stripes(map);
3694 * Due to the rotation, for RAID56 it's better to iterate each stripe
3695 * using their physical offset.
3697 while (physical < physical_end) {
3698 ret = get_raid56_logic_offset(physical, stripe_index, map,
3699 &logical, &stripe_logical);
3700 logical += chunk_logical;
3702 /* it is parity strip */
3703 stripe_logical += chunk_logical;
3704 stripe_end = stripe_logical + increment;
3705 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3714 * Now we're at a data stripe, scrub each extents in the range.
3716 * At this stage, if we ignore the repair part, inside each data
3717 * stripe it is no different than SINGLE profile.
3718 * We can reuse scrub_simple_mirror() here, as the repair part
3719 * is still based on @mirror_num.
3721 ret = scrub_simple_mirror(sctx, root, csum_root, bg, map,
3722 logical, map->stripe_len,
3723 scrub_dev, physical, 1);
3727 logical += increment;
3728 physical += map->stripe_len;
3729 spin_lock(&sctx->stat_lock);
3731 sctx->stat.last_physical =
3732 map->stripes[stripe_index].physical + dev_stripe_len;
3734 sctx->stat.last_physical = physical;
3735 spin_unlock(&sctx->stat_lock);
3740 /* push queued extents */
3742 mutex_lock(&sctx->wr_lock);
3743 scrub_wr_submit(sctx);
3744 mutex_unlock(&sctx->wr_lock);
3746 blk_finish_plug(&plug);
3747 btrfs_free_path(path);
3749 if (sctx->is_dev_replace && ret >= 0) {
3752 ret2 = sync_write_pointer_for_zoned(sctx,
3753 chunk_logical + offset,
3754 map->stripes[stripe_index].physical,
3760 return ret < 0 ? ret : 0;
3763 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3764 struct btrfs_block_group *bg,
3765 struct btrfs_device *scrub_dev,
3769 struct btrfs_fs_info *fs_info = sctx->fs_info;
3770 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3771 struct map_lookup *map;
3772 struct extent_map *em;
3776 read_lock(&map_tree->lock);
3777 em = lookup_extent_mapping(map_tree, bg->start, bg->length);
3778 read_unlock(&map_tree->lock);
3782 * Might have been an unused block group deleted by the cleaner
3783 * kthread or relocation.
3785 spin_lock(&bg->lock);
3786 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
3788 spin_unlock(&bg->lock);
3792 if (em->start != bg->start)
3794 if (em->len < dev_extent_len)
3797 map = em->map_lookup;
3798 for (i = 0; i < map->num_stripes; ++i) {
3799 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3800 map->stripes[i].physical == dev_offset) {
3801 ret = scrub_stripe(sctx, bg, em, scrub_dev, i);
3807 free_extent_map(em);
3812 static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3813 struct btrfs_block_group *cache)
3815 struct btrfs_fs_info *fs_info = cache->fs_info;
3816 struct btrfs_trans_handle *trans;
3818 if (!btrfs_is_zoned(fs_info))
3821 btrfs_wait_block_group_reservations(cache);
3822 btrfs_wait_nocow_writers(cache);
3823 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3825 trans = btrfs_join_transaction(root);
3827 return PTR_ERR(trans);
3828 return btrfs_commit_transaction(trans);
3831 static noinline_for_stack
3832 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3833 struct btrfs_device *scrub_dev, u64 start, u64 end)
3835 struct btrfs_dev_extent *dev_extent = NULL;
3836 struct btrfs_path *path;
3837 struct btrfs_fs_info *fs_info = sctx->fs_info;
3838 struct btrfs_root *root = fs_info->dev_root;
3843 struct extent_buffer *l;
3844 struct btrfs_key key;
3845 struct btrfs_key found_key;
3846 struct btrfs_block_group *cache;
3847 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3849 path = btrfs_alloc_path();
3853 path->reada = READA_FORWARD;
3854 path->search_commit_root = 1;
3855 path->skip_locking = 1;
3857 key.objectid = scrub_dev->devid;
3859 key.type = BTRFS_DEV_EXTENT_KEY;
3864 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3868 if (path->slots[0] >=
3869 btrfs_header_nritems(path->nodes[0])) {
3870 ret = btrfs_next_leaf(root, path);
3883 slot = path->slots[0];
3885 btrfs_item_key_to_cpu(l, &found_key, slot);
3887 if (found_key.objectid != scrub_dev->devid)
3890 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3893 if (found_key.offset >= end)
3896 if (found_key.offset < key.offset)
3899 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3900 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
3902 if (found_key.offset + dev_extent_len <= start)
3905 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3908 * get a reference on the corresponding block group to prevent
3909 * the chunk from going away while we scrub it
3911 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3913 /* some chunks are removed but not committed to disk yet,
3914 * continue scrubbing */
3918 ASSERT(cache->start <= chunk_offset);
3920 * We are using the commit root to search for device extents, so
3921 * that means we could have found a device extent item from a
3922 * block group that was deleted in the current transaction. The
3923 * logical start offset of the deleted block group, stored at
3924 * @chunk_offset, might be part of the logical address range of
3925 * a new block group (which uses different physical extents).
3926 * In this case btrfs_lookup_block_group() has returned the new
3927 * block group, and its start address is less than @chunk_offset.
3929 * We skip such new block groups, because it's pointless to
3930 * process them, as we won't find their extents because we search
3931 * for them using the commit root of the extent tree. For a device
3932 * replace it's also fine to skip it, we won't miss copying them
3933 * to the target device because we have the write duplication
3934 * setup through the regular write path (by btrfs_map_block()),
3935 * and we have committed a transaction when we started the device
3936 * replace, right after setting up the device replace state.
3938 if (cache->start < chunk_offset) {
3939 btrfs_put_block_group(cache);
3943 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3944 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
3945 btrfs_put_block_group(cache);
3951 * Make sure that while we are scrubbing the corresponding block
3952 * group doesn't get its logical address and its device extents
3953 * reused for another block group, which can possibly be of a
3954 * different type and different profile. We do this to prevent
3955 * false error detections and crashes due to bogus attempts to
3958 spin_lock(&cache->lock);
3959 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
3960 spin_unlock(&cache->lock);
3961 btrfs_put_block_group(cache);
3964 btrfs_freeze_block_group(cache);
3965 spin_unlock(&cache->lock);
3968 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3969 * to avoid deadlock caused by:
3970 * btrfs_inc_block_group_ro()
3971 * -> btrfs_wait_for_commit()
3972 * -> btrfs_commit_transaction()
3973 * -> btrfs_scrub_pause()
3975 scrub_pause_on(fs_info);
3978 * Don't do chunk preallocation for scrub.
3980 * This is especially important for SYSTEM bgs, or we can hit
3981 * -EFBIG from btrfs_finish_chunk_alloc() like:
3982 * 1. The only SYSTEM bg is marked RO.
3983 * Since SYSTEM bg is small, that's pretty common.
3984 * 2. New SYSTEM bg will be allocated
3985 * Due to regular version will allocate new chunk.
3986 * 3. New SYSTEM bg is empty and will get cleaned up
3987 * Before cleanup really happens, it's marked RO again.
3988 * 4. Empty SYSTEM bg get scrubbed
3991 * This can easily boost the amount of SYSTEM chunks if cleaner
3992 * thread can't be triggered fast enough, and use up all space
3993 * of btrfs_super_block::sys_chunk_array
3995 * While for dev replace, we need to try our best to mark block
3996 * group RO, to prevent race between:
3997 * - Write duplication
3998 * Contains latest data
4000 * Contains data from commit tree
4002 * If target block group is not marked RO, nocow writes can
4003 * be overwritten by scrub copy, causing data corruption.
4004 * So for dev-replace, it's not allowed to continue if a block
4007 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
4008 if (!ret && sctx->is_dev_replace) {
4009 ret = finish_extent_writes_for_zoned(root, cache);
4011 btrfs_dec_block_group_ro(cache);
4012 scrub_pause_off(fs_info);
4013 btrfs_put_block_group(cache);
4020 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
4021 !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
4023 * btrfs_inc_block_group_ro return -ENOSPC when it
4024 * failed in creating new chunk for metadata.
4025 * It is not a problem for scrub, because
4026 * metadata are always cowed, and our scrub paused
4027 * commit_transactions.
4029 * For RAID56 chunks, we have to mark them read-only
4030 * for scrub, as later we would use our own cache
4031 * out of RAID56 realm.
4032 * Thus we want the RAID56 bg to be marked RO to
4033 * prevent RMW from screwing up out cache.
4036 } else if (ret == -ETXTBSY) {
4038 "skipping scrub of block group %llu due to active swapfile",
4040 scrub_pause_off(fs_info);
4045 "failed setting block group ro: %d", ret);
4046 btrfs_unfreeze_block_group(cache);
4047 btrfs_put_block_group(cache);
4048 scrub_pause_off(fs_info);
4053 * Now the target block is marked RO, wait for nocow writes to
4054 * finish before dev-replace.
4055 * COW is fine, as COW never overwrites extents in commit tree.
4057 if (sctx->is_dev_replace) {
4058 btrfs_wait_nocow_writers(cache);
4059 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
4063 scrub_pause_off(fs_info);
4064 down_write(&dev_replace->rwsem);
4065 dev_replace->cursor_right = found_key.offset + dev_extent_len;
4066 dev_replace->cursor_left = found_key.offset;
4067 dev_replace->item_needs_writeback = 1;
4068 up_write(&dev_replace->rwsem);
4070 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
4074 * flush, submit all pending read and write bios, afterwards
4076 * Note that in the dev replace case, a read request causes
4077 * write requests that are submitted in the read completion
4078 * worker. Therefore in the current situation, it is required
4079 * that all write requests are flushed, so that all read and
4080 * write requests are really completed when bios_in_flight
4083 sctx->flush_all_writes = true;
4085 mutex_lock(&sctx->wr_lock);
4086 scrub_wr_submit(sctx);
4087 mutex_unlock(&sctx->wr_lock);
4089 wait_event(sctx->list_wait,
4090 atomic_read(&sctx->bios_in_flight) == 0);
4092 scrub_pause_on(fs_info);
4095 * must be called before we decrease @scrub_paused.
4096 * make sure we don't block transaction commit while
4097 * we are waiting pending workers finished.
4099 wait_event(sctx->list_wait,
4100 atomic_read(&sctx->workers_pending) == 0);
4101 sctx->flush_all_writes = false;
4103 scrub_pause_off(fs_info);
4105 if (sctx->is_dev_replace &&
4106 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
4107 cache, found_key.offset))
4110 down_write(&dev_replace->rwsem);
4111 dev_replace->cursor_left = dev_replace->cursor_right;
4112 dev_replace->item_needs_writeback = 1;
4113 up_write(&dev_replace->rwsem);
4116 btrfs_dec_block_group_ro(cache);
4119 * We might have prevented the cleaner kthread from deleting
4120 * this block group if it was already unused because we raced
4121 * and set it to RO mode first. So add it back to the unused
4122 * list, otherwise it might not ever be deleted unless a manual
4123 * balance is triggered or it becomes used and unused again.
4125 spin_lock(&cache->lock);
4126 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
4127 !cache->ro && cache->reserved == 0 && cache->used == 0) {
4128 spin_unlock(&cache->lock);
4129 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
4130 btrfs_discard_queue_work(&fs_info->discard_ctl,
4133 btrfs_mark_bg_unused(cache);
4135 spin_unlock(&cache->lock);
4138 btrfs_unfreeze_block_group(cache);
4139 btrfs_put_block_group(cache);
4142 if (sctx->is_dev_replace &&
4143 atomic64_read(&dev_replace->num_write_errors) > 0) {
4147 if (sctx->stat.malloc_errors > 0) {
4152 key.offset = found_key.offset + dev_extent_len;
4153 btrfs_release_path(path);
4156 btrfs_free_path(path);
4161 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4162 struct btrfs_device *scrub_dev)
4168 struct btrfs_fs_info *fs_info = sctx->fs_info;
4170 if (BTRFS_FS_ERROR(fs_info))
4173 /* Seed devices of a new filesystem has their own generation. */
4174 if (scrub_dev->fs_devices != fs_info->fs_devices)
4175 gen = scrub_dev->generation;
4177 gen = fs_info->last_trans_committed;
4179 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4180 ret = btrfs_sb_log_location(scrub_dev, i, 0, &bytenr);
4185 spin_lock(&sctx->stat_lock);
4186 sctx->stat.super_errors++;
4187 spin_unlock(&sctx->stat_lock);
4191 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4192 scrub_dev->commit_total_bytes)
4194 if (!btrfs_check_super_location(scrub_dev, bytenr))
4197 ret = scrub_sectors(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4198 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4203 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4208 static void scrub_workers_put(struct btrfs_fs_info *fs_info)
4210 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
4211 &fs_info->scrub_lock)) {
4212 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
4213 struct workqueue_struct *scrub_wr_comp =
4214 fs_info->scrub_wr_completion_workers;
4215 struct workqueue_struct *scrub_parity =
4216 fs_info->scrub_parity_workers;
4218 fs_info->scrub_workers = NULL;
4219 fs_info->scrub_wr_completion_workers = NULL;
4220 fs_info->scrub_parity_workers = NULL;
4221 mutex_unlock(&fs_info->scrub_lock);
4224 destroy_workqueue(scrub_workers);
4226 destroy_workqueue(scrub_wr_comp);
4228 destroy_workqueue(scrub_parity);
4233 * get a reference count on fs_info->scrub_workers. start worker if necessary
4235 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4238 struct workqueue_struct *scrub_workers = NULL;
4239 struct workqueue_struct *scrub_wr_comp = NULL;
4240 struct workqueue_struct *scrub_parity = NULL;
4241 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4242 int max_active = fs_info->thread_pool_size;
4245 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
4248 scrub_workers = alloc_workqueue("btrfs-scrub", flags,
4249 is_dev_replace ? 1 : max_active);
4251 goto fail_scrub_workers;
4253 scrub_wr_comp = alloc_workqueue("btrfs-scrubwrc", flags, max_active);
4255 goto fail_scrub_wr_completion_workers;
4257 scrub_parity = alloc_workqueue("btrfs-scrubparity", flags, max_active);
4259 goto fail_scrub_parity_workers;
4261 mutex_lock(&fs_info->scrub_lock);
4262 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
4263 ASSERT(fs_info->scrub_workers == NULL &&
4264 fs_info->scrub_wr_completion_workers == NULL &&
4265 fs_info->scrub_parity_workers == NULL);
4266 fs_info->scrub_workers = scrub_workers;
4267 fs_info->scrub_wr_completion_workers = scrub_wr_comp;
4268 fs_info->scrub_parity_workers = scrub_parity;
4269 refcount_set(&fs_info->scrub_workers_refcnt, 1);
4270 mutex_unlock(&fs_info->scrub_lock);
4273 /* Other thread raced in and created the workers for us */
4274 refcount_inc(&fs_info->scrub_workers_refcnt);
4275 mutex_unlock(&fs_info->scrub_lock);
4278 destroy_workqueue(scrub_parity);
4279 fail_scrub_parity_workers:
4280 destroy_workqueue(scrub_wr_comp);
4281 fail_scrub_wr_completion_workers:
4282 destroy_workqueue(scrub_workers);
4287 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4288 u64 end, struct btrfs_scrub_progress *progress,
4289 int readonly, int is_dev_replace)
4291 struct btrfs_dev_lookup_args args = { .devid = devid };
4292 struct scrub_ctx *sctx;
4294 struct btrfs_device *dev;
4295 unsigned int nofs_flag;
4296 bool need_commit = false;
4298 if (btrfs_fs_closing(fs_info))
4301 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
4302 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
4305 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
4306 * value (max nodesize / min sectorsize), thus nodesize should always
4309 ASSERT(fs_info->nodesize <=
4310 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
4312 /* Allocate outside of device_list_mutex */
4313 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4315 return PTR_ERR(sctx);
4317 ret = scrub_workers_get(fs_info, is_dev_replace);
4321 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4322 dev = btrfs_find_device(fs_info->fs_devices, &args);
4323 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4325 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4330 if (!is_dev_replace && !readonly &&
4331 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4332 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4333 btrfs_err_in_rcu(fs_info,
4334 "scrub on devid %llu: filesystem on %s is not writable",
4335 devid, rcu_str_deref(dev->name));
4340 mutex_lock(&fs_info->scrub_lock);
4341 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4342 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4343 mutex_unlock(&fs_info->scrub_lock);
4344 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4349 down_read(&fs_info->dev_replace.rwsem);
4350 if (dev->scrub_ctx ||
4352 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4353 up_read(&fs_info->dev_replace.rwsem);
4354 mutex_unlock(&fs_info->scrub_lock);
4355 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4359 up_read(&fs_info->dev_replace.rwsem);
4361 sctx->readonly = readonly;
4362 dev->scrub_ctx = sctx;
4363 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4366 * checking @scrub_pause_req here, we can avoid
4367 * race between committing transaction and scrubbing.
4369 __scrub_blocked_if_needed(fs_info);
4370 atomic_inc(&fs_info->scrubs_running);
4371 mutex_unlock(&fs_info->scrub_lock);
4374 * In order to avoid deadlock with reclaim when there is a transaction
4375 * trying to pause scrub, make sure we use GFP_NOFS for all the
4376 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
4377 * invoked by our callees. The pausing request is done when the
4378 * transaction commit starts, and it blocks the transaction until scrub
4379 * is paused (done at specific points at scrub_stripe() or right above
4380 * before incrementing fs_info->scrubs_running).
4382 nofs_flag = memalloc_nofs_save();
4383 if (!is_dev_replace) {
4384 u64 old_super_errors;
4386 spin_lock(&sctx->stat_lock);
4387 old_super_errors = sctx->stat.super_errors;
4388 spin_unlock(&sctx->stat_lock);
4390 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4392 * by holding device list mutex, we can
4393 * kick off writing super in log tree sync.
4395 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4396 ret = scrub_supers(sctx, dev);
4397 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4399 spin_lock(&sctx->stat_lock);
4401 * Super block errors found, but we can not commit transaction
4402 * at current context, since btrfs_commit_transaction() needs
4403 * to pause the current running scrub (hold by ourselves).
4405 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
4407 spin_unlock(&sctx->stat_lock);
4411 ret = scrub_enumerate_chunks(sctx, dev, start, end);
4412 memalloc_nofs_restore(nofs_flag);
4414 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4415 atomic_dec(&fs_info->scrubs_running);
4416 wake_up(&fs_info->scrub_pause_wait);
4418 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4421 memcpy(progress, &sctx->stat, sizeof(*progress));
4423 if (!is_dev_replace)
4424 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4425 ret ? "not finished" : "finished", devid, ret);
4427 mutex_lock(&fs_info->scrub_lock);
4428 dev->scrub_ctx = NULL;
4429 mutex_unlock(&fs_info->scrub_lock);
4431 scrub_workers_put(fs_info);
4432 scrub_put_ctx(sctx);
4435 * We found some super block errors before, now try to force a
4436 * transaction commit, as scrub has finished.
4439 struct btrfs_trans_handle *trans;
4441 trans = btrfs_start_transaction(fs_info->tree_root, 0);
4442 if (IS_ERR(trans)) {
4443 ret = PTR_ERR(trans);
4445 "scrub: failed to start transaction to fix super block errors: %d", ret);
4448 ret = btrfs_commit_transaction(trans);
4451 "scrub: failed to commit transaction to fix super block errors: %d", ret);
4455 scrub_workers_put(fs_info);
4457 scrub_free_ctx(sctx);
4462 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4464 mutex_lock(&fs_info->scrub_lock);
4465 atomic_inc(&fs_info->scrub_pause_req);
4466 while (atomic_read(&fs_info->scrubs_paused) !=
4467 atomic_read(&fs_info->scrubs_running)) {
4468 mutex_unlock(&fs_info->scrub_lock);
4469 wait_event(fs_info->scrub_pause_wait,
4470 atomic_read(&fs_info->scrubs_paused) ==
4471 atomic_read(&fs_info->scrubs_running));
4472 mutex_lock(&fs_info->scrub_lock);
4474 mutex_unlock(&fs_info->scrub_lock);
4477 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4479 atomic_dec(&fs_info->scrub_pause_req);
4480 wake_up(&fs_info->scrub_pause_wait);
4483 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4485 mutex_lock(&fs_info->scrub_lock);
4486 if (!atomic_read(&fs_info->scrubs_running)) {
4487 mutex_unlock(&fs_info->scrub_lock);
4491 atomic_inc(&fs_info->scrub_cancel_req);
4492 while (atomic_read(&fs_info->scrubs_running)) {
4493 mutex_unlock(&fs_info->scrub_lock);
4494 wait_event(fs_info->scrub_pause_wait,
4495 atomic_read(&fs_info->scrubs_running) == 0);
4496 mutex_lock(&fs_info->scrub_lock);
4498 atomic_dec(&fs_info->scrub_cancel_req);
4499 mutex_unlock(&fs_info->scrub_lock);
4504 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4506 struct btrfs_fs_info *fs_info = dev->fs_info;
4507 struct scrub_ctx *sctx;
4509 mutex_lock(&fs_info->scrub_lock);
4510 sctx = dev->scrub_ctx;
4512 mutex_unlock(&fs_info->scrub_lock);
4515 atomic_inc(&sctx->cancel_req);
4516 while (dev->scrub_ctx) {
4517 mutex_unlock(&fs_info->scrub_lock);
4518 wait_event(fs_info->scrub_pause_wait,
4519 dev->scrub_ctx == NULL);
4520 mutex_lock(&fs_info->scrub_lock);
4522 mutex_unlock(&fs_info->scrub_lock);
4527 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4528 struct btrfs_scrub_progress *progress)
4530 struct btrfs_dev_lookup_args args = { .devid = devid };
4531 struct btrfs_device *dev;
4532 struct scrub_ctx *sctx = NULL;
4534 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4535 dev = btrfs_find_device(fs_info->fs_devices, &args);
4537 sctx = dev->scrub_ctx;
4539 memcpy(progress, &sctx->stat, sizeof(*progress));
4540 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4542 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4545 static void scrub_find_good_copy(struct btrfs_fs_info *fs_info,
4546 u64 extent_logical, u32 extent_len,
4547 u64 *extent_physical,
4548 struct btrfs_device **extent_dev,
4549 int *extent_mirror_num)
4552 struct btrfs_io_context *bioc = NULL;
4555 mapped_length = extent_len;
4556 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4557 &mapped_length, &bioc, 0);
4558 if (ret || !bioc || mapped_length < extent_len ||
4559 !bioc->stripes[0].dev->bdev) {
4560 btrfs_put_bioc(bioc);
4564 *extent_physical = bioc->stripes[0].physical;
4565 *extent_mirror_num = bioc->mirror_num;
4566 *extent_dev = bioc->stripes[0].dev;
4567 btrfs_put_bioc(bioc);