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>
12 #include "ordered-data.h"
13 #include "transaction.h"
15 #include "extent_io.h"
16 #include "dev-replace.h"
17 #include "check-integrity.h"
18 #include "rcu-string.h"
22 * This is only the first step towards a full-features scrub. It reads all
23 * extent and super block and verifies the checksums. In case a bad checksum
24 * is found or the extent cannot be read, good data will be written back if
27 * Future enhancements:
28 * - In case an unrepairable extent is encountered, track which files are
29 * affected and report them
30 * - track and record media errors, throw out bad devices
31 * - add a mode to also read unallocated space
38 * the following three values only influence the performance.
39 * The last one configures the number of parallel and outstanding I/O
40 * operations. The first two values configure an upper limit for the number
41 * of (dynamically allocated) pages that are added to a bio.
43 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
44 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
45 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
48 * the following value times PAGE_SIZE needs to be large enough to match the
49 * largest node/leaf/sector size that shall be supported.
50 * Values larger than BTRFS_STRIPE_LEN are not supported.
52 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
54 struct scrub_recover {
56 struct btrfs_bio *bbio;
61 struct scrub_block *sblock;
63 struct btrfs_device *dev;
64 struct list_head list;
65 u64 flags; /* extent flags */
69 u64 physical_for_dev_replace;
72 unsigned int mirror_num:8;
73 unsigned int have_csum:1;
74 unsigned int io_error:1;
76 u8 csum[BTRFS_CSUM_SIZE];
78 struct scrub_recover *recover;
83 struct scrub_ctx *sctx;
84 struct btrfs_device *dev;
89 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
90 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
92 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
96 struct btrfs_work work;
100 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
102 atomic_t outstanding_pages;
103 refcount_t refs; /* free mem on transition to zero */
104 struct scrub_ctx *sctx;
105 struct scrub_parity *sparity;
107 unsigned int header_error:1;
108 unsigned int checksum_error:1;
109 unsigned int no_io_error_seen:1;
110 unsigned int generation_error:1; /* also sets header_error */
112 /* The following is for the data used to check parity */
113 /* It is for the data with checksum */
114 unsigned int data_corrected:1;
116 struct btrfs_work work;
119 /* Used for the chunks with parity stripe such RAID5/6 */
120 struct scrub_parity {
121 struct scrub_ctx *sctx;
123 struct btrfs_device *scrub_dev;
135 struct list_head spages;
137 /* Work of parity check and repair */
138 struct btrfs_work work;
140 /* Mark the parity blocks which have data */
141 unsigned long *dbitmap;
144 * Mark the parity blocks which have data, but errors happen when
145 * read data or check data
147 unsigned long *ebitmap;
149 unsigned long bitmap[0];
153 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
154 struct btrfs_fs_info *fs_info;
157 atomic_t bios_in_flight;
158 atomic_t workers_pending;
159 spinlock_t list_lock;
160 wait_queue_head_t list_wait;
162 struct list_head csum_list;
165 int pages_per_rd_bio;
169 struct scrub_bio *wr_curr_bio;
170 struct mutex wr_lock;
171 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
172 struct btrfs_device *wr_tgtdev;
173 bool flush_all_writes;
178 struct btrfs_scrub_progress stat;
179 spinlock_t stat_lock;
182 * Use a ref counter to avoid use-after-free issues. Scrub workers
183 * decrement bios_in_flight and workers_pending and then do a wakeup
184 * on the list_wait wait queue. We must ensure the main scrub task
185 * doesn't free the scrub context before or while the workers are
186 * doing the wakeup() call.
191 struct scrub_warning {
192 struct btrfs_path *path;
193 u64 extent_item_size;
197 struct btrfs_device *dev;
200 struct full_stripe_lock {
207 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
208 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
209 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
210 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
211 struct scrub_block *sblocks_for_recheck);
212 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
213 struct scrub_block *sblock,
214 int retry_failed_mirror);
215 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
216 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
217 struct scrub_block *sblock_good);
218 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
219 struct scrub_block *sblock_good,
220 int page_num, int force_write);
221 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
222 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
224 static int scrub_checksum_data(struct scrub_block *sblock);
225 static int scrub_checksum_tree_block(struct scrub_block *sblock);
226 static int scrub_checksum_super(struct scrub_block *sblock);
227 static void scrub_block_get(struct scrub_block *sblock);
228 static void scrub_block_put(struct scrub_block *sblock);
229 static void scrub_page_get(struct scrub_page *spage);
230 static void scrub_page_put(struct scrub_page *spage);
231 static void scrub_parity_get(struct scrub_parity *sparity);
232 static void scrub_parity_put(struct scrub_parity *sparity);
233 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
234 struct scrub_page *spage);
235 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
236 u64 physical, struct btrfs_device *dev, u64 flags,
237 u64 gen, int mirror_num, u8 *csum, int force,
238 u64 physical_for_dev_replace);
239 static void scrub_bio_end_io(struct bio *bio);
240 static void scrub_bio_end_io_worker(struct btrfs_work *work);
241 static void scrub_block_complete(struct scrub_block *sblock);
242 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
243 u64 extent_logical, u64 extent_len,
244 u64 *extent_physical,
245 struct btrfs_device **extent_dev,
246 int *extent_mirror_num);
247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
248 struct scrub_page *spage);
249 static void scrub_wr_submit(struct scrub_ctx *sctx);
250 static void scrub_wr_bio_end_io(struct bio *bio);
251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
252 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
253 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
254 static void scrub_put_ctx(struct scrub_ctx *sctx);
256 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
258 return page->recover &&
259 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
262 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
264 refcount_inc(&sctx->refs);
265 atomic_inc(&sctx->bios_in_flight);
268 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
270 atomic_dec(&sctx->bios_in_flight);
271 wake_up(&sctx->list_wait);
275 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
277 while (atomic_read(&fs_info->scrub_pause_req)) {
278 mutex_unlock(&fs_info->scrub_lock);
279 wait_event(fs_info->scrub_pause_wait,
280 atomic_read(&fs_info->scrub_pause_req) == 0);
281 mutex_lock(&fs_info->scrub_lock);
285 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
287 atomic_inc(&fs_info->scrubs_paused);
288 wake_up(&fs_info->scrub_pause_wait);
291 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
293 mutex_lock(&fs_info->scrub_lock);
294 __scrub_blocked_if_needed(fs_info);
295 atomic_dec(&fs_info->scrubs_paused);
296 mutex_unlock(&fs_info->scrub_lock);
298 wake_up(&fs_info->scrub_pause_wait);
301 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
303 scrub_pause_on(fs_info);
304 scrub_pause_off(fs_info);
308 * Insert new full stripe lock into full stripe locks tree
310 * Return pointer to existing or newly inserted full_stripe_lock structure if
311 * everything works well.
312 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
314 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
317 static struct full_stripe_lock *insert_full_stripe_lock(
318 struct btrfs_full_stripe_locks_tree *locks_root,
322 struct rb_node *parent = NULL;
323 struct full_stripe_lock *entry;
324 struct full_stripe_lock *ret;
326 lockdep_assert_held(&locks_root->lock);
328 p = &locks_root->root.rb_node;
331 entry = rb_entry(parent, struct full_stripe_lock, node);
332 if (fstripe_logical < entry->logical) {
334 } else if (fstripe_logical > entry->logical) {
342 /* Insert new lock */
343 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
345 return ERR_PTR(-ENOMEM);
346 ret->logical = fstripe_logical;
348 mutex_init(&ret->mutex);
350 rb_link_node(&ret->node, parent, p);
351 rb_insert_color(&ret->node, &locks_root->root);
356 * Search for a full stripe lock of a block group
358 * Return pointer to existing full stripe lock if found
359 * Return NULL if not found
361 static struct full_stripe_lock *search_full_stripe_lock(
362 struct btrfs_full_stripe_locks_tree *locks_root,
365 struct rb_node *node;
366 struct full_stripe_lock *entry;
368 lockdep_assert_held(&locks_root->lock);
370 node = locks_root->root.rb_node;
372 entry = rb_entry(node, struct full_stripe_lock, node);
373 if (fstripe_logical < entry->logical)
374 node = node->rb_left;
375 else if (fstripe_logical > entry->logical)
376 node = node->rb_right;
384 * Helper to get full stripe logical from a normal bytenr.
386 * Caller must ensure @cache is a RAID56 block group.
388 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
394 * Due to chunk item size limit, full stripe length should not be
395 * larger than U32_MAX. Just a sanity check here.
397 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
400 * round_down() can only handle power of 2, while RAID56 full
401 * stripe length can be 64KiB * n, so we need to manually round down.
403 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
404 cache->full_stripe_len + cache->key.objectid;
409 * Lock a full stripe to avoid concurrency of recovery and read
411 * It's only used for profiles with parities (RAID5/6), for other profiles it
414 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
415 * So caller must call unlock_full_stripe() at the same context.
417 * Return <0 if encounters error.
419 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
422 struct btrfs_block_group_cache *bg_cache;
423 struct btrfs_full_stripe_locks_tree *locks_root;
424 struct full_stripe_lock *existing;
429 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
435 /* Profiles not based on parity don't need full stripe lock */
436 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
438 locks_root = &bg_cache->full_stripe_locks_root;
440 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
442 /* Now insert the full stripe lock */
443 mutex_lock(&locks_root->lock);
444 existing = insert_full_stripe_lock(locks_root, fstripe_start);
445 mutex_unlock(&locks_root->lock);
446 if (IS_ERR(existing)) {
447 ret = PTR_ERR(existing);
450 mutex_lock(&existing->mutex);
453 btrfs_put_block_group(bg_cache);
458 * Unlock a full stripe.
460 * NOTE: Caller must ensure it's the same context calling corresponding
461 * lock_full_stripe().
463 * Return 0 if we unlock full stripe without problem.
464 * Return <0 for error
466 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
469 struct btrfs_block_group_cache *bg_cache;
470 struct btrfs_full_stripe_locks_tree *locks_root;
471 struct full_stripe_lock *fstripe_lock;
476 /* If we didn't acquire full stripe lock, no need to continue */
480 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
485 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
488 locks_root = &bg_cache->full_stripe_locks_root;
489 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
491 mutex_lock(&locks_root->lock);
492 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
493 /* Unpaired unlock_full_stripe() detected */
497 mutex_unlock(&locks_root->lock);
501 if (fstripe_lock->refs == 0) {
503 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
504 fstripe_lock->logical);
506 fstripe_lock->refs--;
509 if (fstripe_lock->refs == 0) {
510 rb_erase(&fstripe_lock->node, &locks_root->root);
513 mutex_unlock(&locks_root->lock);
515 mutex_unlock(&fstripe_lock->mutex);
519 btrfs_put_block_group(bg_cache);
523 static void scrub_free_csums(struct scrub_ctx *sctx)
525 while (!list_empty(&sctx->csum_list)) {
526 struct btrfs_ordered_sum *sum;
527 sum = list_first_entry(&sctx->csum_list,
528 struct btrfs_ordered_sum, list);
529 list_del(&sum->list);
534 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
541 /* this can happen when scrub is cancelled */
542 if (sctx->curr != -1) {
543 struct scrub_bio *sbio = sctx->bios[sctx->curr];
545 for (i = 0; i < sbio->page_count; i++) {
546 WARN_ON(!sbio->pagev[i]->page);
547 scrub_block_put(sbio->pagev[i]->sblock);
552 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
553 struct scrub_bio *sbio = sctx->bios[i];
560 kfree(sctx->wr_curr_bio);
561 scrub_free_csums(sctx);
565 static void scrub_put_ctx(struct scrub_ctx *sctx)
567 if (refcount_dec_and_test(&sctx->refs))
568 scrub_free_ctx(sctx);
571 static noinline_for_stack
572 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
574 struct scrub_ctx *sctx;
576 struct btrfs_fs_info *fs_info = dev->fs_info;
578 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
581 refcount_set(&sctx->refs, 1);
582 sctx->is_dev_replace = is_dev_replace;
583 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
585 sctx->fs_info = dev->fs_info;
586 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
587 struct scrub_bio *sbio;
589 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
592 sctx->bios[i] = sbio;
596 sbio->page_count = 0;
597 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
598 scrub_bio_end_io_worker, NULL, NULL);
600 if (i != SCRUB_BIOS_PER_SCTX - 1)
601 sctx->bios[i]->next_free = i + 1;
603 sctx->bios[i]->next_free = -1;
605 sctx->first_free = 0;
606 atomic_set(&sctx->bios_in_flight, 0);
607 atomic_set(&sctx->workers_pending, 0);
608 atomic_set(&sctx->cancel_req, 0);
609 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
610 INIT_LIST_HEAD(&sctx->csum_list);
612 spin_lock_init(&sctx->list_lock);
613 spin_lock_init(&sctx->stat_lock);
614 init_waitqueue_head(&sctx->list_wait);
616 WARN_ON(sctx->wr_curr_bio != NULL);
617 mutex_init(&sctx->wr_lock);
618 sctx->wr_curr_bio = NULL;
619 if (is_dev_replace) {
620 WARN_ON(!fs_info->dev_replace.tgtdev);
621 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
622 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
623 sctx->flush_all_writes = false;
629 scrub_free_ctx(sctx);
630 return ERR_PTR(-ENOMEM);
633 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
641 struct extent_buffer *eb;
642 struct btrfs_inode_item *inode_item;
643 struct scrub_warning *swarn = warn_ctx;
644 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
645 struct inode_fs_paths *ipath = NULL;
646 struct btrfs_root *local_root;
647 struct btrfs_key root_key;
648 struct btrfs_key key;
650 root_key.objectid = root;
651 root_key.type = BTRFS_ROOT_ITEM_KEY;
652 root_key.offset = (u64)-1;
653 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
654 if (IS_ERR(local_root)) {
655 ret = PTR_ERR(local_root);
660 * this makes the path point to (inum INODE_ITEM ioff)
663 key.type = BTRFS_INODE_ITEM_KEY;
666 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
668 btrfs_release_path(swarn->path);
672 eb = swarn->path->nodes[0];
673 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
674 struct btrfs_inode_item);
675 isize = btrfs_inode_size(eb, inode_item);
676 nlink = btrfs_inode_nlink(eb, inode_item);
677 btrfs_release_path(swarn->path);
680 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
681 * uses GFP_NOFS in this context, so we keep it consistent but it does
682 * not seem to be strictly necessary.
684 nofs_flag = memalloc_nofs_save();
685 ipath = init_ipath(4096, local_root, swarn->path);
686 memalloc_nofs_restore(nofs_flag);
688 ret = PTR_ERR(ipath);
692 ret = paths_from_inode(inum, ipath);
698 * we deliberately ignore the bit ipath might have been too small to
699 * hold all of the paths here
701 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
702 btrfs_warn_in_rcu(fs_info,
703 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
704 swarn->errstr, swarn->logical,
705 rcu_str_deref(swarn->dev->name),
708 min(isize - offset, (u64)PAGE_SIZE), nlink,
709 (char *)(unsigned long)ipath->fspath->val[i]);
715 btrfs_warn_in_rcu(fs_info,
716 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
717 swarn->errstr, swarn->logical,
718 rcu_str_deref(swarn->dev->name),
720 root, inum, offset, ret);
726 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
728 struct btrfs_device *dev;
729 struct btrfs_fs_info *fs_info;
730 struct btrfs_path *path;
731 struct btrfs_key found_key;
732 struct extent_buffer *eb;
733 struct btrfs_extent_item *ei;
734 struct scrub_warning swarn;
735 unsigned long ptr = 0;
743 WARN_ON(sblock->page_count < 1);
744 dev = sblock->pagev[0]->dev;
745 fs_info = sblock->sctx->fs_info;
747 path = btrfs_alloc_path();
751 swarn.physical = sblock->pagev[0]->physical;
752 swarn.logical = sblock->pagev[0]->logical;
753 swarn.errstr = errstr;
756 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
761 extent_item_pos = swarn.logical - found_key.objectid;
762 swarn.extent_item_size = found_key.offset;
765 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
766 item_size = btrfs_item_size_nr(eb, path->slots[0]);
768 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
770 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
771 item_size, &ref_root,
773 btrfs_warn_in_rcu(fs_info,
774 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
775 errstr, swarn.logical,
776 rcu_str_deref(dev->name),
778 ref_level ? "node" : "leaf",
779 ret < 0 ? -1 : ref_level,
780 ret < 0 ? -1 : ref_root);
782 btrfs_release_path(path);
784 btrfs_release_path(path);
787 iterate_extent_inodes(fs_info, found_key.objectid,
789 scrub_print_warning_inode, &swarn, false);
793 btrfs_free_path(path);
796 static inline void scrub_get_recover(struct scrub_recover *recover)
798 refcount_inc(&recover->refs);
801 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
802 struct scrub_recover *recover)
804 if (refcount_dec_and_test(&recover->refs)) {
805 btrfs_bio_counter_dec(fs_info);
806 btrfs_put_bbio(recover->bbio);
812 * scrub_handle_errored_block gets called when either verification of the
813 * pages failed or the bio failed to read, e.g. with EIO. In the latter
814 * case, this function handles all pages in the bio, even though only one
816 * The goal of this function is to repair the errored block by using the
817 * contents of one of the mirrors.
819 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
821 struct scrub_ctx *sctx = sblock_to_check->sctx;
822 struct btrfs_device *dev;
823 struct btrfs_fs_info *fs_info;
825 unsigned int failed_mirror_index;
826 unsigned int is_metadata;
827 unsigned int have_csum;
828 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
829 struct scrub_block *sblock_bad;
834 bool full_stripe_locked;
835 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
836 DEFAULT_RATELIMIT_BURST);
838 BUG_ON(sblock_to_check->page_count < 1);
839 fs_info = sctx->fs_info;
840 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
842 * if we find an error in a super block, we just report it.
843 * They will get written with the next transaction commit
846 spin_lock(&sctx->stat_lock);
847 ++sctx->stat.super_errors;
848 spin_unlock(&sctx->stat_lock);
851 logical = sblock_to_check->pagev[0]->logical;
852 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
853 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
854 is_metadata = !(sblock_to_check->pagev[0]->flags &
855 BTRFS_EXTENT_FLAG_DATA);
856 have_csum = sblock_to_check->pagev[0]->have_csum;
857 dev = sblock_to_check->pagev[0]->dev;
860 * For RAID5/6, race can happen for a different device scrub thread.
861 * For data corruption, Parity and Data threads will both try
862 * to recovery the data.
863 * Race can lead to doubly added csum error, or even unrecoverable
866 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
868 spin_lock(&sctx->stat_lock);
870 sctx->stat.malloc_errors++;
871 sctx->stat.read_errors++;
872 sctx->stat.uncorrectable_errors++;
873 spin_unlock(&sctx->stat_lock);
878 * read all mirrors one after the other. This includes to
879 * re-read the extent or metadata block that failed (that was
880 * the cause that this fixup code is called) another time,
881 * page by page this time in order to know which pages
882 * caused I/O errors and which ones are good (for all mirrors).
883 * It is the goal to handle the situation when more than one
884 * mirror contains I/O errors, but the errors do not
885 * overlap, i.e. the data can be repaired by selecting the
886 * pages from those mirrors without I/O error on the
887 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
888 * would be that mirror #1 has an I/O error on the first page,
889 * the second page is good, and mirror #2 has an I/O error on
890 * the second page, but the first page is good.
891 * Then the first page of the first mirror can be repaired by
892 * taking the first page of the second mirror, and the
893 * second page of the second mirror can be repaired by
894 * copying the contents of the 2nd page of the 1st mirror.
895 * One more note: if the pages of one mirror contain I/O
896 * errors, the checksum cannot be verified. In order to get
897 * the best data for repairing, the first attempt is to find
898 * a mirror without I/O errors and with a validated checksum.
899 * Only if this is not possible, the pages are picked from
900 * mirrors with I/O errors without considering the checksum.
901 * If the latter is the case, at the end, the checksum of the
902 * repaired area is verified in order to correctly maintain
906 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
907 sizeof(*sblocks_for_recheck), GFP_NOFS);
908 if (!sblocks_for_recheck) {
909 spin_lock(&sctx->stat_lock);
910 sctx->stat.malloc_errors++;
911 sctx->stat.read_errors++;
912 sctx->stat.uncorrectable_errors++;
913 spin_unlock(&sctx->stat_lock);
914 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
918 /* setup the context, map the logical blocks and alloc the pages */
919 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
921 spin_lock(&sctx->stat_lock);
922 sctx->stat.read_errors++;
923 sctx->stat.uncorrectable_errors++;
924 spin_unlock(&sctx->stat_lock);
925 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
928 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
929 sblock_bad = sblocks_for_recheck + failed_mirror_index;
931 /* build and submit the bios for the failed mirror, check checksums */
932 scrub_recheck_block(fs_info, sblock_bad, 1);
934 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
935 sblock_bad->no_io_error_seen) {
937 * the error disappeared after reading page by page, or
938 * the area was part of a huge bio and other parts of the
939 * bio caused I/O errors, or the block layer merged several
940 * read requests into one and the error is caused by a
941 * different bio (usually one of the two latter cases is
944 spin_lock(&sctx->stat_lock);
945 sctx->stat.unverified_errors++;
946 sblock_to_check->data_corrected = 1;
947 spin_unlock(&sctx->stat_lock);
949 if (sctx->is_dev_replace)
950 scrub_write_block_to_dev_replace(sblock_bad);
954 if (!sblock_bad->no_io_error_seen) {
955 spin_lock(&sctx->stat_lock);
956 sctx->stat.read_errors++;
957 spin_unlock(&sctx->stat_lock);
958 if (__ratelimit(&_rs))
959 scrub_print_warning("i/o error", sblock_to_check);
960 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
961 } else if (sblock_bad->checksum_error) {
962 spin_lock(&sctx->stat_lock);
963 sctx->stat.csum_errors++;
964 spin_unlock(&sctx->stat_lock);
965 if (__ratelimit(&_rs))
966 scrub_print_warning("checksum error", sblock_to_check);
967 btrfs_dev_stat_inc_and_print(dev,
968 BTRFS_DEV_STAT_CORRUPTION_ERRS);
969 } else if (sblock_bad->header_error) {
970 spin_lock(&sctx->stat_lock);
971 sctx->stat.verify_errors++;
972 spin_unlock(&sctx->stat_lock);
973 if (__ratelimit(&_rs))
974 scrub_print_warning("checksum/header error",
976 if (sblock_bad->generation_error)
977 btrfs_dev_stat_inc_and_print(dev,
978 BTRFS_DEV_STAT_GENERATION_ERRS);
980 btrfs_dev_stat_inc_and_print(dev,
981 BTRFS_DEV_STAT_CORRUPTION_ERRS);
984 if (sctx->readonly) {
985 ASSERT(!sctx->is_dev_replace);
990 * now build and submit the bios for the other mirrors, check
992 * First try to pick the mirror which is completely without I/O
993 * errors and also does not have a checksum error.
994 * If one is found, and if a checksum is present, the full block
995 * that is known to contain an error is rewritten. Afterwards
996 * the block is known to be corrected.
997 * If a mirror is found which is completely correct, and no
998 * checksum is present, only those pages are rewritten that had
999 * an I/O error in the block to be repaired, since it cannot be
1000 * determined, which copy of the other pages is better (and it
1001 * could happen otherwise that a correct page would be
1002 * overwritten by a bad one).
1004 for (mirror_index = 0; ;mirror_index++) {
1005 struct scrub_block *sblock_other;
1007 if (mirror_index == failed_mirror_index)
1010 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1011 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1012 if (mirror_index >= BTRFS_MAX_MIRRORS)
1014 if (!sblocks_for_recheck[mirror_index].page_count)
1017 sblock_other = sblocks_for_recheck + mirror_index;
1019 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1020 int max_allowed = r->bbio->num_stripes -
1021 r->bbio->num_tgtdevs;
1023 if (mirror_index >= max_allowed)
1025 if (!sblocks_for_recheck[1].page_count)
1028 ASSERT(failed_mirror_index == 0);
1029 sblock_other = sblocks_for_recheck + 1;
1030 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1033 /* build and submit the bios, check checksums */
1034 scrub_recheck_block(fs_info, sblock_other, 0);
1036 if (!sblock_other->header_error &&
1037 !sblock_other->checksum_error &&
1038 sblock_other->no_io_error_seen) {
1039 if (sctx->is_dev_replace) {
1040 scrub_write_block_to_dev_replace(sblock_other);
1041 goto corrected_error;
1043 ret = scrub_repair_block_from_good_copy(
1044 sblock_bad, sblock_other);
1046 goto corrected_error;
1051 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1052 goto did_not_correct_error;
1055 * In case of I/O errors in the area that is supposed to be
1056 * repaired, continue by picking good copies of those pages.
1057 * Select the good pages from mirrors to rewrite bad pages from
1058 * the area to fix. Afterwards verify the checksum of the block
1059 * that is supposed to be repaired. This verification step is
1060 * only done for the purpose of statistic counting and for the
1061 * final scrub report, whether errors remain.
1062 * A perfect algorithm could make use of the checksum and try
1063 * all possible combinations of pages from the different mirrors
1064 * until the checksum verification succeeds. For example, when
1065 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1066 * of mirror #2 is readable but the final checksum test fails,
1067 * then the 2nd page of mirror #3 could be tried, whether now
1068 * the final checksum succeeds. But this would be a rare
1069 * exception and is therefore not implemented. At least it is
1070 * avoided that the good copy is overwritten.
1071 * A more useful improvement would be to pick the sectors
1072 * without I/O error based on sector sizes (512 bytes on legacy
1073 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1074 * mirror could be repaired by taking 512 byte of a different
1075 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1076 * area are unreadable.
1079 for (page_num = 0; page_num < sblock_bad->page_count;
1081 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1082 struct scrub_block *sblock_other = NULL;
1084 /* skip no-io-error page in scrub */
1085 if (!page_bad->io_error && !sctx->is_dev_replace)
1088 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1090 * In case of dev replace, if raid56 rebuild process
1091 * didn't work out correct data, then copy the content
1092 * in sblock_bad to make sure target device is identical
1093 * to source device, instead of writing garbage data in
1094 * sblock_for_recheck array to target device.
1096 sblock_other = NULL;
1097 } else if (page_bad->io_error) {
1098 /* try to find no-io-error page in mirrors */
1099 for (mirror_index = 0;
1100 mirror_index < BTRFS_MAX_MIRRORS &&
1101 sblocks_for_recheck[mirror_index].page_count > 0;
1103 if (!sblocks_for_recheck[mirror_index].
1104 pagev[page_num]->io_error) {
1105 sblock_other = sblocks_for_recheck +
1114 if (sctx->is_dev_replace) {
1116 * did not find a mirror to fetch the page
1117 * from. scrub_write_page_to_dev_replace()
1118 * handles this case (page->io_error), by
1119 * filling the block with zeros before
1120 * submitting the write request
1123 sblock_other = sblock_bad;
1125 if (scrub_write_page_to_dev_replace(sblock_other,
1127 btrfs_dev_replace_stats_inc(
1128 &fs_info->dev_replace.num_write_errors);
1131 } else if (sblock_other) {
1132 ret = scrub_repair_page_from_good_copy(sblock_bad,
1136 page_bad->io_error = 0;
1142 if (success && !sctx->is_dev_replace) {
1143 if (is_metadata || have_csum) {
1145 * need to verify the checksum now that all
1146 * sectors on disk are repaired (the write
1147 * request for data to be repaired is on its way).
1148 * Just be lazy and use scrub_recheck_block()
1149 * which re-reads the data before the checksum
1150 * is verified, but most likely the data comes out
1151 * of the page cache.
1153 scrub_recheck_block(fs_info, sblock_bad, 1);
1154 if (!sblock_bad->header_error &&
1155 !sblock_bad->checksum_error &&
1156 sblock_bad->no_io_error_seen)
1157 goto corrected_error;
1159 goto did_not_correct_error;
1162 spin_lock(&sctx->stat_lock);
1163 sctx->stat.corrected_errors++;
1164 sblock_to_check->data_corrected = 1;
1165 spin_unlock(&sctx->stat_lock);
1166 btrfs_err_rl_in_rcu(fs_info,
1167 "fixed up error at logical %llu on dev %s",
1168 logical, rcu_str_deref(dev->name));
1171 did_not_correct_error:
1172 spin_lock(&sctx->stat_lock);
1173 sctx->stat.uncorrectable_errors++;
1174 spin_unlock(&sctx->stat_lock);
1175 btrfs_err_rl_in_rcu(fs_info,
1176 "unable to fixup (regular) error at logical %llu on dev %s",
1177 logical, rcu_str_deref(dev->name));
1181 if (sblocks_for_recheck) {
1182 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1184 struct scrub_block *sblock = sblocks_for_recheck +
1186 struct scrub_recover *recover;
1189 for (page_index = 0; page_index < sblock->page_count;
1191 sblock->pagev[page_index]->sblock = NULL;
1192 recover = sblock->pagev[page_index]->recover;
1194 scrub_put_recover(fs_info, recover);
1195 sblock->pagev[page_index]->recover =
1198 scrub_page_put(sblock->pagev[page_index]);
1201 kfree(sblocks_for_recheck);
1204 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1210 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1212 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1214 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1217 return (int)bbio->num_stripes;
1220 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1223 int nstripes, int mirror,
1229 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1231 for (i = 0; i < nstripes; i++) {
1232 if (raid_map[i] == RAID6_Q_STRIPE ||
1233 raid_map[i] == RAID5_P_STRIPE)
1236 if (logical >= raid_map[i] &&
1237 logical < raid_map[i] + mapped_length)
1242 *stripe_offset = logical - raid_map[i];
1244 /* The other RAID type */
1245 *stripe_index = mirror;
1250 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1251 struct scrub_block *sblocks_for_recheck)
1253 struct scrub_ctx *sctx = original_sblock->sctx;
1254 struct btrfs_fs_info *fs_info = sctx->fs_info;
1255 u64 length = original_sblock->page_count * PAGE_SIZE;
1256 u64 logical = original_sblock->pagev[0]->logical;
1257 u64 generation = original_sblock->pagev[0]->generation;
1258 u64 flags = original_sblock->pagev[0]->flags;
1259 u64 have_csum = original_sblock->pagev[0]->have_csum;
1260 struct scrub_recover *recover;
1261 struct btrfs_bio *bbio;
1272 * note: the two members refs and outstanding_pages
1273 * are not used (and not set) in the blocks that are used for
1274 * the recheck procedure
1277 while (length > 0) {
1278 sublen = min_t(u64, length, PAGE_SIZE);
1279 mapped_length = sublen;
1283 * with a length of PAGE_SIZE, each returned stripe
1284 * represents one mirror
1286 btrfs_bio_counter_inc_blocked(fs_info);
1287 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1288 logical, &mapped_length, &bbio);
1289 if (ret || !bbio || mapped_length < sublen) {
1290 btrfs_put_bbio(bbio);
1291 btrfs_bio_counter_dec(fs_info);
1295 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1297 btrfs_put_bbio(bbio);
1298 btrfs_bio_counter_dec(fs_info);
1302 refcount_set(&recover->refs, 1);
1303 recover->bbio = bbio;
1304 recover->map_length = mapped_length;
1306 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1308 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1310 for (mirror_index = 0; mirror_index < nmirrors;
1312 struct scrub_block *sblock;
1313 struct scrub_page *page;
1315 sblock = sblocks_for_recheck + mirror_index;
1316 sblock->sctx = sctx;
1318 page = kzalloc(sizeof(*page), GFP_NOFS);
1321 spin_lock(&sctx->stat_lock);
1322 sctx->stat.malloc_errors++;
1323 spin_unlock(&sctx->stat_lock);
1324 scrub_put_recover(fs_info, recover);
1327 scrub_page_get(page);
1328 sblock->pagev[page_index] = page;
1329 page->sblock = sblock;
1330 page->flags = flags;
1331 page->generation = generation;
1332 page->logical = logical;
1333 page->have_csum = have_csum;
1336 original_sblock->pagev[0]->csum,
1339 scrub_stripe_index_and_offset(logical,
1348 page->physical = bbio->stripes[stripe_index].physical +
1350 page->dev = bbio->stripes[stripe_index].dev;
1352 BUG_ON(page_index >= original_sblock->page_count);
1353 page->physical_for_dev_replace =
1354 original_sblock->pagev[page_index]->
1355 physical_for_dev_replace;
1356 /* for missing devices, dev->bdev is NULL */
1357 page->mirror_num = mirror_index + 1;
1358 sblock->page_count++;
1359 page->page = alloc_page(GFP_NOFS);
1363 scrub_get_recover(recover);
1364 page->recover = recover;
1366 scrub_put_recover(fs_info, recover);
1375 static void scrub_bio_wait_endio(struct bio *bio)
1377 complete(bio->bi_private);
1380 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1382 struct scrub_page *page)
1384 DECLARE_COMPLETION_ONSTACK(done);
1388 bio->bi_iter.bi_sector = page->logical >> 9;
1389 bio->bi_private = &done;
1390 bio->bi_end_io = scrub_bio_wait_endio;
1392 mirror_num = page->sblock->pagev[0]->mirror_num;
1393 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1394 page->recover->map_length,
1399 wait_for_completion_io(&done);
1400 return blk_status_to_errno(bio->bi_status);
1403 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1404 struct scrub_block *sblock)
1406 struct scrub_page *first_page = sblock->pagev[0];
1410 /* All pages in sblock belong to the same stripe on the same device. */
1411 ASSERT(first_page->dev);
1412 if (!first_page->dev->bdev)
1415 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1416 bio_set_dev(bio, first_page->dev->bdev);
1418 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1419 struct scrub_page *page = sblock->pagev[page_num];
1421 WARN_ON(!page->page);
1422 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1425 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1432 scrub_recheck_block_checksum(sblock);
1436 for (page_num = 0; page_num < sblock->page_count; page_num++)
1437 sblock->pagev[page_num]->io_error = 1;
1439 sblock->no_io_error_seen = 0;
1443 * this function will check the on disk data for checksum errors, header
1444 * errors and read I/O errors. If any I/O errors happen, the exact pages
1445 * which are errored are marked as being bad. The goal is to enable scrub
1446 * to take those pages that are not errored from all the mirrors so that
1447 * the pages that are errored in the just handled mirror can be repaired.
1449 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1450 struct scrub_block *sblock,
1451 int retry_failed_mirror)
1455 sblock->no_io_error_seen = 1;
1457 /* short cut for raid56 */
1458 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1459 return scrub_recheck_block_on_raid56(fs_info, sblock);
1461 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1463 struct scrub_page *page = sblock->pagev[page_num];
1465 if (page->dev->bdev == NULL) {
1467 sblock->no_io_error_seen = 0;
1471 WARN_ON(!page->page);
1472 bio = btrfs_io_bio_alloc(1);
1473 bio_set_dev(bio, page->dev->bdev);
1475 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1476 bio->bi_iter.bi_sector = page->physical >> 9;
1477 bio->bi_opf = REQ_OP_READ;
1479 if (btrfsic_submit_bio_wait(bio)) {
1481 sblock->no_io_error_seen = 0;
1487 if (sblock->no_io_error_seen)
1488 scrub_recheck_block_checksum(sblock);
1491 static inline int scrub_check_fsid(u8 fsid[],
1492 struct scrub_page *spage)
1494 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1497 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1501 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1503 sblock->header_error = 0;
1504 sblock->checksum_error = 0;
1505 sblock->generation_error = 0;
1507 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1508 scrub_checksum_data(sblock);
1510 scrub_checksum_tree_block(sblock);
1513 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1514 struct scrub_block *sblock_good)
1519 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1522 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1532 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1533 struct scrub_block *sblock_good,
1534 int page_num, int force_write)
1536 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1537 struct scrub_page *page_good = sblock_good->pagev[page_num];
1538 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1540 BUG_ON(page_bad->page == NULL);
1541 BUG_ON(page_good->page == NULL);
1542 if (force_write || sblock_bad->header_error ||
1543 sblock_bad->checksum_error || page_bad->io_error) {
1547 if (!page_bad->dev->bdev) {
1548 btrfs_warn_rl(fs_info,
1549 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1553 bio = btrfs_io_bio_alloc(1);
1554 bio_set_dev(bio, page_bad->dev->bdev);
1555 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1556 bio->bi_opf = REQ_OP_WRITE;
1558 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1559 if (PAGE_SIZE != ret) {
1564 if (btrfsic_submit_bio_wait(bio)) {
1565 btrfs_dev_stat_inc_and_print(page_bad->dev,
1566 BTRFS_DEV_STAT_WRITE_ERRS);
1567 btrfs_dev_replace_stats_inc(
1568 &fs_info->dev_replace.num_write_errors);
1578 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1580 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1584 * This block is used for the check of the parity on the source device,
1585 * so the data needn't be written into the destination device.
1587 if (sblock->sparity)
1590 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1593 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1595 btrfs_dev_replace_stats_inc(
1596 &fs_info->dev_replace.num_write_errors);
1600 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1603 struct scrub_page *spage = sblock->pagev[page_num];
1605 BUG_ON(spage->page == NULL);
1606 if (spage->io_error) {
1607 void *mapped_buffer = kmap_atomic(spage->page);
1609 clear_page(mapped_buffer);
1610 flush_dcache_page(spage->page);
1611 kunmap_atomic(mapped_buffer);
1613 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1616 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1617 struct scrub_page *spage)
1619 struct scrub_bio *sbio;
1622 mutex_lock(&sctx->wr_lock);
1624 if (!sctx->wr_curr_bio) {
1625 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1627 if (!sctx->wr_curr_bio) {
1628 mutex_unlock(&sctx->wr_lock);
1631 sctx->wr_curr_bio->sctx = sctx;
1632 sctx->wr_curr_bio->page_count = 0;
1634 sbio = sctx->wr_curr_bio;
1635 if (sbio->page_count == 0) {
1638 sbio->physical = spage->physical_for_dev_replace;
1639 sbio->logical = spage->logical;
1640 sbio->dev = sctx->wr_tgtdev;
1643 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1647 bio->bi_private = sbio;
1648 bio->bi_end_io = scrub_wr_bio_end_io;
1649 bio_set_dev(bio, sbio->dev->bdev);
1650 bio->bi_iter.bi_sector = sbio->physical >> 9;
1651 bio->bi_opf = REQ_OP_WRITE;
1653 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1654 spage->physical_for_dev_replace ||
1655 sbio->logical + sbio->page_count * PAGE_SIZE !=
1657 scrub_wr_submit(sctx);
1661 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1662 if (ret != PAGE_SIZE) {
1663 if (sbio->page_count < 1) {
1666 mutex_unlock(&sctx->wr_lock);
1669 scrub_wr_submit(sctx);
1673 sbio->pagev[sbio->page_count] = spage;
1674 scrub_page_get(spage);
1676 if (sbio->page_count == sctx->pages_per_wr_bio)
1677 scrub_wr_submit(sctx);
1678 mutex_unlock(&sctx->wr_lock);
1683 static void scrub_wr_submit(struct scrub_ctx *sctx)
1685 struct scrub_bio *sbio;
1687 if (!sctx->wr_curr_bio)
1690 sbio = sctx->wr_curr_bio;
1691 sctx->wr_curr_bio = NULL;
1692 WARN_ON(!sbio->bio->bi_disk);
1693 scrub_pending_bio_inc(sctx);
1694 /* process all writes in a single worker thread. Then the block layer
1695 * orders the requests before sending them to the driver which
1696 * doubled the write performance on spinning disks when measured
1698 btrfsic_submit_bio(sbio->bio);
1701 static void scrub_wr_bio_end_io(struct bio *bio)
1703 struct scrub_bio *sbio = bio->bi_private;
1704 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1706 sbio->status = bio->bi_status;
1709 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1710 scrub_wr_bio_end_io_worker, NULL, NULL);
1711 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1714 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1716 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1717 struct scrub_ctx *sctx = sbio->sctx;
1720 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1722 struct btrfs_dev_replace *dev_replace =
1723 &sbio->sctx->fs_info->dev_replace;
1725 for (i = 0; i < sbio->page_count; i++) {
1726 struct scrub_page *spage = sbio->pagev[i];
1728 spage->io_error = 1;
1729 btrfs_dev_replace_stats_inc(&dev_replace->
1734 for (i = 0; i < sbio->page_count; i++)
1735 scrub_page_put(sbio->pagev[i]);
1739 scrub_pending_bio_dec(sctx);
1742 static int scrub_checksum(struct scrub_block *sblock)
1748 * No need to initialize these stats currently,
1749 * because this function only use return value
1750 * instead of these stats value.
1755 sblock->header_error = 0;
1756 sblock->generation_error = 0;
1757 sblock->checksum_error = 0;
1759 WARN_ON(sblock->page_count < 1);
1760 flags = sblock->pagev[0]->flags;
1762 if (flags & BTRFS_EXTENT_FLAG_DATA)
1763 ret = scrub_checksum_data(sblock);
1764 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1765 ret = scrub_checksum_tree_block(sblock);
1766 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1767 (void)scrub_checksum_super(sblock);
1771 scrub_handle_errored_block(sblock);
1776 static int scrub_checksum_data(struct scrub_block *sblock)
1778 struct scrub_ctx *sctx = sblock->sctx;
1779 u8 csum[BTRFS_CSUM_SIZE];
1787 BUG_ON(sblock->page_count < 1);
1788 if (!sblock->pagev[0]->have_csum)
1791 on_disk_csum = sblock->pagev[0]->csum;
1792 page = sblock->pagev[0]->page;
1793 buffer = kmap_atomic(page);
1795 len = sctx->fs_info->sectorsize;
1798 u64 l = min_t(u64, len, PAGE_SIZE);
1800 crc = btrfs_csum_data(buffer, crc, l);
1801 kunmap_atomic(buffer);
1806 BUG_ON(index >= sblock->page_count);
1807 BUG_ON(!sblock->pagev[index]->page);
1808 page = sblock->pagev[index]->page;
1809 buffer = kmap_atomic(page);
1812 btrfs_csum_final(crc, csum);
1813 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1814 sblock->checksum_error = 1;
1816 return sblock->checksum_error;
1819 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1821 struct scrub_ctx *sctx = sblock->sctx;
1822 struct btrfs_header *h;
1823 struct btrfs_fs_info *fs_info = sctx->fs_info;
1824 u8 calculated_csum[BTRFS_CSUM_SIZE];
1825 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1827 void *mapped_buffer;
1834 BUG_ON(sblock->page_count < 1);
1835 page = sblock->pagev[0]->page;
1836 mapped_buffer = kmap_atomic(page);
1837 h = (struct btrfs_header *)mapped_buffer;
1838 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1841 * we don't use the getter functions here, as we
1842 * a) don't have an extent buffer and
1843 * b) the page is already kmapped
1845 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1846 sblock->header_error = 1;
1848 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1849 sblock->header_error = 1;
1850 sblock->generation_error = 1;
1853 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1854 sblock->header_error = 1;
1856 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1858 sblock->header_error = 1;
1860 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1861 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1862 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1865 u64 l = min_t(u64, len, mapped_size);
1867 crc = btrfs_csum_data(p, crc, l);
1868 kunmap_atomic(mapped_buffer);
1873 BUG_ON(index >= sblock->page_count);
1874 BUG_ON(!sblock->pagev[index]->page);
1875 page = sblock->pagev[index]->page;
1876 mapped_buffer = kmap_atomic(page);
1877 mapped_size = PAGE_SIZE;
1881 btrfs_csum_final(crc, calculated_csum);
1882 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1883 sblock->checksum_error = 1;
1885 return sblock->header_error || sblock->checksum_error;
1888 static int scrub_checksum_super(struct scrub_block *sblock)
1890 struct btrfs_super_block *s;
1891 struct scrub_ctx *sctx = sblock->sctx;
1892 u8 calculated_csum[BTRFS_CSUM_SIZE];
1893 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1895 void *mapped_buffer;
1904 BUG_ON(sblock->page_count < 1);
1905 page = sblock->pagev[0]->page;
1906 mapped_buffer = kmap_atomic(page);
1907 s = (struct btrfs_super_block *)mapped_buffer;
1908 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1910 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1913 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1916 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1919 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1920 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1921 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1924 u64 l = min_t(u64, len, mapped_size);
1926 crc = btrfs_csum_data(p, crc, l);
1927 kunmap_atomic(mapped_buffer);
1932 BUG_ON(index >= sblock->page_count);
1933 BUG_ON(!sblock->pagev[index]->page);
1934 page = sblock->pagev[index]->page;
1935 mapped_buffer = kmap_atomic(page);
1936 mapped_size = PAGE_SIZE;
1940 btrfs_csum_final(crc, calculated_csum);
1941 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1944 if (fail_cor + fail_gen) {
1946 * if we find an error in a super block, we just report it.
1947 * They will get written with the next transaction commit
1950 spin_lock(&sctx->stat_lock);
1951 ++sctx->stat.super_errors;
1952 spin_unlock(&sctx->stat_lock);
1954 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1955 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1957 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1958 BTRFS_DEV_STAT_GENERATION_ERRS);
1961 return fail_cor + fail_gen;
1964 static void scrub_block_get(struct scrub_block *sblock)
1966 refcount_inc(&sblock->refs);
1969 static void scrub_block_put(struct scrub_block *sblock)
1971 if (refcount_dec_and_test(&sblock->refs)) {
1974 if (sblock->sparity)
1975 scrub_parity_put(sblock->sparity);
1977 for (i = 0; i < sblock->page_count; i++)
1978 scrub_page_put(sblock->pagev[i]);
1983 static void scrub_page_get(struct scrub_page *spage)
1985 atomic_inc(&spage->refs);
1988 static void scrub_page_put(struct scrub_page *spage)
1990 if (atomic_dec_and_test(&spage->refs)) {
1992 __free_page(spage->page);
1997 static void scrub_submit(struct scrub_ctx *sctx)
1999 struct scrub_bio *sbio;
2001 if (sctx->curr == -1)
2004 sbio = sctx->bios[sctx->curr];
2006 scrub_pending_bio_inc(sctx);
2007 btrfsic_submit_bio(sbio->bio);
2010 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2011 struct scrub_page *spage)
2013 struct scrub_block *sblock = spage->sblock;
2014 struct scrub_bio *sbio;
2019 * grab a fresh bio or wait for one to become available
2021 while (sctx->curr == -1) {
2022 spin_lock(&sctx->list_lock);
2023 sctx->curr = sctx->first_free;
2024 if (sctx->curr != -1) {
2025 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2026 sctx->bios[sctx->curr]->next_free = -1;
2027 sctx->bios[sctx->curr]->page_count = 0;
2028 spin_unlock(&sctx->list_lock);
2030 spin_unlock(&sctx->list_lock);
2031 wait_event(sctx->list_wait, sctx->first_free != -1);
2034 sbio = sctx->bios[sctx->curr];
2035 if (sbio->page_count == 0) {
2038 sbio->physical = spage->physical;
2039 sbio->logical = spage->logical;
2040 sbio->dev = spage->dev;
2043 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2047 bio->bi_private = sbio;
2048 bio->bi_end_io = scrub_bio_end_io;
2049 bio_set_dev(bio, sbio->dev->bdev);
2050 bio->bi_iter.bi_sector = sbio->physical >> 9;
2051 bio->bi_opf = REQ_OP_READ;
2053 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2055 sbio->logical + sbio->page_count * PAGE_SIZE !=
2057 sbio->dev != spage->dev) {
2062 sbio->pagev[sbio->page_count] = spage;
2063 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2064 if (ret != PAGE_SIZE) {
2065 if (sbio->page_count < 1) {
2074 scrub_block_get(sblock); /* one for the page added to the bio */
2075 atomic_inc(&sblock->outstanding_pages);
2077 if (sbio->page_count == sctx->pages_per_rd_bio)
2083 static void scrub_missing_raid56_end_io(struct bio *bio)
2085 struct scrub_block *sblock = bio->bi_private;
2086 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2089 sblock->no_io_error_seen = 0;
2093 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2096 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2098 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2099 struct scrub_ctx *sctx = sblock->sctx;
2100 struct btrfs_fs_info *fs_info = sctx->fs_info;
2102 struct btrfs_device *dev;
2104 logical = sblock->pagev[0]->logical;
2105 dev = sblock->pagev[0]->dev;
2107 if (sblock->no_io_error_seen)
2108 scrub_recheck_block_checksum(sblock);
2110 if (!sblock->no_io_error_seen) {
2111 spin_lock(&sctx->stat_lock);
2112 sctx->stat.read_errors++;
2113 spin_unlock(&sctx->stat_lock);
2114 btrfs_err_rl_in_rcu(fs_info,
2115 "IO error rebuilding logical %llu for dev %s",
2116 logical, rcu_str_deref(dev->name));
2117 } else if (sblock->header_error || sblock->checksum_error) {
2118 spin_lock(&sctx->stat_lock);
2119 sctx->stat.uncorrectable_errors++;
2120 spin_unlock(&sctx->stat_lock);
2121 btrfs_err_rl_in_rcu(fs_info,
2122 "failed to rebuild valid logical %llu for dev %s",
2123 logical, rcu_str_deref(dev->name));
2125 scrub_write_block_to_dev_replace(sblock);
2128 scrub_block_put(sblock);
2130 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2131 mutex_lock(&sctx->wr_lock);
2132 scrub_wr_submit(sctx);
2133 mutex_unlock(&sctx->wr_lock);
2136 scrub_pending_bio_dec(sctx);
2139 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2141 struct scrub_ctx *sctx = sblock->sctx;
2142 struct btrfs_fs_info *fs_info = sctx->fs_info;
2143 u64 length = sblock->page_count * PAGE_SIZE;
2144 u64 logical = sblock->pagev[0]->logical;
2145 struct btrfs_bio *bbio = NULL;
2147 struct btrfs_raid_bio *rbio;
2151 btrfs_bio_counter_inc_blocked(fs_info);
2152 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2154 if (ret || !bbio || !bbio->raid_map)
2157 if (WARN_ON(!sctx->is_dev_replace ||
2158 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2160 * We shouldn't be scrubbing a missing device. Even for dev
2161 * replace, we should only get here for RAID 5/6. We either
2162 * managed to mount something with no mirrors remaining or
2163 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2168 bio = btrfs_io_bio_alloc(0);
2169 bio->bi_iter.bi_sector = logical >> 9;
2170 bio->bi_private = sblock;
2171 bio->bi_end_io = scrub_missing_raid56_end_io;
2173 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2177 for (i = 0; i < sblock->page_count; i++) {
2178 struct scrub_page *spage = sblock->pagev[i];
2180 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2183 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2184 scrub_missing_raid56_worker, NULL, NULL);
2185 scrub_block_get(sblock);
2186 scrub_pending_bio_inc(sctx);
2187 raid56_submit_missing_rbio(rbio);
2193 btrfs_bio_counter_dec(fs_info);
2194 btrfs_put_bbio(bbio);
2195 spin_lock(&sctx->stat_lock);
2196 sctx->stat.malloc_errors++;
2197 spin_unlock(&sctx->stat_lock);
2200 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2201 u64 physical, struct btrfs_device *dev, u64 flags,
2202 u64 gen, int mirror_num, u8 *csum, int force,
2203 u64 physical_for_dev_replace)
2205 struct scrub_block *sblock;
2208 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2210 spin_lock(&sctx->stat_lock);
2211 sctx->stat.malloc_errors++;
2212 spin_unlock(&sctx->stat_lock);
2216 /* one ref inside this function, plus one for each page added to
2218 refcount_set(&sblock->refs, 1);
2219 sblock->sctx = sctx;
2220 sblock->no_io_error_seen = 1;
2222 for (index = 0; len > 0; index++) {
2223 struct scrub_page *spage;
2224 u64 l = min_t(u64, len, PAGE_SIZE);
2226 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2229 spin_lock(&sctx->stat_lock);
2230 sctx->stat.malloc_errors++;
2231 spin_unlock(&sctx->stat_lock);
2232 scrub_block_put(sblock);
2235 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2236 scrub_page_get(spage);
2237 sblock->pagev[index] = spage;
2238 spage->sblock = sblock;
2240 spage->flags = flags;
2241 spage->generation = gen;
2242 spage->logical = logical;
2243 spage->physical = physical;
2244 spage->physical_for_dev_replace = physical_for_dev_replace;
2245 spage->mirror_num = mirror_num;
2247 spage->have_csum = 1;
2248 memcpy(spage->csum, csum, sctx->csum_size);
2250 spage->have_csum = 0;
2252 sblock->page_count++;
2253 spage->page = alloc_page(GFP_KERNEL);
2259 physical_for_dev_replace += l;
2262 WARN_ON(sblock->page_count == 0);
2263 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2265 * This case should only be hit for RAID 5/6 device replace. See
2266 * the comment in scrub_missing_raid56_pages() for details.
2268 scrub_missing_raid56_pages(sblock);
2270 for (index = 0; index < sblock->page_count; index++) {
2271 struct scrub_page *spage = sblock->pagev[index];
2274 ret = scrub_add_page_to_rd_bio(sctx, spage);
2276 scrub_block_put(sblock);
2285 /* last one frees, either here or in bio completion for last page */
2286 scrub_block_put(sblock);
2290 static void scrub_bio_end_io(struct bio *bio)
2292 struct scrub_bio *sbio = bio->bi_private;
2293 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2295 sbio->status = bio->bi_status;
2298 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2301 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2303 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2304 struct scrub_ctx *sctx = sbio->sctx;
2307 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2309 for (i = 0; i < sbio->page_count; i++) {
2310 struct scrub_page *spage = sbio->pagev[i];
2312 spage->io_error = 1;
2313 spage->sblock->no_io_error_seen = 0;
2317 /* now complete the scrub_block items that have all pages completed */
2318 for (i = 0; i < sbio->page_count; i++) {
2319 struct scrub_page *spage = sbio->pagev[i];
2320 struct scrub_block *sblock = spage->sblock;
2322 if (atomic_dec_and_test(&sblock->outstanding_pages))
2323 scrub_block_complete(sblock);
2324 scrub_block_put(sblock);
2329 spin_lock(&sctx->list_lock);
2330 sbio->next_free = sctx->first_free;
2331 sctx->first_free = sbio->index;
2332 spin_unlock(&sctx->list_lock);
2334 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2335 mutex_lock(&sctx->wr_lock);
2336 scrub_wr_submit(sctx);
2337 mutex_unlock(&sctx->wr_lock);
2340 scrub_pending_bio_dec(sctx);
2343 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2344 unsigned long *bitmap,
2350 int sectorsize = sparity->sctx->fs_info->sectorsize;
2352 if (len >= sparity->stripe_len) {
2353 bitmap_set(bitmap, 0, sparity->nsectors);
2357 start -= sparity->logic_start;
2358 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2359 offset = div_u64(offset, sectorsize);
2360 nsectors64 = div_u64(len, sectorsize);
2362 ASSERT(nsectors64 < UINT_MAX);
2363 nsectors = (u32)nsectors64;
2365 if (offset + nsectors <= sparity->nsectors) {
2366 bitmap_set(bitmap, offset, nsectors);
2370 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2371 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2374 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2377 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2380 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2383 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2386 static void scrub_block_complete(struct scrub_block *sblock)
2390 if (!sblock->no_io_error_seen) {
2392 scrub_handle_errored_block(sblock);
2395 * if has checksum error, write via repair mechanism in
2396 * dev replace case, otherwise write here in dev replace
2399 corrupted = scrub_checksum(sblock);
2400 if (!corrupted && sblock->sctx->is_dev_replace)
2401 scrub_write_block_to_dev_replace(sblock);
2404 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2405 u64 start = sblock->pagev[0]->logical;
2406 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2409 scrub_parity_mark_sectors_error(sblock->sparity,
2410 start, end - start);
2414 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2416 struct btrfs_ordered_sum *sum = NULL;
2417 unsigned long index;
2418 unsigned long num_sectors;
2420 while (!list_empty(&sctx->csum_list)) {
2421 sum = list_first_entry(&sctx->csum_list,
2422 struct btrfs_ordered_sum, list);
2423 if (sum->bytenr > logical)
2425 if (sum->bytenr + sum->len > logical)
2428 ++sctx->stat.csum_discards;
2429 list_del(&sum->list);
2436 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2437 ASSERT(index < UINT_MAX);
2439 num_sectors = sum->len / sctx->fs_info->sectorsize;
2440 memcpy(csum, sum->sums + index, sctx->csum_size);
2441 if (index == num_sectors - 1) {
2442 list_del(&sum->list);
2448 /* scrub extent tries to collect up to 64 kB for each bio */
2449 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2450 u64 logical, u64 len,
2451 u64 physical, struct btrfs_device *dev, u64 flags,
2452 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2455 u8 csum[BTRFS_CSUM_SIZE];
2458 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2459 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2460 blocksize = map->stripe_len;
2462 blocksize = sctx->fs_info->sectorsize;
2463 spin_lock(&sctx->stat_lock);
2464 sctx->stat.data_extents_scrubbed++;
2465 sctx->stat.data_bytes_scrubbed += len;
2466 spin_unlock(&sctx->stat_lock);
2467 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2468 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2469 blocksize = map->stripe_len;
2471 blocksize = sctx->fs_info->nodesize;
2472 spin_lock(&sctx->stat_lock);
2473 sctx->stat.tree_extents_scrubbed++;
2474 sctx->stat.tree_bytes_scrubbed += len;
2475 spin_unlock(&sctx->stat_lock);
2477 blocksize = sctx->fs_info->sectorsize;
2482 u64 l = min_t(u64, len, blocksize);
2485 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2486 /* push csums to sbio */
2487 have_csum = scrub_find_csum(sctx, logical, csum);
2489 ++sctx->stat.no_csum;
2491 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2492 mirror_num, have_csum ? csum : NULL, 0,
2493 physical_for_dev_replace);
2499 physical_for_dev_replace += l;
2504 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2505 u64 logical, u64 len,
2506 u64 physical, struct btrfs_device *dev,
2507 u64 flags, u64 gen, int mirror_num, u8 *csum)
2509 struct scrub_ctx *sctx = sparity->sctx;
2510 struct scrub_block *sblock;
2513 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2515 spin_lock(&sctx->stat_lock);
2516 sctx->stat.malloc_errors++;
2517 spin_unlock(&sctx->stat_lock);
2521 /* one ref inside this function, plus one for each page added to
2523 refcount_set(&sblock->refs, 1);
2524 sblock->sctx = sctx;
2525 sblock->no_io_error_seen = 1;
2526 sblock->sparity = sparity;
2527 scrub_parity_get(sparity);
2529 for (index = 0; len > 0; index++) {
2530 struct scrub_page *spage;
2531 u64 l = min_t(u64, len, PAGE_SIZE);
2533 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2536 spin_lock(&sctx->stat_lock);
2537 sctx->stat.malloc_errors++;
2538 spin_unlock(&sctx->stat_lock);
2539 scrub_block_put(sblock);
2542 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2543 /* For scrub block */
2544 scrub_page_get(spage);
2545 sblock->pagev[index] = spage;
2546 /* For scrub parity */
2547 scrub_page_get(spage);
2548 list_add_tail(&spage->list, &sparity->spages);
2549 spage->sblock = sblock;
2551 spage->flags = flags;
2552 spage->generation = gen;
2553 spage->logical = logical;
2554 spage->physical = physical;
2555 spage->mirror_num = mirror_num;
2557 spage->have_csum = 1;
2558 memcpy(spage->csum, csum, sctx->csum_size);
2560 spage->have_csum = 0;
2562 sblock->page_count++;
2563 spage->page = alloc_page(GFP_KERNEL);
2571 WARN_ON(sblock->page_count == 0);
2572 for (index = 0; index < sblock->page_count; index++) {
2573 struct scrub_page *spage = sblock->pagev[index];
2576 ret = scrub_add_page_to_rd_bio(sctx, spage);
2578 scrub_block_put(sblock);
2583 /* last one frees, either here or in bio completion for last page */
2584 scrub_block_put(sblock);
2588 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2589 u64 logical, u64 len,
2590 u64 physical, struct btrfs_device *dev,
2591 u64 flags, u64 gen, int mirror_num)
2593 struct scrub_ctx *sctx = sparity->sctx;
2595 u8 csum[BTRFS_CSUM_SIZE];
2598 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2599 scrub_parity_mark_sectors_error(sparity, logical, len);
2603 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2604 blocksize = sparity->stripe_len;
2605 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2606 blocksize = sparity->stripe_len;
2608 blocksize = sctx->fs_info->sectorsize;
2613 u64 l = min_t(u64, len, blocksize);
2616 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2617 /* push csums to sbio */
2618 have_csum = scrub_find_csum(sctx, logical, csum);
2622 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2623 flags, gen, mirror_num,
2624 have_csum ? csum : NULL);
2636 * Given a physical address, this will calculate it's
2637 * logical offset. if this is a parity stripe, it will return
2638 * the most left data stripe's logical offset.
2640 * return 0 if it is a data stripe, 1 means parity stripe.
2642 static int get_raid56_logic_offset(u64 physical, int num,
2643 struct map_lookup *map, u64 *offset,
2653 last_offset = (physical - map->stripes[num].physical) *
2654 nr_data_stripes(map);
2656 *stripe_start = last_offset;
2658 *offset = last_offset;
2659 for (i = 0; i < nr_data_stripes(map); i++) {
2660 *offset = last_offset + i * map->stripe_len;
2662 stripe_nr = div64_u64(*offset, map->stripe_len);
2663 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2665 /* Work out the disk rotation on this stripe-set */
2666 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2667 /* calculate which stripe this data locates */
2669 stripe_index = rot % map->num_stripes;
2670 if (stripe_index == num)
2672 if (stripe_index < num)
2675 *offset = last_offset + j * map->stripe_len;
2679 static void scrub_free_parity(struct scrub_parity *sparity)
2681 struct scrub_ctx *sctx = sparity->sctx;
2682 struct scrub_page *curr, *next;
2685 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2687 spin_lock(&sctx->stat_lock);
2688 sctx->stat.read_errors += nbits;
2689 sctx->stat.uncorrectable_errors += nbits;
2690 spin_unlock(&sctx->stat_lock);
2693 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2694 list_del_init(&curr->list);
2695 scrub_page_put(curr);
2701 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2703 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2705 struct scrub_ctx *sctx = sparity->sctx;
2707 scrub_free_parity(sparity);
2708 scrub_pending_bio_dec(sctx);
2711 static void scrub_parity_bio_endio(struct bio *bio)
2713 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2714 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2717 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2722 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2723 scrub_parity_bio_endio_worker, NULL, NULL);
2724 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2727 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2729 struct scrub_ctx *sctx = sparity->sctx;
2730 struct btrfs_fs_info *fs_info = sctx->fs_info;
2732 struct btrfs_raid_bio *rbio;
2733 struct btrfs_bio *bbio = NULL;
2737 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2741 length = sparity->logic_end - sparity->logic_start;
2743 btrfs_bio_counter_inc_blocked(fs_info);
2744 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2746 if (ret || !bbio || !bbio->raid_map)
2749 bio = btrfs_io_bio_alloc(0);
2750 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2751 bio->bi_private = sparity;
2752 bio->bi_end_io = scrub_parity_bio_endio;
2754 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2755 length, sparity->scrub_dev,
2761 scrub_pending_bio_inc(sctx);
2762 raid56_parity_submit_scrub_rbio(rbio);
2768 btrfs_bio_counter_dec(fs_info);
2769 btrfs_put_bbio(bbio);
2770 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2772 spin_lock(&sctx->stat_lock);
2773 sctx->stat.malloc_errors++;
2774 spin_unlock(&sctx->stat_lock);
2776 scrub_free_parity(sparity);
2779 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2781 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2784 static void scrub_parity_get(struct scrub_parity *sparity)
2786 refcount_inc(&sparity->refs);
2789 static void scrub_parity_put(struct scrub_parity *sparity)
2791 if (!refcount_dec_and_test(&sparity->refs))
2794 scrub_parity_check_and_repair(sparity);
2797 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2798 struct map_lookup *map,
2799 struct btrfs_device *sdev,
2800 struct btrfs_path *path,
2804 struct btrfs_fs_info *fs_info = sctx->fs_info;
2805 struct btrfs_root *root = fs_info->extent_root;
2806 struct btrfs_root *csum_root = fs_info->csum_root;
2807 struct btrfs_extent_item *extent;
2808 struct btrfs_bio *bbio = NULL;
2812 struct extent_buffer *l;
2813 struct btrfs_key key;
2816 u64 extent_physical;
2819 struct btrfs_device *extent_dev;
2820 struct scrub_parity *sparity;
2823 int extent_mirror_num;
2826 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2827 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2828 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2831 spin_lock(&sctx->stat_lock);
2832 sctx->stat.malloc_errors++;
2833 spin_unlock(&sctx->stat_lock);
2837 sparity->stripe_len = map->stripe_len;
2838 sparity->nsectors = nsectors;
2839 sparity->sctx = sctx;
2840 sparity->scrub_dev = sdev;
2841 sparity->logic_start = logic_start;
2842 sparity->logic_end = logic_end;
2843 refcount_set(&sparity->refs, 1);
2844 INIT_LIST_HEAD(&sparity->spages);
2845 sparity->dbitmap = sparity->bitmap;
2846 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2849 while (logic_start < logic_end) {
2850 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2851 key.type = BTRFS_METADATA_ITEM_KEY;
2853 key.type = BTRFS_EXTENT_ITEM_KEY;
2854 key.objectid = logic_start;
2855 key.offset = (u64)-1;
2857 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2862 ret = btrfs_previous_extent_item(root, path, 0);
2866 btrfs_release_path(path);
2867 ret = btrfs_search_slot(NULL, root, &key,
2879 slot = path->slots[0];
2880 if (slot >= btrfs_header_nritems(l)) {
2881 ret = btrfs_next_leaf(root, path);
2890 btrfs_item_key_to_cpu(l, &key, slot);
2892 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2893 key.type != BTRFS_METADATA_ITEM_KEY)
2896 if (key.type == BTRFS_METADATA_ITEM_KEY)
2897 bytes = fs_info->nodesize;
2901 if (key.objectid + bytes <= logic_start)
2904 if (key.objectid >= logic_end) {
2909 while (key.objectid >= logic_start + map->stripe_len)
2910 logic_start += map->stripe_len;
2912 extent = btrfs_item_ptr(l, slot,
2913 struct btrfs_extent_item);
2914 flags = btrfs_extent_flags(l, extent);
2915 generation = btrfs_extent_generation(l, extent);
2917 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2918 (key.objectid < logic_start ||
2919 key.objectid + bytes >
2920 logic_start + map->stripe_len)) {
2922 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2923 key.objectid, logic_start);
2924 spin_lock(&sctx->stat_lock);
2925 sctx->stat.uncorrectable_errors++;
2926 spin_unlock(&sctx->stat_lock);
2930 extent_logical = key.objectid;
2933 if (extent_logical < logic_start) {
2934 extent_len -= logic_start - extent_logical;
2935 extent_logical = logic_start;
2938 if (extent_logical + extent_len >
2939 logic_start + map->stripe_len)
2940 extent_len = logic_start + map->stripe_len -
2943 scrub_parity_mark_sectors_data(sparity, extent_logical,
2946 mapped_length = extent_len;
2948 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2949 extent_logical, &mapped_length, &bbio,
2952 if (!bbio || mapped_length < extent_len)
2956 btrfs_put_bbio(bbio);
2959 extent_physical = bbio->stripes[0].physical;
2960 extent_mirror_num = bbio->mirror_num;
2961 extent_dev = bbio->stripes[0].dev;
2962 btrfs_put_bbio(bbio);
2964 ret = btrfs_lookup_csums_range(csum_root,
2966 extent_logical + extent_len - 1,
2967 &sctx->csum_list, 1);
2971 ret = scrub_extent_for_parity(sparity, extent_logical,
2978 scrub_free_csums(sctx);
2983 if (extent_logical + extent_len <
2984 key.objectid + bytes) {
2985 logic_start += map->stripe_len;
2987 if (logic_start >= logic_end) {
2992 if (logic_start < key.objectid + bytes) {
3001 btrfs_release_path(path);
3006 logic_start += map->stripe_len;
3010 scrub_parity_mark_sectors_error(sparity, logic_start,
3011 logic_end - logic_start);
3012 scrub_parity_put(sparity);
3014 mutex_lock(&sctx->wr_lock);
3015 scrub_wr_submit(sctx);
3016 mutex_unlock(&sctx->wr_lock);
3018 btrfs_release_path(path);
3019 return ret < 0 ? ret : 0;
3022 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3023 struct map_lookup *map,
3024 struct btrfs_device *scrub_dev,
3025 int num, u64 base, u64 length)
3027 struct btrfs_path *path, *ppath;
3028 struct btrfs_fs_info *fs_info = sctx->fs_info;
3029 struct btrfs_root *root = fs_info->extent_root;
3030 struct btrfs_root *csum_root = fs_info->csum_root;
3031 struct btrfs_extent_item *extent;
3032 struct blk_plug plug;
3037 struct extent_buffer *l;
3044 struct reada_control *reada1;
3045 struct reada_control *reada2;
3046 struct btrfs_key key;
3047 struct btrfs_key key_end;
3048 u64 increment = map->stripe_len;
3051 u64 extent_physical;
3055 struct btrfs_device *extent_dev;
3056 int extent_mirror_num;
3059 physical = map->stripes[num].physical;
3061 nstripes = div64_u64(length, map->stripe_len);
3062 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3063 offset = map->stripe_len * num;
3064 increment = map->stripe_len * map->num_stripes;
3066 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3067 int factor = map->num_stripes / map->sub_stripes;
3068 offset = map->stripe_len * (num / map->sub_stripes);
3069 increment = map->stripe_len * factor;
3070 mirror_num = num % map->sub_stripes + 1;
3071 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3072 increment = map->stripe_len;
3073 mirror_num = num % map->num_stripes + 1;
3074 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3075 increment = map->stripe_len;
3076 mirror_num = num % map->num_stripes + 1;
3077 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3078 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3079 increment = map->stripe_len * nr_data_stripes(map);
3082 increment = map->stripe_len;
3086 path = btrfs_alloc_path();
3090 ppath = btrfs_alloc_path();
3092 btrfs_free_path(path);
3097 * work on commit root. The related disk blocks are static as
3098 * long as COW is applied. This means, it is save to rewrite
3099 * them to repair disk errors without any race conditions
3101 path->search_commit_root = 1;
3102 path->skip_locking = 1;
3104 ppath->search_commit_root = 1;
3105 ppath->skip_locking = 1;
3107 * trigger the readahead for extent tree csum tree and wait for
3108 * completion. During readahead, the scrub is officially paused
3109 * to not hold off transaction commits
3111 logical = base + offset;
3112 physical_end = physical + nstripes * map->stripe_len;
3113 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3114 get_raid56_logic_offset(physical_end, num,
3115 map, &logic_end, NULL);
3118 logic_end = logical + increment * nstripes;
3120 wait_event(sctx->list_wait,
3121 atomic_read(&sctx->bios_in_flight) == 0);
3122 scrub_blocked_if_needed(fs_info);
3124 /* FIXME it might be better to start readahead at commit root */
3125 key.objectid = logical;
3126 key.type = BTRFS_EXTENT_ITEM_KEY;
3127 key.offset = (u64)0;
3128 key_end.objectid = logic_end;
3129 key_end.type = BTRFS_METADATA_ITEM_KEY;
3130 key_end.offset = (u64)-1;
3131 reada1 = btrfs_reada_add(root, &key, &key_end);
3133 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3134 key.type = BTRFS_EXTENT_CSUM_KEY;
3135 key.offset = logical;
3136 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3137 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3138 key_end.offset = logic_end;
3139 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3141 if (!IS_ERR(reada1))
3142 btrfs_reada_wait(reada1);
3143 if (!IS_ERR(reada2))
3144 btrfs_reada_wait(reada2);
3148 * collect all data csums for the stripe to avoid seeking during
3149 * the scrub. This might currently (crc32) end up to be about 1MB
3151 blk_start_plug(&plug);
3154 * now find all extents for each stripe and scrub them
3157 while (physical < physical_end) {
3161 if (atomic_read(&fs_info->scrub_cancel_req) ||
3162 atomic_read(&sctx->cancel_req)) {
3167 * check to see if we have to pause
3169 if (atomic_read(&fs_info->scrub_pause_req)) {
3170 /* push queued extents */
3171 sctx->flush_all_writes = true;
3173 mutex_lock(&sctx->wr_lock);
3174 scrub_wr_submit(sctx);
3175 mutex_unlock(&sctx->wr_lock);
3176 wait_event(sctx->list_wait,
3177 atomic_read(&sctx->bios_in_flight) == 0);
3178 sctx->flush_all_writes = false;
3179 scrub_blocked_if_needed(fs_info);
3182 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3183 ret = get_raid56_logic_offset(physical, num, map,
3188 /* it is parity strip */
3189 stripe_logical += base;
3190 stripe_end = stripe_logical + increment;
3191 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3192 ppath, stripe_logical,
3200 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3201 key.type = BTRFS_METADATA_ITEM_KEY;
3203 key.type = BTRFS_EXTENT_ITEM_KEY;
3204 key.objectid = logical;
3205 key.offset = (u64)-1;
3207 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3212 ret = btrfs_previous_extent_item(root, path, 0);
3216 /* there's no smaller item, so stick with the
3218 btrfs_release_path(path);
3219 ret = btrfs_search_slot(NULL, root, &key,
3231 slot = path->slots[0];
3232 if (slot >= btrfs_header_nritems(l)) {
3233 ret = btrfs_next_leaf(root, path);
3242 btrfs_item_key_to_cpu(l, &key, slot);
3244 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3245 key.type != BTRFS_METADATA_ITEM_KEY)
3248 if (key.type == BTRFS_METADATA_ITEM_KEY)
3249 bytes = fs_info->nodesize;
3253 if (key.objectid + bytes <= logical)
3256 if (key.objectid >= logical + map->stripe_len) {
3257 /* out of this device extent */
3258 if (key.objectid >= logic_end)
3263 extent = btrfs_item_ptr(l, slot,
3264 struct btrfs_extent_item);
3265 flags = btrfs_extent_flags(l, extent);
3266 generation = btrfs_extent_generation(l, extent);
3268 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3269 (key.objectid < logical ||
3270 key.objectid + bytes >
3271 logical + map->stripe_len)) {
3273 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3274 key.objectid, logical);
3275 spin_lock(&sctx->stat_lock);
3276 sctx->stat.uncorrectable_errors++;
3277 spin_unlock(&sctx->stat_lock);
3282 extent_logical = key.objectid;
3286 * trim extent to this stripe
3288 if (extent_logical < logical) {
3289 extent_len -= logical - extent_logical;
3290 extent_logical = logical;
3292 if (extent_logical + extent_len >
3293 logical + map->stripe_len) {
3294 extent_len = logical + map->stripe_len -
3298 extent_physical = extent_logical - logical + physical;
3299 extent_dev = scrub_dev;
3300 extent_mirror_num = mirror_num;
3301 if (sctx->is_dev_replace)
3302 scrub_remap_extent(fs_info, extent_logical,
3303 extent_len, &extent_physical,
3305 &extent_mirror_num);
3307 ret = btrfs_lookup_csums_range(csum_root,
3311 &sctx->csum_list, 1);
3315 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3316 extent_physical, extent_dev, flags,
3317 generation, extent_mirror_num,
3318 extent_logical - logical + physical);
3320 scrub_free_csums(sctx);
3325 if (extent_logical + extent_len <
3326 key.objectid + bytes) {
3327 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3329 * loop until we find next data stripe
3330 * or we have finished all stripes.
3333 physical += map->stripe_len;
3334 ret = get_raid56_logic_offset(physical,
3339 if (ret && physical < physical_end) {
3340 stripe_logical += base;
3341 stripe_end = stripe_logical +
3343 ret = scrub_raid56_parity(sctx,
3344 map, scrub_dev, ppath,
3352 physical += map->stripe_len;
3353 logical += increment;
3355 if (logical < key.objectid + bytes) {
3360 if (physical >= physical_end) {
3368 btrfs_release_path(path);
3370 logical += increment;
3371 physical += map->stripe_len;
3372 spin_lock(&sctx->stat_lock);
3374 sctx->stat.last_physical = map->stripes[num].physical +
3377 sctx->stat.last_physical = physical;
3378 spin_unlock(&sctx->stat_lock);
3383 /* push queued extents */
3385 mutex_lock(&sctx->wr_lock);
3386 scrub_wr_submit(sctx);
3387 mutex_unlock(&sctx->wr_lock);
3389 blk_finish_plug(&plug);
3390 btrfs_free_path(path);
3391 btrfs_free_path(ppath);
3392 return ret < 0 ? ret : 0;
3395 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3396 struct btrfs_device *scrub_dev,
3397 u64 chunk_offset, u64 length,
3399 struct btrfs_block_group_cache *cache)
3401 struct btrfs_fs_info *fs_info = sctx->fs_info;
3402 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3403 struct map_lookup *map;
3404 struct extent_map *em;
3408 read_lock(&map_tree->map_tree.lock);
3409 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3410 read_unlock(&map_tree->map_tree.lock);
3414 * Might have been an unused block group deleted by the cleaner
3415 * kthread or relocation.
3417 spin_lock(&cache->lock);
3418 if (!cache->removed)
3420 spin_unlock(&cache->lock);
3425 map = em->map_lookup;
3426 if (em->start != chunk_offset)
3429 if (em->len < length)
3432 for (i = 0; i < map->num_stripes; ++i) {
3433 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3434 map->stripes[i].physical == dev_offset) {
3435 ret = scrub_stripe(sctx, map, scrub_dev, i,
3436 chunk_offset, length);
3442 free_extent_map(em);
3447 static noinline_for_stack
3448 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3449 struct btrfs_device *scrub_dev, u64 start, u64 end)
3451 struct btrfs_dev_extent *dev_extent = NULL;
3452 struct btrfs_path *path;
3453 struct btrfs_fs_info *fs_info = sctx->fs_info;
3454 struct btrfs_root *root = fs_info->dev_root;
3460 struct extent_buffer *l;
3461 struct btrfs_key key;
3462 struct btrfs_key found_key;
3463 struct btrfs_block_group_cache *cache;
3464 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3466 path = btrfs_alloc_path();
3470 path->reada = READA_FORWARD;
3471 path->search_commit_root = 1;
3472 path->skip_locking = 1;
3474 key.objectid = scrub_dev->devid;
3476 key.type = BTRFS_DEV_EXTENT_KEY;
3479 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3483 if (path->slots[0] >=
3484 btrfs_header_nritems(path->nodes[0])) {
3485 ret = btrfs_next_leaf(root, path);
3498 slot = path->slots[0];
3500 btrfs_item_key_to_cpu(l, &found_key, slot);
3502 if (found_key.objectid != scrub_dev->devid)
3505 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3508 if (found_key.offset >= end)
3511 if (found_key.offset < key.offset)
3514 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3515 length = btrfs_dev_extent_length(l, dev_extent);
3517 if (found_key.offset + length <= start)
3520 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3523 * get a reference on the corresponding block group to prevent
3524 * the chunk from going away while we scrub it
3526 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3528 /* some chunks are removed but not committed to disk yet,
3529 * continue scrubbing */
3534 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3535 * to avoid deadlock caused by:
3536 * btrfs_inc_block_group_ro()
3537 * -> btrfs_wait_for_commit()
3538 * -> btrfs_commit_transaction()
3539 * -> btrfs_scrub_pause()
3541 scrub_pause_on(fs_info);
3542 ret = btrfs_inc_block_group_ro(cache);
3543 if (!ret && sctx->is_dev_replace) {
3545 * If we are doing a device replace wait for any tasks
3546 * that started dellaloc right before we set the block
3547 * group to RO mode, as they might have just allocated
3548 * an extent from it or decided they could do a nocow
3549 * write. And if any such tasks did that, wait for their
3550 * ordered extents to complete and then commit the
3551 * current transaction, so that we can later see the new
3552 * extent items in the extent tree - the ordered extents
3553 * create delayed data references (for cow writes) when
3554 * they complete, which will be run and insert the
3555 * corresponding extent items into the extent tree when
3556 * we commit the transaction they used when running
3557 * inode.c:btrfs_finish_ordered_io(). We later use
3558 * the commit root of the extent tree to find extents
3559 * to copy from the srcdev into the tgtdev, and we don't
3560 * want to miss any new extents.
3562 btrfs_wait_block_group_reservations(cache);
3563 btrfs_wait_nocow_writers(cache);
3564 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3565 cache->key.objectid,
3568 struct btrfs_trans_handle *trans;
3570 trans = btrfs_join_transaction(root);
3572 ret = PTR_ERR(trans);
3574 ret = btrfs_commit_transaction(trans);
3576 scrub_pause_off(fs_info);
3577 btrfs_put_block_group(cache);
3582 scrub_pause_off(fs_info);
3586 } else if (ret == -ENOSPC) {
3588 * btrfs_inc_block_group_ro return -ENOSPC when it
3589 * failed in creating new chunk for metadata.
3590 * It is not a problem for scrub/replace, because
3591 * metadata are always cowed, and our scrub paused
3592 * commit_transactions.
3597 "failed setting block group ro: %d", ret);
3598 btrfs_put_block_group(cache);
3602 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3603 dev_replace->cursor_right = found_key.offset + length;
3604 dev_replace->cursor_left = found_key.offset;
3605 dev_replace->item_needs_writeback = 1;
3606 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3607 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3608 found_key.offset, cache);
3611 * flush, submit all pending read and write bios, afterwards
3613 * Note that in the dev replace case, a read request causes
3614 * write requests that are submitted in the read completion
3615 * worker. Therefore in the current situation, it is required
3616 * that all write requests are flushed, so that all read and
3617 * write requests are really completed when bios_in_flight
3620 sctx->flush_all_writes = true;
3622 mutex_lock(&sctx->wr_lock);
3623 scrub_wr_submit(sctx);
3624 mutex_unlock(&sctx->wr_lock);
3626 wait_event(sctx->list_wait,
3627 atomic_read(&sctx->bios_in_flight) == 0);
3629 scrub_pause_on(fs_info);
3632 * must be called before we decrease @scrub_paused.
3633 * make sure we don't block transaction commit while
3634 * we are waiting pending workers finished.
3636 wait_event(sctx->list_wait,
3637 atomic_read(&sctx->workers_pending) == 0);
3638 sctx->flush_all_writes = false;
3640 scrub_pause_off(fs_info);
3642 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3643 dev_replace->cursor_left = dev_replace->cursor_right;
3644 dev_replace->item_needs_writeback = 1;
3645 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3648 btrfs_dec_block_group_ro(cache);
3651 * We might have prevented the cleaner kthread from deleting
3652 * this block group if it was already unused because we raced
3653 * and set it to RO mode first. So add it back to the unused
3654 * list, otherwise it might not ever be deleted unless a manual
3655 * balance is triggered or it becomes used and unused again.
3657 spin_lock(&cache->lock);
3658 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3659 btrfs_block_group_used(&cache->item) == 0) {
3660 spin_unlock(&cache->lock);
3661 btrfs_mark_bg_unused(cache);
3663 spin_unlock(&cache->lock);
3666 btrfs_put_block_group(cache);
3669 if (sctx->is_dev_replace &&
3670 atomic64_read(&dev_replace->num_write_errors) > 0) {
3674 if (sctx->stat.malloc_errors > 0) {
3679 key.offset = found_key.offset + length;
3680 btrfs_release_path(path);
3683 btrfs_free_path(path);
3688 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3689 struct btrfs_device *scrub_dev)
3695 struct btrfs_fs_info *fs_info = sctx->fs_info;
3697 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3700 /* Seed devices of a new filesystem has their own generation. */
3701 if (scrub_dev->fs_devices != fs_info->fs_devices)
3702 gen = scrub_dev->generation;
3704 gen = fs_info->last_trans_committed;
3706 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3707 bytenr = btrfs_sb_offset(i);
3708 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3709 scrub_dev->commit_total_bytes)
3712 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3713 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3718 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3724 * get a reference count on fs_info->scrub_workers. start worker if necessary
3726 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3729 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3730 int max_active = fs_info->thread_pool_size;
3732 if (fs_info->scrub_workers_refcnt == 0) {
3733 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3734 flags, is_dev_replace ? 1 : max_active, 4);
3735 if (!fs_info->scrub_workers)
3736 goto fail_scrub_workers;
3738 fs_info->scrub_wr_completion_workers =
3739 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3741 if (!fs_info->scrub_wr_completion_workers)
3742 goto fail_scrub_wr_completion_workers;
3744 fs_info->scrub_parity_workers =
3745 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3747 if (!fs_info->scrub_parity_workers)
3748 goto fail_scrub_parity_workers;
3750 ++fs_info->scrub_workers_refcnt;
3753 fail_scrub_parity_workers:
3754 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3755 fail_scrub_wr_completion_workers:
3756 btrfs_destroy_workqueue(fs_info->scrub_workers);
3761 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3763 if (--fs_info->scrub_workers_refcnt == 0) {
3764 btrfs_destroy_workqueue(fs_info->scrub_workers);
3765 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3766 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3768 WARN_ON(fs_info->scrub_workers_refcnt < 0);
3771 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3772 u64 end, struct btrfs_scrub_progress *progress,
3773 int readonly, int is_dev_replace)
3775 struct scrub_ctx *sctx;
3777 struct btrfs_device *dev;
3779 if (btrfs_fs_closing(fs_info))
3782 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3784 * in this case scrub is unable to calculate the checksum
3785 * the way scrub is implemented. Do not handle this
3786 * situation at all because it won't ever happen.
3789 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3795 if (fs_info->sectorsize != PAGE_SIZE) {
3796 /* not supported for data w/o checksums */
3797 btrfs_err_rl(fs_info,
3798 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3799 fs_info->sectorsize, PAGE_SIZE);
3803 if (fs_info->nodesize >
3804 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3805 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3807 * would exhaust the array bounds of pagev member in
3808 * struct scrub_block
3811 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3813 SCRUB_MAX_PAGES_PER_BLOCK,
3814 fs_info->sectorsize,
3815 SCRUB_MAX_PAGES_PER_BLOCK);
3820 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3821 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3822 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3824 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3828 if (!is_dev_replace && !readonly &&
3829 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3830 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3831 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3832 rcu_str_deref(dev->name));
3836 mutex_lock(&fs_info->scrub_lock);
3837 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3838 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3839 mutex_unlock(&fs_info->scrub_lock);
3840 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3844 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
3845 if (dev->scrub_ctx ||
3847 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3848 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
3849 mutex_unlock(&fs_info->scrub_lock);
3850 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3851 return -EINPROGRESS;
3853 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
3855 ret = scrub_workers_get(fs_info, is_dev_replace);
3857 mutex_unlock(&fs_info->scrub_lock);
3858 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3862 sctx = scrub_setup_ctx(dev, is_dev_replace);
3864 mutex_unlock(&fs_info->scrub_lock);
3865 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3866 scrub_workers_put(fs_info);
3867 return PTR_ERR(sctx);
3869 sctx->readonly = readonly;
3870 dev->scrub_ctx = sctx;
3871 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3874 * checking @scrub_pause_req here, we can avoid
3875 * race between committing transaction and scrubbing.
3877 __scrub_blocked_if_needed(fs_info);
3878 atomic_inc(&fs_info->scrubs_running);
3879 mutex_unlock(&fs_info->scrub_lock);
3881 if (!is_dev_replace) {
3883 * by holding device list mutex, we can
3884 * kick off writing super in log tree sync.
3886 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3887 ret = scrub_supers(sctx, dev);
3888 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3892 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3894 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3895 atomic_dec(&fs_info->scrubs_running);
3896 wake_up(&fs_info->scrub_pause_wait);
3898 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3901 memcpy(progress, &sctx->stat, sizeof(*progress));
3903 mutex_lock(&fs_info->scrub_lock);
3904 dev->scrub_ctx = NULL;
3905 scrub_workers_put(fs_info);
3906 mutex_unlock(&fs_info->scrub_lock);
3908 scrub_put_ctx(sctx);
3913 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3915 mutex_lock(&fs_info->scrub_lock);
3916 atomic_inc(&fs_info->scrub_pause_req);
3917 while (atomic_read(&fs_info->scrubs_paused) !=
3918 atomic_read(&fs_info->scrubs_running)) {
3919 mutex_unlock(&fs_info->scrub_lock);
3920 wait_event(fs_info->scrub_pause_wait,
3921 atomic_read(&fs_info->scrubs_paused) ==
3922 atomic_read(&fs_info->scrubs_running));
3923 mutex_lock(&fs_info->scrub_lock);
3925 mutex_unlock(&fs_info->scrub_lock);
3928 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3930 atomic_dec(&fs_info->scrub_pause_req);
3931 wake_up(&fs_info->scrub_pause_wait);
3934 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3936 mutex_lock(&fs_info->scrub_lock);
3937 if (!atomic_read(&fs_info->scrubs_running)) {
3938 mutex_unlock(&fs_info->scrub_lock);
3942 atomic_inc(&fs_info->scrub_cancel_req);
3943 while (atomic_read(&fs_info->scrubs_running)) {
3944 mutex_unlock(&fs_info->scrub_lock);
3945 wait_event(fs_info->scrub_pause_wait,
3946 atomic_read(&fs_info->scrubs_running) == 0);
3947 mutex_lock(&fs_info->scrub_lock);
3949 atomic_dec(&fs_info->scrub_cancel_req);
3950 mutex_unlock(&fs_info->scrub_lock);
3955 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3956 struct btrfs_device *dev)
3958 struct scrub_ctx *sctx;
3960 mutex_lock(&fs_info->scrub_lock);
3961 sctx = dev->scrub_ctx;
3963 mutex_unlock(&fs_info->scrub_lock);
3966 atomic_inc(&sctx->cancel_req);
3967 while (dev->scrub_ctx) {
3968 mutex_unlock(&fs_info->scrub_lock);
3969 wait_event(fs_info->scrub_pause_wait,
3970 dev->scrub_ctx == NULL);
3971 mutex_lock(&fs_info->scrub_lock);
3973 mutex_unlock(&fs_info->scrub_lock);
3978 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
3979 struct btrfs_scrub_progress *progress)
3981 struct btrfs_device *dev;
3982 struct scrub_ctx *sctx = NULL;
3984 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3985 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3987 sctx = dev->scrub_ctx;
3989 memcpy(progress, &sctx->stat, sizeof(*progress));
3990 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3992 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3995 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3996 u64 extent_logical, u64 extent_len,
3997 u64 *extent_physical,
3998 struct btrfs_device **extent_dev,
3999 int *extent_mirror_num)
4002 struct btrfs_bio *bbio = NULL;
4005 mapped_length = extent_len;
4006 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4007 &mapped_length, &bbio, 0);
4008 if (ret || !bbio || mapped_length < extent_len ||
4009 !bbio->stripes[0].dev->bdev) {
4010 btrfs_put_bbio(bbio);
4014 *extent_physical = bbio->stripes[0].physical;
4015 *extent_mirror_num = bbio->mirror_num;
4016 *extent_dev = bbio->stripes[0].dev;
4017 btrfs_put_bbio(bbio);