1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
9 #include <crypto/hash.h>
13 #include "ordered-data.h"
14 #include "transaction.h"
16 #include "extent_io.h"
17 #include "dev-replace.h"
18 #include "check-integrity.h"
19 #include "rcu-string.h"
23 * This is only the first step towards a full-features scrub. It reads all
24 * extent and super block and verifies the checksums. In case a bad checksum
25 * is found or the extent cannot be read, good data will be written back if
28 * Future enhancements:
29 * - In case an unrepairable extent is encountered, track which files are
30 * affected and report them
31 * - track and record media errors, throw out bad devices
32 * - add a mode to also read unallocated space
39 * the following three values only influence the performance.
40 * The last one configures the number of parallel and outstanding I/O
41 * operations. The first two values configure an upper limit for the number
42 * of (dynamically allocated) pages that are added to a bio.
44 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
45 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
46 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
49 * the following value times PAGE_SIZE needs to be large enough to match the
50 * largest node/leaf/sector size that shall be supported.
51 * Values larger than BTRFS_STRIPE_LEN are not supported.
53 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
55 struct scrub_recover {
57 struct btrfs_bio *bbio;
62 struct scrub_block *sblock;
64 struct btrfs_device *dev;
65 struct list_head list;
66 u64 flags; /* extent flags */
70 u64 physical_for_dev_replace;
73 unsigned int mirror_num:8;
74 unsigned int have_csum:1;
75 unsigned int io_error:1;
77 u8 csum[BTRFS_CSUM_SIZE];
79 struct scrub_recover *recover;
84 struct scrub_ctx *sctx;
85 struct btrfs_device *dev;
90 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
91 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
93 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
97 struct btrfs_work work;
101 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
103 atomic_t outstanding_pages;
104 refcount_t refs; /* free mem on transition to zero */
105 struct scrub_ctx *sctx;
106 struct scrub_parity *sparity;
108 unsigned int header_error:1;
109 unsigned int checksum_error:1;
110 unsigned int no_io_error_seen:1;
111 unsigned int generation_error:1; /* also sets header_error */
113 /* The following is for the data used to check parity */
114 /* It is for the data with checksum */
115 unsigned int data_corrected:1;
117 struct btrfs_work work;
120 /* Used for the chunks with parity stripe such RAID5/6 */
121 struct scrub_parity {
122 struct scrub_ctx *sctx;
124 struct btrfs_device *scrub_dev;
136 struct list_head spages;
138 /* Work of parity check and repair */
139 struct btrfs_work work;
141 /* Mark the parity blocks which have data */
142 unsigned long *dbitmap;
145 * Mark the parity blocks which have data, but errors happen when
146 * read data or check data
148 unsigned long *ebitmap;
150 unsigned long bitmap[0];
154 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
155 struct btrfs_fs_info *fs_info;
158 atomic_t bios_in_flight;
159 atomic_t workers_pending;
160 spinlock_t list_lock;
161 wait_queue_head_t list_wait;
163 struct list_head csum_list;
166 int pages_per_rd_bio;
170 struct scrub_bio *wr_curr_bio;
171 struct mutex wr_lock;
172 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
173 struct btrfs_device *wr_tgtdev;
174 bool flush_all_writes;
179 struct btrfs_scrub_progress stat;
180 spinlock_t stat_lock;
183 * Use a ref counter to avoid use-after-free issues. Scrub workers
184 * decrement bios_in_flight and workers_pending and then do a wakeup
185 * on the list_wait wait queue. We must ensure the main scrub task
186 * doesn't free the scrub context before or while the workers are
187 * doing the wakeup() call.
192 struct scrub_warning {
193 struct btrfs_path *path;
194 u64 extent_item_size;
198 struct btrfs_device *dev;
201 struct full_stripe_lock {
208 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
209 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
210 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
211 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
212 struct scrub_block *sblocks_for_recheck);
213 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
214 struct scrub_block *sblock,
215 int retry_failed_mirror);
216 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
217 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
218 struct scrub_block *sblock_good);
219 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
220 struct scrub_block *sblock_good,
221 int page_num, int force_write);
222 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
223 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
225 static int scrub_checksum_data(struct scrub_block *sblock);
226 static int scrub_checksum_tree_block(struct scrub_block *sblock);
227 static int scrub_checksum_super(struct scrub_block *sblock);
228 static void scrub_block_get(struct scrub_block *sblock);
229 static void scrub_block_put(struct scrub_block *sblock);
230 static void scrub_page_get(struct scrub_page *spage);
231 static void scrub_page_put(struct scrub_page *spage);
232 static void scrub_parity_get(struct scrub_parity *sparity);
233 static void scrub_parity_put(struct scrub_parity *sparity);
234 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
235 struct scrub_page *spage);
236 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
237 u64 physical, struct btrfs_device *dev, u64 flags,
238 u64 gen, int mirror_num, u8 *csum, int force,
239 u64 physical_for_dev_replace);
240 static void scrub_bio_end_io(struct bio *bio);
241 static void scrub_bio_end_io_worker(struct btrfs_work *work);
242 static void scrub_block_complete(struct scrub_block *sblock);
243 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
244 u64 extent_logical, u64 extent_len,
245 u64 *extent_physical,
246 struct btrfs_device **extent_dev,
247 int *extent_mirror_num);
248 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
249 struct scrub_page *spage);
250 static void scrub_wr_submit(struct scrub_ctx *sctx);
251 static void scrub_wr_bio_end_io(struct bio *bio);
252 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
253 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
254 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
255 static void scrub_put_ctx(struct scrub_ctx *sctx);
257 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
259 return page->recover &&
260 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
263 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
265 refcount_inc(&sctx->refs);
266 atomic_inc(&sctx->bios_in_flight);
269 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
271 atomic_dec(&sctx->bios_in_flight);
272 wake_up(&sctx->list_wait);
276 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
278 while (atomic_read(&fs_info->scrub_pause_req)) {
279 mutex_unlock(&fs_info->scrub_lock);
280 wait_event(fs_info->scrub_pause_wait,
281 atomic_read(&fs_info->scrub_pause_req) == 0);
282 mutex_lock(&fs_info->scrub_lock);
286 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
288 atomic_inc(&fs_info->scrubs_paused);
289 wake_up(&fs_info->scrub_pause_wait);
292 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
294 mutex_lock(&fs_info->scrub_lock);
295 __scrub_blocked_if_needed(fs_info);
296 atomic_dec(&fs_info->scrubs_paused);
297 mutex_unlock(&fs_info->scrub_lock);
299 wake_up(&fs_info->scrub_pause_wait);
302 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
304 scrub_pause_on(fs_info);
305 scrub_pause_off(fs_info);
309 * Insert new full stripe lock into full stripe locks tree
311 * Return pointer to existing or newly inserted full_stripe_lock structure if
312 * everything works well.
313 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
315 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
318 static struct full_stripe_lock *insert_full_stripe_lock(
319 struct btrfs_full_stripe_locks_tree *locks_root,
323 struct rb_node *parent = NULL;
324 struct full_stripe_lock *entry;
325 struct full_stripe_lock *ret;
327 lockdep_assert_held(&locks_root->lock);
329 p = &locks_root->root.rb_node;
332 entry = rb_entry(parent, struct full_stripe_lock, node);
333 if (fstripe_logical < entry->logical) {
335 } else if (fstripe_logical > entry->logical) {
346 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
348 return ERR_PTR(-ENOMEM);
349 ret->logical = fstripe_logical;
351 mutex_init(&ret->mutex);
353 rb_link_node(&ret->node, parent, p);
354 rb_insert_color(&ret->node, &locks_root->root);
359 * Search for a full stripe lock of a block group
361 * Return pointer to existing full stripe lock if found
362 * Return NULL if not found
364 static struct full_stripe_lock *search_full_stripe_lock(
365 struct btrfs_full_stripe_locks_tree *locks_root,
368 struct rb_node *node;
369 struct full_stripe_lock *entry;
371 lockdep_assert_held(&locks_root->lock);
373 node = locks_root->root.rb_node;
375 entry = rb_entry(node, struct full_stripe_lock, node);
376 if (fstripe_logical < entry->logical)
377 node = node->rb_left;
378 else if (fstripe_logical > entry->logical)
379 node = node->rb_right;
387 * Helper to get full stripe logical from a normal bytenr.
389 * Caller must ensure @cache is a RAID56 block group.
391 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
397 * Due to chunk item size limit, full stripe length should not be
398 * larger than U32_MAX. Just a sanity check here.
400 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
403 * round_down() can only handle power of 2, while RAID56 full
404 * stripe length can be 64KiB * n, so we need to manually round down.
406 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
407 cache->full_stripe_len + cache->key.objectid;
412 * Lock a full stripe to avoid concurrency of recovery and read
414 * It's only used for profiles with parities (RAID5/6), for other profiles it
417 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
418 * So caller must call unlock_full_stripe() at the same context.
420 * Return <0 if encounters error.
422 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
425 struct btrfs_block_group_cache *bg_cache;
426 struct btrfs_full_stripe_locks_tree *locks_root;
427 struct full_stripe_lock *existing;
432 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
438 /* Profiles not based on parity don't need full stripe lock */
439 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
441 locks_root = &bg_cache->full_stripe_locks_root;
443 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
445 /* Now insert the full stripe lock */
446 mutex_lock(&locks_root->lock);
447 existing = insert_full_stripe_lock(locks_root, fstripe_start);
448 mutex_unlock(&locks_root->lock);
449 if (IS_ERR(existing)) {
450 ret = PTR_ERR(existing);
453 mutex_lock(&existing->mutex);
456 btrfs_put_block_group(bg_cache);
461 * Unlock a full stripe.
463 * NOTE: Caller must ensure it's the same context calling corresponding
464 * lock_full_stripe().
466 * Return 0 if we unlock full stripe without problem.
467 * Return <0 for error
469 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
472 struct btrfs_block_group_cache *bg_cache;
473 struct btrfs_full_stripe_locks_tree *locks_root;
474 struct full_stripe_lock *fstripe_lock;
479 /* If we didn't acquire full stripe lock, no need to continue */
483 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
488 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
491 locks_root = &bg_cache->full_stripe_locks_root;
492 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
494 mutex_lock(&locks_root->lock);
495 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
496 /* Unpaired unlock_full_stripe() detected */
500 mutex_unlock(&locks_root->lock);
504 if (fstripe_lock->refs == 0) {
506 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
507 fstripe_lock->logical);
509 fstripe_lock->refs--;
512 if (fstripe_lock->refs == 0) {
513 rb_erase(&fstripe_lock->node, &locks_root->root);
516 mutex_unlock(&locks_root->lock);
518 mutex_unlock(&fstripe_lock->mutex);
522 btrfs_put_block_group(bg_cache);
526 static void scrub_free_csums(struct scrub_ctx *sctx)
528 while (!list_empty(&sctx->csum_list)) {
529 struct btrfs_ordered_sum *sum;
530 sum = list_first_entry(&sctx->csum_list,
531 struct btrfs_ordered_sum, list);
532 list_del(&sum->list);
537 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
544 /* this can happen when scrub is cancelled */
545 if (sctx->curr != -1) {
546 struct scrub_bio *sbio = sctx->bios[sctx->curr];
548 for (i = 0; i < sbio->page_count; i++) {
549 WARN_ON(!sbio->pagev[i]->page);
550 scrub_block_put(sbio->pagev[i]->sblock);
555 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
556 struct scrub_bio *sbio = sctx->bios[i];
563 kfree(sctx->wr_curr_bio);
564 scrub_free_csums(sctx);
568 static void scrub_put_ctx(struct scrub_ctx *sctx)
570 if (refcount_dec_and_test(&sctx->refs))
571 scrub_free_ctx(sctx);
574 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
575 struct btrfs_fs_info *fs_info, int is_dev_replace)
577 struct scrub_ctx *sctx;
580 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
583 refcount_set(&sctx->refs, 1);
584 sctx->is_dev_replace = is_dev_replace;
585 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
587 sctx->fs_info = fs_info;
588 INIT_LIST_HEAD(&sctx->csum_list);
589 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
590 struct scrub_bio *sbio;
592 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
595 sctx->bios[i] = sbio;
599 sbio->page_count = 0;
600 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
601 scrub_bio_end_io_worker, NULL, NULL);
603 if (i != SCRUB_BIOS_PER_SCTX - 1)
604 sctx->bios[i]->next_free = i + 1;
606 sctx->bios[i]->next_free = -1;
608 sctx->first_free = 0;
609 atomic_set(&sctx->bios_in_flight, 0);
610 atomic_set(&sctx->workers_pending, 0);
611 atomic_set(&sctx->cancel_req, 0);
612 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
614 spin_lock_init(&sctx->list_lock);
615 spin_lock_init(&sctx->stat_lock);
616 init_waitqueue_head(&sctx->list_wait);
618 WARN_ON(sctx->wr_curr_bio != NULL);
619 mutex_init(&sctx->wr_lock);
620 sctx->wr_curr_bio = NULL;
621 if (is_dev_replace) {
622 WARN_ON(!fs_info->dev_replace.tgtdev);
623 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
624 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
625 sctx->flush_all_writes = false;
631 scrub_free_ctx(sctx);
632 return ERR_PTR(-ENOMEM);
635 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
643 struct extent_buffer *eb;
644 struct btrfs_inode_item *inode_item;
645 struct scrub_warning *swarn = warn_ctx;
646 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
647 struct inode_fs_paths *ipath = NULL;
648 struct btrfs_root *local_root;
649 struct btrfs_key root_key;
650 struct btrfs_key key;
652 root_key.objectid = root;
653 root_key.type = BTRFS_ROOT_ITEM_KEY;
654 root_key.offset = (u64)-1;
655 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
656 if (IS_ERR(local_root)) {
657 ret = PTR_ERR(local_root);
662 * this makes the path point to (inum INODE_ITEM ioff)
665 key.type = BTRFS_INODE_ITEM_KEY;
668 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
670 btrfs_release_path(swarn->path);
674 eb = swarn->path->nodes[0];
675 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
676 struct btrfs_inode_item);
677 isize = btrfs_inode_size(eb, inode_item);
678 nlink = btrfs_inode_nlink(eb, inode_item);
679 btrfs_release_path(swarn->path);
682 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
683 * uses GFP_NOFS in this context, so we keep it consistent but it does
684 * not seem to be strictly necessary.
686 nofs_flag = memalloc_nofs_save();
687 ipath = init_ipath(4096, local_root, swarn->path);
688 memalloc_nofs_restore(nofs_flag);
690 ret = PTR_ERR(ipath);
694 ret = paths_from_inode(inum, ipath);
700 * we deliberately ignore the bit ipath might have been too small to
701 * hold all of the paths here
703 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
704 btrfs_warn_in_rcu(fs_info,
705 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
706 swarn->errstr, swarn->logical,
707 rcu_str_deref(swarn->dev->name),
710 min(isize - offset, (u64)PAGE_SIZE), nlink,
711 (char *)(unsigned long)ipath->fspath->val[i]);
717 btrfs_warn_in_rcu(fs_info,
718 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
719 swarn->errstr, swarn->logical,
720 rcu_str_deref(swarn->dev->name),
722 root, inum, offset, ret);
728 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
730 struct btrfs_device *dev;
731 struct btrfs_fs_info *fs_info;
732 struct btrfs_path *path;
733 struct btrfs_key found_key;
734 struct extent_buffer *eb;
735 struct btrfs_extent_item *ei;
736 struct scrub_warning swarn;
737 unsigned long ptr = 0;
745 WARN_ON(sblock->page_count < 1);
746 dev = sblock->pagev[0]->dev;
747 fs_info = sblock->sctx->fs_info;
749 path = btrfs_alloc_path();
753 swarn.physical = sblock->pagev[0]->physical;
754 swarn.logical = sblock->pagev[0]->logical;
755 swarn.errstr = errstr;
758 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
763 extent_item_pos = swarn.logical - found_key.objectid;
764 swarn.extent_item_size = found_key.offset;
767 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
768 item_size = btrfs_item_size_nr(eb, path->slots[0]);
770 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
772 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
773 item_size, &ref_root,
775 btrfs_warn_in_rcu(fs_info,
776 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
777 errstr, swarn.logical,
778 rcu_str_deref(dev->name),
780 ref_level ? "node" : "leaf",
781 ret < 0 ? -1 : ref_level,
782 ret < 0 ? -1 : ref_root);
784 btrfs_release_path(path);
786 btrfs_release_path(path);
789 iterate_extent_inodes(fs_info, found_key.objectid,
791 scrub_print_warning_inode, &swarn, false);
795 btrfs_free_path(path);
798 static inline void scrub_get_recover(struct scrub_recover *recover)
800 refcount_inc(&recover->refs);
803 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
804 struct scrub_recover *recover)
806 if (refcount_dec_and_test(&recover->refs)) {
807 btrfs_bio_counter_dec(fs_info);
808 btrfs_put_bbio(recover->bbio);
814 * scrub_handle_errored_block gets called when either verification of the
815 * pages failed or the bio failed to read, e.g. with EIO. In the latter
816 * case, this function handles all pages in the bio, even though only one
818 * The goal of this function is to repair the errored block by using the
819 * contents of one of the mirrors.
821 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
823 struct scrub_ctx *sctx = sblock_to_check->sctx;
824 struct btrfs_device *dev;
825 struct btrfs_fs_info *fs_info;
827 unsigned int failed_mirror_index;
828 unsigned int is_metadata;
829 unsigned int have_csum;
830 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
831 struct scrub_block *sblock_bad;
836 bool full_stripe_locked;
837 unsigned int nofs_flag;
838 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
839 DEFAULT_RATELIMIT_BURST);
841 BUG_ON(sblock_to_check->page_count < 1);
842 fs_info = sctx->fs_info;
843 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
845 * if we find an error in a super block, we just report it.
846 * They will get written with the next transaction commit
849 spin_lock(&sctx->stat_lock);
850 ++sctx->stat.super_errors;
851 spin_unlock(&sctx->stat_lock);
854 logical = sblock_to_check->pagev[0]->logical;
855 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
856 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
857 is_metadata = !(sblock_to_check->pagev[0]->flags &
858 BTRFS_EXTENT_FLAG_DATA);
859 have_csum = sblock_to_check->pagev[0]->have_csum;
860 dev = sblock_to_check->pagev[0]->dev;
863 * We must use GFP_NOFS because the scrub task might be waiting for a
864 * worker task executing this function and in turn a transaction commit
865 * might be waiting the scrub task to pause (which needs to wait for all
866 * the worker tasks to complete before pausing).
867 * We do allocations in the workers through insert_full_stripe_lock()
868 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
871 nofs_flag = memalloc_nofs_save();
873 * For RAID5/6, race can happen for a different device scrub thread.
874 * For data corruption, Parity and Data threads will both try
875 * to recovery the data.
876 * Race can lead to doubly added csum error, or even unrecoverable
879 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
881 memalloc_nofs_restore(nofs_flag);
882 spin_lock(&sctx->stat_lock);
884 sctx->stat.malloc_errors++;
885 sctx->stat.read_errors++;
886 sctx->stat.uncorrectable_errors++;
887 spin_unlock(&sctx->stat_lock);
892 * read all mirrors one after the other. This includes to
893 * re-read the extent or metadata block that failed (that was
894 * the cause that this fixup code is called) another time,
895 * page by page this time in order to know which pages
896 * caused I/O errors and which ones are good (for all mirrors).
897 * It is the goal to handle the situation when more than one
898 * mirror contains I/O errors, but the errors do not
899 * overlap, i.e. the data can be repaired by selecting the
900 * pages from those mirrors without I/O error on the
901 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
902 * would be that mirror #1 has an I/O error on the first page,
903 * the second page is good, and mirror #2 has an I/O error on
904 * the second page, but the first page is good.
905 * Then the first page of the first mirror can be repaired by
906 * taking the first page of the second mirror, and the
907 * second page of the second mirror can be repaired by
908 * copying the contents of the 2nd page of the 1st mirror.
909 * One more note: if the pages of one mirror contain I/O
910 * errors, the checksum cannot be verified. In order to get
911 * the best data for repairing, the first attempt is to find
912 * a mirror without I/O errors and with a validated checksum.
913 * Only if this is not possible, the pages are picked from
914 * mirrors with I/O errors without considering the checksum.
915 * If the latter is the case, at the end, the checksum of the
916 * repaired area is verified in order to correctly maintain
920 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
921 sizeof(*sblocks_for_recheck), GFP_KERNEL);
922 if (!sblocks_for_recheck) {
923 spin_lock(&sctx->stat_lock);
924 sctx->stat.malloc_errors++;
925 sctx->stat.read_errors++;
926 sctx->stat.uncorrectable_errors++;
927 spin_unlock(&sctx->stat_lock);
928 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
932 /* setup the context, map the logical blocks and alloc the pages */
933 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
935 spin_lock(&sctx->stat_lock);
936 sctx->stat.read_errors++;
937 sctx->stat.uncorrectable_errors++;
938 spin_unlock(&sctx->stat_lock);
939 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
942 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
943 sblock_bad = sblocks_for_recheck + failed_mirror_index;
945 /* build and submit the bios for the failed mirror, check checksums */
946 scrub_recheck_block(fs_info, sblock_bad, 1);
948 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
949 sblock_bad->no_io_error_seen) {
951 * the error disappeared after reading page by page, or
952 * the area was part of a huge bio and other parts of the
953 * bio caused I/O errors, or the block layer merged several
954 * read requests into one and the error is caused by a
955 * different bio (usually one of the two latter cases is
958 spin_lock(&sctx->stat_lock);
959 sctx->stat.unverified_errors++;
960 sblock_to_check->data_corrected = 1;
961 spin_unlock(&sctx->stat_lock);
963 if (sctx->is_dev_replace)
964 scrub_write_block_to_dev_replace(sblock_bad);
968 if (!sblock_bad->no_io_error_seen) {
969 spin_lock(&sctx->stat_lock);
970 sctx->stat.read_errors++;
971 spin_unlock(&sctx->stat_lock);
972 if (__ratelimit(&_rs))
973 scrub_print_warning("i/o error", sblock_to_check);
974 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
975 } else if (sblock_bad->checksum_error) {
976 spin_lock(&sctx->stat_lock);
977 sctx->stat.csum_errors++;
978 spin_unlock(&sctx->stat_lock);
979 if (__ratelimit(&_rs))
980 scrub_print_warning("checksum error", sblock_to_check);
981 btrfs_dev_stat_inc_and_print(dev,
982 BTRFS_DEV_STAT_CORRUPTION_ERRS);
983 } else if (sblock_bad->header_error) {
984 spin_lock(&sctx->stat_lock);
985 sctx->stat.verify_errors++;
986 spin_unlock(&sctx->stat_lock);
987 if (__ratelimit(&_rs))
988 scrub_print_warning("checksum/header error",
990 if (sblock_bad->generation_error)
991 btrfs_dev_stat_inc_and_print(dev,
992 BTRFS_DEV_STAT_GENERATION_ERRS);
994 btrfs_dev_stat_inc_and_print(dev,
995 BTRFS_DEV_STAT_CORRUPTION_ERRS);
998 if (sctx->readonly) {
999 ASSERT(!sctx->is_dev_replace);
1004 * now build and submit the bios for the other mirrors, check
1006 * First try to pick the mirror which is completely without I/O
1007 * errors and also does not have a checksum error.
1008 * If one is found, and if a checksum is present, the full block
1009 * that is known to contain an error is rewritten. Afterwards
1010 * the block is known to be corrected.
1011 * If a mirror is found which is completely correct, and no
1012 * checksum is present, only those pages are rewritten that had
1013 * an I/O error in the block to be repaired, since it cannot be
1014 * determined, which copy of the other pages is better (and it
1015 * could happen otherwise that a correct page would be
1016 * overwritten by a bad one).
1018 for (mirror_index = 0; ;mirror_index++) {
1019 struct scrub_block *sblock_other;
1021 if (mirror_index == failed_mirror_index)
1024 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1025 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1026 if (mirror_index >= BTRFS_MAX_MIRRORS)
1028 if (!sblocks_for_recheck[mirror_index].page_count)
1031 sblock_other = sblocks_for_recheck + mirror_index;
1033 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1034 int max_allowed = r->bbio->num_stripes -
1035 r->bbio->num_tgtdevs;
1037 if (mirror_index >= max_allowed)
1039 if (!sblocks_for_recheck[1].page_count)
1042 ASSERT(failed_mirror_index == 0);
1043 sblock_other = sblocks_for_recheck + 1;
1044 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1047 /* build and submit the bios, check checksums */
1048 scrub_recheck_block(fs_info, sblock_other, 0);
1050 if (!sblock_other->header_error &&
1051 !sblock_other->checksum_error &&
1052 sblock_other->no_io_error_seen) {
1053 if (sctx->is_dev_replace) {
1054 scrub_write_block_to_dev_replace(sblock_other);
1055 goto corrected_error;
1057 ret = scrub_repair_block_from_good_copy(
1058 sblock_bad, sblock_other);
1060 goto corrected_error;
1065 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1066 goto did_not_correct_error;
1069 * In case of I/O errors in the area that is supposed to be
1070 * repaired, continue by picking good copies of those pages.
1071 * Select the good pages from mirrors to rewrite bad pages from
1072 * the area to fix. Afterwards verify the checksum of the block
1073 * that is supposed to be repaired. This verification step is
1074 * only done for the purpose of statistic counting and for the
1075 * final scrub report, whether errors remain.
1076 * A perfect algorithm could make use of the checksum and try
1077 * all possible combinations of pages from the different mirrors
1078 * until the checksum verification succeeds. For example, when
1079 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1080 * of mirror #2 is readable but the final checksum test fails,
1081 * then the 2nd page of mirror #3 could be tried, whether now
1082 * the final checksum succeeds. But this would be a rare
1083 * exception and is therefore not implemented. At least it is
1084 * avoided that the good copy is overwritten.
1085 * A more useful improvement would be to pick the sectors
1086 * without I/O error based on sector sizes (512 bytes on legacy
1087 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1088 * mirror could be repaired by taking 512 byte of a different
1089 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1090 * area are unreadable.
1093 for (page_num = 0; page_num < sblock_bad->page_count;
1095 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1096 struct scrub_block *sblock_other = NULL;
1098 /* skip no-io-error page in scrub */
1099 if (!page_bad->io_error && !sctx->is_dev_replace)
1102 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1104 * In case of dev replace, if raid56 rebuild process
1105 * didn't work out correct data, then copy the content
1106 * in sblock_bad to make sure target device is identical
1107 * to source device, instead of writing garbage data in
1108 * sblock_for_recheck array to target device.
1110 sblock_other = NULL;
1111 } else if (page_bad->io_error) {
1112 /* try to find no-io-error page in mirrors */
1113 for (mirror_index = 0;
1114 mirror_index < BTRFS_MAX_MIRRORS &&
1115 sblocks_for_recheck[mirror_index].page_count > 0;
1117 if (!sblocks_for_recheck[mirror_index].
1118 pagev[page_num]->io_error) {
1119 sblock_other = sblocks_for_recheck +
1128 if (sctx->is_dev_replace) {
1130 * did not find a mirror to fetch the page
1131 * from. scrub_write_page_to_dev_replace()
1132 * handles this case (page->io_error), by
1133 * filling the block with zeros before
1134 * submitting the write request
1137 sblock_other = sblock_bad;
1139 if (scrub_write_page_to_dev_replace(sblock_other,
1142 &fs_info->dev_replace.num_write_errors);
1145 } else if (sblock_other) {
1146 ret = scrub_repair_page_from_good_copy(sblock_bad,
1150 page_bad->io_error = 0;
1156 if (success && !sctx->is_dev_replace) {
1157 if (is_metadata || have_csum) {
1159 * need to verify the checksum now that all
1160 * sectors on disk are repaired (the write
1161 * request for data to be repaired is on its way).
1162 * Just be lazy and use scrub_recheck_block()
1163 * which re-reads the data before the checksum
1164 * is verified, but most likely the data comes out
1165 * of the page cache.
1167 scrub_recheck_block(fs_info, sblock_bad, 1);
1168 if (!sblock_bad->header_error &&
1169 !sblock_bad->checksum_error &&
1170 sblock_bad->no_io_error_seen)
1171 goto corrected_error;
1173 goto did_not_correct_error;
1176 spin_lock(&sctx->stat_lock);
1177 sctx->stat.corrected_errors++;
1178 sblock_to_check->data_corrected = 1;
1179 spin_unlock(&sctx->stat_lock);
1180 btrfs_err_rl_in_rcu(fs_info,
1181 "fixed up error at logical %llu on dev %s",
1182 logical, rcu_str_deref(dev->name));
1185 did_not_correct_error:
1186 spin_lock(&sctx->stat_lock);
1187 sctx->stat.uncorrectable_errors++;
1188 spin_unlock(&sctx->stat_lock);
1189 btrfs_err_rl_in_rcu(fs_info,
1190 "unable to fixup (regular) error at logical %llu on dev %s",
1191 logical, rcu_str_deref(dev->name));
1195 if (sblocks_for_recheck) {
1196 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1198 struct scrub_block *sblock = sblocks_for_recheck +
1200 struct scrub_recover *recover;
1203 for (page_index = 0; page_index < sblock->page_count;
1205 sblock->pagev[page_index]->sblock = NULL;
1206 recover = sblock->pagev[page_index]->recover;
1208 scrub_put_recover(fs_info, recover);
1209 sblock->pagev[page_index]->recover =
1212 scrub_page_put(sblock->pagev[page_index]);
1215 kfree(sblocks_for_recheck);
1218 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1219 memalloc_nofs_restore(nofs_flag);
1225 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1227 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1229 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1232 return (int)bbio->num_stripes;
1235 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1238 int nstripes, int mirror,
1244 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1246 for (i = 0; i < nstripes; i++) {
1247 if (raid_map[i] == RAID6_Q_STRIPE ||
1248 raid_map[i] == RAID5_P_STRIPE)
1251 if (logical >= raid_map[i] &&
1252 logical < raid_map[i] + mapped_length)
1257 *stripe_offset = logical - raid_map[i];
1259 /* The other RAID type */
1260 *stripe_index = mirror;
1265 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1266 struct scrub_block *sblocks_for_recheck)
1268 struct scrub_ctx *sctx = original_sblock->sctx;
1269 struct btrfs_fs_info *fs_info = sctx->fs_info;
1270 u64 length = original_sblock->page_count * PAGE_SIZE;
1271 u64 logical = original_sblock->pagev[0]->logical;
1272 u64 generation = original_sblock->pagev[0]->generation;
1273 u64 flags = original_sblock->pagev[0]->flags;
1274 u64 have_csum = original_sblock->pagev[0]->have_csum;
1275 struct scrub_recover *recover;
1276 struct btrfs_bio *bbio;
1287 * note: the two members refs and outstanding_pages
1288 * are not used (and not set) in the blocks that are used for
1289 * the recheck procedure
1292 while (length > 0) {
1293 sublen = min_t(u64, length, PAGE_SIZE);
1294 mapped_length = sublen;
1298 * with a length of PAGE_SIZE, each returned stripe
1299 * represents one mirror
1301 btrfs_bio_counter_inc_blocked(fs_info);
1302 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1303 logical, &mapped_length, &bbio);
1304 if (ret || !bbio || mapped_length < sublen) {
1305 btrfs_put_bbio(bbio);
1306 btrfs_bio_counter_dec(fs_info);
1310 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1312 btrfs_put_bbio(bbio);
1313 btrfs_bio_counter_dec(fs_info);
1317 refcount_set(&recover->refs, 1);
1318 recover->bbio = bbio;
1319 recover->map_length = mapped_length;
1321 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1323 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1325 for (mirror_index = 0; mirror_index < nmirrors;
1327 struct scrub_block *sblock;
1328 struct scrub_page *page;
1330 sblock = sblocks_for_recheck + mirror_index;
1331 sblock->sctx = sctx;
1333 page = kzalloc(sizeof(*page), GFP_NOFS);
1336 spin_lock(&sctx->stat_lock);
1337 sctx->stat.malloc_errors++;
1338 spin_unlock(&sctx->stat_lock);
1339 scrub_put_recover(fs_info, recover);
1342 scrub_page_get(page);
1343 sblock->pagev[page_index] = page;
1344 page->sblock = sblock;
1345 page->flags = flags;
1346 page->generation = generation;
1347 page->logical = logical;
1348 page->have_csum = have_csum;
1351 original_sblock->pagev[0]->csum,
1354 scrub_stripe_index_and_offset(logical,
1363 page->physical = bbio->stripes[stripe_index].physical +
1365 page->dev = bbio->stripes[stripe_index].dev;
1367 BUG_ON(page_index >= original_sblock->page_count);
1368 page->physical_for_dev_replace =
1369 original_sblock->pagev[page_index]->
1370 physical_for_dev_replace;
1371 /* for missing devices, dev->bdev is NULL */
1372 page->mirror_num = mirror_index + 1;
1373 sblock->page_count++;
1374 page->page = alloc_page(GFP_NOFS);
1378 scrub_get_recover(recover);
1379 page->recover = recover;
1381 scrub_put_recover(fs_info, recover);
1390 static void scrub_bio_wait_endio(struct bio *bio)
1392 complete(bio->bi_private);
1395 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1397 struct scrub_page *page)
1399 DECLARE_COMPLETION_ONSTACK(done);
1403 bio->bi_iter.bi_sector = page->logical >> 9;
1404 bio->bi_private = &done;
1405 bio->bi_end_io = scrub_bio_wait_endio;
1407 mirror_num = page->sblock->pagev[0]->mirror_num;
1408 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1409 page->recover->map_length,
1414 wait_for_completion_io(&done);
1415 return blk_status_to_errno(bio->bi_status);
1418 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1419 struct scrub_block *sblock)
1421 struct scrub_page *first_page = sblock->pagev[0];
1425 /* All pages in sblock belong to the same stripe on the same device. */
1426 ASSERT(first_page->dev);
1427 if (!first_page->dev->bdev)
1430 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1431 bio_set_dev(bio, first_page->dev->bdev);
1433 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1434 struct scrub_page *page = sblock->pagev[page_num];
1436 WARN_ON(!page->page);
1437 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1440 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1447 scrub_recheck_block_checksum(sblock);
1451 for (page_num = 0; page_num < sblock->page_count; page_num++)
1452 sblock->pagev[page_num]->io_error = 1;
1454 sblock->no_io_error_seen = 0;
1458 * this function will check the on disk data for checksum errors, header
1459 * errors and read I/O errors. If any I/O errors happen, the exact pages
1460 * which are errored are marked as being bad. The goal is to enable scrub
1461 * to take those pages that are not errored from all the mirrors so that
1462 * the pages that are errored in the just handled mirror can be repaired.
1464 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1465 struct scrub_block *sblock,
1466 int retry_failed_mirror)
1470 sblock->no_io_error_seen = 1;
1472 /* short cut for raid56 */
1473 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1474 return scrub_recheck_block_on_raid56(fs_info, sblock);
1476 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1478 struct scrub_page *page = sblock->pagev[page_num];
1480 if (page->dev->bdev == NULL) {
1482 sblock->no_io_error_seen = 0;
1486 WARN_ON(!page->page);
1487 bio = btrfs_io_bio_alloc(1);
1488 bio_set_dev(bio, page->dev->bdev);
1490 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1491 bio->bi_iter.bi_sector = page->physical >> 9;
1492 bio->bi_opf = REQ_OP_READ;
1494 if (btrfsic_submit_bio_wait(bio)) {
1496 sblock->no_io_error_seen = 0;
1502 if (sblock->no_io_error_seen)
1503 scrub_recheck_block_checksum(sblock);
1506 static inline int scrub_check_fsid(u8 fsid[],
1507 struct scrub_page *spage)
1509 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1512 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1516 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1518 sblock->header_error = 0;
1519 sblock->checksum_error = 0;
1520 sblock->generation_error = 0;
1522 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1523 scrub_checksum_data(sblock);
1525 scrub_checksum_tree_block(sblock);
1528 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1529 struct scrub_block *sblock_good)
1534 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1537 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1547 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1548 struct scrub_block *sblock_good,
1549 int page_num, int force_write)
1551 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1552 struct scrub_page *page_good = sblock_good->pagev[page_num];
1553 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1555 BUG_ON(page_bad->page == NULL);
1556 BUG_ON(page_good->page == NULL);
1557 if (force_write || sblock_bad->header_error ||
1558 sblock_bad->checksum_error || page_bad->io_error) {
1562 if (!page_bad->dev->bdev) {
1563 btrfs_warn_rl(fs_info,
1564 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1568 bio = btrfs_io_bio_alloc(1);
1569 bio_set_dev(bio, page_bad->dev->bdev);
1570 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1571 bio->bi_opf = REQ_OP_WRITE;
1573 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1574 if (PAGE_SIZE != ret) {
1579 if (btrfsic_submit_bio_wait(bio)) {
1580 btrfs_dev_stat_inc_and_print(page_bad->dev,
1581 BTRFS_DEV_STAT_WRITE_ERRS);
1582 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1592 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1594 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1598 * This block is used for the check of the parity on the source device,
1599 * so the data needn't be written into the destination device.
1601 if (sblock->sparity)
1604 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1607 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1609 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1613 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1616 struct scrub_page *spage = sblock->pagev[page_num];
1618 BUG_ON(spage->page == NULL);
1619 if (spage->io_error) {
1620 void *mapped_buffer = kmap_atomic(spage->page);
1622 clear_page(mapped_buffer);
1623 flush_dcache_page(spage->page);
1624 kunmap_atomic(mapped_buffer);
1626 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1629 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1630 struct scrub_page *spage)
1632 struct scrub_bio *sbio;
1635 mutex_lock(&sctx->wr_lock);
1637 if (!sctx->wr_curr_bio) {
1638 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1640 if (!sctx->wr_curr_bio) {
1641 mutex_unlock(&sctx->wr_lock);
1644 sctx->wr_curr_bio->sctx = sctx;
1645 sctx->wr_curr_bio->page_count = 0;
1647 sbio = sctx->wr_curr_bio;
1648 if (sbio->page_count == 0) {
1651 sbio->physical = spage->physical_for_dev_replace;
1652 sbio->logical = spage->logical;
1653 sbio->dev = sctx->wr_tgtdev;
1656 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1660 bio->bi_private = sbio;
1661 bio->bi_end_io = scrub_wr_bio_end_io;
1662 bio_set_dev(bio, sbio->dev->bdev);
1663 bio->bi_iter.bi_sector = sbio->physical >> 9;
1664 bio->bi_opf = REQ_OP_WRITE;
1666 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1667 spage->physical_for_dev_replace ||
1668 sbio->logical + sbio->page_count * PAGE_SIZE !=
1670 scrub_wr_submit(sctx);
1674 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1675 if (ret != PAGE_SIZE) {
1676 if (sbio->page_count < 1) {
1679 mutex_unlock(&sctx->wr_lock);
1682 scrub_wr_submit(sctx);
1686 sbio->pagev[sbio->page_count] = spage;
1687 scrub_page_get(spage);
1689 if (sbio->page_count == sctx->pages_per_wr_bio)
1690 scrub_wr_submit(sctx);
1691 mutex_unlock(&sctx->wr_lock);
1696 static void scrub_wr_submit(struct scrub_ctx *sctx)
1698 struct scrub_bio *sbio;
1700 if (!sctx->wr_curr_bio)
1703 sbio = sctx->wr_curr_bio;
1704 sctx->wr_curr_bio = NULL;
1705 WARN_ON(!sbio->bio->bi_disk);
1706 scrub_pending_bio_inc(sctx);
1707 /* process all writes in a single worker thread. Then the block layer
1708 * orders the requests before sending them to the driver which
1709 * doubled the write performance on spinning disks when measured
1711 btrfsic_submit_bio(sbio->bio);
1714 static void scrub_wr_bio_end_io(struct bio *bio)
1716 struct scrub_bio *sbio = bio->bi_private;
1717 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1719 sbio->status = bio->bi_status;
1722 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1723 scrub_wr_bio_end_io_worker, NULL, NULL);
1724 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1727 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1729 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1730 struct scrub_ctx *sctx = sbio->sctx;
1733 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1735 struct btrfs_dev_replace *dev_replace =
1736 &sbio->sctx->fs_info->dev_replace;
1738 for (i = 0; i < sbio->page_count; i++) {
1739 struct scrub_page *spage = sbio->pagev[i];
1741 spage->io_error = 1;
1742 atomic64_inc(&dev_replace->num_write_errors);
1746 for (i = 0; i < sbio->page_count; i++)
1747 scrub_page_put(sbio->pagev[i]);
1751 scrub_pending_bio_dec(sctx);
1754 static int scrub_checksum(struct scrub_block *sblock)
1760 * No need to initialize these stats currently,
1761 * because this function only use return value
1762 * instead of these stats value.
1767 sblock->header_error = 0;
1768 sblock->generation_error = 0;
1769 sblock->checksum_error = 0;
1771 WARN_ON(sblock->page_count < 1);
1772 flags = sblock->pagev[0]->flags;
1774 if (flags & BTRFS_EXTENT_FLAG_DATA)
1775 ret = scrub_checksum_data(sblock);
1776 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1777 ret = scrub_checksum_tree_block(sblock);
1778 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1779 (void)scrub_checksum_super(sblock);
1783 scrub_handle_errored_block(sblock);
1788 static int scrub_checksum_data(struct scrub_block *sblock)
1790 struct scrub_ctx *sctx = sblock->sctx;
1791 struct btrfs_fs_info *fs_info = sctx->fs_info;
1792 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1793 u8 csum[BTRFS_CSUM_SIZE];
1800 BUG_ON(sblock->page_count < 1);
1801 if (!sblock->pagev[0]->have_csum)
1804 shash->tfm = fs_info->csum_shash;
1805 crypto_shash_init(shash);
1807 on_disk_csum = sblock->pagev[0]->csum;
1808 page = sblock->pagev[0]->page;
1809 buffer = kmap_atomic(page);
1811 len = sctx->fs_info->sectorsize;
1814 u64 l = min_t(u64, len, PAGE_SIZE);
1816 crypto_shash_update(shash, buffer, l);
1817 kunmap_atomic(buffer);
1822 BUG_ON(index >= sblock->page_count);
1823 BUG_ON(!sblock->pagev[index]->page);
1824 page = sblock->pagev[index]->page;
1825 buffer = kmap_atomic(page);
1828 crypto_shash_final(shash, csum);
1829 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1830 sblock->checksum_error = 1;
1832 return sblock->checksum_error;
1835 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1837 struct scrub_ctx *sctx = sblock->sctx;
1838 struct btrfs_header *h;
1839 struct btrfs_fs_info *fs_info = sctx->fs_info;
1840 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1841 u8 calculated_csum[BTRFS_CSUM_SIZE];
1842 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1844 void *mapped_buffer;
1850 shash->tfm = fs_info->csum_shash;
1851 crypto_shash_init(shash);
1853 BUG_ON(sblock->page_count < 1);
1854 page = sblock->pagev[0]->page;
1855 mapped_buffer = kmap_atomic(page);
1856 h = (struct btrfs_header *)mapped_buffer;
1857 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1860 * we don't use the getter functions here, as we
1861 * a) don't have an extent buffer and
1862 * b) the page is already kmapped
1864 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1865 sblock->header_error = 1;
1867 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1868 sblock->header_error = 1;
1869 sblock->generation_error = 1;
1872 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1873 sblock->header_error = 1;
1875 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1877 sblock->header_error = 1;
1879 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1880 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1881 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1884 u64 l = min_t(u64, len, mapped_size);
1886 crypto_shash_update(shash, p, l);
1887 kunmap_atomic(mapped_buffer);
1892 BUG_ON(index >= sblock->page_count);
1893 BUG_ON(!sblock->pagev[index]->page);
1894 page = sblock->pagev[index]->page;
1895 mapped_buffer = kmap_atomic(page);
1896 mapped_size = PAGE_SIZE;
1900 crypto_shash_final(shash, calculated_csum);
1901 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1902 sblock->checksum_error = 1;
1904 return sblock->header_error || sblock->checksum_error;
1907 static int scrub_checksum_super(struct scrub_block *sblock)
1909 struct btrfs_super_block *s;
1910 struct scrub_ctx *sctx = sblock->sctx;
1911 struct btrfs_fs_info *fs_info = sctx->fs_info;
1912 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1913 u8 calculated_csum[BTRFS_CSUM_SIZE];
1914 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1916 void *mapped_buffer;
1924 shash->tfm = fs_info->csum_shash;
1925 crypto_shash_init(shash);
1927 BUG_ON(sblock->page_count < 1);
1928 page = sblock->pagev[0]->page;
1929 mapped_buffer = kmap_atomic(page);
1930 s = (struct btrfs_super_block *)mapped_buffer;
1931 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1933 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1936 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1939 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1942 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1943 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1944 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1947 u64 l = min_t(u64, len, mapped_size);
1949 crypto_shash_update(shash, p, l);
1950 kunmap_atomic(mapped_buffer);
1955 BUG_ON(index >= sblock->page_count);
1956 BUG_ON(!sblock->pagev[index]->page);
1957 page = sblock->pagev[index]->page;
1958 mapped_buffer = kmap_atomic(page);
1959 mapped_size = PAGE_SIZE;
1963 crypto_shash_final(shash, calculated_csum);
1964 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1967 if (fail_cor + fail_gen) {
1969 * if we find an error in a super block, we just report it.
1970 * They will get written with the next transaction commit
1973 spin_lock(&sctx->stat_lock);
1974 ++sctx->stat.super_errors;
1975 spin_unlock(&sctx->stat_lock);
1977 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1978 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1980 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1981 BTRFS_DEV_STAT_GENERATION_ERRS);
1984 return fail_cor + fail_gen;
1987 static void scrub_block_get(struct scrub_block *sblock)
1989 refcount_inc(&sblock->refs);
1992 static void scrub_block_put(struct scrub_block *sblock)
1994 if (refcount_dec_and_test(&sblock->refs)) {
1997 if (sblock->sparity)
1998 scrub_parity_put(sblock->sparity);
2000 for (i = 0; i < sblock->page_count; i++)
2001 scrub_page_put(sblock->pagev[i]);
2006 static void scrub_page_get(struct scrub_page *spage)
2008 atomic_inc(&spage->refs);
2011 static void scrub_page_put(struct scrub_page *spage)
2013 if (atomic_dec_and_test(&spage->refs)) {
2015 __free_page(spage->page);
2020 static void scrub_submit(struct scrub_ctx *sctx)
2022 struct scrub_bio *sbio;
2024 if (sctx->curr == -1)
2027 sbio = sctx->bios[sctx->curr];
2029 scrub_pending_bio_inc(sctx);
2030 btrfsic_submit_bio(sbio->bio);
2033 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2034 struct scrub_page *spage)
2036 struct scrub_block *sblock = spage->sblock;
2037 struct scrub_bio *sbio;
2042 * grab a fresh bio or wait for one to become available
2044 while (sctx->curr == -1) {
2045 spin_lock(&sctx->list_lock);
2046 sctx->curr = sctx->first_free;
2047 if (sctx->curr != -1) {
2048 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2049 sctx->bios[sctx->curr]->next_free = -1;
2050 sctx->bios[sctx->curr]->page_count = 0;
2051 spin_unlock(&sctx->list_lock);
2053 spin_unlock(&sctx->list_lock);
2054 wait_event(sctx->list_wait, sctx->first_free != -1);
2057 sbio = sctx->bios[sctx->curr];
2058 if (sbio->page_count == 0) {
2061 sbio->physical = spage->physical;
2062 sbio->logical = spage->logical;
2063 sbio->dev = spage->dev;
2066 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2070 bio->bi_private = sbio;
2071 bio->bi_end_io = scrub_bio_end_io;
2072 bio_set_dev(bio, sbio->dev->bdev);
2073 bio->bi_iter.bi_sector = sbio->physical >> 9;
2074 bio->bi_opf = REQ_OP_READ;
2076 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2078 sbio->logical + sbio->page_count * PAGE_SIZE !=
2080 sbio->dev != spage->dev) {
2085 sbio->pagev[sbio->page_count] = spage;
2086 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2087 if (ret != PAGE_SIZE) {
2088 if (sbio->page_count < 1) {
2097 scrub_block_get(sblock); /* one for the page added to the bio */
2098 atomic_inc(&sblock->outstanding_pages);
2100 if (sbio->page_count == sctx->pages_per_rd_bio)
2106 static void scrub_missing_raid56_end_io(struct bio *bio)
2108 struct scrub_block *sblock = bio->bi_private;
2109 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2112 sblock->no_io_error_seen = 0;
2116 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2119 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2121 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2122 struct scrub_ctx *sctx = sblock->sctx;
2123 struct btrfs_fs_info *fs_info = sctx->fs_info;
2125 struct btrfs_device *dev;
2127 logical = sblock->pagev[0]->logical;
2128 dev = sblock->pagev[0]->dev;
2130 if (sblock->no_io_error_seen)
2131 scrub_recheck_block_checksum(sblock);
2133 if (!sblock->no_io_error_seen) {
2134 spin_lock(&sctx->stat_lock);
2135 sctx->stat.read_errors++;
2136 spin_unlock(&sctx->stat_lock);
2137 btrfs_err_rl_in_rcu(fs_info,
2138 "IO error rebuilding logical %llu for dev %s",
2139 logical, rcu_str_deref(dev->name));
2140 } else if (sblock->header_error || sblock->checksum_error) {
2141 spin_lock(&sctx->stat_lock);
2142 sctx->stat.uncorrectable_errors++;
2143 spin_unlock(&sctx->stat_lock);
2144 btrfs_err_rl_in_rcu(fs_info,
2145 "failed to rebuild valid logical %llu for dev %s",
2146 logical, rcu_str_deref(dev->name));
2148 scrub_write_block_to_dev_replace(sblock);
2151 scrub_block_put(sblock);
2153 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2154 mutex_lock(&sctx->wr_lock);
2155 scrub_wr_submit(sctx);
2156 mutex_unlock(&sctx->wr_lock);
2159 scrub_pending_bio_dec(sctx);
2162 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2164 struct scrub_ctx *sctx = sblock->sctx;
2165 struct btrfs_fs_info *fs_info = sctx->fs_info;
2166 u64 length = sblock->page_count * PAGE_SIZE;
2167 u64 logical = sblock->pagev[0]->logical;
2168 struct btrfs_bio *bbio = NULL;
2170 struct btrfs_raid_bio *rbio;
2174 btrfs_bio_counter_inc_blocked(fs_info);
2175 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2177 if (ret || !bbio || !bbio->raid_map)
2180 if (WARN_ON(!sctx->is_dev_replace ||
2181 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2183 * We shouldn't be scrubbing a missing device. Even for dev
2184 * replace, we should only get here for RAID 5/6. We either
2185 * managed to mount something with no mirrors remaining or
2186 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2191 bio = btrfs_io_bio_alloc(0);
2192 bio->bi_iter.bi_sector = logical >> 9;
2193 bio->bi_private = sblock;
2194 bio->bi_end_io = scrub_missing_raid56_end_io;
2196 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2200 for (i = 0; i < sblock->page_count; i++) {
2201 struct scrub_page *spage = sblock->pagev[i];
2203 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2206 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2207 scrub_missing_raid56_worker, NULL, NULL);
2208 scrub_block_get(sblock);
2209 scrub_pending_bio_inc(sctx);
2210 raid56_submit_missing_rbio(rbio);
2216 btrfs_bio_counter_dec(fs_info);
2217 btrfs_put_bbio(bbio);
2218 spin_lock(&sctx->stat_lock);
2219 sctx->stat.malloc_errors++;
2220 spin_unlock(&sctx->stat_lock);
2223 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2224 u64 physical, struct btrfs_device *dev, u64 flags,
2225 u64 gen, int mirror_num, u8 *csum, int force,
2226 u64 physical_for_dev_replace)
2228 struct scrub_block *sblock;
2231 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2233 spin_lock(&sctx->stat_lock);
2234 sctx->stat.malloc_errors++;
2235 spin_unlock(&sctx->stat_lock);
2239 /* one ref inside this function, plus one for each page added to
2241 refcount_set(&sblock->refs, 1);
2242 sblock->sctx = sctx;
2243 sblock->no_io_error_seen = 1;
2245 for (index = 0; len > 0; index++) {
2246 struct scrub_page *spage;
2247 u64 l = min_t(u64, len, PAGE_SIZE);
2249 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2252 spin_lock(&sctx->stat_lock);
2253 sctx->stat.malloc_errors++;
2254 spin_unlock(&sctx->stat_lock);
2255 scrub_block_put(sblock);
2258 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2259 scrub_page_get(spage);
2260 sblock->pagev[index] = spage;
2261 spage->sblock = sblock;
2263 spage->flags = flags;
2264 spage->generation = gen;
2265 spage->logical = logical;
2266 spage->physical = physical;
2267 spage->physical_for_dev_replace = physical_for_dev_replace;
2268 spage->mirror_num = mirror_num;
2270 spage->have_csum = 1;
2271 memcpy(spage->csum, csum, sctx->csum_size);
2273 spage->have_csum = 0;
2275 sblock->page_count++;
2276 spage->page = alloc_page(GFP_KERNEL);
2282 physical_for_dev_replace += l;
2285 WARN_ON(sblock->page_count == 0);
2286 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2288 * This case should only be hit for RAID 5/6 device replace. See
2289 * the comment in scrub_missing_raid56_pages() for details.
2291 scrub_missing_raid56_pages(sblock);
2293 for (index = 0; index < sblock->page_count; index++) {
2294 struct scrub_page *spage = sblock->pagev[index];
2297 ret = scrub_add_page_to_rd_bio(sctx, spage);
2299 scrub_block_put(sblock);
2308 /* last one frees, either here or in bio completion for last page */
2309 scrub_block_put(sblock);
2313 static void scrub_bio_end_io(struct bio *bio)
2315 struct scrub_bio *sbio = bio->bi_private;
2316 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2318 sbio->status = bio->bi_status;
2321 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2324 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2326 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2327 struct scrub_ctx *sctx = sbio->sctx;
2330 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2332 for (i = 0; i < sbio->page_count; i++) {
2333 struct scrub_page *spage = sbio->pagev[i];
2335 spage->io_error = 1;
2336 spage->sblock->no_io_error_seen = 0;
2340 /* now complete the scrub_block items that have all pages completed */
2341 for (i = 0; i < sbio->page_count; i++) {
2342 struct scrub_page *spage = sbio->pagev[i];
2343 struct scrub_block *sblock = spage->sblock;
2345 if (atomic_dec_and_test(&sblock->outstanding_pages))
2346 scrub_block_complete(sblock);
2347 scrub_block_put(sblock);
2352 spin_lock(&sctx->list_lock);
2353 sbio->next_free = sctx->first_free;
2354 sctx->first_free = sbio->index;
2355 spin_unlock(&sctx->list_lock);
2357 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2358 mutex_lock(&sctx->wr_lock);
2359 scrub_wr_submit(sctx);
2360 mutex_unlock(&sctx->wr_lock);
2363 scrub_pending_bio_dec(sctx);
2366 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2367 unsigned long *bitmap,
2373 int sectorsize = sparity->sctx->fs_info->sectorsize;
2375 if (len >= sparity->stripe_len) {
2376 bitmap_set(bitmap, 0, sparity->nsectors);
2380 start -= sparity->logic_start;
2381 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2382 offset = div_u64(offset, sectorsize);
2383 nsectors64 = div_u64(len, sectorsize);
2385 ASSERT(nsectors64 < UINT_MAX);
2386 nsectors = (u32)nsectors64;
2388 if (offset + nsectors <= sparity->nsectors) {
2389 bitmap_set(bitmap, offset, nsectors);
2393 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2394 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2397 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2400 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2403 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2406 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2409 static void scrub_block_complete(struct scrub_block *sblock)
2413 if (!sblock->no_io_error_seen) {
2415 scrub_handle_errored_block(sblock);
2418 * if has checksum error, write via repair mechanism in
2419 * dev replace case, otherwise write here in dev replace
2422 corrupted = scrub_checksum(sblock);
2423 if (!corrupted && sblock->sctx->is_dev_replace)
2424 scrub_write_block_to_dev_replace(sblock);
2427 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2428 u64 start = sblock->pagev[0]->logical;
2429 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2432 scrub_parity_mark_sectors_error(sblock->sparity,
2433 start, end - start);
2437 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2439 struct btrfs_ordered_sum *sum = NULL;
2440 unsigned long index;
2441 unsigned long num_sectors;
2443 while (!list_empty(&sctx->csum_list)) {
2444 sum = list_first_entry(&sctx->csum_list,
2445 struct btrfs_ordered_sum, list);
2446 if (sum->bytenr > logical)
2448 if (sum->bytenr + sum->len > logical)
2451 ++sctx->stat.csum_discards;
2452 list_del(&sum->list);
2459 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2460 ASSERT(index < UINT_MAX);
2462 num_sectors = sum->len / sctx->fs_info->sectorsize;
2463 memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2464 if (index == num_sectors - 1) {
2465 list_del(&sum->list);
2471 /* scrub extent tries to collect up to 64 kB for each bio */
2472 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2473 u64 logical, u64 len,
2474 u64 physical, struct btrfs_device *dev, u64 flags,
2475 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2478 u8 csum[BTRFS_CSUM_SIZE];
2481 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2482 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2483 blocksize = map->stripe_len;
2485 blocksize = sctx->fs_info->sectorsize;
2486 spin_lock(&sctx->stat_lock);
2487 sctx->stat.data_extents_scrubbed++;
2488 sctx->stat.data_bytes_scrubbed += len;
2489 spin_unlock(&sctx->stat_lock);
2490 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2491 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2492 blocksize = map->stripe_len;
2494 blocksize = sctx->fs_info->nodesize;
2495 spin_lock(&sctx->stat_lock);
2496 sctx->stat.tree_extents_scrubbed++;
2497 sctx->stat.tree_bytes_scrubbed += len;
2498 spin_unlock(&sctx->stat_lock);
2500 blocksize = sctx->fs_info->sectorsize;
2505 u64 l = min_t(u64, len, blocksize);
2508 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2509 /* push csums to sbio */
2510 have_csum = scrub_find_csum(sctx, logical, csum);
2512 ++sctx->stat.no_csum;
2514 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2515 mirror_num, have_csum ? csum : NULL, 0,
2516 physical_for_dev_replace);
2522 physical_for_dev_replace += l;
2527 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2528 u64 logical, u64 len,
2529 u64 physical, struct btrfs_device *dev,
2530 u64 flags, u64 gen, int mirror_num, u8 *csum)
2532 struct scrub_ctx *sctx = sparity->sctx;
2533 struct scrub_block *sblock;
2536 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2538 spin_lock(&sctx->stat_lock);
2539 sctx->stat.malloc_errors++;
2540 spin_unlock(&sctx->stat_lock);
2544 /* one ref inside this function, plus one for each page added to
2546 refcount_set(&sblock->refs, 1);
2547 sblock->sctx = sctx;
2548 sblock->no_io_error_seen = 1;
2549 sblock->sparity = sparity;
2550 scrub_parity_get(sparity);
2552 for (index = 0; len > 0; index++) {
2553 struct scrub_page *spage;
2554 u64 l = min_t(u64, len, PAGE_SIZE);
2556 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2559 spin_lock(&sctx->stat_lock);
2560 sctx->stat.malloc_errors++;
2561 spin_unlock(&sctx->stat_lock);
2562 scrub_block_put(sblock);
2565 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2566 /* For scrub block */
2567 scrub_page_get(spage);
2568 sblock->pagev[index] = spage;
2569 /* For scrub parity */
2570 scrub_page_get(spage);
2571 list_add_tail(&spage->list, &sparity->spages);
2572 spage->sblock = sblock;
2574 spage->flags = flags;
2575 spage->generation = gen;
2576 spage->logical = logical;
2577 spage->physical = physical;
2578 spage->mirror_num = mirror_num;
2580 spage->have_csum = 1;
2581 memcpy(spage->csum, csum, sctx->csum_size);
2583 spage->have_csum = 0;
2585 sblock->page_count++;
2586 spage->page = alloc_page(GFP_KERNEL);
2594 WARN_ON(sblock->page_count == 0);
2595 for (index = 0; index < sblock->page_count; index++) {
2596 struct scrub_page *spage = sblock->pagev[index];
2599 ret = scrub_add_page_to_rd_bio(sctx, spage);
2601 scrub_block_put(sblock);
2606 /* last one frees, either here or in bio completion for last page */
2607 scrub_block_put(sblock);
2611 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2612 u64 logical, u64 len,
2613 u64 physical, struct btrfs_device *dev,
2614 u64 flags, u64 gen, int mirror_num)
2616 struct scrub_ctx *sctx = sparity->sctx;
2618 u8 csum[BTRFS_CSUM_SIZE];
2621 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2622 scrub_parity_mark_sectors_error(sparity, logical, len);
2626 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2627 blocksize = sparity->stripe_len;
2628 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2629 blocksize = sparity->stripe_len;
2631 blocksize = sctx->fs_info->sectorsize;
2636 u64 l = min_t(u64, len, blocksize);
2639 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2640 /* push csums to sbio */
2641 have_csum = scrub_find_csum(sctx, logical, csum);
2645 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2646 flags, gen, mirror_num,
2647 have_csum ? csum : NULL);
2659 * Given a physical address, this will calculate it's
2660 * logical offset. if this is a parity stripe, it will return
2661 * the most left data stripe's logical offset.
2663 * return 0 if it is a data stripe, 1 means parity stripe.
2665 static int get_raid56_logic_offset(u64 physical, int num,
2666 struct map_lookup *map, u64 *offset,
2675 const int data_stripes = nr_data_stripes(map);
2677 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2679 *stripe_start = last_offset;
2681 *offset = last_offset;
2682 for (i = 0; i < data_stripes; i++) {
2683 *offset = last_offset + i * map->stripe_len;
2685 stripe_nr = div64_u64(*offset, map->stripe_len);
2686 stripe_nr = div_u64(stripe_nr, data_stripes);
2688 /* Work out the disk rotation on this stripe-set */
2689 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2690 /* calculate which stripe this data locates */
2692 stripe_index = rot % map->num_stripes;
2693 if (stripe_index == num)
2695 if (stripe_index < num)
2698 *offset = last_offset + j * map->stripe_len;
2702 static void scrub_free_parity(struct scrub_parity *sparity)
2704 struct scrub_ctx *sctx = sparity->sctx;
2705 struct scrub_page *curr, *next;
2708 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2710 spin_lock(&sctx->stat_lock);
2711 sctx->stat.read_errors += nbits;
2712 sctx->stat.uncorrectable_errors += nbits;
2713 spin_unlock(&sctx->stat_lock);
2716 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2717 list_del_init(&curr->list);
2718 scrub_page_put(curr);
2724 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2726 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2728 struct scrub_ctx *sctx = sparity->sctx;
2730 scrub_free_parity(sparity);
2731 scrub_pending_bio_dec(sctx);
2734 static void scrub_parity_bio_endio(struct bio *bio)
2736 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2737 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2740 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2745 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2746 scrub_parity_bio_endio_worker, NULL, NULL);
2747 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2750 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2752 struct scrub_ctx *sctx = sparity->sctx;
2753 struct btrfs_fs_info *fs_info = sctx->fs_info;
2755 struct btrfs_raid_bio *rbio;
2756 struct btrfs_bio *bbio = NULL;
2760 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2764 length = sparity->logic_end - sparity->logic_start;
2766 btrfs_bio_counter_inc_blocked(fs_info);
2767 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2769 if (ret || !bbio || !bbio->raid_map)
2772 bio = btrfs_io_bio_alloc(0);
2773 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2774 bio->bi_private = sparity;
2775 bio->bi_end_io = scrub_parity_bio_endio;
2777 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2778 length, sparity->scrub_dev,
2784 scrub_pending_bio_inc(sctx);
2785 raid56_parity_submit_scrub_rbio(rbio);
2791 btrfs_bio_counter_dec(fs_info);
2792 btrfs_put_bbio(bbio);
2793 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2795 spin_lock(&sctx->stat_lock);
2796 sctx->stat.malloc_errors++;
2797 spin_unlock(&sctx->stat_lock);
2799 scrub_free_parity(sparity);
2802 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2804 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2807 static void scrub_parity_get(struct scrub_parity *sparity)
2809 refcount_inc(&sparity->refs);
2812 static void scrub_parity_put(struct scrub_parity *sparity)
2814 if (!refcount_dec_and_test(&sparity->refs))
2817 scrub_parity_check_and_repair(sparity);
2820 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2821 struct map_lookup *map,
2822 struct btrfs_device *sdev,
2823 struct btrfs_path *path,
2827 struct btrfs_fs_info *fs_info = sctx->fs_info;
2828 struct btrfs_root *root = fs_info->extent_root;
2829 struct btrfs_root *csum_root = fs_info->csum_root;
2830 struct btrfs_extent_item *extent;
2831 struct btrfs_bio *bbio = NULL;
2835 struct extent_buffer *l;
2836 struct btrfs_key key;
2839 u64 extent_physical;
2842 struct btrfs_device *extent_dev;
2843 struct scrub_parity *sparity;
2846 int extent_mirror_num;
2849 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2850 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2851 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2854 spin_lock(&sctx->stat_lock);
2855 sctx->stat.malloc_errors++;
2856 spin_unlock(&sctx->stat_lock);
2860 sparity->stripe_len = map->stripe_len;
2861 sparity->nsectors = nsectors;
2862 sparity->sctx = sctx;
2863 sparity->scrub_dev = sdev;
2864 sparity->logic_start = logic_start;
2865 sparity->logic_end = logic_end;
2866 refcount_set(&sparity->refs, 1);
2867 INIT_LIST_HEAD(&sparity->spages);
2868 sparity->dbitmap = sparity->bitmap;
2869 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2872 while (logic_start < logic_end) {
2873 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2874 key.type = BTRFS_METADATA_ITEM_KEY;
2876 key.type = BTRFS_EXTENT_ITEM_KEY;
2877 key.objectid = logic_start;
2878 key.offset = (u64)-1;
2880 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2885 ret = btrfs_previous_extent_item(root, path, 0);
2889 btrfs_release_path(path);
2890 ret = btrfs_search_slot(NULL, root, &key,
2902 slot = path->slots[0];
2903 if (slot >= btrfs_header_nritems(l)) {
2904 ret = btrfs_next_leaf(root, path);
2913 btrfs_item_key_to_cpu(l, &key, slot);
2915 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2916 key.type != BTRFS_METADATA_ITEM_KEY)
2919 if (key.type == BTRFS_METADATA_ITEM_KEY)
2920 bytes = fs_info->nodesize;
2924 if (key.objectid + bytes <= logic_start)
2927 if (key.objectid >= logic_end) {
2932 while (key.objectid >= logic_start + map->stripe_len)
2933 logic_start += map->stripe_len;
2935 extent = btrfs_item_ptr(l, slot,
2936 struct btrfs_extent_item);
2937 flags = btrfs_extent_flags(l, extent);
2938 generation = btrfs_extent_generation(l, extent);
2940 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2941 (key.objectid < logic_start ||
2942 key.objectid + bytes >
2943 logic_start + map->stripe_len)) {
2945 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2946 key.objectid, logic_start);
2947 spin_lock(&sctx->stat_lock);
2948 sctx->stat.uncorrectable_errors++;
2949 spin_unlock(&sctx->stat_lock);
2953 extent_logical = key.objectid;
2956 if (extent_logical < logic_start) {
2957 extent_len -= logic_start - extent_logical;
2958 extent_logical = logic_start;
2961 if (extent_logical + extent_len >
2962 logic_start + map->stripe_len)
2963 extent_len = logic_start + map->stripe_len -
2966 scrub_parity_mark_sectors_data(sparity, extent_logical,
2969 mapped_length = extent_len;
2971 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2972 extent_logical, &mapped_length, &bbio,
2975 if (!bbio || mapped_length < extent_len)
2979 btrfs_put_bbio(bbio);
2982 extent_physical = bbio->stripes[0].physical;
2983 extent_mirror_num = bbio->mirror_num;
2984 extent_dev = bbio->stripes[0].dev;
2985 btrfs_put_bbio(bbio);
2987 ret = btrfs_lookup_csums_range(csum_root,
2989 extent_logical + extent_len - 1,
2990 &sctx->csum_list, 1);
2994 ret = scrub_extent_for_parity(sparity, extent_logical,
3001 scrub_free_csums(sctx);
3006 if (extent_logical + extent_len <
3007 key.objectid + bytes) {
3008 logic_start += map->stripe_len;
3010 if (logic_start >= logic_end) {
3015 if (logic_start < key.objectid + bytes) {
3024 btrfs_release_path(path);
3029 logic_start += map->stripe_len;
3033 scrub_parity_mark_sectors_error(sparity, logic_start,
3034 logic_end - logic_start);
3035 scrub_parity_put(sparity);
3037 mutex_lock(&sctx->wr_lock);
3038 scrub_wr_submit(sctx);
3039 mutex_unlock(&sctx->wr_lock);
3041 btrfs_release_path(path);
3042 return ret < 0 ? ret : 0;
3045 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3046 struct map_lookup *map,
3047 struct btrfs_device *scrub_dev,
3048 int num, u64 base, u64 length)
3050 struct btrfs_path *path, *ppath;
3051 struct btrfs_fs_info *fs_info = sctx->fs_info;
3052 struct btrfs_root *root = fs_info->extent_root;
3053 struct btrfs_root *csum_root = fs_info->csum_root;
3054 struct btrfs_extent_item *extent;
3055 struct blk_plug plug;
3060 struct extent_buffer *l;
3067 struct reada_control *reada1;
3068 struct reada_control *reada2;
3069 struct btrfs_key key;
3070 struct btrfs_key key_end;
3071 u64 increment = map->stripe_len;
3074 u64 extent_physical;
3078 struct btrfs_device *extent_dev;
3079 int extent_mirror_num;
3082 physical = map->stripes[num].physical;
3084 nstripes = div64_u64(length, map->stripe_len);
3085 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3086 offset = map->stripe_len * num;
3087 increment = map->stripe_len * map->num_stripes;
3089 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3090 int factor = map->num_stripes / map->sub_stripes;
3091 offset = map->stripe_len * (num / map->sub_stripes);
3092 increment = map->stripe_len * factor;
3093 mirror_num = num % map->sub_stripes + 1;
3094 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3095 increment = map->stripe_len;
3096 mirror_num = num % map->num_stripes + 1;
3097 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3098 increment = map->stripe_len;
3099 mirror_num = num % map->num_stripes + 1;
3100 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3101 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3102 increment = map->stripe_len * nr_data_stripes(map);
3105 increment = map->stripe_len;
3109 path = btrfs_alloc_path();
3113 ppath = btrfs_alloc_path();
3115 btrfs_free_path(path);
3120 * work on commit root. The related disk blocks are static as
3121 * long as COW is applied. This means, it is save to rewrite
3122 * them to repair disk errors without any race conditions
3124 path->search_commit_root = 1;
3125 path->skip_locking = 1;
3127 ppath->search_commit_root = 1;
3128 ppath->skip_locking = 1;
3130 * trigger the readahead for extent tree csum tree and wait for
3131 * completion. During readahead, the scrub is officially paused
3132 * to not hold off transaction commits
3134 logical = base + offset;
3135 physical_end = physical + nstripes * map->stripe_len;
3136 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3137 get_raid56_logic_offset(physical_end, num,
3138 map, &logic_end, NULL);
3141 logic_end = logical + increment * nstripes;
3143 wait_event(sctx->list_wait,
3144 atomic_read(&sctx->bios_in_flight) == 0);
3145 scrub_blocked_if_needed(fs_info);
3147 /* FIXME it might be better to start readahead at commit root */
3148 key.objectid = logical;
3149 key.type = BTRFS_EXTENT_ITEM_KEY;
3150 key.offset = (u64)0;
3151 key_end.objectid = logic_end;
3152 key_end.type = BTRFS_METADATA_ITEM_KEY;
3153 key_end.offset = (u64)-1;
3154 reada1 = btrfs_reada_add(root, &key, &key_end);
3156 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3157 key.type = BTRFS_EXTENT_CSUM_KEY;
3158 key.offset = logical;
3159 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3160 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3161 key_end.offset = logic_end;
3162 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3164 if (!IS_ERR(reada1))
3165 btrfs_reada_wait(reada1);
3166 if (!IS_ERR(reada2))
3167 btrfs_reada_wait(reada2);
3171 * collect all data csums for the stripe to avoid seeking during
3172 * the scrub. This might currently (crc32) end up to be about 1MB
3174 blk_start_plug(&plug);
3177 * now find all extents for each stripe and scrub them
3180 while (physical < physical_end) {
3184 if (atomic_read(&fs_info->scrub_cancel_req) ||
3185 atomic_read(&sctx->cancel_req)) {
3190 * check to see if we have to pause
3192 if (atomic_read(&fs_info->scrub_pause_req)) {
3193 /* push queued extents */
3194 sctx->flush_all_writes = true;
3196 mutex_lock(&sctx->wr_lock);
3197 scrub_wr_submit(sctx);
3198 mutex_unlock(&sctx->wr_lock);
3199 wait_event(sctx->list_wait,
3200 atomic_read(&sctx->bios_in_flight) == 0);
3201 sctx->flush_all_writes = false;
3202 scrub_blocked_if_needed(fs_info);
3205 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3206 ret = get_raid56_logic_offset(physical, num, map,
3211 /* it is parity strip */
3212 stripe_logical += base;
3213 stripe_end = stripe_logical + increment;
3214 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3215 ppath, stripe_logical,
3223 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3224 key.type = BTRFS_METADATA_ITEM_KEY;
3226 key.type = BTRFS_EXTENT_ITEM_KEY;
3227 key.objectid = logical;
3228 key.offset = (u64)-1;
3230 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3235 ret = btrfs_previous_extent_item(root, path, 0);
3239 /* there's no smaller item, so stick with the
3241 btrfs_release_path(path);
3242 ret = btrfs_search_slot(NULL, root, &key,
3254 slot = path->slots[0];
3255 if (slot >= btrfs_header_nritems(l)) {
3256 ret = btrfs_next_leaf(root, path);
3265 btrfs_item_key_to_cpu(l, &key, slot);
3267 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3268 key.type != BTRFS_METADATA_ITEM_KEY)
3271 if (key.type == BTRFS_METADATA_ITEM_KEY)
3272 bytes = fs_info->nodesize;
3276 if (key.objectid + bytes <= logical)
3279 if (key.objectid >= logical + map->stripe_len) {
3280 /* out of this device extent */
3281 if (key.objectid >= logic_end)
3286 extent = btrfs_item_ptr(l, slot,
3287 struct btrfs_extent_item);
3288 flags = btrfs_extent_flags(l, extent);
3289 generation = btrfs_extent_generation(l, extent);
3291 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3292 (key.objectid < logical ||
3293 key.objectid + bytes >
3294 logical + map->stripe_len)) {
3296 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3297 key.objectid, logical);
3298 spin_lock(&sctx->stat_lock);
3299 sctx->stat.uncorrectable_errors++;
3300 spin_unlock(&sctx->stat_lock);
3305 extent_logical = key.objectid;
3309 * trim extent to this stripe
3311 if (extent_logical < logical) {
3312 extent_len -= logical - extent_logical;
3313 extent_logical = logical;
3315 if (extent_logical + extent_len >
3316 logical + map->stripe_len) {
3317 extent_len = logical + map->stripe_len -
3321 extent_physical = extent_logical - logical + physical;
3322 extent_dev = scrub_dev;
3323 extent_mirror_num = mirror_num;
3324 if (sctx->is_dev_replace)
3325 scrub_remap_extent(fs_info, extent_logical,
3326 extent_len, &extent_physical,
3328 &extent_mirror_num);
3330 ret = btrfs_lookup_csums_range(csum_root,
3334 &sctx->csum_list, 1);
3338 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3339 extent_physical, extent_dev, flags,
3340 generation, extent_mirror_num,
3341 extent_logical - logical + physical);
3343 scrub_free_csums(sctx);
3348 if (extent_logical + extent_len <
3349 key.objectid + bytes) {
3350 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3352 * loop until we find next data stripe
3353 * or we have finished all stripes.
3356 physical += map->stripe_len;
3357 ret = get_raid56_logic_offset(physical,
3362 if (ret && physical < physical_end) {
3363 stripe_logical += base;
3364 stripe_end = stripe_logical +
3366 ret = scrub_raid56_parity(sctx,
3367 map, scrub_dev, ppath,
3375 physical += map->stripe_len;
3376 logical += increment;
3378 if (logical < key.objectid + bytes) {
3383 if (physical >= physical_end) {
3391 btrfs_release_path(path);
3393 logical += increment;
3394 physical += map->stripe_len;
3395 spin_lock(&sctx->stat_lock);
3397 sctx->stat.last_physical = map->stripes[num].physical +
3400 sctx->stat.last_physical = physical;
3401 spin_unlock(&sctx->stat_lock);
3406 /* push queued extents */
3408 mutex_lock(&sctx->wr_lock);
3409 scrub_wr_submit(sctx);
3410 mutex_unlock(&sctx->wr_lock);
3412 blk_finish_plug(&plug);
3413 btrfs_free_path(path);
3414 btrfs_free_path(ppath);
3415 return ret < 0 ? ret : 0;
3418 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3419 struct btrfs_device *scrub_dev,
3420 u64 chunk_offset, u64 length,
3422 struct btrfs_block_group_cache *cache)
3424 struct btrfs_fs_info *fs_info = sctx->fs_info;
3425 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3426 struct map_lookup *map;
3427 struct extent_map *em;
3431 read_lock(&map_tree->lock);
3432 em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3433 read_unlock(&map_tree->lock);
3437 * Might have been an unused block group deleted by the cleaner
3438 * kthread or relocation.
3440 spin_lock(&cache->lock);
3441 if (!cache->removed)
3443 spin_unlock(&cache->lock);
3448 map = em->map_lookup;
3449 if (em->start != chunk_offset)
3452 if (em->len < length)
3455 for (i = 0; i < map->num_stripes; ++i) {
3456 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3457 map->stripes[i].physical == dev_offset) {
3458 ret = scrub_stripe(sctx, map, scrub_dev, i,
3459 chunk_offset, length);
3465 free_extent_map(em);
3470 static noinline_for_stack
3471 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3472 struct btrfs_device *scrub_dev, u64 start, u64 end)
3474 struct btrfs_dev_extent *dev_extent = NULL;
3475 struct btrfs_path *path;
3476 struct btrfs_fs_info *fs_info = sctx->fs_info;
3477 struct btrfs_root *root = fs_info->dev_root;
3483 struct extent_buffer *l;
3484 struct btrfs_key key;
3485 struct btrfs_key found_key;
3486 struct btrfs_block_group_cache *cache;
3487 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3489 path = btrfs_alloc_path();
3493 path->reada = READA_FORWARD;
3494 path->search_commit_root = 1;
3495 path->skip_locking = 1;
3497 key.objectid = scrub_dev->devid;
3499 key.type = BTRFS_DEV_EXTENT_KEY;
3502 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3506 if (path->slots[0] >=
3507 btrfs_header_nritems(path->nodes[0])) {
3508 ret = btrfs_next_leaf(root, path);
3521 slot = path->slots[0];
3523 btrfs_item_key_to_cpu(l, &found_key, slot);
3525 if (found_key.objectid != scrub_dev->devid)
3528 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3531 if (found_key.offset >= end)
3534 if (found_key.offset < key.offset)
3537 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3538 length = btrfs_dev_extent_length(l, dev_extent);
3540 if (found_key.offset + length <= start)
3543 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3546 * get a reference on the corresponding block group to prevent
3547 * the chunk from going away while we scrub it
3549 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3551 /* some chunks are removed but not committed to disk yet,
3552 * continue scrubbing */
3557 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3558 * to avoid deadlock caused by:
3559 * btrfs_inc_block_group_ro()
3560 * -> btrfs_wait_for_commit()
3561 * -> btrfs_commit_transaction()
3562 * -> btrfs_scrub_pause()
3564 scrub_pause_on(fs_info);
3565 ret = btrfs_inc_block_group_ro(cache);
3566 if (!ret && sctx->is_dev_replace) {
3568 * If we are doing a device replace wait for any tasks
3569 * that started delalloc right before we set the block
3570 * group to RO mode, as they might have just allocated
3571 * an extent from it or decided they could do a nocow
3572 * write. And if any such tasks did that, wait for their
3573 * ordered extents to complete and then commit the
3574 * current transaction, so that we can later see the new
3575 * extent items in the extent tree - the ordered extents
3576 * create delayed data references (for cow writes) when
3577 * they complete, which will be run and insert the
3578 * corresponding extent items into the extent tree when
3579 * we commit the transaction they used when running
3580 * inode.c:btrfs_finish_ordered_io(). We later use
3581 * the commit root of the extent tree to find extents
3582 * to copy from the srcdev into the tgtdev, and we don't
3583 * want to miss any new extents.
3585 btrfs_wait_block_group_reservations(cache);
3586 btrfs_wait_nocow_writers(cache);
3587 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3588 cache->key.objectid,
3591 struct btrfs_trans_handle *trans;
3593 trans = btrfs_join_transaction(root);
3595 ret = PTR_ERR(trans);
3597 ret = btrfs_commit_transaction(trans);
3599 scrub_pause_off(fs_info);
3600 btrfs_put_block_group(cache);
3605 scrub_pause_off(fs_info);
3609 } else if (ret == -ENOSPC) {
3611 * btrfs_inc_block_group_ro return -ENOSPC when it
3612 * failed in creating new chunk for metadata.
3613 * It is not a problem for scrub/replace, because
3614 * metadata are always cowed, and our scrub paused
3615 * commit_transactions.
3620 "failed setting block group ro: %d", ret);
3621 btrfs_put_block_group(cache);
3625 down_write(&fs_info->dev_replace.rwsem);
3626 dev_replace->cursor_right = found_key.offset + length;
3627 dev_replace->cursor_left = found_key.offset;
3628 dev_replace->item_needs_writeback = 1;
3629 up_write(&dev_replace->rwsem);
3631 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3632 found_key.offset, cache);
3635 * flush, submit all pending read and write bios, afterwards
3637 * Note that in the dev replace case, a read request causes
3638 * write requests that are submitted in the read completion
3639 * worker. Therefore in the current situation, it is required
3640 * that all write requests are flushed, so that all read and
3641 * write requests are really completed when bios_in_flight
3644 sctx->flush_all_writes = true;
3646 mutex_lock(&sctx->wr_lock);
3647 scrub_wr_submit(sctx);
3648 mutex_unlock(&sctx->wr_lock);
3650 wait_event(sctx->list_wait,
3651 atomic_read(&sctx->bios_in_flight) == 0);
3653 scrub_pause_on(fs_info);
3656 * must be called before we decrease @scrub_paused.
3657 * make sure we don't block transaction commit while
3658 * we are waiting pending workers finished.
3660 wait_event(sctx->list_wait,
3661 atomic_read(&sctx->workers_pending) == 0);
3662 sctx->flush_all_writes = false;
3664 scrub_pause_off(fs_info);
3666 down_write(&fs_info->dev_replace.rwsem);
3667 dev_replace->cursor_left = dev_replace->cursor_right;
3668 dev_replace->item_needs_writeback = 1;
3669 up_write(&fs_info->dev_replace.rwsem);
3672 btrfs_dec_block_group_ro(cache);
3675 * We might have prevented the cleaner kthread from deleting
3676 * this block group if it was already unused because we raced
3677 * and set it to RO mode first. So add it back to the unused
3678 * list, otherwise it might not ever be deleted unless a manual
3679 * balance is triggered or it becomes used and unused again.
3681 spin_lock(&cache->lock);
3682 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3683 btrfs_block_group_used(&cache->item) == 0) {
3684 spin_unlock(&cache->lock);
3685 btrfs_mark_bg_unused(cache);
3687 spin_unlock(&cache->lock);
3690 btrfs_put_block_group(cache);
3693 if (sctx->is_dev_replace &&
3694 atomic64_read(&dev_replace->num_write_errors) > 0) {
3698 if (sctx->stat.malloc_errors > 0) {
3703 key.offset = found_key.offset + length;
3704 btrfs_release_path(path);
3707 btrfs_free_path(path);
3712 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3713 struct btrfs_device *scrub_dev)
3719 struct btrfs_fs_info *fs_info = sctx->fs_info;
3721 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3724 /* Seed devices of a new filesystem has their own generation. */
3725 if (scrub_dev->fs_devices != fs_info->fs_devices)
3726 gen = scrub_dev->generation;
3728 gen = fs_info->last_trans_committed;
3730 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3731 bytenr = btrfs_sb_offset(i);
3732 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3733 scrub_dev->commit_total_bytes)
3736 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3737 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3742 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3748 * get a reference count on fs_info->scrub_workers. start worker if necessary
3750 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3753 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3754 int max_active = fs_info->thread_pool_size;
3756 lockdep_assert_held(&fs_info->scrub_lock);
3758 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3759 ASSERT(fs_info->scrub_workers == NULL);
3760 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3761 flags, is_dev_replace ? 1 : max_active, 4);
3762 if (!fs_info->scrub_workers)
3763 goto fail_scrub_workers;
3765 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3766 fs_info->scrub_wr_completion_workers =
3767 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3769 if (!fs_info->scrub_wr_completion_workers)
3770 goto fail_scrub_wr_completion_workers;
3772 ASSERT(fs_info->scrub_parity_workers == NULL);
3773 fs_info->scrub_parity_workers =
3774 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3776 if (!fs_info->scrub_parity_workers)
3777 goto fail_scrub_parity_workers;
3779 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3781 refcount_inc(&fs_info->scrub_workers_refcnt);
3785 fail_scrub_parity_workers:
3786 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3787 fail_scrub_wr_completion_workers:
3788 btrfs_destroy_workqueue(fs_info->scrub_workers);
3793 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3794 u64 end, struct btrfs_scrub_progress *progress,
3795 int readonly, int is_dev_replace)
3797 struct scrub_ctx *sctx;
3799 struct btrfs_device *dev;
3800 unsigned int nofs_flag;
3801 struct btrfs_workqueue *scrub_workers = NULL;
3802 struct btrfs_workqueue *scrub_wr_comp = NULL;
3803 struct btrfs_workqueue *scrub_parity = NULL;
3805 if (btrfs_fs_closing(fs_info))
3808 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3810 * in this case scrub is unable to calculate the checksum
3811 * the way scrub is implemented. Do not handle this
3812 * situation at all because it won't ever happen.
3815 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3821 if (fs_info->sectorsize != PAGE_SIZE) {
3822 /* not supported for data w/o checksums */
3823 btrfs_err_rl(fs_info,
3824 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3825 fs_info->sectorsize, PAGE_SIZE);
3829 if (fs_info->nodesize >
3830 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3831 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3833 * would exhaust the array bounds of pagev member in
3834 * struct scrub_block
3837 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3839 SCRUB_MAX_PAGES_PER_BLOCK,
3840 fs_info->sectorsize,
3841 SCRUB_MAX_PAGES_PER_BLOCK);
3845 /* Allocate outside of device_list_mutex */
3846 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3848 return PTR_ERR(sctx);
3850 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3851 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3852 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3854 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3859 if (!is_dev_replace && !readonly &&
3860 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3861 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3862 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3863 rcu_str_deref(dev->name));
3868 mutex_lock(&fs_info->scrub_lock);
3869 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3870 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3871 mutex_unlock(&fs_info->scrub_lock);
3872 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3877 down_read(&fs_info->dev_replace.rwsem);
3878 if (dev->scrub_ctx ||
3880 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3881 up_read(&fs_info->dev_replace.rwsem);
3882 mutex_unlock(&fs_info->scrub_lock);
3883 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3887 up_read(&fs_info->dev_replace.rwsem);
3889 ret = scrub_workers_get(fs_info, is_dev_replace);
3891 mutex_unlock(&fs_info->scrub_lock);
3892 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3896 sctx->readonly = readonly;
3897 dev->scrub_ctx = sctx;
3898 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3901 * checking @scrub_pause_req here, we can avoid
3902 * race between committing transaction and scrubbing.
3904 __scrub_blocked_if_needed(fs_info);
3905 atomic_inc(&fs_info->scrubs_running);
3906 mutex_unlock(&fs_info->scrub_lock);
3909 * In order to avoid deadlock with reclaim when there is a transaction
3910 * trying to pause scrub, make sure we use GFP_NOFS for all the
3911 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3912 * invoked by our callees. The pausing request is done when the
3913 * transaction commit starts, and it blocks the transaction until scrub
3914 * is paused (done at specific points at scrub_stripe() or right above
3915 * before incrementing fs_info->scrubs_running).
3917 nofs_flag = memalloc_nofs_save();
3918 if (!is_dev_replace) {
3919 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3921 * by holding device list mutex, we can
3922 * kick off writing super in log tree sync.
3924 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3925 ret = scrub_supers(sctx, dev);
3926 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3930 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3931 memalloc_nofs_restore(nofs_flag);
3933 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3934 atomic_dec(&fs_info->scrubs_running);
3935 wake_up(&fs_info->scrub_pause_wait);
3937 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3940 memcpy(progress, &sctx->stat, sizeof(*progress));
3942 if (!is_dev_replace)
3943 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3944 ret ? "not finished" : "finished", devid, ret);
3946 mutex_lock(&fs_info->scrub_lock);
3947 dev->scrub_ctx = NULL;
3948 if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3949 scrub_workers = fs_info->scrub_workers;
3950 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3951 scrub_parity = fs_info->scrub_parity_workers;
3953 fs_info->scrub_workers = NULL;
3954 fs_info->scrub_wr_completion_workers = NULL;
3955 fs_info->scrub_parity_workers = NULL;
3957 mutex_unlock(&fs_info->scrub_lock);
3959 btrfs_destroy_workqueue(scrub_workers);
3960 btrfs_destroy_workqueue(scrub_wr_comp);
3961 btrfs_destroy_workqueue(scrub_parity);
3962 scrub_put_ctx(sctx);
3967 scrub_free_ctx(sctx);
3972 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3974 mutex_lock(&fs_info->scrub_lock);
3975 atomic_inc(&fs_info->scrub_pause_req);
3976 while (atomic_read(&fs_info->scrubs_paused) !=
3977 atomic_read(&fs_info->scrubs_running)) {
3978 mutex_unlock(&fs_info->scrub_lock);
3979 wait_event(fs_info->scrub_pause_wait,
3980 atomic_read(&fs_info->scrubs_paused) ==
3981 atomic_read(&fs_info->scrubs_running));
3982 mutex_lock(&fs_info->scrub_lock);
3984 mutex_unlock(&fs_info->scrub_lock);
3987 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3989 atomic_dec(&fs_info->scrub_pause_req);
3990 wake_up(&fs_info->scrub_pause_wait);
3993 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3995 mutex_lock(&fs_info->scrub_lock);
3996 if (!atomic_read(&fs_info->scrubs_running)) {
3997 mutex_unlock(&fs_info->scrub_lock);
4001 atomic_inc(&fs_info->scrub_cancel_req);
4002 while (atomic_read(&fs_info->scrubs_running)) {
4003 mutex_unlock(&fs_info->scrub_lock);
4004 wait_event(fs_info->scrub_pause_wait,
4005 atomic_read(&fs_info->scrubs_running) == 0);
4006 mutex_lock(&fs_info->scrub_lock);
4008 atomic_dec(&fs_info->scrub_cancel_req);
4009 mutex_unlock(&fs_info->scrub_lock);
4014 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4016 struct btrfs_fs_info *fs_info = dev->fs_info;
4017 struct scrub_ctx *sctx;
4019 mutex_lock(&fs_info->scrub_lock);
4020 sctx = dev->scrub_ctx;
4022 mutex_unlock(&fs_info->scrub_lock);
4025 atomic_inc(&sctx->cancel_req);
4026 while (dev->scrub_ctx) {
4027 mutex_unlock(&fs_info->scrub_lock);
4028 wait_event(fs_info->scrub_pause_wait,
4029 dev->scrub_ctx == NULL);
4030 mutex_lock(&fs_info->scrub_lock);
4032 mutex_unlock(&fs_info->scrub_lock);
4037 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4038 struct btrfs_scrub_progress *progress)
4040 struct btrfs_device *dev;
4041 struct scrub_ctx *sctx = NULL;
4043 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4044 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4046 sctx = dev->scrub_ctx;
4048 memcpy(progress, &sctx->stat, sizeof(*progress));
4049 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4051 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4054 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4055 u64 extent_logical, u64 extent_len,
4056 u64 *extent_physical,
4057 struct btrfs_device **extent_dev,
4058 int *extent_mirror_num)
4061 struct btrfs_bio *bbio = NULL;
4064 mapped_length = extent_len;
4065 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4066 &mapped_length, &bbio, 0);
4067 if (ret || !bbio || mapped_length < extent_len ||
4068 !bbio->stripes[0].dev->bdev) {
4069 btrfs_put_bbio(bbio);
4073 *extent_physical = bbio->stripes[0].physical;
4074 *extent_mirror_num = bbio->mirror_num;
4075 *extent_dev = bbio->stripes[0].dev;
4076 btrfs_put_bbio(bbio);