1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
9 #include <crypto/hash.h>
14 #include "ordered-data.h"
15 #include "transaction.h"
17 #include "extent_io.h"
18 #include "dev-replace.h"
19 #include "check-integrity.h"
20 #include "rcu-string.h"
22 #include "block-group.h"
25 * This is only the first step towards a full-features scrub. It reads all
26 * extent and super block and verifies the checksums. In case a bad checksum
27 * is found or the extent cannot be read, good data will be written back if
30 * Future enhancements:
31 * - In case an unrepairable extent is encountered, track which files are
32 * affected and report them
33 * - track and record media errors, throw out bad devices
34 * - add a mode to also read unallocated space
41 * the following three values only influence the performance.
42 * The last one configures the number of parallel and outstanding I/O
43 * operations. The first two values configure an upper limit for the number
44 * of (dynamically allocated) pages that are added to a bio.
46 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
47 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
48 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
51 * the following value times PAGE_SIZE needs to be large enough to match the
52 * largest node/leaf/sector size that shall be supported.
53 * Values larger than BTRFS_STRIPE_LEN are not supported.
55 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
57 struct scrub_recover {
59 struct btrfs_bio *bbio;
64 struct scrub_block *sblock;
66 struct btrfs_device *dev;
67 struct list_head list;
68 u64 flags; /* extent flags */
72 u64 physical_for_dev_replace;
75 unsigned int mirror_num:8;
76 unsigned int have_csum:1;
77 unsigned int io_error:1;
79 u8 csum[BTRFS_CSUM_SIZE];
81 struct scrub_recover *recover;
86 struct scrub_ctx *sctx;
87 struct btrfs_device *dev;
92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
95 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
99 struct btrfs_work work;
103 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
105 atomic_t outstanding_pages;
106 refcount_t refs; /* free mem on transition to zero */
107 struct scrub_ctx *sctx;
108 struct scrub_parity *sparity;
110 unsigned int header_error:1;
111 unsigned int checksum_error:1;
112 unsigned int no_io_error_seen:1;
113 unsigned int generation_error:1; /* also sets header_error */
115 /* The following is for the data used to check parity */
116 /* It is for the data with checksum */
117 unsigned int data_corrected:1;
119 struct btrfs_work work;
122 /* Used for the chunks with parity stripe such RAID5/6 */
123 struct scrub_parity {
124 struct scrub_ctx *sctx;
126 struct btrfs_device *scrub_dev;
138 struct list_head spages;
140 /* Work of parity check and repair */
141 struct btrfs_work work;
143 /* Mark the parity blocks which have data */
144 unsigned long *dbitmap;
147 * Mark the parity blocks which have data, but errors happen when
148 * read data or check data
150 unsigned long *ebitmap;
152 unsigned long bitmap[];
156 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
157 struct btrfs_fs_info *fs_info;
160 atomic_t bios_in_flight;
161 atomic_t workers_pending;
162 spinlock_t list_lock;
163 wait_queue_head_t list_wait;
165 struct list_head csum_list;
168 int pages_per_rd_bio;
172 struct scrub_bio *wr_curr_bio;
173 struct mutex wr_lock;
174 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
175 struct btrfs_device *wr_tgtdev;
176 bool flush_all_writes;
181 struct btrfs_scrub_progress stat;
182 spinlock_t stat_lock;
185 * Use a ref counter to avoid use-after-free issues. Scrub workers
186 * decrement bios_in_flight and workers_pending and then do a wakeup
187 * on the list_wait wait queue. We must ensure the main scrub task
188 * doesn't free the scrub context before or while the workers are
189 * doing the wakeup() call.
194 struct scrub_warning {
195 struct btrfs_path *path;
196 u64 extent_item_size;
200 struct btrfs_device *dev;
203 struct full_stripe_lock {
210 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
211 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
212 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
213 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
214 struct scrub_block *sblocks_for_recheck);
215 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
216 struct scrub_block *sblock,
217 int retry_failed_mirror);
218 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
219 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
220 struct scrub_block *sblock_good);
221 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
222 struct scrub_block *sblock_good,
223 int page_num, int force_write);
224 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
225 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
227 static int scrub_checksum_data(struct scrub_block *sblock);
228 static int scrub_checksum_tree_block(struct scrub_block *sblock);
229 static int scrub_checksum_super(struct scrub_block *sblock);
230 static void scrub_block_get(struct scrub_block *sblock);
231 static void scrub_block_put(struct scrub_block *sblock);
232 static void scrub_page_get(struct scrub_page *spage);
233 static void scrub_page_put(struct scrub_page *spage);
234 static void scrub_parity_get(struct scrub_parity *sparity);
235 static void scrub_parity_put(struct scrub_parity *sparity);
236 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
237 struct scrub_page *spage);
238 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
239 u64 physical, struct btrfs_device *dev, u64 flags,
240 u64 gen, int mirror_num, u8 *csum, int force,
241 u64 physical_for_dev_replace);
242 static void scrub_bio_end_io(struct bio *bio);
243 static void scrub_bio_end_io_worker(struct btrfs_work *work);
244 static void scrub_block_complete(struct scrub_block *sblock);
245 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
246 u64 extent_logical, u64 extent_len,
247 u64 *extent_physical,
248 struct btrfs_device **extent_dev,
249 int *extent_mirror_num);
250 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
251 struct scrub_page *spage);
252 static void scrub_wr_submit(struct scrub_ctx *sctx);
253 static void scrub_wr_bio_end_io(struct bio *bio);
254 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
255 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
256 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
257 static void scrub_put_ctx(struct scrub_ctx *sctx);
259 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
261 return page->recover &&
262 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
265 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
267 refcount_inc(&sctx->refs);
268 atomic_inc(&sctx->bios_in_flight);
271 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
273 atomic_dec(&sctx->bios_in_flight);
274 wake_up(&sctx->list_wait);
278 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
280 while (atomic_read(&fs_info->scrub_pause_req)) {
281 mutex_unlock(&fs_info->scrub_lock);
282 wait_event(fs_info->scrub_pause_wait,
283 atomic_read(&fs_info->scrub_pause_req) == 0);
284 mutex_lock(&fs_info->scrub_lock);
288 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
290 atomic_inc(&fs_info->scrubs_paused);
291 wake_up(&fs_info->scrub_pause_wait);
294 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
296 mutex_lock(&fs_info->scrub_lock);
297 __scrub_blocked_if_needed(fs_info);
298 atomic_dec(&fs_info->scrubs_paused);
299 mutex_unlock(&fs_info->scrub_lock);
301 wake_up(&fs_info->scrub_pause_wait);
304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
306 scrub_pause_on(fs_info);
307 scrub_pause_off(fs_info);
311 * Insert new full stripe lock into full stripe locks tree
313 * Return pointer to existing or newly inserted full_stripe_lock structure if
314 * everything works well.
315 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
317 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
320 static struct full_stripe_lock *insert_full_stripe_lock(
321 struct btrfs_full_stripe_locks_tree *locks_root,
325 struct rb_node *parent = NULL;
326 struct full_stripe_lock *entry;
327 struct full_stripe_lock *ret;
329 lockdep_assert_held(&locks_root->lock);
331 p = &locks_root->root.rb_node;
334 entry = rb_entry(parent, struct full_stripe_lock, node);
335 if (fstripe_logical < entry->logical) {
337 } else if (fstripe_logical > entry->logical) {
348 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
350 return ERR_PTR(-ENOMEM);
351 ret->logical = fstripe_logical;
353 mutex_init(&ret->mutex);
355 rb_link_node(&ret->node, parent, p);
356 rb_insert_color(&ret->node, &locks_root->root);
361 * Search for a full stripe lock of a block group
363 * Return pointer to existing full stripe lock if found
364 * Return NULL if not found
366 static struct full_stripe_lock *search_full_stripe_lock(
367 struct btrfs_full_stripe_locks_tree *locks_root,
370 struct rb_node *node;
371 struct full_stripe_lock *entry;
373 lockdep_assert_held(&locks_root->lock);
375 node = locks_root->root.rb_node;
377 entry = rb_entry(node, struct full_stripe_lock, node);
378 if (fstripe_logical < entry->logical)
379 node = node->rb_left;
380 else if (fstripe_logical > entry->logical)
381 node = node->rb_right;
389 * Helper to get full stripe logical from a normal bytenr.
391 * Caller must ensure @cache is a RAID56 block group.
393 static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
398 * Due to chunk item size limit, full stripe length should not be
399 * larger than U32_MAX. Just a sanity check here.
401 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
404 * round_down() can only handle power of 2, while RAID56 full
405 * stripe length can be 64KiB * n, so we need to manually round down.
407 ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
408 cache->full_stripe_len + cache->start;
413 * Lock a full stripe to avoid concurrency of recovery and read
415 * It's only used for profiles with parities (RAID5/6), for other profiles it
418 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
419 * So caller must call unlock_full_stripe() at the same context.
421 * Return <0 if encounters error.
423 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
426 struct btrfs_block_group *bg_cache;
427 struct btrfs_full_stripe_locks_tree *locks_root;
428 struct full_stripe_lock *existing;
433 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
439 /* Profiles not based on parity don't need full stripe lock */
440 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
442 locks_root = &bg_cache->full_stripe_locks_root;
444 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
446 /* Now insert the full stripe lock */
447 mutex_lock(&locks_root->lock);
448 existing = insert_full_stripe_lock(locks_root, fstripe_start);
449 mutex_unlock(&locks_root->lock);
450 if (IS_ERR(existing)) {
451 ret = PTR_ERR(existing);
454 mutex_lock(&existing->mutex);
457 btrfs_put_block_group(bg_cache);
462 * Unlock a full stripe.
464 * NOTE: Caller must ensure it's the same context calling corresponding
465 * lock_full_stripe().
467 * Return 0 if we unlock full stripe without problem.
468 * Return <0 for error
470 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
473 struct btrfs_block_group *bg_cache;
474 struct btrfs_full_stripe_locks_tree *locks_root;
475 struct full_stripe_lock *fstripe_lock;
480 /* If we didn't acquire full stripe lock, no need to continue */
484 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
489 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
492 locks_root = &bg_cache->full_stripe_locks_root;
493 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
495 mutex_lock(&locks_root->lock);
496 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
497 /* Unpaired unlock_full_stripe() detected */
501 mutex_unlock(&locks_root->lock);
505 if (fstripe_lock->refs == 0) {
507 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
508 fstripe_lock->logical);
510 fstripe_lock->refs--;
513 if (fstripe_lock->refs == 0) {
514 rb_erase(&fstripe_lock->node, &locks_root->root);
517 mutex_unlock(&locks_root->lock);
519 mutex_unlock(&fstripe_lock->mutex);
523 btrfs_put_block_group(bg_cache);
527 static void scrub_free_csums(struct scrub_ctx *sctx)
529 while (!list_empty(&sctx->csum_list)) {
530 struct btrfs_ordered_sum *sum;
531 sum = list_first_entry(&sctx->csum_list,
532 struct btrfs_ordered_sum, list);
533 list_del(&sum->list);
538 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
545 /* this can happen when scrub is cancelled */
546 if (sctx->curr != -1) {
547 struct scrub_bio *sbio = sctx->bios[sctx->curr];
549 for (i = 0; i < sbio->page_count; i++) {
550 WARN_ON(!sbio->pagev[i]->page);
551 scrub_block_put(sbio->pagev[i]->sblock);
556 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
557 struct scrub_bio *sbio = sctx->bios[i];
564 kfree(sctx->wr_curr_bio);
565 scrub_free_csums(sctx);
569 static void scrub_put_ctx(struct scrub_ctx *sctx)
571 if (refcount_dec_and_test(&sctx->refs))
572 scrub_free_ctx(sctx);
575 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
576 struct btrfs_fs_info *fs_info, int is_dev_replace)
578 struct scrub_ctx *sctx;
581 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
584 refcount_set(&sctx->refs, 1);
585 sctx->is_dev_replace = is_dev_replace;
586 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
588 sctx->fs_info = fs_info;
589 INIT_LIST_HEAD(&sctx->csum_list);
590 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
591 struct scrub_bio *sbio;
593 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
596 sctx->bios[i] = sbio;
600 sbio->page_count = 0;
601 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
604 if (i != SCRUB_BIOS_PER_SCTX - 1)
605 sctx->bios[i]->next_free = i + 1;
607 sctx->bios[i]->next_free = -1;
609 sctx->first_free = 0;
610 atomic_set(&sctx->bios_in_flight, 0);
611 atomic_set(&sctx->workers_pending, 0);
612 atomic_set(&sctx->cancel_req, 0);
613 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
615 spin_lock_init(&sctx->list_lock);
616 spin_lock_init(&sctx->stat_lock);
617 init_waitqueue_head(&sctx->list_wait);
619 WARN_ON(sctx->wr_curr_bio != NULL);
620 mutex_init(&sctx->wr_lock);
621 sctx->wr_curr_bio = NULL;
622 if (is_dev_replace) {
623 WARN_ON(!fs_info->dev_replace.tgtdev);
624 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
625 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
626 sctx->flush_all_writes = false;
632 scrub_free_ctx(sctx);
633 return ERR_PTR(-ENOMEM);
636 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
644 struct extent_buffer *eb;
645 struct btrfs_inode_item *inode_item;
646 struct scrub_warning *swarn = warn_ctx;
647 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
648 struct inode_fs_paths *ipath = NULL;
649 struct btrfs_root *local_root;
650 struct btrfs_key key;
652 local_root = btrfs_get_fs_root(fs_info, root, true);
653 if (IS_ERR(local_root)) {
654 ret = PTR_ERR(local_root);
659 * this makes the path point to (inum INODE_ITEM ioff)
662 key.type = BTRFS_INODE_ITEM_KEY;
665 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
667 btrfs_put_root(local_root);
668 btrfs_release_path(swarn->path);
672 eb = swarn->path->nodes[0];
673 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
674 struct btrfs_inode_item);
675 isize = btrfs_inode_size(eb, inode_item);
676 nlink = btrfs_inode_nlink(eb, inode_item);
677 btrfs_release_path(swarn->path);
680 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
681 * uses GFP_NOFS in this context, so we keep it consistent but it does
682 * not seem to be strictly necessary.
684 nofs_flag = memalloc_nofs_save();
685 ipath = init_ipath(4096, local_root, swarn->path);
686 memalloc_nofs_restore(nofs_flag);
688 btrfs_put_root(local_root);
689 ret = PTR_ERR(ipath);
693 ret = paths_from_inode(inum, ipath);
699 * we deliberately ignore the bit ipath might have been too small to
700 * hold all of the paths here
702 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
703 btrfs_warn_in_rcu(fs_info,
704 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
705 swarn->errstr, swarn->logical,
706 rcu_str_deref(swarn->dev->name),
709 min(isize - offset, (u64)PAGE_SIZE), nlink,
710 (char *)(unsigned long)ipath->fspath->val[i]);
712 btrfs_put_root(local_root);
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 clear_page(page_address(spage->page));
1622 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1625 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1626 struct scrub_page *spage)
1628 struct scrub_bio *sbio;
1631 mutex_lock(&sctx->wr_lock);
1633 if (!sctx->wr_curr_bio) {
1634 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1636 if (!sctx->wr_curr_bio) {
1637 mutex_unlock(&sctx->wr_lock);
1640 sctx->wr_curr_bio->sctx = sctx;
1641 sctx->wr_curr_bio->page_count = 0;
1643 sbio = sctx->wr_curr_bio;
1644 if (sbio->page_count == 0) {
1647 sbio->physical = spage->physical_for_dev_replace;
1648 sbio->logical = spage->logical;
1649 sbio->dev = sctx->wr_tgtdev;
1652 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1656 bio->bi_private = sbio;
1657 bio->bi_end_io = scrub_wr_bio_end_io;
1658 bio_set_dev(bio, sbio->dev->bdev);
1659 bio->bi_iter.bi_sector = sbio->physical >> 9;
1660 bio->bi_opf = REQ_OP_WRITE;
1662 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1663 spage->physical_for_dev_replace ||
1664 sbio->logical + sbio->page_count * PAGE_SIZE !=
1666 scrub_wr_submit(sctx);
1670 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1671 if (ret != PAGE_SIZE) {
1672 if (sbio->page_count < 1) {
1675 mutex_unlock(&sctx->wr_lock);
1678 scrub_wr_submit(sctx);
1682 sbio->pagev[sbio->page_count] = spage;
1683 scrub_page_get(spage);
1685 if (sbio->page_count == sctx->pages_per_wr_bio)
1686 scrub_wr_submit(sctx);
1687 mutex_unlock(&sctx->wr_lock);
1692 static void scrub_wr_submit(struct scrub_ctx *sctx)
1694 struct scrub_bio *sbio;
1696 if (!sctx->wr_curr_bio)
1699 sbio = sctx->wr_curr_bio;
1700 sctx->wr_curr_bio = NULL;
1701 WARN_ON(!sbio->bio->bi_disk);
1702 scrub_pending_bio_inc(sctx);
1703 /* process all writes in a single worker thread. Then the block layer
1704 * orders the requests before sending them to the driver which
1705 * doubled the write performance on spinning disks when measured
1707 btrfsic_submit_bio(sbio->bio);
1710 static void scrub_wr_bio_end_io(struct bio *bio)
1712 struct scrub_bio *sbio = bio->bi_private;
1713 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1715 sbio->status = bio->bi_status;
1718 btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1719 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1722 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1724 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1725 struct scrub_ctx *sctx = sbio->sctx;
1728 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1730 struct btrfs_dev_replace *dev_replace =
1731 &sbio->sctx->fs_info->dev_replace;
1733 for (i = 0; i < sbio->page_count; i++) {
1734 struct scrub_page *spage = sbio->pagev[i];
1736 spage->io_error = 1;
1737 atomic64_inc(&dev_replace->num_write_errors);
1741 for (i = 0; i < sbio->page_count; i++)
1742 scrub_page_put(sbio->pagev[i]);
1746 scrub_pending_bio_dec(sctx);
1749 static int scrub_checksum(struct scrub_block *sblock)
1755 * No need to initialize these stats currently,
1756 * because this function only use return value
1757 * instead of these stats value.
1762 sblock->header_error = 0;
1763 sblock->generation_error = 0;
1764 sblock->checksum_error = 0;
1766 WARN_ON(sblock->page_count < 1);
1767 flags = sblock->pagev[0]->flags;
1769 if (flags & BTRFS_EXTENT_FLAG_DATA)
1770 ret = scrub_checksum_data(sblock);
1771 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1772 ret = scrub_checksum_tree_block(sblock);
1773 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1774 (void)scrub_checksum_super(sblock);
1778 scrub_handle_errored_block(sblock);
1783 static int scrub_checksum_data(struct scrub_block *sblock)
1785 struct scrub_ctx *sctx = sblock->sctx;
1786 struct btrfs_fs_info *fs_info = sctx->fs_info;
1787 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1788 u8 csum[BTRFS_CSUM_SIZE];
1795 BUG_ON(sblock->page_count < 1);
1796 if (!sblock->pagev[0]->have_csum)
1799 shash->tfm = fs_info->csum_shash;
1800 crypto_shash_init(shash);
1802 on_disk_csum = sblock->pagev[0]->csum;
1803 page = sblock->pagev[0]->page;
1804 buffer = page_address(page);
1806 len = sctx->fs_info->sectorsize;
1809 u64 l = min_t(u64, len, PAGE_SIZE);
1811 crypto_shash_update(shash, buffer, l);
1816 BUG_ON(index >= sblock->page_count);
1817 BUG_ON(!sblock->pagev[index]->page);
1818 page = sblock->pagev[index]->page;
1819 buffer = page_address(page);
1822 crypto_shash_final(shash, csum);
1823 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1824 sblock->checksum_error = 1;
1826 return sblock->checksum_error;
1829 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1831 struct scrub_ctx *sctx = sblock->sctx;
1832 struct btrfs_header *h;
1833 struct btrfs_fs_info *fs_info = sctx->fs_info;
1834 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1835 u8 calculated_csum[BTRFS_CSUM_SIZE];
1836 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1838 void *mapped_buffer;
1844 shash->tfm = fs_info->csum_shash;
1845 crypto_shash_init(shash);
1847 BUG_ON(sblock->page_count < 1);
1848 page = sblock->pagev[0]->page;
1849 mapped_buffer = page_address(page);
1850 h = (struct btrfs_header *)mapped_buffer;
1851 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1854 * we don't use the getter functions here, as we
1855 * a) don't have an extent buffer and
1856 * b) the page is already kmapped
1858 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1859 sblock->header_error = 1;
1861 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1862 sblock->header_error = 1;
1863 sblock->generation_error = 1;
1866 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1867 sblock->header_error = 1;
1869 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1871 sblock->header_error = 1;
1873 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1874 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1875 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1878 u64 l = min_t(u64, len, mapped_size);
1880 crypto_shash_update(shash, p, l);
1885 BUG_ON(index >= sblock->page_count);
1886 BUG_ON(!sblock->pagev[index]->page);
1887 page = sblock->pagev[index]->page;
1888 mapped_buffer = page_address(page);
1889 mapped_size = PAGE_SIZE;
1893 crypto_shash_final(shash, calculated_csum);
1894 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1895 sblock->checksum_error = 1;
1897 return sblock->header_error || sblock->checksum_error;
1900 static int scrub_checksum_super(struct scrub_block *sblock)
1902 struct btrfs_super_block *s;
1903 struct scrub_ctx *sctx = sblock->sctx;
1904 struct btrfs_fs_info *fs_info = sctx->fs_info;
1905 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1906 u8 calculated_csum[BTRFS_CSUM_SIZE];
1907 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1909 void *mapped_buffer;
1917 shash->tfm = fs_info->csum_shash;
1918 crypto_shash_init(shash);
1920 BUG_ON(sblock->page_count < 1);
1921 page = sblock->pagev[0]->page;
1922 mapped_buffer = page_address(page);
1923 s = (struct btrfs_super_block *)mapped_buffer;
1924 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1926 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1929 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1932 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1935 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1936 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1937 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1940 u64 l = min_t(u64, len, mapped_size);
1942 crypto_shash_update(shash, p, l);
1947 BUG_ON(index >= sblock->page_count);
1948 BUG_ON(!sblock->pagev[index]->page);
1949 page = sblock->pagev[index]->page;
1950 mapped_buffer = page_address(page);
1951 mapped_size = PAGE_SIZE;
1955 crypto_shash_final(shash, calculated_csum);
1956 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1959 if (fail_cor + fail_gen) {
1961 * if we find an error in a super block, we just report it.
1962 * They will get written with the next transaction commit
1965 spin_lock(&sctx->stat_lock);
1966 ++sctx->stat.super_errors;
1967 spin_unlock(&sctx->stat_lock);
1969 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1970 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1972 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1973 BTRFS_DEV_STAT_GENERATION_ERRS);
1976 return fail_cor + fail_gen;
1979 static void scrub_block_get(struct scrub_block *sblock)
1981 refcount_inc(&sblock->refs);
1984 static void scrub_block_put(struct scrub_block *sblock)
1986 if (refcount_dec_and_test(&sblock->refs)) {
1989 if (sblock->sparity)
1990 scrub_parity_put(sblock->sparity);
1992 for (i = 0; i < sblock->page_count; i++)
1993 scrub_page_put(sblock->pagev[i]);
1998 static void scrub_page_get(struct scrub_page *spage)
2000 atomic_inc(&spage->refs);
2003 static void scrub_page_put(struct scrub_page *spage)
2005 if (atomic_dec_and_test(&spage->refs)) {
2007 __free_page(spage->page);
2012 static void scrub_submit(struct scrub_ctx *sctx)
2014 struct scrub_bio *sbio;
2016 if (sctx->curr == -1)
2019 sbio = sctx->bios[sctx->curr];
2021 scrub_pending_bio_inc(sctx);
2022 btrfsic_submit_bio(sbio->bio);
2025 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2026 struct scrub_page *spage)
2028 struct scrub_block *sblock = spage->sblock;
2029 struct scrub_bio *sbio;
2034 * grab a fresh bio or wait for one to become available
2036 while (sctx->curr == -1) {
2037 spin_lock(&sctx->list_lock);
2038 sctx->curr = sctx->first_free;
2039 if (sctx->curr != -1) {
2040 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2041 sctx->bios[sctx->curr]->next_free = -1;
2042 sctx->bios[sctx->curr]->page_count = 0;
2043 spin_unlock(&sctx->list_lock);
2045 spin_unlock(&sctx->list_lock);
2046 wait_event(sctx->list_wait, sctx->first_free != -1);
2049 sbio = sctx->bios[sctx->curr];
2050 if (sbio->page_count == 0) {
2053 sbio->physical = spage->physical;
2054 sbio->logical = spage->logical;
2055 sbio->dev = spage->dev;
2058 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2062 bio->bi_private = sbio;
2063 bio->bi_end_io = scrub_bio_end_io;
2064 bio_set_dev(bio, sbio->dev->bdev);
2065 bio->bi_iter.bi_sector = sbio->physical >> 9;
2066 bio->bi_opf = REQ_OP_READ;
2068 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2070 sbio->logical + sbio->page_count * PAGE_SIZE !=
2072 sbio->dev != spage->dev) {
2077 sbio->pagev[sbio->page_count] = spage;
2078 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2079 if (ret != PAGE_SIZE) {
2080 if (sbio->page_count < 1) {
2089 scrub_block_get(sblock); /* one for the page added to the bio */
2090 atomic_inc(&sblock->outstanding_pages);
2092 if (sbio->page_count == sctx->pages_per_rd_bio)
2098 static void scrub_missing_raid56_end_io(struct bio *bio)
2100 struct scrub_block *sblock = bio->bi_private;
2101 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2104 sblock->no_io_error_seen = 0;
2108 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2111 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2113 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2114 struct scrub_ctx *sctx = sblock->sctx;
2115 struct btrfs_fs_info *fs_info = sctx->fs_info;
2117 struct btrfs_device *dev;
2119 logical = sblock->pagev[0]->logical;
2120 dev = sblock->pagev[0]->dev;
2122 if (sblock->no_io_error_seen)
2123 scrub_recheck_block_checksum(sblock);
2125 if (!sblock->no_io_error_seen) {
2126 spin_lock(&sctx->stat_lock);
2127 sctx->stat.read_errors++;
2128 spin_unlock(&sctx->stat_lock);
2129 btrfs_err_rl_in_rcu(fs_info,
2130 "IO error rebuilding logical %llu for dev %s",
2131 logical, rcu_str_deref(dev->name));
2132 } else if (sblock->header_error || sblock->checksum_error) {
2133 spin_lock(&sctx->stat_lock);
2134 sctx->stat.uncorrectable_errors++;
2135 spin_unlock(&sctx->stat_lock);
2136 btrfs_err_rl_in_rcu(fs_info,
2137 "failed to rebuild valid logical %llu for dev %s",
2138 logical, rcu_str_deref(dev->name));
2140 scrub_write_block_to_dev_replace(sblock);
2143 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2144 mutex_lock(&sctx->wr_lock);
2145 scrub_wr_submit(sctx);
2146 mutex_unlock(&sctx->wr_lock);
2149 scrub_block_put(sblock);
2150 scrub_pending_bio_dec(sctx);
2153 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2155 struct scrub_ctx *sctx = sblock->sctx;
2156 struct btrfs_fs_info *fs_info = sctx->fs_info;
2157 u64 length = sblock->page_count * PAGE_SIZE;
2158 u64 logical = sblock->pagev[0]->logical;
2159 struct btrfs_bio *bbio = NULL;
2161 struct btrfs_raid_bio *rbio;
2165 btrfs_bio_counter_inc_blocked(fs_info);
2166 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2168 if (ret || !bbio || !bbio->raid_map)
2171 if (WARN_ON(!sctx->is_dev_replace ||
2172 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2174 * We shouldn't be scrubbing a missing device. Even for dev
2175 * replace, we should only get here for RAID 5/6. We either
2176 * managed to mount something with no mirrors remaining or
2177 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2182 bio = btrfs_io_bio_alloc(0);
2183 bio->bi_iter.bi_sector = logical >> 9;
2184 bio->bi_private = sblock;
2185 bio->bi_end_io = scrub_missing_raid56_end_io;
2187 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2191 for (i = 0; i < sblock->page_count; i++) {
2192 struct scrub_page *spage = sblock->pagev[i];
2194 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2197 btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2198 scrub_block_get(sblock);
2199 scrub_pending_bio_inc(sctx);
2200 raid56_submit_missing_rbio(rbio);
2206 btrfs_bio_counter_dec(fs_info);
2207 btrfs_put_bbio(bbio);
2208 spin_lock(&sctx->stat_lock);
2209 sctx->stat.malloc_errors++;
2210 spin_unlock(&sctx->stat_lock);
2213 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2214 u64 physical, struct btrfs_device *dev, u64 flags,
2215 u64 gen, int mirror_num, u8 *csum, int force,
2216 u64 physical_for_dev_replace)
2218 struct scrub_block *sblock;
2221 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2223 spin_lock(&sctx->stat_lock);
2224 sctx->stat.malloc_errors++;
2225 spin_unlock(&sctx->stat_lock);
2229 /* one ref inside this function, plus one for each page added to
2231 refcount_set(&sblock->refs, 1);
2232 sblock->sctx = sctx;
2233 sblock->no_io_error_seen = 1;
2235 for (index = 0; len > 0; index++) {
2236 struct scrub_page *spage;
2237 u64 l = min_t(u64, len, PAGE_SIZE);
2239 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2242 spin_lock(&sctx->stat_lock);
2243 sctx->stat.malloc_errors++;
2244 spin_unlock(&sctx->stat_lock);
2245 scrub_block_put(sblock);
2248 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2249 scrub_page_get(spage);
2250 sblock->pagev[index] = spage;
2251 spage->sblock = sblock;
2253 spage->flags = flags;
2254 spage->generation = gen;
2255 spage->logical = logical;
2256 spage->physical = physical;
2257 spage->physical_for_dev_replace = physical_for_dev_replace;
2258 spage->mirror_num = mirror_num;
2260 spage->have_csum = 1;
2261 memcpy(spage->csum, csum, sctx->csum_size);
2263 spage->have_csum = 0;
2265 sblock->page_count++;
2266 spage->page = alloc_page(GFP_KERNEL);
2272 physical_for_dev_replace += l;
2275 WARN_ON(sblock->page_count == 0);
2276 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2278 * This case should only be hit for RAID 5/6 device replace. See
2279 * the comment in scrub_missing_raid56_pages() for details.
2281 scrub_missing_raid56_pages(sblock);
2283 for (index = 0; index < sblock->page_count; index++) {
2284 struct scrub_page *spage = sblock->pagev[index];
2287 ret = scrub_add_page_to_rd_bio(sctx, spage);
2289 scrub_block_put(sblock);
2298 /* last one frees, either here or in bio completion for last page */
2299 scrub_block_put(sblock);
2303 static void scrub_bio_end_io(struct bio *bio)
2305 struct scrub_bio *sbio = bio->bi_private;
2306 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2308 sbio->status = bio->bi_status;
2311 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2314 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2316 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2317 struct scrub_ctx *sctx = sbio->sctx;
2320 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2322 for (i = 0; i < sbio->page_count; i++) {
2323 struct scrub_page *spage = sbio->pagev[i];
2325 spage->io_error = 1;
2326 spage->sblock->no_io_error_seen = 0;
2330 /* now complete the scrub_block items that have all pages completed */
2331 for (i = 0; i < sbio->page_count; i++) {
2332 struct scrub_page *spage = sbio->pagev[i];
2333 struct scrub_block *sblock = spage->sblock;
2335 if (atomic_dec_and_test(&sblock->outstanding_pages))
2336 scrub_block_complete(sblock);
2337 scrub_block_put(sblock);
2342 spin_lock(&sctx->list_lock);
2343 sbio->next_free = sctx->first_free;
2344 sctx->first_free = sbio->index;
2345 spin_unlock(&sctx->list_lock);
2347 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2348 mutex_lock(&sctx->wr_lock);
2349 scrub_wr_submit(sctx);
2350 mutex_unlock(&sctx->wr_lock);
2353 scrub_pending_bio_dec(sctx);
2356 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2357 unsigned long *bitmap,
2363 int sectorsize = sparity->sctx->fs_info->sectorsize;
2365 if (len >= sparity->stripe_len) {
2366 bitmap_set(bitmap, 0, sparity->nsectors);
2370 start -= sparity->logic_start;
2371 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2372 offset = div_u64(offset, sectorsize);
2373 nsectors64 = div_u64(len, sectorsize);
2375 ASSERT(nsectors64 < UINT_MAX);
2376 nsectors = (u32)nsectors64;
2378 if (offset + nsectors <= sparity->nsectors) {
2379 bitmap_set(bitmap, offset, nsectors);
2383 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2384 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2387 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2390 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2393 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2396 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2399 static void scrub_block_complete(struct scrub_block *sblock)
2403 if (!sblock->no_io_error_seen) {
2405 scrub_handle_errored_block(sblock);
2408 * if has checksum error, write via repair mechanism in
2409 * dev replace case, otherwise write here in dev replace
2412 corrupted = scrub_checksum(sblock);
2413 if (!corrupted && sblock->sctx->is_dev_replace)
2414 scrub_write_block_to_dev_replace(sblock);
2417 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2418 u64 start = sblock->pagev[0]->logical;
2419 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2422 scrub_parity_mark_sectors_error(sblock->sparity,
2423 start, end - start);
2427 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2429 struct btrfs_ordered_sum *sum = NULL;
2430 unsigned long index;
2431 unsigned long num_sectors;
2433 while (!list_empty(&sctx->csum_list)) {
2434 sum = list_first_entry(&sctx->csum_list,
2435 struct btrfs_ordered_sum, list);
2436 if (sum->bytenr > logical)
2438 if (sum->bytenr + sum->len > logical)
2441 ++sctx->stat.csum_discards;
2442 list_del(&sum->list);
2449 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2450 ASSERT(index < UINT_MAX);
2452 num_sectors = sum->len / sctx->fs_info->sectorsize;
2453 memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2454 if (index == num_sectors - 1) {
2455 list_del(&sum->list);
2461 /* scrub extent tries to collect up to 64 kB for each bio */
2462 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2463 u64 logical, u64 len,
2464 u64 physical, struct btrfs_device *dev, u64 flags,
2465 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2468 u8 csum[BTRFS_CSUM_SIZE];
2471 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2472 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2473 blocksize = map->stripe_len;
2475 blocksize = sctx->fs_info->sectorsize;
2476 spin_lock(&sctx->stat_lock);
2477 sctx->stat.data_extents_scrubbed++;
2478 sctx->stat.data_bytes_scrubbed += len;
2479 spin_unlock(&sctx->stat_lock);
2480 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2481 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2482 blocksize = map->stripe_len;
2484 blocksize = sctx->fs_info->nodesize;
2485 spin_lock(&sctx->stat_lock);
2486 sctx->stat.tree_extents_scrubbed++;
2487 sctx->stat.tree_bytes_scrubbed += len;
2488 spin_unlock(&sctx->stat_lock);
2490 blocksize = sctx->fs_info->sectorsize;
2495 u64 l = min_t(u64, len, blocksize);
2498 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2499 /* push csums to sbio */
2500 have_csum = scrub_find_csum(sctx, logical, csum);
2502 ++sctx->stat.no_csum;
2504 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2505 mirror_num, have_csum ? csum : NULL, 0,
2506 physical_for_dev_replace);
2512 physical_for_dev_replace += l;
2517 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2518 u64 logical, u64 len,
2519 u64 physical, struct btrfs_device *dev,
2520 u64 flags, u64 gen, int mirror_num, u8 *csum)
2522 struct scrub_ctx *sctx = sparity->sctx;
2523 struct scrub_block *sblock;
2526 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2528 spin_lock(&sctx->stat_lock);
2529 sctx->stat.malloc_errors++;
2530 spin_unlock(&sctx->stat_lock);
2534 /* one ref inside this function, plus one for each page added to
2536 refcount_set(&sblock->refs, 1);
2537 sblock->sctx = sctx;
2538 sblock->no_io_error_seen = 1;
2539 sblock->sparity = sparity;
2540 scrub_parity_get(sparity);
2542 for (index = 0; len > 0; index++) {
2543 struct scrub_page *spage;
2544 u64 l = min_t(u64, len, PAGE_SIZE);
2546 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2549 spin_lock(&sctx->stat_lock);
2550 sctx->stat.malloc_errors++;
2551 spin_unlock(&sctx->stat_lock);
2552 scrub_block_put(sblock);
2555 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2556 /* For scrub block */
2557 scrub_page_get(spage);
2558 sblock->pagev[index] = spage;
2559 /* For scrub parity */
2560 scrub_page_get(spage);
2561 list_add_tail(&spage->list, &sparity->spages);
2562 spage->sblock = sblock;
2564 spage->flags = flags;
2565 spage->generation = gen;
2566 spage->logical = logical;
2567 spage->physical = physical;
2568 spage->mirror_num = mirror_num;
2570 spage->have_csum = 1;
2571 memcpy(spage->csum, csum, sctx->csum_size);
2573 spage->have_csum = 0;
2575 sblock->page_count++;
2576 spage->page = alloc_page(GFP_KERNEL);
2584 WARN_ON(sblock->page_count == 0);
2585 for (index = 0; index < sblock->page_count; index++) {
2586 struct scrub_page *spage = sblock->pagev[index];
2589 ret = scrub_add_page_to_rd_bio(sctx, spage);
2591 scrub_block_put(sblock);
2596 /* last one frees, either here or in bio completion for last page */
2597 scrub_block_put(sblock);
2601 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2602 u64 logical, u64 len,
2603 u64 physical, struct btrfs_device *dev,
2604 u64 flags, u64 gen, int mirror_num)
2606 struct scrub_ctx *sctx = sparity->sctx;
2608 u8 csum[BTRFS_CSUM_SIZE];
2611 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2612 scrub_parity_mark_sectors_error(sparity, logical, len);
2616 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2617 blocksize = sparity->stripe_len;
2618 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2619 blocksize = sparity->stripe_len;
2621 blocksize = sctx->fs_info->sectorsize;
2626 u64 l = min_t(u64, len, blocksize);
2629 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2630 /* push csums to sbio */
2631 have_csum = scrub_find_csum(sctx, logical, csum);
2635 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2636 flags, gen, mirror_num,
2637 have_csum ? csum : NULL);
2649 * Given a physical address, this will calculate it's
2650 * logical offset. if this is a parity stripe, it will return
2651 * the most left data stripe's logical offset.
2653 * return 0 if it is a data stripe, 1 means parity stripe.
2655 static int get_raid56_logic_offset(u64 physical, int num,
2656 struct map_lookup *map, u64 *offset,
2665 const int data_stripes = nr_data_stripes(map);
2667 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2669 *stripe_start = last_offset;
2671 *offset = last_offset;
2672 for (i = 0; i < data_stripes; i++) {
2673 *offset = last_offset + i * map->stripe_len;
2675 stripe_nr = div64_u64(*offset, map->stripe_len);
2676 stripe_nr = div_u64(stripe_nr, data_stripes);
2678 /* Work out the disk rotation on this stripe-set */
2679 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2680 /* calculate which stripe this data locates */
2682 stripe_index = rot % map->num_stripes;
2683 if (stripe_index == num)
2685 if (stripe_index < num)
2688 *offset = last_offset + j * map->stripe_len;
2692 static void scrub_free_parity(struct scrub_parity *sparity)
2694 struct scrub_ctx *sctx = sparity->sctx;
2695 struct scrub_page *curr, *next;
2698 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2700 spin_lock(&sctx->stat_lock);
2701 sctx->stat.read_errors += nbits;
2702 sctx->stat.uncorrectable_errors += nbits;
2703 spin_unlock(&sctx->stat_lock);
2706 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2707 list_del_init(&curr->list);
2708 scrub_page_put(curr);
2714 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2716 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2718 struct scrub_ctx *sctx = sparity->sctx;
2720 scrub_free_parity(sparity);
2721 scrub_pending_bio_dec(sctx);
2724 static void scrub_parity_bio_endio(struct bio *bio)
2726 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2727 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2730 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2735 btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2737 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2740 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2742 struct scrub_ctx *sctx = sparity->sctx;
2743 struct btrfs_fs_info *fs_info = sctx->fs_info;
2745 struct btrfs_raid_bio *rbio;
2746 struct btrfs_bio *bbio = NULL;
2750 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2754 length = sparity->logic_end - sparity->logic_start;
2756 btrfs_bio_counter_inc_blocked(fs_info);
2757 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2759 if (ret || !bbio || !bbio->raid_map)
2762 bio = btrfs_io_bio_alloc(0);
2763 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2764 bio->bi_private = sparity;
2765 bio->bi_end_io = scrub_parity_bio_endio;
2767 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2768 length, sparity->scrub_dev,
2774 scrub_pending_bio_inc(sctx);
2775 raid56_parity_submit_scrub_rbio(rbio);
2781 btrfs_bio_counter_dec(fs_info);
2782 btrfs_put_bbio(bbio);
2783 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2785 spin_lock(&sctx->stat_lock);
2786 sctx->stat.malloc_errors++;
2787 spin_unlock(&sctx->stat_lock);
2789 scrub_free_parity(sparity);
2792 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2794 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2797 static void scrub_parity_get(struct scrub_parity *sparity)
2799 refcount_inc(&sparity->refs);
2802 static void scrub_parity_put(struct scrub_parity *sparity)
2804 if (!refcount_dec_and_test(&sparity->refs))
2807 scrub_parity_check_and_repair(sparity);
2810 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2811 struct map_lookup *map,
2812 struct btrfs_device *sdev,
2813 struct btrfs_path *path,
2817 struct btrfs_fs_info *fs_info = sctx->fs_info;
2818 struct btrfs_root *root = fs_info->extent_root;
2819 struct btrfs_root *csum_root = fs_info->csum_root;
2820 struct btrfs_extent_item *extent;
2821 struct btrfs_bio *bbio = NULL;
2825 struct extent_buffer *l;
2826 struct btrfs_key key;
2829 u64 extent_physical;
2832 struct btrfs_device *extent_dev;
2833 struct scrub_parity *sparity;
2836 int extent_mirror_num;
2839 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2840 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2841 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2844 spin_lock(&sctx->stat_lock);
2845 sctx->stat.malloc_errors++;
2846 spin_unlock(&sctx->stat_lock);
2850 sparity->stripe_len = map->stripe_len;
2851 sparity->nsectors = nsectors;
2852 sparity->sctx = sctx;
2853 sparity->scrub_dev = sdev;
2854 sparity->logic_start = logic_start;
2855 sparity->logic_end = logic_end;
2856 refcount_set(&sparity->refs, 1);
2857 INIT_LIST_HEAD(&sparity->spages);
2858 sparity->dbitmap = sparity->bitmap;
2859 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2862 while (logic_start < logic_end) {
2863 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2864 key.type = BTRFS_METADATA_ITEM_KEY;
2866 key.type = BTRFS_EXTENT_ITEM_KEY;
2867 key.objectid = logic_start;
2868 key.offset = (u64)-1;
2870 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2875 ret = btrfs_previous_extent_item(root, path, 0);
2879 btrfs_release_path(path);
2880 ret = btrfs_search_slot(NULL, root, &key,
2892 slot = path->slots[0];
2893 if (slot >= btrfs_header_nritems(l)) {
2894 ret = btrfs_next_leaf(root, path);
2903 btrfs_item_key_to_cpu(l, &key, slot);
2905 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2906 key.type != BTRFS_METADATA_ITEM_KEY)
2909 if (key.type == BTRFS_METADATA_ITEM_KEY)
2910 bytes = fs_info->nodesize;
2914 if (key.objectid + bytes <= logic_start)
2917 if (key.objectid >= logic_end) {
2922 while (key.objectid >= logic_start + map->stripe_len)
2923 logic_start += map->stripe_len;
2925 extent = btrfs_item_ptr(l, slot,
2926 struct btrfs_extent_item);
2927 flags = btrfs_extent_flags(l, extent);
2928 generation = btrfs_extent_generation(l, extent);
2930 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2931 (key.objectid < logic_start ||
2932 key.objectid + bytes >
2933 logic_start + map->stripe_len)) {
2935 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2936 key.objectid, logic_start);
2937 spin_lock(&sctx->stat_lock);
2938 sctx->stat.uncorrectable_errors++;
2939 spin_unlock(&sctx->stat_lock);
2943 extent_logical = key.objectid;
2946 if (extent_logical < logic_start) {
2947 extent_len -= logic_start - extent_logical;
2948 extent_logical = logic_start;
2951 if (extent_logical + extent_len >
2952 logic_start + map->stripe_len)
2953 extent_len = logic_start + map->stripe_len -
2956 scrub_parity_mark_sectors_data(sparity, extent_logical,
2959 mapped_length = extent_len;
2961 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2962 extent_logical, &mapped_length, &bbio,
2965 if (!bbio || mapped_length < extent_len)
2969 btrfs_put_bbio(bbio);
2972 extent_physical = bbio->stripes[0].physical;
2973 extent_mirror_num = bbio->mirror_num;
2974 extent_dev = bbio->stripes[0].dev;
2975 btrfs_put_bbio(bbio);
2977 ret = btrfs_lookup_csums_range(csum_root,
2979 extent_logical + extent_len - 1,
2980 &sctx->csum_list, 1);
2984 ret = scrub_extent_for_parity(sparity, extent_logical,
2991 scrub_free_csums(sctx);
2996 if (extent_logical + extent_len <
2997 key.objectid + bytes) {
2998 logic_start += map->stripe_len;
3000 if (logic_start >= logic_end) {
3005 if (logic_start < key.objectid + bytes) {
3014 btrfs_release_path(path);
3019 logic_start += map->stripe_len;
3023 scrub_parity_mark_sectors_error(sparity, logic_start,
3024 logic_end - logic_start);
3025 scrub_parity_put(sparity);
3027 mutex_lock(&sctx->wr_lock);
3028 scrub_wr_submit(sctx);
3029 mutex_unlock(&sctx->wr_lock);
3031 btrfs_release_path(path);
3032 return ret < 0 ? ret : 0;
3035 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3036 struct map_lookup *map,
3037 struct btrfs_device *scrub_dev,
3038 int num, u64 base, u64 length,
3039 struct btrfs_block_group *cache)
3041 struct btrfs_path *path, *ppath;
3042 struct btrfs_fs_info *fs_info = sctx->fs_info;
3043 struct btrfs_root *root = fs_info->extent_root;
3044 struct btrfs_root *csum_root = fs_info->csum_root;
3045 struct btrfs_extent_item *extent;
3046 struct blk_plug plug;
3051 struct extent_buffer *l;
3058 struct reada_control *reada1;
3059 struct reada_control *reada2;
3060 struct btrfs_key key;
3061 struct btrfs_key key_end;
3062 u64 increment = map->stripe_len;
3065 u64 extent_physical;
3069 struct btrfs_device *extent_dev;
3070 int extent_mirror_num;
3073 physical = map->stripes[num].physical;
3075 nstripes = div64_u64(length, map->stripe_len);
3076 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3077 offset = map->stripe_len * num;
3078 increment = map->stripe_len * map->num_stripes;
3080 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3081 int factor = map->num_stripes / map->sub_stripes;
3082 offset = map->stripe_len * (num / map->sub_stripes);
3083 increment = map->stripe_len * factor;
3084 mirror_num = num % map->sub_stripes + 1;
3085 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3086 increment = map->stripe_len;
3087 mirror_num = num % map->num_stripes + 1;
3088 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3089 increment = map->stripe_len;
3090 mirror_num = num % map->num_stripes + 1;
3091 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3092 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3093 increment = map->stripe_len * nr_data_stripes(map);
3096 increment = map->stripe_len;
3100 path = btrfs_alloc_path();
3104 ppath = btrfs_alloc_path();
3106 btrfs_free_path(path);
3111 * work on commit root. The related disk blocks are static as
3112 * long as COW is applied. This means, it is save to rewrite
3113 * them to repair disk errors without any race conditions
3115 path->search_commit_root = 1;
3116 path->skip_locking = 1;
3118 ppath->search_commit_root = 1;
3119 ppath->skip_locking = 1;
3121 * trigger the readahead for extent tree csum tree and wait for
3122 * completion. During readahead, the scrub is officially paused
3123 * to not hold off transaction commits
3125 logical = base + offset;
3126 physical_end = physical + nstripes * map->stripe_len;
3127 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3128 get_raid56_logic_offset(physical_end, num,
3129 map, &logic_end, NULL);
3132 logic_end = logical + increment * nstripes;
3134 wait_event(sctx->list_wait,
3135 atomic_read(&sctx->bios_in_flight) == 0);
3136 scrub_blocked_if_needed(fs_info);
3138 /* FIXME it might be better to start readahead at commit root */
3139 key.objectid = logical;
3140 key.type = BTRFS_EXTENT_ITEM_KEY;
3141 key.offset = (u64)0;
3142 key_end.objectid = logic_end;
3143 key_end.type = BTRFS_METADATA_ITEM_KEY;
3144 key_end.offset = (u64)-1;
3145 reada1 = btrfs_reada_add(root, &key, &key_end);
3147 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3148 key.type = BTRFS_EXTENT_CSUM_KEY;
3149 key.offset = logical;
3150 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3151 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3152 key_end.offset = logic_end;
3153 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3155 if (!IS_ERR(reada1))
3156 btrfs_reada_wait(reada1);
3157 if (!IS_ERR(reada2))
3158 btrfs_reada_wait(reada2);
3162 * collect all data csums for the stripe to avoid seeking during
3163 * the scrub. This might currently (crc32) end up to be about 1MB
3165 blk_start_plug(&plug);
3168 * now find all extents for each stripe and scrub them
3171 while (physical < physical_end) {
3175 if (atomic_read(&fs_info->scrub_cancel_req) ||
3176 atomic_read(&sctx->cancel_req)) {
3181 * check to see if we have to pause
3183 if (atomic_read(&fs_info->scrub_pause_req)) {
3184 /* push queued extents */
3185 sctx->flush_all_writes = true;
3187 mutex_lock(&sctx->wr_lock);
3188 scrub_wr_submit(sctx);
3189 mutex_unlock(&sctx->wr_lock);
3190 wait_event(sctx->list_wait,
3191 atomic_read(&sctx->bios_in_flight) == 0);
3192 sctx->flush_all_writes = false;
3193 scrub_blocked_if_needed(fs_info);
3196 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3197 ret = get_raid56_logic_offset(physical, num, map,
3202 /* it is parity strip */
3203 stripe_logical += base;
3204 stripe_end = stripe_logical + increment;
3205 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3206 ppath, stripe_logical,
3214 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3215 key.type = BTRFS_METADATA_ITEM_KEY;
3217 key.type = BTRFS_EXTENT_ITEM_KEY;
3218 key.objectid = logical;
3219 key.offset = (u64)-1;
3221 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3226 ret = btrfs_previous_extent_item(root, path, 0);
3230 /* there's no smaller item, so stick with the
3232 btrfs_release_path(path);
3233 ret = btrfs_search_slot(NULL, root, &key,
3245 slot = path->slots[0];
3246 if (slot >= btrfs_header_nritems(l)) {
3247 ret = btrfs_next_leaf(root, path);
3256 btrfs_item_key_to_cpu(l, &key, slot);
3258 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3259 key.type != BTRFS_METADATA_ITEM_KEY)
3262 if (key.type == BTRFS_METADATA_ITEM_KEY)
3263 bytes = fs_info->nodesize;
3267 if (key.objectid + bytes <= logical)
3270 if (key.objectid >= logical + map->stripe_len) {
3271 /* out of this device extent */
3272 if (key.objectid >= logic_end)
3278 * If our block group was removed in the meanwhile, just
3279 * stop scrubbing since there is no point in continuing.
3280 * Continuing would prevent reusing its device extents
3281 * for new block groups for a long time.
3283 spin_lock(&cache->lock);
3284 if (cache->removed) {
3285 spin_unlock(&cache->lock);
3289 spin_unlock(&cache->lock);
3291 extent = btrfs_item_ptr(l, slot,
3292 struct btrfs_extent_item);
3293 flags = btrfs_extent_flags(l, extent);
3294 generation = btrfs_extent_generation(l, extent);
3296 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3297 (key.objectid < logical ||
3298 key.objectid + bytes >
3299 logical + map->stripe_len)) {
3301 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3302 key.objectid, logical);
3303 spin_lock(&sctx->stat_lock);
3304 sctx->stat.uncorrectable_errors++;
3305 spin_unlock(&sctx->stat_lock);
3310 extent_logical = key.objectid;
3314 * trim extent to this stripe
3316 if (extent_logical < logical) {
3317 extent_len -= logical - extent_logical;
3318 extent_logical = logical;
3320 if (extent_logical + extent_len >
3321 logical + map->stripe_len) {
3322 extent_len = logical + map->stripe_len -
3326 extent_physical = extent_logical - logical + physical;
3327 extent_dev = scrub_dev;
3328 extent_mirror_num = mirror_num;
3329 if (sctx->is_dev_replace)
3330 scrub_remap_extent(fs_info, extent_logical,
3331 extent_len, &extent_physical,
3333 &extent_mirror_num);
3335 if (flags & BTRFS_EXTENT_FLAG_DATA) {
3336 ret = btrfs_lookup_csums_range(csum_root,
3338 extent_logical + extent_len - 1,
3339 &sctx->csum_list, 1);
3344 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3345 extent_physical, extent_dev, flags,
3346 generation, extent_mirror_num,
3347 extent_logical - logical + physical);
3349 scrub_free_csums(sctx);
3354 if (extent_logical + extent_len <
3355 key.objectid + bytes) {
3356 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3358 * loop until we find next data stripe
3359 * or we have finished all stripes.
3362 physical += map->stripe_len;
3363 ret = get_raid56_logic_offset(physical,
3368 if (ret && physical < physical_end) {
3369 stripe_logical += base;
3370 stripe_end = stripe_logical +
3372 ret = scrub_raid56_parity(sctx,
3373 map, scrub_dev, ppath,
3381 physical += map->stripe_len;
3382 logical += increment;
3384 if (logical < key.objectid + bytes) {
3389 if (physical >= physical_end) {
3397 btrfs_release_path(path);
3399 logical += increment;
3400 physical += map->stripe_len;
3401 spin_lock(&sctx->stat_lock);
3403 sctx->stat.last_physical = map->stripes[num].physical +
3406 sctx->stat.last_physical = physical;
3407 spin_unlock(&sctx->stat_lock);
3412 /* push queued extents */
3414 mutex_lock(&sctx->wr_lock);
3415 scrub_wr_submit(sctx);
3416 mutex_unlock(&sctx->wr_lock);
3418 blk_finish_plug(&plug);
3419 btrfs_free_path(path);
3420 btrfs_free_path(ppath);
3421 return ret < 0 ? ret : 0;
3424 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3425 struct btrfs_device *scrub_dev,
3426 u64 chunk_offset, u64 length,
3428 struct btrfs_block_group *cache)
3430 struct btrfs_fs_info *fs_info = sctx->fs_info;
3431 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3432 struct map_lookup *map;
3433 struct extent_map *em;
3437 read_lock(&map_tree->lock);
3438 em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3439 read_unlock(&map_tree->lock);
3443 * Might have been an unused block group deleted by the cleaner
3444 * kthread or relocation.
3446 spin_lock(&cache->lock);
3447 if (!cache->removed)
3449 spin_unlock(&cache->lock);
3454 map = em->map_lookup;
3455 if (em->start != chunk_offset)
3458 if (em->len < length)
3461 for (i = 0; i < map->num_stripes; ++i) {
3462 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3463 map->stripes[i].physical == dev_offset) {
3464 ret = scrub_stripe(sctx, map, scrub_dev, i,
3465 chunk_offset, length, cache);
3471 free_extent_map(em);
3476 static noinline_for_stack
3477 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3478 struct btrfs_device *scrub_dev, u64 start, u64 end)
3480 struct btrfs_dev_extent *dev_extent = NULL;
3481 struct btrfs_path *path;
3482 struct btrfs_fs_info *fs_info = sctx->fs_info;
3483 struct btrfs_root *root = fs_info->dev_root;
3489 struct extent_buffer *l;
3490 struct btrfs_key key;
3491 struct btrfs_key found_key;
3492 struct btrfs_block_group *cache;
3493 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3495 path = btrfs_alloc_path();
3499 path->reada = READA_FORWARD;
3500 path->search_commit_root = 1;
3501 path->skip_locking = 1;
3503 key.objectid = scrub_dev->devid;
3505 key.type = BTRFS_DEV_EXTENT_KEY;
3508 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3512 if (path->slots[0] >=
3513 btrfs_header_nritems(path->nodes[0])) {
3514 ret = btrfs_next_leaf(root, path);
3527 slot = path->slots[0];
3529 btrfs_item_key_to_cpu(l, &found_key, slot);
3531 if (found_key.objectid != scrub_dev->devid)
3534 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3537 if (found_key.offset >= end)
3540 if (found_key.offset < key.offset)
3543 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3544 length = btrfs_dev_extent_length(l, dev_extent);
3546 if (found_key.offset + length <= start)
3549 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3552 * get a reference on the corresponding block group to prevent
3553 * the chunk from going away while we scrub it
3555 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3557 /* some chunks are removed but not committed to disk yet,
3558 * continue scrubbing */
3563 * Make sure that while we are scrubbing the corresponding block
3564 * group doesn't get its logical address and its device extents
3565 * reused for another block group, which can possibly be of a
3566 * different type and different profile. We do this to prevent
3567 * false error detections and crashes due to bogus attempts to
3570 spin_lock(&cache->lock);
3571 if (cache->removed) {
3572 spin_unlock(&cache->lock);
3573 btrfs_put_block_group(cache);
3576 btrfs_freeze_block_group(cache);
3577 spin_unlock(&cache->lock);
3580 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3581 * to avoid deadlock caused by:
3582 * btrfs_inc_block_group_ro()
3583 * -> btrfs_wait_for_commit()
3584 * -> btrfs_commit_transaction()
3585 * -> btrfs_scrub_pause()
3587 scrub_pause_on(fs_info);
3590 * Don't do chunk preallocation for scrub.
3592 * This is especially important for SYSTEM bgs, or we can hit
3593 * -EFBIG from btrfs_finish_chunk_alloc() like:
3594 * 1. The only SYSTEM bg is marked RO.
3595 * Since SYSTEM bg is small, that's pretty common.
3596 * 2. New SYSTEM bg will be allocated
3597 * Due to regular version will allocate new chunk.
3598 * 3. New SYSTEM bg is empty and will get cleaned up
3599 * Before cleanup really happens, it's marked RO again.
3600 * 4. Empty SYSTEM bg get scrubbed
3603 * This can easily boost the amount of SYSTEM chunks if cleaner
3604 * thread can't be triggered fast enough, and use up all space
3605 * of btrfs_super_block::sys_chunk_array
3607 * While for dev replace, we need to try our best to mark block
3608 * group RO, to prevent race between:
3609 * - Write duplication
3610 * Contains latest data
3612 * Contains data from commit tree
3614 * If target block group is not marked RO, nocow writes can
3615 * be overwritten by scrub copy, causing data corruption.
3616 * So for dev-replace, it's not allowed to continue if a block
3619 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3622 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3624 * btrfs_inc_block_group_ro return -ENOSPC when it
3625 * failed in creating new chunk for metadata.
3626 * It is not a problem for scrub, because
3627 * metadata are always cowed, and our scrub paused
3628 * commit_transactions.
3633 "failed setting block group ro: %d", ret);
3634 btrfs_unfreeze_block_group(cache);
3635 btrfs_put_block_group(cache);
3636 scrub_pause_off(fs_info);
3641 * Now the target block is marked RO, wait for nocow writes to
3642 * finish before dev-replace.
3643 * COW is fine, as COW never overwrites extents in commit tree.
3645 if (sctx->is_dev_replace) {
3646 btrfs_wait_nocow_writers(cache);
3647 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3651 scrub_pause_off(fs_info);
3652 down_write(&dev_replace->rwsem);
3653 dev_replace->cursor_right = found_key.offset + length;
3654 dev_replace->cursor_left = found_key.offset;
3655 dev_replace->item_needs_writeback = 1;
3656 up_write(&dev_replace->rwsem);
3658 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3659 found_key.offset, cache);
3662 * flush, submit all pending read and write bios, afterwards
3664 * Note that in the dev replace case, a read request causes
3665 * write requests that are submitted in the read completion
3666 * worker. Therefore in the current situation, it is required
3667 * that all write requests are flushed, so that all read and
3668 * write requests are really completed when bios_in_flight
3671 sctx->flush_all_writes = true;
3673 mutex_lock(&sctx->wr_lock);
3674 scrub_wr_submit(sctx);
3675 mutex_unlock(&sctx->wr_lock);
3677 wait_event(sctx->list_wait,
3678 atomic_read(&sctx->bios_in_flight) == 0);
3680 scrub_pause_on(fs_info);
3683 * must be called before we decrease @scrub_paused.
3684 * make sure we don't block transaction commit while
3685 * we are waiting pending workers finished.
3687 wait_event(sctx->list_wait,
3688 atomic_read(&sctx->workers_pending) == 0);
3689 sctx->flush_all_writes = false;
3691 scrub_pause_off(fs_info);
3693 down_write(&dev_replace->rwsem);
3694 dev_replace->cursor_left = dev_replace->cursor_right;
3695 dev_replace->item_needs_writeback = 1;
3696 up_write(&dev_replace->rwsem);
3699 btrfs_dec_block_group_ro(cache);
3702 * We might have prevented the cleaner kthread from deleting
3703 * this block group if it was already unused because we raced
3704 * and set it to RO mode first. So add it back to the unused
3705 * list, otherwise it might not ever be deleted unless a manual
3706 * balance is triggered or it becomes used and unused again.
3708 spin_lock(&cache->lock);
3709 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3711 spin_unlock(&cache->lock);
3712 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3713 btrfs_discard_queue_work(&fs_info->discard_ctl,
3716 btrfs_mark_bg_unused(cache);
3718 spin_unlock(&cache->lock);
3721 btrfs_unfreeze_block_group(cache);
3722 btrfs_put_block_group(cache);
3725 if (sctx->is_dev_replace &&
3726 atomic64_read(&dev_replace->num_write_errors) > 0) {
3730 if (sctx->stat.malloc_errors > 0) {
3735 key.offset = found_key.offset + length;
3736 btrfs_release_path(path);
3739 btrfs_free_path(path);
3744 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3745 struct btrfs_device *scrub_dev)
3751 struct btrfs_fs_info *fs_info = sctx->fs_info;
3753 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3756 /* Seed devices of a new filesystem has their own generation. */
3757 if (scrub_dev->fs_devices != fs_info->fs_devices)
3758 gen = scrub_dev->generation;
3760 gen = fs_info->last_trans_committed;
3762 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3763 bytenr = btrfs_sb_offset(i);
3764 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3765 scrub_dev->commit_total_bytes)
3768 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3769 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3774 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3780 * get a reference count on fs_info->scrub_workers. start worker if necessary
3782 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3785 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3786 int max_active = fs_info->thread_pool_size;
3788 lockdep_assert_held(&fs_info->scrub_lock);
3790 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3791 ASSERT(fs_info->scrub_workers == NULL);
3792 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3793 flags, is_dev_replace ? 1 : max_active, 4);
3794 if (!fs_info->scrub_workers)
3795 goto fail_scrub_workers;
3797 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3798 fs_info->scrub_wr_completion_workers =
3799 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3801 if (!fs_info->scrub_wr_completion_workers)
3802 goto fail_scrub_wr_completion_workers;
3804 ASSERT(fs_info->scrub_parity_workers == NULL);
3805 fs_info->scrub_parity_workers =
3806 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3808 if (!fs_info->scrub_parity_workers)
3809 goto fail_scrub_parity_workers;
3811 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3813 refcount_inc(&fs_info->scrub_workers_refcnt);
3817 fail_scrub_parity_workers:
3818 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3819 fail_scrub_wr_completion_workers:
3820 btrfs_destroy_workqueue(fs_info->scrub_workers);
3825 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3826 u64 end, struct btrfs_scrub_progress *progress,
3827 int readonly, int is_dev_replace)
3829 struct scrub_ctx *sctx;
3831 struct btrfs_device *dev;
3832 unsigned int nofs_flag;
3833 struct btrfs_workqueue *scrub_workers = NULL;
3834 struct btrfs_workqueue *scrub_wr_comp = NULL;
3835 struct btrfs_workqueue *scrub_parity = NULL;
3837 if (btrfs_fs_closing(fs_info))
3840 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3842 * in this case scrub is unable to calculate the checksum
3843 * the way scrub is implemented. Do not handle this
3844 * situation at all because it won't ever happen.
3847 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3853 if (fs_info->sectorsize != PAGE_SIZE) {
3854 /* not supported for data w/o checksums */
3855 btrfs_err_rl(fs_info,
3856 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3857 fs_info->sectorsize, PAGE_SIZE);
3861 if (fs_info->nodesize >
3862 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3863 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3865 * would exhaust the array bounds of pagev member in
3866 * struct scrub_block
3869 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3871 SCRUB_MAX_PAGES_PER_BLOCK,
3872 fs_info->sectorsize,
3873 SCRUB_MAX_PAGES_PER_BLOCK);
3877 /* Allocate outside of device_list_mutex */
3878 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3880 return PTR_ERR(sctx);
3882 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3883 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3884 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3886 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3891 if (!is_dev_replace && !readonly &&
3892 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3893 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3894 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3895 rcu_str_deref(dev->name));
3900 mutex_lock(&fs_info->scrub_lock);
3901 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3902 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3903 mutex_unlock(&fs_info->scrub_lock);
3904 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3909 down_read(&fs_info->dev_replace.rwsem);
3910 if (dev->scrub_ctx ||
3912 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3913 up_read(&fs_info->dev_replace.rwsem);
3914 mutex_unlock(&fs_info->scrub_lock);
3915 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3919 up_read(&fs_info->dev_replace.rwsem);
3921 ret = scrub_workers_get(fs_info, is_dev_replace);
3923 mutex_unlock(&fs_info->scrub_lock);
3924 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3928 sctx->readonly = readonly;
3929 dev->scrub_ctx = sctx;
3930 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3933 * checking @scrub_pause_req here, we can avoid
3934 * race between committing transaction and scrubbing.
3936 __scrub_blocked_if_needed(fs_info);
3937 atomic_inc(&fs_info->scrubs_running);
3938 mutex_unlock(&fs_info->scrub_lock);
3941 * In order to avoid deadlock with reclaim when there is a transaction
3942 * trying to pause scrub, make sure we use GFP_NOFS for all the
3943 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3944 * invoked by our callees. The pausing request is done when the
3945 * transaction commit starts, and it blocks the transaction until scrub
3946 * is paused (done at specific points at scrub_stripe() or right above
3947 * before incrementing fs_info->scrubs_running).
3949 nofs_flag = memalloc_nofs_save();
3950 if (!is_dev_replace) {
3951 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3953 * by holding device list mutex, we can
3954 * kick off writing super in log tree sync.
3956 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3957 ret = scrub_supers(sctx, dev);
3958 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3962 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3963 memalloc_nofs_restore(nofs_flag);
3965 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3966 atomic_dec(&fs_info->scrubs_running);
3967 wake_up(&fs_info->scrub_pause_wait);
3969 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3972 memcpy(progress, &sctx->stat, sizeof(*progress));
3974 if (!is_dev_replace)
3975 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3976 ret ? "not finished" : "finished", devid, ret);
3978 mutex_lock(&fs_info->scrub_lock);
3979 dev->scrub_ctx = NULL;
3980 if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3981 scrub_workers = fs_info->scrub_workers;
3982 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3983 scrub_parity = fs_info->scrub_parity_workers;
3985 fs_info->scrub_workers = NULL;
3986 fs_info->scrub_wr_completion_workers = NULL;
3987 fs_info->scrub_parity_workers = NULL;
3989 mutex_unlock(&fs_info->scrub_lock);
3991 btrfs_destroy_workqueue(scrub_workers);
3992 btrfs_destroy_workqueue(scrub_wr_comp);
3993 btrfs_destroy_workqueue(scrub_parity);
3994 scrub_put_ctx(sctx);
3999 scrub_free_ctx(sctx);
4004 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4006 mutex_lock(&fs_info->scrub_lock);
4007 atomic_inc(&fs_info->scrub_pause_req);
4008 while (atomic_read(&fs_info->scrubs_paused) !=
4009 atomic_read(&fs_info->scrubs_running)) {
4010 mutex_unlock(&fs_info->scrub_lock);
4011 wait_event(fs_info->scrub_pause_wait,
4012 atomic_read(&fs_info->scrubs_paused) ==
4013 atomic_read(&fs_info->scrubs_running));
4014 mutex_lock(&fs_info->scrub_lock);
4016 mutex_unlock(&fs_info->scrub_lock);
4019 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4021 atomic_dec(&fs_info->scrub_pause_req);
4022 wake_up(&fs_info->scrub_pause_wait);
4025 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4027 mutex_lock(&fs_info->scrub_lock);
4028 if (!atomic_read(&fs_info->scrubs_running)) {
4029 mutex_unlock(&fs_info->scrub_lock);
4033 atomic_inc(&fs_info->scrub_cancel_req);
4034 while (atomic_read(&fs_info->scrubs_running)) {
4035 mutex_unlock(&fs_info->scrub_lock);
4036 wait_event(fs_info->scrub_pause_wait,
4037 atomic_read(&fs_info->scrubs_running) == 0);
4038 mutex_lock(&fs_info->scrub_lock);
4040 atomic_dec(&fs_info->scrub_cancel_req);
4041 mutex_unlock(&fs_info->scrub_lock);
4046 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4048 struct btrfs_fs_info *fs_info = dev->fs_info;
4049 struct scrub_ctx *sctx;
4051 mutex_lock(&fs_info->scrub_lock);
4052 sctx = dev->scrub_ctx;
4054 mutex_unlock(&fs_info->scrub_lock);
4057 atomic_inc(&sctx->cancel_req);
4058 while (dev->scrub_ctx) {
4059 mutex_unlock(&fs_info->scrub_lock);
4060 wait_event(fs_info->scrub_pause_wait,
4061 dev->scrub_ctx == NULL);
4062 mutex_lock(&fs_info->scrub_lock);
4064 mutex_unlock(&fs_info->scrub_lock);
4069 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4070 struct btrfs_scrub_progress *progress)
4072 struct btrfs_device *dev;
4073 struct scrub_ctx *sctx = NULL;
4075 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4076 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4078 sctx = dev->scrub_ctx;
4080 memcpy(progress, &sctx->stat, sizeof(*progress));
4081 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4083 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4086 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4087 u64 extent_logical, u64 extent_len,
4088 u64 *extent_physical,
4089 struct btrfs_device **extent_dev,
4090 int *extent_mirror_num)
4093 struct btrfs_bio *bbio = NULL;
4096 mapped_length = extent_len;
4097 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4098 &mapped_length, &bbio, 0);
4099 if (ret || !bbio || mapped_length < extent_len ||
4100 !bbio->stripes[0].dev->bdev) {
4101 btrfs_put_bbio(bbio);
4105 *extent_physical = bbio->stripes[0].physical;
4106 *extent_mirror_num = bbio->mirror_num;
4107 *extent_dev = bbio->stripes[0].dev;
4108 btrfs_put_bbio(bbio);