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 root_key;
651 struct btrfs_key key;
653 root_key.objectid = root;
654 root_key.type = BTRFS_ROOT_ITEM_KEY;
655 root_key.offset = (u64)-1;
656 local_root = btrfs_get_fs_root(fs_info, &root_key, true);
657 if (IS_ERR(local_root)) {
658 ret = PTR_ERR(local_root);
663 * this makes the path point to (inum INODE_ITEM ioff)
666 key.type = BTRFS_INODE_ITEM_KEY;
669 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
671 btrfs_put_root(local_root);
672 btrfs_release_path(swarn->path);
676 eb = swarn->path->nodes[0];
677 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
678 struct btrfs_inode_item);
679 isize = btrfs_inode_size(eb, inode_item);
680 nlink = btrfs_inode_nlink(eb, inode_item);
681 btrfs_release_path(swarn->path);
684 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
685 * uses GFP_NOFS in this context, so we keep it consistent but it does
686 * not seem to be strictly necessary.
688 nofs_flag = memalloc_nofs_save();
689 ipath = init_ipath(4096, local_root, swarn->path);
690 memalloc_nofs_restore(nofs_flag);
692 btrfs_put_root(local_root);
693 ret = PTR_ERR(ipath);
697 ret = paths_from_inode(inum, ipath);
703 * we deliberately ignore the bit ipath might have been too small to
704 * hold all of the paths here
706 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
707 btrfs_warn_in_rcu(fs_info,
708 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
709 swarn->errstr, swarn->logical,
710 rcu_str_deref(swarn->dev->name),
713 min(isize - offset, (u64)PAGE_SIZE), nlink,
714 (char *)(unsigned long)ipath->fspath->val[i]);
716 btrfs_put_root(local_root);
721 btrfs_warn_in_rcu(fs_info,
722 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
723 swarn->errstr, swarn->logical,
724 rcu_str_deref(swarn->dev->name),
726 root, inum, offset, ret);
732 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
734 struct btrfs_device *dev;
735 struct btrfs_fs_info *fs_info;
736 struct btrfs_path *path;
737 struct btrfs_key found_key;
738 struct extent_buffer *eb;
739 struct btrfs_extent_item *ei;
740 struct scrub_warning swarn;
741 unsigned long ptr = 0;
749 WARN_ON(sblock->page_count < 1);
750 dev = sblock->pagev[0]->dev;
751 fs_info = sblock->sctx->fs_info;
753 path = btrfs_alloc_path();
757 swarn.physical = sblock->pagev[0]->physical;
758 swarn.logical = sblock->pagev[0]->logical;
759 swarn.errstr = errstr;
762 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
767 extent_item_pos = swarn.logical - found_key.objectid;
768 swarn.extent_item_size = found_key.offset;
771 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
772 item_size = btrfs_item_size_nr(eb, path->slots[0]);
774 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
776 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
777 item_size, &ref_root,
779 btrfs_warn_in_rcu(fs_info,
780 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
781 errstr, swarn.logical,
782 rcu_str_deref(dev->name),
784 ref_level ? "node" : "leaf",
785 ret < 0 ? -1 : ref_level,
786 ret < 0 ? -1 : ref_root);
788 btrfs_release_path(path);
790 btrfs_release_path(path);
793 iterate_extent_inodes(fs_info, found_key.objectid,
795 scrub_print_warning_inode, &swarn, false);
799 btrfs_free_path(path);
802 static inline void scrub_get_recover(struct scrub_recover *recover)
804 refcount_inc(&recover->refs);
807 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
808 struct scrub_recover *recover)
810 if (refcount_dec_and_test(&recover->refs)) {
811 btrfs_bio_counter_dec(fs_info);
812 btrfs_put_bbio(recover->bbio);
818 * scrub_handle_errored_block gets called when either verification of the
819 * pages failed or the bio failed to read, e.g. with EIO. In the latter
820 * case, this function handles all pages in the bio, even though only one
822 * The goal of this function is to repair the errored block by using the
823 * contents of one of the mirrors.
825 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
827 struct scrub_ctx *sctx = sblock_to_check->sctx;
828 struct btrfs_device *dev;
829 struct btrfs_fs_info *fs_info;
831 unsigned int failed_mirror_index;
832 unsigned int is_metadata;
833 unsigned int have_csum;
834 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
835 struct scrub_block *sblock_bad;
840 bool full_stripe_locked;
841 unsigned int nofs_flag;
842 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
843 DEFAULT_RATELIMIT_BURST);
845 BUG_ON(sblock_to_check->page_count < 1);
846 fs_info = sctx->fs_info;
847 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
849 * if we find an error in a super block, we just report it.
850 * They will get written with the next transaction commit
853 spin_lock(&sctx->stat_lock);
854 ++sctx->stat.super_errors;
855 spin_unlock(&sctx->stat_lock);
858 logical = sblock_to_check->pagev[0]->logical;
859 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
860 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
861 is_metadata = !(sblock_to_check->pagev[0]->flags &
862 BTRFS_EXTENT_FLAG_DATA);
863 have_csum = sblock_to_check->pagev[0]->have_csum;
864 dev = sblock_to_check->pagev[0]->dev;
867 * We must use GFP_NOFS because the scrub task might be waiting for a
868 * worker task executing this function and in turn a transaction commit
869 * might be waiting the scrub task to pause (which needs to wait for all
870 * the worker tasks to complete before pausing).
871 * We do allocations in the workers through insert_full_stripe_lock()
872 * and scrub_add_page_to_wr_bio(), which happens down the call chain of
875 nofs_flag = memalloc_nofs_save();
877 * For RAID5/6, race can happen for a different device scrub thread.
878 * For data corruption, Parity and Data threads will both try
879 * to recovery the data.
880 * Race can lead to doubly added csum error, or even unrecoverable
883 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
885 memalloc_nofs_restore(nofs_flag);
886 spin_lock(&sctx->stat_lock);
888 sctx->stat.malloc_errors++;
889 sctx->stat.read_errors++;
890 sctx->stat.uncorrectable_errors++;
891 spin_unlock(&sctx->stat_lock);
896 * read all mirrors one after the other. This includes to
897 * re-read the extent or metadata block that failed (that was
898 * the cause that this fixup code is called) another time,
899 * page by page this time in order to know which pages
900 * caused I/O errors and which ones are good (for all mirrors).
901 * It is the goal to handle the situation when more than one
902 * mirror contains I/O errors, but the errors do not
903 * overlap, i.e. the data can be repaired by selecting the
904 * pages from those mirrors without I/O error on the
905 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
906 * would be that mirror #1 has an I/O error on the first page,
907 * the second page is good, and mirror #2 has an I/O error on
908 * the second page, but the first page is good.
909 * Then the first page of the first mirror can be repaired by
910 * taking the first page of the second mirror, and the
911 * second page of the second mirror can be repaired by
912 * copying the contents of the 2nd page of the 1st mirror.
913 * One more note: if the pages of one mirror contain I/O
914 * errors, the checksum cannot be verified. In order to get
915 * the best data for repairing, the first attempt is to find
916 * a mirror without I/O errors and with a validated checksum.
917 * Only if this is not possible, the pages are picked from
918 * mirrors with I/O errors without considering the checksum.
919 * If the latter is the case, at the end, the checksum of the
920 * repaired area is verified in order to correctly maintain
924 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
925 sizeof(*sblocks_for_recheck), GFP_KERNEL);
926 if (!sblocks_for_recheck) {
927 spin_lock(&sctx->stat_lock);
928 sctx->stat.malloc_errors++;
929 sctx->stat.read_errors++;
930 sctx->stat.uncorrectable_errors++;
931 spin_unlock(&sctx->stat_lock);
932 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
936 /* setup the context, map the logical blocks and alloc the pages */
937 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
939 spin_lock(&sctx->stat_lock);
940 sctx->stat.read_errors++;
941 sctx->stat.uncorrectable_errors++;
942 spin_unlock(&sctx->stat_lock);
943 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
946 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
947 sblock_bad = sblocks_for_recheck + failed_mirror_index;
949 /* build and submit the bios for the failed mirror, check checksums */
950 scrub_recheck_block(fs_info, sblock_bad, 1);
952 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
953 sblock_bad->no_io_error_seen) {
955 * the error disappeared after reading page by page, or
956 * the area was part of a huge bio and other parts of the
957 * bio caused I/O errors, or the block layer merged several
958 * read requests into one and the error is caused by a
959 * different bio (usually one of the two latter cases is
962 spin_lock(&sctx->stat_lock);
963 sctx->stat.unverified_errors++;
964 sblock_to_check->data_corrected = 1;
965 spin_unlock(&sctx->stat_lock);
967 if (sctx->is_dev_replace)
968 scrub_write_block_to_dev_replace(sblock_bad);
972 if (!sblock_bad->no_io_error_seen) {
973 spin_lock(&sctx->stat_lock);
974 sctx->stat.read_errors++;
975 spin_unlock(&sctx->stat_lock);
976 if (__ratelimit(&_rs))
977 scrub_print_warning("i/o error", sblock_to_check);
978 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
979 } else if (sblock_bad->checksum_error) {
980 spin_lock(&sctx->stat_lock);
981 sctx->stat.csum_errors++;
982 spin_unlock(&sctx->stat_lock);
983 if (__ratelimit(&_rs))
984 scrub_print_warning("checksum error", sblock_to_check);
985 btrfs_dev_stat_inc_and_print(dev,
986 BTRFS_DEV_STAT_CORRUPTION_ERRS);
987 } else if (sblock_bad->header_error) {
988 spin_lock(&sctx->stat_lock);
989 sctx->stat.verify_errors++;
990 spin_unlock(&sctx->stat_lock);
991 if (__ratelimit(&_rs))
992 scrub_print_warning("checksum/header error",
994 if (sblock_bad->generation_error)
995 btrfs_dev_stat_inc_and_print(dev,
996 BTRFS_DEV_STAT_GENERATION_ERRS);
998 btrfs_dev_stat_inc_and_print(dev,
999 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1002 if (sctx->readonly) {
1003 ASSERT(!sctx->is_dev_replace);
1008 * now build and submit the bios for the other mirrors, check
1010 * First try to pick the mirror which is completely without I/O
1011 * errors and also does not have a checksum error.
1012 * If one is found, and if a checksum is present, the full block
1013 * that is known to contain an error is rewritten. Afterwards
1014 * the block is known to be corrected.
1015 * If a mirror is found which is completely correct, and no
1016 * checksum is present, only those pages are rewritten that had
1017 * an I/O error in the block to be repaired, since it cannot be
1018 * determined, which copy of the other pages is better (and it
1019 * could happen otherwise that a correct page would be
1020 * overwritten by a bad one).
1022 for (mirror_index = 0; ;mirror_index++) {
1023 struct scrub_block *sblock_other;
1025 if (mirror_index == failed_mirror_index)
1028 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1029 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1030 if (mirror_index >= BTRFS_MAX_MIRRORS)
1032 if (!sblocks_for_recheck[mirror_index].page_count)
1035 sblock_other = sblocks_for_recheck + mirror_index;
1037 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1038 int max_allowed = r->bbio->num_stripes -
1039 r->bbio->num_tgtdevs;
1041 if (mirror_index >= max_allowed)
1043 if (!sblocks_for_recheck[1].page_count)
1046 ASSERT(failed_mirror_index == 0);
1047 sblock_other = sblocks_for_recheck + 1;
1048 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1051 /* build and submit the bios, check checksums */
1052 scrub_recheck_block(fs_info, sblock_other, 0);
1054 if (!sblock_other->header_error &&
1055 !sblock_other->checksum_error &&
1056 sblock_other->no_io_error_seen) {
1057 if (sctx->is_dev_replace) {
1058 scrub_write_block_to_dev_replace(sblock_other);
1059 goto corrected_error;
1061 ret = scrub_repair_block_from_good_copy(
1062 sblock_bad, sblock_other);
1064 goto corrected_error;
1069 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1070 goto did_not_correct_error;
1073 * In case of I/O errors in the area that is supposed to be
1074 * repaired, continue by picking good copies of those pages.
1075 * Select the good pages from mirrors to rewrite bad pages from
1076 * the area to fix. Afterwards verify the checksum of the block
1077 * that is supposed to be repaired. This verification step is
1078 * only done for the purpose of statistic counting and for the
1079 * final scrub report, whether errors remain.
1080 * A perfect algorithm could make use of the checksum and try
1081 * all possible combinations of pages from the different mirrors
1082 * until the checksum verification succeeds. For example, when
1083 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1084 * of mirror #2 is readable but the final checksum test fails,
1085 * then the 2nd page of mirror #3 could be tried, whether now
1086 * the final checksum succeeds. But this would be a rare
1087 * exception and is therefore not implemented. At least it is
1088 * avoided that the good copy is overwritten.
1089 * A more useful improvement would be to pick the sectors
1090 * without I/O error based on sector sizes (512 bytes on legacy
1091 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1092 * mirror could be repaired by taking 512 byte of a different
1093 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1094 * area are unreadable.
1097 for (page_num = 0; page_num < sblock_bad->page_count;
1099 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1100 struct scrub_block *sblock_other = NULL;
1102 /* skip no-io-error page in scrub */
1103 if (!page_bad->io_error && !sctx->is_dev_replace)
1106 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1108 * In case of dev replace, if raid56 rebuild process
1109 * didn't work out correct data, then copy the content
1110 * in sblock_bad to make sure target device is identical
1111 * to source device, instead of writing garbage data in
1112 * sblock_for_recheck array to target device.
1114 sblock_other = NULL;
1115 } else if (page_bad->io_error) {
1116 /* try to find no-io-error page in mirrors */
1117 for (mirror_index = 0;
1118 mirror_index < BTRFS_MAX_MIRRORS &&
1119 sblocks_for_recheck[mirror_index].page_count > 0;
1121 if (!sblocks_for_recheck[mirror_index].
1122 pagev[page_num]->io_error) {
1123 sblock_other = sblocks_for_recheck +
1132 if (sctx->is_dev_replace) {
1134 * did not find a mirror to fetch the page
1135 * from. scrub_write_page_to_dev_replace()
1136 * handles this case (page->io_error), by
1137 * filling the block with zeros before
1138 * submitting the write request
1141 sblock_other = sblock_bad;
1143 if (scrub_write_page_to_dev_replace(sblock_other,
1146 &fs_info->dev_replace.num_write_errors);
1149 } else if (sblock_other) {
1150 ret = scrub_repair_page_from_good_copy(sblock_bad,
1154 page_bad->io_error = 0;
1160 if (success && !sctx->is_dev_replace) {
1161 if (is_metadata || have_csum) {
1163 * need to verify the checksum now that all
1164 * sectors on disk are repaired (the write
1165 * request for data to be repaired is on its way).
1166 * Just be lazy and use scrub_recheck_block()
1167 * which re-reads the data before the checksum
1168 * is verified, but most likely the data comes out
1169 * of the page cache.
1171 scrub_recheck_block(fs_info, sblock_bad, 1);
1172 if (!sblock_bad->header_error &&
1173 !sblock_bad->checksum_error &&
1174 sblock_bad->no_io_error_seen)
1175 goto corrected_error;
1177 goto did_not_correct_error;
1180 spin_lock(&sctx->stat_lock);
1181 sctx->stat.corrected_errors++;
1182 sblock_to_check->data_corrected = 1;
1183 spin_unlock(&sctx->stat_lock);
1184 btrfs_err_rl_in_rcu(fs_info,
1185 "fixed up error at logical %llu on dev %s",
1186 logical, rcu_str_deref(dev->name));
1189 did_not_correct_error:
1190 spin_lock(&sctx->stat_lock);
1191 sctx->stat.uncorrectable_errors++;
1192 spin_unlock(&sctx->stat_lock);
1193 btrfs_err_rl_in_rcu(fs_info,
1194 "unable to fixup (regular) error at logical %llu on dev %s",
1195 logical, rcu_str_deref(dev->name));
1199 if (sblocks_for_recheck) {
1200 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1202 struct scrub_block *sblock = sblocks_for_recheck +
1204 struct scrub_recover *recover;
1207 for (page_index = 0; page_index < sblock->page_count;
1209 sblock->pagev[page_index]->sblock = NULL;
1210 recover = sblock->pagev[page_index]->recover;
1212 scrub_put_recover(fs_info, recover);
1213 sblock->pagev[page_index]->recover =
1216 scrub_page_put(sblock->pagev[page_index]);
1219 kfree(sblocks_for_recheck);
1222 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1223 memalloc_nofs_restore(nofs_flag);
1229 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1231 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1233 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1236 return (int)bbio->num_stripes;
1239 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1242 int nstripes, int mirror,
1248 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1250 for (i = 0; i < nstripes; i++) {
1251 if (raid_map[i] == RAID6_Q_STRIPE ||
1252 raid_map[i] == RAID5_P_STRIPE)
1255 if (logical >= raid_map[i] &&
1256 logical < raid_map[i] + mapped_length)
1261 *stripe_offset = logical - raid_map[i];
1263 /* The other RAID type */
1264 *stripe_index = mirror;
1269 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1270 struct scrub_block *sblocks_for_recheck)
1272 struct scrub_ctx *sctx = original_sblock->sctx;
1273 struct btrfs_fs_info *fs_info = sctx->fs_info;
1274 u64 length = original_sblock->page_count * PAGE_SIZE;
1275 u64 logical = original_sblock->pagev[0]->logical;
1276 u64 generation = original_sblock->pagev[0]->generation;
1277 u64 flags = original_sblock->pagev[0]->flags;
1278 u64 have_csum = original_sblock->pagev[0]->have_csum;
1279 struct scrub_recover *recover;
1280 struct btrfs_bio *bbio;
1291 * note: the two members refs and outstanding_pages
1292 * are not used (and not set) in the blocks that are used for
1293 * the recheck procedure
1296 while (length > 0) {
1297 sublen = min_t(u64, length, PAGE_SIZE);
1298 mapped_length = sublen;
1302 * with a length of PAGE_SIZE, each returned stripe
1303 * represents one mirror
1305 btrfs_bio_counter_inc_blocked(fs_info);
1306 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1307 logical, &mapped_length, &bbio);
1308 if (ret || !bbio || mapped_length < sublen) {
1309 btrfs_put_bbio(bbio);
1310 btrfs_bio_counter_dec(fs_info);
1314 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1316 btrfs_put_bbio(bbio);
1317 btrfs_bio_counter_dec(fs_info);
1321 refcount_set(&recover->refs, 1);
1322 recover->bbio = bbio;
1323 recover->map_length = mapped_length;
1325 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1327 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1329 for (mirror_index = 0; mirror_index < nmirrors;
1331 struct scrub_block *sblock;
1332 struct scrub_page *page;
1334 sblock = sblocks_for_recheck + mirror_index;
1335 sblock->sctx = sctx;
1337 page = kzalloc(sizeof(*page), GFP_NOFS);
1340 spin_lock(&sctx->stat_lock);
1341 sctx->stat.malloc_errors++;
1342 spin_unlock(&sctx->stat_lock);
1343 scrub_put_recover(fs_info, recover);
1346 scrub_page_get(page);
1347 sblock->pagev[page_index] = page;
1348 page->sblock = sblock;
1349 page->flags = flags;
1350 page->generation = generation;
1351 page->logical = logical;
1352 page->have_csum = have_csum;
1355 original_sblock->pagev[0]->csum,
1358 scrub_stripe_index_and_offset(logical,
1367 page->physical = bbio->stripes[stripe_index].physical +
1369 page->dev = bbio->stripes[stripe_index].dev;
1371 BUG_ON(page_index >= original_sblock->page_count);
1372 page->physical_for_dev_replace =
1373 original_sblock->pagev[page_index]->
1374 physical_for_dev_replace;
1375 /* for missing devices, dev->bdev is NULL */
1376 page->mirror_num = mirror_index + 1;
1377 sblock->page_count++;
1378 page->page = alloc_page(GFP_NOFS);
1382 scrub_get_recover(recover);
1383 page->recover = recover;
1385 scrub_put_recover(fs_info, recover);
1394 static void scrub_bio_wait_endio(struct bio *bio)
1396 complete(bio->bi_private);
1399 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1401 struct scrub_page *page)
1403 DECLARE_COMPLETION_ONSTACK(done);
1407 bio->bi_iter.bi_sector = page->logical >> 9;
1408 bio->bi_private = &done;
1409 bio->bi_end_io = scrub_bio_wait_endio;
1411 mirror_num = page->sblock->pagev[0]->mirror_num;
1412 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1413 page->recover->map_length,
1418 wait_for_completion_io(&done);
1419 return blk_status_to_errno(bio->bi_status);
1422 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1423 struct scrub_block *sblock)
1425 struct scrub_page *first_page = sblock->pagev[0];
1429 /* All pages in sblock belong to the same stripe on the same device. */
1430 ASSERT(first_page->dev);
1431 if (!first_page->dev->bdev)
1434 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1435 bio_set_dev(bio, first_page->dev->bdev);
1437 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1438 struct scrub_page *page = sblock->pagev[page_num];
1440 WARN_ON(!page->page);
1441 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1444 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1451 scrub_recheck_block_checksum(sblock);
1455 for (page_num = 0; page_num < sblock->page_count; page_num++)
1456 sblock->pagev[page_num]->io_error = 1;
1458 sblock->no_io_error_seen = 0;
1462 * this function will check the on disk data for checksum errors, header
1463 * errors and read I/O errors. If any I/O errors happen, the exact pages
1464 * which are errored are marked as being bad. The goal is to enable scrub
1465 * to take those pages that are not errored from all the mirrors so that
1466 * the pages that are errored in the just handled mirror can be repaired.
1468 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1469 struct scrub_block *sblock,
1470 int retry_failed_mirror)
1474 sblock->no_io_error_seen = 1;
1476 /* short cut for raid56 */
1477 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1478 return scrub_recheck_block_on_raid56(fs_info, sblock);
1480 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1482 struct scrub_page *page = sblock->pagev[page_num];
1484 if (page->dev->bdev == NULL) {
1486 sblock->no_io_error_seen = 0;
1490 WARN_ON(!page->page);
1491 bio = btrfs_io_bio_alloc(1);
1492 bio_set_dev(bio, page->dev->bdev);
1494 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1495 bio->bi_iter.bi_sector = page->physical >> 9;
1496 bio->bi_opf = REQ_OP_READ;
1498 if (btrfsic_submit_bio_wait(bio)) {
1500 sblock->no_io_error_seen = 0;
1506 if (sblock->no_io_error_seen)
1507 scrub_recheck_block_checksum(sblock);
1510 static inline int scrub_check_fsid(u8 fsid[],
1511 struct scrub_page *spage)
1513 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1516 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1520 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1522 sblock->header_error = 0;
1523 sblock->checksum_error = 0;
1524 sblock->generation_error = 0;
1526 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1527 scrub_checksum_data(sblock);
1529 scrub_checksum_tree_block(sblock);
1532 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1533 struct scrub_block *sblock_good)
1538 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1541 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1551 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1552 struct scrub_block *sblock_good,
1553 int page_num, int force_write)
1555 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1556 struct scrub_page *page_good = sblock_good->pagev[page_num];
1557 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1559 BUG_ON(page_bad->page == NULL);
1560 BUG_ON(page_good->page == NULL);
1561 if (force_write || sblock_bad->header_error ||
1562 sblock_bad->checksum_error || page_bad->io_error) {
1566 if (!page_bad->dev->bdev) {
1567 btrfs_warn_rl(fs_info,
1568 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1572 bio = btrfs_io_bio_alloc(1);
1573 bio_set_dev(bio, page_bad->dev->bdev);
1574 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1575 bio->bi_opf = REQ_OP_WRITE;
1577 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1578 if (PAGE_SIZE != ret) {
1583 if (btrfsic_submit_bio_wait(bio)) {
1584 btrfs_dev_stat_inc_and_print(page_bad->dev,
1585 BTRFS_DEV_STAT_WRITE_ERRS);
1586 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1596 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1598 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1602 * This block is used for the check of the parity on the source device,
1603 * so the data needn't be written into the destination device.
1605 if (sblock->sparity)
1608 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1611 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1613 atomic64_inc(&fs_info->dev_replace.num_write_errors);
1617 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1620 struct scrub_page *spage = sblock->pagev[page_num];
1622 BUG_ON(spage->page == NULL);
1623 if (spage->io_error) {
1624 void *mapped_buffer = kmap_atomic(spage->page);
1626 clear_page(mapped_buffer);
1627 flush_dcache_page(spage->page);
1628 kunmap_atomic(mapped_buffer);
1630 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1633 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1634 struct scrub_page *spage)
1636 struct scrub_bio *sbio;
1639 mutex_lock(&sctx->wr_lock);
1641 if (!sctx->wr_curr_bio) {
1642 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1644 if (!sctx->wr_curr_bio) {
1645 mutex_unlock(&sctx->wr_lock);
1648 sctx->wr_curr_bio->sctx = sctx;
1649 sctx->wr_curr_bio->page_count = 0;
1651 sbio = sctx->wr_curr_bio;
1652 if (sbio->page_count == 0) {
1655 sbio->physical = spage->physical_for_dev_replace;
1656 sbio->logical = spage->logical;
1657 sbio->dev = sctx->wr_tgtdev;
1660 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1664 bio->bi_private = sbio;
1665 bio->bi_end_io = scrub_wr_bio_end_io;
1666 bio_set_dev(bio, sbio->dev->bdev);
1667 bio->bi_iter.bi_sector = sbio->physical >> 9;
1668 bio->bi_opf = REQ_OP_WRITE;
1670 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1671 spage->physical_for_dev_replace ||
1672 sbio->logical + sbio->page_count * PAGE_SIZE !=
1674 scrub_wr_submit(sctx);
1678 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1679 if (ret != PAGE_SIZE) {
1680 if (sbio->page_count < 1) {
1683 mutex_unlock(&sctx->wr_lock);
1686 scrub_wr_submit(sctx);
1690 sbio->pagev[sbio->page_count] = spage;
1691 scrub_page_get(spage);
1693 if (sbio->page_count == sctx->pages_per_wr_bio)
1694 scrub_wr_submit(sctx);
1695 mutex_unlock(&sctx->wr_lock);
1700 static void scrub_wr_submit(struct scrub_ctx *sctx)
1702 struct scrub_bio *sbio;
1704 if (!sctx->wr_curr_bio)
1707 sbio = sctx->wr_curr_bio;
1708 sctx->wr_curr_bio = NULL;
1709 WARN_ON(!sbio->bio->bi_disk);
1710 scrub_pending_bio_inc(sctx);
1711 /* process all writes in a single worker thread. Then the block layer
1712 * orders the requests before sending them to the driver which
1713 * doubled the write performance on spinning disks when measured
1715 btrfsic_submit_bio(sbio->bio);
1718 static void scrub_wr_bio_end_io(struct bio *bio)
1720 struct scrub_bio *sbio = bio->bi_private;
1721 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1723 sbio->status = bio->bi_status;
1726 btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1727 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1730 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1732 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1733 struct scrub_ctx *sctx = sbio->sctx;
1736 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1738 struct btrfs_dev_replace *dev_replace =
1739 &sbio->sctx->fs_info->dev_replace;
1741 for (i = 0; i < sbio->page_count; i++) {
1742 struct scrub_page *spage = sbio->pagev[i];
1744 spage->io_error = 1;
1745 atomic64_inc(&dev_replace->num_write_errors);
1749 for (i = 0; i < sbio->page_count; i++)
1750 scrub_page_put(sbio->pagev[i]);
1754 scrub_pending_bio_dec(sctx);
1757 static int scrub_checksum(struct scrub_block *sblock)
1763 * No need to initialize these stats currently,
1764 * because this function only use return value
1765 * instead of these stats value.
1770 sblock->header_error = 0;
1771 sblock->generation_error = 0;
1772 sblock->checksum_error = 0;
1774 WARN_ON(sblock->page_count < 1);
1775 flags = sblock->pagev[0]->flags;
1777 if (flags & BTRFS_EXTENT_FLAG_DATA)
1778 ret = scrub_checksum_data(sblock);
1779 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1780 ret = scrub_checksum_tree_block(sblock);
1781 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1782 (void)scrub_checksum_super(sblock);
1786 scrub_handle_errored_block(sblock);
1791 static int scrub_checksum_data(struct scrub_block *sblock)
1793 struct scrub_ctx *sctx = sblock->sctx;
1794 struct btrfs_fs_info *fs_info = sctx->fs_info;
1795 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1796 u8 csum[BTRFS_CSUM_SIZE];
1803 BUG_ON(sblock->page_count < 1);
1804 if (!sblock->pagev[0]->have_csum)
1807 shash->tfm = fs_info->csum_shash;
1808 crypto_shash_init(shash);
1810 on_disk_csum = sblock->pagev[0]->csum;
1811 page = sblock->pagev[0]->page;
1812 buffer = kmap_atomic(page);
1814 len = sctx->fs_info->sectorsize;
1817 u64 l = min_t(u64, len, PAGE_SIZE);
1819 crypto_shash_update(shash, buffer, l);
1820 kunmap_atomic(buffer);
1825 BUG_ON(index >= sblock->page_count);
1826 BUG_ON(!sblock->pagev[index]->page);
1827 page = sblock->pagev[index]->page;
1828 buffer = kmap_atomic(page);
1831 crypto_shash_final(shash, csum);
1832 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1833 sblock->checksum_error = 1;
1835 return sblock->checksum_error;
1838 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1840 struct scrub_ctx *sctx = sblock->sctx;
1841 struct btrfs_header *h;
1842 struct btrfs_fs_info *fs_info = sctx->fs_info;
1843 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1844 u8 calculated_csum[BTRFS_CSUM_SIZE];
1845 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1847 void *mapped_buffer;
1853 shash->tfm = fs_info->csum_shash;
1854 crypto_shash_init(shash);
1856 BUG_ON(sblock->page_count < 1);
1857 page = sblock->pagev[0]->page;
1858 mapped_buffer = kmap_atomic(page);
1859 h = (struct btrfs_header *)mapped_buffer;
1860 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1863 * we don't use the getter functions here, as we
1864 * a) don't have an extent buffer and
1865 * b) the page is already kmapped
1867 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1868 sblock->header_error = 1;
1870 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1871 sblock->header_error = 1;
1872 sblock->generation_error = 1;
1875 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1876 sblock->header_error = 1;
1878 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1880 sblock->header_error = 1;
1882 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1883 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1884 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1887 u64 l = min_t(u64, len, mapped_size);
1889 crypto_shash_update(shash, p, l);
1890 kunmap_atomic(mapped_buffer);
1895 BUG_ON(index >= sblock->page_count);
1896 BUG_ON(!sblock->pagev[index]->page);
1897 page = sblock->pagev[index]->page;
1898 mapped_buffer = kmap_atomic(page);
1899 mapped_size = PAGE_SIZE;
1903 crypto_shash_final(shash, calculated_csum);
1904 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1905 sblock->checksum_error = 1;
1907 return sblock->header_error || sblock->checksum_error;
1910 static int scrub_checksum_super(struct scrub_block *sblock)
1912 struct btrfs_super_block *s;
1913 struct scrub_ctx *sctx = sblock->sctx;
1914 struct btrfs_fs_info *fs_info = sctx->fs_info;
1915 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1916 u8 calculated_csum[BTRFS_CSUM_SIZE];
1917 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1919 void *mapped_buffer;
1927 shash->tfm = fs_info->csum_shash;
1928 crypto_shash_init(shash);
1930 BUG_ON(sblock->page_count < 1);
1931 page = sblock->pagev[0]->page;
1932 mapped_buffer = kmap_atomic(page);
1933 s = (struct btrfs_super_block *)mapped_buffer;
1934 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1936 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1939 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1942 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1945 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1946 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1947 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1950 u64 l = min_t(u64, len, mapped_size);
1952 crypto_shash_update(shash, p, l);
1953 kunmap_atomic(mapped_buffer);
1958 BUG_ON(index >= sblock->page_count);
1959 BUG_ON(!sblock->pagev[index]->page);
1960 page = sblock->pagev[index]->page;
1961 mapped_buffer = kmap_atomic(page);
1962 mapped_size = PAGE_SIZE;
1966 crypto_shash_final(shash, calculated_csum);
1967 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1970 if (fail_cor + fail_gen) {
1972 * if we find an error in a super block, we just report it.
1973 * They will get written with the next transaction commit
1976 spin_lock(&sctx->stat_lock);
1977 ++sctx->stat.super_errors;
1978 spin_unlock(&sctx->stat_lock);
1980 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1981 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1983 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1984 BTRFS_DEV_STAT_GENERATION_ERRS);
1987 return fail_cor + fail_gen;
1990 static void scrub_block_get(struct scrub_block *sblock)
1992 refcount_inc(&sblock->refs);
1995 static void scrub_block_put(struct scrub_block *sblock)
1997 if (refcount_dec_and_test(&sblock->refs)) {
2000 if (sblock->sparity)
2001 scrub_parity_put(sblock->sparity);
2003 for (i = 0; i < sblock->page_count; i++)
2004 scrub_page_put(sblock->pagev[i]);
2009 static void scrub_page_get(struct scrub_page *spage)
2011 atomic_inc(&spage->refs);
2014 static void scrub_page_put(struct scrub_page *spage)
2016 if (atomic_dec_and_test(&spage->refs)) {
2018 __free_page(spage->page);
2023 static void scrub_submit(struct scrub_ctx *sctx)
2025 struct scrub_bio *sbio;
2027 if (sctx->curr == -1)
2030 sbio = sctx->bios[sctx->curr];
2032 scrub_pending_bio_inc(sctx);
2033 btrfsic_submit_bio(sbio->bio);
2036 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2037 struct scrub_page *spage)
2039 struct scrub_block *sblock = spage->sblock;
2040 struct scrub_bio *sbio;
2045 * grab a fresh bio or wait for one to become available
2047 while (sctx->curr == -1) {
2048 spin_lock(&sctx->list_lock);
2049 sctx->curr = sctx->first_free;
2050 if (sctx->curr != -1) {
2051 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2052 sctx->bios[sctx->curr]->next_free = -1;
2053 sctx->bios[sctx->curr]->page_count = 0;
2054 spin_unlock(&sctx->list_lock);
2056 spin_unlock(&sctx->list_lock);
2057 wait_event(sctx->list_wait, sctx->first_free != -1);
2060 sbio = sctx->bios[sctx->curr];
2061 if (sbio->page_count == 0) {
2064 sbio->physical = spage->physical;
2065 sbio->logical = spage->logical;
2066 sbio->dev = spage->dev;
2069 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2073 bio->bi_private = sbio;
2074 bio->bi_end_io = scrub_bio_end_io;
2075 bio_set_dev(bio, sbio->dev->bdev);
2076 bio->bi_iter.bi_sector = sbio->physical >> 9;
2077 bio->bi_opf = REQ_OP_READ;
2079 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2081 sbio->logical + sbio->page_count * PAGE_SIZE !=
2083 sbio->dev != spage->dev) {
2088 sbio->pagev[sbio->page_count] = spage;
2089 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2090 if (ret != PAGE_SIZE) {
2091 if (sbio->page_count < 1) {
2100 scrub_block_get(sblock); /* one for the page added to the bio */
2101 atomic_inc(&sblock->outstanding_pages);
2103 if (sbio->page_count == sctx->pages_per_rd_bio)
2109 static void scrub_missing_raid56_end_io(struct bio *bio)
2111 struct scrub_block *sblock = bio->bi_private;
2112 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2115 sblock->no_io_error_seen = 0;
2119 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2122 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2124 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2125 struct scrub_ctx *sctx = sblock->sctx;
2126 struct btrfs_fs_info *fs_info = sctx->fs_info;
2128 struct btrfs_device *dev;
2130 logical = sblock->pagev[0]->logical;
2131 dev = sblock->pagev[0]->dev;
2133 if (sblock->no_io_error_seen)
2134 scrub_recheck_block_checksum(sblock);
2136 if (!sblock->no_io_error_seen) {
2137 spin_lock(&sctx->stat_lock);
2138 sctx->stat.read_errors++;
2139 spin_unlock(&sctx->stat_lock);
2140 btrfs_err_rl_in_rcu(fs_info,
2141 "IO error rebuilding logical %llu for dev %s",
2142 logical, rcu_str_deref(dev->name));
2143 } else if (sblock->header_error || sblock->checksum_error) {
2144 spin_lock(&sctx->stat_lock);
2145 sctx->stat.uncorrectable_errors++;
2146 spin_unlock(&sctx->stat_lock);
2147 btrfs_err_rl_in_rcu(fs_info,
2148 "failed to rebuild valid logical %llu for dev %s",
2149 logical, rcu_str_deref(dev->name));
2151 scrub_write_block_to_dev_replace(sblock);
2154 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2155 mutex_lock(&sctx->wr_lock);
2156 scrub_wr_submit(sctx);
2157 mutex_unlock(&sctx->wr_lock);
2160 scrub_block_put(sblock);
2161 scrub_pending_bio_dec(sctx);
2164 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2166 struct scrub_ctx *sctx = sblock->sctx;
2167 struct btrfs_fs_info *fs_info = sctx->fs_info;
2168 u64 length = sblock->page_count * PAGE_SIZE;
2169 u64 logical = sblock->pagev[0]->logical;
2170 struct btrfs_bio *bbio = NULL;
2172 struct btrfs_raid_bio *rbio;
2176 btrfs_bio_counter_inc_blocked(fs_info);
2177 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2179 if (ret || !bbio || !bbio->raid_map)
2182 if (WARN_ON(!sctx->is_dev_replace ||
2183 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2185 * We shouldn't be scrubbing a missing device. Even for dev
2186 * replace, we should only get here for RAID 5/6. We either
2187 * managed to mount something with no mirrors remaining or
2188 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2193 bio = btrfs_io_bio_alloc(0);
2194 bio->bi_iter.bi_sector = logical >> 9;
2195 bio->bi_private = sblock;
2196 bio->bi_end_io = scrub_missing_raid56_end_io;
2198 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2202 for (i = 0; i < sblock->page_count; i++) {
2203 struct scrub_page *spage = sblock->pagev[i];
2205 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2208 btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2209 scrub_block_get(sblock);
2210 scrub_pending_bio_inc(sctx);
2211 raid56_submit_missing_rbio(rbio);
2217 btrfs_bio_counter_dec(fs_info);
2218 btrfs_put_bbio(bbio);
2219 spin_lock(&sctx->stat_lock);
2220 sctx->stat.malloc_errors++;
2221 spin_unlock(&sctx->stat_lock);
2224 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2225 u64 physical, struct btrfs_device *dev, u64 flags,
2226 u64 gen, int mirror_num, u8 *csum, int force,
2227 u64 physical_for_dev_replace)
2229 struct scrub_block *sblock;
2232 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2234 spin_lock(&sctx->stat_lock);
2235 sctx->stat.malloc_errors++;
2236 spin_unlock(&sctx->stat_lock);
2240 /* one ref inside this function, plus one for each page added to
2242 refcount_set(&sblock->refs, 1);
2243 sblock->sctx = sctx;
2244 sblock->no_io_error_seen = 1;
2246 for (index = 0; len > 0; index++) {
2247 struct scrub_page *spage;
2248 u64 l = min_t(u64, len, PAGE_SIZE);
2250 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2253 spin_lock(&sctx->stat_lock);
2254 sctx->stat.malloc_errors++;
2255 spin_unlock(&sctx->stat_lock);
2256 scrub_block_put(sblock);
2259 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2260 scrub_page_get(spage);
2261 sblock->pagev[index] = spage;
2262 spage->sblock = sblock;
2264 spage->flags = flags;
2265 spage->generation = gen;
2266 spage->logical = logical;
2267 spage->physical = physical;
2268 spage->physical_for_dev_replace = physical_for_dev_replace;
2269 spage->mirror_num = mirror_num;
2271 spage->have_csum = 1;
2272 memcpy(spage->csum, csum, sctx->csum_size);
2274 spage->have_csum = 0;
2276 sblock->page_count++;
2277 spage->page = alloc_page(GFP_KERNEL);
2283 physical_for_dev_replace += l;
2286 WARN_ON(sblock->page_count == 0);
2287 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2289 * This case should only be hit for RAID 5/6 device replace. See
2290 * the comment in scrub_missing_raid56_pages() for details.
2292 scrub_missing_raid56_pages(sblock);
2294 for (index = 0; index < sblock->page_count; index++) {
2295 struct scrub_page *spage = sblock->pagev[index];
2298 ret = scrub_add_page_to_rd_bio(sctx, spage);
2300 scrub_block_put(sblock);
2309 /* last one frees, either here or in bio completion for last page */
2310 scrub_block_put(sblock);
2314 static void scrub_bio_end_io(struct bio *bio)
2316 struct scrub_bio *sbio = bio->bi_private;
2317 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2319 sbio->status = bio->bi_status;
2322 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2325 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2327 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2328 struct scrub_ctx *sctx = sbio->sctx;
2331 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2333 for (i = 0; i < sbio->page_count; i++) {
2334 struct scrub_page *spage = sbio->pagev[i];
2336 spage->io_error = 1;
2337 spage->sblock->no_io_error_seen = 0;
2341 /* now complete the scrub_block items that have all pages completed */
2342 for (i = 0; i < sbio->page_count; i++) {
2343 struct scrub_page *spage = sbio->pagev[i];
2344 struct scrub_block *sblock = spage->sblock;
2346 if (atomic_dec_and_test(&sblock->outstanding_pages))
2347 scrub_block_complete(sblock);
2348 scrub_block_put(sblock);
2353 spin_lock(&sctx->list_lock);
2354 sbio->next_free = sctx->first_free;
2355 sctx->first_free = sbio->index;
2356 spin_unlock(&sctx->list_lock);
2358 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2359 mutex_lock(&sctx->wr_lock);
2360 scrub_wr_submit(sctx);
2361 mutex_unlock(&sctx->wr_lock);
2364 scrub_pending_bio_dec(sctx);
2367 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2368 unsigned long *bitmap,
2374 int sectorsize = sparity->sctx->fs_info->sectorsize;
2376 if (len >= sparity->stripe_len) {
2377 bitmap_set(bitmap, 0, sparity->nsectors);
2381 start -= sparity->logic_start;
2382 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2383 offset = div_u64(offset, sectorsize);
2384 nsectors64 = div_u64(len, sectorsize);
2386 ASSERT(nsectors64 < UINT_MAX);
2387 nsectors = (u32)nsectors64;
2389 if (offset + nsectors <= sparity->nsectors) {
2390 bitmap_set(bitmap, offset, nsectors);
2394 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2395 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2398 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2401 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2404 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2407 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2410 static void scrub_block_complete(struct scrub_block *sblock)
2414 if (!sblock->no_io_error_seen) {
2416 scrub_handle_errored_block(sblock);
2419 * if has checksum error, write via repair mechanism in
2420 * dev replace case, otherwise write here in dev replace
2423 corrupted = scrub_checksum(sblock);
2424 if (!corrupted && sblock->sctx->is_dev_replace)
2425 scrub_write_block_to_dev_replace(sblock);
2428 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2429 u64 start = sblock->pagev[0]->logical;
2430 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2433 scrub_parity_mark_sectors_error(sblock->sparity,
2434 start, end - start);
2438 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2440 struct btrfs_ordered_sum *sum = NULL;
2441 unsigned long index;
2442 unsigned long num_sectors;
2444 while (!list_empty(&sctx->csum_list)) {
2445 sum = list_first_entry(&sctx->csum_list,
2446 struct btrfs_ordered_sum, list);
2447 if (sum->bytenr > logical)
2449 if (sum->bytenr + sum->len > logical)
2452 ++sctx->stat.csum_discards;
2453 list_del(&sum->list);
2460 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2461 ASSERT(index < UINT_MAX);
2463 num_sectors = sum->len / sctx->fs_info->sectorsize;
2464 memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2465 if (index == num_sectors - 1) {
2466 list_del(&sum->list);
2472 /* scrub extent tries to collect up to 64 kB for each bio */
2473 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2474 u64 logical, u64 len,
2475 u64 physical, struct btrfs_device *dev, u64 flags,
2476 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2479 u8 csum[BTRFS_CSUM_SIZE];
2482 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2483 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2484 blocksize = map->stripe_len;
2486 blocksize = sctx->fs_info->sectorsize;
2487 spin_lock(&sctx->stat_lock);
2488 sctx->stat.data_extents_scrubbed++;
2489 sctx->stat.data_bytes_scrubbed += len;
2490 spin_unlock(&sctx->stat_lock);
2491 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2492 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2493 blocksize = map->stripe_len;
2495 blocksize = sctx->fs_info->nodesize;
2496 spin_lock(&sctx->stat_lock);
2497 sctx->stat.tree_extents_scrubbed++;
2498 sctx->stat.tree_bytes_scrubbed += len;
2499 spin_unlock(&sctx->stat_lock);
2501 blocksize = sctx->fs_info->sectorsize;
2506 u64 l = min_t(u64, len, blocksize);
2509 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2510 /* push csums to sbio */
2511 have_csum = scrub_find_csum(sctx, logical, csum);
2513 ++sctx->stat.no_csum;
2515 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2516 mirror_num, have_csum ? csum : NULL, 0,
2517 physical_for_dev_replace);
2523 physical_for_dev_replace += l;
2528 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2529 u64 logical, u64 len,
2530 u64 physical, struct btrfs_device *dev,
2531 u64 flags, u64 gen, int mirror_num, u8 *csum)
2533 struct scrub_ctx *sctx = sparity->sctx;
2534 struct scrub_block *sblock;
2537 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2539 spin_lock(&sctx->stat_lock);
2540 sctx->stat.malloc_errors++;
2541 spin_unlock(&sctx->stat_lock);
2545 /* one ref inside this function, plus one for each page added to
2547 refcount_set(&sblock->refs, 1);
2548 sblock->sctx = sctx;
2549 sblock->no_io_error_seen = 1;
2550 sblock->sparity = sparity;
2551 scrub_parity_get(sparity);
2553 for (index = 0; len > 0; index++) {
2554 struct scrub_page *spage;
2555 u64 l = min_t(u64, len, PAGE_SIZE);
2557 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2560 spin_lock(&sctx->stat_lock);
2561 sctx->stat.malloc_errors++;
2562 spin_unlock(&sctx->stat_lock);
2563 scrub_block_put(sblock);
2566 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2567 /* For scrub block */
2568 scrub_page_get(spage);
2569 sblock->pagev[index] = spage;
2570 /* For scrub parity */
2571 scrub_page_get(spage);
2572 list_add_tail(&spage->list, &sparity->spages);
2573 spage->sblock = sblock;
2575 spage->flags = flags;
2576 spage->generation = gen;
2577 spage->logical = logical;
2578 spage->physical = physical;
2579 spage->mirror_num = mirror_num;
2581 spage->have_csum = 1;
2582 memcpy(spage->csum, csum, sctx->csum_size);
2584 spage->have_csum = 0;
2586 sblock->page_count++;
2587 spage->page = alloc_page(GFP_KERNEL);
2595 WARN_ON(sblock->page_count == 0);
2596 for (index = 0; index < sblock->page_count; index++) {
2597 struct scrub_page *spage = sblock->pagev[index];
2600 ret = scrub_add_page_to_rd_bio(sctx, spage);
2602 scrub_block_put(sblock);
2607 /* last one frees, either here or in bio completion for last page */
2608 scrub_block_put(sblock);
2612 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2613 u64 logical, u64 len,
2614 u64 physical, struct btrfs_device *dev,
2615 u64 flags, u64 gen, int mirror_num)
2617 struct scrub_ctx *sctx = sparity->sctx;
2619 u8 csum[BTRFS_CSUM_SIZE];
2622 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2623 scrub_parity_mark_sectors_error(sparity, logical, len);
2627 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2628 blocksize = sparity->stripe_len;
2629 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2630 blocksize = sparity->stripe_len;
2632 blocksize = sctx->fs_info->sectorsize;
2637 u64 l = min_t(u64, len, blocksize);
2640 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2641 /* push csums to sbio */
2642 have_csum = scrub_find_csum(sctx, logical, csum);
2646 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2647 flags, gen, mirror_num,
2648 have_csum ? csum : NULL);
2660 * Given a physical address, this will calculate it's
2661 * logical offset. if this is a parity stripe, it will return
2662 * the most left data stripe's logical offset.
2664 * return 0 if it is a data stripe, 1 means parity stripe.
2666 static int get_raid56_logic_offset(u64 physical, int num,
2667 struct map_lookup *map, u64 *offset,
2676 const int data_stripes = nr_data_stripes(map);
2678 last_offset = (physical - map->stripes[num].physical) * data_stripes;
2680 *stripe_start = last_offset;
2682 *offset = last_offset;
2683 for (i = 0; i < data_stripes; i++) {
2684 *offset = last_offset + i * map->stripe_len;
2686 stripe_nr = div64_u64(*offset, map->stripe_len);
2687 stripe_nr = div_u64(stripe_nr, data_stripes);
2689 /* Work out the disk rotation on this stripe-set */
2690 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2691 /* calculate which stripe this data locates */
2693 stripe_index = rot % map->num_stripes;
2694 if (stripe_index == num)
2696 if (stripe_index < num)
2699 *offset = last_offset + j * map->stripe_len;
2703 static void scrub_free_parity(struct scrub_parity *sparity)
2705 struct scrub_ctx *sctx = sparity->sctx;
2706 struct scrub_page *curr, *next;
2709 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2711 spin_lock(&sctx->stat_lock);
2712 sctx->stat.read_errors += nbits;
2713 sctx->stat.uncorrectable_errors += nbits;
2714 spin_unlock(&sctx->stat_lock);
2717 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2718 list_del_init(&curr->list);
2719 scrub_page_put(curr);
2725 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2727 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2729 struct scrub_ctx *sctx = sparity->sctx;
2731 scrub_free_parity(sparity);
2732 scrub_pending_bio_dec(sctx);
2735 static void scrub_parity_bio_endio(struct bio *bio)
2737 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2738 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2741 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2746 btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2748 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2751 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2753 struct scrub_ctx *sctx = sparity->sctx;
2754 struct btrfs_fs_info *fs_info = sctx->fs_info;
2756 struct btrfs_raid_bio *rbio;
2757 struct btrfs_bio *bbio = NULL;
2761 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2765 length = sparity->logic_end - sparity->logic_start;
2767 btrfs_bio_counter_inc_blocked(fs_info);
2768 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2770 if (ret || !bbio || !bbio->raid_map)
2773 bio = btrfs_io_bio_alloc(0);
2774 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2775 bio->bi_private = sparity;
2776 bio->bi_end_io = scrub_parity_bio_endio;
2778 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2779 length, sparity->scrub_dev,
2785 scrub_pending_bio_inc(sctx);
2786 raid56_parity_submit_scrub_rbio(rbio);
2792 btrfs_bio_counter_dec(fs_info);
2793 btrfs_put_bbio(bbio);
2794 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2796 spin_lock(&sctx->stat_lock);
2797 sctx->stat.malloc_errors++;
2798 spin_unlock(&sctx->stat_lock);
2800 scrub_free_parity(sparity);
2803 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2805 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2808 static void scrub_parity_get(struct scrub_parity *sparity)
2810 refcount_inc(&sparity->refs);
2813 static void scrub_parity_put(struct scrub_parity *sparity)
2815 if (!refcount_dec_and_test(&sparity->refs))
2818 scrub_parity_check_and_repair(sparity);
2821 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2822 struct map_lookup *map,
2823 struct btrfs_device *sdev,
2824 struct btrfs_path *path,
2828 struct btrfs_fs_info *fs_info = sctx->fs_info;
2829 struct btrfs_root *root = fs_info->extent_root;
2830 struct btrfs_root *csum_root = fs_info->csum_root;
2831 struct btrfs_extent_item *extent;
2832 struct btrfs_bio *bbio = NULL;
2836 struct extent_buffer *l;
2837 struct btrfs_key key;
2840 u64 extent_physical;
2843 struct btrfs_device *extent_dev;
2844 struct scrub_parity *sparity;
2847 int extent_mirror_num;
2850 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2851 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2852 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2855 spin_lock(&sctx->stat_lock);
2856 sctx->stat.malloc_errors++;
2857 spin_unlock(&sctx->stat_lock);
2861 sparity->stripe_len = map->stripe_len;
2862 sparity->nsectors = nsectors;
2863 sparity->sctx = sctx;
2864 sparity->scrub_dev = sdev;
2865 sparity->logic_start = logic_start;
2866 sparity->logic_end = logic_end;
2867 refcount_set(&sparity->refs, 1);
2868 INIT_LIST_HEAD(&sparity->spages);
2869 sparity->dbitmap = sparity->bitmap;
2870 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2873 while (logic_start < logic_end) {
2874 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2875 key.type = BTRFS_METADATA_ITEM_KEY;
2877 key.type = BTRFS_EXTENT_ITEM_KEY;
2878 key.objectid = logic_start;
2879 key.offset = (u64)-1;
2881 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2886 ret = btrfs_previous_extent_item(root, path, 0);
2890 btrfs_release_path(path);
2891 ret = btrfs_search_slot(NULL, root, &key,
2903 slot = path->slots[0];
2904 if (slot >= btrfs_header_nritems(l)) {
2905 ret = btrfs_next_leaf(root, path);
2914 btrfs_item_key_to_cpu(l, &key, slot);
2916 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2917 key.type != BTRFS_METADATA_ITEM_KEY)
2920 if (key.type == BTRFS_METADATA_ITEM_KEY)
2921 bytes = fs_info->nodesize;
2925 if (key.objectid + bytes <= logic_start)
2928 if (key.objectid >= logic_end) {
2933 while (key.objectid >= logic_start + map->stripe_len)
2934 logic_start += map->stripe_len;
2936 extent = btrfs_item_ptr(l, slot,
2937 struct btrfs_extent_item);
2938 flags = btrfs_extent_flags(l, extent);
2939 generation = btrfs_extent_generation(l, extent);
2941 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2942 (key.objectid < logic_start ||
2943 key.objectid + bytes >
2944 logic_start + map->stripe_len)) {
2946 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2947 key.objectid, logic_start);
2948 spin_lock(&sctx->stat_lock);
2949 sctx->stat.uncorrectable_errors++;
2950 spin_unlock(&sctx->stat_lock);
2954 extent_logical = key.objectid;
2957 if (extent_logical < logic_start) {
2958 extent_len -= logic_start - extent_logical;
2959 extent_logical = logic_start;
2962 if (extent_logical + extent_len >
2963 logic_start + map->stripe_len)
2964 extent_len = logic_start + map->stripe_len -
2967 scrub_parity_mark_sectors_data(sparity, extent_logical,
2970 mapped_length = extent_len;
2972 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2973 extent_logical, &mapped_length, &bbio,
2976 if (!bbio || mapped_length < extent_len)
2980 btrfs_put_bbio(bbio);
2983 extent_physical = bbio->stripes[0].physical;
2984 extent_mirror_num = bbio->mirror_num;
2985 extent_dev = bbio->stripes[0].dev;
2986 btrfs_put_bbio(bbio);
2988 ret = btrfs_lookup_csums_range(csum_root,
2990 extent_logical + extent_len - 1,
2991 &sctx->csum_list, 1);
2995 ret = scrub_extent_for_parity(sparity, extent_logical,
3002 scrub_free_csums(sctx);
3007 if (extent_logical + extent_len <
3008 key.objectid + bytes) {
3009 logic_start += map->stripe_len;
3011 if (logic_start >= logic_end) {
3016 if (logic_start < key.objectid + bytes) {
3025 btrfs_release_path(path);
3030 logic_start += map->stripe_len;
3034 scrub_parity_mark_sectors_error(sparity, logic_start,
3035 logic_end - logic_start);
3036 scrub_parity_put(sparity);
3038 mutex_lock(&sctx->wr_lock);
3039 scrub_wr_submit(sctx);
3040 mutex_unlock(&sctx->wr_lock);
3042 btrfs_release_path(path);
3043 return ret < 0 ? ret : 0;
3046 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3047 struct map_lookup *map,
3048 struct btrfs_device *scrub_dev,
3049 int num, u64 base, u64 length)
3051 struct btrfs_path *path, *ppath;
3052 struct btrfs_fs_info *fs_info = sctx->fs_info;
3053 struct btrfs_root *root = fs_info->extent_root;
3054 struct btrfs_root *csum_root = fs_info->csum_root;
3055 struct btrfs_extent_item *extent;
3056 struct blk_plug plug;
3061 struct extent_buffer *l;
3068 struct reada_control *reada1;
3069 struct reada_control *reada2;
3070 struct btrfs_key key;
3071 struct btrfs_key key_end;
3072 u64 increment = map->stripe_len;
3075 u64 extent_physical;
3079 struct btrfs_device *extent_dev;
3080 int extent_mirror_num;
3083 physical = map->stripes[num].physical;
3085 nstripes = div64_u64(length, map->stripe_len);
3086 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3087 offset = map->stripe_len * num;
3088 increment = map->stripe_len * map->num_stripes;
3090 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3091 int factor = map->num_stripes / map->sub_stripes;
3092 offset = map->stripe_len * (num / map->sub_stripes);
3093 increment = map->stripe_len * factor;
3094 mirror_num = num % map->sub_stripes + 1;
3095 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3096 increment = map->stripe_len;
3097 mirror_num = num % map->num_stripes + 1;
3098 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3099 increment = map->stripe_len;
3100 mirror_num = num % map->num_stripes + 1;
3101 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3102 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3103 increment = map->stripe_len * nr_data_stripes(map);
3106 increment = map->stripe_len;
3110 path = btrfs_alloc_path();
3114 ppath = btrfs_alloc_path();
3116 btrfs_free_path(path);
3121 * work on commit root. The related disk blocks are static as
3122 * long as COW is applied. This means, it is save to rewrite
3123 * them to repair disk errors without any race conditions
3125 path->search_commit_root = 1;
3126 path->skip_locking = 1;
3128 ppath->search_commit_root = 1;
3129 ppath->skip_locking = 1;
3131 * trigger the readahead for extent tree csum tree and wait for
3132 * completion. During readahead, the scrub is officially paused
3133 * to not hold off transaction commits
3135 logical = base + offset;
3136 physical_end = physical + nstripes * map->stripe_len;
3137 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3138 get_raid56_logic_offset(physical_end, num,
3139 map, &logic_end, NULL);
3142 logic_end = logical + increment * nstripes;
3144 wait_event(sctx->list_wait,
3145 atomic_read(&sctx->bios_in_flight) == 0);
3146 scrub_blocked_if_needed(fs_info);
3148 /* FIXME it might be better to start readahead at commit root */
3149 key.objectid = logical;
3150 key.type = BTRFS_EXTENT_ITEM_KEY;
3151 key.offset = (u64)0;
3152 key_end.objectid = logic_end;
3153 key_end.type = BTRFS_METADATA_ITEM_KEY;
3154 key_end.offset = (u64)-1;
3155 reada1 = btrfs_reada_add(root, &key, &key_end);
3157 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3158 key.type = BTRFS_EXTENT_CSUM_KEY;
3159 key.offset = logical;
3160 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3161 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3162 key_end.offset = logic_end;
3163 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3165 if (!IS_ERR(reada1))
3166 btrfs_reada_wait(reada1);
3167 if (!IS_ERR(reada2))
3168 btrfs_reada_wait(reada2);
3172 * collect all data csums for the stripe to avoid seeking during
3173 * the scrub. This might currently (crc32) end up to be about 1MB
3175 blk_start_plug(&plug);
3178 * now find all extents for each stripe and scrub them
3181 while (physical < physical_end) {
3185 if (atomic_read(&fs_info->scrub_cancel_req) ||
3186 atomic_read(&sctx->cancel_req)) {
3191 * check to see if we have to pause
3193 if (atomic_read(&fs_info->scrub_pause_req)) {
3194 /* push queued extents */
3195 sctx->flush_all_writes = true;
3197 mutex_lock(&sctx->wr_lock);
3198 scrub_wr_submit(sctx);
3199 mutex_unlock(&sctx->wr_lock);
3200 wait_event(sctx->list_wait,
3201 atomic_read(&sctx->bios_in_flight) == 0);
3202 sctx->flush_all_writes = false;
3203 scrub_blocked_if_needed(fs_info);
3206 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3207 ret = get_raid56_logic_offset(physical, num, map,
3212 /* it is parity strip */
3213 stripe_logical += base;
3214 stripe_end = stripe_logical + increment;
3215 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3216 ppath, stripe_logical,
3224 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3225 key.type = BTRFS_METADATA_ITEM_KEY;
3227 key.type = BTRFS_EXTENT_ITEM_KEY;
3228 key.objectid = logical;
3229 key.offset = (u64)-1;
3231 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3236 ret = btrfs_previous_extent_item(root, path, 0);
3240 /* there's no smaller item, so stick with the
3242 btrfs_release_path(path);
3243 ret = btrfs_search_slot(NULL, root, &key,
3255 slot = path->slots[0];
3256 if (slot >= btrfs_header_nritems(l)) {
3257 ret = btrfs_next_leaf(root, path);
3266 btrfs_item_key_to_cpu(l, &key, slot);
3268 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3269 key.type != BTRFS_METADATA_ITEM_KEY)
3272 if (key.type == BTRFS_METADATA_ITEM_KEY)
3273 bytes = fs_info->nodesize;
3277 if (key.objectid + bytes <= logical)
3280 if (key.objectid >= logical + map->stripe_len) {
3281 /* out of this device extent */
3282 if (key.objectid >= logic_end)
3287 extent = btrfs_item_ptr(l, slot,
3288 struct btrfs_extent_item);
3289 flags = btrfs_extent_flags(l, extent);
3290 generation = btrfs_extent_generation(l, extent);
3292 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3293 (key.objectid < logical ||
3294 key.objectid + bytes >
3295 logical + map->stripe_len)) {
3297 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3298 key.objectid, logical);
3299 spin_lock(&sctx->stat_lock);
3300 sctx->stat.uncorrectable_errors++;
3301 spin_unlock(&sctx->stat_lock);
3306 extent_logical = key.objectid;
3310 * trim extent to this stripe
3312 if (extent_logical < logical) {
3313 extent_len -= logical - extent_logical;
3314 extent_logical = logical;
3316 if (extent_logical + extent_len >
3317 logical + map->stripe_len) {
3318 extent_len = logical + map->stripe_len -
3322 extent_physical = extent_logical - logical + physical;
3323 extent_dev = scrub_dev;
3324 extent_mirror_num = mirror_num;
3325 if (sctx->is_dev_replace)
3326 scrub_remap_extent(fs_info, extent_logical,
3327 extent_len, &extent_physical,
3329 &extent_mirror_num);
3331 ret = btrfs_lookup_csums_range(csum_root,
3335 &sctx->csum_list, 1);
3339 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3340 extent_physical, extent_dev, flags,
3341 generation, extent_mirror_num,
3342 extent_logical - logical + physical);
3344 scrub_free_csums(sctx);
3349 if (extent_logical + extent_len <
3350 key.objectid + bytes) {
3351 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3353 * loop until we find next data stripe
3354 * or we have finished all stripes.
3357 physical += map->stripe_len;
3358 ret = get_raid56_logic_offset(physical,
3363 if (ret && physical < physical_end) {
3364 stripe_logical += base;
3365 stripe_end = stripe_logical +
3367 ret = scrub_raid56_parity(sctx,
3368 map, scrub_dev, ppath,
3376 physical += map->stripe_len;
3377 logical += increment;
3379 if (logical < key.objectid + bytes) {
3384 if (physical >= physical_end) {
3392 btrfs_release_path(path);
3394 logical += increment;
3395 physical += map->stripe_len;
3396 spin_lock(&sctx->stat_lock);
3398 sctx->stat.last_physical = map->stripes[num].physical +
3401 sctx->stat.last_physical = physical;
3402 spin_unlock(&sctx->stat_lock);
3407 /* push queued extents */
3409 mutex_lock(&sctx->wr_lock);
3410 scrub_wr_submit(sctx);
3411 mutex_unlock(&sctx->wr_lock);
3413 blk_finish_plug(&plug);
3414 btrfs_free_path(path);
3415 btrfs_free_path(ppath);
3416 return ret < 0 ? ret : 0;
3419 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3420 struct btrfs_device *scrub_dev,
3421 u64 chunk_offset, u64 length,
3423 struct btrfs_block_group *cache)
3425 struct btrfs_fs_info *fs_info = sctx->fs_info;
3426 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3427 struct map_lookup *map;
3428 struct extent_map *em;
3432 read_lock(&map_tree->lock);
3433 em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3434 read_unlock(&map_tree->lock);
3438 * Might have been an unused block group deleted by the cleaner
3439 * kthread or relocation.
3441 spin_lock(&cache->lock);
3442 if (!cache->removed)
3444 spin_unlock(&cache->lock);
3449 map = em->map_lookup;
3450 if (em->start != chunk_offset)
3453 if (em->len < length)
3456 for (i = 0; i < map->num_stripes; ++i) {
3457 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3458 map->stripes[i].physical == dev_offset) {
3459 ret = scrub_stripe(sctx, map, scrub_dev, i,
3460 chunk_offset, length);
3466 free_extent_map(em);
3471 static noinline_for_stack
3472 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3473 struct btrfs_device *scrub_dev, u64 start, u64 end)
3475 struct btrfs_dev_extent *dev_extent = NULL;
3476 struct btrfs_path *path;
3477 struct btrfs_fs_info *fs_info = sctx->fs_info;
3478 struct btrfs_root *root = fs_info->dev_root;
3484 struct extent_buffer *l;
3485 struct btrfs_key key;
3486 struct btrfs_key found_key;
3487 struct btrfs_block_group *cache;
3488 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3490 path = btrfs_alloc_path();
3494 path->reada = READA_FORWARD;
3495 path->search_commit_root = 1;
3496 path->skip_locking = 1;
3498 key.objectid = scrub_dev->devid;
3500 key.type = BTRFS_DEV_EXTENT_KEY;
3503 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3507 if (path->slots[0] >=
3508 btrfs_header_nritems(path->nodes[0])) {
3509 ret = btrfs_next_leaf(root, path);
3522 slot = path->slots[0];
3524 btrfs_item_key_to_cpu(l, &found_key, slot);
3526 if (found_key.objectid != scrub_dev->devid)
3529 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3532 if (found_key.offset >= end)
3535 if (found_key.offset < key.offset)
3538 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3539 length = btrfs_dev_extent_length(l, dev_extent);
3541 if (found_key.offset + length <= start)
3544 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3547 * get a reference on the corresponding block group to prevent
3548 * the chunk from going away while we scrub it
3550 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3552 /* some chunks are removed but not committed to disk yet,
3553 * continue scrubbing */
3558 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3559 * to avoid deadlock caused by:
3560 * btrfs_inc_block_group_ro()
3561 * -> btrfs_wait_for_commit()
3562 * -> btrfs_commit_transaction()
3563 * -> btrfs_scrub_pause()
3565 scrub_pause_on(fs_info);
3568 * Don't do chunk preallocation for scrub.
3570 * This is especially important for SYSTEM bgs, or we can hit
3571 * -EFBIG from btrfs_finish_chunk_alloc() like:
3572 * 1. The only SYSTEM bg is marked RO.
3573 * Since SYSTEM bg is small, that's pretty common.
3574 * 2. New SYSTEM bg will be allocated
3575 * Due to regular version will allocate new chunk.
3576 * 3. New SYSTEM bg is empty and will get cleaned up
3577 * Before cleanup really happens, it's marked RO again.
3578 * 4. Empty SYSTEM bg get scrubbed
3581 * This can easily boost the amount of SYSTEM chunks if cleaner
3582 * thread can't be triggered fast enough, and use up all space
3583 * of btrfs_super_block::sys_chunk_array
3585 * While for dev replace, we need to try our best to mark block
3586 * group RO, to prevent race between:
3587 * - Write duplication
3588 * Contains latest data
3590 * Contains data from commit tree
3592 * If target block group is not marked RO, nocow writes can
3593 * be overwritten by scrub copy, causing data corruption.
3594 * So for dev-replace, it's not allowed to continue if a block
3597 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3600 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3602 * btrfs_inc_block_group_ro return -ENOSPC when it
3603 * failed in creating new chunk for metadata.
3604 * It is not a problem for scrub, because
3605 * metadata are always cowed, and our scrub paused
3606 * commit_transactions.
3611 "failed setting block group ro: %d", ret);
3612 btrfs_put_block_group(cache);
3613 scrub_pause_off(fs_info);
3618 * Now the target block is marked RO, wait for nocow writes to
3619 * finish before dev-replace.
3620 * COW is fine, as COW never overwrites extents in commit tree.
3622 if (sctx->is_dev_replace) {
3623 btrfs_wait_nocow_writers(cache);
3624 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3628 scrub_pause_off(fs_info);
3629 down_write(&dev_replace->rwsem);
3630 dev_replace->cursor_right = found_key.offset + length;
3631 dev_replace->cursor_left = found_key.offset;
3632 dev_replace->item_needs_writeback = 1;
3633 up_write(&dev_replace->rwsem);
3635 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3636 found_key.offset, cache);
3639 * flush, submit all pending read and write bios, afterwards
3641 * Note that in the dev replace case, a read request causes
3642 * write requests that are submitted in the read completion
3643 * worker. Therefore in the current situation, it is required
3644 * that all write requests are flushed, so that all read and
3645 * write requests are really completed when bios_in_flight
3648 sctx->flush_all_writes = true;
3650 mutex_lock(&sctx->wr_lock);
3651 scrub_wr_submit(sctx);
3652 mutex_unlock(&sctx->wr_lock);
3654 wait_event(sctx->list_wait,
3655 atomic_read(&sctx->bios_in_flight) == 0);
3657 scrub_pause_on(fs_info);
3660 * must be called before we decrease @scrub_paused.
3661 * make sure we don't block transaction commit while
3662 * we are waiting pending workers finished.
3664 wait_event(sctx->list_wait,
3665 atomic_read(&sctx->workers_pending) == 0);
3666 sctx->flush_all_writes = false;
3668 scrub_pause_off(fs_info);
3670 down_write(&dev_replace->rwsem);
3671 dev_replace->cursor_left = dev_replace->cursor_right;
3672 dev_replace->item_needs_writeback = 1;
3673 up_write(&dev_replace->rwsem);
3676 btrfs_dec_block_group_ro(cache);
3679 * We might have prevented the cleaner kthread from deleting
3680 * this block group if it was already unused because we raced
3681 * and set it to RO mode first. So add it back to the unused
3682 * list, otherwise it might not ever be deleted unless a manual
3683 * balance is triggered or it becomes used and unused again.
3685 spin_lock(&cache->lock);
3686 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3688 spin_unlock(&cache->lock);
3689 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3690 btrfs_discard_queue_work(&fs_info->discard_ctl,
3693 btrfs_mark_bg_unused(cache);
3695 spin_unlock(&cache->lock);
3698 btrfs_put_block_group(cache);
3701 if (sctx->is_dev_replace &&
3702 atomic64_read(&dev_replace->num_write_errors) > 0) {
3706 if (sctx->stat.malloc_errors > 0) {
3711 key.offset = found_key.offset + length;
3712 btrfs_release_path(path);
3715 btrfs_free_path(path);
3720 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3721 struct btrfs_device *scrub_dev)
3727 struct btrfs_fs_info *fs_info = sctx->fs_info;
3729 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3732 /* Seed devices of a new filesystem has their own generation. */
3733 if (scrub_dev->fs_devices != fs_info->fs_devices)
3734 gen = scrub_dev->generation;
3736 gen = fs_info->last_trans_committed;
3738 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3739 bytenr = btrfs_sb_offset(i);
3740 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3741 scrub_dev->commit_total_bytes)
3744 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3745 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3750 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3756 * get a reference count on fs_info->scrub_workers. start worker if necessary
3758 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3761 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3762 int max_active = fs_info->thread_pool_size;
3764 lockdep_assert_held(&fs_info->scrub_lock);
3766 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3767 ASSERT(fs_info->scrub_workers == NULL);
3768 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3769 flags, is_dev_replace ? 1 : max_active, 4);
3770 if (!fs_info->scrub_workers)
3771 goto fail_scrub_workers;
3773 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3774 fs_info->scrub_wr_completion_workers =
3775 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3777 if (!fs_info->scrub_wr_completion_workers)
3778 goto fail_scrub_wr_completion_workers;
3780 ASSERT(fs_info->scrub_parity_workers == NULL);
3781 fs_info->scrub_parity_workers =
3782 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3784 if (!fs_info->scrub_parity_workers)
3785 goto fail_scrub_parity_workers;
3787 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3789 refcount_inc(&fs_info->scrub_workers_refcnt);
3793 fail_scrub_parity_workers:
3794 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3795 fail_scrub_wr_completion_workers:
3796 btrfs_destroy_workqueue(fs_info->scrub_workers);
3801 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3802 u64 end, struct btrfs_scrub_progress *progress,
3803 int readonly, int is_dev_replace)
3805 struct scrub_ctx *sctx;
3807 struct btrfs_device *dev;
3808 unsigned int nofs_flag;
3809 struct btrfs_workqueue *scrub_workers = NULL;
3810 struct btrfs_workqueue *scrub_wr_comp = NULL;
3811 struct btrfs_workqueue *scrub_parity = NULL;
3813 if (btrfs_fs_closing(fs_info))
3816 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3818 * in this case scrub is unable to calculate the checksum
3819 * the way scrub is implemented. Do not handle this
3820 * situation at all because it won't ever happen.
3823 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3829 if (fs_info->sectorsize != PAGE_SIZE) {
3830 /* not supported for data w/o checksums */
3831 btrfs_err_rl(fs_info,
3832 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3833 fs_info->sectorsize, PAGE_SIZE);
3837 if (fs_info->nodesize >
3838 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3839 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3841 * would exhaust the array bounds of pagev member in
3842 * struct scrub_block
3845 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3847 SCRUB_MAX_PAGES_PER_BLOCK,
3848 fs_info->sectorsize,
3849 SCRUB_MAX_PAGES_PER_BLOCK);
3853 /* Allocate outside of device_list_mutex */
3854 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3856 return PTR_ERR(sctx);
3858 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3859 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3860 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3862 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3867 if (!is_dev_replace && !readonly &&
3868 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3869 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3870 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3871 rcu_str_deref(dev->name));
3876 mutex_lock(&fs_info->scrub_lock);
3877 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3878 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3879 mutex_unlock(&fs_info->scrub_lock);
3880 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3885 down_read(&fs_info->dev_replace.rwsem);
3886 if (dev->scrub_ctx ||
3888 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3889 up_read(&fs_info->dev_replace.rwsem);
3890 mutex_unlock(&fs_info->scrub_lock);
3891 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3895 up_read(&fs_info->dev_replace.rwsem);
3897 ret = scrub_workers_get(fs_info, is_dev_replace);
3899 mutex_unlock(&fs_info->scrub_lock);
3900 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3904 sctx->readonly = readonly;
3905 dev->scrub_ctx = sctx;
3906 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3909 * checking @scrub_pause_req here, we can avoid
3910 * race between committing transaction and scrubbing.
3912 __scrub_blocked_if_needed(fs_info);
3913 atomic_inc(&fs_info->scrubs_running);
3914 mutex_unlock(&fs_info->scrub_lock);
3917 * In order to avoid deadlock with reclaim when there is a transaction
3918 * trying to pause scrub, make sure we use GFP_NOFS for all the
3919 * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3920 * invoked by our callees. The pausing request is done when the
3921 * transaction commit starts, and it blocks the transaction until scrub
3922 * is paused (done at specific points at scrub_stripe() or right above
3923 * before incrementing fs_info->scrubs_running).
3925 nofs_flag = memalloc_nofs_save();
3926 if (!is_dev_replace) {
3927 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3929 * by holding device list mutex, we can
3930 * kick off writing super in log tree sync.
3932 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3933 ret = scrub_supers(sctx, dev);
3934 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3938 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3939 memalloc_nofs_restore(nofs_flag);
3941 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3942 atomic_dec(&fs_info->scrubs_running);
3943 wake_up(&fs_info->scrub_pause_wait);
3945 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3948 memcpy(progress, &sctx->stat, sizeof(*progress));
3950 if (!is_dev_replace)
3951 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3952 ret ? "not finished" : "finished", devid, ret);
3954 mutex_lock(&fs_info->scrub_lock);
3955 dev->scrub_ctx = NULL;
3956 if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3957 scrub_workers = fs_info->scrub_workers;
3958 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3959 scrub_parity = fs_info->scrub_parity_workers;
3961 fs_info->scrub_workers = NULL;
3962 fs_info->scrub_wr_completion_workers = NULL;
3963 fs_info->scrub_parity_workers = NULL;
3965 mutex_unlock(&fs_info->scrub_lock);
3967 btrfs_destroy_workqueue(scrub_workers);
3968 btrfs_destroy_workqueue(scrub_wr_comp);
3969 btrfs_destroy_workqueue(scrub_parity);
3970 scrub_put_ctx(sctx);
3975 scrub_free_ctx(sctx);
3980 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3982 mutex_lock(&fs_info->scrub_lock);
3983 atomic_inc(&fs_info->scrub_pause_req);
3984 while (atomic_read(&fs_info->scrubs_paused) !=
3985 atomic_read(&fs_info->scrubs_running)) {
3986 mutex_unlock(&fs_info->scrub_lock);
3987 wait_event(fs_info->scrub_pause_wait,
3988 atomic_read(&fs_info->scrubs_paused) ==
3989 atomic_read(&fs_info->scrubs_running));
3990 mutex_lock(&fs_info->scrub_lock);
3992 mutex_unlock(&fs_info->scrub_lock);
3995 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3997 atomic_dec(&fs_info->scrub_pause_req);
3998 wake_up(&fs_info->scrub_pause_wait);
4001 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4003 mutex_lock(&fs_info->scrub_lock);
4004 if (!atomic_read(&fs_info->scrubs_running)) {
4005 mutex_unlock(&fs_info->scrub_lock);
4009 atomic_inc(&fs_info->scrub_cancel_req);
4010 while (atomic_read(&fs_info->scrubs_running)) {
4011 mutex_unlock(&fs_info->scrub_lock);
4012 wait_event(fs_info->scrub_pause_wait,
4013 atomic_read(&fs_info->scrubs_running) == 0);
4014 mutex_lock(&fs_info->scrub_lock);
4016 atomic_dec(&fs_info->scrub_cancel_req);
4017 mutex_unlock(&fs_info->scrub_lock);
4022 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4024 struct btrfs_fs_info *fs_info = dev->fs_info;
4025 struct scrub_ctx *sctx;
4027 mutex_lock(&fs_info->scrub_lock);
4028 sctx = dev->scrub_ctx;
4030 mutex_unlock(&fs_info->scrub_lock);
4033 atomic_inc(&sctx->cancel_req);
4034 while (dev->scrub_ctx) {
4035 mutex_unlock(&fs_info->scrub_lock);
4036 wait_event(fs_info->scrub_pause_wait,
4037 dev->scrub_ctx == NULL);
4038 mutex_lock(&fs_info->scrub_lock);
4040 mutex_unlock(&fs_info->scrub_lock);
4045 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4046 struct btrfs_scrub_progress *progress)
4048 struct btrfs_device *dev;
4049 struct scrub_ctx *sctx = NULL;
4051 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4052 dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4054 sctx = dev->scrub_ctx;
4056 memcpy(progress, &sctx->stat, sizeof(*progress));
4057 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4059 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4062 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4063 u64 extent_logical, u64 extent_len,
4064 u64 *extent_physical,
4065 struct btrfs_device **extent_dev,
4066 int *extent_mirror_num)
4069 struct btrfs_bio *bbio = NULL;
4072 mapped_length = extent_len;
4073 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4074 &mapped_length, &bbio, 0);
4075 if (ret || !bbio || mapped_length < extent_len ||
4076 !bbio->stripes[0].dev->bdev) {
4077 btrfs_put_bbio(bbio);
4081 *extent_physical = bbio->stripes[0].physical;
4082 *extent_mirror_num = bbio->mirror_num;
4083 *extent_dev = bbio->stripes[0].dev;
4084 btrfs_put_bbio(bbio);