1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
12 #include "ordered-data.h"
13 #include "transaction.h"
15 #include "extent_io.h"
16 #include "dev-replace.h"
17 #include "check-integrity.h"
18 #include "rcu-string.h"
22 * This is only the first step towards a full-features scrub. It reads all
23 * extent and super block and verifies the checksums. In case a bad checksum
24 * is found or the extent cannot be read, good data will be written back if
27 * Future enhancements:
28 * - In case an unrepairable extent is encountered, track which files are
29 * affected and report them
30 * - track and record media errors, throw out bad devices
31 * - add a mode to also read unallocated space
38 * the following three values only influence the performance.
39 * The last one configures the number of parallel and outstanding I/O
40 * operations. The first two values configure an upper limit for the number
41 * of (dynamically allocated) pages that are added to a bio.
43 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
44 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
45 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
48 * the following value times PAGE_SIZE needs to be large enough to match the
49 * largest node/leaf/sector size that shall be supported.
50 * Values larger than BTRFS_STRIPE_LEN are not supported.
52 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
54 struct scrub_recover {
56 struct btrfs_bio *bbio;
61 struct scrub_block *sblock;
63 struct btrfs_device *dev;
64 struct list_head list;
65 u64 flags; /* extent flags */
69 u64 physical_for_dev_replace;
72 unsigned int mirror_num:8;
73 unsigned int have_csum:1;
74 unsigned int io_error:1;
76 u8 csum[BTRFS_CSUM_SIZE];
78 struct scrub_recover *recover;
83 struct scrub_ctx *sctx;
84 struct btrfs_device *dev;
89 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
90 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
92 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
96 struct btrfs_work work;
100 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
102 atomic_t outstanding_pages;
103 refcount_t refs; /* free mem on transition to zero */
104 struct scrub_ctx *sctx;
105 struct scrub_parity *sparity;
107 unsigned int header_error:1;
108 unsigned int checksum_error:1;
109 unsigned int no_io_error_seen:1;
110 unsigned int generation_error:1; /* also sets header_error */
112 /* The following is for the data used to check parity */
113 /* It is for the data with checksum */
114 unsigned int data_corrected:1;
116 struct btrfs_work work;
119 /* Used for the chunks with parity stripe such RAID5/6 */
120 struct scrub_parity {
121 struct scrub_ctx *sctx;
123 struct btrfs_device *scrub_dev;
135 struct list_head spages;
137 /* Work of parity check and repair */
138 struct btrfs_work work;
140 /* Mark the parity blocks which have data */
141 unsigned long *dbitmap;
144 * Mark the parity blocks which have data, but errors happen when
145 * read data or check data
147 unsigned long *ebitmap;
149 unsigned long bitmap[0];
153 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
154 struct btrfs_fs_info *fs_info;
157 atomic_t bios_in_flight;
158 atomic_t workers_pending;
159 spinlock_t list_lock;
160 wait_queue_head_t list_wait;
162 struct list_head csum_list;
165 int pages_per_rd_bio;
169 struct scrub_bio *wr_curr_bio;
170 struct mutex wr_lock;
171 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
172 struct btrfs_device *wr_tgtdev;
173 bool flush_all_writes;
178 struct btrfs_scrub_progress stat;
179 spinlock_t stat_lock;
182 * Use a ref counter to avoid use-after-free issues. Scrub workers
183 * decrement bios_in_flight and workers_pending and then do a wakeup
184 * on the list_wait wait queue. We must ensure the main scrub task
185 * doesn't free the scrub context before or while the workers are
186 * doing the wakeup() call.
191 struct scrub_fixup_nodatasum {
192 struct scrub_ctx *sctx;
193 struct btrfs_device *dev;
195 struct btrfs_root *root;
196 struct btrfs_work work;
200 struct scrub_nocow_inode {
204 struct list_head list;
207 struct scrub_copy_nocow_ctx {
208 struct scrub_ctx *sctx;
212 u64 physical_for_dev_replace;
213 struct list_head inodes;
214 struct btrfs_work work;
217 struct scrub_warning {
218 struct btrfs_path *path;
219 u64 extent_item_size;
223 struct btrfs_device *dev;
226 struct full_stripe_lock {
233 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
234 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
235 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
236 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
237 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
238 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
239 struct scrub_block *sblocks_for_recheck);
240 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
241 struct scrub_block *sblock,
242 int retry_failed_mirror);
243 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
244 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
245 struct scrub_block *sblock_good);
246 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
247 struct scrub_block *sblock_good,
248 int page_num, int force_write);
249 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
250 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
252 static int scrub_checksum_data(struct scrub_block *sblock);
253 static int scrub_checksum_tree_block(struct scrub_block *sblock);
254 static int scrub_checksum_super(struct scrub_block *sblock);
255 static void scrub_block_get(struct scrub_block *sblock);
256 static void scrub_block_put(struct scrub_block *sblock);
257 static void scrub_page_get(struct scrub_page *spage);
258 static void scrub_page_put(struct scrub_page *spage);
259 static void scrub_parity_get(struct scrub_parity *sparity);
260 static void scrub_parity_put(struct scrub_parity *sparity);
261 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
262 struct scrub_page *spage);
263 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
264 u64 physical, struct btrfs_device *dev, u64 flags,
265 u64 gen, int mirror_num, u8 *csum, int force,
266 u64 physical_for_dev_replace);
267 static void scrub_bio_end_io(struct bio *bio);
268 static void scrub_bio_end_io_worker(struct btrfs_work *work);
269 static void scrub_block_complete(struct scrub_block *sblock);
270 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
271 u64 extent_logical, u64 extent_len,
272 u64 *extent_physical,
273 struct btrfs_device **extent_dev,
274 int *extent_mirror_num);
275 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
276 struct scrub_page *spage);
277 static void scrub_wr_submit(struct scrub_ctx *sctx);
278 static void scrub_wr_bio_end_io(struct bio *bio);
279 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
280 static int write_page_nocow(struct scrub_ctx *sctx,
281 u64 physical_for_dev_replace, struct page *page);
282 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
283 struct scrub_copy_nocow_ctx *ctx);
284 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
285 int mirror_num, u64 physical_for_dev_replace);
286 static void copy_nocow_pages_worker(struct btrfs_work *work);
287 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
288 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
289 static void scrub_put_ctx(struct scrub_ctx *sctx);
291 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
293 return page->recover &&
294 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
297 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
299 refcount_inc(&sctx->refs);
300 atomic_inc(&sctx->bios_in_flight);
303 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
305 atomic_dec(&sctx->bios_in_flight);
306 wake_up(&sctx->list_wait);
310 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
312 while (atomic_read(&fs_info->scrub_pause_req)) {
313 mutex_unlock(&fs_info->scrub_lock);
314 wait_event(fs_info->scrub_pause_wait,
315 atomic_read(&fs_info->scrub_pause_req) == 0);
316 mutex_lock(&fs_info->scrub_lock);
320 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
322 atomic_inc(&fs_info->scrubs_paused);
323 wake_up(&fs_info->scrub_pause_wait);
326 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
328 mutex_lock(&fs_info->scrub_lock);
329 __scrub_blocked_if_needed(fs_info);
330 atomic_dec(&fs_info->scrubs_paused);
331 mutex_unlock(&fs_info->scrub_lock);
333 wake_up(&fs_info->scrub_pause_wait);
336 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
338 scrub_pause_on(fs_info);
339 scrub_pause_off(fs_info);
343 * Insert new full stripe lock into full stripe locks tree
345 * Return pointer to existing or newly inserted full_stripe_lock structure if
346 * everything works well.
347 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
349 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
352 static struct full_stripe_lock *insert_full_stripe_lock(
353 struct btrfs_full_stripe_locks_tree *locks_root,
357 struct rb_node *parent = NULL;
358 struct full_stripe_lock *entry;
359 struct full_stripe_lock *ret;
361 lockdep_assert_held(&locks_root->lock);
363 p = &locks_root->root.rb_node;
366 entry = rb_entry(parent, struct full_stripe_lock, node);
367 if (fstripe_logical < entry->logical) {
369 } else if (fstripe_logical > entry->logical) {
377 /* Insert new lock */
378 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
380 return ERR_PTR(-ENOMEM);
381 ret->logical = fstripe_logical;
383 mutex_init(&ret->mutex);
385 rb_link_node(&ret->node, parent, p);
386 rb_insert_color(&ret->node, &locks_root->root);
391 * Search for a full stripe lock of a block group
393 * Return pointer to existing full stripe lock if found
394 * Return NULL if not found
396 static struct full_stripe_lock *search_full_stripe_lock(
397 struct btrfs_full_stripe_locks_tree *locks_root,
400 struct rb_node *node;
401 struct full_stripe_lock *entry;
403 lockdep_assert_held(&locks_root->lock);
405 node = locks_root->root.rb_node;
407 entry = rb_entry(node, struct full_stripe_lock, node);
408 if (fstripe_logical < entry->logical)
409 node = node->rb_left;
410 else if (fstripe_logical > entry->logical)
411 node = node->rb_right;
419 * Helper to get full stripe logical from a normal bytenr.
421 * Caller must ensure @cache is a RAID56 block group.
423 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
429 * Due to chunk item size limit, full stripe length should not be
430 * larger than U32_MAX. Just a sanity check here.
432 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
435 * round_down() can only handle power of 2, while RAID56 full
436 * stripe length can be 64KiB * n, so we need to manually round down.
438 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
439 cache->full_stripe_len + cache->key.objectid;
444 * Lock a full stripe to avoid concurrency of recovery and read
446 * It's only used for profiles with parities (RAID5/6), for other profiles it
449 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
450 * So caller must call unlock_full_stripe() at the same context.
452 * Return <0 if encounters error.
454 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
457 struct btrfs_block_group_cache *bg_cache;
458 struct btrfs_full_stripe_locks_tree *locks_root;
459 struct full_stripe_lock *existing;
464 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
470 /* Profiles not based on parity don't need full stripe lock */
471 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
473 locks_root = &bg_cache->full_stripe_locks_root;
475 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
477 /* Now insert the full stripe lock */
478 mutex_lock(&locks_root->lock);
479 existing = insert_full_stripe_lock(locks_root, fstripe_start);
480 mutex_unlock(&locks_root->lock);
481 if (IS_ERR(existing)) {
482 ret = PTR_ERR(existing);
485 mutex_lock(&existing->mutex);
488 btrfs_put_block_group(bg_cache);
493 * Unlock a full stripe.
495 * NOTE: Caller must ensure it's the same context calling corresponding
496 * lock_full_stripe().
498 * Return 0 if we unlock full stripe without problem.
499 * Return <0 for error
501 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
504 struct btrfs_block_group_cache *bg_cache;
505 struct btrfs_full_stripe_locks_tree *locks_root;
506 struct full_stripe_lock *fstripe_lock;
511 /* If we didn't acquire full stripe lock, no need to continue */
515 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
520 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
523 locks_root = &bg_cache->full_stripe_locks_root;
524 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
526 mutex_lock(&locks_root->lock);
527 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
528 /* Unpaired unlock_full_stripe() detected */
532 mutex_unlock(&locks_root->lock);
536 if (fstripe_lock->refs == 0) {
538 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
539 fstripe_lock->logical);
541 fstripe_lock->refs--;
544 if (fstripe_lock->refs == 0) {
545 rb_erase(&fstripe_lock->node, &locks_root->root);
548 mutex_unlock(&locks_root->lock);
550 mutex_unlock(&fstripe_lock->mutex);
554 btrfs_put_block_group(bg_cache);
559 * used for workers that require transaction commits (i.e., for the
562 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
564 struct btrfs_fs_info *fs_info = sctx->fs_info;
566 refcount_inc(&sctx->refs);
568 * increment scrubs_running to prevent cancel requests from
569 * completing as long as a worker is running. we must also
570 * increment scrubs_paused to prevent deadlocking on pause
571 * requests used for transactions commits (as the worker uses a
572 * transaction context). it is safe to regard the worker
573 * as paused for all matters practical. effectively, we only
574 * avoid cancellation requests from completing.
576 mutex_lock(&fs_info->scrub_lock);
577 atomic_inc(&fs_info->scrubs_running);
578 atomic_inc(&fs_info->scrubs_paused);
579 mutex_unlock(&fs_info->scrub_lock);
582 * check if @scrubs_running=@scrubs_paused condition
583 * inside wait_event() is not an atomic operation.
584 * which means we may inc/dec @scrub_running/paused
585 * at any time. Let's wake up @scrub_pause_wait as
586 * much as we can to let commit transaction blocked less.
588 wake_up(&fs_info->scrub_pause_wait);
590 atomic_inc(&sctx->workers_pending);
593 /* used for workers that require transaction commits */
594 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
596 struct btrfs_fs_info *fs_info = sctx->fs_info;
599 * see scrub_pending_trans_workers_inc() why we're pretending
600 * to be paused in the scrub counters
602 mutex_lock(&fs_info->scrub_lock);
603 atomic_dec(&fs_info->scrubs_running);
604 atomic_dec(&fs_info->scrubs_paused);
605 mutex_unlock(&fs_info->scrub_lock);
606 atomic_dec(&sctx->workers_pending);
607 wake_up(&fs_info->scrub_pause_wait);
608 wake_up(&sctx->list_wait);
612 static void scrub_free_csums(struct scrub_ctx *sctx)
614 while (!list_empty(&sctx->csum_list)) {
615 struct btrfs_ordered_sum *sum;
616 sum = list_first_entry(&sctx->csum_list,
617 struct btrfs_ordered_sum, list);
618 list_del(&sum->list);
623 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
630 /* this can happen when scrub is cancelled */
631 if (sctx->curr != -1) {
632 struct scrub_bio *sbio = sctx->bios[sctx->curr];
634 for (i = 0; i < sbio->page_count; i++) {
635 WARN_ON(!sbio->pagev[i]->page);
636 scrub_block_put(sbio->pagev[i]->sblock);
641 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
642 struct scrub_bio *sbio = sctx->bios[i];
649 kfree(sctx->wr_curr_bio);
650 scrub_free_csums(sctx);
654 static void scrub_put_ctx(struct scrub_ctx *sctx)
656 if (refcount_dec_and_test(&sctx->refs))
657 scrub_free_ctx(sctx);
660 static noinline_for_stack
661 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
663 struct scrub_ctx *sctx;
665 struct btrfs_fs_info *fs_info = dev->fs_info;
667 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
670 refcount_set(&sctx->refs, 1);
671 sctx->is_dev_replace = is_dev_replace;
672 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
674 sctx->fs_info = dev->fs_info;
675 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
676 struct scrub_bio *sbio;
678 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
681 sctx->bios[i] = sbio;
685 sbio->page_count = 0;
686 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
687 scrub_bio_end_io_worker, NULL, NULL);
689 if (i != SCRUB_BIOS_PER_SCTX - 1)
690 sctx->bios[i]->next_free = i + 1;
692 sctx->bios[i]->next_free = -1;
694 sctx->first_free = 0;
695 atomic_set(&sctx->bios_in_flight, 0);
696 atomic_set(&sctx->workers_pending, 0);
697 atomic_set(&sctx->cancel_req, 0);
698 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
699 INIT_LIST_HEAD(&sctx->csum_list);
701 spin_lock_init(&sctx->list_lock);
702 spin_lock_init(&sctx->stat_lock);
703 init_waitqueue_head(&sctx->list_wait);
705 WARN_ON(sctx->wr_curr_bio != NULL);
706 mutex_init(&sctx->wr_lock);
707 sctx->wr_curr_bio = NULL;
708 if (is_dev_replace) {
709 WARN_ON(!fs_info->dev_replace.tgtdev);
710 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
711 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
712 sctx->flush_all_writes = false;
718 scrub_free_ctx(sctx);
719 return ERR_PTR(-ENOMEM);
722 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
730 struct extent_buffer *eb;
731 struct btrfs_inode_item *inode_item;
732 struct scrub_warning *swarn = warn_ctx;
733 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
734 struct inode_fs_paths *ipath = NULL;
735 struct btrfs_root *local_root;
736 struct btrfs_key root_key;
737 struct btrfs_key key;
739 root_key.objectid = root;
740 root_key.type = BTRFS_ROOT_ITEM_KEY;
741 root_key.offset = (u64)-1;
742 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
743 if (IS_ERR(local_root)) {
744 ret = PTR_ERR(local_root);
749 * this makes the path point to (inum INODE_ITEM ioff)
752 key.type = BTRFS_INODE_ITEM_KEY;
755 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
757 btrfs_release_path(swarn->path);
761 eb = swarn->path->nodes[0];
762 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
763 struct btrfs_inode_item);
764 isize = btrfs_inode_size(eb, inode_item);
765 nlink = btrfs_inode_nlink(eb, inode_item);
766 btrfs_release_path(swarn->path);
769 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
770 * uses GFP_NOFS in this context, so we keep it consistent but it does
771 * not seem to be strictly necessary.
773 nofs_flag = memalloc_nofs_save();
774 ipath = init_ipath(4096, local_root, swarn->path);
775 memalloc_nofs_restore(nofs_flag);
777 ret = PTR_ERR(ipath);
781 ret = paths_from_inode(inum, ipath);
787 * we deliberately ignore the bit ipath might have been too small to
788 * hold all of the paths here
790 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
791 btrfs_warn_in_rcu(fs_info,
792 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
793 swarn->errstr, swarn->logical,
794 rcu_str_deref(swarn->dev->name),
797 min(isize - offset, (u64)PAGE_SIZE), nlink,
798 (char *)(unsigned long)ipath->fspath->val[i]);
804 btrfs_warn_in_rcu(fs_info,
805 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
806 swarn->errstr, swarn->logical,
807 rcu_str_deref(swarn->dev->name),
809 root, inum, offset, ret);
815 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
817 struct btrfs_device *dev;
818 struct btrfs_fs_info *fs_info;
819 struct btrfs_path *path;
820 struct btrfs_key found_key;
821 struct extent_buffer *eb;
822 struct btrfs_extent_item *ei;
823 struct scrub_warning swarn;
824 unsigned long ptr = 0;
832 WARN_ON(sblock->page_count < 1);
833 dev = sblock->pagev[0]->dev;
834 fs_info = sblock->sctx->fs_info;
836 path = btrfs_alloc_path();
840 swarn.physical = sblock->pagev[0]->physical;
841 swarn.logical = sblock->pagev[0]->logical;
842 swarn.errstr = errstr;
845 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
850 extent_item_pos = swarn.logical - found_key.objectid;
851 swarn.extent_item_size = found_key.offset;
854 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
855 item_size = btrfs_item_size_nr(eb, path->slots[0]);
857 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
859 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
860 item_size, &ref_root,
862 btrfs_warn_in_rcu(fs_info,
863 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
864 errstr, swarn.logical,
865 rcu_str_deref(dev->name),
867 ref_level ? "node" : "leaf",
868 ret < 0 ? -1 : ref_level,
869 ret < 0 ? -1 : ref_root);
871 btrfs_release_path(path);
873 btrfs_release_path(path);
876 iterate_extent_inodes(fs_info, found_key.objectid,
878 scrub_print_warning_inode, &swarn, false);
882 btrfs_free_path(path);
885 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
887 struct page *page = NULL;
889 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
892 struct btrfs_key key;
893 struct inode *inode = NULL;
894 struct btrfs_fs_info *fs_info;
895 u64 end = offset + PAGE_SIZE - 1;
896 struct btrfs_root *local_root;
900 key.type = BTRFS_ROOT_ITEM_KEY;
901 key.offset = (u64)-1;
903 fs_info = fixup->root->fs_info;
904 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
906 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
907 if (IS_ERR(local_root)) {
908 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
909 return PTR_ERR(local_root);
912 key.type = BTRFS_INODE_ITEM_KEY;
915 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
916 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
918 return PTR_ERR(inode);
920 index = offset >> PAGE_SHIFT;
922 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
928 if (PageUptodate(page)) {
929 if (PageDirty(page)) {
931 * we need to write the data to the defect sector. the
932 * data that was in that sector is not in memory,
933 * because the page was modified. we must not write the
934 * modified page to that sector.
936 * TODO: what could be done here: wait for the delalloc
937 * runner to write out that page (might involve
938 * COW) and see whether the sector is still
939 * referenced afterwards.
941 * For the meantime, we'll treat this error
942 * incorrectable, although there is a chance that a
943 * later scrub will find the bad sector again and that
944 * there's no dirty page in memory, then.
949 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
950 fixup->logical, page,
951 offset - page_offset(page),
957 * we need to get good data first. the general readpage path
958 * will call repair_io_failure for us, we just have to make
959 * sure we read the bad mirror.
961 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
964 /* set_extent_bits should give proper error */
971 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
974 wait_on_page_locked(page);
976 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
977 end, EXTENT_DAMAGED, 0, NULL);
979 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
992 if (ret == 0 && corrected) {
994 * we only need to call readpage for one of the inodes belonging
995 * to this extent. so make iterate_extent_inodes stop
1003 static void scrub_fixup_nodatasum(struct btrfs_work *work)
1005 struct btrfs_fs_info *fs_info;
1007 struct scrub_fixup_nodatasum *fixup;
1008 struct scrub_ctx *sctx;
1009 struct btrfs_trans_handle *trans = NULL;
1010 struct btrfs_path *path;
1011 int uncorrectable = 0;
1013 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1015 fs_info = fixup->root->fs_info;
1017 path = btrfs_alloc_path();
1019 spin_lock(&sctx->stat_lock);
1020 ++sctx->stat.malloc_errors;
1021 spin_unlock(&sctx->stat_lock);
1026 trans = btrfs_join_transaction(fixup->root);
1027 if (IS_ERR(trans)) {
1033 * the idea is to trigger a regular read through the standard path. we
1034 * read a page from the (failed) logical address by specifying the
1035 * corresponding copynum of the failed sector. thus, that readpage is
1037 * that is the point where on-the-fly error correction will kick in
1038 * (once it's finished) and rewrite the failed sector if a good copy
1041 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1042 scrub_fixup_readpage, fixup, false);
1049 spin_lock(&sctx->stat_lock);
1050 ++sctx->stat.corrected_errors;
1051 spin_unlock(&sctx->stat_lock);
1054 if (trans && !IS_ERR(trans))
1055 btrfs_end_transaction(trans);
1056 if (uncorrectable) {
1057 spin_lock(&sctx->stat_lock);
1058 ++sctx->stat.uncorrectable_errors;
1059 spin_unlock(&sctx->stat_lock);
1060 btrfs_dev_replace_stats_inc(
1061 &fs_info->dev_replace.num_uncorrectable_read_errors);
1062 btrfs_err_rl_in_rcu(fs_info,
1063 "unable to fixup (nodatasum) error at logical %llu on dev %s",
1064 fixup->logical, rcu_str_deref(fixup->dev->name));
1067 btrfs_free_path(path);
1070 scrub_pending_trans_workers_dec(sctx);
1073 static inline void scrub_get_recover(struct scrub_recover *recover)
1075 refcount_inc(&recover->refs);
1078 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1079 struct scrub_recover *recover)
1081 if (refcount_dec_and_test(&recover->refs)) {
1082 btrfs_bio_counter_dec(fs_info);
1083 btrfs_put_bbio(recover->bbio);
1089 * scrub_handle_errored_block gets called when either verification of the
1090 * pages failed or the bio failed to read, e.g. with EIO. In the latter
1091 * case, this function handles all pages in the bio, even though only one
1093 * The goal of this function is to repair the errored block by using the
1094 * contents of one of the mirrors.
1096 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1098 struct scrub_ctx *sctx = sblock_to_check->sctx;
1099 struct btrfs_device *dev;
1100 struct btrfs_fs_info *fs_info;
1102 unsigned int failed_mirror_index;
1103 unsigned int is_metadata;
1104 unsigned int have_csum;
1105 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1106 struct scrub_block *sblock_bad;
1111 bool full_stripe_locked;
1112 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1113 DEFAULT_RATELIMIT_BURST);
1115 BUG_ON(sblock_to_check->page_count < 1);
1116 fs_info = sctx->fs_info;
1117 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1119 * if we find an error in a super block, we just report it.
1120 * They will get written with the next transaction commit
1123 spin_lock(&sctx->stat_lock);
1124 ++sctx->stat.super_errors;
1125 spin_unlock(&sctx->stat_lock);
1128 logical = sblock_to_check->pagev[0]->logical;
1129 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1130 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1131 is_metadata = !(sblock_to_check->pagev[0]->flags &
1132 BTRFS_EXTENT_FLAG_DATA);
1133 have_csum = sblock_to_check->pagev[0]->have_csum;
1134 dev = sblock_to_check->pagev[0]->dev;
1137 * For RAID5/6, race can happen for a different device scrub thread.
1138 * For data corruption, Parity and Data threads will both try
1139 * to recovery the data.
1140 * Race can lead to doubly added csum error, or even unrecoverable
1143 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1145 spin_lock(&sctx->stat_lock);
1147 sctx->stat.malloc_errors++;
1148 sctx->stat.read_errors++;
1149 sctx->stat.uncorrectable_errors++;
1150 spin_unlock(&sctx->stat_lock);
1155 * read all mirrors one after the other. This includes to
1156 * re-read the extent or metadata block that failed (that was
1157 * the cause that this fixup code is called) another time,
1158 * page by page this time in order to know which pages
1159 * caused I/O errors and which ones are good (for all mirrors).
1160 * It is the goal to handle the situation when more than one
1161 * mirror contains I/O errors, but the errors do not
1162 * overlap, i.e. the data can be repaired by selecting the
1163 * pages from those mirrors without I/O error on the
1164 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1165 * would be that mirror #1 has an I/O error on the first page,
1166 * the second page is good, and mirror #2 has an I/O error on
1167 * the second page, but the first page is good.
1168 * Then the first page of the first mirror can be repaired by
1169 * taking the first page of the second mirror, and the
1170 * second page of the second mirror can be repaired by
1171 * copying the contents of the 2nd page of the 1st mirror.
1172 * One more note: if the pages of one mirror contain I/O
1173 * errors, the checksum cannot be verified. In order to get
1174 * the best data for repairing, the first attempt is to find
1175 * a mirror without I/O errors and with a validated checksum.
1176 * Only if this is not possible, the pages are picked from
1177 * mirrors with I/O errors without considering the checksum.
1178 * If the latter is the case, at the end, the checksum of the
1179 * repaired area is verified in order to correctly maintain
1183 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1184 sizeof(*sblocks_for_recheck), GFP_NOFS);
1185 if (!sblocks_for_recheck) {
1186 spin_lock(&sctx->stat_lock);
1187 sctx->stat.malloc_errors++;
1188 sctx->stat.read_errors++;
1189 sctx->stat.uncorrectable_errors++;
1190 spin_unlock(&sctx->stat_lock);
1191 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1195 /* setup the context, map the logical blocks and alloc the pages */
1196 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1198 spin_lock(&sctx->stat_lock);
1199 sctx->stat.read_errors++;
1200 sctx->stat.uncorrectable_errors++;
1201 spin_unlock(&sctx->stat_lock);
1202 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1205 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1206 sblock_bad = sblocks_for_recheck + failed_mirror_index;
1208 /* build and submit the bios for the failed mirror, check checksums */
1209 scrub_recheck_block(fs_info, sblock_bad, 1);
1211 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1212 sblock_bad->no_io_error_seen) {
1214 * the error disappeared after reading page by page, or
1215 * the area was part of a huge bio and other parts of the
1216 * bio caused I/O errors, or the block layer merged several
1217 * read requests into one and the error is caused by a
1218 * different bio (usually one of the two latter cases is
1221 spin_lock(&sctx->stat_lock);
1222 sctx->stat.unverified_errors++;
1223 sblock_to_check->data_corrected = 1;
1224 spin_unlock(&sctx->stat_lock);
1226 if (sctx->is_dev_replace)
1227 scrub_write_block_to_dev_replace(sblock_bad);
1231 if (!sblock_bad->no_io_error_seen) {
1232 spin_lock(&sctx->stat_lock);
1233 sctx->stat.read_errors++;
1234 spin_unlock(&sctx->stat_lock);
1235 if (__ratelimit(&_rs))
1236 scrub_print_warning("i/o error", sblock_to_check);
1237 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1238 } else if (sblock_bad->checksum_error) {
1239 spin_lock(&sctx->stat_lock);
1240 sctx->stat.csum_errors++;
1241 spin_unlock(&sctx->stat_lock);
1242 if (__ratelimit(&_rs))
1243 scrub_print_warning("checksum error", sblock_to_check);
1244 btrfs_dev_stat_inc_and_print(dev,
1245 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1246 } else if (sblock_bad->header_error) {
1247 spin_lock(&sctx->stat_lock);
1248 sctx->stat.verify_errors++;
1249 spin_unlock(&sctx->stat_lock);
1250 if (__ratelimit(&_rs))
1251 scrub_print_warning("checksum/header error",
1253 if (sblock_bad->generation_error)
1254 btrfs_dev_stat_inc_and_print(dev,
1255 BTRFS_DEV_STAT_GENERATION_ERRS);
1257 btrfs_dev_stat_inc_and_print(dev,
1258 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1261 if (sctx->readonly) {
1262 ASSERT(!sctx->is_dev_replace);
1267 * NOTE: Even for nodatasum case, it's still possible that it's a
1268 * compressed data extent, thus scrub_fixup_nodatasum(), which write
1269 * inode page cache onto disk, could cause serious data corruption.
1271 * So here we could only read from disk, and hope our recovery could
1272 * reach disk before the newer write.
1274 if (0 && !is_metadata && !have_csum) {
1275 struct scrub_fixup_nodatasum *fixup_nodatasum;
1277 WARN_ON(sctx->is_dev_replace);
1280 * !is_metadata and !have_csum, this means that the data
1281 * might not be COWed, that it might be modified
1282 * concurrently. The general strategy to work on the
1283 * commit root does not help in the case when COW is not
1286 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1287 if (!fixup_nodatasum)
1288 goto did_not_correct_error;
1289 fixup_nodatasum->sctx = sctx;
1290 fixup_nodatasum->dev = dev;
1291 fixup_nodatasum->logical = logical;
1292 fixup_nodatasum->root = fs_info->extent_root;
1293 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1294 scrub_pending_trans_workers_inc(sctx);
1295 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1296 scrub_fixup_nodatasum, NULL, NULL);
1297 btrfs_queue_work(fs_info->scrub_workers,
1298 &fixup_nodatasum->work);
1303 * now build and submit the bios for the other mirrors, check
1305 * First try to pick the mirror which is completely without I/O
1306 * errors and also does not have a checksum error.
1307 * If one is found, and if a checksum is present, the full block
1308 * that is known to contain an error is rewritten. Afterwards
1309 * the block is known to be corrected.
1310 * If a mirror is found which is completely correct, and no
1311 * checksum is present, only those pages are rewritten that had
1312 * an I/O error in the block to be repaired, since it cannot be
1313 * determined, which copy of the other pages is better (and it
1314 * could happen otherwise that a correct page would be
1315 * overwritten by a bad one).
1317 for (mirror_index = 0; ;mirror_index++) {
1318 struct scrub_block *sblock_other;
1320 if (mirror_index == failed_mirror_index)
1323 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1324 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1325 if (mirror_index >= BTRFS_MAX_MIRRORS)
1327 if (!sblocks_for_recheck[mirror_index].page_count)
1330 sblock_other = sblocks_for_recheck + mirror_index;
1332 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1333 int max_allowed = r->bbio->num_stripes -
1334 r->bbio->num_tgtdevs;
1336 if (mirror_index >= max_allowed)
1338 if (!sblocks_for_recheck[1].page_count)
1341 ASSERT(failed_mirror_index == 0);
1342 sblock_other = sblocks_for_recheck + 1;
1343 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1346 /* build and submit the bios, check checksums */
1347 scrub_recheck_block(fs_info, sblock_other, 0);
1349 if (!sblock_other->header_error &&
1350 !sblock_other->checksum_error &&
1351 sblock_other->no_io_error_seen) {
1352 if (sctx->is_dev_replace) {
1353 scrub_write_block_to_dev_replace(sblock_other);
1354 goto corrected_error;
1356 ret = scrub_repair_block_from_good_copy(
1357 sblock_bad, sblock_other);
1359 goto corrected_error;
1364 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1365 goto did_not_correct_error;
1368 * In case of I/O errors in the area that is supposed to be
1369 * repaired, continue by picking good copies of those pages.
1370 * Select the good pages from mirrors to rewrite bad pages from
1371 * the area to fix. Afterwards verify the checksum of the block
1372 * that is supposed to be repaired. This verification step is
1373 * only done for the purpose of statistic counting and for the
1374 * final scrub report, whether errors remain.
1375 * A perfect algorithm could make use of the checksum and try
1376 * all possible combinations of pages from the different mirrors
1377 * until the checksum verification succeeds. For example, when
1378 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1379 * of mirror #2 is readable but the final checksum test fails,
1380 * then the 2nd page of mirror #3 could be tried, whether now
1381 * the final checksum succeeds. But this would be a rare
1382 * exception and is therefore not implemented. At least it is
1383 * avoided that the good copy is overwritten.
1384 * A more useful improvement would be to pick the sectors
1385 * without I/O error based on sector sizes (512 bytes on legacy
1386 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1387 * mirror could be repaired by taking 512 byte of a different
1388 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1389 * area are unreadable.
1392 for (page_num = 0; page_num < sblock_bad->page_count;
1394 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1395 struct scrub_block *sblock_other = NULL;
1397 /* skip no-io-error page in scrub */
1398 if (!page_bad->io_error && !sctx->is_dev_replace)
1401 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1403 * In case of dev replace, if raid56 rebuild process
1404 * didn't work out correct data, then copy the content
1405 * in sblock_bad to make sure target device is identical
1406 * to source device, instead of writing garbage data in
1407 * sblock_for_recheck array to target device.
1409 sblock_other = NULL;
1410 } else if (page_bad->io_error) {
1411 /* try to find no-io-error page in mirrors */
1412 for (mirror_index = 0;
1413 mirror_index < BTRFS_MAX_MIRRORS &&
1414 sblocks_for_recheck[mirror_index].page_count > 0;
1416 if (!sblocks_for_recheck[mirror_index].
1417 pagev[page_num]->io_error) {
1418 sblock_other = sblocks_for_recheck +
1427 if (sctx->is_dev_replace) {
1429 * did not find a mirror to fetch the page
1430 * from. scrub_write_page_to_dev_replace()
1431 * handles this case (page->io_error), by
1432 * filling the block with zeros before
1433 * submitting the write request
1436 sblock_other = sblock_bad;
1438 if (scrub_write_page_to_dev_replace(sblock_other,
1440 btrfs_dev_replace_stats_inc(
1441 &fs_info->dev_replace.num_write_errors);
1444 } else if (sblock_other) {
1445 ret = scrub_repair_page_from_good_copy(sblock_bad,
1449 page_bad->io_error = 0;
1455 if (success && !sctx->is_dev_replace) {
1456 if (is_metadata || have_csum) {
1458 * need to verify the checksum now that all
1459 * sectors on disk are repaired (the write
1460 * request for data to be repaired is on its way).
1461 * Just be lazy and use scrub_recheck_block()
1462 * which re-reads the data before the checksum
1463 * is verified, but most likely the data comes out
1464 * of the page cache.
1466 scrub_recheck_block(fs_info, sblock_bad, 1);
1467 if (!sblock_bad->header_error &&
1468 !sblock_bad->checksum_error &&
1469 sblock_bad->no_io_error_seen)
1470 goto corrected_error;
1472 goto did_not_correct_error;
1475 spin_lock(&sctx->stat_lock);
1476 sctx->stat.corrected_errors++;
1477 sblock_to_check->data_corrected = 1;
1478 spin_unlock(&sctx->stat_lock);
1479 btrfs_err_rl_in_rcu(fs_info,
1480 "fixed up error at logical %llu on dev %s",
1481 logical, rcu_str_deref(dev->name));
1484 did_not_correct_error:
1485 spin_lock(&sctx->stat_lock);
1486 sctx->stat.uncorrectable_errors++;
1487 spin_unlock(&sctx->stat_lock);
1488 btrfs_err_rl_in_rcu(fs_info,
1489 "unable to fixup (regular) error at logical %llu on dev %s",
1490 logical, rcu_str_deref(dev->name));
1494 if (sblocks_for_recheck) {
1495 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1497 struct scrub_block *sblock = sblocks_for_recheck +
1499 struct scrub_recover *recover;
1502 for (page_index = 0; page_index < sblock->page_count;
1504 sblock->pagev[page_index]->sblock = NULL;
1505 recover = sblock->pagev[page_index]->recover;
1507 scrub_put_recover(fs_info, recover);
1508 sblock->pagev[page_index]->recover =
1511 scrub_page_put(sblock->pagev[page_index]);
1514 kfree(sblocks_for_recheck);
1517 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1523 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1525 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1527 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1530 return (int)bbio->num_stripes;
1533 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1536 int nstripes, int mirror,
1542 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1544 for (i = 0; i < nstripes; i++) {
1545 if (raid_map[i] == RAID6_Q_STRIPE ||
1546 raid_map[i] == RAID5_P_STRIPE)
1549 if (logical >= raid_map[i] &&
1550 logical < raid_map[i] + mapped_length)
1555 *stripe_offset = logical - raid_map[i];
1557 /* The other RAID type */
1558 *stripe_index = mirror;
1563 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1564 struct scrub_block *sblocks_for_recheck)
1566 struct scrub_ctx *sctx = original_sblock->sctx;
1567 struct btrfs_fs_info *fs_info = sctx->fs_info;
1568 u64 length = original_sblock->page_count * PAGE_SIZE;
1569 u64 logical = original_sblock->pagev[0]->logical;
1570 u64 generation = original_sblock->pagev[0]->generation;
1571 u64 flags = original_sblock->pagev[0]->flags;
1572 u64 have_csum = original_sblock->pagev[0]->have_csum;
1573 struct scrub_recover *recover;
1574 struct btrfs_bio *bbio;
1585 * note: the two members refs and outstanding_pages
1586 * are not used (and not set) in the blocks that are used for
1587 * the recheck procedure
1590 while (length > 0) {
1591 sublen = min_t(u64, length, PAGE_SIZE);
1592 mapped_length = sublen;
1596 * with a length of PAGE_SIZE, each returned stripe
1597 * represents one mirror
1599 btrfs_bio_counter_inc_blocked(fs_info);
1600 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1601 logical, &mapped_length, &bbio);
1602 if (ret || !bbio || mapped_length < sublen) {
1603 btrfs_put_bbio(bbio);
1604 btrfs_bio_counter_dec(fs_info);
1608 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1610 btrfs_put_bbio(bbio);
1611 btrfs_bio_counter_dec(fs_info);
1615 refcount_set(&recover->refs, 1);
1616 recover->bbio = bbio;
1617 recover->map_length = mapped_length;
1619 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1621 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1623 for (mirror_index = 0; mirror_index < nmirrors;
1625 struct scrub_block *sblock;
1626 struct scrub_page *page;
1628 sblock = sblocks_for_recheck + mirror_index;
1629 sblock->sctx = sctx;
1631 page = kzalloc(sizeof(*page), GFP_NOFS);
1634 spin_lock(&sctx->stat_lock);
1635 sctx->stat.malloc_errors++;
1636 spin_unlock(&sctx->stat_lock);
1637 scrub_put_recover(fs_info, recover);
1640 scrub_page_get(page);
1641 sblock->pagev[page_index] = page;
1642 page->sblock = sblock;
1643 page->flags = flags;
1644 page->generation = generation;
1645 page->logical = logical;
1646 page->have_csum = have_csum;
1649 original_sblock->pagev[0]->csum,
1652 scrub_stripe_index_and_offset(logical,
1661 page->physical = bbio->stripes[stripe_index].physical +
1663 page->dev = bbio->stripes[stripe_index].dev;
1665 BUG_ON(page_index >= original_sblock->page_count);
1666 page->physical_for_dev_replace =
1667 original_sblock->pagev[page_index]->
1668 physical_for_dev_replace;
1669 /* for missing devices, dev->bdev is NULL */
1670 page->mirror_num = mirror_index + 1;
1671 sblock->page_count++;
1672 page->page = alloc_page(GFP_NOFS);
1676 scrub_get_recover(recover);
1677 page->recover = recover;
1679 scrub_put_recover(fs_info, recover);
1688 static void scrub_bio_wait_endio(struct bio *bio)
1690 complete(bio->bi_private);
1693 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1695 struct scrub_page *page)
1697 DECLARE_COMPLETION_ONSTACK(done);
1701 bio->bi_iter.bi_sector = page->logical >> 9;
1702 bio->bi_private = &done;
1703 bio->bi_end_io = scrub_bio_wait_endio;
1705 mirror_num = page->sblock->pagev[0]->mirror_num;
1706 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1707 page->recover->map_length,
1712 wait_for_completion_io(&done);
1713 return blk_status_to_errno(bio->bi_status);
1716 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1717 struct scrub_block *sblock)
1719 struct scrub_page *first_page = sblock->pagev[0];
1723 /* All pages in sblock belong to the same stripe on the same device. */
1724 ASSERT(first_page->dev);
1725 if (!first_page->dev->bdev)
1728 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1729 bio_set_dev(bio, first_page->dev->bdev);
1731 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1732 struct scrub_page *page = sblock->pagev[page_num];
1734 WARN_ON(!page->page);
1735 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1738 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1745 scrub_recheck_block_checksum(sblock);
1749 for (page_num = 0; page_num < sblock->page_count; page_num++)
1750 sblock->pagev[page_num]->io_error = 1;
1752 sblock->no_io_error_seen = 0;
1756 * this function will check the on disk data for checksum errors, header
1757 * errors and read I/O errors. If any I/O errors happen, the exact pages
1758 * which are errored are marked as being bad. The goal is to enable scrub
1759 * to take those pages that are not errored from all the mirrors so that
1760 * the pages that are errored in the just handled mirror can be repaired.
1762 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1763 struct scrub_block *sblock,
1764 int retry_failed_mirror)
1768 sblock->no_io_error_seen = 1;
1770 /* short cut for raid56 */
1771 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1772 return scrub_recheck_block_on_raid56(fs_info, sblock);
1774 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1776 struct scrub_page *page = sblock->pagev[page_num];
1778 if (page->dev->bdev == NULL) {
1780 sblock->no_io_error_seen = 0;
1784 WARN_ON(!page->page);
1785 bio = btrfs_io_bio_alloc(1);
1786 bio_set_dev(bio, page->dev->bdev);
1788 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1789 bio->bi_iter.bi_sector = page->physical >> 9;
1790 bio->bi_opf = REQ_OP_READ;
1792 if (btrfsic_submit_bio_wait(bio)) {
1794 sblock->no_io_error_seen = 0;
1800 if (sblock->no_io_error_seen)
1801 scrub_recheck_block_checksum(sblock);
1804 static inline int scrub_check_fsid(u8 fsid[],
1805 struct scrub_page *spage)
1807 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1810 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1814 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1816 sblock->header_error = 0;
1817 sblock->checksum_error = 0;
1818 sblock->generation_error = 0;
1820 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1821 scrub_checksum_data(sblock);
1823 scrub_checksum_tree_block(sblock);
1826 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1827 struct scrub_block *sblock_good)
1832 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1835 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1845 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1846 struct scrub_block *sblock_good,
1847 int page_num, int force_write)
1849 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1850 struct scrub_page *page_good = sblock_good->pagev[page_num];
1851 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1853 BUG_ON(page_bad->page == NULL);
1854 BUG_ON(page_good->page == NULL);
1855 if (force_write || sblock_bad->header_error ||
1856 sblock_bad->checksum_error || page_bad->io_error) {
1860 if (!page_bad->dev->bdev) {
1861 btrfs_warn_rl(fs_info,
1862 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1866 bio = btrfs_io_bio_alloc(1);
1867 bio_set_dev(bio, page_bad->dev->bdev);
1868 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1869 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1871 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1872 if (PAGE_SIZE != ret) {
1877 if (btrfsic_submit_bio_wait(bio)) {
1878 btrfs_dev_stat_inc_and_print(page_bad->dev,
1879 BTRFS_DEV_STAT_WRITE_ERRS);
1880 btrfs_dev_replace_stats_inc(
1881 &fs_info->dev_replace.num_write_errors);
1891 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1893 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1897 * This block is used for the check of the parity on the source device,
1898 * so the data needn't be written into the destination device.
1900 if (sblock->sparity)
1903 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1906 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1908 btrfs_dev_replace_stats_inc(
1909 &fs_info->dev_replace.num_write_errors);
1913 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1916 struct scrub_page *spage = sblock->pagev[page_num];
1918 BUG_ON(spage->page == NULL);
1919 if (spage->io_error) {
1920 void *mapped_buffer = kmap_atomic(spage->page);
1922 clear_page(mapped_buffer);
1923 flush_dcache_page(spage->page);
1924 kunmap_atomic(mapped_buffer);
1926 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1929 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1930 struct scrub_page *spage)
1932 struct scrub_bio *sbio;
1935 mutex_lock(&sctx->wr_lock);
1937 if (!sctx->wr_curr_bio) {
1938 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1940 if (!sctx->wr_curr_bio) {
1941 mutex_unlock(&sctx->wr_lock);
1944 sctx->wr_curr_bio->sctx = sctx;
1945 sctx->wr_curr_bio->page_count = 0;
1947 sbio = sctx->wr_curr_bio;
1948 if (sbio->page_count == 0) {
1951 sbio->physical = spage->physical_for_dev_replace;
1952 sbio->logical = spage->logical;
1953 sbio->dev = sctx->wr_tgtdev;
1956 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1960 bio->bi_private = sbio;
1961 bio->bi_end_io = scrub_wr_bio_end_io;
1962 bio_set_dev(bio, sbio->dev->bdev);
1963 bio->bi_iter.bi_sector = sbio->physical >> 9;
1964 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1966 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1967 spage->physical_for_dev_replace ||
1968 sbio->logical + sbio->page_count * PAGE_SIZE !=
1970 scrub_wr_submit(sctx);
1974 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1975 if (ret != PAGE_SIZE) {
1976 if (sbio->page_count < 1) {
1979 mutex_unlock(&sctx->wr_lock);
1982 scrub_wr_submit(sctx);
1986 sbio->pagev[sbio->page_count] = spage;
1987 scrub_page_get(spage);
1989 if (sbio->page_count == sctx->pages_per_wr_bio)
1990 scrub_wr_submit(sctx);
1991 mutex_unlock(&sctx->wr_lock);
1996 static void scrub_wr_submit(struct scrub_ctx *sctx)
1998 struct scrub_bio *sbio;
2000 if (!sctx->wr_curr_bio)
2003 sbio = sctx->wr_curr_bio;
2004 sctx->wr_curr_bio = NULL;
2005 WARN_ON(!sbio->bio->bi_disk);
2006 scrub_pending_bio_inc(sctx);
2007 /* process all writes in a single worker thread. Then the block layer
2008 * orders the requests before sending them to the driver which
2009 * doubled the write performance on spinning disks when measured
2011 btrfsic_submit_bio(sbio->bio);
2014 static void scrub_wr_bio_end_io(struct bio *bio)
2016 struct scrub_bio *sbio = bio->bi_private;
2017 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2019 sbio->status = bio->bi_status;
2022 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2023 scrub_wr_bio_end_io_worker, NULL, NULL);
2024 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2027 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2029 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2030 struct scrub_ctx *sctx = sbio->sctx;
2033 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2035 struct btrfs_dev_replace *dev_replace =
2036 &sbio->sctx->fs_info->dev_replace;
2038 for (i = 0; i < sbio->page_count; i++) {
2039 struct scrub_page *spage = sbio->pagev[i];
2041 spage->io_error = 1;
2042 btrfs_dev_replace_stats_inc(&dev_replace->
2047 for (i = 0; i < sbio->page_count; i++)
2048 scrub_page_put(sbio->pagev[i]);
2052 scrub_pending_bio_dec(sctx);
2055 static int scrub_checksum(struct scrub_block *sblock)
2061 * No need to initialize these stats currently,
2062 * because this function only use return value
2063 * instead of these stats value.
2068 sblock->header_error = 0;
2069 sblock->generation_error = 0;
2070 sblock->checksum_error = 0;
2072 WARN_ON(sblock->page_count < 1);
2073 flags = sblock->pagev[0]->flags;
2075 if (flags & BTRFS_EXTENT_FLAG_DATA)
2076 ret = scrub_checksum_data(sblock);
2077 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2078 ret = scrub_checksum_tree_block(sblock);
2079 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2080 (void)scrub_checksum_super(sblock);
2084 scrub_handle_errored_block(sblock);
2089 static int scrub_checksum_data(struct scrub_block *sblock)
2091 struct scrub_ctx *sctx = sblock->sctx;
2092 u8 csum[BTRFS_CSUM_SIZE];
2100 BUG_ON(sblock->page_count < 1);
2101 if (!sblock->pagev[0]->have_csum)
2104 on_disk_csum = sblock->pagev[0]->csum;
2105 page = sblock->pagev[0]->page;
2106 buffer = kmap_atomic(page);
2108 len = sctx->fs_info->sectorsize;
2111 u64 l = min_t(u64, len, PAGE_SIZE);
2113 crc = btrfs_csum_data(buffer, crc, l);
2114 kunmap_atomic(buffer);
2119 BUG_ON(index >= sblock->page_count);
2120 BUG_ON(!sblock->pagev[index]->page);
2121 page = sblock->pagev[index]->page;
2122 buffer = kmap_atomic(page);
2125 btrfs_csum_final(crc, csum);
2126 if (memcmp(csum, on_disk_csum, sctx->csum_size))
2127 sblock->checksum_error = 1;
2129 return sblock->checksum_error;
2132 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2134 struct scrub_ctx *sctx = sblock->sctx;
2135 struct btrfs_header *h;
2136 struct btrfs_fs_info *fs_info = sctx->fs_info;
2137 u8 calculated_csum[BTRFS_CSUM_SIZE];
2138 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2140 void *mapped_buffer;
2147 BUG_ON(sblock->page_count < 1);
2148 page = sblock->pagev[0]->page;
2149 mapped_buffer = kmap_atomic(page);
2150 h = (struct btrfs_header *)mapped_buffer;
2151 memcpy(on_disk_csum, h->csum, sctx->csum_size);
2154 * we don't use the getter functions here, as we
2155 * a) don't have an extent buffer and
2156 * b) the page is already kmapped
2158 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2159 sblock->header_error = 1;
2161 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2162 sblock->header_error = 1;
2163 sblock->generation_error = 1;
2166 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2167 sblock->header_error = 1;
2169 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2171 sblock->header_error = 1;
2173 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2174 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2175 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2178 u64 l = min_t(u64, len, mapped_size);
2180 crc = btrfs_csum_data(p, crc, l);
2181 kunmap_atomic(mapped_buffer);
2186 BUG_ON(index >= sblock->page_count);
2187 BUG_ON(!sblock->pagev[index]->page);
2188 page = sblock->pagev[index]->page;
2189 mapped_buffer = kmap_atomic(page);
2190 mapped_size = PAGE_SIZE;
2194 btrfs_csum_final(crc, calculated_csum);
2195 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2196 sblock->checksum_error = 1;
2198 return sblock->header_error || sblock->checksum_error;
2201 static int scrub_checksum_super(struct scrub_block *sblock)
2203 struct btrfs_super_block *s;
2204 struct scrub_ctx *sctx = sblock->sctx;
2205 u8 calculated_csum[BTRFS_CSUM_SIZE];
2206 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2208 void *mapped_buffer;
2217 BUG_ON(sblock->page_count < 1);
2218 page = sblock->pagev[0]->page;
2219 mapped_buffer = kmap_atomic(page);
2220 s = (struct btrfs_super_block *)mapped_buffer;
2221 memcpy(on_disk_csum, s->csum, sctx->csum_size);
2223 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2226 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2229 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2232 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2233 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2234 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2237 u64 l = min_t(u64, len, mapped_size);
2239 crc = btrfs_csum_data(p, crc, l);
2240 kunmap_atomic(mapped_buffer);
2245 BUG_ON(index >= sblock->page_count);
2246 BUG_ON(!sblock->pagev[index]->page);
2247 page = sblock->pagev[index]->page;
2248 mapped_buffer = kmap_atomic(page);
2249 mapped_size = PAGE_SIZE;
2253 btrfs_csum_final(crc, calculated_csum);
2254 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2257 if (fail_cor + fail_gen) {
2259 * if we find an error in a super block, we just report it.
2260 * They will get written with the next transaction commit
2263 spin_lock(&sctx->stat_lock);
2264 ++sctx->stat.super_errors;
2265 spin_unlock(&sctx->stat_lock);
2267 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2268 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2270 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2271 BTRFS_DEV_STAT_GENERATION_ERRS);
2274 return fail_cor + fail_gen;
2277 static void scrub_block_get(struct scrub_block *sblock)
2279 refcount_inc(&sblock->refs);
2282 static void scrub_block_put(struct scrub_block *sblock)
2284 if (refcount_dec_and_test(&sblock->refs)) {
2287 if (sblock->sparity)
2288 scrub_parity_put(sblock->sparity);
2290 for (i = 0; i < sblock->page_count; i++)
2291 scrub_page_put(sblock->pagev[i]);
2296 static void scrub_page_get(struct scrub_page *spage)
2298 atomic_inc(&spage->refs);
2301 static void scrub_page_put(struct scrub_page *spage)
2303 if (atomic_dec_and_test(&spage->refs)) {
2305 __free_page(spage->page);
2310 static void scrub_submit(struct scrub_ctx *sctx)
2312 struct scrub_bio *sbio;
2314 if (sctx->curr == -1)
2317 sbio = sctx->bios[sctx->curr];
2319 scrub_pending_bio_inc(sctx);
2320 btrfsic_submit_bio(sbio->bio);
2323 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2324 struct scrub_page *spage)
2326 struct scrub_block *sblock = spage->sblock;
2327 struct scrub_bio *sbio;
2332 * grab a fresh bio or wait for one to become available
2334 while (sctx->curr == -1) {
2335 spin_lock(&sctx->list_lock);
2336 sctx->curr = sctx->first_free;
2337 if (sctx->curr != -1) {
2338 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2339 sctx->bios[sctx->curr]->next_free = -1;
2340 sctx->bios[sctx->curr]->page_count = 0;
2341 spin_unlock(&sctx->list_lock);
2343 spin_unlock(&sctx->list_lock);
2344 wait_event(sctx->list_wait, sctx->first_free != -1);
2347 sbio = sctx->bios[sctx->curr];
2348 if (sbio->page_count == 0) {
2351 sbio->physical = spage->physical;
2352 sbio->logical = spage->logical;
2353 sbio->dev = spage->dev;
2356 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2360 bio->bi_private = sbio;
2361 bio->bi_end_io = scrub_bio_end_io;
2362 bio_set_dev(bio, sbio->dev->bdev);
2363 bio->bi_iter.bi_sector = sbio->physical >> 9;
2364 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2366 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2368 sbio->logical + sbio->page_count * PAGE_SIZE !=
2370 sbio->dev != spage->dev) {
2375 sbio->pagev[sbio->page_count] = spage;
2376 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2377 if (ret != PAGE_SIZE) {
2378 if (sbio->page_count < 1) {
2387 scrub_block_get(sblock); /* one for the page added to the bio */
2388 atomic_inc(&sblock->outstanding_pages);
2390 if (sbio->page_count == sctx->pages_per_rd_bio)
2396 static void scrub_missing_raid56_end_io(struct bio *bio)
2398 struct scrub_block *sblock = bio->bi_private;
2399 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2402 sblock->no_io_error_seen = 0;
2406 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2409 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2411 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2412 struct scrub_ctx *sctx = sblock->sctx;
2413 struct btrfs_fs_info *fs_info = sctx->fs_info;
2415 struct btrfs_device *dev;
2417 logical = sblock->pagev[0]->logical;
2418 dev = sblock->pagev[0]->dev;
2420 if (sblock->no_io_error_seen)
2421 scrub_recheck_block_checksum(sblock);
2423 if (!sblock->no_io_error_seen) {
2424 spin_lock(&sctx->stat_lock);
2425 sctx->stat.read_errors++;
2426 spin_unlock(&sctx->stat_lock);
2427 btrfs_err_rl_in_rcu(fs_info,
2428 "IO error rebuilding logical %llu for dev %s",
2429 logical, rcu_str_deref(dev->name));
2430 } else if (sblock->header_error || sblock->checksum_error) {
2431 spin_lock(&sctx->stat_lock);
2432 sctx->stat.uncorrectable_errors++;
2433 spin_unlock(&sctx->stat_lock);
2434 btrfs_err_rl_in_rcu(fs_info,
2435 "failed to rebuild valid logical %llu for dev %s",
2436 logical, rcu_str_deref(dev->name));
2438 scrub_write_block_to_dev_replace(sblock);
2441 scrub_block_put(sblock);
2443 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2444 mutex_lock(&sctx->wr_lock);
2445 scrub_wr_submit(sctx);
2446 mutex_unlock(&sctx->wr_lock);
2449 scrub_pending_bio_dec(sctx);
2452 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2454 struct scrub_ctx *sctx = sblock->sctx;
2455 struct btrfs_fs_info *fs_info = sctx->fs_info;
2456 u64 length = sblock->page_count * PAGE_SIZE;
2457 u64 logical = sblock->pagev[0]->logical;
2458 struct btrfs_bio *bbio = NULL;
2460 struct btrfs_raid_bio *rbio;
2464 btrfs_bio_counter_inc_blocked(fs_info);
2465 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2467 if (ret || !bbio || !bbio->raid_map)
2470 if (WARN_ON(!sctx->is_dev_replace ||
2471 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2473 * We shouldn't be scrubbing a missing device. Even for dev
2474 * replace, we should only get here for RAID 5/6. We either
2475 * managed to mount something with no mirrors remaining or
2476 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2481 bio = btrfs_io_bio_alloc(0);
2482 bio->bi_iter.bi_sector = logical >> 9;
2483 bio->bi_private = sblock;
2484 bio->bi_end_io = scrub_missing_raid56_end_io;
2486 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2490 for (i = 0; i < sblock->page_count; i++) {
2491 struct scrub_page *spage = sblock->pagev[i];
2493 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2496 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2497 scrub_missing_raid56_worker, NULL, NULL);
2498 scrub_block_get(sblock);
2499 scrub_pending_bio_inc(sctx);
2500 raid56_submit_missing_rbio(rbio);
2506 btrfs_bio_counter_dec(fs_info);
2507 btrfs_put_bbio(bbio);
2508 spin_lock(&sctx->stat_lock);
2509 sctx->stat.malloc_errors++;
2510 spin_unlock(&sctx->stat_lock);
2513 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2514 u64 physical, struct btrfs_device *dev, u64 flags,
2515 u64 gen, int mirror_num, u8 *csum, int force,
2516 u64 physical_for_dev_replace)
2518 struct scrub_block *sblock;
2521 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2523 spin_lock(&sctx->stat_lock);
2524 sctx->stat.malloc_errors++;
2525 spin_unlock(&sctx->stat_lock);
2529 /* one ref inside this function, plus one for each page added to
2531 refcount_set(&sblock->refs, 1);
2532 sblock->sctx = sctx;
2533 sblock->no_io_error_seen = 1;
2535 for (index = 0; len > 0; index++) {
2536 struct scrub_page *spage;
2537 u64 l = min_t(u64, len, PAGE_SIZE);
2539 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2542 spin_lock(&sctx->stat_lock);
2543 sctx->stat.malloc_errors++;
2544 spin_unlock(&sctx->stat_lock);
2545 scrub_block_put(sblock);
2548 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2549 scrub_page_get(spage);
2550 sblock->pagev[index] = spage;
2551 spage->sblock = sblock;
2553 spage->flags = flags;
2554 spage->generation = gen;
2555 spage->logical = logical;
2556 spage->physical = physical;
2557 spage->physical_for_dev_replace = physical_for_dev_replace;
2558 spage->mirror_num = mirror_num;
2560 spage->have_csum = 1;
2561 memcpy(spage->csum, csum, sctx->csum_size);
2563 spage->have_csum = 0;
2565 sblock->page_count++;
2566 spage->page = alloc_page(GFP_KERNEL);
2572 physical_for_dev_replace += l;
2575 WARN_ON(sblock->page_count == 0);
2576 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2578 * This case should only be hit for RAID 5/6 device replace. See
2579 * the comment in scrub_missing_raid56_pages() for details.
2581 scrub_missing_raid56_pages(sblock);
2583 for (index = 0; index < sblock->page_count; index++) {
2584 struct scrub_page *spage = sblock->pagev[index];
2587 ret = scrub_add_page_to_rd_bio(sctx, spage);
2589 scrub_block_put(sblock);
2598 /* last one frees, either here or in bio completion for last page */
2599 scrub_block_put(sblock);
2603 static void scrub_bio_end_io(struct bio *bio)
2605 struct scrub_bio *sbio = bio->bi_private;
2606 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2608 sbio->status = bio->bi_status;
2611 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2614 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2616 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2617 struct scrub_ctx *sctx = sbio->sctx;
2620 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2622 for (i = 0; i < sbio->page_count; i++) {
2623 struct scrub_page *spage = sbio->pagev[i];
2625 spage->io_error = 1;
2626 spage->sblock->no_io_error_seen = 0;
2630 /* now complete the scrub_block items that have all pages completed */
2631 for (i = 0; i < sbio->page_count; i++) {
2632 struct scrub_page *spage = sbio->pagev[i];
2633 struct scrub_block *sblock = spage->sblock;
2635 if (atomic_dec_and_test(&sblock->outstanding_pages))
2636 scrub_block_complete(sblock);
2637 scrub_block_put(sblock);
2642 spin_lock(&sctx->list_lock);
2643 sbio->next_free = sctx->first_free;
2644 sctx->first_free = sbio->index;
2645 spin_unlock(&sctx->list_lock);
2647 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2648 mutex_lock(&sctx->wr_lock);
2649 scrub_wr_submit(sctx);
2650 mutex_unlock(&sctx->wr_lock);
2653 scrub_pending_bio_dec(sctx);
2656 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2657 unsigned long *bitmap,
2663 int sectorsize = sparity->sctx->fs_info->sectorsize;
2665 if (len >= sparity->stripe_len) {
2666 bitmap_set(bitmap, 0, sparity->nsectors);
2670 start -= sparity->logic_start;
2671 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2672 offset = div_u64(offset, sectorsize);
2673 nsectors64 = div_u64(len, sectorsize);
2675 ASSERT(nsectors64 < UINT_MAX);
2676 nsectors = (u32)nsectors64;
2678 if (offset + nsectors <= sparity->nsectors) {
2679 bitmap_set(bitmap, offset, nsectors);
2683 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2684 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2687 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2690 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2693 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2696 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2699 static void scrub_block_complete(struct scrub_block *sblock)
2703 if (!sblock->no_io_error_seen) {
2705 scrub_handle_errored_block(sblock);
2708 * if has checksum error, write via repair mechanism in
2709 * dev replace case, otherwise write here in dev replace
2712 corrupted = scrub_checksum(sblock);
2713 if (!corrupted && sblock->sctx->is_dev_replace)
2714 scrub_write_block_to_dev_replace(sblock);
2717 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2718 u64 start = sblock->pagev[0]->logical;
2719 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2722 scrub_parity_mark_sectors_error(sblock->sparity,
2723 start, end - start);
2727 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2729 struct btrfs_ordered_sum *sum = NULL;
2730 unsigned long index;
2731 unsigned long num_sectors;
2733 while (!list_empty(&sctx->csum_list)) {
2734 sum = list_first_entry(&sctx->csum_list,
2735 struct btrfs_ordered_sum, list);
2736 if (sum->bytenr > logical)
2738 if (sum->bytenr + sum->len > logical)
2741 ++sctx->stat.csum_discards;
2742 list_del(&sum->list);
2749 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2750 ASSERT(index < UINT_MAX);
2752 num_sectors = sum->len / sctx->fs_info->sectorsize;
2753 memcpy(csum, sum->sums + index, sctx->csum_size);
2754 if (index == num_sectors - 1) {
2755 list_del(&sum->list);
2761 /* scrub extent tries to collect up to 64 kB for each bio */
2762 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2763 u64 logical, u64 len,
2764 u64 physical, struct btrfs_device *dev, u64 flags,
2765 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2768 u8 csum[BTRFS_CSUM_SIZE];
2771 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2772 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2773 blocksize = map->stripe_len;
2775 blocksize = sctx->fs_info->sectorsize;
2776 spin_lock(&sctx->stat_lock);
2777 sctx->stat.data_extents_scrubbed++;
2778 sctx->stat.data_bytes_scrubbed += len;
2779 spin_unlock(&sctx->stat_lock);
2780 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2781 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2782 blocksize = map->stripe_len;
2784 blocksize = sctx->fs_info->nodesize;
2785 spin_lock(&sctx->stat_lock);
2786 sctx->stat.tree_extents_scrubbed++;
2787 sctx->stat.tree_bytes_scrubbed += len;
2788 spin_unlock(&sctx->stat_lock);
2790 blocksize = sctx->fs_info->sectorsize;
2795 u64 l = min_t(u64, len, blocksize);
2798 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2799 /* push csums to sbio */
2800 have_csum = scrub_find_csum(sctx, logical, csum);
2802 ++sctx->stat.no_csum;
2803 if (0 && sctx->is_dev_replace && !have_csum) {
2804 ret = copy_nocow_pages(sctx, logical, l,
2806 physical_for_dev_replace);
2807 goto behind_scrub_pages;
2810 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2811 mirror_num, have_csum ? csum : NULL, 0,
2812 physical_for_dev_replace);
2819 physical_for_dev_replace += l;
2824 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2825 u64 logical, u64 len,
2826 u64 physical, struct btrfs_device *dev,
2827 u64 flags, u64 gen, int mirror_num, u8 *csum)
2829 struct scrub_ctx *sctx = sparity->sctx;
2830 struct scrub_block *sblock;
2833 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2835 spin_lock(&sctx->stat_lock);
2836 sctx->stat.malloc_errors++;
2837 spin_unlock(&sctx->stat_lock);
2841 /* one ref inside this function, plus one for each page added to
2843 refcount_set(&sblock->refs, 1);
2844 sblock->sctx = sctx;
2845 sblock->no_io_error_seen = 1;
2846 sblock->sparity = sparity;
2847 scrub_parity_get(sparity);
2849 for (index = 0; len > 0; index++) {
2850 struct scrub_page *spage;
2851 u64 l = min_t(u64, len, PAGE_SIZE);
2853 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2856 spin_lock(&sctx->stat_lock);
2857 sctx->stat.malloc_errors++;
2858 spin_unlock(&sctx->stat_lock);
2859 scrub_block_put(sblock);
2862 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2863 /* For scrub block */
2864 scrub_page_get(spage);
2865 sblock->pagev[index] = spage;
2866 /* For scrub parity */
2867 scrub_page_get(spage);
2868 list_add_tail(&spage->list, &sparity->spages);
2869 spage->sblock = sblock;
2871 spage->flags = flags;
2872 spage->generation = gen;
2873 spage->logical = logical;
2874 spage->physical = physical;
2875 spage->mirror_num = mirror_num;
2877 spage->have_csum = 1;
2878 memcpy(spage->csum, csum, sctx->csum_size);
2880 spage->have_csum = 0;
2882 sblock->page_count++;
2883 spage->page = alloc_page(GFP_KERNEL);
2891 WARN_ON(sblock->page_count == 0);
2892 for (index = 0; index < sblock->page_count; index++) {
2893 struct scrub_page *spage = sblock->pagev[index];
2896 ret = scrub_add_page_to_rd_bio(sctx, spage);
2898 scrub_block_put(sblock);
2903 /* last one frees, either here or in bio completion for last page */
2904 scrub_block_put(sblock);
2908 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2909 u64 logical, u64 len,
2910 u64 physical, struct btrfs_device *dev,
2911 u64 flags, u64 gen, int mirror_num)
2913 struct scrub_ctx *sctx = sparity->sctx;
2915 u8 csum[BTRFS_CSUM_SIZE];
2918 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2919 scrub_parity_mark_sectors_error(sparity, logical, len);
2923 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2924 blocksize = sparity->stripe_len;
2925 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2926 blocksize = sparity->stripe_len;
2928 blocksize = sctx->fs_info->sectorsize;
2933 u64 l = min_t(u64, len, blocksize);
2936 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2937 /* push csums to sbio */
2938 have_csum = scrub_find_csum(sctx, logical, csum);
2942 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2943 flags, gen, mirror_num,
2944 have_csum ? csum : NULL);
2956 * Given a physical address, this will calculate it's
2957 * logical offset. if this is a parity stripe, it will return
2958 * the most left data stripe's logical offset.
2960 * return 0 if it is a data stripe, 1 means parity stripe.
2962 static int get_raid56_logic_offset(u64 physical, int num,
2963 struct map_lookup *map, u64 *offset,
2973 last_offset = (physical - map->stripes[num].physical) *
2974 nr_data_stripes(map);
2976 *stripe_start = last_offset;
2978 *offset = last_offset;
2979 for (i = 0; i < nr_data_stripes(map); i++) {
2980 *offset = last_offset + i * map->stripe_len;
2982 stripe_nr = div64_u64(*offset, map->stripe_len);
2983 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2985 /* Work out the disk rotation on this stripe-set */
2986 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2987 /* calculate which stripe this data locates */
2989 stripe_index = rot % map->num_stripes;
2990 if (stripe_index == num)
2992 if (stripe_index < num)
2995 *offset = last_offset + j * map->stripe_len;
2999 static void scrub_free_parity(struct scrub_parity *sparity)
3001 struct scrub_ctx *sctx = sparity->sctx;
3002 struct scrub_page *curr, *next;
3005 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
3007 spin_lock(&sctx->stat_lock);
3008 sctx->stat.read_errors += nbits;
3009 sctx->stat.uncorrectable_errors += nbits;
3010 spin_unlock(&sctx->stat_lock);
3013 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
3014 list_del_init(&curr->list);
3015 scrub_page_put(curr);
3021 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
3023 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3025 struct scrub_ctx *sctx = sparity->sctx;
3027 scrub_free_parity(sparity);
3028 scrub_pending_bio_dec(sctx);
3031 static void scrub_parity_bio_endio(struct bio *bio)
3033 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3034 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3037 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3042 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3043 scrub_parity_bio_endio_worker, NULL, NULL);
3044 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3047 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3049 struct scrub_ctx *sctx = sparity->sctx;
3050 struct btrfs_fs_info *fs_info = sctx->fs_info;
3052 struct btrfs_raid_bio *rbio;
3053 struct btrfs_bio *bbio = NULL;
3057 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3061 length = sparity->logic_end - sparity->logic_start;
3063 btrfs_bio_counter_inc_blocked(fs_info);
3064 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3066 if (ret || !bbio || !bbio->raid_map)
3069 bio = btrfs_io_bio_alloc(0);
3070 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3071 bio->bi_private = sparity;
3072 bio->bi_end_io = scrub_parity_bio_endio;
3074 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3075 length, sparity->scrub_dev,
3081 scrub_pending_bio_inc(sctx);
3082 raid56_parity_submit_scrub_rbio(rbio);
3088 btrfs_bio_counter_dec(fs_info);
3089 btrfs_put_bbio(bbio);
3090 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3092 spin_lock(&sctx->stat_lock);
3093 sctx->stat.malloc_errors++;
3094 spin_unlock(&sctx->stat_lock);
3096 scrub_free_parity(sparity);
3099 static inline int scrub_calc_parity_bitmap_len(int nsectors)
3101 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3104 static void scrub_parity_get(struct scrub_parity *sparity)
3106 refcount_inc(&sparity->refs);
3109 static void scrub_parity_put(struct scrub_parity *sparity)
3111 if (!refcount_dec_and_test(&sparity->refs))
3114 scrub_parity_check_and_repair(sparity);
3117 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3118 struct map_lookup *map,
3119 struct btrfs_device *sdev,
3120 struct btrfs_path *path,
3124 struct btrfs_fs_info *fs_info = sctx->fs_info;
3125 struct btrfs_root *root = fs_info->extent_root;
3126 struct btrfs_root *csum_root = fs_info->csum_root;
3127 struct btrfs_extent_item *extent;
3128 struct btrfs_bio *bbio = NULL;
3132 struct extent_buffer *l;
3133 struct btrfs_key key;
3136 u64 extent_physical;
3139 struct btrfs_device *extent_dev;
3140 struct scrub_parity *sparity;
3143 int extent_mirror_num;
3146 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3147 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3148 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3151 spin_lock(&sctx->stat_lock);
3152 sctx->stat.malloc_errors++;
3153 spin_unlock(&sctx->stat_lock);
3157 sparity->stripe_len = map->stripe_len;
3158 sparity->nsectors = nsectors;
3159 sparity->sctx = sctx;
3160 sparity->scrub_dev = sdev;
3161 sparity->logic_start = logic_start;
3162 sparity->logic_end = logic_end;
3163 refcount_set(&sparity->refs, 1);
3164 INIT_LIST_HEAD(&sparity->spages);
3165 sparity->dbitmap = sparity->bitmap;
3166 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3169 while (logic_start < logic_end) {
3170 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3171 key.type = BTRFS_METADATA_ITEM_KEY;
3173 key.type = BTRFS_EXTENT_ITEM_KEY;
3174 key.objectid = logic_start;
3175 key.offset = (u64)-1;
3177 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3182 ret = btrfs_previous_extent_item(root, path, 0);
3186 btrfs_release_path(path);
3187 ret = btrfs_search_slot(NULL, root, &key,
3199 slot = path->slots[0];
3200 if (slot >= btrfs_header_nritems(l)) {
3201 ret = btrfs_next_leaf(root, path);
3210 btrfs_item_key_to_cpu(l, &key, slot);
3212 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3213 key.type != BTRFS_METADATA_ITEM_KEY)
3216 if (key.type == BTRFS_METADATA_ITEM_KEY)
3217 bytes = fs_info->nodesize;
3221 if (key.objectid + bytes <= logic_start)
3224 if (key.objectid >= logic_end) {
3229 while (key.objectid >= logic_start + map->stripe_len)
3230 logic_start += map->stripe_len;
3232 extent = btrfs_item_ptr(l, slot,
3233 struct btrfs_extent_item);
3234 flags = btrfs_extent_flags(l, extent);
3235 generation = btrfs_extent_generation(l, extent);
3237 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3238 (key.objectid < logic_start ||
3239 key.objectid + bytes >
3240 logic_start + map->stripe_len)) {
3242 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3243 key.objectid, logic_start);
3244 spin_lock(&sctx->stat_lock);
3245 sctx->stat.uncorrectable_errors++;
3246 spin_unlock(&sctx->stat_lock);
3250 extent_logical = key.objectid;
3253 if (extent_logical < logic_start) {
3254 extent_len -= logic_start - extent_logical;
3255 extent_logical = logic_start;
3258 if (extent_logical + extent_len >
3259 logic_start + map->stripe_len)
3260 extent_len = logic_start + map->stripe_len -
3263 scrub_parity_mark_sectors_data(sparity, extent_logical,
3266 mapped_length = extent_len;
3268 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3269 extent_logical, &mapped_length, &bbio,
3272 if (!bbio || mapped_length < extent_len)
3276 btrfs_put_bbio(bbio);
3279 extent_physical = bbio->stripes[0].physical;
3280 extent_mirror_num = bbio->mirror_num;
3281 extent_dev = bbio->stripes[0].dev;
3282 btrfs_put_bbio(bbio);
3284 ret = btrfs_lookup_csums_range(csum_root,
3286 extent_logical + extent_len - 1,
3287 &sctx->csum_list, 1);
3291 ret = scrub_extent_for_parity(sparity, extent_logical,
3298 scrub_free_csums(sctx);
3303 if (extent_logical + extent_len <
3304 key.objectid + bytes) {
3305 logic_start += map->stripe_len;
3307 if (logic_start >= logic_end) {
3312 if (logic_start < key.objectid + bytes) {
3321 btrfs_release_path(path);
3326 logic_start += map->stripe_len;
3330 scrub_parity_mark_sectors_error(sparity, logic_start,
3331 logic_end - logic_start);
3332 scrub_parity_put(sparity);
3334 mutex_lock(&sctx->wr_lock);
3335 scrub_wr_submit(sctx);
3336 mutex_unlock(&sctx->wr_lock);
3338 btrfs_release_path(path);
3339 return ret < 0 ? ret : 0;
3342 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3343 struct map_lookup *map,
3344 struct btrfs_device *scrub_dev,
3345 int num, u64 base, u64 length,
3348 struct btrfs_path *path, *ppath;
3349 struct btrfs_fs_info *fs_info = sctx->fs_info;
3350 struct btrfs_root *root = fs_info->extent_root;
3351 struct btrfs_root *csum_root = fs_info->csum_root;
3352 struct btrfs_extent_item *extent;
3353 struct blk_plug plug;
3358 struct extent_buffer *l;
3365 struct reada_control *reada1;
3366 struct reada_control *reada2;
3367 struct btrfs_key key;
3368 struct btrfs_key key_end;
3369 u64 increment = map->stripe_len;
3372 u64 extent_physical;
3376 struct btrfs_device *extent_dev;
3377 int extent_mirror_num;
3380 physical = map->stripes[num].physical;
3382 nstripes = div64_u64(length, map->stripe_len);
3383 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3384 offset = map->stripe_len * num;
3385 increment = map->stripe_len * map->num_stripes;
3387 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3388 int factor = map->num_stripes / map->sub_stripes;
3389 offset = map->stripe_len * (num / map->sub_stripes);
3390 increment = map->stripe_len * factor;
3391 mirror_num = num % map->sub_stripes + 1;
3392 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3393 increment = map->stripe_len;
3394 mirror_num = num % map->num_stripes + 1;
3395 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3396 increment = map->stripe_len;
3397 mirror_num = num % map->num_stripes + 1;
3398 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3399 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3400 increment = map->stripe_len * nr_data_stripes(map);
3403 increment = map->stripe_len;
3407 path = btrfs_alloc_path();
3411 ppath = btrfs_alloc_path();
3413 btrfs_free_path(path);
3418 * work on commit root. The related disk blocks are static as
3419 * long as COW is applied. This means, it is save to rewrite
3420 * them to repair disk errors without any race conditions
3422 path->search_commit_root = 1;
3423 path->skip_locking = 1;
3425 ppath->search_commit_root = 1;
3426 ppath->skip_locking = 1;
3428 * trigger the readahead for extent tree csum tree and wait for
3429 * completion. During readahead, the scrub is officially paused
3430 * to not hold off transaction commits
3432 logical = base + offset;
3433 physical_end = physical + nstripes * map->stripe_len;
3434 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3435 get_raid56_logic_offset(physical_end, num,
3436 map, &logic_end, NULL);
3439 logic_end = logical + increment * nstripes;
3441 wait_event(sctx->list_wait,
3442 atomic_read(&sctx->bios_in_flight) == 0);
3443 scrub_blocked_if_needed(fs_info);
3445 /* FIXME it might be better to start readahead at commit root */
3446 key.objectid = logical;
3447 key.type = BTRFS_EXTENT_ITEM_KEY;
3448 key.offset = (u64)0;
3449 key_end.objectid = logic_end;
3450 key_end.type = BTRFS_METADATA_ITEM_KEY;
3451 key_end.offset = (u64)-1;
3452 reada1 = btrfs_reada_add(root, &key, &key_end);
3454 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3455 key.type = BTRFS_EXTENT_CSUM_KEY;
3456 key.offset = logical;
3457 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3458 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3459 key_end.offset = logic_end;
3460 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3462 if (!IS_ERR(reada1))
3463 btrfs_reada_wait(reada1);
3464 if (!IS_ERR(reada2))
3465 btrfs_reada_wait(reada2);
3469 * collect all data csums for the stripe to avoid seeking during
3470 * the scrub. This might currently (crc32) end up to be about 1MB
3472 blk_start_plug(&plug);
3475 * now find all extents for each stripe and scrub them
3478 while (physical < physical_end) {
3482 if (atomic_read(&fs_info->scrub_cancel_req) ||
3483 atomic_read(&sctx->cancel_req)) {
3488 * check to see if we have to pause
3490 if (atomic_read(&fs_info->scrub_pause_req)) {
3491 /* push queued extents */
3492 sctx->flush_all_writes = true;
3494 mutex_lock(&sctx->wr_lock);
3495 scrub_wr_submit(sctx);
3496 mutex_unlock(&sctx->wr_lock);
3497 wait_event(sctx->list_wait,
3498 atomic_read(&sctx->bios_in_flight) == 0);
3499 sctx->flush_all_writes = false;
3500 scrub_blocked_if_needed(fs_info);
3503 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3504 ret = get_raid56_logic_offset(physical, num, map,
3509 /* it is parity strip */
3510 stripe_logical += base;
3511 stripe_end = stripe_logical + increment;
3512 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3513 ppath, stripe_logical,
3521 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3522 key.type = BTRFS_METADATA_ITEM_KEY;
3524 key.type = BTRFS_EXTENT_ITEM_KEY;
3525 key.objectid = logical;
3526 key.offset = (u64)-1;
3528 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3533 ret = btrfs_previous_extent_item(root, path, 0);
3537 /* there's no smaller item, so stick with the
3539 btrfs_release_path(path);
3540 ret = btrfs_search_slot(NULL, root, &key,
3552 slot = path->slots[0];
3553 if (slot >= btrfs_header_nritems(l)) {
3554 ret = btrfs_next_leaf(root, path);
3563 btrfs_item_key_to_cpu(l, &key, slot);
3565 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3566 key.type != BTRFS_METADATA_ITEM_KEY)
3569 if (key.type == BTRFS_METADATA_ITEM_KEY)
3570 bytes = fs_info->nodesize;
3574 if (key.objectid + bytes <= logical)
3577 if (key.objectid >= logical + map->stripe_len) {
3578 /* out of this device extent */
3579 if (key.objectid >= logic_end)
3584 extent = btrfs_item_ptr(l, slot,
3585 struct btrfs_extent_item);
3586 flags = btrfs_extent_flags(l, extent);
3587 generation = btrfs_extent_generation(l, extent);
3589 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3590 (key.objectid < logical ||
3591 key.objectid + bytes >
3592 logical + map->stripe_len)) {
3594 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3595 key.objectid, logical);
3596 spin_lock(&sctx->stat_lock);
3597 sctx->stat.uncorrectable_errors++;
3598 spin_unlock(&sctx->stat_lock);
3603 extent_logical = key.objectid;
3607 * trim extent to this stripe
3609 if (extent_logical < logical) {
3610 extent_len -= logical - extent_logical;
3611 extent_logical = logical;
3613 if (extent_logical + extent_len >
3614 logical + map->stripe_len) {
3615 extent_len = logical + map->stripe_len -
3619 extent_physical = extent_logical - logical + physical;
3620 extent_dev = scrub_dev;
3621 extent_mirror_num = mirror_num;
3623 scrub_remap_extent(fs_info, extent_logical,
3624 extent_len, &extent_physical,
3626 &extent_mirror_num);
3628 ret = btrfs_lookup_csums_range(csum_root,
3632 &sctx->csum_list, 1);
3636 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3637 extent_physical, extent_dev, flags,
3638 generation, extent_mirror_num,
3639 extent_logical - logical + physical);
3641 scrub_free_csums(sctx);
3646 if (extent_logical + extent_len <
3647 key.objectid + bytes) {
3648 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3650 * loop until we find next data stripe
3651 * or we have finished all stripes.
3654 physical += map->stripe_len;
3655 ret = get_raid56_logic_offset(physical,
3660 if (ret && physical < physical_end) {
3661 stripe_logical += base;
3662 stripe_end = stripe_logical +
3664 ret = scrub_raid56_parity(sctx,
3665 map, scrub_dev, ppath,
3673 physical += map->stripe_len;
3674 logical += increment;
3676 if (logical < key.objectid + bytes) {
3681 if (physical >= physical_end) {
3689 btrfs_release_path(path);
3691 logical += increment;
3692 physical += map->stripe_len;
3693 spin_lock(&sctx->stat_lock);
3695 sctx->stat.last_physical = map->stripes[num].physical +
3698 sctx->stat.last_physical = physical;
3699 spin_unlock(&sctx->stat_lock);
3704 /* push queued extents */
3706 mutex_lock(&sctx->wr_lock);
3707 scrub_wr_submit(sctx);
3708 mutex_unlock(&sctx->wr_lock);
3710 blk_finish_plug(&plug);
3711 btrfs_free_path(path);
3712 btrfs_free_path(ppath);
3713 return ret < 0 ? ret : 0;
3716 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3717 struct btrfs_device *scrub_dev,
3718 u64 chunk_offset, u64 length,
3720 struct btrfs_block_group_cache *cache,
3723 struct btrfs_fs_info *fs_info = sctx->fs_info;
3724 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3725 struct map_lookup *map;
3726 struct extent_map *em;
3730 read_lock(&map_tree->map_tree.lock);
3731 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3732 read_unlock(&map_tree->map_tree.lock);
3736 * Might have been an unused block group deleted by the cleaner
3737 * kthread or relocation.
3739 spin_lock(&cache->lock);
3740 if (!cache->removed)
3742 spin_unlock(&cache->lock);
3747 map = em->map_lookup;
3748 if (em->start != chunk_offset)
3751 if (em->len < length)
3754 for (i = 0; i < map->num_stripes; ++i) {
3755 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3756 map->stripes[i].physical == dev_offset) {
3757 ret = scrub_stripe(sctx, map, scrub_dev, i,
3758 chunk_offset, length,
3765 free_extent_map(em);
3770 static noinline_for_stack
3771 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3772 struct btrfs_device *scrub_dev, u64 start, u64 end,
3775 struct btrfs_dev_extent *dev_extent = NULL;
3776 struct btrfs_path *path;
3777 struct btrfs_fs_info *fs_info = sctx->fs_info;
3778 struct btrfs_root *root = fs_info->dev_root;
3784 struct extent_buffer *l;
3785 struct btrfs_key key;
3786 struct btrfs_key found_key;
3787 struct btrfs_block_group_cache *cache;
3788 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3790 path = btrfs_alloc_path();
3794 path->reada = READA_FORWARD;
3795 path->search_commit_root = 1;
3796 path->skip_locking = 1;
3798 key.objectid = scrub_dev->devid;
3800 key.type = BTRFS_DEV_EXTENT_KEY;
3803 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3807 if (path->slots[0] >=
3808 btrfs_header_nritems(path->nodes[0])) {
3809 ret = btrfs_next_leaf(root, path);
3822 slot = path->slots[0];
3824 btrfs_item_key_to_cpu(l, &found_key, slot);
3826 if (found_key.objectid != scrub_dev->devid)
3829 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3832 if (found_key.offset >= end)
3835 if (found_key.offset < key.offset)
3838 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3839 length = btrfs_dev_extent_length(l, dev_extent);
3841 if (found_key.offset + length <= start)
3844 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3847 * get a reference on the corresponding block group to prevent
3848 * the chunk from going away while we scrub it
3850 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3852 /* some chunks are removed but not committed to disk yet,
3853 * continue scrubbing */
3858 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3859 * to avoid deadlock caused by:
3860 * btrfs_inc_block_group_ro()
3861 * -> btrfs_wait_for_commit()
3862 * -> btrfs_commit_transaction()
3863 * -> btrfs_scrub_pause()
3865 scrub_pause_on(fs_info);
3866 ret = btrfs_inc_block_group_ro(fs_info, cache);
3867 if (!ret && is_dev_replace) {
3869 * If we are doing a device replace wait for any tasks
3870 * that started dellaloc right before we set the block
3871 * group to RO mode, as they might have just allocated
3872 * an extent from it or decided they could do a nocow
3873 * write. And if any such tasks did that, wait for their
3874 * ordered extents to complete and then commit the
3875 * current transaction, so that we can later see the new
3876 * extent items in the extent tree - the ordered extents
3877 * create delayed data references (for cow writes) when
3878 * they complete, which will be run and insert the
3879 * corresponding extent items into the extent tree when
3880 * we commit the transaction they used when running
3881 * inode.c:btrfs_finish_ordered_io(). We later use
3882 * the commit root of the extent tree to find extents
3883 * to copy from the srcdev into the tgtdev, and we don't
3884 * want to miss any new extents.
3886 btrfs_wait_block_group_reservations(cache);
3887 btrfs_wait_nocow_writers(cache);
3888 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3889 cache->key.objectid,
3892 struct btrfs_trans_handle *trans;
3894 trans = btrfs_join_transaction(root);
3896 ret = PTR_ERR(trans);
3898 ret = btrfs_commit_transaction(trans);
3900 scrub_pause_off(fs_info);
3901 btrfs_put_block_group(cache);
3906 scrub_pause_off(fs_info);
3910 } else if (ret == -ENOSPC) {
3912 * btrfs_inc_block_group_ro return -ENOSPC when it
3913 * failed in creating new chunk for metadata.
3914 * It is not a problem for scrub/replace, because
3915 * metadata are always cowed, and our scrub paused
3916 * commit_transactions.
3921 "failed setting block group ro: %d", ret);
3922 btrfs_put_block_group(cache);
3926 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3927 dev_replace->cursor_right = found_key.offset + length;
3928 dev_replace->cursor_left = found_key.offset;
3929 dev_replace->item_needs_writeback = 1;
3930 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3931 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3932 found_key.offset, cache, is_dev_replace);
3935 * flush, submit all pending read and write bios, afterwards
3937 * Note that in the dev replace case, a read request causes
3938 * write requests that are submitted in the read completion
3939 * worker. Therefore in the current situation, it is required
3940 * that all write requests are flushed, so that all read and
3941 * write requests are really completed when bios_in_flight
3944 sctx->flush_all_writes = true;
3946 mutex_lock(&sctx->wr_lock);
3947 scrub_wr_submit(sctx);
3948 mutex_unlock(&sctx->wr_lock);
3950 wait_event(sctx->list_wait,
3951 atomic_read(&sctx->bios_in_flight) == 0);
3953 scrub_pause_on(fs_info);
3956 * must be called before we decrease @scrub_paused.
3957 * make sure we don't block transaction commit while
3958 * we are waiting pending workers finished.
3960 wait_event(sctx->list_wait,
3961 atomic_read(&sctx->workers_pending) == 0);
3962 sctx->flush_all_writes = false;
3964 scrub_pause_off(fs_info);
3966 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3967 dev_replace->cursor_left = dev_replace->cursor_right;
3968 dev_replace->item_needs_writeback = 1;
3969 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3972 btrfs_dec_block_group_ro(cache);
3975 * We might have prevented the cleaner kthread from deleting
3976 * this block group if it was already unused because we raced
3977 * and set it to RO mode first. So add it back to the unused
3978 * list, otherwise it might not ever be deleted unless a manual
3979 * balance is triggered or it becomes used and unused again.
3981 spin_lock(&cache->lock);
3982 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3983 btrfs_block_group_used(&cache->item) == 0) {
3984 spin_unlock(&cache->lock);
3985 spin_lock(&fs_info->unused_bgs_lock);
3986 if (list_empty(&cache->bg_list)) {
3987 btrfs_get_block_group(cache);
3988 trace_btrfs_add_unused_block_group(cache);
3989 list_add_tail(&cache->bg_list,
3990 &fs_info->unused_bgs);
3992 spin_unlock(&fs_info->unused_bgs_lock);
3994 spin_unlock(&cache->lock);
3997 btrfs_put_block_group(cache);
4000 if (is_dev_replace &&
4001 atomic64_read(&dev_replace->num_write_errors) > 0) {
4005 if (sctx->stat.malloc_errors > 0) {
4010 key.offset = found_key.offset + length;
4011 btrfs_release_path(path);
4014 btrfs_free_path(path);
4019 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4020 struct btrfs_device *scrub_dev)
4026 struct btrfs_fs_info *fs_info = sctx->fs_info;
4028 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4031 /* Seed devices of a new filesystem has their own generation. */
4032 if (scrub_dev->fs_devices != fs_info->fs_devices)
4033 gen = scrub_dev->generation;
4035 gen = fs_info->last_trans_committed;
4037 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4038 bytenr = btrfs_sb_offset(i);
4039 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4040 scrub_dev->commit_total_bytes)
4043 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4044 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4049 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4055 * get a reference count on fs_info->scrub_workers. start worker if necessary
4057 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4060 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4061 int max_active = fs_info->thread_pool_size;
4063 if (fs_info->scrub_workers_refcnt == 0) {
4064 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
4065 flags, is_dev_replace ? 1 : max_active, 4);
4066 if (!fs_info->scrub_workers)
4067 goto fail_scrub_workers;
4069 fs_info->scrub_wr_completion_workers =
4070 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4072 if (!fs_info->scrub_wr_completion_workers)
4073 goto fail_scrub_wr_completion_workers;
4075 fs_info->scrub_nocow_workers =
4076 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4077 if (!fs_info->scrub_nocow_workers)
4078 goto fail_scrub_nocow_workers;
4079 fs_info->scrub_parity_workers =
4080 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4082 if (!fs_info->scrub_parity_workers)
4083 goto fail_scrub_parity_workers;
4085 ++fs_info->scrub_workers_refcnt;
4088 fail_scrub_parity_workers:
4089 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4090 fail_scrub_nocow_workers:
4091 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4092 fail_scrub_wr_completion_workers:
4093 btrfs_destroy_workqueue(fs_info->scrub_workers);
4098 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4100 if (--fs_info->scrub_workers_refcnt == 0) {
4101 btrfs_destroy_workqueue(fs_info->scrub_workers);
4102 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4103 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4104 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4106 WARN_ON(fs_info->scrub_workers_refcnt < 0);
4109 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4110 u64 end, struct btrfs_scrub_progress *progress,
4111 int readonly, int is_dev_replace)
4113 struct scrub_ctx *sctx;
4115 struct btrfs_device *dev;
4116 struct rcu_string *name;
4118 if (btrfs_fs_closing(fs_info))
4121 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4123 * in this case scrub is unable to calculate the checksum
4124 * the way scrub is implemented. Do not handle this
4125 * situation at all because it won't ever happen.
4128 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4134 if (fs_info->sectorsize != PAGE_SIZE) {
4135 /* not supported for data w/o checksums */
4136 btrfs_err_rl(fs_info,
4137 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4138 fs_info->sectorsize, PAGE_SIZE);
4142 if (fs_info->nodesize >
4143 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4144 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4146 * would exhaust the array bounds of pagev member in
4147 * struct scrub_block
4150 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4152 SCRUB_MAX_PAGES_PER_BLOCK,
4153 fs_info->sectorsize,
4154 SCRUB_MAX_PAGES_PER_BLOCK);
4159 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4160 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4161 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4163 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4167 if (!is_dev_replace && !readonly &&
4168 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4169 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4171 name = rcu_dereference(dev->name);
4172 btrfs_err(fs_info, "scrub: device %s is not writable",
4178 mutex_lock(&fs_info->scrub_lock);
4179 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4180 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4181 mutex_unlock(&fs_info->scrub_lock);
4182 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4186 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
4187 if (dev->scrub_ctx ||
4189 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4190 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4191 mutex_unlock(&fs_info->scrub_lock);
4192 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4193 return -EINPROGRESS;
4195 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4197 ret = scrub_workers_get(fs_info, is_dev_replace);
4199 mutex_unlock(&fs_info->scrub_lock);
4200 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4204 sctx = scrub_setup_ctx(dev, is_dev_replace);
4206 mutex_unlock(&fs_info->scrub_lock);
4207 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4208 scrub_workers_put(fs_info);
4209 return PTR_ERR(sctx);
4211 sctx->readonly = readonly;
4212 dev->scrub_ctx = sctx;
4213 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4216 * checking @scrub_pause_req here, we can avoid
4217 * race between committing transaction and scrubbing.
4219 __scrub_blocked_if_needed(fs_info);
4220 atomic_inc(&fs_info->scrubs_running);
4221 mutex_unlock(&fs_info->scrub_lock);
4223 if (!is_dev_replace) {
4225 * by holding device list mutex, we can
4226 * kick off writing super in log tree sync.
4228 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4229 ret = scrub_supers(sctx, dev);
4230 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4234 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4237 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4238 atomic_dec(&fs_info->scrubs_running);
4239 wake_up(&fs_info->scrub_pause_wait);
4241 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4244 memcpy(progress, &sctx->stat, sizeof(*progress));
4246 mutex_lock(&fs_info->scrub_lock);
4247 dev->scrub_ctx = NULL;
4248 scrub_workers_put(fs_info);
4249 mutex_unlock(&fs_info->scrub_lock);
4251 scrub_put_ctx(sctx);
4256 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4258 mutex_lock(&fs_info->scrub_lock);
4259 atomic_inc(&fs_info->scrub_pause_req);
4260 while (atomic_read(&fs_info->scrubs_paused) !=
4261 atomic_read(&fs_info->scrubs_running)) {
4262 mutex_unlock(&fs_info->scrub_lock);
4263 wait_event(fs_info->scrub_pause_wait,
4264 atomic_read(&fs_info->scrubs_paused) ==
4265 atomic_read(&fs_info->scrubs_running));
4266 mutex_lock(&fs_info->scrub_lock);
4268 mutex_unlock(&fs_info->scrub_lock);
4271 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4273 atomic_dec(&fs_info->scrub_pause_req);
4274 wake_up(&fs_info->scrub_pause_wait);
4277 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4279 mutex_lock(&fs_info->scrub_lock);
4280 if (!atomic_read(&fs_info->scrubs_running)) {
4281 mutex_unlock(&fs_info->scrub_lock);
4285 atomic_inc(&fs_info->scrub_cancel_req);
4286 while (atomic_read(&fs_info->scrubs_running)) {
4287 mutex_unlock(&fs_info->scrub_lock);
4288 wait_event(fs_info->scrub_pause_wait,
4289 atomic_read(&fs_info->scrubs_running) == 0);
4290 mutex_lock(&fs_info->scrub_lock);
4292 atomic_dec(&fs_info->scrub_cancel_req);
4293 mutex_unlock(&fs_info->scrub_lock);
4298 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4299 struct btrfs_device *dev)
4301 struct scrub_ctx *sctx;
4303 mutex_lock(&fs_info->scrub_lock);
4304 sctx = dev->scrub_ctx;
4306 mutex_unlock(&fs_info->scrub_lock);
4309 atomic_inc(&sctx->cancel_req);
4310 while (dev->scrub_ctx) {
4311 mutex_unlock(&fs_info->scrub_lock);
4312 wait_event(fs_info->scrub_pause_wait,
4313 dev->scrub_ctx == NULL);
4314 mutex_lock(&fs_info->scrub_lock);
4316 mutex_unlock(&fs_info->scrub_lock);
4321 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4322 struct btrfs_scrub_progress *progress)
4324 struct btrfs_device *dev;
4325 struct scrub_ctx *sctx = NULL;
4327 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4328 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4330 sctx = dev->scrub_ctx;
4332 memcpy(progress, &sctx->stat, sizeof(*progress));
4333 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4335 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4338 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4339 u64 extent_logical, u64 extent_len,
4340 u64 *extent_physical,
4341 struct btrfs_device **extent_dev,
4342 int *extent_mirror_num)
4345 struct btrfs_bio *bbio = NULL;
4348 mapped_length = extent_len;
4349 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4350 &mapped_length, &bbio, 0);
4351 if (ret || !bbio || mapped_length < extent_len ||
4352 !bbio->stripes[0].dev->bdev) {
4353 btrfs_put_bbio(bbio);
4357 *extent_physical = bbio->stripes[0].physical;
4358 *extent_mirror_num = bbio->mirror_num;
4359 *extent_dev = bbio->stripes[0].dev;
4360 btrfs_put_bbio(bbio);
4363 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4364 int mirror_num, u64 physical_for_dev_replace)
4366 struct scrub_copy_nocow_ctx *nocow_ctx;
4367 struct btrfs_fs_info *fs_info = sctx->fs_info;
4369 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4371 spin_lock(&sctx->stat_lock);
4372 sctx->stat.malloc_errors++;
4373 spin_unlock(&sctx->stat_lock);
4377 scrub_pending_trans_workers_inc(sctx);
4379 nocow_ctx->sctx = sctx;
4380 nocow_ctx->logical = logical;
4381 nocow_ctx->len = len;
4382 nocow_ctx->mirror_num = mirror_num;
4383 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4384 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4385 copy_nocow_pages_worker, NULL, NULL);
4386 INIT_LIST_HEAD(&nocow_ctx->inodes);
4387 btrfs_queue_work(fs_info->scrub_nocow_workers,
4393 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4395 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4396 struct scrub_nocow_inode *nocow_inode;
4398 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4401 nocow_inode->inum = inum;
4402 nocow_inode->offset = offset;
4403 nocow_inode->root = root;
4404 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4408 #define COPY_COMPLETE 1
4410 static void copy_nocow_pages_worker(struct btrfs_work *work)
4412 struct scrub_copy_nocow_ctx *nocow_ctx =
4413 container_of(work, struct scrub_copy_nocow_ctx, work);
4414 struct scrub_ctx *sctx = nocow_ctx->sctx;
4415 struct btrfs_fs_info *fs_info = sctx->fs_info;
4416 struct btrfs_root *root = fs_info->extent_root;
4417 u64 logical = nocow_ctx->logical;
4418 u64 len = nocow_ctx->len;
4419 int mirror_num = nocow_ctx->mirror_num;
4420 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4422 struct btrfs_trans_handle *trans = NULL;
4423 struct btrfs_path *path;
4424 int not_written = 0;
4426 path = btrfs_alloc_path();
4428 spin_lock(&sctx->stat_lock);
4429 sctx->stat.malloc_errors++;
4430 spin_unlock(&sctx->stat_lock);
4435 trans = btrfs_join_transaction(root);
4436 if (IS_ERR(trans)) {
4441 ret = iterate_inodes_from_logical(logical, fs_info, path,
4442 record_inode_for_nocow, nocow_ctx, false);
4443 if (ret != 0 && ret != -ENOENT) {
4445 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4446 logical, physical_for_dev_replace, len, mirror_num,
4452 btrfs_end_transaction(trans);
4454 while (!list_empty(&nocow_ctx->inodes)) {
4455 struct scrub_nocow_inode *entry;
4456 entry = list_first_entry(&nocow_ctx->inodes,
4457 struct scrub_nocow_inode,
4459 list_del_init(&entry->list);
4460 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4461 entry->root, nocow_ctx);
4463 if (ret == COPY_COMPLETE) {
4471 while (!list_empty(&nocow_ctx->inodes)) {
4472 struct scrub_nocow_inode *entry;
4473 entry = list_first_entry(&nocow_ctx->inodes,
4474 struct scrub_nocow_inode,
4476 list_del_init(&entry->list);
4479 if (trans && !IS_ERR(trans))
4480 btrfs_end_transaction(trans);
4482 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4483 num_uncorrectable_read_errors);
4485 btrfs_free_path(path);
4488 scrub_pending_trans_workers_dec(sctx);
4491 static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4494 struct extent_state *cached_state = NULL;
4495 struct btrfs_ordered_extent *ordered;
4496 struct extent_io_tree *io_tree;
4497 struct extent_map *em;
4498 u64 lockstart = start, lockend = start + len - 1;
4501 io_tree = &inode->io_tree;
4503 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4504 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4506 btrfs_put_ordered_extent(ordered);
4511 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4518 * This extent does not actually cover the logical extent anymore,
4519 * move on to the next inode.
4521 if (em->block_start > logical ||
4522 em->block_start + em->block_len < logical + len ||
4523 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4524 free_extent_map(em);
4528 free_extent_map(em);
4531 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state);
4535 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4536 struct scrub_copy_nocow_ctx *nocow_ctx)
4538 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4539 struct btrfs_key key;
4540 struct inode *inode;
4542 struct btrfs_root *local_root;
4543 struct extent_io_tree *io_tree;
4544 u64 physical_for_dev_replace;
4545 u64 nocow_ctx_logical;
4546 u64 len = nocow_ctx->len;
4547 unsigned long index;
4552 key.objectid = root;
4553 key.type = BTRFS_ROOT_ITEM_KEY;
4554 key.offset = (u64)-1;
4556 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4558 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4559 if (IS_ERR(local_root)) {
4560 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4561 return PTR_ERR(local_root);
4564 key.type = BTRFS_INODE_ITEM_KEY;
4565 key.objectid = inum;
4567 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4568 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4570 return PTR_ERR(inode);
4572 /* Avoid truncate/dio/punch hole.. */
4574 inode_dio_wait(inode);
4576 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4577 io_tree = &BTRFS_I(inode)->io_tree;
4578 nocow_ctx_logical = nocow_ctx->logical;
4580 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4583 ret = ret > 0 ? 0 : ret;
4587 while (len >= PAGE_SIZE) {
4588 index = offset >> PAGE_SHIFT;
4590 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4592 btrfs_err(fs_info, "find_or_create_page() failed");
4597 if (PageUptodate(page)) {
4598 if (PageDirty(page))
4601 ClearPageError(page);
4602 err = extent_read_full_page(io_tree, page,
4604 nocow_ctx->mirror_num);
4612 * If the page has been remove from the page cache,
4613 * the data on it is meaningless, because it may be
4614 * old one, the new data may be written into the new
4615 * page in the page cache.
4617 if (page->mapping != inode->i_mapping) {
4622 if (!PageUptodate(page)) {
4628 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4631 ret = ret > 0 ? 0 : ret;
4635 err = write_page_nocow(nocow_ctx->sctx,
4636 physical_for_dev_replace, page);
4646 offset += PAGE_SIZE;
4647 physical_for_dev_replace += PAGE_SIZE;
4648 nocow_ctx_logical += PAGE_SIZE;
4651 ret = COPY_COMPLETE;
4653 inode_unlock(inode);
4658 static int write_page_nocow(struct scrub_ctx *sctx,
4659 u64 physical_for_dev_replace, struct page *page)
4662 struct btrfs_device *dev;
4664 dev = sctx->wr_tgtdev;
4668 btrfs_warn_rl(dev->fs_info,
4669 "scrub write_page_nocow(bdev == NULL) is unexpected");
4672 bio = btrfs_io_bio_alloc(1);
4673 bio->bi_iter.bi_size = 0;
4674 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4675 bio_set_dev(bio, dev->bdev);
4676 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4677 /* bio_add_page won't fail on a freshly allocated bio */
4678 bio_add_page(bio, page, PAGE_SIZE, 0);
4680 if (btrfsic_submit_bio_wait(bio)) {
4682 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);