2 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
19 #include <linux/blkdev.h>
20 #include <linux/ratelimit.h>
21 #include <linux/sched/mm.h>
25 #include "ordered-data.h"
26 #include "transaction.h"
28 #include "extent_io.h"
29 #include "dev-replace.h"
30 #include "check-integrity.h"
31 #include "rcu-string.h"
35 * This is only the first step towards a full-features scrub. It reads all
36 * extent and super block and verifies the checksums. In case a bad checksum
37 * is found or the extent cannot be read, good data will be written back if
40 * Future enhancements:
41 * - In case an unrepairable extent is encountered, track which files are
42 * affected and report them
43 * - track and record media errors, throw out bad devices
44 * - add a mode to also read unallocated space
51 * the following three values only influence the performance.
52 * The last one configures the number of parallel and outstanding I/O
53 * operations. The first two values configure an upper limit for the number
54 * of (dynamically allocated) pages that are added to a bio.
56 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
57 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
58 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
61 * the following value times PAGE_SIZE needs to be large enough to match the
62 * largest node/leaf/sector size that shall be supported.
63 * Values larger than BTRFS_STRIPE_LEN are not supported.
65 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
67 struct scrub_recover {
69 struct btrfs_bio *bbio;
74 struct scrub_block *sblock;
76 struct btrfs_device *dev;
77 struct list_head list;
78 u64 flags; /* extent flags */
82 u64 physical_for_dev_replace;
85 unsigned int mirror_num:8;
86 unsigned int have_csum:1;
87 unsigned int io_error:1;
89 u8 csum[BTRFS_CSUM_SIZE];
91 struct scrub_recover *recover;
96 struct scrub_ctx *sctx;
97 struct btrfs_device *dev;
102 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
103 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
105 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
109 struct btrfs_work work;
113 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
115 atomic_t outstanding_pages;
116 refcount_t refs; /* free mem on transition to zero */
117 struct scrub_ctx *sctx;
118 struct scrub_parity *sparity;
120 unsigned int header_error:1;
121 unsigned int checksum_error:1;
122 unsigned int no_io_error_seen:1;
123 unsigned int generation_error:1; /* also sets header_error */
125 /* The following is for the data used to check parity */
126 /* It is for the data with checksum */
127 unsigned int data_corrected:1;
129 struct btrfs_work work;
132 /* Used for the chunks with parity stripe such RAID5/6 */
133 struct scrub_parity {
134 struct scrub_ctx *sctx;
136 struct btrfs_device *scrub_dev;
148 struct list_head spages;
150 /* Work of parity check and repair */
151 struct btrfs_work work;
153 /* Mark the parity blocks which have data */
154 unsigned long *dbitmap;
157 * Mark the parity blocks which have data, but errors happen when
158 * read data or check data
160 unsigned long *ebitmap;
162 unsigned long bitmap[0];
166 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
167 struct btrfs_fs_info *fs_info;
170 atomic_t bios_in_flight;
171 atomic_t workers_pending;
172 spinlock_t list_lock;
173 wait_queue_head_t list_wait;
175 struct list_head csum_list;
178 int pages_per_rd_bio;
182 struct scrub_bio *wr_curr_bio;
183 struct mutex wr_lock;
184 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
185 atomic_t flush_all_writes;
186 struct btrfs_device *wr_tgtdev;
191 struct btrfs_scrub_progress stat;
192 spinlock_t stat_lock;
195 * Use a ref counter to avoid use-after-free issues. Scrub workers
196 * decrement bios_in_flight and workers_pending and then do a wakeup
197 * on the list_wait wait queue. We must ensure the main scrub task
198 * doesn't free the scrub context before or while the workers are
199 * doing the wakeup() call.
204 struct scrub_fixup_nodatasum {
205 struct scrub_ctx *sctx;
206 struct btrfs_device *dev;
208 struct btrfs_root *root;
209 struct btrfs_work work;
213 struct scrub_nocow_inode {
217 struct list_head list;
220 struct scrub_copy_nocow_ctx {
221 struct scrub_ctx *sctx;
225 u64 physical_for_dev_replace;
226 struct list_head inodes;
227 struct btrfs_work work;
230 struct scrub_warning {
231 struct btrfs_path *path;
232 u64 extent_item_size;
236 struct btrfs_device *dev;
239 struct full_stripe_lock {
246 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
247 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
248 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
249 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
250 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
251 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
252 struct scrub_block *sblocks_for_recheck);
253 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
254 struct scrub_block *sblock,
255 int retry_failed_mirror);
256 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
257 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
258 struct scrub_block *sblock_good);
259 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
260 struct scrub_block *sblock_good,
261 int page_num, int force_write);
262 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
263 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
265 static int scrub_checksum_data(struct scrub_block *sblock);
266 static int scrub_checksum_tree_block(struct scrub_block *sblock);
267 static int scrub_checksum_super(struct scrub_block *sblock);
268 static void scrub_block_get(struct scrub_block *sblock);
269 static void scrub_block_put(struct scrub_block *sblock);
270 static void scrub_page_get(struct scrub_page *spage);
271 static void scrub_page_put(struct scrub_page *spage);
272 static void scrub_parity_get(struct scrub_parity *sparity);
273 static void scrub_parity_put(struct scrub_parity *sparity);
274 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
275 struct scrub_page *spage);
276 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
277 u64 physical, struct btrfs_device *dev, u64 flags,
278 u64 gen, int mirror_num, u8 *csum, int force,
279 u64 physical_for_dev_replace);
280 static void scrub_bio_end_io(struct bio *bio);
281 static void scrub_bio_end_io_worker(struct btrfs_work *work);
282 static void scrub_block_complete(struct scrub_block *sblock);
283 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
284 u64 extent_logical, u64 extent_len,
285 u64 *extent_physical,
286 struct btrfs_device **extent_dev,
287 int *extent_mirror_num);
288 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
289 struct scrub_page *spage);
290 static void scrub_wr_submit(struct scrub_ctx *sctx);
291 static void scrub_wr_bio_end_io(struct bio *bio);
292 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
293 static int write_page_nocow(struct scrub_ctx *sctx,
294 u64 physical_for_dev_replace, struct page *page);
295 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
296 struct scrub_copy_nocow_ctx *ctx);
297 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
298 int mirror_num, u64 physical_for_dev_replace);
299 static void copy_nocow_pages_worker(struct btrfs_work *work);
300 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
301 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
302 static void scrub_put_ctx(struct scrub_ctx *sctx);
305 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
307 refcount_inc(&sctx->refs);
308 atomic_inc(&sctx->bios_in_flight);
311 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
313 atomic_dec(&sctx->bios_in_flight);
314 wake_up(&sctx->list_wait);
318 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
320 while (atomic_read(&fs_info->scrub_pause_req)) {
321 mutex_unlock(&fs_info->scrub_lock);
322 wait_event(fs_info->scrub_pause_wait,
323 atomic_read(&fs_info->scrub_pause_req) == 0);
324 mutex_lock(&fs_info->scrub_lock);
328 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
330 atomic_inc(&fs_info->scrubs_paused);
331 wake_up(&fs_info->scrub_pause_wait);
334 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
336 mutex_lock(&fs_info->scrub_lock);
337 __scrub_blocked_if_needed(fs_info);
338 atomic_dec(&fs_info->scrubs_paused);
339 mutex_unlock(&fs_info->scrub_lock);
341 wake_up(&fs_info->scrub_pause_wait);
344 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
346 scrub_pause_on(fs_info);
347 scrub_pause_off(fs_info);
351 * Insert new full stripe lock into full stripe locks tree
353 * Return pointer to existing or newly inserted full_stripe_lock structure if
354 * everything works well.
355 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
357 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
360 static struct full_stripe_lock *insert_full_stripe_lock(
361 struct btrfs_full_stripe_locks_tree *locks_root,
365 struct rb_node *parent = NULL;
366 struct full_stripe_lock *entry;
367 struct full_stripe_lock *ret;
369 WARN_ON(!mutex_is_locked(&locks_root->lock));
371 p = &locks_root->root.rb_node;
374 entry = rb_entry(parent, struct full_stripe_lock, node);
375 if (fstripe_logical < entry->logical) {
377 } else if (fstripe_logical > entry->logical) {
385 /* Insert new lock */
386 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
388 return ERR_PTR(-ENOMEM);
389 ret->logical = fstripe_logical;
391 mutex_init(&ret->mutex);
393 rb_link_node(&ret->node, parent, p);
394 rb_insert_color(&ret->node, &locks_root->root);
399 * Search for a full stripe lock of a block group
401 * Return pointer to existing full stripe lock if found
402 * Return NULL if not found
404 static struct full_stripe_lock *search_full_stripe_lock(
405 struct btrfs_full_stripe_locks_tree *locks_root,
408 struct rb_node *node;
409 struct full_stripe_lock *entry;
411 WARN_ON(!mutex_is_locked(&locks_root->lock));
413 node = locks_root->root.rb_node;
415 entry = rb_entry(node, struct full_stripe_lock, node);
416 if (fstripe_logical < entry->logical)
417 node = node->rb_left;
418 else if (fstripe_logical > entry->logical)
419 node = node->rb_right;
427 * Helper to get full stripe logical from a normal bytenr.
429 * Caller must ensure @cache is a RAID56 block group.
431 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
437 * Due to chunk item size limit, full stripe length should not be
438 * larger than U32_MAX. Just a sanity check here.
440 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
443 * round_down() can only handle power of 2, while RAID56 full
444 * stripe length can be 64KiB * n, so we need to manually round down.
446 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
447 cache->full_stripe_len + cache->key.objectid;
452 * Lock a full stripe to avoid concurrency of recovery and read
454 * It's only used for profiles with parities (RAID5/6), for other profiles it
457 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
458 * So caller must call unlock_full_stripe() at the same context.
460 * Return <0 if encounters error.
462 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
465 struct btrfs_block_group_cache *bg_cache;
466 struct btrfs_full_stripe_locks_tree *locks_root;
467 struct full_stripe_lock *existing;
472 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
478 /* Profiles not based on parity don't need full stripe lock */
479 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
481 locks_root = &bg_cache->full_stripe_locks_root;
483 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
485 /* Now insert the full stripe lock */
486 mutex_lock(&locks_root->lock);
487 existing = insert_full_stripe_lock(locks_root, fstripe_start);
488 mutex_unlock(&locks_root->lock);
489 if (IS_ERR(existing)) {
490 ret = PTR_ERR(existing);
493 mutex_lock(&existing->mutex);
496 btrfs_put_block_group(bg_cache);
501 * Unlock a full stripe.
503 * NOTE: Caller must ensure it's the same context calling corresponding
504 * lock_full_stripe().
506 * Return 0 if we unlock full stripe without problem.
507 * Return <0 for error
509 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
512 struct btrfs_block_group_cache *bg_cache;
513 struct btrfs_full_stripe_locks_tree *locks_root;
514 struct full_stripe_lock *fstripe_lock;
519 /* If we didn't acquire full stripe lock, no need to continue */
523 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
528 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
531 locks_root = &bg_cache->full_stripe_locks_root;
532 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
534 mutex_lock(&locks_root->lock);
535 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
536 /* Unpaired unlock_full_stripe() detected */
540 mutex_unlock(&locks_root->lock);
544 if (fstripe_lock->refs == 0) {
546 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
547 fstripe_lock->logical);
549 fstripe_lock->refs--;
552 if (fstripe_lock->refs == 0) {
553 rb_erase(&fstripe_lock->node, &locks_root->root);
556 mutex_unlock(&locks_root->lock);
558 mutex_unlock(&fstripe_lock->mutex);
562 btrfs_put_block_group(bg_cache);
567 * used for workers that require transaction commits (i.e., for the
570 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
572 struct btrfs_fs_info *fs_info = sctx->fs_info;
574 refcount_inc(&sctx->refs);
576 * increment scrubs_running to prevent cancel requests from
577 * completing as long as a worker is running. we must also
578 * increment scrubs_paused to prevent deadlocking on pause
579 * requests used for transactions commits (as the worker uses a
580 * transaction context). it is safe to regard the worker
581 * as paused for all matters practical. effectively, we only
582 * avoid cancellation requests from completing.
584 mutex_lock(&fs_info->scrub_lock);
585 atomic_inc(&fs_info->scrubs_running);
586 atomic_inc(&fs_info->scrubs_paused);
587 mutex_unlock(&fs_info->scrub_lock);
590 * check if @scrubs_running=@scrubs_paused condition
591 * inside wait_event() is not an atomic operation.
592 * which means we may inc/dec @scrub_running/paused
593 * at any time. Let's wake up @scrub_pause_wait as
594 * much as we can to let commit transaction blocked less.
596 wake_up(&fs_info->scrub_pause_wait);
598 atomic_inc(&sctx->workers_pending);
601 /* used for workers that require transaction commits */
602 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
604 struct btrfs_fs_info *fs_info = sctx->fs_info;
607 * see scrub_pending_trans_workers_inc() why we're pretending
608 * to be paused in the scrub counters
610 mutex_lock(&fs_info->scrub_lock);
611 atomic_dec(&fs_info->scrubs_running);
612 atomic_dec(&fs_info->scrubs_paused);
613 mutex_unlock(&fs_info->scrub_lock);
614 atomic_dec(&sctx->workers_pending);
615 wake_up(&fs_info->scrub_pause_wait);
616 wake_up(&sctx->list_wait);
620 static void scrub_free_csums(struct scrub_ctx *sctx)
622 while (!list_empty(&sctx->csum_list)) {
623 struct btrfs_ordered_sum *sum;
624 sum = list_first_entry(&sctx->csum_list,
625 struct btrfs_ordered_sum, list);
626 list_del(&sum->list);
631 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
638 /* this can happen when scrub is cancelled */
639 if (sctx->curr != -1) {
640 struct scrub_bio *sbio = sctx->bios[sctx->curr];
642 for (i = 0; i < sbio->page_count; i++) {
643 WARN_ON(!sbio->pagev[i]->page);
644 scrub_block_put(sbio->pagev[i]->sblock);
649 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
650 struct scrub_bio *sbio = sctx->bios[i];
657 kfree(sctx->wr_curr_bio);
658 scrub_free_csums(sctx);
662 static void scrub_put_ctx(struct scrub_ctx *sctx)
664 if (refcount_dec_and_test(&sctx->refs))
665 scrub_free_ctx(sctx);
668 static noinline_for_stack
669 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
671 struct scrub_ctx *sctx;
673 struct btrfs_fs_info *fs_info = dev->fs_info;
675 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
678 refcount_set(&sctx->refs, 1);
679 sctx->is_dev_replace = is_dev_replace;
680 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
682 sctx->fs_info = dev->fs_info;
683 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
684 struct scrub_bio *sbio;
686 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
689 sctx->bios[i] = sbio;
693 sbio->page_count = 0;
694 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
695 scrub_bio_end_io_worker, NULL, NULL);
697 if (i != SCRUB_BIOS_PER_SCTX - 1)
698 sctx->bios[i]->next_free = i + 1;
700 sctx->bios[i]->next_free = -1;
702 sctx->first_free = 0;
703 atomic_set(&sctx->bios_in_flight, 0);
704 atomic_set(&sctx->workers_pending, 0);
705 atomic_set(&sctx->cancel_req, 0);
706 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
707 INIT_LIST_HEAD(&sctx->csum_list);
709 spin_lock_init(&sctx->list_lock);
710 spin_lock_init(&sctx->stat_lock);
711 init_waitqueue_head(&sctx->list_wait);
713 WARN_ON(sctx->wr_curr_bio != NULL);
714 mutex_init(&sctx->wr_lock);
715 sctx->wr_curr_bio = NULL;
716 if (is_dev_replace) {
717 WARN_ON(!fs_info->dev_replace.tgtdev);
718 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
719 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
720 atomic_set(&sctx->flush_all_writes, 0);
726 scrub_free_ctx(sctx);
727 return ERR_PTR(-ENOMEM);
730 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
738 struct extent_buffer *eb;
739 struct btrfs_inode_item *inode_item;
740 struct scrub_warning *swarn = warn_ctx;
741 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
742 struct inode_fs_paths *ipath = NULL;
743 struct btrfs_root *local_root;
744 struct btrfs_key root_key;
745 struct btrfs_key key;
747 root_key.objectid = root;
748 root_key.type = BTRFS_ROOT_ITEM_KEY;
749 root_key.offset = (u64)-1;
750 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
751 if (IS_ERR(local_root)) {
752 ret = PTR_ERR(local_root);
757 * this makes the path point to (inum INODE_ITEM ioff)
760 key.type = BTRFS_INODE_ITEM_KEY;
763 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
765 btrfs_release_path(swarn->path);
769 eb = swarn->path->nodes[0];
770 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
771 struct btrfs_inode_item);
772 isize = btrfs_inode_size(eb, inode_item);
773 nlink = btrfs_inode_nlink(eb, inode_item);
774 btrfs_release_path(swarn->path);
777 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
778 * uses GFP_NOFS in this context, so we keep it consistent but it does
779 * not seem to be strictly necessary.
781 nofs_flag = memalloc_nofs_save();
782 ipath = init_ipath(4096, local_root, swarn->path);
783 memalloc_nofs_restore(nofs_flag);
785 ret = PTR_ERR(ipath);
789 ret = paths_from_inode(inum, ipath);
795 * we deliberately ignore the bit ipath might have been too small to
796 * hold all of the paths here
798 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
799 btrfs_warn_in_rcu(fs_info,
800 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
801 swarn->errstr, swarn->logical,
802 rcu_str_deref(swarn->dev->name),
803 (unsigned long long)swarn->sector,
805 min(isize - offset, (u64)PAGE_SIZE), nlink,
806 (char *)(unsigned long)ipath->fspath->val[i]);
812 btrfs_warn_in_rcu(fs_info,
813 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
814 swarn->errstr, swarn->logical,
815 rcu_str_deref(swarn->dev->name),
816 (unsigned long long)swarn->sector,
817 root, inum, offset, ret);
823 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
825 struct btrfs_device *dev;
826 struct btrfs_fs_info *fs_info;
827 struct btrfs_path *path;
828 struct btrfs_key found_key;
829 struct extent_buffer *eb;
830 struct btrfs_extent_item *ei;
831 struct scrub_warning swarn;
832 unsigned long ptr = 0;
840 WARN_ON(sblock->page_count < 1);
841 dev = sblock->pagev[0]->dev;
842 fs_info = sblock->sctx->fs_info;
844 path = btrfs_alloc_path();
848 swarn.sector = (sblock->pagev[0]->physical) >> 9;
849 swarn.logical = sblock->pagev[0]->logical;
850 swarn.errstr = errstr;
853 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
858 extent_item_pos = swarn.logical - found_key.objectid;
859 swarn.extent_item_size = found_key.offset;
862 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
863 item_size = btrfs_item_size_nr(eb, path->slots[0]);
865 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
867 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
868 item_size, &ref_root,
870 btrfs_warn_in_rcu(fs_info,
871 "%s at logical %llu on dev %s, sector %llu: metadata %s (level %d) in tree %llu",
872 errstr, swarn.logical,
873 rcu_str_deref(dev->name),
874 (unsigned long long)swarn.sector,
875 ref_level ? "node" : "leaf",
876 ret < 0 ? -1 : ref_level,
877 ret < 0 ? -1 : ref_root);
879 btrfs_release_path(path);
881 btrfs_release_path(path);
884 iterate_extent_inodes(fs_info, found_key.objectid,
886 scrub_print_warning_inode, &swarn);
890 btrfs_free_path(path);
893 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
895 struct page *page = NULL;
897 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
900 struct btrfs_key key;
901 struct inode *inode = NULL;
902 struct btrfs_fs_info *fs_info;
903 u64 end = offset + PAGE_SIZE - 1;
904 struct btrfs_root *local_root;
908 key.type = BTRFS_ROOT_ITEM_KEY;
909 key.offset = (u64)-1;
911 fs_info = fixup->root->fs_info;
912 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
914 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
915 if (IS_ERR(local_root)) {
916 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
917 return PTR_ERR(local_root);
920 key.type = BTRFS_INODE_ITEM_KEY;
923 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
924 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
926 return PTR_ERR(inode);
928 index = offset >> PAGE_SHIFT;
930 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
936 if (PageUptodate(page)) {
937 if (PageDirty(page)) {
939 * we need to write the data to the defect sector. the
940 * data that was in that sector is not in memory,
941 * because the page was modified. we must not write the
942 * modified page to that sector.
944 * TODO: what could be done here: wait for the delalloc
945 * runner to write out that page (might involve
946 * COW) and see whether the sector is still
947 * referenced afterwards.
949 * For the meantime, we'll treat this error
950 * incorrectable, although there is a chance that a
951 * later scrub will find the bad sector again and that
952 * there's no dirty page in memory, then.
957 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
958 fixup->logical, page,
959 offset - page_offset(page),
965 * we need to get good data first. the general readpage path
966 * will call repair_io_failure for us, we just have to make
967 * sure we read the bad mirror.
969 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
972 /* set_extent_bits should give proper error */
979 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
982 wait_on_page_locked(page);
984 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
985 end, EXTENT_DAMAGED, 0, NULL);
987 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
1000 if (ret == 0 && corrected) {
1002 * we only need to call readpage for one of the inodes belonging
1003 * to this extent. so make iterate_extent_inodes stop
1011 static void scrub_fixup_nodatasum(struct btrfs_work *work)
1013 struct btrfs_fs_info *fs_info;
1015 struct scrub_fixup_nodatasum *fixup;
1016 struct scrub_ctx *sctx;
1017 struct btrfs_trans_handle *trans = NULL;
1018 struct btrfs_path *path;
1019 int uncorrectable = 0;
1021 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1023 fs_info = fixup->root->fs_info;
1025 path = btrfs_alloc_path();
1027 spin_lock(&sctx->stat_lock);
1028 ++sctx->stat.malloc_errors;
1029 spin_unlock(&sctx->stat_lock);
1034 trans = btrfs_join_transaction(fixup->root);
1035 if (IS_ERR(trans)) {
1041 * the idea is to trigger a regular read through the standard path. we
1042 * read a page from the (failed) logical address by specifying the
1043 * corresponding copynum of the failed sector. thus, that readpage is
1045 * that is the point where on-the-fly error correction will kick in
1046 * (once it's finished) and rewrite the failed sector if a good copy
1049 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1050 scrub_fixup_readpage, fixup);
1057 spin_lock(&sctx->stat_lock);
1058 ++sctx->stat.corrected_errors;
1059 spin_unlock(&sctx->stat_lock);
1062 if (trans && !IS_ERR(trans))
1063 btrfs_end_transaction(trans);
1064 if (uncorrectable) {
1065 spin_lock(&sctx->stat_lock);
1066 ++sctx->stat.uncorrectable_errors;
1067 spin_unlock(&sctx->stat_lock);
1068 btrfs_dev_replace_stats_inc(
1069 &fs_info->dev_replace.num_uncorrectable_read_errors);
1070 btrfs_err_rl_in_rcu(fs_info,
1071 "unable to fixup (nodatasum) error at logical %llu on dev %s",
1072 fixup->logical, rcu_str_deref(fixup->dev->name));
1075 btrfs_free_path(path);
1078 scrub_pending_trans_workers_dec(sctx);
1081 static inline void scrub_get_recover(struct scrub_recover *recover)
1083 refcount_inc(&recover->refs);
1086 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1087 struct scrub_recover *recover)
1089 if (refcount_dec_and_test(&recover->refs)) {
1090 btrfs_bio_counter_dec(fs_info);
1091 btrfs_put_bbio(recover->bbio);
1097 * scrub_handle_errored_block gets called when either verification of the
1098 * pages failed or the bio failed to read, e.g. with EIO. In the latter
1099 * case, this function handles all pages in the bio, even though only one
1101 * The goal of this function is to repair the errored block by using the
1102 * contents of one of the mirrors.
1104 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1106 struct scrub_ctx *sctx = sblock_to_check->sctx;
1107 struct btrfs_device *dev;
1108 struct btrfs_fs_info *fs_info;
1111 unsigned int failed_mirror_index;
1112 unsigned int is_metadata;
1113 unsigned int have_csum;
1114 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1115 struct scrub_block *sblock_bad;
1120 bool full_stripe_locked;
1121 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1122 DEFAULT_RATELIMIT_BURST);
1124 BUG_ON(sblock_to_check->page_count < 1);
1125 fs_info = sctx->fs_info;
1126 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1128 * if we find an error in a super block, we just report it.
1129 * They will get written with the next transaction commit
1132 spin_lock(&sctx->stat_lock);
1133 ++sctx->stat.super_errors;
1134 spin_unlock(&sctx->stat_lock);
1137 length = sblock_to_check->page_count * PAGE_SIZE;
1138 logical = sblock_to_check->pagev[0]->logical;
1139 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1140 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1141 is_metadata = !(sblock_to_check->pagev[0]->flags &
1142 BTRFS_EXTENT_FLAG_DATA);
1143 have_csum = sblock_to_check->pagev[0]->have_csum;
1144 dev = sblock_to_check->pagev[0]->dev;
1147 * For RAID5/6, race can happen for a different device scrub thread.
1148 * For data corruption, Parity and Data threads will both try
1149 * to recovery the data.
1150 * Race can lead to doubly added csum error, or even unrecoverable
1153 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1155 spin_lock(&sctx->stat_lock);
1157 sctx->stat.malloc_errors++;
1158 sctx->stat.read_errors++;
1159 sctx->stat.uncorrectable_errors++;
1160 spin_unlock(&sctx->stat_lock);
1164 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
1165 sblocks_for_recheck = NULL;
1166 goto nodatasum_case;
1170 * read all mirrors one after the other. This includes to
1171 * re-read the extent or metadata block that failed (that was
1172 * the cause that this fixup code is called) another time,
1173 * page by page this time in order to know which pages
1174 * caused I/O errors and which ones are good (for all mirrors).
1175 * It is the goal to handle the situation when more than one
1176 * mirror contains I/O errors, but the errors do not
1177 * overlap, i.e. the data can be repaired by selecting the
1178 * pages from those mirrors without I/O error on the
1179 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1180 * would be that mirror #1 has an I/O error on the first page,
1181 * the second page is good, and mirror #2 has an I/O error on
1182 * the second page, but the first page is good.
1183 * Then the first page of the first mirror can be repaired by
1184 * taking the first page of the second mirror, and the
1185 * second page of the second mirror can be repaired by
1186 * copying the contents of the 2nd page of the 1st mirror.
1187 * One more note: if the pages of one mirror contain I/O
1188 * errors, the checksum cannot be verified. In order to get
1189 * the best data for repairing, the first attempt is to find
1190 * a mirror without I/O errors and with a validated checksum.
1191 * Only if this is not possible, the pages are picked from
1192 * mirrors with I/O errors without considering the checksum.
1193 * If the latter is the case, at the end, the checksum of the
1194 * repaired area is verified in order to correctly maintain
1198 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1199 sizeof(*sblocks_for_recheck), GFP_NOFS);
1200 if (!sblocks_for_recheck) {
1201 spin_lock(&sctx->stat_lock);
1202 sctx->stat.malloc_errors++;
1203 sctx->stat.read_errors++;
1204 sctx->stat.uncorrectable_errors++;
1205 spin_unlock(&sctx->stat_lock);
1206 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1210 /* setup the context, map the logical blocks and alloc the pages */
1211 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1213 spin_lock(&sctx->stat_lock);
1214 sctx->stat.read_errors++;
1215 sctx->stat.uncorrectable_errors++;
1216 spin_unlock(&sctx->stat_lock);
1217 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1220 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1221 sblock_bad = sblocks_for_recheck + failed_mirror_index;
1223 /* build and submit the bios for the failed mirror, check checksums */
1224 scrub_recheck_block(fs_info, sblock_bad, 1);
1226 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1227 sblock_bad->no_io_error_seen) {
1229 * the error disappeared after reading page by page, or
1230 * the area was part of a huge bio and other parts of the
1231 * bio caused I/O errors, or the block layer merged several
1232 * read requests into one and the error is caused by a
1233 * different bio (usually one of the two latter cases is
1236 spin_lock(&sctx->stat_lock);
1237 sctx->stat.unverified_errors++;
1238 sblock_to_check->data_corrected = 1;
1239 spin_unlock(&sctx->stat_lock);
1241 if (sctx->is_dev_replace)
1242 scrub_write_block_to_dev_replace(sblock_bad);
1246 if (!sblock_bad->no_io_error_seen) {
1247 spin_lock(&sctx->stat_lock);
1248 sctx->stat.read_errors++;
1249 spin_unlock(&sctx->stat_lock);
1250 if (__ratelimit(&_rs))
1251 scrub_print_warning("i/o error", sblock_to_check);
1252 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1253 } else if (sblock_bad->checksum_error) {
1254 spin_lock(&sctx->stat_lock);
1255 sctx->stat.csum_errors++;
1256 spin_unlock(&sctx->stat_lock);
1257 if (__ratelimit(&_rs))
1258 scrub_print_warning("checksum error", sblock_to_check);
1259 btrfs_dev_stat_inc_and_print(dev,
1260 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1261 } else if (sblock_bad->header_error) {
1262 spin_lock(&sctx->stat_lock);
1263 sctx->stat.verify_errors++;
1264 spin_unlock(&sctx->stat_lock);
1265 if (__ratelimit(&_rs))
1266 scrub_print_warning("checksum/header error",
1268 if (sblock_bad->generation_error)
1269 btrfs_dev_stat_inc_and_print(dev,
1270 BTRFS_DEV_STAT_GENERATION_ERRS);
1272 btrfs_dev_stat_inc_and_print(dev,
1273 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1276 if (sctx->readonly) {
1277 ASSERT(!sctx->is_dev_replace);
1281 if (!is_metadata && !have_csum) {
1282 struct scrub_fixup_nodatasum *fixup_nodatasum;
1284 WARN_ON(sctx->is_dev_replace);
1289 * !is_metadata and !have_csum, this means that the data
1290 * might not be COWed, that it might be modified
1291 * concurrently. The general strategy to work on the
1292 * commit root does not help in the case when COW is not
1295 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1296 if (!fixup_nodatasum)
1297 goto did_not_correct_error;
1298 fixup_nodatasum->sctx = sctx;
1299 fixup_nodatasum->dev = dev;
1300 fixup_nodatasum->logical = logical;
1301 fixup_nodatasum->root = fs_info->extent_root;
1302 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1303 scrub_pending_trans_workers_inc(sctx);
1304 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1305 scrub_fixup_nodatasum, NULL, NULL);
1306 btrfs_queue_work(fs_info->scrub_workers,
1307 &fixup_nodatasum->work);
1312 * now build and submit the bios for the other mirrors, check
1314 * First try to pick the mirror which is completely without I/O
1315 * errors and also does not have a checksum error.
1316 * If one is found, and if a checksum is present, the full block
1317 * that is known to contain an error is rewritten. Afterwards
1318 * the block is known to be corrected.
1319 * If a mirror is found which is completely correct, and no
1320 * checksum is present, only those pages are rewritten that had
1321 * an I/O error in the block to be repaired, since it cannot be
1322 * determined, which copy of the other pages is better (and it
1323 * could happen otherwise that a correct page would be
1324 * overwritten by a bad one).
1326 for (mirror_index = 0;
1327 mirror_index < BTRFS_MAX_MIRRORS &&
1328 sblocks_for_recheck[mirror_index].page_count > 0;
1330 struct scrub_block *sblock_other;
1332 if (mirror_index == failed_mirror_index)
1334 sblock_other = sblocks_for_recheck + mirror_index;
1336 /* build and submit the bios, check checksums */
1337 scrub_recheck_block(fs_info, sblock_other, 0);
1339 if (!sblock_other->header_error &&
1340 !sblock_other->checksum_error &&
1341 sblock_other->no_io_error_seen) {
1342 if (sctx->is_dev_replace) {
1343 scrub_write_block_to_dev_replace(sblock_other);
1344 goto corrected_error;
1346 ret = scrub_repair_block_from_good_copy(
1347 sblock_bad, sblock_other);
1349 goto corrected_error;
1354 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1355 goto did_not_correct_error;
1358 * In case of I/O errors in the area that is supposed to be
1359 * repaired, continue by picking good copies of those pages.
1360 * Select the good pages from mirrors to rewrite bad pages from
1361 * the area to fix. Afterwards verify the checksum of the block
1362 * that is supposed to be repaired. This verification step is
1363 * only done for the purpose of statistic counting and for the
1364 * final scrub report, whether errors remain.
1365 * A perfect algorithm could make use of the checksum and try
1366 * all possible combinations of pages from the different mirrors
1367 * until the checksum verification succeeds. For example, when
1368 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1369 * of mirror #2 is readable but the final checksum test fails,
1370 * then the 2nd page of mirror #3 could be tried, whether now
1371 * the final checksum succeeds. But this would be a rare
1372 * exception and is therefore not implemented. At least it is
1373 * avoided that the good copy is overwritten.
1374 * A more useful improvement would be to pick the sectors
1375 * without I/O error based on sector sizes (512 bytes on legacy
1376 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1377 * mirror could be repaired by taking 512 byte of a different
1378 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1379 * area are unreadable.
1382 for (page_num = 0; page_num < sblock_bad->page_count;
1384 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1385 struct scrub_block *sblock_other = NULL;
1387 /* skip no-io-error page in scrub */
1388 if (!page_bad->io_error && !sctx->is_dev_replace)
1391 /* try to find no-io-error page in mirrors */
1392 if (page_bad->io_error) {
1393 for (mirror_index = 0;
1394 mirror_index < BTRFS_MAX_MIRRORS &&
1395 sblocks_for_recheck[mirror_index].page_count > 0;
1397 if (!sblocks_for_recheck[mirror_index].
1398 pagev[page_num]->io_error) {
1399 sblock_other = sblocks_for_recheck +
1408 if (sctx->is_dev_replace) {
1410 * did not find a mirror to fetch the page
1411 * from. scrub_write_page_to_dev_replace()
1412 * handles this case (page->io_error), by
1413 * filling the block with zeros before
1414 * submitting the write request
1417 sblock_other = sblock_bad;
1419 if (scrub_write_page_to_dev_replace(sblock_other,
1421 btrfs_dev_replace_stats_inc(
1422 &fs_info->dev_replace.num_write_errors);
1425 } else if (sblock_other) {
1426 ret = scrub_repair_page_from_good_copy(sblock_bad,
1430 page_bad->io_error = 0;
1436 if (success && !sctx->is_dev_replace) {
1437 if (is_metadata || have_csum) {
1439 * need to verify the checksum now that all
1440 * sectors on disk are repaired (the write
1441 * request for data to be repaired is on its way).
1442 * Just be lazy and use scrub_recheck_block()
1443 * which re-reads the data before the checksum
1444 * is verified, but most likely the data comes out
1445 * of the page cache.
1447 scrub_recheck_block(fs_info, sblock_bad, 1);
1448 if (!sblock_bad->header_error &&
1449 !sblock_bad->checksum_error &&
1450 sblock_bad->no_io_error_seen)
1451 goto corrected_error;
1453 goto did_not_correct_error;
1456 spin_lock(&sctx->stat_lock);
1457 sctx->stat.corrected_errors++;
1458 sblock_to_check->data_corrected = 1;
1459 spin_unlock(&sctx->stat_lock);
1460 btrfs_err_rl_in_rcu(fs_info,
1461 "fixed up error at logical %llu on dev %s",
1462 logical, rcu_str_deref(dev->name));
1465 did_not_correct_error:
1466 spin_lock(&sctx->stat_lock);
1467 sctx->stat.uncorrectable_errors++;
1468 spin_unlock(&sctx->stat_lock);
1469 btrfs_err_rl_in_rcu(fs_info,
1470 "unable to fixup (regular) error at logical %llu on dev %s",
1471 logical, rcu_str_deref(dev->name));
1475 if (sblocks_for_recheck) {
1476 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1478 struct scrub_block *sblock = sblocks_for_recheck +
1480 struct scrub_recover *recover;
1483 for (page_index = 0; page_index < sblock->page_count;
1485 sblock->pagev[page_index]->sblock = NULL;
1486 recover = sblock->pagev[page_index]->recover;
1488 scrub_put_recover(fs_info, recover);
1489 sblock->pagev[page_index]->recover =
1492 scrub_page_put(sblock->pagev[page_index]);
1495 kfree(sblocks_for_recheck);
1498 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1504 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1506 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1508 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1511 return (int)bbio->num_stripes;
1514 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1517 int nstripes, int mirror,
1523 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1525 for (i = 0; i < nstripes; i++) {
1526 if (raid_map[i] == RAID6_Q_STRIPE ||
1527 raid_map[i] == RAID5_P_STRIPE)
1530 if (logical >= raid_map[i] &&
1531 logical < raid_map[i] + mapped_length)
1536 *stripe_offset = logical - raid_map[i];
1538 /* The other RAID type */
1539 *stripe_index = mirror;
1544 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1545 struct scrub_block *sblocks_for_recheck)
1547 struct scrub_ctx *sctx = original_sblock->sctx;
1548 struct btrfs_fs_info *fs_info = sctx->fs_info;
1549 u64 length = original_sblock->page_count * PAGE_SIZE;
1550 u64 logical = original_sblock->pagev[0]->logical;
1551 u64 generation = original_sblock->pagev[0]->generation;
1552 u64 flags = original_sblock->pagev[0]->flags;
1553 u64 have_csum = original_sblock->pagev[0]->have_csum;
1554 struct scrub_recover *recover;
1555 struct btrfs_bio *bbio;
1566 * note: the two members refs and outstanding_pages
1567 * are not used (and not set) in the blocks that are used for
1568 * the recheck procedure
1571 while (length > 0) {
1572 sublen = min_t(u64, length, PAGE_SIZE);
1573 mapped_length = sublen;
1577 * with a length of PAGE_SIZE, each returned stripe
1578 * represents one mirror
1580 btrfs_bio_counter_inc_blocked(fs_info);
1581 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1582 logical, &mapped_length, &bbio);
1583 if (ret || !bbio || mapped_length < sublen) {
1584 btrfs_put_bbio(bbio);
1585 btrfs_bio_counter_dec(fs_info);
1589 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1591 btrfs_put_bbio(bbio);
1592 btrfs_bio_counter_dec(fs_info);
1596 refcount_set(&recover->refs, 1);
1597 recover->bbio = bbio;
1598 recover->map_length = mapped_length;
1600 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1602 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1604 for (mirror_index = 0; mirror_index < nmirrors;
1606 struct scrub_block *sblock;
1607 struct scrub_page *page;
1609 sblock = sblocks_for_recheck + mirror_index;
1610 sblock->sctx = sctx;
1612 page = kzalloc(sizeof(*page), GFP_NOFS);
1615 spin_lock(&sctx->stat_lock);
1616 sctx->stat.malloc_errors++;
1617 spin_unlock(&sctx->stat_lock);
1618 scrub_put_recover(fs_info, recover);
1621 scrub_page_get(page);
1622 sblock->pagev[page_index] = page;
1623 page->sblock = sblock;
1624 page->flags = flags;
1625 page->generation = generation;
1626 page->logical = logical;
1627 page->have_csum = have_csum;
1630 original_sblock->pagev[0]->csum,
1633 scrub_stripe_index_and_offset(logical,
1642 page->physical = bbio->stripes[stripe_index].physical +
1644 page->dev = bbio->stripes[stripe_index].dev;
1646 BUG_ON(page_index >= original_sblock->page_count);
1647 page->physical_for_dev_replace =
1648 original_sblock->pagev[page_index]->
1649 physical_for_dev_replace;
1650 /* for missing devices, dev->bdev is NULL */
1651 page->mirror_num = mirror_index + 1;
1652 sblock->page_count++;
1653 page->page = alloc_page(GFP_NOFS);
1657 scrub_get_recover(recover);
1658 page->recover = recover;
1660 scrub_put_recover(fs_info, recover);
1669 struct scrub_bio_ret {
1670 struct completion event;
1671 blk_status_t status;
1674 static void scrub_bio_wait_endio(struct bio *bio)
1676 struct scrub_bio_ret *ret = bio->bi_private;
1678 ret->status = bio->bi_status;
1679 complete(&ret->event);
1682 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1684 return page->recover &&
1685 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1688 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1690 struct scrub_page *page)
1692 struct scrub_bio_ret done;
1695 init_completion(&done.event);
1697 bio->bi_iter.bi_sector = page->logical >> 9;
1698 bio->bi_private = &done;
1699 bio->bi_end_io = scrub_bio_wait_endio;
1701 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1702 page->recover->map_length,
1703 page->mirror_num, 0);
1707 wait_for_completion(&done.event);
1715 * this function will check the on disk data for checksum errors, header
1716 * errors and read I/O errors. If any I/O errors happen, the exact pages
1717 * which are errored are marked as being bad. The goal is to enable scrub
1718 * to take those pages that are not errored from all the mirrors so that
1719 * the pages that are errored in the just handled mirror can be repaired.
1721 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1722 struct scrub_block *sblock,
1723 int retry_failed_mirror)
1727 sblock->no_io_error_seen = 1;
1729 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1731 struct scrub_page *page = sblock->pagev[page_num];
1733 if (page->dev->bdev == NULL) {
1735 sblock->no_io_error_seen = 0;
1739 WARN_ON(!page->page);
1740 bio = btrfs_io_bio_alloc(1);
1741 bio->bi_bdev = page->dev->bdev;
1743 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1744 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1745 if (scrub_submit_raid56_bio_wait(fs_info, bio, page)) {
1747 sblock->no_io_error_seen = 0;
1750 bio->bi_iter.bi_sector = page->physical >> 9;
1751 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1753 if (btrfsic_submit_bio_wait(bio)) {
1755 sblock->no_io_error_seen = 0;
1762 if (sblock->no_io_error_seen)
1763 scrub_recheck_block_checksum(sblock);
1766 static inline int scrub_check_fsid(u8 fsid[],
1767 struct scrub_page *spage)
1769 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1772 ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1776 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1778 sblock->header_error = 0;
1779 sblock->checksum_error = 0;
1780 sblock->generation_error = 0;
1782 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1783 scrub_checksum_data(sblock);
1785 scrub_checksum_tree_block(sblock);
1788 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1789 struct scrub_block *sblock_good)
1794 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1797 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1807 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1808 struct scrub_block *sblock_good,
1809 int page_num, int force_write)
1811 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1812 struct scrub_page *page_good = sblock_good->pagev[page_num];
1813 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1815 BUG_ON(page_bad->page == NULL);
1816 BUG_ON(page_good->page == NULL);
1817 if (force_write || sblock_bad->header_error ||
1818 sblock_bad->checksum_error || page_bad->io_error) {
1822 if (!page_bad->dev->bdev) {
1823 btrfs_warn_rl(fs_info,
1824 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1828 bio = btrfs_io_bio_alloc(1);
1829 bio->bi_bdev = page_bad->dev->bdev;
1830 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1831 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1833 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1834 if (PAGE_SIZE != ret) {
1839 if (btrfsic_submit_bio_wait(bio)) {
1840 btrfs_dev_stat_inc_and_print(page_bad->dev,
1841 BTRFS_DEV_STAT_WRITE_ERRS);
1842 btrfs_dev_replace_stats_inc(
1843 &fs_info->dev_replace.num_write_errors);
1853 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1855 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1859 * This block is used for the check of the parity on the source device,
1860 * so the data needn't be written into the destination device.
1862 if (sblock->sparity)
1865 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1868 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1870 btrfs_dev_replace_stats_inc(
1871 &fs_info->dev_replace.num_write_errors);
1875 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1878 struct scrub_page *spage = sblock->pagev[page_num];
1880 BUG_ON(spage->page == NULL);
1881 if (spage->io_error) {
1882 void *mapped_buffer = kmap_atomic(spage->page);
1884 clear_page(mapped_buffer);
1885 flush_dcache_page(spage->page);
1886 kunmap_atomic(mapped_buffer);
1888 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1891 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1892 struct scrub_page *spage)
1894 struct scrub_bio *sbio;
1897 mutex_lock(&sctx->wr_lock);
1899 if (!sctx->wr_curr_bio) {
1900 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1902 if (!sctx->wr_curr_bio) {
1903 mutex_unlock(&sctx->wr_lock);
1906 sctx->wr_curr_bio->sctx = sctx;
1907 sctx->wr_curr_bio->page_count = 0;
1909 sbio = sctx->wr_curr_bio;
1910 if (sbio->page_count == 0) {
1913 sbio->physical = spage->physical_for_dev_replace;
1914 sbio->logical = spage->logical;
1915 sbio->dev = sctx->wr_tgtdev;
1918 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1922 bio->bi_private = sbio;
1923 bio->bi_end_io = scrub_wr_bio_end_io;
1924 bio->bi_bdev = sbio->dev->bdev;
1925 bio->bi_iter.bi_sector = sbio->physical >> 9;
1926 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1928 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1929 spage->physical_for_dev_replace ||
1930 sbio->logical + sbio->page_count * PAGE_SIZE !=
1932 scrub_wr_submit(sctx);
1936 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1937 if (ret != PAGE_SIZE) {
1938 if (sbio->page_count < 1) {
1941 mutex_unlock(&sctx->wr_lock);
1944 scrub_wr_submit(sctx);
1948 sbio->pagev[sbio->page_count] = spage;
1949 scrub_page_get(spage);
1951 if (sbio->page_count == sctx->pages_per_wr_bio)
1952 scrub_wr_submit(sctx);
1953 mutex_unlock(&sctx->wr_lock);
1958 static void scrub_wr_submit(struct scrub_ctx *sctx)
1960 struct scrub_bio *sbio;
1962 if (!sctx->wr_curr_bio)
1965 sbio = sctx->wr_curr_bio;
1966 sctx->wr_curr_bio = NULL;
1967 WARN_ON(!sbio->bio->bi_bdev);
1968 scrub_pending_bio_inc(sctx);
1969 /* process all writes in a single worker thread. Then the block layer
1970 * orders the requests before sending them to the driver which
1971 * doubled the write performance on spinning disks when measured
1973 btrfsic_submit_bio(sbio->bio);
1976 static void scrub_wr_bio_end_io(struct bio *bio)
1978 struct scrub_bio *sbio = bio->bi_private;
1979 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1981 sbio->status = bio->bi_status;
1984 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1985 scrub_wr_bio_end_io_worker, NULL, NULL);
1986 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1989 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1991 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1992 struct scrub_ctx *sctx = sbio->sctx;
1995 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1997 struct btrfs_dev_replace *dev_replace =
1998 &sbio->sctx->fs_info->dev_replace;
2000 for (i = 0; i < sbio->page_count; i++) {
2001 struct scrub_page *spage = sbio->pagev[i];
2003 spage->io_error = 1;
2004 btrfs_dev_replace_stats_inc(&dev_replace->
2009 for (i = 0; i < sbio->page_count; i++)
2010 scrub_page_put(sbio->pagev[i]);
2014 scrub_pending_bio_dec(sctx);
2017 static int scrub_checksum(struct scrub_block *sblock)
2023 * No need to initialize these stats currently,
2024 * because this function only use return value
2025 * instead of these stats value.
2030 sblock->header_error = 0;
2031 sblock->generation_error = 0;
2032 sblock->checksum_error = 0;
2034 WARN_ON(sblock->page_count < 1);
2035 flags = sblock->pagev[0]->flags;
2037 if (flags & BTRFS_EXTENT_FLAG_DATA)
2038 ret = scrub_checksum_data(sblock);
2039 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2040 ret = scrub_checksum_tree_block(sblock);
2041 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2042 (void)scrub_checksum_super(sblock);
2046 scrub_handle_errored_block(sblock);
2051 static int scrub_checksum_data(struct scrub_block *sblock)
2053 struct scrub_ctx *sctx = sblock->sctx;
2054 u8 csum[BTRFS_CSUM_SIZE];
2062 BUG_ON(sblock->page_count < 1);
2063 if (!sblock->pagev[0]->have_csum)
2066 on_disk_csum = sblock->pagev[0]->csum;
2067 page = sblock->pagev[0]->page;
2068 buffer = kmap_atomic(page);
2070 len = sctx->fs_info->sectorsize;
2073 u64 l = min_t(u64, len, PAGE_SIZE);
2075 crc = btrfs_csum_data(buffer, crc, l);
2076 kunmap_atomic(buffer);
2081 BUG_ON(index >= sblock->page_count);
2082 BUG_ON(!sblock->pagev[index]->page);
2083 page = sblock->pagev[index]->page;
2084 buffer = kmap_atomic(page);
2087 btrfs_csum_final(crc, csum);
2088 if (memcmp(csum, on_disk_csum, sctx->csum_size))
2089 sblock->checksum_error = 1;
2091 return sblock->checksum_error;
2094 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2096 struct scrub_ctx *sctx = sblock->sctx;
2097 struct btrfs_header *h;
2098 struct btrfs_fs_info *fs_info = sctx->fs_info;
2099 u8 calculated_csum[BTRFS_CSUM_SIZE];
2100 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2102 void *mapped_buffer;
2109 BUG_ON(sblock->page_count < 1);
2110 page = sblock->pagev[0]->page;
2111 mapped_buffer = kmap_atomic(page);
2112 h = (struct btrfs_header *)mapped_buffer;
2113 memcpy(on_disk_csum, h->csum, sctx->csum_size);
2116 * we don't use the getter functions here, as we
2117 * a) don't have an extent buffer and
2118 * b) the page is already kmapped
2120 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2121 sblock->header_error = 1;
2123 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2124 sblock->header_error = 1;
2125 sblock->generation_error = 1;
2128 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2129 sblock->header_error = 1;
2131 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2133 sblock->header_error = 1;
2135 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2136 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2137 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2140 u64 l = min_t(u64, len, mapped_size);
2142 crc = btrfs_csum_data(p, crc, l);
2143 kunmap_atomic(mapped_buffer);
2148 BUG_ON(index >= sblock->page_count);
2149 BUG_ON(!sblock->pagev[index]->page);
2150 page = sblock->pagev[index]->page;
2151 mapped_buffer = kmap_atomic(page);
2152 mapped_size = PAGE_SIZE;
2156 btrfs_csum_final(crc, calculated_csum);
2157 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2158 sblock->checksum_error = 1;
2160 return sblock->header_error || sblock->checksum_error;
2163 static int scrub_checksum_super(struct scrub_block *sblock)
2165 struct btrfs_super_block *s;
2166 struct scrub_ctx *sctx = sblock->sctx;
2167 u8 calculated_csum[BTRFS_CSUM_SIZE];
2168 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2170 void *mapped_buffer;
2179 BUG_ON(sblock->page_count < 1);
2180 page = sblock->pagev[0]->page;
2181 mapped_buffer = kmap_atomic(page);
2182 s = (struct btrfs_super_block *)mapped_buffer;
2183 memcpy(on_disk_csum, s->csum, sctx->csum_size);
2185 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2188 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2191 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2194 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2195 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2196 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2199 u64 l = min_t(u64, len, mapped_size);
2201 crc = btrfs_csum_data(p, crc, l);
2202 kunmap_atomic(mapped_buffer);
2207 BUG_ON(index >= sblock->page_count);
2208 BUG_ON(!sblock->pagev[index]->page);
2209 page = sblock->pagev[index]->page;
2210 mapped_buffer = kmap_atomic(page);
2211 mapped_size = PAGE_SIZE;
2215 btrfs_csum_final(crc, calculated_csum);
2216 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2219 if (fail_cor + fail_gen) {
2221 * if we find an error in a super block, we just report it.
2222 * They will get written with the next transaction commit
2225 spin_lock(&sctx->stat_lock);
2226 ++sctx->stat.super_errors;
2227 spin_unlock(&sctx->stat_lock);
2229 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2230 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2232 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2233 BTRFS_DEV_STAT_GENERATION_ERRS);
2236 return fail_cor + fail_gen;
2239 static void scrub_block_get(struct scrub_block *sblock)
2241 refcount_inc(&sblock->refs);
2244 static void scrub_block_put(struct scrub_block *sblock)
2246 if (refcount_dec_and_test(&sblock->refs)) {
2249 if (sblock->sparity)
2250 scrub_parity_put(sblock->sparity);
2252 for (i = 0; i < sblock->page_count; i++)
2253 scrub_page_put(sblock->pagev[i]);
2258 static void scrub_page_get(struct scrub_page *spage)
2260 atomic_inc(&spage->refs);
2263 static void scrub_page_put(struct scrub_page *spage)
2265 if (atomic_dec_and_test(&spage->refs)) {
2267 __free_page(spage->page);
2272 static void scrub_submit(struct scrub_ctx *sctx)
2274 struct scrub_bio *sbio;
2276 if (sctx->curr == -1)
2279 sbio = sctx->bios[sctx->curr];
2281 scrub_pending_bio_inc(sctx);
2282 btrfsic_submit_bio(sbio->bio);
2285 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2286 struct scrub_page *spage)
2288 struct scrub_block *sblock = spage->sblock;
2289 struct scrub_bio *sbio;
2294 * grab a fresh bio or wait for one to become available
2296 while (sctx->curr == -1) {
2297 spin_lock(&sctx->list_lock);
2298 sctx->curr = sctx->first_free;
2299 if (sctx->curr != -1) {
2300 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2301 sctx->bios[sctx->curr]->next_free = -1;
2302 sctx->bios[sctx->curr]->page_count = 0;
2303 spin_unlock(&sctx->list_lock);
2305 spin_unlock(&sctx->list_lock);
2306 wait_event(sctx->list_wait, sctx->first_free != -1);
2309 sbio = sctx->bios[sctx->curr];
2310 if (sbio->page_count == 0) {
2313 sbio->physical = spage->physical;
2314 sbio->logical = spage->logical;
2315 sbio->dev = spage->dev;
2318 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2322 bio->bi_private = sbio;
2323 bio->bi_end_io = scrub_bio_end_io;
2324 bio->bi_bdev = sbio->dev->bdev;
2325 bio->bi_iter.bi_sector = sbio->physical >> 9;
2326 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2328 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2330 sbio->logical + sbio->page_count * PAGE_SIZE !=
2332 sbio->dev != spage->dev) {
2337 sbio->pagev[sbio->page_count] = spage;
2338 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2339 if (ret != PAGE_SIZE) {
2340 if (sbio->page_count < 1) {
2349 scrub_block_get(sblock); /* one for the page added to the bio */
2350 atomic_inc(&sblock->outstanding_pages);
2352 if (sbio->page_count == sctx->pages_per_rd_bio)
2358 static void scrub_missing_raid56_end_io(struct bio *bio)
2360 struct scrub_block *sblock = bio->bi_private;
2361 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2364 sblock->no_io_error_seen = 0;
2368 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2371 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2373 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2374 struct scrub_ctx *sctx = sblock->sctx;
2375 struct btrfs_fs_info *fs_info = sctx->fs_info;
2377 struct btrfs_device *dev;
2379 logical = sblock->pagev[0]->logical;
2380 dev = sblock->pagev[0]->dev;
2382 if (sblock->no_io_error_seen)
2383 scrub_recheck_block_checksum(sblock);
2385 if (!sblock->no_io_error_seen) {
2386 spin_lock(&sctx->stat_lock);
2387 sctx->stat.read_errors++;
2388 spin_unlock(&sctx->stat_lock);
2389 btrfs_err_rl_in_rcu(fs_info,
2390 "IO error rebuilding logical %llu for dev %s",
2391 logical, rcu_str_deref(dev->name));
2392 } else if (sblock->header_error || sblock->checksum_error) {
2393 spin_lock(&sctx->stat_lock);
2394 sctx->stat.uncorrectable_errors++;
2395 spin_unlock(&sctx->stat_lock);
2396 btrfs_err_rl_in_rcu(fs_info,
2397 "failed to rebuild valid logical %llu for dev %s",
2398 logical, rcu_str_deref(dev->name));
2400 scrub_write_block_to_dev_replace(sblock);
2403 scrub_block_put(sblock);
2405 if (sctx->is_dev_replace &&
2406 atomic_read(&sctx->flush_all_writes)) {
2407 mutex_lock(&sctx->wr_lock);
2408 scrub_wr_submit(sctx);
2409 mutex_unlock(&sctx->wr_lock);
2412 scrub_pending_bio_dec(sctx);
2415 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2417 struct scrub_ctx *sctx = sblock->sctx;
2418 struct btrfs_fs_info *fs_info = sctx->fs_info;
2419 u64 length = sblock->page_count * PAGE_SIZE;
2420 u64 logical = sblock->pagev[0]->logical;
2421 struct btrfs_bio *bbio = NULL;
2423 struct btrfs_raid_bio *rbio;
2427 btrfs_bio_counter_inc_blocked(fs_info);
2428 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2430 if (ret || !bbio || !bbio->raid_map)
2433 if (WARN_ON(!sctx->is_dev_replace ||
2434 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2436 * We shouldn't be scrubbing a missing device. Even for dev
2437 * replace, we should only get here for RAID 5/6. We either
2438 * managed to mount something with no mirrors remaining or
2439 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2444 bio = btrfs_io_bio_alloc(0);
2445 bio->bi_iter.bi_sector = logical >> 9;
2446 bio->bi_private = sblock;
2447 bio->bi_end_io = scrub_missing_raid56_end_io;
2449 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2453 for (i = 0; i < sblock->page_count; i++) {
2454 struct scrub_page *spage = sblock->pagev[i];
2456 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2459 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2460 scrub_missing_raid56_worker, NULL, NULL);
2461 scrub_block_get(sblock);
2462 scrub_pending_bio_inc(sctx);
2463 raid56_submit_missing_rbio(rbio);
2469 btrfs_bio_counter_dec(fs_info);
2470 btrfs_put_bbio(bbio);
2471 spin_lock(&sctx->stat_lock);
2472 sctx->stat.malloc_errors++;
2473 spin_unlock(&sctx->stat_lock);
2476 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2477 u64 physical, struct btrfs_device *dev, u64 flags,
2478 u64 gen, int mirror_num, u8 *csum, int force,
2479 u64 physical_for_dev_replace)
2481 struct scrub_block *sblock;
2484 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2486 spin_lock(&sctx->stat_lock);
2487 sctx->stat.malloc_errors++;
2488 spin_unlock(&sctx->stat_lock);
2492 /* one ref inside this function, plus one for each page added to
2494 refcount_set(&sblock->refs, 1);
2495 sblock->sctx = sctx;
2496 sblock->no_io_error_seen = 1;
2498 for (index = 0; len > 0; index++) {
2499 struct scrub_page *spage;
2500 u64 l = min_t(u64, len, PAGE_SIZE);
2502 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2505 spin_lock(&sctx->stat_lock);
2506 sctx->stat.malloc_errors++;
2507 spin_unlock(&sctx->stat_lock);
2508 scrub_block_put(sblock);
2511 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2512 scrub_page_get(spage);
2513 sblock->pagev[index] = spage;
2514 spage->sblock = sblock;
2516 spage->flags = flags;
2517 spage->generation = gen;
2518 spage->logical = logical;
2519 spage->physical = physical;
2520 spage->physical_for_dev_replace = physical_for_dev_replace;
2521 spage->mirror_num = mirror_num;
2523 spage->have_csum = 1;
2524 memcpy(spage->csum, csum, sctx->csum_size);
2526 spage->have_csum = 0;
2528 sblock->page_count++;
2529 spage->page = alloc_page(GFP_KERNEL);
2535 physical_for_dev_replace += l;
2538 WARN_ON(sblock->page_count == 0);
2541 * This case should only be hit for RAID 5/6 device replace. See
2542 * the comment in scrub_missing_raid56_pages() for details.
2544 scrub_missing_raid56_pages(sblock);
2546 for (index = 0; index < sblock->page_count; index++) {
2547 struct scrub_page *spage = sblock->pagev[index];
2550 ret = scrub_add_page_to_rd_bio(sctx, spage);
2552 scrub_block_put(sblock);
2561 /* last one frees, either here or in bio completion for last page */
2562 scrub_block_put(sblock);
2566 static void scrub_bio_end_io(struct bio *bio)
2568 struct scrub_bio *sbio = bio->bi_private;
2569 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2571 sbio->status = bio->bi_status;
2574 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2577 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2579 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2580 struct scrub_ctx *sctx = sbio->sctx;
2583 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2585 for (i = 0; i < sbio->page_count; i++) {
2586 struct scrub_page *spage = sbio->pagev[i];
2588 spage->io_error = 1;
2589 spage->sblock->no_io_error_seen = 0;
2593 /* now complete the scrub_block items that have all pages completed */
2594 for (i = 0; i < sbio->page_count; i++) {
2595 struct scrub_page *spage = sbio->pagev[i];
2596 struct scrub_block *sblock = spage->sblock;
2598 if (atomic_dec_and_test(&sblock->outstanding_pages))
2599 scrub_block_complete(sblock);
2600 scrub_block_put(sblock);
2605 spin_lock(&sctx->list_lock);
2606 sbio->next_free = sctx->first_free;
2607 sctx->first_free = sbio->index;
2608 spin_unlock(&sctx->list_lock);
2610 if (sctx->is_dev_replace &&
2611 atomic_read(&sctx->flush_all_writes)) {
2612 mutex_lock(&sctx->wr_lock);
2613 scrub_wr_submit(sctx);
2614 mutex_unlock(&sctx->wr_lock);
2617 scrub_pending_bio_dec(sctx);
2620 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2621 unsigned long *bitmap,
2626 int sectorsize = sparity->sctx->fs_info->sectorsize;
2628 if (len >= sparity->stripe_len) {
2629 bitmap_set(bitmap, 0, sparity->nsectors);
2633 start -= sparity->logic_start;
2634 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2635 offset = div_u64(offset, sectorsize);
2636 nsectors = (int)len / sectorsize;
2638 if (offset + nsectors <= sparity->nsectors) {
2639 bitmap_set(bitmap, offset, nsectors);
2643 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2644 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2647 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2650 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2653 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2656 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2659 static void scrub_block_complete(struct scrub_block *sblock)
2663 if (!sblock->no_io_error_seen) {
2665 scrub_handle_errored_block(sblock);
2668 * if has checksum error, write via repair mechanism in
2669 * dev replace case, otherwise write here in dev replace
2672 corrupted = scrub_checksum(sblock);
2673 if (!corrupted && sblock->sctx->is_dev_replace)
2674 scrub_write_block_to_dev_replace(sblock);
2677 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2678 u64 start = sblock->pagev[0]->logical;
2679 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2682 scrub_parity_mark_sectors_error(sblock->sparity,
2683 start, end - start);
2687 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2689 struct btrfs_ordered_sum *sum = NULL;
2690 unsigned long index;
2691 unsigned long num_sectors;
2693 while (!list_empty(&sctx->csum_list)) {
2694 sum = list_first_entry(&sctx->csum_list,
2695 struct btrfs_ordered_sum, list);
2696 if (sum->bytenr > logical)
2698 if (sum->bytenr + sum->len > logical)
2701 ++sctx->stat.csum_discards;
2702 list_del(&sum->list);
2709 index = ((u32)(logical - sum->bytenr)) / sctx->fs_info->sectorsize;
2710 num_sectors = sum->len / sctx->fs_info->sectorsize;
2711 memcpy(csum, sum->sums + index, sctx->csum_size);
2712 if (index == num_sectors - 1) {
2713 list_del(&sum->list);
2719 /* scrub extent tries to collect up to 64 kB for each bio */
2720 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2721 u64 physical, struct btrfs_device *dev, u64 flags,
2722 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2725 u8 csum[BTRFS_CSUM_SIZE];
2728 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2729 blocksize = sctx->fs_info->sectorsize;
2730 spin_lock(&sctx->stat_lock);
2731 sctx->stat.data_extents_scrubbed++;
2732 sctx->stat.data_bytes_scrubbed += len;
2733 spin_unlock(&sctx->stat_lock);
2734 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2735 blocksize = sctx->fs_info->nodesize;
2736 spin_lock(&sctx->stat_lock);
2737 sctx->stat.tree_extents_scrubbed++;
2738 sctx->stat.tree_bytes_scrubbed += len;
2739 spin_unlock(&sctx->stat_lock);
2741 blocksize = sctx->fs_info->sectorsize;
2746 u64 l = min_t(u64, len, blocksize);
2749 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2750 /* push csums to sbio */
2751 have_csum = scrub_find_csum(sctx, logical, csum);
2753 ++sctx->stat.no_csum;
2754 if (sctx->is_dev_replace && !have_csum) {
2755 ret = copy_nocow_pages(sctx, logical, l,
2757 physical_for_dev_replace);
2758 goto behind_scrub_pages;
2761 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2762 mirror_num, have_csum ? csum : NULL, 0,
2763 physical_for_dev_replace);
2770 physical_for_dev_replace += l;
2775 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2776 u64 logical, u64 len,
2777 u64 physical, struct btrfs_device *dev,
2778 u64 flags, u64 gen, int mirror_num, u8 *csum)
2780 struct scrub_ctx *sctx = sparity->sctx;
2781 struct scrub_block *sblock;
2784 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2786 spin_lock(&sctx->stat_lock);
2787 sctx->stat.malloc_errors++;
2788 spin_unlock(&sctx->stat_lock);
2792 /* one ref inside this function, plus one for each page added to
2794 refcount_set(&sblock->refs, 1);
2795 sblock->sctx = sctx;
2796 sblock->no_io_error_seen = 1;
2797 sblock->sparity = sparity;
2798 scrub_parity_get(sparity);
2800 for (index = 0; len > 0; index++) {
2801 struct scrub_page *spage;
2802 u64 l = min_t(u64, len, PAGE_SIZE);
2804 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2807 spin_lock(&sctx->stat_lock);
2808 sctx->stat.malloc_errors++;
2809 spin_unlock(&sctx->stat_lock);
2810 scrub_block_put(sblock);
2813 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2814 /* For scrub block */
2815 scrub_page_get(spage);
2816 sblock->pagev[index] = spage;
2817 /* For scrub parity */
2818 scrub_page_get(spage);
2819 list_add_tail(&spage->list, &sparity->spages);
2820 spage->sblock = sblock;
2822 spage->flags = flags;
2823 spage->generation = gen;
2824 spage->logical = logical;
2825 spage->physical = physical;
2826 spage->mirror_num = mirror_num;
2828 spage->have_csum = 1;
2829 memcpy(spage->csum, csum, sctx->csum_size);
2831 spage->have_csum = 0;
2833 sblock->page_count++;
2834 spage->page = alloc_page(GFP_KERNEL);
2842 WARN_ON(sblock->page_count == 0);
2843 for (index = 0; index < sblock->page_count; index++) {
2844 struct scrub_page *spage = sblock->pagev[index];
2847 ret = scrub_add_page_to_rd_bio(sctx, spage);
2849 scrub_block_put(sblock);
2854 /* last one frees, either here or in bio completion for last page */
2855 scrub_block_put(sblock);
2859 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2860 u64 logical, u64 len,
2861 u64 physical, struct btrfs_device *dev,
2862 u64 flags, u64 gen, int mirror_num)
2864 struct scrub_ctx *sctx = sparity->sctx;
2866 u8 csum[BTRFS_CSUM_SIZE];
2870 scrub_parity_mark_sectors_error(sparity, logical, len);
2874 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2875 blocksize = sctx->fs_info->sectorsize;
2876 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2877 blocksize = sctx->fs_info->nodesize;
2879 blocksize = sctx->fs_info->sectorsize;
2884 u64 l = min_t(u64, len, blocksize);
2887 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2888 /* push csums to sbio */
2889 have_csum = scrub_find_csum(sctx, logical, csum);
2893 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2894 flags, gen, mirror_num,
2895 have_csum ? csum : NULL);
2907 * Given a physical address, this will calculate it's
2908 * logical offset. if this is a parity stripe, it will return
2909 * the most left data stripe's logical offset.
2911 * return 0 if it is a data stripe, 1 means parity stripe.
2913 static int get_raid56_logic_offset(u64 physical, int num,
2914 struct map_lookup *map, u64 *offset,
2924 last_offset = (physical - map->stripes[num].physical) *
2925 nr_data_stripes(map);
2927 *stripe_start = last_offset;
2929 *offset = last_offset;
2930 for (i = 0; i < nr_data_stripes(map); i++) {
2931 *offset = last_offset + i * map->stripe_len;
2933 stripe_nr = div64_u64(*offset, map->stripe_len);
2934 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2936 /* Work out the disk rotation on this stripe-set */
2937 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2938 /* calculate which stripe this data locates */
2940 stripe_index = rot % map->num_stripes;
2941 if (stripe_index == num)
2943 if (stripe_index < num)
2946 *offset = last_offset + j * map->stripe_len;
2950 static void scrub_free_parity(struct scrub_parity *sparity)
2952 struct scrub_ctx *sctx = sparity->sctx;
2953 struct scrub_page *curr, *next;
2956 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2958 spin_lock(&sctx->stat_lock);
2959 sctx->stat.read_errors += nbits;
2960 sctx->stat.uncorrectable_errors += nbits;
2961 spin_unlock(&sctx->stat_lock);
2964 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2965 list_del_init(&curr->list);
2966 scrub_page_put(curr);
2972 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2974 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2976 struct scrub_ctx *sctx = sparity->sctx;
2978 scrub_free_parity(sparity);
2979 scrub_pending_bio_dec(sctx);
2982 static void scrub_parity_bio_endio(struct bio *bio)
2984 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2985 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2988 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2993 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2994 scrub_parity_bio_endio_worker, NULL, NULL);
2995 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2998 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3000 struct scrub_ctx *sctx = sparity->sctx;
3001 struct btrfs_fs_info *fs_info = sctx->fs_info;
3003 struct btrfs_raid_bio *rbio;
3004 struct btrfs_bio *bbio = NULL;
3008 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3012 length = sparity->logic_end - sparity->logic_start;
3014 btrfs_bio_counter_inc_blocked(fs_info);
3015 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3017 if (ret || !bbio || !bbio->raid_map)
3020 bio = btrfs_io_bio_alloc(0);
3021 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3022 bio->bi_private = sparity;
3023 bio->bi_end_io = scrub_parity_bio_endio;
3025 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3026 length, sparity->scrub_dev,
3032 scrub_pending_bio_inc(sctx);
3033 raid56_parity_submit_scrub_rbio(rbio);
3039 btrfs_bio_counter_dec(fs_info);
3040 btrfs_put_bbio(bbio);
3041 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3043 spin_lock(&sctx->stat_lock);
3044 sctx->stat.malloc_errors++;
3045 spin_unlock(&sctx->stat_lock);
3047 scrub_free_parity(sparity);
3050 static inline int scrub_calc_parity_bitmap_len(int nsectors)
3052 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3055 static void scrub_parity_get(struct scrub_parity *sparity)
3057 refcount_inc(&sparity->refs);
3060 static void scrub_parity_put(struct scrub_parity *sparity)
3062 if (!refcount_dec_and_test(&sparity->refs))
3065 scrub_parity_check_and_repair(sparity);
3068 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3069 struct map_lookup *map,
3070 struct btrfs_device *sdev,
3071 struct btrfs_path *path,
3075 struct btrfs_fs_info *fs_info = sctx->fs_info;
3076 struct btrfs_root *root = fs_info->extent_root;
3077 struct btrfs_root *csum_root = fs_info->csum_root;
3078 struct btrfs_extent_item *extent;
3079 struct btrfs_bio *bbio = NULL;
3083 struct extent_buffer *l;
3084 struct btrfs_key key;
3087 u64 extent_physical;
3090 struct btrfs_device *extent_dev;
3091 struct scrub_parity *sparity;
3094 int extent_mirror_num;
3097 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3098 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3099 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3102 spin_lock(&sctx->stat_lock);
3103 sctx->stat.malloc_errors++;
3104 spin_unlock(&sctx->stat_lock);
3108 sparity->stripe_len = map->stripe_len;
3109 sparity->nsectors = nsectors;
3110 sparity->sctx = sctx;
3111 sparity->scrub_dev = sdev;
3112 sparity->logic_start = logic_start;
3113 sparity->logic_end = logic_end;
3114 refcount_set(&sparity->refs, 1);
3115 INIT_LIST_HEAD(&sparity->spages);
3116 sparity->dbitmap = sparity->bitmap;
3117 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3120 while (logic_start < logic_end) {
3121 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3122 key.type = BTRFS_METADATA_ITEM_KEY;
3124 key.type = BTRFS_EXTENT_ITEM_KEY;
3125 key.objectid = logic_start;
3126 key.offset = (u64)-1;
3128 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3133 ret = btrfs_previous_extent_item(root, path, 0);
3137 btrfs_release_path(path);
3138 ret = btrfs_search_slot(NULL, root, &key,
3150 slot = path->slots[0];
3151 if (slot >= btrfs_header_nritems(l)) {
3152 ret = btrfs_next_leaf(root, path);
3161 btrfs_item_key_to_cpu(l, &key, slot);
3163 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3164 key.type != BTRFS_METADATA_ITEM_KEY)
3167 if (key.type == BTRFS_METADATA_ITEM_KEY)
3168 bytes = fs_info->nodesize;
3172 if (key.objectid + bytes <= logic_start)
3175 if (key.objectid >= logic_end) {
3180 while (key.objectid >= logic_start + map->stripe_len)
3181 logic_start += map->stripe_len;
3183 extent = btrfs_item_ptr(l, slot,
3184 struct btrfs_extent_item);
3185 flags = btrfs_extent_flags(l, extent);
3186 generation = btrfs_extent_generation(l, extent);
3188 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3189 (key.objectid < logic_start ||
3190 key.objectid + bytes >
3191 logic_start + map->stripe_len)) {
3193 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3194 key.objectid, logic_start);
3195 spin_lock(&sctx->stat_lock);
3196 sctx->stat.uncorrectable_errors++;
3197 spin_unlock(&sctx->stat_lock);
3201 extent_logical = key.objectid;
3204 if (extent_logical < logic_start) {
3205 extent_len -= logic_start - extent_logical;
3206 extent_logical = logic_start;
3209 if (extent_logical + extent_len >
3210 logic_start + map->stripe_len)
3211 extent_len = logic_start + map->stripe_len -
3214 scrub_parity_mark_sectors_data(sparity, extent_logical,
3217 mapped_length = extent_len;
3219 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3220 extent_logical, &mapped_length, &bbio,
3223 if (!bbio || mapped_length < extent_len)
3227 btrfs_put_bbio(bbio);
3230 extent_physical = bbio->stripes[0].physical;
3231 extent_mirror_num = bbio->mirror_num;
3232 extent_dev = bbio->stripes[0].dev;
3233 btrfs_put_bbio(bbio);
3235 ret = btrfs_lookup_csums_range(csum_root,
3237 extent_logical + extent_len - 1,
3238 &sctx->csum_list, 1);
3242 ret = scrub_extent_for_parity(sparity, extent_logical,
3249 scrub_free_csums(sctx);
3254 if (extent_logical + extent_len <
3255 key.objectid + bytes) {
3256 logic_start += map->stripe_len;
3258 if (logic_start >= logic_end) {
3263 if (logic_start < key.objectid + bytes) {
3272 btrfs_release_path(path);
3277 logic_start += map->stripe_len;
3281 scrub_parity_mark_sectors_error(sparity, logic_start,
3282 logic_end - logic_start);
3283 scrub_parity_put(sparity);
3285 mutex_lock(&sctx->wr_lock);
3286 scrub_wr_submit(sctx);
3287 mutex_unlock(&sctx->wr_lock);
3289 btrfs_release_path(path);
3290 return ret < 0 ? ret : 0;
3293 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3294 struct map_lookup *map,
3295 struct btrfs_device *scrub_dev,
3296 int num, u64 base, u64 length,
3299 struct btrfs_path *path, *ppath;
3300 struct btrfs_fs_info *fs_info = sctx->fs_info;
3301 struct btrfs_root *root = fs_info->extent_root;
3302 struct btrfs_root *csum_root = fs_info->csum_root;
3303 struct btrfs_extent_item *extent;
3304 struct blk_plug plug;
3309 struct extent_buffer *l;
3316 struct reada_control *reada1;
3317 struct reada_control *reada2;
3318 struct btrfs_key key;
3319 struct btrfs_key key_end;
3320 u64 increment = map->stripe_len;
3323 u64 extent_physical;
3327 struct btrfs_device *extent_dev;
3328 int extent_mirror_num;
3331 physical = map->stripes[num].physical;
3333 nstripes = div64_u64(length, map->stripe_len);
3334 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3335 offset = map->stripe_len * num;
3336 increment = map->stripe_len * map->num_stripes;
3338 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3339 int factor = map->num_stripes / map->sub_stripes;
3340 offset = map->stripe_len * (num / map->sub_stripes);
3341 increment = map->stripe_len * factor;
3342 mirror_num = num % map->sub_stripes + 1;
3343 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3344 increment = map->stripe_len;
3345 mirror_num = num % map->num_stripes + 1;
3346 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3347 increment = map->stripe_len;
3348 mirror_num = num % map->num_stripes + 1;
3349 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3350 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3351 increment = map->stripe_len * nr_data_stripes(map);
3354 increment = map->stripe_len;
3358 path = btrfs_alloc_path();
3362 ppath = btrfs_alloc_path();
3364 btrfs_free_path(path);
3369 * work on commit root. The related disk blocks are static as
3370 * long as COW is applied. This means, it is save to rewrite
3371 * them to repair disk errors without any race conditions
3373 path->search_commit_root = 1;
3374 path->skip_locking = 1;
3376 ppath->search_commit_root = 1;
3377 ppath->skip_locking = 1;
3379 * trigger the readahead for extent tree csum tree and wait for
3380 * completion. During readahead, the scrub is officially paused
3381 * to not hold off transaction commits
3383 logical = base + offset;
3384 physical_end = physical + nstripes * map->stripe_len;
3385 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3386 get_raid56_logic_offset(physical_end, num,
3387 map, &logic_end, NULL);
3390 logic_end = logical + increment * nstripes;
3392 wait_event(sctx->list_wait,
3393 atomic_read(&sctx->bios_in_flight) == 0);
3394 scrub_blocked_if_needed(fs_info);
3396 /* FIXME it might be better to start readahead at commit root */
3397 key.objectid = logical;
3398 key.type = BTRFS_EXTENT_ITEM_KEY;
3399 key.offset = (u64)0;
3400 key_end.objectid = logic_end;
3401 key_end.type = BTRFS_METADATA_ITEM_KEY;
3402 key_end.offset = (u64)-1;
3403 reada1 = btrfs_reada_add(root, &key, &key_end);
3405 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3406 key.type = BTRFS_EXTENT_CSUM_KEY;
3407 key.offset = logical;
3408 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3409 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3410 key_end.offset = logic_end;
3411 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3413 if (!IS_ERR(reada1))
3414 btrfs_reada_wait(reada1);
3415 if (!IS_ERR(reada2))
3416 btrfs_reada_wait(reada2);
3420 * collect all data csums for the stripe to avoid seeking during
3421 * the scrub. This might currently (crc32) end up to be about 1MB
3423 blk_start_plug(&plug);
3426 * now find all extents for each stripe and scrub them
3429 while (physical < physical_end) {
3433 if (atomic_read(&fs_info->scrub_cancel_req) ||
3434 atomic_read(&sctx->cancel_req)) {
3439 * check to see if we have to pause
3441 if (atomic_read(&fs_info->scrub_pause_req)) {
3442 /* push queued extents */
3443 atomic_set(&sctx->flush_all_writes, 1);
3445 mutex_lock(&sctx->wr_lock);
3446 scrub_wr_submit(sctx);
3447 mutex_unlock(&sctx->wr_lock);
3448 wait_event(sctx->list_wait,
3449 atomic_read(&sctx->bios_in_flight) == 0);
3450 atomic_set(&sctx->flush_all_writes, 0);
3451 scrub_blocked_if_needed(fs_info);
3454 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3455 ret = get_raid56_logic_offset(physical, num, map,
3460 /* it is parity strip */
3461 stripe_logical += base;
3462 stripe_end = stripe_logical + increment;
3463 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3464 ppath, stripe_logical,
3472 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3473 key.type = BTRFS_METADATA_ITEM_KEY;
3475 key.type = BTRFS_EXTENT_ITEM_KEY;
3476 key.objectid = logical;
3477 key.offset = (u64)-1;
3479 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3484 ret = btrfs_previous_extent_item(root, path, 0);
3488 /* there's no smaller item, so stick with the
3490 btrfs_release_path(path);
3491 ret = btrfs_search_slot(NULL, root, &key,
3503 slot = path->slots[0];
3504 if (slot >= btrfs_header_nritems(l)) {
3505 ret = btrfs_next_leaf(root, path);
3514 btrfs_item_key_to_cpu(l, &key, slot);
3516 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3517 key.type != BTRFS_METADATA_ITEM_KEY)
3520 if (key.type == BTRFS_METADATA_ITEM_KEY)
3521 bytes = fs_info->nodesize;
3525 if (key.objectid + bytes <= logical)
3528 if (key.objectid >= logical + map->stripe_len) {
3529 /* out of this device extent */
3530 if (key.objectid >= logic_end)
3535 extent = btrfs_item_ptr(l, slot,
3536 struct btrfs_extent_item);
3537 flags = btrfs_extent_flags(l, extent);
3538 generation = btrfs_extent_generation(l, extent);
3540 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3541 (key.objectid < logical ||
3542 key.objectid + bytes >
3543 logical + map->stripe_len)) {
3545 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3546 key.objectid, logical);
3547 spin_lock(&sctx->stat_lock);
3548 sctx->stat.uncorrectable_errors++;
3549 spin_unlock(&sctx->stat_lock);
3554 extent_logical = key.objectid;
3558 * trim extent to this stripe
3560 if (extent_logical < logical) {
3561 extent_len -= logical - extent_logical;
3562 extent_logical = logical;
3564 if (extent_logical + extent_len >
3565 logical + map->stripe_len) {
3566 extent_len = logical + map->stripe_len -
3570 extent_physical = extent_logical - logical + physical;
3571 extent_dev = scrub_dev;
3572 extent_mirror_num = mirror_num;
3574 scrub_remap_extent(fs_info, extent_logical,
3575 extent_len, &extent_physical,
3577 &extent_mirror_num);
3579 ret = btrfs_lookup_csums_range(csum_root,
3583 &sctx->csum_list, 1);
3587 ret = scrub_extent(sctx, extent_logical, extent_len,
3588 extent_physical, extent_dev, flags,
3589 generation, extent_mirror_num,
3590 extent_logical - logical + physical);
3592 scrub_free_csums(sctx);
3597 if (extent_logical + extent_len <
3598 key.objectid + bytes) {
3599 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3601 * loop until we find next data stripe
3602 * or we have finished all stripes.
3605 physical += map->stripe_len;
3606 ret = get_raid56_logic_offset(physical,
3611 if (ret && physical < physical_end) {
3612 stripe_logical += base;
3613 stripe_end = stripe_logical +
3615 ret = scrub_raid56_parity(sctx,
3616 map, scrub_dev, ppath,
3624 physical += map->stripe_len;
3625 logical += increment;
3627 if (logical < key.objectid + bytes) {
3632 if (physical >= physical_end) {
3640 btrfs_release_path(path);
3642 logical += increment;
3643 physical += map->stripe_len;
3644 spin_lock(&sctx->stat_lock);
3646 sctx->stat.last_physical = map->stripes[num].physical +
3649 sctx->stat.last_physical = physical;
3650 spin_unlock(&sctx->stat_lock);
3655 /* push queued extents */
3657 mutex_lock(&sctx->wr_lock);
3658 scrub_wr_submit(sctx);
3659 mutex_unlock(&sctx->wr_lock);
3661 blk_finish_plug(&plug);
3662 btrfs_free_path(path);
3663 btrfs_free_path(ppath);
3664 return ret < 0 ? ret : 0;
3667 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3668 struct btrfs_device *scrub_dev,
3669 u64 chunk_offset, u64 length,
3671 struct btrfs_block_group_cache *cache,
3674 struct btrfs_fs_info *fs_info = sctx->fs_info;
3675 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3676 struct map_lookup *map;
3677 struct extent_map *em;
3681 read_lock(&map_tree->map_tree.lock);
3682 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3683 read_unlock(&map_tree->map_tree.lock);
3687 * Might have been an unused block group deleted by the cleaner
3688 * kthread or relocation.
3690 spin_lock(&cache->lock);
3691 if (!cache->removed)
3693 spin_unlock(&cache->lock);
3698 map = em->map_lookup;
3699 if (em->start != chunk_offset)
3702 if (em->len < length)
3705 for (i = 0; i < map->num_stripes; ++i) {
3706 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3707 map->stripes[i].physical == dev_offset) {
3708 ret = scrub_stripe(sctx, map, scrub_dev, i,
3709 chunk_offset, length,
3716 free_extent_map(em);
3721 static noinline_for_stack
3722 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3723 struct btrfs_device *scrub_dev, u64 start, u64 end,
3726 struct btrfs_dev_extent *dev_extent = NULL;
3727 struct btrfs_path *path;
3728 struct btrfs_fs_info *fs_info = sctx->fs_info;
3729 struct btrfs_root *root = fs_info->dev_root;
3735 struct extent_buffer *l;
3736 struct btrfs_key key;
3737 struct btrfs_key found_key;
3738 struct btrfs_block_group_cache *cache;
3739 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3741 path = btrfs_alloc_path();
3745 path->reada = READA_FORWARD;
3746 path->search_commit_root = 1;
3747 path->skip_locking = 1;
3749 key.objectid = scrub_dev->devid;
3751 key.type = BTRFS_DEV_EXTENT_KEY;
3754 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3758 if (path->slots[0] >=
3759 btrfs_header_nritems(path->nodes[0])) {
3760 ret = btrfs_next_leaf(root, path);
3773 slot = path->slots[0];
3775 btrfs_item_key_to_cpu(l, &found_key, slot);
3777 if (found_key.objectid != scrub_dev->devid)
3780 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3783 if (found_key.offset >= end)
3786 if (found_key.offset < key.offset)
3789 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3790 length = btrfs_dev_extent_length(l, dev_extent);
3792 if (found_key.offset + length <= start)
3795 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3798 * get a reference on the corresponding block group to prevent
3799 * the chunk from going away while we scrub it
3801 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3803 /* some chunks are removed but not committed to disk yet,
3804 * continue scrubbing */
3809 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3810 * to avoid deadlock caused by:
3811 * btrfs_inc_block_group_ro()
3812 * -> btrfs_wait_for_commit()
3813 * -> btrfs_commit_transaction()
3814 * -> btrfs_scrub_pause()
3816 scrub_pause_on(fs_info);
3817 ret = btrfs_inc_block_group_ro(fs_info, cache);
3818 if (!ret && is_dev_replace) {
3820 * If we are doing a device replace wait for any tasks
3821 * that started dellaloc right before we set the block
3822 * group to RO mode, as they might have just allocated
3823 * an extent from it or decided they could do a nocow
3824 * write. And if any such tasks did that, wait for their
3825 * ordered extents to complete and then commit the
3826 * current transaction, so that we can later see the new
3827 * extent items in the extent tree - the ordered extents
3828 * create delayed data references (for cow writes) when
3829 * they complete, which will be run and insert the
3830 * corresponding extent items into the extent tree when
3831 * we commit the transaction they used when running
3832 * inode.c:btrfs_finish_ordered_io(). We later use
3833 * the commit root of the extent tree to find extents
3834 * to copy from the srcdev into the tgtdev, and we don't
3835 * want to miss any new extents.
3837 btrfs_wait_block_group_reservations(cache);
3838 btrfs_wait_nocow_writers(cache);
3839 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3840 cache->key.objectid,
3843 struct btrfs_trans_handle *trans;
3845 trans = btrfs_join_transaction(root);
3847 ret = PTR_ERR(trans);
3849 ret = btrfs_commit_transaction(trans);
3851 scrub_pause_off(fs_info);
3852 btrfs_put_block_group(cache);
3857 scrub_pause_off(fs_info);
3861 } else if (ret == -ENOSPC) {
3863 * btrfs_inc_block_group_ro return -ENOSPC when it
3864 * failed in creating new chunk for metadata.
3865 * It is not a problem for scrub/replace, because
3866 * metadata are always cowed, and our scrub paused
3867 * commit_transactions.
3872 "failed setting block group ro: %d", ret);
3873 btrfs_put_block_group(cache);
3877 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3878 dev_replace->cursor_right = found_key.offset + length;
3879 dev_replace->cursor_left = found_key.offset;
3880 dev_replace->item_needs_writeback = 1;
3881 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3882 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3883 found_key.offset, cache, is_dev_replace);
3886 * flush, submit all pending read and write bios, afterwards
3888 * Note that in the dev replace case, a read request causes
3889 * write requests that are submitted in the read completion
3890 * worker. Therefore in the current situation, it is required
3891 * that all write requests are flushed, so that all read and
3892 * write requests are really completed when bios_in_flight
3895 atomic_set(&sctx->flush_all_writes, 1);
3897 mutex_lock(&sctx->wr_lock);
3898 scrub_wr_submit(sctx);
3899 mutex_unlock(&sctx->wr_lock);
3901 wait_event(sctx->list_wait,
3902 atomic_read(&sctx->bios_in_flight) == 0);
3904 scrub_pause_on(fs_info);
3907 * must be called before we decrease @scrub_paused.
3908 * make sure we don't block transaction commit while
3909 * we are waiting pending workers finished.
3911 wait_event(sctx->list_wait,
3912 atomic_read(&sctx->workers_pending) == 0);
3913 atomic_set(&sctx->flush_all_writes, 0);
3915 scrub_pause_off(fs_info);
3917 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3918 dev_replace->cursor_left = dev_replace->cursor_right;
3919 dev_replace->item_needs_writeback = 1;
3920 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3923 btrfs_dec_block_group_ro(cache);
3926 * We might have prevented the cleaner kthread from deleting
3927 * this block group if it was already unused because we raced
3928 * and set it to RO mode first. So add it back to the unused
3929 * list, otherwise it might not ever be deleted unless a manual
3930 * balance is triggered or it becomes used and unused again.
3932 spin_lock(&cache->lock);
3933 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3934 btrfs_block_group_used(&cache->item) == 0) {
3935 spin_unlock(&cache->lock);
3936 spin_lock(&fs_info->unused_bgs_lock);
3937 if (list_empty(&cache->bg_list)) {
3938 btrfs_get_block_group(cache);
3939 list_add_tail(&cache->bg_list,
3940 &fs_info->unused_bgs);
3942 spin_unlock(&fs_info->unused_bgs_lock);
3944 spin_unlock(&cache->lock);
3947 btrfs_put_block_group(cache);
3950 if (is_dev_replace &&
3951 atomic64_read(&dev_replace->num_write_errors) > 0) {
3955 if (sctx->stat.malloc_errors > 0) {
3960 key.offset = found_key.offset + length;
3961 btrfs_release_path(path);
3964 btrfs_free_path(path);
3969 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3970 struct btrfs_device *scrub_dev)
3976 struct btrfs_fs_info *fs_info = sctx->fs_info;
3978 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3981 /* Seed devices of a new filesystem has their own generation. */
3982 if (scrub_dev->fs_devices != fs_info->fs_devices)
3983 gen = scrub_dev->generation;
3985 gen = fs_info->last_trans_committed;
3987 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3988 bytenr = btrfs_sb_offset(i);
3989 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3990 scrub_dev->commit_total_bytes)
3993 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3994 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3999 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4005 * get a reference count on fs_info->scrub_workers. start worker if necessary
4007 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4010 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4011 int max_active = fs_info->thread_pool_size;
4013 if (fs_info->scrub_workers_refcnt == 0) {
4015 fs_info->scrub_workers =
4016 btrfs_alloc_workqueue(fs_info, "scrub", flags,
4019 fs_info->scrub_workers =
4020 btrfs_alloc_workqueue(fs_info, "scrub", flags,
4022 if (!fs_info->scrub_workers)
4023 goto fail_scrub_workers;
4025 fs_info->scrub_wr_completion_workers =
4026 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4028 if (!fs_info->scrub_wr_completion_workers)
4029 goto fail_scrub_wr_completion_workers;
4031 fs_info->scrub_nocow_workers =
4032 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4033 if (!fs_info->scrub_nocow_workers)
4034 goto fail_scrub_nocow_workers;
4035 fs_info->scrub_parity_workers =
4036 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4038 if (!fs_info->scrub_parity_workers)
4039 goto fail_scrub_parity_workers;
4041 ++fs_info->scrub_workers_refcnt;
4044 fail_scrub_parity_workers:
4045 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4046 fail_scrub_nocow_workers:
4047 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4048 fail_scrub_wr_completion_workers:
4049 btrfs_destroy_workqueue(fs_info->scrub_workers);
4054 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4056 if (--fs_info->scrub_workers_refcnt == 0) {
4057 btrfs_destroy_workqueue(fs_info->scrub_workers);
4058 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4059 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4060 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4062 WARN_ON(fs_info->scrub_workers_refcnt < 0);
4065 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4066 u64 end, struct btrfs_scrub_progress *progress,
4067 int readonly, int is_dev_replace)
4069 struct scrub_ctx *sctx;
4071 struct btrfs_device *dev;
4072 struct rcu_string *name;
4074 if (btrfs_fs_closing(fs_info))
4077 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4079 * in this case scrub is unable to calculate the checksum
4080 * the way scrub is implemented. Do not handle this
4081 * situation at all because it won't ever happen.
4084 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4090 if (fs_info->sectorsize != PAGE_SIZE) {
4091 /* not supported for data w/o checksums */
4092 btrfs_err_rl(fs_info,
4093 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4094 fs_info->sectorsize, PAGE_SIZE);
4098 if (fs_info->nodesize >
4099 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4100 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4102 * would exhaust the array bounds of pagev member in
4103 * struct scrub_block
4106 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4108 SCRUB_MAX_PAGES_PER_BLOCK,
4109 fs_info->sectorsize,
4110 SCRUB_MAX_PAGES_PER_BLOCK);
4115 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4116 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4117 if (!dev || (dev->missing && !is_dev_replace)) {
4118 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4122 if (!is_dev_replace && !readonly && !dev->writeable) {
4123 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4125 name = rcu_dereference(dev->name);
4126 btrfs_err(fs_info, "scrub: device %s is not writable",
4132 mutex_lock(&fs_info->scrub_lock);
4133 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
4134 mutex_unlock(&fs_info->scrub_lock);
4135 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4139 btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
4140 if (dev->scrub_device ||
4142 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4143 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
4144 mutex_unlock(&fs_info->scrub_lock);
4145 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4146 return -EINPROGRESS;
4148 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
4150 ret = scrub_workers_get(fs_info, is_dev_replace);
4152 mutex_unlock(&fs_info->scrub_lock);
4153 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4157 sctx = scrub_setup_ctx(dev, is_dev_replace);
4159 mutex_unlock(&fs_info->scrub_lock);
4160 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4161 scrub_workers_put(fs_info);
4162 return PTR_ERR(sctx);
4164 sctx->readonly = readonly;
4165 dev->scrub_device = sctx;
4166 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4169 * checking @scrub_pause_req here, we can avoid
4170 * race between committing transaction and scrubbing.
4172 __scrub_blocked_if_needed(fs_info);
4173 atomic_inc(&fs_info->scrubs_running);
4174 mutex_unlock(&fs_info->scrub_lock);
4176 if (!is_dev_replace) {
4178 * by holding device list mutex, we can
4179 * kick off writing super in log tree sync.
4181 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4182 ret = scrub_supers(sctx, dev);
4183 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4187 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4190 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4191 atomic_dec(&fs_info->scrubs_running);
4192 wake_up(&fs_info->scrub_pause_wait);
4194 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4197 memcpy(progress, &sctx->stat, sizeof(*progress));
4199 mutex_lock(&fs_info->scrub_lock);
4200 dev->scrub_device = NULL;
4201 scrub_workers_put(fs_info);
4202 mutex_unlock(&fs_info->scrub_lock);
4204 scrub_put_ctx(sctx);
4209 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4211 mutex_lock(&fs_info->scrub_lock);
4212 atomic_inc(&fs_info->scrub_pause_req);
4213 while (atomic_read(&fs_info->scrubs_paused) !=
4214 atomic_read(&fs_info->scrubs_running)) {
4215 mutex_unlock(&fs_info->scrub_lock);
4216 wait_event(fs_info->scrub_pause_wait,
4217 atomic_read(&fs_info->scrubs_paused) ==
4218 atomic_read(&fs_info->scrubs_running));
4219 mutex_lock(&fs_info->scrub_lock);
4221 mutex_unlock(&fs_info->scrub_lock);
4224 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4226 atomic_dec(&fs_info->scrub_pause_req);
4227 wake_up(&fs_info->scrub_pause_wait);
4230 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4232 mutex_lock(&fs_info->scrub_lock);
4233 if (!atomic_read(&fs_info->scrubs_running)) {
4234 mutex_unlock(&fs_info->scrub_lock);
4238 atomic_inc(&fs_info->scrub_cancel_req);
4239 while (atomic_read(&fs_info->scrubs_running)) {
4240 mutex_unlock(&fs_info->scrub_lock);
4241 wait_event(fs_info->scrub_pause_wait,
4242 atomic_read(&fs_info->scrubs_running) == 0);
4243 mutex_lock(&fs_info->scrub_lock);
4245 atomic_dec(&fs_info->scrub_cancel_req);
4246 mutex_unlock(&fs_info->scrub_lock);
4251 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4252 struct btrfs_device *dev)
4254 struct scrub_ctx *sctx;
4256 mutex_lock(&fs_info->scrub_lock);
4257 sctx = dev->scrub_device;
4259 mutex_unlock(&fs_info->scrub_lock);
4262 atomic_inc(&sctx->cancel_req);
4263 while (dev->scrub_device) {
4264 mutex_unlock(&fs_info->scrub_lock);
4265 wait_event(fs_info->scrub_pause_wait,
4266 dev->scrub_device == NULL);
4267 mutex_lock(&fs_info->scrub_lock);
4269 mutex_unlock(&fs_info->scrub_lock);
4274 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4275 struct btrfs_scrub_progress *progress)
4277 struct btrfs_device *dev;
4278 struct scrub_ctx *sctx = NULL;
4280 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4281 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4283 sctx = dev->scrub_device;
4285 memcpy(progress, &sctx->stat, sizeof(*progress));
4286 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4288 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4291 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4292 u64 extent_logical, u64 extent_len,
4293 u64 *extent_physical,
4294 struct btrfs_device **extent_dev,
4295 int *extent_mirror_num)
4298 struct btrfs_bio *bbio = NULL;
4301 mapped_length = extent_len;
4302 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4303 &mapped_length, &bbio, 0);
4304 if (ret || !bbio || mapped_length < extent_len ||
4305 !bbio->stripes[0].dev->bdev) {
4306 btrfs_put_bbio(bbio);
4310 *extent_physical = bbio->stripes[0].physical;
4311 *extent_mirror_num = bbio->mirror_num;
4312 *extent_dev = bbio->stripes[0].dev;
4313 btrfs_put_bbio(bbio);
4316 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4317 int mirror_num, u64 physical_for_dev_replace)
4319 struct scrub_copy_nocow_ctx *nocow_ctx;
4320 struct btrfs_fs_info *fs_info = sctx->fs_info;
4322 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4324 spin_lock(&sctx->stat_lock);
4325 sctx->stat.malloc_errors++;
4326 spin_unlock(&sctx->stat_lock);
4330 scrub_pending_trans_workers_inc(sctx);
4332 nocow_ctx->sctx = sctx;
4333 nocow_ctx->logical = logical;
4334 nocow_ctx->len = len;
4335 nocow_ctx->mirror_num = mirror_num;
4336 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4337 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4338 copy_nocow_pages_worker, NULL, NULL);
4339 INIT_LIST_HEAD(&nocow_ctx->inodes);
4340 btrfs_queue_work(fs_info->scrub_nocow_workers,
4346 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4348 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4349 struct scrub_nocow_inode *nocow_inode;
4351 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4354 nocow_inode->inum = inum;
4355 nocow_inode->offset = offset;
4356 nocow_inode->root = root;
4357 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4361 #define COPY_COMPLETE 1
4363 static void copy_nocow_pages_worker(struct btrfs_work *work)
4365 struct scrub_copy_nocow_ctx *nocow_ctx =
4366 container_of(work, struct scrub_copy_nocow_ctx, work);
4367 struct scrub_ctx *sctx = nocow_ctx->sctx;
4368 struct btrfs_fs_info *fs_info = sctx->fs_info;
4369 struct btrfs_root *root = fs_info->extent_root;
4370 u64 logical = nocow_ctx->logical;
4371 u64 len = nocow_ctx->len;
4372 int mirror_num = nocow_ctx->mirror_num;
4373 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4375 struct btrfs_trans_handle *trans = NULL;
4376 struct btrfs_path *path;
4377 int not_written = 0;
4379 path = btrfs_alloc_path();
4381 spin_lock(&sctx->stat_lock);
4382 sctx->stat.malloc_errors++;
4383 spin_unlock(&sctx->stat_lock);
4388 trans = btrfs_join_transaction(root);
4389 if (IS_ERR(trans)) {
4394 ret = iterate_inodes_from_logical(logical, fs_info, path,
4395 record_inode_for_nocow, nocow_ctx);
4396 if (ret != 0 && ret != -ENOENT) {
4398 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4399 logical, physical_for_dev_replace, len, mirror_num,
4405 btrfs_end_transaction(trans);
4407 while (!list_empty(&nocow_ctx->inodes)) {
4408 struct scrub_nocow_inode *entry;
4409 entry = list_first_entry(&nocow_ctx->inodes,
4410 struct scrub_nocow_inode,
4412 list_del_init(&entry->list);
4413 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4414 entry->root, nocow_ctx);
4416 if (ret == COPY_COMPLETE) {
4424 while (!list_empty(&nocow_ctx->inodes)) {
4425 struct scrub_nocow_inode *entry;
4426 entry = list_first_entry(&nocow_ctx->inodes,
4427 struct scrub_nocow_inode,
4429 list_del_init(&entry->list);
4432 if (trans && !IS_ERR(trans))
4433 btrfs_end_transaction(trans);
4435 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4436 num_uncorrectable_read_errors);
4438 btrfs_free_path(path);
4441 scrub_pending_trans_workers_dec(sctx);
4444 static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4447 struct extent_state *cached_state = NULL;
4448 struct btrfs_ordered_extent *ordered;
4449 struct extent_io_tree *io_tree;
4450 struct extent_map *em;
4451 u64 lockstart = start, lockend = start + len - 1;
4454 io_tree = &inode->io_tree;
4456 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4457 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4459 btrfs_put_ordered_extent(ordered);
4464 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4471 * This extent does not actually cover the logical extent anymore,
4472 * move on to the next inode.
4474 if (em->block_start > logical ||
4475 em->block_start + em->block_len < logical + len) {
4476 free_extent_map(em);
4480 free_extent_map(em);
4483 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4488 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4489 struct scrub_copy_nocow_ctx *nocow_ctx)
4491 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4492 struct btrfs_key key;
4493 struct inode *inode;
4495 struct btrfs_root *local_root;
4496 struct extent_io_tree *io_tree;
4497 u64 physical_for_dev_replace;
4498 u64 nocow_ctx_logical;
4499 u64 len = nocow_ctx->len;
4500 unsigned long index;
4505 key.objectid = root;
4506 key.type = BTRFS_ROOT_ITEM_KEY;
4507 key.offset = (u64)-1;
4509 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4511 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4512 if (IS_ERR(local_root)) {
4513 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4514 return PTR_ERR(local_root);
4517 key.type = BTRFS_INODE_ITEM_KEY;
4518 key.objectid = inum;
4520 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4521 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4523 return PTR_ERR(inode);
4525 /* Avoid truncate/dio/punch hole.. */
4527 inode_dio_wait(inode);
4529 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4530 io_tree = &BTRFS_I(inode)->io_tree;
4531 nocow_ctx_logical = nocow_ctx->logical;
4533 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4536 ret = ret > 0 ? 0 : ret;
4540 while (len >= PAGE_SIZE) {
4541 index = offset >> PAGE_SHIFT;
4543 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4545 btrfs_err(fs_info, "find_or_create_page() failed");
4550 if (PageUptodate(page)) {
4551 if (PageDirty(page))
4554 ClearPageError(page);
4555 err = extent_read_full_page(io_tree, page,
4557 nocow_ctx->mirror_num);
4565 * If the page has been remove from the page cache,
4566 * the data on it is meaningless, because it may be
4567 * old one, the new data may be written into the new
4568 * page in the page cache.
4570 if (page->mapping != inode->i_mapping) {
4575 if (!PageUptodate(page)) {
4581 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4584 ret = ret > 0 ? 0 : ret;
4588 err = write_page_nocow(nocow_ctx->sctx,
4589 physical_for_dev_replace, page);
4599 offset += PAGE_SIZE;
4600 physical_for_dev_replace += PAGE_SIZE;
4601 nocow_ctx_logical += PAGE_SIZE;
4604 ret = COPY_COMPLETE;
4606 inode_unlock(inode);
4611 static int write_page_nocow(struct scrub_ctx *sctx,
4612 u64 physical_for_dev_replace, struct page *page)
4615 struct btrfs_device *dev;
4618 dev = sctx->wr_tgtdev;
4622 btrfs_warn_rl(dev->fs_info,
4623 "scrub write_page_nocow(bdev == NULL) is unexpected");
4626 bio = btrfs_io_bio_alloc(1);
4627 bio->bi_iter.bi_size = 0;
4628 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4629 bio->bi_bdev = dev->bdev;
4630 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4631 ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4632 if (ret != PAGE_SIZE) {
4635 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4639 if (btrfsic_submit_bio_wait(bio))
4640 goto leave_with_eio;