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 struct btrfs_device *wr_tgtdev;
186 bool flush_all_writes;
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);
304 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
306 return page->recover &&
307 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
310 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
312 refcount_inc(&sctx->refs);
313 atomic_inc(&sctx->bios_in_flight);
316 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
318 atomic_dec(&sctx->bios_in_flight);
319 wake_up(&sctx->list_wait);
323 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
325 while (atomic_read(&fs_info->scrub_pause_req)) {
326 mutex_unlock(&fs_info->scrub_lock);
327 wait_event(fs_info->scrub_pause_wait,
328 atomic_read(&fs_info->scrub_pause_req) == 0);
329 mutex_lock(&fs_info->scrub_lock);
333 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
335 atomic_inc(&fs_info->scrubs_paused);
336 wake_up(&fs_info->scrub_pause_wait);
339 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
341 mutex_lock(&fs_info->scrub_lock);
342 __scrub_blocked_if_needed(fs_info);
343 atomic_dec(&fs_info->scrubs_paused);
344 mutex_unlock(&fs_info->scrub_lock);
346 wake_up(&fs_info->scrub_pause_wait);
349 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
351 scrub_pause_on(fs_info);
352 scrub_pause_off(fs_info);
356 * Insert new full stripe lock into full stripe locks tree
358 * Return pointer to existing or newly inserted full_stripe_lock structure if
359 * everything works well.
360 * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
362 * NOTE: caller must hold full_stripe_locks_root->lock before calling this
365 static struct full_stripe_lock *insert_full_stripe_lock(
366 struct btrfs_full_stripe_locks_tree *locks_root,
370 struct rb_node *parent = NULL;
371 struct full_stripe_lock *entry;
372 struct full_stripe_lock *ret;
374 lockdep_assert_held(&locks_root->lock);
376 p = &locks_root->root.rb_node;
379 entry = rb_entry(parent, struct full_stripe_lock, node);
380 if (fstripe_logical < entry->logical) {
382 } else if (fstripe_logical > entry->logical) {
390 /* Insert new lock */
391 ret = kmalloc(sizeof(*ret), GFP_KERNEL);
393 return ERR_PTR(-ENOMEM);
394 ret->logical = fstripe_logical;
396 mutex_init(&ret->mutex);
398 rb_link_node(&ret->node, parent, p);
399 rb_insert_color(&ret->node, &locks_root->root);
404 * Search for a full stripe lock of a block group
406 * Return pointer to existing full stripe lock if found
407 * Return NULL if not found
409 static struct full_stripe_lock *search_full_stripe_lock(
410 struct btrfs_full_stripe_locks_tree *locks_root,
413 struct rb_node *node;
414 struct full_stripe_lock *entry;
416 lockdep_assert_held(&locks_root->lock);
418 node = locks_root->root.rb_node;
420 entry = rb_entry(node, struct full_stripe_lock, node);
421 if (fstripe_logical < entry->logical)
422 node = node->rb_left;
423 else if (fstripe_logical > entry->logical)
424 node = node->rb_right;
432 * Helper to get full stripe logical from a normal bytenr.
434 * Caller must ensure @cache is a RAID56 block group.
436 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
442 * Due to chunk item size limit, full stripe length should not be
443 * larger than U32_MAX. Just a sanity check here.
445 WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
448 * round_down() can only handle power of 2, while RAID56 full
449 * stripe length can be 64KiB * n, so we need to manually round down.
451 ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
452 cache->full_stripe_len + cache->key.objectid;
457 * Lock a full stripe to avoid concurrency of recovery and read
459 * It's only used for profiles with parities (RAID5/6), for other profiles it
462 * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
463 * So caller must call unlock_full_stripe() at the same context.
465 * Return <0 if encounters error.
467 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
470 struct btrfs_block_group_cache *bg_cache;
471 struct btrfs_full_stripe_locks_tree *locks_root;
472 struct full_stripe_lock *existing;
477 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
483 /* Profiles not based on parity don't need full stripe lock */
484 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
486 locks_root = &bg_cache->full_stripe_locks_root;
488 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
490 /* Now insert the full stripe lock */
491 mutex_lock(&locks_root->lock);
492 existing = insert_full_stripe_lock(locks_root, fstripe_start);
493 mutex_unlock(&locks_root->lock);
494 if (IS_ERR(existing)) {
495 ret = PTR_ERR(existing);
498 mutex_lock(&existing->mutex);
501 btrfs_put_block_group(bg_cache);
506 * Unlock a full stripe.
508 * NOTE: Caller must ensure it's the same context calling corresponding
509 * lock_full_stripe().
511 * Return 0 if we unlock full stripe without problem.
512 * Return <0 for error
514 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
517 struct btrfs_block_group_cache *bg_cache;
518 struct btrfs_full_stripe_locks_tree *locks_root;
519 struct full_stripe_lock *fstripe_lock;
524 /* If we didn't acquire full stripe lock, no need to continue */
528 bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
533 if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
536 locks_root = &bg_cache->full_stripe_locks_root;
537 fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
539 mutex_lock(&locks_root->lock);
540 fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
541 /* Unpaired unlock_full_stripe() detected */
545 mutex_unlock(&locks_root->lock);
549 if (fstripe_lock->refs == 0) {
551 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
552 fstripe_lock->logical);
554 fstripe_lock->refs--;
557 if (fstripe_lock->refs == 0) {
558 rb_erase(&fstripe_lock->node, &locks_root->root);
561 mutex_unlock(&locks_root->lock);
563 mutex_unlock(&fstripe_lock->mutex);
567 btrfs_put_block_group(bg_cache);
572 * used for workers that require transaction commits (i.e., for the
575 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
577 struct btrfs_fs_info *fs_info = sctx->fs_info;
579 refcount_inc(&sctx->refs);
581 * increment scrubs_running to prevent cancel requests from
582 * completing as long as a worker is running. we must also
583 * increment scrubs_paused to prevent deadlocking on pause
584 * requests used for transactions commits (as the worker uses a
585 * transaction context). it is safe to regard the worker
586 * as paused for all matters practical. effectively, we only
587 * avoid cancellation requests from completing.
589 mutex_lock(&fs_info->scrub_lock);
590 atomic_inc(&fs_info->scrubs_running);
591 atomic_inc(&fs_info->scrubs_paused);
592 mutex_unlock(&fs_info->scrub_lock);
595 * check if @scrubs_running=@scrubs_paused condition
596 * inside wait_event() is not an atomic operation.
597 * which means we may inc/dec @scrub_running/paused
598 * at any time. Let's wake up @scrub_pause_wait as
599 * much as we can to let commit transaction blocked less.
601 wake_up(&fs_info->scrub_pause_wait);
603 atomic_inc(&sctx->workers_pending);
606 /* used for workers that require transaction commits */
607 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
609 struct btrfs_fs_info *fs_info = sctx->fs_info;
612 * see scrub_pending_trans_workers_inc() why we're pretending
613 * to be paused in the scrub counters
615 mutex_lock(&fs_info->scrub_lock);
616 atomic_dec(&fs_info->scrubs_running);
617 atomic_dec(&fs_info->scrubs_paused);
618 mutex_unlock(&fs_info->scrub_lock);
619 atomic_dec(&sctx->workers_pending);
620 wake_up(&fs_info->scrub_pause_wait);
621 wake_up(&sctx->list_wait);
625 static void scrub_free_csums(struct scrub_ctx *sctx)
627 while (!list_empty(&sctx->csum_list)) {
628 struct btrfs_ordered_sum *sum;
629 sum = list_first_entry(&sctx->csum_list,
630 struct btrfs_ordered_sum, list);
631 list_del(&sum->list);
636 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
643 /* this can happen when scrub is cancelled */
644 if (sctx->curr != -1) {
645 struct scrub_bio *sbio = sctx->bios[sctx->curr];
647 for (i = 0; i < sbio->page_count; i++) {
648 WARN_ON(!sbio->pagev[i]->page);
649 scrub_block_put(sbio->pagev[i]->sblock);
654 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
655 struct scrub_bio *sbio = sctx->bios[i];
662 kfree(sctx->wr_curr_bio);
663 scrub_free_csums(sctx);
667 static void scrub_put_ctx(struct scrub_ctx *sctx)
669 if (refcount_dec_and_test(&sctx->refs))
670 scrub_free_ctx(sctx);
673 static noinline_for_stack
674 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
676 struct scrub_ctx *sctx;
678 struct btrfs_fs_info *fs_info = dev->fs_info;
680 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
683 refcount_set(&sctx->refs, 1);
684 sctx->is_dev_replace = is_dev_replace;
685 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
687 sctx->fs_info = dev->fs_info;
688 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
689 struct scrub_bio *sbio;
691 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
694 sctx->bios[i] = sbio;
698 sbio->page_count = 0;
699 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
700 scrub_bio_end_io_worker, NULL, NULL);
702 if (i != SCRUB_BIOS_PER_SCTX - 1)
703 sctx->bios[i]->next_free = i + 1;
705 sctx->bios[i]->next_free = -1;
707 sctx->first_free = 0;
708 atomic_set(&sctx->bios_in_flight, 0);
709 atomic_set(&sctx->workers_pending, 0);
710 atomic_set(&sctx->cancel_req, 0);
711 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
712 INIT_LIST_HEAD(&sctx->csum_list);
714 spin_lock_init(&sctx->list_lock);
715 spin_lock_init(&sctx->stat_lock);
716 init_waitqueue_head(&sctx->list_wait);
718 WARN_ON(sctx->wr_curr_bio != NULL);
719 mutex_init(&sctx->wr_lock);
720 sctx->wr_curr_bio = NULL;
721 if (is_dev_replace) {
722 WARN_ON(!fs_info->dev_replace.tgtdev);
723 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
724 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
725 sctx->flush_all_writes = false;
731 scrub_free_ctx(sctx);
732 return ERR_PTR(-ENOMEM);
735 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
743 struct extent_buffer *eb;
744 struct btrfs_inode_item *inode_item;
745 struct scrub_warning *swarn = warn_ctx;
746 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
747 struct inode_fs_paths *ipath = NULL;
748 struct btrfs_root *local_root;
749 struct btrfs_key root_key;
750 struct btrfs_key key;
752 root_key.objectid = root;
753 root_key.type = BTRFS_ROOT_ITEM_KEY;
754 root_key.offset = (u64)-1;
755 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
756 if (IS_ERR(local_root)) {
757 ret = PTR_ERR(local_root);
762 * this makes the path point to (inum INODE_ITEM ioff)
765 key.type = BTRFS_INODE_ITEM_KEY;
768 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
770 btrfs_release_path(swarn->path);
774 eb = swarn->path->nodes[0];
775 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
776 struct btrfs_inode_item);
777 isize = btrfs_inode_size(eb, inode_item);
778 nlink = btrfs_inode_nlink(eb, inode_item);
779 btrfs_release_path(swarn->path);
782 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
783 * uses GFP_NOFS in this context, so we keep it consistent but it does
784 * not seem to be strictly necessary.
786 nofs_flag = memalloc_nofs_save();
787 ipath = init_ipath(4096, local_root, swarn->path);
788 memalloc_nofs_restore(nofs_flag);
790 ret = PTR_ERR(ipath);
794 ret = paths_from_inode(inum, ipath);
800 * we deliberately ignore the bit ipath might have been too small to
801 * hold all of the paths here
803 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
804 btrfs_warn_in_rcu(fs_info,
805 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
806 swarn->errstr, swarn->logical,
807 rcu_str_deref(swarn->dev->name),
810 min(isize - offset, (u64)PAGE_SIZE), nlink,
811 (char *)(unsigned long)ipath->fspath->val[i]);
817 btrfs_warn_in_rcu(fs_info,
818 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
819 swarn->errstr, swarn->logical,
820 rcu_str_deref(swarn->dev->name),
822 root, inum, offset, ret);
828 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
830 struct btrfs_device *dev;
831 struct btrfs_fs_info *fs_info;
832 struct btrfs_path *path;
833 struct btrfs_key found_key;
834 struct extent_buffer *eb;
835 struct btrfs_extent_item *ei;
836 struct scrub_warning swarn;
837 unsigned long ptr = 0;
845 WARN_ON(sblock->page_count < 1);
846 dev = sblock->pagev[0]->dev;
847 fs_info = sblock->sctx->fs_info;
849 path = btrfs_alloc_path();
853 swarn.physical = sblock->pagev[0]->physical;
854 swarn.logical = sblock->pagev[0]->logical;
855 swarn.errstr = errstr;
858 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
863 extent_item_pos = swarn.logical - found_key.objectid;
864 swarn.extent_item_size = found_key.offset;
867 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
868 item_size = btrfs_item_size_nr(eb, path->slots[0]);
870 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
872 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
873 item_size, &ref_root,
875 btrfs_warn_in_rcu(fs_info,
876 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
877 errstr, swarn.logical,
878 rcu_str_deref(dev->name),
880 ref_level ? "node" : "leaf",
881 ret < 0 ? -1 : ref_level,
882 ret < 0 ? -1 : ref_root);
884 btrfs_release_path(path);
886 btrfs_release_path(path);
889 iterate_extent_inodes(fs_info, found_key.objectid,
891 scrub_print_warning_inode, &swarn, false);
895 btrfs_free_path(path);
898 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
900 struct page *page = NULL;
902 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
905 struct btrfs_key key;
906 struct inode *inode = NULL;
907 struct btrfs_fs_info *fs_info;
908 u64 end = offset + PAGE_SIZE - 1;
909 struct btrfs_root *local_root;
913 key.type = BTRFS_ROOT_ITEM_KEY;
914 key.offset = (u64)-1;
916 fs_info = fixup->root->fs_info;
917 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
919 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
920 if (IS_ERR(local_root)) {
921 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
922 return PTR_ERR(local_root);
925 key.type = BTRFS_INODE_ITEM_KEY;
928 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
929 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
931 return PTR_ERR(inode);
933 index = offset >> PAGE_SHIFT;
935 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
941 if (PageUptodate(page)) {
942 if (PageDirty(page)) {
944 * we need to write the data to the defect sector. the
945 * data that was in that sector is not in memory,
946 * because the page was modified. we must not write the
947 * modified page to that sector.
949 * TODO: what could be done here: wait for the delalloc
950 * runner to write out that page (might involve
951 * COW) and see whether the sector is still
952 * referenced afterwards.
954 * For the meantime, we'll treat this error
955 * incorrectable, although there is a chance that a
956 * later scrub will find the bad sector again and that
957 * there's no dirty page in memory, then.
962 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
963 fixup->logical, page,
964 offset - page_offset(page),
970 * we need to get good data first. the general readpage path
971 * will call repair_io_failure for us, we just have to make
972 * sure we read the bad mirror.
974 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
977 /* set_extent_bits should give proper error */
984 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
987 wait_on_page_locked(page);
989 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
990 end, EXTENT_DAMAGED, 0, NULL);
992 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
1005 if (ret == 0 && corrected) {
1007 * we only need to call readpage for one of the inodes belonging
1008 * to this extent. so make iterate_extent_inodes stop
1016 static void scrub_fixup_nodatasum(struct btrfs_work *work)
1018 struct btrfs_fs_info *fs_info;
1020 struct scrub_fixup_nodatasum *fixup;
1021 struct scrub_ctx *sctx;
1022 struct btrfs_trans_handle *trans = NULL;
1023 struct btrfs_path *path;
1024 int uncorrectable = 0;
1026 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1028 fs_info = fixup->root->fs_info;
1030 path = btrfs_alloc_path();
1032 spin_lock(&sctx->stat_lock);
1033 ++sctx->stat.malloc_errors;
1034 spin_unlock(&sctx->stat_lock);
1039 trans = btrfs_join_transaction(fixup->root);
1040 if (IS_ERR(trans)) {
1046 * the idea is to trigger a regular read through the standard path. we
1047 * read a page from the (failed) logical address by specifying the
1048 * corresponding copynum of the failed sector. thus, that readpage is
1050 * that is the point where on-the-fly error correction will kick in
1051 * (once it's finished) and rewrite the failed sector if a good copy
1054 ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1055 scrub_fixup_readpage, fixup, false);
1062 spin_lock(&sctx->stat_lock);
1063 ++sctx->stat.corrected_errors;
1064 spin_unlock(&sctx->stat_lock);
1067 if (trans && !IS_ERR(trans))
1068 btrfs_end_transaction(trans);
1069 if (uncorrectable) {
1070 spin_lock(&sctx->stat_lock);
1071 ++sctx->stat.uncorrectable_errors;
1072 spin_unlock(&sctx->stat_lock);
1073 btrfs_dev_replace_stats_inc(
1074 &fs_info->dev_replace.num_uncorrectable_read_errors);
1075 btrfs_err_rl_in_rcu(fs_info,
1076 "unable to fixup (nodatasum) error at logical %llu on dev %s",
1077 fixup->logical, rcu_str_deref(fixup->dev->name));
1080 btrfs_free_path(path);
1083 scrub_pending_trans_workers_dec(sctx);
1086 static inline void scrub_get_recover(struct scrub_recover *recover)
1088 refcount_inc(&recover->refs);
1091 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1092 struct scrub_recover *recover)
1094 if (refcount_dec_and_test(&recover->refs)) {
1095 btrfs_bio_counter_dec(fs_info);
1096 btrfs_put_bbio(recover->bbio);
1102 * scrub_handle_errored_block gets called when either verification of the
1103 * pages failed or the bio failed to read, e.g. with EIO. In the latter
1104 * case, this function handles all pages in the bio, even though only one
1106 * The goal of this function is to repair the errored block by using the
1107 * contents of one of the mirrors.
1109 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1111 struct scrub_ctx *sctx = sblock_to_check->sctx;
1112 struct btrfs_device *dev;
1113 struct btrfs_fs_info *fs_info;
1115 unsigned int failed_mirror_index;
1116 unsigned int is_metadata;
1117 unsigned int have_csum;
1118 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1119 struct scrub_block *sblock_bad;
1124 bool full_stripe_locked;
1125 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1126 DEFAULT_RATELIMIT_BURST);
1128 BUG_ON(sblock_to_check->page_count < 1);
1129 fs_info = sctx->fs_info;
1130 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1132 * if we find an error in a super block, we just report it.
1133 * They will get written with the next transaction commit
1136 spin_lock(&sctx->stat_lock);
1137 ++sctx->stat.super_errors;
1138 spin_unlock(&sctx->stat_lock);
1141 logical = sblock_to_check->pagev[0]->logical;
1142 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1143 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1144 is_metadata = !(sblock_to_check->pagev[0]->flags &
1145 BTRFS_EXTENT_FLAG_DATA);
1146 have_csum = sblock_to_check->pagev[0]->have_csum;
1147 dev = sblock_to_check->pagev[0]->dev;
1150 * For RAID5/6, race can happen for a different device scrub thread.
1151 * For data corruption, Parity and Data threads will both try
1152 * to recovery the data.
1153 * Race can lead to doubly added csum error, or even unrecoverable
1156 ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1158 spin_lock(&sctx->stat_lock);
1160 sctx->stat.malloc_errors++;
1161 sctx->stat.read_errors++;
1162 sctx->stat.uncorrectable_errors++;
1163 spin_unlock(&sctx->stat_lock);
1167 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
1168 sblocks_for_recheck = NULL;
1169 goto nodatasum_case;
1173 * read all mirrors one after the other. This includes to
1174 * re-read the extent or metadata block that failed (that was
1175 * the cause that this fixup code is called) another time,
1176 * page by page this time in order to know which pages
1177 * caused I/O errors and which ones are good (for all mirrors).
1178 * It is the goal to handle the situation when more than one
1179 * mirror contains I/O errors, but the errors do not
1180 * overlap, i.e. the data can be repaired by selecting the
1181 * pages from those mirrors without I/O error on the
1182 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1183 * would be that mirror #1 has an I/O error on the first page,
1184 * the second page is good, and mirror #2 has an I/O error on
1185 * the second page, but the first page is good.
1186 * Then the first page of the first mirror can be repaired by
1187 * taking the first page of the second mirror, and the
1188 * second page of the second mirror can be repaired by
1189 * copying the contents of the 2nd page of the 1st mirror.
1190 * One more note: if the pages of one mirror contain I/O
1191 * errors, the checksum cannot be verified. In order to get
1192 * the best data for repairing, the first attempt is to find
1193 * a mirror without I/O errors and with a validated checksum.
1194 * Only if this is not possible, the pages are picked from
1195 * mirrors with I/O errors without considering the checksum.
1196 * If the latter is the case, at the end, the checksum of the
1197 * repaired area is verified in order to correctly maintain
1201 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1202 sizeof(*sblocks_for_recheck), GFP_NOFS);
1203 if (!sblocks_for_recheck) {
1204 spin_lock(&sctx->stat_lock);
1205 sctx->stat.malloc_errors++;
1206 sctx->stat.read_errors++;
1207 sctx->stat.uncorrectable_errors++;
1208 spin_unlock(&sctx->stat_lock);
1209 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1213 /* setup the context, map the logical blocks and alloc the pages */
1214 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1216 spin_lock(&sctx->stat_lock);
1217 sctx->stat.read_errors++;
1218 sctx->stat.uncorrectable_errors++;
1219 spin_unlock(&sctx->stat_lock);
1220 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1223 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1224 sblock_bad = sblocks_for_recheck + failed_mirror_index;
1226 /* build and submit the bios for the failed mirror, check checksums */
1227 scrub_recheck_block(fs_info, sblock_bad, 1);
1229 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1230 sblock_bad->no_io_error_seen) {
1232 * the error disappeared after reading page by page, or
1233 * the area was part of a huge bio and other parts of the
1234 * bio caused I/O errors, or the block layer merged several
1235 * read requests into one and the error is caused by a
1236 * different bio (usually one of the two latter cases is
1239 spin_lock(&sctx->stat_lock);
1240 sctx->stat.unverified_errors++;
1241 sblock_to_check->data_corrected = 1;
1242 spin_unlock(&sctx->stat_lock);
1244 if (sctx->is_dev_replace)
1245 scrub_write_block_to_dev_replace(sblock_bad);
1249 if (!sblock_bad->no_io_error_seen) {
1250 spin_lock(&sctx->stat_lock);
1251 sctx->stat.read_errors++;
1252 spin_unlock(&sctx->stat_lock);
1253 if (__ratelimit(&_rs))
1254 scrub_print_warning("i/o error", sblock_to_check);
1255 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1256 } else if (sblock_bad->checksum_error) {
1257 spin_lock(&sctx->stat_lock);
1258 sctx->stat.csum_errors++;
1259 spin_unlock(&sctx->stat_lock);
1260 if (__ratelimit(&_rs))
1261 scrub_print_warning("checksum error", sblock_to_check);
1262 btrfs_dev_stat_inc_and_print(dev,
1263 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1264 } else if (sblock_bad->header_error) {
1265 spin_lock(&sctx->stat_lock);
1266 sctx->stat.verify_errors++;
1267 spin_unlock(&sctx->stat_lock);
1268 if (__ratelimit(&_rs))
1269 scrub_print_warning("checksum/header error",
1271 if (sblock_bad->generation_error)
1272 btrfs_dev_stat_inc_and_print(dev,
1273 BTRFS_DEV_STAT_GENERATION_ERRS);
1275 btrfs_dev_stat_inc_and_print(dev,
1276 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1279 if (sctx->readonly) {
1280 ASSERT(!sctx->is_dev_replace);
1284 if (!is_metadata && !have_csum) {
1285 struct scrub_fixup_nodatasum *fixup_nodatasum;
1287 WARN_ON(sctx->is_dev_replace);
1292 * !is_metadata and !have_csum, this means that the data
1293 * might not be COWed, that it might be modified
1294 * concurrently. The general strategy to work on the
1295 * commit root does not help in the case when COW is not
1298 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1299 if (!fixup_nodatasum)
1300 goto did_not_correct_error;
1301 fixup_nodatasum->sctx = sctx;
1302 fixup_nodatasum->dev = dev;
1303 fixup_nodatasum->logical = logical;
1304 fixup_nodatasum->root = fs_info->extent_root;
1305 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1306 scrub_pending_trans_workers_inc(sctx);
1307 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1308 scrub_fixup_nodatasum, NULL, NULL);
1309 btrfs_queue_work(fs_info->scrub_workers,
1310 &fixup_nodatasum->work);
1315 * now build and submit the bios for the other mirrors, check
1317 * First try to pick the mirror which is completely without I/O
1318 * errors and also does not have a checksum error.
1319 * If one is found, and if a checksum is present, the full block
1320 * that is known to contain an error is rewritten. Afterwards
1321 * the block is known to be corrected.
1322 * If a mirror is found which is completely correct, and no
1323 * checksum is present, only those pages are rewritten that had
1324 * an I/O error in the block to be repaired, since it cannot be
1325 * determined, which copy of the other pages is better (and it
1326 * could happen otherwise that a correct page would be
1327 * overwritten by a bad one).
1329 for (mirror_index = 0; ;mirror_index++) {
1330 struct scrub_block *sblock_other;
1332 if (mirror_index == failed_mirror_index)
1335 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1336 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1337 if (mirror_index >= BTRFS_MAX_MIRRORS)
1339 if (!sblocks_for_recheck[mirror_index].page_count)
1342 sblock_other = sblocks_for_recheck + mirror_index;
1344 struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1345 int max_allowed = r->bbio->num_stripes -
1346 r->bbio->num_tgtdevs;
1348 if (mirror_index >= max_allowed)
1350 if (!sblocks_for_recheck[1].page_count)
1353 ASSERT(failed_mirror_index == 0);
1354 sblock_other = sblocks_for_recheck + 1;
1355 sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1358 /* build and submit the bios, check checksums */
1359 scrub_recheck_block(fs_info, sblock_other, 0);
1361 if (!sblock_other->header_error &&
1362 !sblock_other->checksum_error &&
1363 sblock_other->no_io_error_seen) {
1364 if (sctx->is_dev_replace) {
1365 scrub_write_block_to_dev_replace(sblock_other);
1366 goto corrected_error;
1368 ret = scrub_repair_block_from_good_copy(
1369 sblock_bad, sblock_other);
1371 goto corrected_error;
1376 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1377 goto did_not_correct_error;
1380 * In case of I/O errors in the area that is supposed to be
1381 * repaired, continue by picking good copies of those pages.
1382 * Select the good pages from mirrors to rewrite bad pages from
1383 * the area to fix. Afterwards verify the checksum of the block
1384 * that is supposed to be repaired. This verification step is
1385 * only done for the purpose of statistic counting and for the
1386 * final scrub report, whether errors remain.
1387 * A perfect algorithm could make use of the checksum and try
1388 * all possible combinations of pages from the different mirrors
1389 * until the checksum verification succeeds. For example, when
1390 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1391 * of mirror #2 is readable but the final checksum test fails,
1392 * then the 2nd page of mirror #3 could be tried, whether now
1393 * the final checksum succeeds. But this would be a rare
1394 * exception and is therefore not implemented. At least it is
1395 * avoided that the good copy is overwritten.
1396 * A more useful improvement would be to pick the sectors
1397 * without I/O error based on sector sizes (512 bytes on legacy
1398 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1399 * mirror could be repaired by taking 512 byte of a different
1400 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1401 * area are unreadable.
1404 for (page_num = 0; page_num < sblock_bad->page_count;
1406 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1407 struct scrub_block *sblock_other = NULL;
1409 /* skip no-io-error page in scrub */
1410 if (!page_bad->io_error && !sctx->is_dev_replace)
1413 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1415 * In case of dev replace, if raid56 rebuild process
1416 * didn't work out correct data, then copy the content
1417 * in sblock_bad to make sure target device is identical
1418 * to source device, instead of writing garbage data in
1419 * sblock_for_recheck array to target device.
1421 sblock_other = NULL;
1422 } else if (page_bad->io_error) {
1423 /* try to find no-io-error page in mirrors */
1424 for (mirror_index = 0;
1425 mirror_index < BTRFS_MAX_MIRRORS &&
1426 sblocks_for_recheck[mirror_index].page_count > 0;
1428 if (!sblocks_for_recheck[mirror_index].
1429 pagev[page_num]->io_error) {
1430 sblock_other = sblocks_for_recheck +
1439 if (sctx->is_dev_replace) {
1441 * did not find a mirror to fetch the page
1442 * from. scrub_write_page_to_dev_replace()
1443 * handles this case (page->io_error), by
1444 * filling the block with zeros before
1445 * submitting the write request
1448 sblock_other = sblock_bad;
1450 if (scrub_write_page_to_dev_replace(sblock_other,
1452 btrfs_dev_replace_stats_inc(
1453 &fs_info->dev_replace.num_write_errors);
1456 } else if (sblock_other) {
1457 ret = scrub_repair_page_from_good_copy(sblock_bad,
1461 page_bad->io_error = 0;
1467 if (success && !sctx->is_dev_replace) {
1468 if (is_metadata || have_csum) {
1470 * need to verify the checksum now that all
1471 * sectors on disk are repaired (the write
1472 * request for data to be repaired is on its way).
1473 * Just be lazy and use scrub_recheck_block()
1474 * which re-reads the data before the checksum
1475 * is verified, but most likely the data comes out
1476 * of the page cache.
1478 scrub_recheck_block(fs_info, sblock_bad, 1);
1479 if (!sblock_bad->header_error &&
1480 !sblock_bad->checksum_error &&
1481 sblock_bad->no_io_error_seen)
1482 goto corrected_error;
1484 goto did_not_correct_error;
1487 spin_lock(&sctx->stat_lock);
1488 sctx->stat.corrected_errors++;
1489 sblock_to_check->data_corrected = 1;
1490 spin_unlock(&sctx->stat_lock);
1491 btrfs_err_rl_in_rcu(fs_info,
1492 "fixed up error at logical %llu on dev %s",
1493 logical, rcu_str_deref(dev->name));
1496 did_not_correct_error:
1497 spin_lock(&sctx->stat_lock);
1498 sctx->stat.uncorrectable_errors++;
1499 spin_unlock(&sctx->stat_lock);
1500 btrfs_err_rl_in_rcu(fs_info,
1501 "unable to fixup (regular) error at logical %llu on dev %s",
1502 logical, rcu_str_deref(dev->name));
1506 if (sblocks_for_recheck) {
1507 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1509 struct scrub_block *sblock = sblocks_for_recheck +
1511 struct scrub_recover *recover;
1514 for (page_index = 0; page_index < sblock->page_count;
1516 sblock->pagev[page_index]->sblock = NULL;
1517 recover = sblock->pagev[page_index]->recover;
1519 scrub_put_recover(fs_info, recover);
1520 sblock->pagev[page_index]->recover =
1523 scrub_page_put(sblock->pagev[page_index]);
1526 kfree(sblocks_for_recheck);
1529 ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1535 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1537 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1539 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1542 return (int)bbio->num_stripes;
1545 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1548 int nstripes, int mirror,
1554 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1556 for (i = 0; i < nstripes; i++) {
1557 if (raid_map[i] == RAID6_Q_STRIPE ||
1558 raid_map[i] == RAID5_P_STRIPE)
1561 if (logical >= raid_map[i] &&
1562 logical < raid_map[i] + mapped_length)
1567 *stripe_offset = logical - raid_map[i];
1569 /* The other RAID type */
1570 *stripe_index = mirror;
1575 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1576 struct scrub_block *sblocks_for_recheck)
1578 struct scrub_ctx *sctx = original_sblock->sctx;
1579 struct btrfs_fs_info *fs_info = sctx->fs_info;
1580 u64 length = original_sblock->page_count * PAGE_SIZE;
1581 u64 logical = original_sblock->pagev[0]->logical;
1582 u64 generation = original_sblock->pagev[0]->generation;
1583 u64 flags = original_sblock->pagev[0]->flags;
1584 u64 have_csum = original_sblock->pagev[0]->have_csum;
1585 struct scrub_recover *recover;
1586 struct btrfs_bio *bbio;
1597 * note: the two members refs and outstanding_pages
1598 * are not used (and not set) in the blocks that are used for
1599 * the recheck procedure
1602 while (length > 0) {
1603 sublen = min_t(u64, length, PAGE_SIZE);
1604 mapped_length = sublen;
1608 * with a length of PAGE_SIZE, each returned stripe
1609 * represents one mirror
1611 btrfs_bio_counter_inc_blocked(fs_info);
1612 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1613 logical, &mapped_length, &bbio);
1614 if (ret || !bbio || mapped_length < sublen) {
1615 btrfs_put_bbio(bbio);
1616 btrfs_bio_counter_dec(fs_info);
1620 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1622 btrfs_put_bbio(bbio);
1623 btrfs_bio_counter_dec(fs_info);
1627 refcount_set(&recover->refs, 1);
1628 recover->bbio = bbio;
1629 recover->map_length = mapped_length;
1631 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1633 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1635 for (mirror_index = 0; mirror_index < nmirrors;
1637 struct scrub_block *sblock;
1638 struct scrub_page *page;
1640 sblock = sblocks_for_recheck + mirror_index;
1641 sblock->sctx = sctx;
1643 page = kzalloc(sizeof(*page), GFP_NOFS);
1646 spin_lock(&sctx->stat_lock);
1647 sctx->stat.malloc_errors++;
1648 spin_unlock(&sctx->stat_lock);
1649 scrub_put_recover(fs_info, recover);
1652 scrub_page_get(page);
1653 sblock->pagev[page_index] = page;
1654 page->sblock = sblock;
1655 page->flags = flags;
1656 page->generation = generation;
1657 page->logical = logical;
1658 page->have_csum = have_csum;
1661 original_sblock->pagev[0]->csum,
1664 scrub_stripe_index_and_offset(logical,
1673 page->physical = bbio->stripes[stripe_index].physical +
1675 page->dev = bbio->stripes[stripe_index].dev;
1677 BUG_ON(page_index >= original_sblock->page_count);
1678 page->physical_for_dev_replace =
1679 original_sblock->pagev[page_index]->
1680 physical_for_dev_replace;
1681 /* for missing devices, dev->bdev is NULL */
1682 page->mirror_num = mirror_index + 1;
1683 sblock->page_count++;
1684 page->page = alloc_page(GFP_NOFS);
1688 scrub_get_recover(recover);
1689 page->recover = recover;
1691 scrub_put_recover(fs_info, recover);
1700 static void scrub_bio_wait_endio(struct bio *bio)
1702 complete(bio->bi_private);
1705 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1707 struct scrub_page *page)
1709 DECLARE_COMPLETION_ONSTACK(done);
1713 bio->bi_iter.bi_sector = page->logical >> 9;
1714 bio->bi_private = &done;
1715 bio->bi_end_io = scrub_bio_wait_endio;
1717 mirror_num = page->sblock->pagev[0]->mirror_num;
1718 ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1719 page->recover->map_length,
1724 wait_for_completion_io(&done);
1725 return blk_status_to_errno(bio->bi_status);
1728 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1729 struct scrub_block *sblock)
1731 struct scrub_page *first_page = sblock->pagev[0];
1735 /* All pages in sblock belong to the same stripe on the same device. */
1736 ASSERT(first_page->dev);
1737 if (!first_page->dev->bdev)
1740 bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1741 bio_set_dev(bio, first_page->dev->bdev);
1743 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1744 struct scrub_page *page = sblock->pagev[page_num];
1746 WARN_ON(!page->page);
1747 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1750 if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1757 scrub_recheck_block_checksum(sblock);
1761 for (page_num = 0; page_num < sblock->page_count; page_num++)
1762 sblock->pagev[page_num]->io_error = 1;
1764 sblock->no_io_error_seen = 0;
1768 * this function will check the on disk data for checksum errors, header
1769 * errors and read I/O errors. If any I/O errors happen, the exact pages
1770 * which are errored are marked as being bad. The goal is to enable scrub
1771 * to take those pages that are not errored from all the mirrors so that
1772 * the pages that are errored in the just handled mirror can be repaired.
1774 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1775 struct scrub_block *sblock,
1776 int retry_failed_mirror)
1780 sblock->no_io_error_seen = 1;
1782 /* short cut for raid56 */
1783 if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1784 return scrub_recheck_block_on_raid56(fs_info, sblock);
1786 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1788 struct scrub_page *page = sblock->pagev[page_num];
1790 if (page->dev->bdev == NULL) {
1792 sblock->no_io_error_seen = 0;
1796 WARN_ON(!page->page);
1797 bio = btrfs_io_bio_alloc(1);
1798 bio_set_dev(bio, page->dev->bdev);
1800 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1801 bio->bi_iter.bi_sector = page->physical >> 9;
1802 bio->bi_opf = REQ_OP_READ;
1804 if (btrfsic_submit_bio_wait(bio)) {
1806 sblock->no_io_error_seen = 0;
1812 if (sblock->no_io_error_seen)
1813 scrub_recheck_block_checksum(sblock);
1816 static inline int scrub_check_fsid(u8 fsid[],
1817 struct scrub_page *spage)
1819 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1822 ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1826 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1828 sblock->header_error = 0;
1829 sblock->checksum_error = 0;
1830 sblock->generation_error = 0;
1832 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1833 scrub_checksum_data(sblock);
1835 scrub_checksum_tree_block(sblock);
1838 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1839 struct scrub_block *sblock_good)
1844 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1847 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1857 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1858 struct scrub_block *sblock_good,
1859 int page_num, int force_write)
1861 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1862 struct scrub_page *page_good = sblock_good->pagev[page_num];
1863 struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1865 BUG_ON(page_bad->page == NULL);
1866 BUG_ON(page_good->page == NULL);
1867 if (force_write || sblock_bad->header_error ||
1868 sblock_bad->checksum_error || page_bad->io_error) {
1872 if (!page_bad->dev->bdev) {
1873 btrfs_warn_rl(fs_info,
1874 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1878 bio = btrfs_io_bio_alloc(1);
1879 bio_set_dev(bio, page_bad->dev->bdev);
1880 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1881 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1883 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1884 if (PAGE_SIZE != ret) {
1889 if (btrfsic_submit_bio_wait(bio)) {
1890 btrfs_dev_stat_inc_and_print(page_bad->dev,
1891 BTRFS_DEV_STAT_WRITE_ERRS);
1892 btrfs_dev_replace_stats_inc(
1893 &fs_info->dev_replace.num_write_errors);
1903 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1905 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1909 * This block is used for the check of the parity on the source device,
1910 * so the data needn't be written into the destination device.
1912 if (sblock->sparity)
1915 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1918 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1920 btrfs_dev_replace_stats_inc(
1921 &fs_info->dev_replace.num_write_errors);
1925 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1928 struct scrub_page *spage = sblock->pagev[page_num];
1930 BUG_ON(spage->page == NULL);
1931 if (spage->io_error) {
1932 void *mapped_buffer = kmap_atomic(spage->page);
1934 clear_page(mapped_buffer);
1935 flush_dcache_page(spage->page);
1936 kunmap_atomic(mapped_buffer);
1938 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1941 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1942 struct scrub_page *spage)
1944 struct scrub_bio *sbio;
1947 mutex_lock(&sctx->wr_lock);
1949 if (!sctx->wr_curr_bio) {
1950 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1952 if (!sctx->wr_curr_bio) {
1953 mutex_unlock(&sctx->wr_lock);
1956 sctx->wr_curr_bio->sctx = sctx;
1957 sctx->wr_curr_bio->page_count = 0;
1959 sbio = sctx->wr_curr_bio;
1960 if (sbio->page_count == 0) {
1963 sbio->physical = spage->physical_for_dev_replace;
1964 sbio->logical = spage->logical;
1965 sbio->dev = sctx->wr_tgtdev;
1968 bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1972 bio->bi_private = sbio;
1973 bio->bi_end_io = scrub_wr_bio_end_io;
1974 bio_set_dev(bio, sbio->dev->bdev);
1975 bio->bi_iter.bi_sector = sbio->physical >> 9;
1976 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1978 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1979 spage->physical_for_dev_replace ||
1980 sbio->logical + sbio->page_count * PAGE_SIZE !=
1982 scrub_wr_submit(sctx);
1986 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1987 if (ret != PAGE_SIZE) {
1988 if (sbio->page_count < 1) {
1991 mutex_unlock(&sctx->wr_lock);
1994 scrub_wr_submit(sctx);
1998 sbio->pagev[sbio->page_count] = spage;
1999 scrub_page_get(spage);
2001 if (sbio->page_count == sctx->pages_per_wr_bio)
2002 scrub_wr_submit(sctx);
2003 mutex_unlock(&sctx->wr_lock);
2008 static void scrub_wr_submit(struct scrub_ctx *sctx)
2010 struct scrub_bio *sbio;
2012 if (!sctx->wr_curr_bio)
2015 sbio = sctx->wr_curr_bio;
2016 sctx->wr_curr_bio = NULL;
2017 WARN_ON(!sbio->bio->bi_disk);
2018 scrub_pending_bio_inc(sctx);
2019 /* process all writes in a single worker thread. Then the block layer
2020 * orders the requests before sending them to the driver which
2021 * doubled the write performance on spinning disks when measured
2023 btrfsic_submit_bio(sbio->bio);
2026 static void scrub_wr_bio_end_io(struct bio *bio)
2028 struct scrub_bio *sbio = bio->bi_private;
2029 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2031 sbio->status = bio->bi_status;
2034 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2035 scrub_wr_bio_end_io_worker, NULL, NULL);
2036 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2039 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2041 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2042 struct scrub_ctx *sctx = sbio->sctx;
2045 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2047 struct btrfs_dev_replace *dev_replace =
2048 &sbio->sctx->fs_info->dev_replace;
2050 for (i = 0; i < sbio->page_count; i++) {
2051 struct scrub_page *spage = sbio->pagev[i];
2053 spage->io_error = 1;
2054 btrfs_dev_replace_stats_inc(&dev_replace->
2059 for (i = 0; i < sbio->page_count; i++)
2060 scrub_page_put(sbio->pagev[i]);
2064 scrub_pending_bio_dec(sctx);
2067 static int scrub_checksum(struct scrub_block *sblock)
2073 * No need to initialize these stats currently,
2074 * because this function only use return value
2075 * instead of these stats value.
2080 sblock->header_error = 0;
2081 sblock->generation_error = 0;
2082 sblock->checksum_error = 0;
2084 WARN_ON(sblock->page_count < 1);
2085 flags = sblock->pagev[0]->flags;
2087 if (flags & BTRFS_EXTENT_FLAG_DATA)
2088 ret = scrub_checksum_data(sblock);
2089 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2090 ret = scrub_checksum_tree_block(sblock);
2091 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2092 (void)scrub_checksum_super(sblock);
2096 scrub_handle_errored_block(sblock);
2101 static int scrub_checksum_data(struct scrub_block *sblock)
2103 struct scrub_ctx *sctx = sblock->sctx;
2104 u8 csum[BTRFS_CSUM_SIZE];
2112 BUG_ON(sblock->page_count < 1);
2113 if (!sblock->pagev[0]->have_csum)
2116 on_disk_csum = sblock->pagev[0]->csum;
2117 page = sblock->pagev[0]->page;
2118 buffer = kmap_atomic(page);
2120 len = sctx->fs_info->sectorsize;
2123 u64 l = min_t(u64, len, PAGE_SIZE);
2125 crc = btrfs_csum_data(buffer, crc, l);
2126 kunmap_atomic(buffer);
2131 BUG_ON(index >= sblock->page_count);
2132 BUG_ON(!sblock->pagev[index]->page);
2133 page = sblock->pagev[index]->page;
2134 buffer = kmap_atomic(page);
2137 btrfs_csum_final(crc, csum);
2138 if (memcmp(csum, on_disk_csum, sctx->csum_size))
2139 sblock->checksum_error = 1;
2141 return sblock->checksum_error;
2144 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2146 struct scrub_ctx *sctx = sblock->sctx;
2147 struct btrfs_header *h;
2148 struct btrfs_fs_info *fs_info = sctx->fs_info;
2149 u8 calculated_csum[BTRFS_CSUM_SIZE];
2150 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2152 void *mapped_buffer;
2159 BUG_ON(sblock->page_count < 1);
2160 page = sblock->pagev[0]->page;
2161 mapped_buffer = kmap_atomic(page);
2162 h = (struct btrfs_header *)mapped_buffer;
2163 memcpy(on_disk_csum, h->csum, sctx->csum_size);
2166 * we don't use the getter functions here, as we
2167 * a) don't have an extent buffer and
2168 * b) the page is already kmapped
2170 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2171 sblock->header_error = 1;
2173 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2174 sblock->header_error = 1;
2175 sblock->generation_error = 1;
2178 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2179 sblock->header_error = 1;
2181 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2183 sblock->header_error = 1;
2185 len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2186 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2187 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2190 u64 l = min_t(u64, len, mapped_size);
2192 crc = btrfs_csum_data(p, crc, l);
2193 kunmap_atomic(mapped_buffer);
2198 BUG_ON(index >= sblock->page_count);
2199 BUG_ON(!sblock->pagev[index]->page);
2200 page = sblock->pagev[index]->page;
2201 mapped_buffer = kmap_atomic(page);
2202 mapped_size = PAGE_SIZE;
2206 btrfs_csum_final(crc, calculated_csum);
2207 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2208 sblock->checksum_error = 1;
2210 return sblock->header_error || sblock->checksum_error;
2213 static int scrub_checksum_super(struct scrub_block *sblock)
2215 struct btrfs_super_block *s;
2216 struct scrub_ctx *sctx = sblock->sctx;
2217 u8 calculated_csum[BTRFS_CSUM_SIZE];
2218 u8 on_disk_csum[BTRFS_CSUM_SIZE];
2220 void *mapped_buffer;
2229 BUG_ON(sblock->page_count < 1);
2230 page = sblock->pagev[0]->page;
2231 mapped_buffer = kmap_atomic(page);
2232 s = (struct btrfs_super_block *)mapped_buffer;
2233 memcpy(on_disk_csum, s->csum, sctx->csum_size);
2235 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2238 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2241 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2244 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2245 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2246 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2249 u64 l = min_t(u64, len, mapped_size);
2251 crc = btrfs_csum_data(p, crc, l);
2252 kunmap_atomic(mapped_buffer);
2257 BUG_ON(index >= sblock->page_count);
2258 BUG_ON(!sblock->pagev[index]->page);
2259 page = sblock->pagev[index]->page;
2260 mapped_buffer = kmap_atomic(page);
2261 mapped_size = PAGE_SIZE;
2265 btrfs_csum_final(crc, calculated_csum);
2266 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2269 if (fail_cor + fail_gen) {
2271 * if we find an error in a super block, we just report it.
2272 * They will get written with the next transaction commit
2275 spin_lock(&sctx->stat_lock);
2276 ++sctx->stat.super_errors;
2277 spin_unlock(&sctx->stat_lock);
2279 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2280 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2282 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2283 BTRFS_DEV_STAT_GENERATION_ERRS);
2286 return fail_cor + fail_gen;
2289 static void scrub_block_get(struct scrub_block *sblock)
2291 refcount_inc(&sblock->refs);
2294 static void scrub_block_put(struct scrub_block *sblock)
2296 if (refcount_dec_and_test(&sblock->refs)) {
2299 if (sblock->sparity)
2300 scrub_parity_put(sblock->sparity);
2302 for (i = 0; i < sblock->page_count; i++)
2303 scrub_page_put(sblock->pagev[i]);
2308 static void scrub_page_get(struct scrub_page *spage)
2310 atomic_inc(&spage->refs);
2313 static void scrub_page_put(struct scrub_page *spage)
2315 if (atomic_dec_and_test(&spage->refs)) {
2317 __free_page(spage->page);
2322 static void scrub_submit(struct scrub_ctx *sctx)
2324 struct scrub_bio *sbio;
2326 if (sctx->curr == -1)
2329 sbio = sctx->bios[sctx->curr];
2331 scrub_pending_bio_inc(sctx);
2332 btrfsic_submit_bio(sbio->bio);
2335 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2336 struct scrub_page *spage)
2338 struct scrub_block *sblock = spage->sblock;
2339 struct scrub_bio *sbio;
2344 * grab a fresh bio or wait for one to become available
2346 while (sctx->curr == -1) {
2347 spin_lock(&sctx->list_lock);
2348 sctx->curr = sctx->first_free;
2349 if (sctx->curr != -1) {
2350 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2351 sctx->bios[sctx->curr]->next_free = -1;
2352 sctx->bios[sctx->curr]->page_count = 0;
2353 spin_unlock(&sctx->list_lock);
2355 spin_unlock(&sctx->list_lock);
2356 wait_event(sctx->list_wait, sctx->first_free != -1);
2359 sbio = sctx->bios[sctx->curr];
2360 if (sbio->page_count == 0) {
2363 sbio->physical = spage->physical;
2364 sbio->logical = spage->logical;
2365 sbio->dev = spage->dev;
2368 bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2372 bio->bi_private = sbio;
2373 bio->bi_end_io = scrub_bio_end_io;
2374 bio_set_dev(bio, sbio->dev->bdev);
2375 bio->bi_iter.bi_sector = sbio->physical >> 9;
2376 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2378 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2380 sbio->logical + sbio->page_count * PAGE_SIZE !=
2382 sbio->dev != spage->dev) {
2387 sbio->pagev[sbio->page_count] = spage;
2388 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2389 if (ret != PAGE_SIZE) {
2390 if (sbio->page_count < 1) {
2399 scrub_block_get(sblock); /* one for the page added to the bio */
2400 atomic_inc(&sblock->outstanding_pages);
2402 if (sbio->page_count == sctx->pages_per_rd_bio)
2408 static void scrub_missing_raid56_end_io(struct bio *bio)
2410 struct scrub_block *sblock = bio->bi_private;
2411 struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2414 sblock->no_io_error_seen = 0;
2418 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2421 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2423 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2424 struct scrub_ctx *sctx = sblock->sctx;
2425 struct btrfs_fs_info *fs_info = sctx->fs_info;
2427 struct btrfs_device *dev;
2429 logical = sblock->pagev[0]->logical;
2430 dev = sblock->pagev[0]->dev;
2432 if (sblock->no_io_error_seen)
2433 scrub_recheck_block_checksum(sblock);
2435 if (!sblock->no_io_error_seen) {
2436 spin_lock(&sctx->stat_lock);
2437 sctx->stat.read_errors++;
2438 spin_unlock(&sctx->stat_lock);
2439 btrfs_err_rl_in_rcu(fs_info,
2440 "IO error rebuilding logical %llu for dev %s",
2441 logical, rcu_str_deref(dev->name));
2442 } else if (sblock->header_error || sblock->checksum_error) {
2443 spin_lock(&sctx->stat_lock);
2444 sctx->stat.uncorrectable_errors++;
2445 spin_unlock(&sctx->stat_lock);
2446 btrfs_err_rl_in_rcu(fs_info,
2447 "failed to rebuild valid logical %llu for dev %s",
2448 logical, rcu_str_deref(dev->name));
2450 scrub_write_block_to_dev_replace(sblock);
2453 scrub_block_put(sblock);
2455 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2456 mutex_lock(&sctx->wr_lock);
2457 scrub_wr_submit(sctx);
2458 mutex_unlock(&sctx->wr_lock);
2461 scrub_pending_bio_dec(sctx);
2464 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2466 struct scrub_ctx *sctx = sblock->sctx;
2467 struct btrfs_fs_info *fs_info = sctx->fs_info;
2468 u64 length = sblock->page_count * PAGE_SIZE;
2469 u64 logical = sblock->pagev[0]->logical;
2470 struct btrfs_bio *bbio = NULL;
2472 struct btrfs_raid_bio *rbio;
2476 btrfs_bio_counter_inc_blocked(fs_info);
2477 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2479 if (ret || !bbio || !bbio->raid_map)
2482 if (WARN_ON(!sctx->is_dev_replace ||
2483 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2485 * We shouldn't be scrubbing a missing device. Even for dev
2486 * replace, we should only get here for RAID 5/6. We either
2487 * managed to mount something with no mirrors remaining or
2488 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2493 bio = btrfs_io_bio_alloc(0);
2494 bio->bi_iter.bi_sector = logical >> 9;
2495 bio->bi_private = sblock;
2496 bio->bi_end_io = scrub_missing_raid56_end_io;
2498 rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2502 for (i = 0; i < sblock->page_count; i++) {
2503 struct scrub_page *spage = sblock->pagev[i];
2505 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2508 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2509 scrub_missing_raid56_worker, NULL, NULL);
2510 scrub_block_get(sblock);
2511 scrub_pending_bio_inc(sctx);
2512 raid56_submit_missing_rbio(rbio);
2518 btrfs_bio_counter_dec(fs_info);
2519 btrfs_put_bbio(bbio);
2520 spin_lock(&sctx->stat_lock);
2521 sctx->stat.malloc_errors++;
2522 spin_unlock(&sctx->stat_lock);
2525 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2526 u64 physical, struct btrfs_device *dev, u64 flags,
2527 u64 gen, int mirror_num, u8 *csum, int force,
2528 u64 physical_for_dev_replace)
2530 struct scrub_block *sblock;
2533 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2535 spin_lock(&sctx->stat_lock);
2536 sctx->stat.malloc_errors++;
2537 spin_unlock(&sctx->stat_lock);
2541 /* one ref inside this function, plus one for each page added to
2543 refcount_set(&sblock->refs, 1);
2544 sblock->sctx = sctx;
2545 sblock->no_io_error_seen = 1;
2547 for (index = 0; len > 0; index++) {
2548 struct scrub_page *spage;
2549 u64 l = min_t(u64, len, PAGE_SIZE);
2551 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2554 spin_lock(&sctx->stat_lock);
2555 sctx->stat.malloc_errors++;
2556 spin_unlock(&sctx->stat_lock);
2557 scrub_block_put(sblock);
2560 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2561 scrub_page_get(spage);
2562 sblock->pagev[index] = spage;
2563 spage->sblock = sblock;
2565 spage->flags = flags;
2566 spage->generation = gen;
2567 spage->logical = logical;
2568 spage->physical = physical;
2569 spage->physical_for_dev_replace = physical_for_dev_replace;
2570 spage->mirror_num = mirror_num;
2572 spage->have_csum = 1;
2573 memcpy(spage->csum, csum, sctx->csum_size);
2575 spage->have_csum = 0;
2577 sblock->page_count++;
2578 spage->page = alloc_page(GFP_KERNEL);
2584 physical_for_dev_replace += l;
2587 WARN_ON(sblock->page_count == 0);
2588 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2590 * This case should only be hit for RAID 5/6 device replace. See
2591 * the comment in scrub_missing_raid56_pages() for details.
2593 scrub_missing_raid56_pages(sblock);
2595 for (index = 0; index < sblock->page_count; index++) {
2596 struct scrub_page *spage = sblock->pagev[index];
2599 ret = scrub_add_page_to_rd_bio(sctx, spage);
2601 scrub_block_put(sblock);
2610 /* last one frees, either here or in bio completion for last page */
2611 scrub_block_put(sblock);
2615 static void scrub_bio_end_io(struct bio *bio)
2617 struct scrub_bio *sbio = bio->bi_private;
2618 struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2620 sbio->status = bio->bi_status;
2623 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2626 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2628 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2629 struct scrub_ctx *sctx = sbio->sctx;
2632 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2634 for (i = 0; i < sbio->page_count; i++) {
2635 struct scrub_page *spage = sbio->pagev[i];
2637 spage->io_error = 1;
2638 spage->sblock->no_io_error_seen = 0;
2642 /* now complete the scrub_block items that have all pages completed */
2643 for (i = 0; i < sbio->page_count; i++) {
2644 struct scrub_page *spage = sbio->pagev[i];
2645 struct scrub_block *sblock = spage->sblock;
2647 if (atomic_dec_and_test(&sblock->outstanding_pages))
2648 scrub_block_complete(sblock);
2649 scrub_block_put(sblock);
2654 spin_lock(&sctx->list_lock);
2655 sbio->next_free = sctx->first_free;
2656 sctx->first_free = sbio->index;
2657 spin_unlock(&sctx->list_lock);
2659 if (sctx->is_dev_replace && sctx->flush_all_writes) {
2660 mutex_lock(&sctx->wr_lock);
2661 scrub_wr_submit(sctx);
2662 mutex_unlock(&sctx->wr_lock);
2665 scrub_pending_bio_dec(sctx);
2668 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2669 unsigned long *bitmap,
2675 int sectorsize = sparity->sctx->fs_info->sectorsize;
2677 if (len >= sparity->stripe_len) {
2678 bitmap_set(bitmap, 0, sparity->nsectors);
2682 start -= sparity->logic_start;
2683 start = div64_u64_rem(start, sparity->stripe_len, &offset);
2684 offset = div_u64(offset, sectorsize);
2685 nsectors64 = div_u64(len, sectorsize);
2687 ASSERT(nsectors64 < UINT_MAX);
2688 nsectors = (u32)nsectors64;
2690 if (offset + nsectors <= sparity->nsectors) {
2691 bitmap_set(bitmap, offset, nsectors);
2695 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2696 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2699 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2702 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2705 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2708 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2711 static void scrub_block_complete(struct scrub_block *sblock)
2715 if (!sblock->no_io_error_seen) {
2717 scrub_handle_errored_block(sblock);
2720 * if has checksum error, write via repair mechanism in
2721 * dev replace case, otherwise write here in dev replace
2724 corrupted = scrub_checksum(sblock);
2725 if (!corrupted && sblock->sctx->is_dev_replace)
2726 scrub_write_block_to_dev_replace(sblock);
2729 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2730 u64 start = sblock->pagev[0]->logical;
2731 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2734 scrub_parity_mark_sectors_error(sblock->sparity,
2735 start, end - start);
2739 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2741 struct btrfs_ordered_sum *sum = NULL;
2742 unsigned long index;
2743 unsigned long num_sectors;
2745 while (!list_empty(&sctx->csum_list)) {
2746 sum = list_first_entry(&sctx->csum_list,
2747 struct btrfs_ordered_sum, list);
2748 if (sum->bytenr > logical)
2750 if (sum->bytenr + sum->len > logical)
2753 ++sctx->stat.csum_discards;
2754 list_del(&sum->list);
2761 index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2762 ASSERT(index < UINT_MAX);
2764 num_sectors = sum->len / sctx->fs_info->sectorsize;
2765 memcpy(csum, sum->sums + index, sctx->csum_size);
2766 if (index == num_sectors - 1) {
2767 list_del(&sum->list);
2773 /* scrub extent tries to collect up to 64 kB for each bio */
2774 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2775 u64 logical, u64 len,
2776 u64 physical, struct btrfs_device *dev, u64 flags,
2777 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2780 u8 csum[BTRFS_CSUM_SIZE];
2783 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2784 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2785 blocksize = map->stripe_len;
2787 blocksize = sctx->fs_info->sectorsize;
2788 spin_lock(&sctx->stat_lock);
2789 sctx->stat.data_extents_scrubbed++;
2790 sctx->stat.data_bytes_scrubbed += len;
2791 spin_unlock(&sctx->stat_lock);
2792 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2793 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2794 blocksize = map->stripe_len;
2796 blocksize = sctx->fs_info->nodesize;
2797 spin_lock(&sctx->stat_lock);
2798 sctx->stat.tree_extents_scrubbed++;
2799 sctx->stat.tree_bytes_scrubbed += len;
2800 spin_unlock(&sctx->stat_lock);
2802 blocksize = sctx->fs_info->sectorsize;
2807 u64 l = min_t(u64, len, blocksize);
2810 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2811 /* push csums to sbio */
2812 have_csum = scrub_find_csum(sctx, logical, csum);
2814 ++sctx->stat.no_csum;
2815 if (sctx->is_dev_replace && !have_csum) {
2816 ret = copy_nocow_pages(sctx, logical, l,
2818 physical_for_dev_replace);
2819 goto behind_scrub_pages;
2822 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2823 mirror_num, have_csum ? csum : NULL, 0,
2824 physical_for_dev_replace);
2831 physical_for_dev_replace += l;
2836 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2837 u64 logical, u64 len,
2838 u64 physical, struct btrfs_device *dev,
2839 u64 flags, u64 gen, int mirror_num, u8 *csum)
2841 struct scrub_ctx *sctx = sparity->sctx;
2842 struct scrub_block *sblock;
2845 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2847 spin_lock(&sctx->stat_lock);
2848 sctx->stat.malloc_errors++;
2849 spin_unlock(&sctx->stat_lock);
2853 /* one ref inside this function, plus one for each page added to
2855 refcount_set(&sblock->refs, 1);
2856 sblock->sctx = sctx;
2857 sblock->no_io_error_seen = 1;
2858 sblock->sparity = sparity;
2859 scrub_parity_get(sparity);
2861 for (index = 0; len > 0; index++) {
2862 struct scrub_page *spage;
2863 u64 l = min_t(u64, len, PAGE_SIZE);
2865 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2868 spin_lock(&sctx->stat_lock);
2869 sctx->stat.malloc_errors++;
2870 spin_unlock(&sctx->stat_lock);
2871 scrub_block_put(sblock);
2874 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2875 /* For scrub block */
2876 scrub_page_get(spage);
2877 sblock->pagev[index] = spage;
2878 /* For scrub parity */
2879 scrub_page_get(spage);
2880 list_add_tail(&spage->list, &sparity->spages);
2881 spage->sblock = sblock;
2883 spage->flags = flags;
2884 spage->generation = gen;
2885 spage->logical = logical;
2886 spage->physical = physical;
2887 spage->mirror_num = mirror_num;
2889 spage->have_csum = 1;
2890 memcpy(spage->csum, csum, sctx->csum_size);
2892 spage->have_csum = 0;
2894 sblock->page_count++;
2895 spage->page = alloc_page(GFP_KERNEL);
2903 WARN_ON(sblock->page_count == 0);
2904 for (index = 0; index < sblock->page_count; index++) {
2905 struct scrub_page *spage = sblock->pagev[index];
2908 ret = scrub_add_page_to_rd_bio(sctx, spage);
2910 scrub_block_put(sblock);
2915 /* last one frees, either here or in bio completion for last page */
2916 scrub_block_put(sblock);
2920 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2921 u64 logical, u64 len,
2922 u64 physical, struct btrfs_device *dev,
2923 u64 flags, u64 gen, int mirror_num)
2925 struct scrub_ctx *sctx = sparity->sctx;
2927 u8 csum[BTRFS_CSUM_SIZE];
2930 if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2931 scrub_parity_mark_sectors_error(sparity, logical, len);
2935 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2936 blocksize = sparity->stripe_len;
2937 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2938 blocksize = sparity->stripe_len;
2940 blocksize = sctx->fs_info->sectorsize;
2945 u64 l = min_t(u64, len, blocksize);
2948 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2949 /* push csums to sbio */
2950 have_csum = scrub_find_csum(sctx, logical, csum);
2954 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2955 flags, gen, mirror_num,
2956 have_csum ? csum : NULL);
2968 * Given a physical address, this will calculate it's
2969 * logical offset. if this is a parity stripe, it will return
2970 * the most left data stripe's logical offset.
2972 * return 0 if it is a data stripe, 1 means parity stripe.
2974 static int get_raid56_logic_offset(u64 physical, int num,
2975 struct map_lookup *map, u64 *offset,
2985 last_offset = (physical - map->stripes[num].physical) *
2986 nr_data_stripes(map);
2988 *stripe_start = last_offset;
2990 *offset = last_offset;
2991 for (i = 0; i < nr_data_stripes(map); i++) {
2992 *offset = last_offset + i * map->stripe_len;
2994 stripe_nr = div64_u64(*offset, map->stripe_len);
2995 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2997 /* Work out the disk rotation on this stripe-set */
2998 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2999 /* calculate which stripe this data locates */
3001 stripe_index = rot % map->num_stripes;
3002 if (stripe_index == num)
3004 if (stripe_index < num)
3007 *offset = last_offset + j * map->stripe_len;
3011 static void scrub_free_parity(struct scrub_parity *sparity)
3013 struct scrub_ctx *sctx = sparity->sctx;
3014 struct scrub_page *curr, *next;
3017 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
3019 spin_lock(&sctx->stat_lock);
3020 sctx->stat.read_errors += nbits;
3021 sctx->stat.uncorrectable_errors += nbits;
3022 spin_unlock(&sctx->stat_lock);
3025 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
3026 list_del_init(&curr->list);
3027 scrub_page_put(curr);
3033 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
3035 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3037 struct scrub_ctx *sctx = sparity->sctx;
3039 scrub_free_parity(sparity);
3040 scrub_pending_bio_dec(sctx);
3043 static void scrub_parity_bio_endio(struct bio *bio)
3045 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3046 struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3049 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3054 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3055 scrub_parity_bio_endio_worker, NULL, NULL);
3056 btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3059 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3061 struct scrub_ctx *sctx = sparity->sctx;
3062 struct btrfs_fs_info *fs_info = sctx->fs_info;
3064 struct btrfs_raid_bio *rbio;
3065 struct btrfs_bio *bbio = NULL;
3069 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3073 length = sparity->logic_end - sparity->logic_start;
3075 btrfs_bio_counter_inc_blocked(fs_info);
3076 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3078 if (ret || !bbio || !bbio->raid_map)
3081 bio = btrfs_io_bio_alloc(0);
3082 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3083 bio->bi_private = sparity;
3084 bio->bi_end_io = scrub_parity_bio_endio;
3086 rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3087 length, sparity->scrub_dev,
3093 scrub_pending_bio_inc(sctx);
3094 raid56_parity_submit_scrub_rbio(rbio);
3100 btrfs_bio_counter_dec(fs_info);
3101 btrfs_put_bbio(bbio);
3102 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3104 spin_lock(&sctx->stat_lock);
3105 sctx->stat.malloc_errors++;
3106 spin_unlock(&sctx->stat_lock);
3108 scrub_free_parity(sparity);
3111 static inline int scrub_calc_parity_bitmap_len(int nsectors)
3113 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3116 static void scrub_parity_get(struct scrub_parity *sparity)
3118 refcount_inc(&sparity->refs);
3121 static void scrub_parity_put(struct scrub_parity *sparity)
3123 if (!refcount_dec_and_test(&sparity->refs))
3126 scrub_parity_check_and_repair(sparity);
3129 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3130 struct map_lookup *map,
3131 struct btrfs_device *sdev,
3132 struct btrfs_path *path,
3136 struct btrfs_fs_info *fs_info = sctx->fs_info;
3137 struct btrfs_root *root = fs_info->extent_root;
3138 struct btrfs_root *csum_root = fs_info->csum_root;
3139 struct btrfs_extent_item *extent;
3140 struct btrfs_bio *bbio = NULL;
3144 struct extent_buffer *l;
3145 struct btrfs_key key;
3148 u64 extent_physical;
3151 struct btrfs_device *extent_dev;
3152 struct scrub_parity *sparity;
3155 int extent_mirror_num;
3158 nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3159 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3160 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3163 spin_lock(&sctx->stat_lock);
3164 sctx->stat.malloc_errors++;
3165 spin_unlock(&sctx->stat_lock);
3169 sparity->stripe_len = map->stripe_len;
3170 sparity->nsectors = nsectors;
3171 sparity->sctx = sctx;
3172 sparity->scrub_dev = sdev;
3173 sparity->logic_start = logic_start;
3174 sparity->logic_end = logic_end;
3175 refcount_set(&sparity->refs, 1);
3176 INIT_LIST_HEAD(&sparity->spages);
3177 sparity->dbitmap = sparity->bitmap;
3178 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3181 while (logic_start < logic_end) {
3182 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3183 key.type = BTRFS_METADATA_ITEM_KEY;
3185 key.type = BTRFS_EXTENT_ITEM_KEY;
3186 key.objectid = logic_start;
3187 key.offset = (u64)-1;
3189 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3194 ret = btrfs_previous_extent_item(root, path, 0);
3198 btrfs_release_path(path);
3199 ret = btrfs_search_slot(NULL, root, &key,
3211 slot = path->slots[0];
3212 if (slot >= btrfs_header_nritems(l)) {
3213 ret = btrfs_next_leaf(root, path);
3222 btrfs_item_key_to_cpu(l, &key, slot);
3224 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3225 key.type != BTRFS_METADATA_ITEM_KEY)
3228 if (key.type == BTRFS_METADATA_ITEM_KEY)
3229 bytes = fs_info->nodesize;
3233 if (key.objectid + bytes <= logic_start)
3236 if (key.objectid >= logic_end) {
3241 while (key.objectid >= logic_start + map->stripe_len)
3242 logic_start += map->stripe_len;
3244 extent = btrfs_item_ptr(l, slot,
3245 struct btrfs_extent_item);
3246 flags = btrfs_extent_flags(l, extent);
3247 generation = btrfs_extent_generation(l, extent);
3249 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3250 (key.objectid < logic_start ||
3251 key.objectid + bytes >
3252 logic_start + map->stripe_len)) {
3254 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3255 key.objectid, logic_start);
3256 spin_lock(&sctx->stat_lock);
3257 sctx->stat.uncorrectable_errors++;
3258 spin_unlock(&sctx->stat_lock);
3262 extent_logical = key.objectid;
3265 if (extent_logical < logic_start) {
3266 extent_len -= logic_start - extent_logical;
3267 extent_logical = logic_start;
3270 if (extent_logical + extent_len >
3271 logic_start + map->stripe_len)
3272 extent_len = logic_start + map->stripe_len -
3275 scrub_parity_mark_sectors_data(sparity, extent_logical,
3278 mapped_length = extent_len;
3280 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3281 extent_logical, &mapped_length, &bbio,
3284 if (!bbio || mapped_length < extent_len)
3288 btrfs_put_bbio(bbio);
3291 extent_physical = bbio->stripes[0].physical;
3292 extent_mirror_num = bbio->mirror_num;
3293 extent_dev = bbio->stripes[0].dev;
3294 btrfs_put_bbio(bbio);
3296 ret = btrfs_lookup_csums_range(csum_root,
3298 extent_logical + extent_len - 1,
3299 &sctx->csum_list, 1);
3303 ret = scrub_extent_for_parity(sparity, extent_logical,
3310 scrub_free_csums(sctx);
3315 if (extent_logical + extent_len <
3316 key.objectid + bytes) {
3317 logic_start += map->stripe_len;
3319 if (logic_start >= logic_end) {
3324 if (logic_start < key.objectid + bytes) {
3333 btrfs_release_path(path);
3338 logic_start += map->stripe_len;
3342 scrub_parity_mark_sectors_error(sparity, logic_start,
3343 logic_end - logic_start);
3344 scrub_parity_put(sparity);
3346 mutex_lock(&sctx->wr_lock);
3347 scrub_wr_submit(sctx);
3348 mutex_unlock(&sctx->wr_lock);
3350 btrfs_release_path(path);
3351 return ret < 0 ? ret : 0;
3354 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3355 struct map_lookup *map,
3356 struct btrfs_device *scrub_dev,
3357 int num, u64 base, u64 length,
3360 struct btrfs_path *path, *ppath;
3361 struct btrfs_fs_info *fs_info = sctx->fs_info;
3362 struct btrfs_root *root = fs_info->extent_root;
3363 struct btrfs_root *csum_root = fs_info->csum_root;
3364 struct btrfs_extent_item *extent;
3365 struct blk_plug plug;
3370 struct extent_buffer *l;
3377 struct reada_control *reada1;
3378 struct reada_control *reada2;
3379 struct btrfs_key key;
3380 struct btrfs_key key_end;
3381 u64 increment = map->stripe_len;
3384 u64 extent_physical;
3388 struct btrfs_device *extent_dev;
3389 int extent_mirror_num;
3392 physical = map->stripes[num].physical;
3394 nstripes = div64_u64(length, map->stripe_len);
3395 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3396 offset = map->stripe_len * num;
3397 increment = map->stripe_len * map->num_stripes;
3399 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3400 int factor = map->num_stripes / map->sub_stripes;
3401 offset = map->stripe_len * (num / map->sub_stripes);
3402 increment = map->stripe_len * factor;
3403 mirror_num = num % map->sub_stripes + 1;
3404 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3405 increment = map->stripe_len;
3406 mirror_num = num % map->num_stripes + 1;
3407 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3408 increment = map->stripe_len;
3409 mirror_num = num % map->num_stripes + 1;
3410 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3411 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3412 increment = map->stripe_len * nr_data_stripes(map);
3415 increment = map->stripe_len;
3419 path = btrfs_alloc_path();
3423 ppath = btrfs_alloc_path();
3425 btrfs_free_path(path);
3430 * work on commit root. The related disk blocks are static as
3431 * long as COW is applied. This means, it is save to rewrite
3432 * them to repair disk errors without any race conditions
3434 path->search_commit_root = 1;
3435 path->skip_locking = 1;
3437 ppath->search_commit_root = 1;
3438 ppath->skip_locking = 1;
3440 * trigger the readahead for extent tree csum tree and wait for
3441 * completion. During readahead, the scrub is officially paused
3442 * to not hold off transaction commits
3444 logical = base + offset;
3445 physical_end = physical + nstripes * map->stripe_len;
3446 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3447 get_raid56_logic_offset(physical_end, num,
3448 map, &logic_end, NULL);
3451 logic_end = logical + increment * nstripes;
3453 wait_event(sctx->list_wait,
3454 atomic_read(&sctx->bios_in_flight) == 0);
3455 scrub_blocked_if_needed(fs_info);
3457 /* FIXME it might be better to start readahead at commit root */
3458 key.objectid = logical;
3459 key.type = BTRFS_EXTENT_ITEM_KEY;
3460 key.offset = (u64)0;
3461 key_end.objectid = logic_end;
3462 key_end.type = BTRFS_METADATA_ITEM_KEY;
3463 key_end.offset = (u64)-1;
3464 reada1 = btrfs_reada_add(root, &key, &key_end);
3466 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3467 key.type = BTRFS_EXTENT_CSUM_KEY;
3468 key.offset = logical;
3469 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3470 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3471 key_end.offset = logic_end;
3472 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3474 if (!IS_ERR(reada1))
3475 btrfs_reada_wait(reada1);
3476 if (!IS_ERR(reada2))
3477 btrfs_reada_wait(reada2);
3481 * collect all data csums for the stripe to avoid seeking during
3482 * the scrub. This might currently (crc32) end up to be about 1MB
3484 blk_start_plug(&plug);
3487 * now find all extents for each stripe and scrub them
3490 while (physical < physical_end) {
3494 if (atomic_read(&fs_info->scrub_cancel_req) ||
3495 atomic_read(&sctx->cancel_req)) {
3500 * check to see if we have to pause
3502 if (atomic_read(&fs_info->scrub_pause_req)) {
3503 /* push queued extents */
3504 sctx->flush_all_writes = true;
3506 mutex_lock(&sctx->wr_lock);
3507 scrub_wr_submit(sctx);
3508 mutex_unlock(&sctx->wr_lock);
3509 wait_event(sctx->list_wait,
3510 atomic_read(&sctx->bios_in_flight) == 0);
3511 sctx->flush_all_writes = false;
3512 scrub_blocked_if_needed(fs_info);
3515 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3516 ret = get_raid56_logic_offset(physical, num, map,
3521 /* it is parity strip */
3522 stripe_logical += base;
3523 stripe_end = stripe_logical + increment;
3524 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3525 ppath, stripe_logical,
3533 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3534 key.type = BTRFS_METADATA_ITEM_KEY;
3536 key.type = BTRFS_EXTENT_ITEM_KEY;
3537 key.objectid = logical;
3538 key.offset = (u64)-1;
3540 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3545 ret = btrfs_previous_extent_item(root, path, 0);
3549 /* there's no smaller item, so stick with the
3551 btrfs_release_path(path);
3552 ret = btrfs_search_slot(NULL, root, &key,
3564 slot = path->slots[0];
3565 if (slot >= btrfs_header_nritems(l)) {
3566 ret = btrfs_next_leaf(root, path);
3575 btrfs_item_key_to_cpu(l, &key, slot);
3577 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3578 key.type != BTRFS_METADATA_ITEM_KEY)
3581 if (key.type == BTRFS_METADATA_ITEM_KEY)
3582 bytes = fs_info->nodesize;
3586 if (key.objectid + bytes <= logical)
3589 if (key.objectid >= logical + map->stripe_len) {
3590 /* out of this device extent */
3591 if (key.objectid >= logic_end)
3596 extent = btrfs_item_ptr(l, slot,
3597 struct btrfs_extent_item);
3598 flags = btrfs_extent_flags(l, extent);
3599 generation = btrfs_extent_generation(l, extent);
3601 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3602 (key.objectid < logical ||
3603 key.objectid + bytes >
3604 logical + map->stripe_len)) {
3606 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3607 key.objectid, logical);
3608 spin_lock(&sctx->stat_lock);
3609 sctx->stat.uncorrectable_errors++;
3610 spin_unlock(&sctx->stat_lock);
3615 extent_logical = key.objectid;
3619 * trim extent to this stripe
3621 if (extent_logical < logical) {
3622 extent_len -= logical - extent_logical;
3623 extent_logical = logical;
3625 if (extent_logical + extent_len >
3626 logical + map->stripe_len) {
3627 extent_len = logical + map->stripe_len -
3631 extent_physical = extent_logical - logical + physical;
3632 extent_dev = scrub_dev;
3633 extent_mirror_num = mirror_num;
3635 scrub_remap_extent(fs_info, extent_logical,
3636 extent_len, &extent_physical,
3638 &extent_mirror_num);
3640 ret = btrfs_lookup_csums_range(csum_root,
3644 &sctx->csum_list, 1);
3648 ret = scrub_extent(sctx, map, extent_logical, extent_len,
3649 extent_physical, extent_dev, flags,
3650 generation, extent_mirror_num,
3651 extent_logical - logical + physical);
3653 scrub_free_csums(sctx);
3658 if (extent_logical + extent_len <
3659 key.objectid + bytes) {
3660 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3662 * loop until we find next data stripe
3663 * or we have finished all stripes.
3666 physical += map->stripe_len;
3667 ret = get_raid56_logic_offset(physical,
3672 if (ret && physical < physical_end) {
3673 stripe_logical += base;
3674 stripe_end = stripe_logical +
3676 ret = scrub_raid56_parity(sctx,
3677 map, scrub_dev, ppath,
3685 physical += map->stripe_len;
3686 logical += increment;
3688 if (logical < key.objectid + bytes) {
3693 if (physical >= physical_end) {
3701 btrfs_release_path(path);
3703 logical += increment;
3704 physical += map->stripe_len;
3705 spin_lock(&sctx->stat_lock);
3707 sctx->stat.last_physical = map->stripes[num].physical +
3710 sctx->stat.last_physical = physical;
3711 spin_unlock(&sctx->stat_lock);
3716 /* push queued extents */
3718 mutex_lock(&sctx->wr_lock);
3719 scrub_wr_submit(sctx);
3720 mutex_unlock(&sctx->wr_lock);
3722 blk_finish_plug(&plug);
3723 btrfs_free_path(path);
3724 btrfs_free_path(ppath);
3725 return ret < 0 ? ret : 0;
3728 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3729 struct btrfs_device *scrub_dev,
3730 u64 chunk_offset, u64 length,
3732 struct btrfs_block_group_cache *cache,
3735 struct btrfs_fs_info *fs_info = sctx->fs_info;
3736 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3737 struct map_lookup *map;
3738 struct extent_map *em;
3742 read_lock(&map_tree->map_tree.lock);
3743 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3744 read_unlock(&map_tree->map_tree.lock);
3748 * Might have been an unused block group deleted by the cleaner
3749 * kthread or relocation.
3751 spin_lock(&cache->lock);
3752 if (!cache->removed)
3754 spin_unlock(&cache->lock);
3759 map = em->map_lookup;
3760 if (em->start != chunk_offset)
3763 if (em->len < length)
3766 for (i = 0; i < map->num_stripes; ++i) {
3767 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3768 map->stripes[i].physical == dev_offset) {
3769 ret = scrub_stripe(sctx, map, scrub_dev, i,
3770 chunk_offset, length,
3777 free_extent_map(em);
3782 static noinline_for_stack
3783 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3784 struct btrfs_device *scrub_dev, u64 start, u64 end,
3787 struct btrfs_dev_extent *dev_extent = NULL;
3788 struct btrfs_path *path;
3789 struct btrfs_fs_info *fs_info = sctx->fs_info;
3790 struct btrfs_root *root = fs_info->dev_root;
3796 struct extent_buffer *l;
3797 struct btrfs_key key;
3798 struct btrfs_key found_key;
3799 struct btrfs_block_group_cache *cache;
3800 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3802 path = btrfs_alloc_path();
3806 path->reada = READA_FORWARD;
3807 path->search_commit_root = 1;
3808 path->skip_locking = 1;
3810 key.objectid = scrub_dev->devid;
3812 key.type = BTRFS_DEV_EXTENT_KEY;
3815 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3819 if (path->slots[0] >=
3820 btrfs_header_nritems(path->nodes[0])) {
3821 ret = btrfs_next_leaf(root, path);
3834 slot = path->slots[0];
3836 btrfs_item_key_to_cpu(l, &found_key, slot);
3838 if (found_key.objectid != scrub_dev->devid)
3841 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3844 if (found_key.offset >= end)
3847 if (found_key.offset < key.offset)
3850 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3851 length = btrfs_dev_extent_length(l, dev_extent);
3853 if (found_key.offset + length <= start)
3856 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3859 * get a reference on the corresponding block group to prevent
3860 * the chunk from going away while we scrub it
3862 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3864 /* some chunks are removed but not committed to disk yet,
3865 * continue scrubbing */
3870 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3871 * to avoid deadlock caused by:
3872 * btrfs_inc_block_group_ro()
3873 * -> btrfs_wait_for_commit()
3874 * -> btrfs_commit_transaction()
3875 * -> btrfs_scrub_pause()
3877 scrub_pause_on(fs_info);
3878 ret = btrfs_inc_block_group_ro(fs_info, cache);
3879 if (!ret && is_dev_replace) {
3881 * If we are doing a device replace wait for any tasks
3882 * that started dellaloc right before we set the block
3883 * group to RO mode, as they might have just allocated
3884 * an extent from it or decided they could do a nocow
3885 * write. And if any such tasks did that, wait for their
3886 * ordered extents to complete and then commit the
3887 * current transaction, so that we can later see the new
3888 * extent items in the extent tree - the ordered extents
3889 * create delayed data references (for cow writes) when
3890 * they complete, which will be run and insert the
3891 * corresponding extent items into the extent tree when
3892 * we commit the transaction they used when running
3893 * inode.c:btrfs_finish_ordered_io(). We later use
3894 * the commit root of the extent tree to find extents
3895 * to copy from the srcdev into the tgtdev, and we don't
3896 * want to miss any new extents.
3898 btrfs_wait_block_group_reservations(cache);
3899 btrfs_wait_nocow_writers(cache);
3900 ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3901 cache->key.objectid,
3904 struct btrfs_trans_handle *trans;
3906 trans = btrfs_join_transaction(root);
3908 ret = PTR_ERR(trans);
3910 ret = btrfs_commit_transaction(trans);
3912 scrub_pause_off(fs_info);
3913 btrfs_put_block_group(cache);
3918 scrub_pause_off(fs_info);
3922 } else if (ret == -ENOSPC) {
3924 * btrfs_inc_block_group_ro return -ENOSPC when it
3925 * failed in creating new chunk for metadata.
3926 * It is not a problem for scrub/replace, because
3927 * metadata are always cowed, and our scrub paused
3928 * commit_transactions.
3933 "failed setting block group ro: %d", ret);
3934 btrfs_put_block_group(cache);
3938 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3939 dev_replace->cursor_right = found_key.offset + length;
3940 dev_replace->cursor_left = found_key.offset;
3941 dev_replace->item_needs_writeback = 1;
3942 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3943 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3944 found_key.offset, cache, is_dev_replace);
3947 * flush, submit all pending read and write bios, afterwards
3949 * Note that in the dev replace case, a read request causes
3950 * write requests that are submitted in the read completion
3951 * worker. Therefore in the current situation, it is required
3952 * that all write requests are flushed, so that all read and
3953 * write requests are really completed when bios_in_flight
3956 sctx->flush_all_writes = true;
3958 mutex_lock(&sctx->wr_lock);
3959 scrub_wr_submit(sctx);
3960 mutex_unlock(&sctx->wr_lock);
3962 wait_event(sctx->list_wait,
3963 atomic_read(&sctx->bios_in_flight) == 0);
3965 scrub_pause_on(fs_info);
3968 * must be called before we decrease @scrub_paused.
3969 * make sure we don't block transaction commit while
3970 * we are waiting pending workers finished.
3972 wait_event(sctx->list_wait,
3973 atomic_read(&sctx->workers_pending) == 0);
3974 sctx->flush_all_writes = false;
3976 scrub_pause_off(fs_info);
3978 btrfs_dev_replace_write_lock(&fs_info->dev_replace);
3979 dev_replace->cursor_left = dev_replace->cursor_right;
3980 dev_replace->item_needs_writeback = 1;
3981 btrfs_dev_replace_write_unlock(&fs_info->dev_replace);
3984 btrfs_dec_block_group_ro(cache);
3987 * We might have prevented the cleaner kthread from deleting
3988 * this block group if it was already unused because we raced
3989 * and set it to RO mode first. So add it back to the unused
3990 * list, otherwise it might not ever be deleted unless a manual
3991 * balance is triggered or it becomes used and unused again.
3993 spin_lock(&cache->lock);
3994 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3995 btrfs_block_group_used(&cache->item) == 0) {
3996 spin_unlock(&cache->lock);
3997 spin_lock(&fs_info->unused_bgs_lock);
3998 if (list_empty(&cache->bg_list)) {
3999 btrfs_get_block_group(cache);
4000 list_add_tail(&cache->bg_list,
4001 &fs_info->unused_bgs);
4003 spin_unlock(&fs_info->unused_bgs_lock);
4005 spin_unlock(&cache->lock);
4008 btrfs_put_block_group(cache);
4011 if (is_dev_replace &&
4012 atomic64_read(&dev_replace->num_write_errors) > 0) {
4016 if (sctx->stat.malloc_errors > 0) {
4021 key.offset = found_key.offset + length;
4022 btrfs_release_path(path);
4025 btrfs_free_path(path);
4030 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
4031 struct btrfs_device *scrub_dev)
4037 struct btrfs_fs_info *fs_info = sctx->fs_info;
4039 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4042 /* Seed devices of a new filesystem has their own generation. */
4043 if (scrub_dev->fs_devices != fs_info->fs_devices)
4044 gen = scrub_dev->generation;
4046 gen = fs_info->last_trans_committed;
4048 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4049 bytenr = btrfs_sb_offset(i);
4050 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4051 scrub_dev->commit_total_bytes)
4054 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4055 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4060 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4066 * get a reference count on fs_info->scrub_workers. start worker if necessary
4068 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4071 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4072 int max_active = fs_info->thread_pool_size;
4074 if (fs_info->scrub_workers_refcnt == 0) {
4075 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
4076 flags, is_dev_replace ? 1 : max_active, 4);
4077 if (!fs_info->scrub_workers)
4078 goto fail_scrub_workers;
4080 fs_info->scrub_wr_completion_workers =
4081 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4083 if (!fs_info->scrub_wr_completion_workers)
4084 goto fail_scrub_wr_completion_workers;
4086 fs_info->scrub_nocow_workers =
4087 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4088 if (!fs_info->scrub_nocow_workers)
4089 goto fail_scrub_nocow_workers;
4090 fs_info->scrub_parity_workers =
4091 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4093 if (!fs_info->scrub_parity_workers)
4094 goto fail_scrub_parity_workers;
4096 ++fs_info->scrub_workers_refcnt;
4099 fail_scrub_parity_workers:
4100 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4101 fail_scrub_nocow_workers:
4102 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4103 fail_scrub_wr_completion_workers:
4104 btrfs_destroy_workqueue(fs_info->scrub_workers);
4109 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4111 if (--fs_info->scrub_workers_refcnt == 0) {
4112 btrfs_destroy_workqueue(fs_info->scrub_workers);
4113 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4114 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4115 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4117 WARN_ON(fs_info->scrub_workers_refcnt < 0);
4120 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4121 u64 end, struct btrfs_scrub_progress *progress,
4122 int readonly, int is_dev_replace)
4124 struct scrub_ctx *sctx;
4126 struct btrfs_device *dev;
4127 struct rcu_string *name;
4129 if (btrfs_fs_closing(fs_info))
4132 if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4134 * in this case scrub is unable to calculate the checksum
4135 * the way scrub is implemented. Do not handle this
4136 * situation at all because it won't ever happen.
4139 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4145 if (fs_info->sectorsize != PAGE_SIZE) {
4146 /* not supported for data w/o checksums */
4147 btrfs_err_rl(fs_info,
4148 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4149 fs_info->sectorsize, PAGE_SIZE);
4153 if (fs_info->nodesize >
4154 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4155 fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4157 * would exhaust the array bounds of pagev member in
4158 * struct scrub_block
4161 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4163 SCRUB_MAX_PAGES_PER_BLOCK,
4164 fs_info->sectorsize,
4165 SCRUB_MAX_PAGES_PER_BLOCK);
4170 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4171 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4172 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4174 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4178 if (!is_dev_replace && !readonly &&
4179 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4180 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4182 name = rcu_dereference(dev->name);
4183 btrfs_err(fs_info, "scrub: device %s is not writable",
4189 mutex_lock(&fs_info->scrub_lock);
4190 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4191 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4192 mutex_unlock(&fs_info->scrub_lock);
4193 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4197 btrfs_dev_replace_read_lock(&fs_info->dev_replace);
4198 if (dev->scrub_ctx ||
4200 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4201 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4202 mutex_unlock(&fs_info->scrub_lock);
4203 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4204 return -EINPROGRESS;
4206 btrfs_dev_replace_read_unlock(&fs_info->dev_replace);
4208 ret = scrub_workers_get(fs_info, is_dev_replace);
4210 mutex_unlock(&fs_info->scrub_lock);
4211 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4215 sctx = scrub_setup_ctx(dev, is_dev_replace);
4217 mutex_unlock(&fs_info->scrub_lock);
4218 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4219 scrub_workers_put(fs_info);
4220 return PTR_ERR(sctx);
4222 sctx->readonly = readonly;
4223 dev->scrub_ctx = sctx;
4224 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4227 * checking @scrub_pause_req here, we can avoid
4228 * race between committing transaction and scrubbing.
4230 __scrub_blocked_if_needed(fs_info);
4231 atomic_inc(&fs_info->scrubs_running);
4232 mutex_unlock(&fs_info->scrub_lock);
4234 if (!is_dev_replace) {
4236 * by holding device list mutex, we can
4237 * kick off writing super in log tree sync.
4239 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4240 ret = scrub_supers(sctx, dev);
4241 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4245 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4248 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4249 atomic_dec(&fs_info->scrubs_running);
4250 wake_up(&fs_info->scrub_pause_wait);
4252 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4255 memcpy(progress, &sctx->stat, sizeof(*progress));
4257 mutex_lock(&fs_info->scrub_lock);
4258 dev->scrub_ctx = NULL;
4259 scrub_workers_put(fs_info);
4260 mutex_unlock(&fs_info->scrub_lock);
4262 scrub_put_ctx(sctx);
4267 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4269 mutex_lock(&fs_info->scrub_lock);
4270 atomic_inc(&fs_info->scrub_pause_req);
4271 while (atomic_read(&fs_info->scrubs_paused) !=
4272 atomic_read(&fs_info->scrubs_running)) {
4273 mutex_unlock(&fs_info->scrub_lock);
4274 wait_event(fs_info->scrub_pause_wait,
4275 atomic_read(&fs_info->scrubs_paused) ==
4276 atomic_read(&fs_info->scrubs_running));
4277 mutex_lock(&fs_info->scrub_lock);
4279 mutex_unlock(&fs_info->scrub_lock);
4282 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4284 atomic_dec(&fs_info->scrub_pause_req);
4285 wake_up(&fs_info->scrub_pause_wait);
4288 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4290 mutex_lock(&fs_info->scrub_lock);
4291 if (!atomic_read(&fs_info->scrubs_running)) {
4292 mutex_unlock(&fs_info->scrub_lock);
4296 atomic_inc(&fs_info->scrub_cancel_req);
4297 while (atomic_read(&fs_info->scrubs_running)) {
4298 mutex_unlock(&fs_info->scrub_lock);
4299 wait_event(fs_info->scrub_pause_wait,
4300 atomic_read(&fs_info->scrubs_running) == 0);
4301 mutex_lock(&fs_info->scrub_lock);
4303 atomic_dec(&fs_info->scrub_cancel_req);
4304 mutex_unlock(&fs_info->scrub_lock);
4309 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4310 struct btrfs_device *dev)
4312 struct scrub_ctx *sctx;
4314 mutex_lock(&fs_info->scrub_lock);
4315 sctx = dev->scrub_ctx;
4317 mutex_unlock(&fs_info->scrub_lock);
4320 atomic_inc(&sctx->cancel_req);
4321 while (dev->scrub_ctx) {
4322 mutex_unlock(&fs_info->scrub_lock);
4323 wait_event(fs_info->scrub_pause_wait,
4324 dev->scrub_ctx == NULL);
4325 mutex_lock(&fs_info->scrub_lock);
4327 mutex_unlock(&fs_info->scrub_lock);
4332 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4333 struct btrfs_scrub_progress *progress)
4335 struct btrfs_device *dev;
4336 struct scrub_ctx *sctx = NULL;
4338 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4339 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4341 sctx = dev->scrub_ctx;
4343 memcpy(progress, &sctx->stat, sizeof(*progress));
4344 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4346 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4349 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4350 u64 extent_logical, u64 extent_len,
4351 u64 *extent_physical,
4352 struct btrfs_device **extent_dev,
4353 int *extent_mirror_num)
4356 struct btrfs_bio *bbio = NULL;
4359 mapped_length = extent_len;
4360 ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4361 &mapped_length, &bbio, 0);
4362 if (ret || !bbio || mapped_length < extent_len ||
4363 !bbio->stripes[0].dev->bdev) {
4364 btrfs_put_bbio(bbio);
4368 *extent_physical = bbio->stripes[0].physical;
4369 *extent_mirror_num = bbio->mirror_num;
4370 *extent_dev = bbio->stripes[0].dev;
4371 btrfs_put_bbio(bbio);
4374 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4375 int mirror_num, u64 physical_for_dev_replace)
4377 struct scrub_copy_nocow_ctx *nocow_ctx;
4378 struct btrfs_fs_info *fs_info = sctx->fs_info;
4380 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4382 spin_lock(&sctx->stat_lock);
4383 sctx->stat.malloc_errors++;
4384 spin_unlock(&sctx->stat_lock);
4388 scrub_pending_trans_workers_inc(sctx);
4390 nocow_ctx->sctx = sctx;
4391 nocow_ctx->logical = logical;
4392 nocow_ctx->len = len;
4393 nocow_ctx->mirror_num = mirror_num;
4394 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4395 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4396 copy_nocow_pages_worker, NULL, NULL);
4397 INIT_LIST_HEAD(&nocow_ctx->inodes);
4398 btrfs_queue_work(fs_info->scrub_nocow_workers,
4404 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4406 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4407 struct scrub_nocow_inode *nocow_inode;
4409 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4412 nocow_inode->inum = inum;
4413 nocow_inode->offset = offset;
4414 nocow_inode->root = root;
4415 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4419 #define COPY_COMPLETE 1
4421 static void copy_nocow_pages_worker(struct btrfs_work *work)
4423 struct scrub_copy_nocow_ctx *nocow_ctx =
4424 container_of(work, struct scrub_copy_nocow_ctx, work);
4425 struct scrub_ctx *sctx = nocow_ctx->sctx;
4426 struct btrfs_fs_info *fs_info = sctx->fs_info;
4427 struct btrfs_root *root = fs_info->extent_root;
4428 u64 logical = nocow_ctx->logical;
4429 u64 len = nocow_ctx->len;
4430 int mirror_num = nocow_ctx->mirror_num;
4431 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4433 struct btrfs_trans_handle *trans = NULL;
4434 struct btrfs_path *path;
4435 int not_written = 0;
4437 path = btrfs_alloc_path();
4439 spin_lock(&sctx->stat_lock);
4440 sctx->stat.malloc_errors++;
4441 spin_unlock(&sctx->stat_lock);
4446 trans = btrfs_join_transaction(root);
4447 if (IS_ERR(trans)) {
4452 ret = iterate_inodes_from_logical(logical, fs_info, path,
4453 record_inode_for_nocow, nocow_ctx, false);
4454 if (ret != 0 && ret != -ENOENT) {
4456 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4457 logical, physical_for_dev_replace, len, mirror_num,
4463 btrfs_end_transaction(trans);
4465 while (!list_empty(&nocow_ctx->inodes)) {
4466 struct scrub_nocow_inode *entry;
4467 entry = list_first_entry(&nocow_ctx->inodes,
4468 struct scrub_nocow_inode,
4470 list_del_init(&entry->list);
4471 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4472 entry->root, nocow_ctx);
4474 if (ret == COPY_COMPLETE) {
4482 while (!list_empty(&nocow_ctx->inodes)) {
4483 struct scrub_nocow_inode *entry;
4484 entry = list_first_entry(&nocow_ctx->inodes,
4485 struct scrub_nocow_inode,
4487 list_del_init(&entry->list);
4490 if (trans && !IS_ERR(trans))
4491 btrfs_end_transaction(trans);
4493 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4494 num_uncorrectable_read_errors);
4496 btrfs_free_path(path);
4499 scrub_pending_trans_workers_dec(sctx);
4502 static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4505 struct extent_state *cached_state = NULL;
4506 struct btrfs_ordered_extent *ordered;
4507 struct extent_io_tree *io_tree;
4508 struct extent_map *em;
4509 u64 lockstart = start, lockend = start + len - 1;
4512 io_tree = &inode->io_tree;
4514 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4515 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4517 btrfs_put_ordered_extent(ordered);
4522 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4529 * This extent does not actually cover the logical extent anymore,
4530 * move on to the next inode.
4532 if (em->block_start > logical ||
4533 em->block_start + em->block_len < logical + len ||
4534 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4535 free_extent_map(em);
4539 free_extent_map(em);
4542 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state);
4546 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4547 struct scrub_copy_nocow_ctx *nocow_ctx)
4549 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4550 struct btrfs_key key;
4551 struct inode *inode;
4553 struct btrfs_root *local_root;
4554 struct extent_io_tree *io_tree;
4555 u64 physical_for_dev_replace;
4556 u64 nocow_ctx_logical;
4557 u64 len = nocow_ctx->len;
4558 unsigned long index;
4563 key.objectid = root;
4564 key.type = BTRFS_ROOT_ITEM_KEY;
4565 key.offset = (u64)-1;
4567 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4569 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4570 if (IS_ERR(local_root)) {
4571 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4572 return PTR_ERR(local_root);
4575 key.type = BTRFS_INODE_ITEM_KEY;
4576 key.objectid = inum;
4578 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4579 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4581 return PTR_ERR(inode);
4583 /* Avoid truncate/dio/punch hole.. */
4585 inode_dio_wait(inode);
4587 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4588 io_tree = &BTRFS_I(inode)->io_tree;
4589 nocow_ctx_logical = nocow_ctx->logical;
4591 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4594 ret = ret > 0 ? 0 : ret;
4598 while (len >= PAGE_SIZE) {
4599 index = offset >> PAGE_SHIFT;
4601 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4603 btrfs_err(fs_info, "find_or_create_page() failed");
4608 if (PageUptodate(page)) {
4609 if (PageDirty(page))
4612 ClearPageError(page);
4613 err = extent_read_full_page(io_tree, page,
4615 nocow_ctx->mirror_num);
4623 * If the page has been remove from the page cache,
4624 * the data on it is meaningless, because it may be
4625 * old one, the new data may be written into the new
4626 * page in the page cache.
4628 if (page->mapping != inode->i_mapping) {
4633 if (!PageUptodate(page)) {
4639 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4642 ret = ret > 0 ? 0 : ret;
4646 err = write_page_nocow(nocow_ctx->sctx,
4647 physical_for_dev_replace, page);
4657 offset += PAGE_SIZE;
4658 physical_for_dev_replace += PAGE_SIZE;
4659 nocow_ctx_logical += PAGE_SIZE;
4662 ret = COPY_COMPLETE;
4664 inode_unlock(inode);
4669 static int write_page_nocow(struct scrub_ctx *sctx,
4670 u64 physical_for_dev_replace, struct page *page)
4673 struct btrfs_device *dev;
4675 dev = sctx->wr_tgtdev;
4679 btrfs_warn_rl(dev->fs_info,
4680 "scrub write_page_nocow(bdev == NULL) is unexpected");
4683 bio = btrfs_io_bio_alloc(1);
4684 bio->bi_iter.bi_size = 0;
4685 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4686 bio_set_dev(bio, dev->bdev);
4687 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4688 /* bio_add_page won't fail on a freshly allocated bio */
4689 bio_add_page(bio, page, PAGE_SIZE, 0);
4691 if (btrfsic_submit_bio_wait(bio)) {
4693 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);