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>
24 #include "ordered-data.h"
25 #include "transaction.h"
27 #include "extent_io.h"
28 #include "dev-replace.h"
29 #include "check-integrity.h"
30 #include "rcu-string.h"
34 * This is only the first step towards a full-features scrub. It reads all
35 * extent and super block and verifies the checksums. In case a bad checksum
36 * is found or the extent cannot be read, good data will be written back if
39 * Future enhancements:
40 * - In case an unrepairable extent is encountered, track which files are
41 * affected and report them
42 * - track and record media errors, throw out bad devices
43 * - add a mode to also read unallocated space
50 * the following three values only influence the performance.
51 * The last one configures the number of parallel and outstanding I/O
52 * operations. The first two values configure an upper limit for the number
53 * of (dynamically allocated) pages that are added to a bio.
55 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
56 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
57 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
60 * the following value times PAGE_SIZE needs to be large enough to match the
61 * largest node/leaf/sector size that shall be supported.
62 * Values larger than BTRFS_STRIPE_LEN are not supported.
64 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
66 struct scrub_recover {
68 struct btrfs_bio *bbio;
73 struct scrub_block *sblock;
75 struct btrfs_device *dev;
76 struct list_head list;
77 u64 flags; /* extent flags */
81 u64 physical_for_dev_replace;
84 unsigned int mirror_num:8;
85 unsigned int have_csum:1;
86 unsigned int io_error:1;
88 u8 csum[BTRFS_CSUM_SIZE];
90 struct scrub_recover *recover;
95 struct scrub_ctx *sctx;
96 struct btrfs_device *dev;
101 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
102 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
104 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
108 struct btrfs_work work;
112 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
114 atomic_t outstanding_pages;
115 atomic_t refs; /* free mem on transition to zero */
116 struct scrub_ctx *sctx;
117 struct scrub_parity *sparity;
119 unsigned int header_error:1;
120 unsigned int checksum_error:1;
121 unsigned int no_io_error_seen:1;
122 unsigned int generation_error:1; /* also sets header_error */
124 /* The following is for the data used to check parity */
125 /* It is for the data with checksum */
126 unsigned int data_corrected:1;
128 struct btrfs_work work;
131 /* Used for the chunks with parity stripe such RAID5/6 */
132 struct scrub_parity {
133 struct scrub_ctx *sctx;
135 struct btrfs_device *scrub_dev;
147 struct list_head spages;
149 /* Work of parity check and repair */
150 struct btrfs_work work;
152 /* Mark the parity blocks which have data */
153 unsigned long *dbitmap;
156 * Mark the parity blocks which have data, but errors happen when
157 * read data or check data
159 unsigned long *ebitmap;
161 unsigned long bitmap[0];
164 struct scrub_wr_ctx {
165 struct scrub_bio *wr_curr_bio;
166 struct btrfs_device *tgtdev;
167 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
168 atomic_t flush_all_writes;
169 struct mutex wr_lock;
173 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
174 struct btrfs_root *dev_root;
177 atomic_t bios_in_flight;
178 atomic_t workers_pending;
179 spinlock_t list_lock;
180 wait_queue_head_t list_wait;
182 struct list_head csum_list;
185 int pages_per_rd_bio;
190 struct scrub_wr_ctx wr_ctx;
195 struct btrfs_scrub_progress stat;
196 spinlock_t stat_lock;
199 * Use a ref counter to avoid use-after-free issues. Scrub workers
200 * decrement bios_in_flight and workers_pending and then do a wakeup
201 * on the list_wait wait queue. We must ensure the main scrub task
202 * doesn't free the scrub context before or while the workers are
203 * doing the wakeup() call.
208 struct scrub_fixup_nodatasum {
209 struct scrub_ctx *sctx;
210 struct btrfs_device *dev;
212 struct btrfs_root *root;
213 struct btrfs_work work;
217 struct scrub_nocow_inode {
221 struct list_head list;
224 struct scrub_copy_nocow_ctx {
225 struct scrub_ctx *sctx;
229 u64 physical_for_dev_replace;
230 struct list_head inodes;
231 struct btrfs_work work;
234 struct scrub_warning {
235 struct btrfs_path *path;
236 u64 extent_item_size;
240 struct btrfs_device *dev;
243 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
244 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
245 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
246 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
247 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
248 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
249 struct scrub_block *sblocks_for_recheck);
250 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
251 struct scrub_block *sblock,
252 int retry_failed_mirror);
253 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
254 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
255 struct scrub_block *sblock_good);
256 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
257 struct scrub_block *sblock_good,
258 int page_num, int force_write);
259 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
260 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
262 static int scrub_checksum_data(struct scrub_block *sblock);
263 static int scrub_checksum_tree_block(struct scrub_block *sblock);
264 static int scrub_checksum_super(struct scrub_block *sblock);
265 static void scrub_block_get(struct scrub_block *sblock);
266 static void scrub_block_put(struct scrub_block *sblock);
267 static void scrub_page_get(struct scrub_page *spage);
268 static void scrub_page_put(struct scrub_page *spage);
269 static void scrub_parity_get(struct scrub_parity *sparity);
270 static void scrub_parity_put(struct scrub_parity *sparity);
271 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
272 struct scrub_page *spage);
273 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
274 u64 physical, struct btrfs_device *dev, u64 flags,
275 u64 gen, int mirror_num, u8 *csum, int force,
276 u64 physical_for_dev_replace);
277 static void scrub_bio_end_io(struct bio *bio);
278 static void scrub_bio_end_io_worker(struct btrfs_work *work);
279 static void scrub_block_complete(struct scrub_block *sblock);
280 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
281 u64 extent_logical, u64 extent_len,
282 u64 *extent_physical,
283 struct btrfs_device **extent_dev,
284 int *extent_mirror_num);
285 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
286 struct scrub_wr_ctx *wr_ctx,
287 struct btrfs_fs_info *fs_info,
288 struct btrfs_device *dev,
290 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
291 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
292 struct scrub_page *spage);
293 static void scrub_wr_submit(struct scrub_ctx *sctx);
294 static void scrub_wr_bio_end_io(struct bio *bio);
295 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
296 static int write_page_nocow(struct scrub_ctx *sctx,
297 u64 physical_for_dev_replace, struct page *page);
298 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
299 struct scrub_copy_nocow_ctx *ctx);
300 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
301 int mirror_num, u64 physical_for_dev_replace);
302 static void copy_nocow_pages_worker(struct btrfs_work *work);
303 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
305 static void scrub_put_ctx(struct scrub_ctx *sctx);
308 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
310 atomic_inc(&sctx->refs);
311 atomic_inc(&sctx->bios_in_flight);
314 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
316 atomic_dec(&sctx->bios_in_flight);
317 wake_up(&sctx->list_wait);
321 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
323 while (atomic_read(&fs_info->scrub_pause_req)) {
324 mutex_unlock(&fs_info->scrub_lock);
325 wait_event(fs_info->scrub_pause_wait,
326 atomic_read(&fs_info->scrub_pause_req) == 0);
327 mutex_lock(&fs_info->scrub_lock);
331 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
333 atomic_inc(&fs_info->scrubs_paused);
334 wake_up(&fs_info->scrub_pause_wait);
337 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
339 mutex_lock(&fs_info->scrub_lock);
340 __scrub_blocked_if_needed(fs_info);
341 atomic_dec(&fs_info->scrubs_paused);
342 mutex_unlock(&fs_info->scrub_lock);
344 wake_up(&fs_info->scrub_pause_wait);
347 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
349 scrub_pause_on(fs_info);
350 scrub_pause_off(fs_info);
354 * used for workers that require transaction commits (i.e., for the
357 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
359 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
361 atomic_inc(&sctx->refs);
363 * increment scrubs_running to prevent cancel requests from
364 * completing as long as a worker is running. we must also
365 * increment scrubs_paused to prevent deadlocking on pause
366 * requests used for transactions commits (as the worker uses a
367 * transaction context). it is safe to regard the worker
368 * as paused for all matters practical. effectively, we only
369 * avoid cancellation requests from completing.
371 mutex_lock(&fs_info->scrub_lock);
372 atomic_inc(&fs_info->scrubs_running);
373 atomic_inc(&fs_info->scrubs_paused);
374 mutex_unlock(&fs_info->scrub_lock);
377 * check if @scrubs_running=@scrubs_paused condition
378 * inside wait_event() is not an atomic operation.
379 * which means we may inc/dec @scrub_running/paused
380 * at any time. Let's wake up @scrub_pause_wait as
381 * much as we can to let commit transaction blocked less.
383 wake_up(&fs_info->scrub_pause_wait);
385 atomic_inc(&sctx->workers_pending);
388 /* used for workers that require transaction commits */
389 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
391 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
394 * see scrub_pending_trans_workers_inc() why we're pretending
395 * to be paused in the scrub counters
397 mutex_lock(&fs_info->scrub_lock);
398 atomic_dec(&fs_info->scrubs_running);
399 atomic_dec(&fs_info->scrubs_paused);
400 mutex_unlock(&fs_info->scrub_lock);
401 atomic_dec(&sctx->workers_pending);
402 wake_up(&fs_info->scrub_pause_wait);
403 wake_up(&sctx->list_wait);
407 static void scrub_free_csums(struct scrub_ctx *sctx)
409 while (!list_empty(&sctx->csum_list)) {
410 struct btrfs_ordered_sum *sum;
411 sum = list_first_entry(&sctx->csum_list,
412 struct btrfs_ordered_sum, list);
413 list_del(&sum->list);
418 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
425 scrub_free_wr_ctx(&sctx->wr_ctx);
427 /* this can happen when scrub is cancelled */
428 if (sctx->curr != -1) {
429 struct scrub_bio *sbio = sctx->bios[sctx->curr];
431 for (i = 0; i < sbio->page_count; i++) {
432 WARN_ON(!sbio->pagev[i]->page);
433 scrub_block_put(sbio->pagev[i]->sblock);
438 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
439 struct scrub_bio *sbio = sctx->bios[i];
446 scrub_free_csums(sctx);
450 static void scrub_put_ctx(struct scrub_ctx *sctx)
452 if (atomic_dec_and_test(&sctx->refs))
453 scrub_free_ctx(sctx);
456 static noinline_for_stack
457 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
459 struct scrub_ctx *sctx;
461 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
464 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
467 atomic_set(&sctx->refs, 1);
468 sctx->is_dev_replace = is_dev_replace;
469 sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
471 sctx->dev_root = dev->dev_root;
472 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
473 struct scrub_bio *sbio;
475 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
478 sctx->bios[i] = sbio;
482 sbio->page_count = 0;
483 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
484 scrub_bio_end_io_worker, NULL, NULL);
486 if (i != SCRUB_BIOS_PER_SCTX - 1)
487 sctx->bios[i]->next_free = i + 1;
489 sctx->bios[i]->next_free = -1;
491 sctx->first_free = 0;
492 sctx->nodesize = dev->dev_root->nodesize;
493 sctx->sectorsize = dev->dev_root->sectorsize;
494 atomic_set(&sctx->bios_in_flight, 0);
495 atomic_set(&sctx->workers_pending, 0);
496 atomic_set(&sctx->cancel_req, 0);
497 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
498 INIT_LIST_HEAD(&sctx->csum_list);
500 spin_lock_init(&sctx->list_lock);
501 spin_lock_init(&sctx->stat_lock);
502 init_waitqueue_head(&sctx->list_wait);
504 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
505 fs_info->dev_replace.tgtdev, is_dev_replace);
507 scrub_free_ctx(sctx);
513 scrub_free_ctx(sctx);
514 return ERR_PTR(-ENOMEM);
517 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
524 struct extent_buffer *eb;
525 struct btrfs_inode_item *inode_item;
526 struct scrub_warning *swarn = warn_ctx;
527 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
528 struct inode_fs_paths *ipath = NULL;
529 struct btrfs_root *local_root;
530 struct btrfs_key root_key;
531 struct btrfs_key key;
533 root_key.objectid = root;
534 root_key.type = BTRFS_ROOT_ITEM_KEY;
535 root_key.offset = (u64)-1;
536 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
537 if (IS_ERR(local_root)) {
538 ret = PTR_ERR(local_root);
543 * this makes the path point to (inum INODE_ITEM ioff)
546 key.type = BTRFS_INODE_ITEM_KEY;
549 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
551 btrfs_release_path(swarn->path);
555 eb = swarn->path->nodes[0];
556 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
557 struct btrfs_inode_item);
558 isize = btrfs_inode_size(eb, inode_item);
559 nlink = btrfs_inode_nlink(eb, inode_item);
560 btrfs_release_path(swarn->path);
562 ipath = init_ipath(4096, local_root, swarn->path);
564 ret = PTR_ERR(ipath);
568 ret = paths_from_inode(inum, ipath);
574 * we deliberately ignore the bit ipath might have been too small to
575 * hold all of the paths here
577 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
578 btrfs_warn_in_rcu(fs_info,
579 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
580 swarn->errstr, swarn->logical,
581 rcu_str_deref(swarn->dev->name),
582 (unsigned long long)swarn->sector,
584 min(isize - offset, (u64)PAGE_SIZE), nlink,
585 (char *)(unsigned long)ipath->fspath->val[i]);
591 btrfs_warn_in_rcu(fs_info,
592 "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
593 swarn->errstr, swarn->logical,
594 rcu_str_deref(swarn->dev->name),
595 (unsigned long long)swarn->sector,
596 root, inum, offset, ret);
602 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
604 struct btrfs_device *dev;
605 struct btrfs_fs_info *fs_info;
606 struct btrfs_path *path;
607 struct btrfs_key found_key;
608 struct extent_buffer *eb;
609 struct btrfs_extent_item *ei;
610 struct scrub_warning swarn;
611 unsigned long ptr = 0;
619 WARN_ON(sblock->page_count < 1);
620 dev = sblock->pagev[0]->dev;
621 fs_info = sblock->sctx->dev_root->fs_info;
623 path = btrfs_alloc_path();
627 swarn.sector = (sblock->pagev[0]->physical) >> 9;
628 swarn.logical = sblock->pagev[0]->logical;
629 swarn.errstr = errstr;
632 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
637 extent_item_pos = swarn.logical - found_key.objectid;
638 swarn.extent_item_size = found_key.offset;
641 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
642 item_size = btrfs_item_size_nr(eb, path->slots[0]);
644 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
646 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
647 item_size, &ref_root,
649 btrfs_warn_in_rcu(fs_info,
650 "%s at logical %llu on dev %s, sector %llu: metadata %s (level %d) in tree %llu",
651 errstr, swarn.logical,
652 rcu_str_deref(dev->name),
653 (unsigned long long)swarn.sector,
654 ref_level ? "node" : "leaf",
655 ret < 0 ? -1 : ref_level,
656 ret < 0 ? -1 : ref_root);
658 btrfs_release_path(path);
660 btrfs_release_path(path);
663 iterate_extent_inodes(fs_info, found_key.objectid,
665 scrub_print_warning_inode, &swarn);
669 btrfs_free_path(path);
672 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
674 struct page *page = NULL;
676 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
679 struct btrfs_key key;
680 struct inode *inode = NULL;
681 struct btrfs_fs_info *fs_info;
682 u64 end = offset + PAGE_SIZE - 1;
683 struct btrfs_root *local_root;
687 key.type = BTRFS_ROOT_ITEM_KEY;
688 key.offset = (u64)-1;
690 fs_info = fixup->root->fs_info;
691 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
693 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
694 if (IS_ERR(local_root)) {
695 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
696 return PTR_ERR(local_root);
699 key.type = BTRFS_INODE_ITEM_KEY;
702 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
703 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
705 return PTR_ERR(inode);
707 index = offset >> PAGE_SHIFT;
709 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
715 if (PageUptodate(page)) {
716 if (PageDirty(page)) {
718 * we need to write the data to the defect sector. the
719 * data that was in that sector is not in memory,
720 * because the page was modified. we must not write the
721 * modified page to that sector.
723 * TODO: what could be done here: wait for the delalloc
724 * runner to write out that page (might involve
725 * COW) and see whether the sector is still
726 * referenced afterwards.
728 * For the meantime, we'll treat this error
729 * incorrectable, although there is a chance that a
730 * later scrub will find the bad sector again and that
731 * there's no dirty page in memory, then.
736 ret = repair_io_failure(inode, offset, PAGE_SIZE,
737 fixup->logical, page,
738 offset - page_offset(page),
744 * we need to get good data first. the general readpage path
745 * will call repair_io_failure for us, we just have to make
746 * sure we read the bad mirror.
748 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
751 /* set_extent_bits should give proper error */
758 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
761 wait_on_page_locked(page);
763 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
764 end, EXTENT_DAMAGED, 0, NULL);
766 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
779 if (ret == 0 && corrected) {
781 * we only need to call readpage for one of the inodes belonging
782 * to this extent. so make iterate_extent_inodes stop
790 static void scrub_fixup_nodatasum(struct btrfs_work *work)
793 struct scrub_fixup_nodatasum *fixup;
794 struct scrub_ctx *sctx;
795 struct btrfs_trans_handle *trans = NULL;
796 struct btrfs_path *path;
797 int uncorrectable = 0;
799 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
802 path = btrfs_alloc_path();
804 spin_lock(&sctx->stat_lock);
805 ++sctx->stat.malloc_errors;
806 spin_unlock(&sctx->stat_lock);
811 trans = btrfs_join_transaction(fixup->root);
818 * the idea is to trigger a regular read through the standard path. we
819 * read a page from the (failed) logical address by specifying the
820 * corresponding copynum of the failed sector. thus, that readpage is
822 * that is the point where on-the-fly error correction will kick in
823 * (once it's finished) and rewrite the failed sector if a good copy
826 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
827 path, scrub_fixup_readpage,
835 spin_lock(&sctx->stat_lock);
836 ++sctx->stat.corrected_errors;
837 spin_unlock(&sctx->stat_lock);
840 if (trans && !IS_ERR(trans))
841 btrfs_end_transaction(trans, fixup->root);
843 spin_lock(&sctx->stat_lock);
844 ++sctx->stat.uncorrectable_errors;
845 spin_unlock(&sctx->stat_lock);
846 btrfs_dev_replace_stats_inc(
847 &sctx->dev_root->fs_info->dev_replace.
848 num_uncorrectable_read_errors);
849 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
850 "unable to fixup (nodatasum) error at logical %llu on dev %s",
851 fixup->logical, rcu_str_deref(fixup->dev->name));
854 btrfs_free_path(path);
857 scrub_pending_trans_workers_dec(sctx);
860 static inline void scrub_get_recover(struct scrub_recover *recover)
862 atomic_inc(&recover->refs);
865 static inline void scrub_put_recover(struct scrub_recover *recover)
867 if (atomic_dec_and_test(&recover->refs)) {
868 btrfs_put_bbio(recover->bbio);
874 * scrub_handle_errored_block gets called when either verification of the
875 * pages failed or the bio failed to read, e.g. with EIO. In the latter
876 * case, this function handles all pages in the bio, even though only one
878 * The goal of this function is to repair the errored block by using the
879 * contents of one of the mirrors.
881 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
883 struct scrub_ctx *sctx = sblock_to_check->sctx;
884 struct btrfs_device *dev;
885 struct btrfs_fs_info *fs_info;
888 unsigned int failed_mirror_index;
889 unsigned int is_metadata;
890 unsigned int have_csum;
891 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
892 struct scrub_block *sblock_bad;
897 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
898 DEFAULT_RATELIMIT_BURST);
900 BUG_ON(sblock_to_check->page_count < 1);
901 fs_info = sctx->dev_root->fs_info;
902 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
904 * if we find an error in a super block, we just report it.
905 * They will get written with the next transaction commit
908 spin_lock(&sctx->stat_lock);
909 ++sctx->stat.super_errors;
910 spin_unlock(&sctx->stat_lock);
913 length = sblock_to_check->page_count * PAGE_SIZE;
914 logical = sblock_to_check->pagev[0]->logical;
915 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
916 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
917 is_metadata = !(sblock_to_check->pagev[0]->flags &
918 BTRFS_EXTENT_FLAG_DATA);
919 have_csum = sblock_to_check->pagev[0]->have_csum;
920 dev = sblock_to_check->pagev[0]->dev;
922 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
923 sblocks_for_recheck = NULL;
928 * read all mirrors one after the other. This includes to
929 * re-read the extent or metadata block that failed (that was
930 * the cause that this fixup code is called) another time,
931 * page by page this time in order to know which pages
932 * caused I/O errors and which ones are good (for all mirrors).
933 * It is the goal to handle the situation when more than one
934 * mirror contains I/O errors, but the errors do not
935 * overlap, i.e. the data can be repaired by selecting the
936 * pages from those mirrors without I/O error on the
937 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
938 * would be that mirror #1 has an I/O error on the first page,
939 * the second page is good, and mirror #2 has an I/O error on
940 * the second page, but the first page is good.
941 * Then the first page of the first mirror can be repaired by
942 * taking the first page of the second mirror, and the
943 * second page of the second mirror can be repaired by
944 * copying the contents of the 2nd page of the 1st mirror.
945 * One more note: if the pages of one mirror contain I/O
946 * errors, the checksum cannot be verified. In order to get
947 * the best data for repairing, the first attempt is to find
948 * a mirror without I/O errors and with a validated checksum.
949 * Only if this is not possible, the pages are picked from
950 * mirrors with I/O errors without considering the checksum.
951 * If the latter is the case, at the end, the checksum of the
952 * repaired area is verified in order to correctly maintain
956 sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
957 sizeof(*sblocks_for_recheck), GFP_NOFS);
958 if (!sblocks_for_recheck) {
959 spin_lock(&sctx->stat_lock);
960 sctx->stat.malloc_errors++;
961 sctx->stat.read_errors++;
962 sctx->stat.uncorrectable_errors++;
963 spin_unlock(&sctx->stat_lock);
964 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
968 /* setup the context, map the logical blocks and alloc the pages */
969 ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
971 spin_lock(&sctx->stat_lock);
972 sctx->stat.read_errors++;
973 sctx->stat.uncorrectable_errors++;
974 spin_unlock(&sctx->stat_lock);
975 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
978 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
979 sblock_bad = sblocks_for_recheck + failed_mirror_index;
981 /* build and submit the bios for the failed mirror, check checksums */
982 scrub_recheck_block(fs_info, sblock_bad, 1);
984 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
985 sblock_bad->no_io_error_seen) {
987 * the error disappeared after reading page by page, or
988 * the area was part of a huge bio and other parts of the
989 * bio caused I/O errors, or the block layer merged several
990 * read requests into one and the error is caused by a
991 * different bio (usually one of the two latter cases is
994 spin_lock(&sctx->stat_lock);
995 sctx->stat.unverified_errors++;
996 sblock_to_check->data_corrected = 1;
997 spin_unlock(&sctx->stat_lock);
999 if (sctx->is_dev_replace)
1000 scrub_write_block_to_dev_replace(sblock_bad);
1004 if (!sblock_bad->no_io_error_seen) {
1005 spin_lock(&sctx->stat_lock);
1006 sctx->stat.read_errors++;
1007 spin_unlock(&sctx->stat_lock);
1008 if (__ratelimit(&_rs))
1009 scrub_print_warning("i/o error", sblock_to_check);
1010 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1011 } else if (sblock_bad->checksum_error) {
1012 spin_lock(&sctx->stat_lock);
1013 sctx->stat.csum_errors++;
1014 spin_unlock(&sctx->stat_lock);
1015 if (__ratelimit(&_rs))
1016 scrub_print_warning("checksum error", sblock_to_check);
1017 btrfs_dev_stat_inc_and_print(dev,
1018 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1019 } else if (sblock_bad->header_error) {
1020 spin_lock(&sctx->stat_lock);
1021 sctx->stat.verify_errors++;
1022 spin_unlock(&sctx->stat_lock);
1023 if (__ratelimit(&_rs))
1024 scrub_print_warning("checksum/header error",
1026 if (sblock_bad->generation_error)
1027 btrfs_dev_stat_inc_and_print(dev,
1028 BTRFS_DEV_STAT_GENERATION_ERRS);
1030 btrfs_dev_stat_inc_and_print(dev,
1031 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1034 if (sctx->readonly) {
1035 ASSERT(!sctx->is_dev_replace);
1039 if (!is_metadata && !have_csum) {
1040 struct scrub_fixup_nodatasum *fixup_nodatasum;
1042 WARN_ON(sctx->is_dev_replace);
1047 * !is_metadata and !have_csum, this means that the data
1048 * might not be COWed, that it might be modified
1049 * concurrently. The general strategy to work on the
1050 * commit root does not help in the case when COW is not
1053 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1054 if (!fixup_nodatasum)
1055 goto did_not_correct_error;
1056 fixup_nodatasum->sctx = sctx;
1057 fixup_nodatasum->dev = dev;
1058 fixup_nodatasum->logical = logical;
1059 fixup_nodatasum->root = fs_info->extent_root;
1060 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1061 scrub_pending_trans_workers_inc(sctx);
1062 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1063 scrub_fixup_nodatasum, NULL, NULL);
1064 btrfs_queue_work(fs_info->scrub_workers,
1065 &fixup_nodatasum->work);
1070 * now build and submit the bios for the other mirrors, check
1072 * First try to pick the mirror which is completely without I/O
1073 * errors and also does not have a checksum error.
1074 * If one is found, and if a checksum is present, the full block
1075 * that is known to contain an error is rewritten. Afterwards
1076 * the block is known to be corrected.
1077 * If a mirror is found which is completely correct, and no
1078 * checksum is present, only those pages are rewritten that had
1079 * an I/O error in the block to be repaired, since it cannot be
1080 * determined, which copy of the other pages is better (and it
1081 * could happen otherwise that a correct page would be
1082 * overwritten by a bad one).
1084 for (mirror_index = 0;
1085 mirror_index < BTRFS_MAX_MIRRORS &&
1086 sblocks_for_recheck[mirror_index].page_count > 0;
1088 struct scrub_block *sblock_other;
1090 if (mirror_index == failed_mirror_index)
1092 sblock_other = sblocks_for_recheck + mirror_index;
1094 /* build and submit the bios, check checksums */
1095 scrub_recheck_block(fs_info, sblock_other, 0);
1097 if (!sblock_other->header_error &&
1098 !sblock_other->checksum_error &&
1099 sblock_other->no_io_error_seen) {
1100 if (sctx->is_dev_replace) {
1101 scrub_write_block_to_dev_replace(sblock_other);
1102 goto corrected_error;
1104 ret = scrub_repair_block_from_good_copy(
1105 sblock_bad, sblock_other);
1107 goto corrected_error;
1112 if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1113 goto did_not_correct_error;
1116 * In case of I/O errors in the area that is supposed to be
1117 * repaired, continue by picking good copies of those pages.
1118 * Select the good pages from mirrors to rewrite bad pages from
1119 * the area to fix. Afterwards verify the checksum of the block
1120 * that is supposed to be repaired. This verification step is
1121 * only done for the purpose of statistic counting and for the
1122 * final scrub report, whether errors remain.
1123 * A perfect algorithm could make use of the checksum and try
1124 * all possible combinations of pages from the different mirrors
1125 * until the checksum verification succeeds. For example, when
1126 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1127 * of mirror #2 is readable but the final checksum test fails,
1128 * then the 2nd page of mirror #3 could be tried, whether now
1129 * the final checksum succeeds. But this would be a rare
1130 * exception and is therefore not implemented. At least it is
1131 * avoided that the good copy is overwritten.
1132 * A more useful improvement would be to pick the sectors
1133 * without I/O error based on sector sizes (512 bytes on legacy
1134 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1135 * mirror could be repaired by taking 512 byte of a different
1136 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1137 * area are unreadable.
1140 for (page_num = 0; page_num < sblock_bad->page_count;
1142 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1143 struct scrub_block *sblock_other = NULL;
1145 /* skip no-io-error page in scrub */
1146 if (!page_bad->io_error && !sctx->is_dev_replace)
1149 /* try to find no-io-error page in mirrors */
1150 if (page_bad->io_error) {
1151 for (mirror_index = 0;
1152 mirror_index < BTRFS_MAX_MIRRORS &&
1153 sblocks_for_recheck[mirror_index].page_count > 0;
1155 if (!sblocks_for_recheck[mirror_index].
1156 pagev[page_num]->io_error) {
1157 sblock_other = sblocks_for_recheck +
1166 if (sctx->is_dev_replace) {
1168 * did not find a mirror to fetch the page
1169 * from. scrub_write_page_to_dev_replace()
1170 * handles this case (page->io_error), by
1171 * filling the block with zeros before
1172 * submitting the write request
1175 sblock_other = sblock_bad;
1177 if (scrub_write_page_to_dev_replace(sblock_other,
1179 btrfs_dev_replace_stats_inc(
1181 fs_info->dev_replace.
1185 } else if (sblock_other) {
1186 ret = scrub_repair_page_from_good_copy(sblock_bad,
1190 page_bad->io_error = 0;
1196 if (success && !sctx->is_dev_replace) {
1197 if (is_metadata || have_csum) {
1199 * need to verify the checksum now that all
1200 * sectors on disk are repaired (the write
1201 * request for data to be repaired is on its way).
1202 * Just be lazy and use scrub_recheck_block()
1203 * which re-reads the data before the checksum
1204 * is verified, but most likely the data comes out
1205 * of the page cache.
1207 scrub_recheck_block(fs_info, sblock_bad, 1);
1208 if (!sblock_bad->header_error &&
1209 !sblock_bad->checksum_error &&
1210 sblock_bad->no_io_error_seen)
1211 goto corrected_error;
1213 goto did_not_correct_error;
1216 spin_lock(&sctx->stat_lock);
1217 sctx->stat.corrected_errors++;
1218 sblock_to_check->data_corrected = 1;
1219 spin_unlock(&sctx->stat_lock);
1220 btrfs_err_rl_in_rcu(fs_info,
1221 "fixed up error at logical %llu on dev %s",
1222 logical, rcu_str_deref(dev->name));
1225 did_not_correct_error:
1226 spin_lock(&sctx->stat_lock);
1227 sctx->stat.uncorrectable_errors++;
1228 spin_unlock(&sctx->stat_lock);
1229 btrfs_err_rl_in_rcu(fs_info,
1230 "unable to fixup (regular) error at logical %llu on dev %s",
1231 logical, rcu_str_deref(dev->name));
1235 if (sblocks_for_recheck) {
1236 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1238 struct scrub_block *sblock = sblocks_for_recheck +
1240 struct scrub_recover *recover;
1243 for (page_index = 0; page_index < sblock->page_count;
1245 sblock->pagev[page_index]->sblock = NULL;
1246 recover = sblock->pagev[page_index]->recover;
1248 scrub_put_recover(recover);
1249 sblock->pagev[page_index]->recover =
1252 scrub_page_put(sblock->pagev[page_index]);
1255 kfree(sblocks_for_recheck);
1261 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1263 if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1265 else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1268 return (int)bbio->num_stripes;
1271 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1274 int nstripes, int mirror,
1280 if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1282 for (i = 0; i < nstripes; i++) {
1283 if (raid_map[i] == RAID6_Q_STRIPE ||
1284 raid_map[i] == RAID5_P_STRIPE)
1287 if (logical >= raid_map[i] &&
1288 logical < raid_map[i] + mapped_length)
1293 *stripe_offset = logical - raid_map[i];
1295 /* The other RAID type */
1296 *stripe_index = mirror;
1301 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1302 struct scrub_block *sblocks_for_recheck)
1304 struct scrub_ctx *sctx = original_sblock->sctx;
1305 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
1306 u64 length = original_sblock->page_count * PAGE_SIZE;
1307 u64 logical = original_sblock->pagev[0]->logical;
1308 u64 generation = original_sblock->pagev[0]->generation;
1309 u64 flags = original_sblock->pagev[0]->flags;
1310 u64 have_csum = original_sblock->pagev[0]->have_csum;
1311 struct scrub_recover *recover;
1312 struct btrfs_bio *bbio;
1323 * note: the two members refs and outstanding_pages
1324 * are not used (and not set) in the blocks that are used for
1325 * the recheck procedure
1328 while (length > 0) {
1329 sublen = min_t(u64, length, PAGE_SIZE);
1330 mapped_length = sublen;
1334 * with a length of PAGE_SIZE, each returned stripe
1335 * represents one mirror
1337 ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical,
1338 &mapped_length, &bbio, 0, 1);
1339 if (ret || !bbio || mapped_length < sublen) {
1340 btrfs_put_bbio(bbio);
1344 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1346 btrfs_put_bbio(bbio);
1350 atomic_set(&recover->refs, 1);
1351 recover->bbio = bbio;
1352 recover->map_length = mapped_length;
1354 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1356 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1358 for (mirror_index = 0; mirror_index < nmirrors;
1360 struct scrub_block *sblock;
1361 struct scrub_page *page;
1363 sblock = sblocks_for_recheck + mirror_index;
1364 sblock->sctx = sctx;
1366 page = kzalloc(sizeof(*page), GFP_NOFS);
1369 spin_lock(&sctx->stat_lock);
1370 sctx->stat.malloc_errors++;
1371 spin_unlock(&sctx->stat_lock);
1372 scrub_put_recover(recover);
1375 scrub_page_get(page);
1376 sblock->pagev[page_index] = page;
1377 page->sblock = sblock;
1378 page->flags = flags;
1379 page->generation = generation;
1380 page->logical = logical;
1381 page->have_csum = have_csum;
1384 original_sblock->pagev[0]->csum,
1387 scrub_stripe_index_and_offset(logical,
1396 page->physical = bbio->stripes[stripe_index].physical +
1398 page->dev = bbio->stripes[stripe_index].dev;
1400 BUG_ON(page_index >= original_sblock->page_count);
1401 page->physical_for_dev_replace =
1402 original_sblock->pagev[page_index]->
1403 physical_for_dev_replace;
1404 /* for missing devices, dev->bdev is NULL */
1405 page->mirror_num = mirror_index + 1;
1406 sblock->page_count++;
1407 page->page = alloc_page(GFP_NOFS);
1411 scrub_get_recover(recover);
1412 page->recover = recover;
1414 scrub_put_recover(recover);
1423 struct scrub_bio_ret {
1424 struct completion event;
1428 static void scrub_bio_wait_endio(struct bio *bio)
1430 struct scrub_bio_ret *ret = bio->bi_private;
1432 ret->error = bio->bi_error;
1433 complete(&ret->event);
1436 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1438 return page->recover &&
1439 (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1442 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1444 struct scrub_page *page)
1446 struct scrub_bio_ret done;
1449 init_completion(&done.event);
1451 bio->bi_iter.bi_sector = page->logical >> 9;
1452 bio->bi_private = &done;
1453 bio->bi_end_io = scrub_bio_wait_endio;
1455 ret = raid56_parity_recover(fs_info->fs_root, bio, page->recover->bbio,
1456 page->recover->map_length,
1457 page->mirror_num, 0);
1461 wait_for_completion(&done.event);
1469 * this function will check the on disk data for checksum errors, header
1470 * errors and read I/O errors. If any I/O errors happen, the exact pages
1471 * which are errored are marked as being bad. The goal is to enable scrub
1472 * to take those pages that are not errored from all the mirrors so that
1473 * the pages that are errored in the just handled mirror can be repaired.
1475 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1476 struct scrub_block *sblock,
1477 int retry_failed_mirror)
1481 sblock->no_io_error_seen = 1;
1483 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1485 struct scrub_page *page = sblock->pagev[page_num];
1487 if (page->dev->bdev == NULL) {
1489 sblock->no_io_error_seen = 0;
1493 WARN_ON(!page->page);
1494 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1497 sblock->no_io_error_seen = 0;
1500 bio->bi_bdev = page->dev->bdev;
1502 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1503 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1504 if (scrub_submit_raid56_bio_wait(fs_info, bio, page))
1505 sblock->no_io_error_seen = 0;
1507 bio->bi_iter.bi_sector = page->physical >> 9;
1508 bio_set_op_attrs(bio, REQ_OP_READ, 0);
1510 if (btrfsic_submit_bio_wait(bio))
1511 sblock->no_io_error_seen = 0;
1517 if (sblock->no_io_error_seen)
1518 scrub_recheck_block_checksum(sblock);
1521 static inline int scrub_check_fsid(u8 fsid[],
1522 struct scrub_page *spage)
1524 struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1527 ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1531 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1533 sblock->header_error = 0;
1534 sblock->checksum_error = 0;
1535 sblock->generation_error = 0;
1537 if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1538 scrub_checksum_data(sblock);
1540 scrub_checksum_tree_block(sblock);
1543 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1544 struct scrub_block *sblock_good)
1549 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1552 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1562 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1563 struct scrub_block *sblock_good,
1564 int page_num, int force_write)
1566 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1567 struct scrub_page *page_good = sblock_good->pagev[page_num];
1569 BUG_ON(page_bad->page == NULL);
1570 BUG_ON(page_good->page == NULL);
1571 if (force_write || sblock_bad->header_error ||
1572 sblock_bad->checksum_error || page_bad->io_error) {
1576 if (!page_bad->dev->bdev) {
1577 btrfs_warn_rl(sblock_bad->sctx->dev_root->fs_info,
1578 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1582 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1585 bio->bi_bdev = page_bad->dev->bdev;
1586 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1587 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1589 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1590 if (PAGE_SIZE != ret) {
1595 if (btrfsic_submit_bio_wait(bio)) {
1596 btrfs_dev_stat_inc_and_print(page_bad->dev,
1597 BTRFS_DEV_STAT_WRITE_ERRS);
1598 btrfs_dev_replace_stats_inc(
1599 &sblock_bad->sctx->dev_root->fs_info->
1600 dev_replace.num_write_errors);
1610 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1615 * This block is used for the check of the parity on the source device,
1616 * so the data needn't be written into the destination device.
1618 if (sblock->sparity)
1621 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1624 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1626 btrfs_dev_replace_stats_inc(
1627 &sblock->sctx->dev_root->fs_info->dev_replace.
1632 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1635 struct scrub_page *spage = sblock->pagev[page_num];
1637 BUG_ON(spage->page == NULL);
1638 if (spage->io_error) {
1639 void *mapped_buffer = kmap_atomic(spage->page);
1641 memset(mapped_buffer, 0, PAGE_SIZE);
1642 flush_dcache_page(spage->page);
1643 kunmap_atomic(mapped_buffer);
1645 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1648 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1649 struct scrub_page *spage)
1651 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1652 struct scrub_bio *sbio;
1655 mutex_lock(&wr_ctx->wr_lock);
1657 if (!wr_ctx->wr_curr_bio) {
1658 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1660 if (!wr_ctx->wr_curr_bio) {
1661 mutex_unlock(&wr_ctx->wr_lock);
1664 wr_ctx->wr_curr_bio->sctx = sctx;
1665 wr_ctx->wr_curr_bio->page_count = 0;
1667 sbio = wr_ctx->wr_curr_bio;
1668 if (sbio->page_count == 0) {
1671 sbio->physical = spage->physical_for_dev_replace;
1672 sbio->logical = spage->logical;
1673 sbio->dev = wr_ctx->tgtdev;
1676 bio = btrfs_io_bio_alloc(GFP_KERNEL,
1677 wr_ctx->pages_per_wr_bio);
1679 mutex_unlock(&wr_ctx->wr_lock);
1685 bio->bi_private = sbio;
1686 bio->bi_end_io = scrub_wr_bio_end_io;
1687 bio->bi_bdev = sbio->dev->bdev;
1688 bio->bi_iter.bi_sector = sbio->physical >> 9;
1689 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1691 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1692 spage->physical_for_dev_replace ||
1693 sbio->logical + sbio->page_count * PAGE_SIZE !=
1695 scrub_wr_submit(sctx);
1699 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1700 if (ret != PAGE_SIZE) {
1701 if (sbio->page_count < 1) {
1704 mutex_unlock(&wr_ctx->wr_lock);
1707 scrub_wr_submit(sctx);
1711 sbio->pagev[sbio->page_count] = spage;
1712 scrub_page_get(spage);
1714 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1715 scrub_wr_submit(sctx);
1716 mutex_unlock(&wr_ctx->wr_lock);
1721 static void scrub_wr_submit(struct scrub_ctx *sctx)
1723 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1724 struct scrub_bio *sbio;
1726 if (!wr_ctx->wr_curr_bio)
1729 sbio = wr_ctx->wr_curr_bio;
1730 wr_ctx->wr_curr_bio = NULL;
1731 WARN_ON(!sbio->bio->bi_bdev);
1732 scrub_pending_bio_inc(sctx);
1733 /* process all writes in a single worker thread. Then the block layer
1734 * orders the requests before sending them to the driver which
1735 * doubled the write performance on spinning disks when measured
1737 btrfsic_submit_bio(sbio->bio);
1740 static void scrub_wr_bio_end_io(struct bio *bio)
1742 struct scrub_bio *sbio = bio->bi_private;
1743 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1745 sbio->err = bio->bi_error;
1748 btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1749 scrub_wr_bio_end_io_worker, NULL, NULL);
1750 btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1753 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1755 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1756 struct scrub_ctx *sctx = sbio->sctx;
1759 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1761 struct btrfs_dev_replace *dev_replace =
1762 &sbio->sctx->dev_root->fs_info->dev_replace;
1764 for (i = 0; i < sbio->page_count; i++) {
1765 struct scrub_page *spage = sbio->pagev[i];
1767 spage->io_error = 1;
1768 btrfs_dev_replace_stats_inc(&dev_replace->
1773 for (i = 0; i < sbio->page_count; i++)
1774 scrub_page_put(sbio->pagev[i]);
1778 scrub_pending_bio_dec(sctx);
1781 static int scrub_checksum(struct scrub_block *sblock)
1787 * No need to initialize these stats currently,
1788 * because this function only use return value
1789 * instead of these stats value.
1794 sblock->header_error = 0;
1795 sblock->generation_error = 0;
1796 sblock->checksum_error = 0;
1798 WARN_ON(sblock->page_count < 1);
1799 flags = sblock->pagev[0]->flags;
1801 if (flags & BTRFS_EXTENT_FLAG_DATA)
1802 ret = scrub_checksum_data(sblock);
1803 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1804 ret = scrub_checksum_tree_block(sblock);
1805 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1806 (void)scrub_checksum_super(sblock);
1810 scrub_handle_errored_block(sblock);
1815 static int scrub_checksum_data(struct scrub_block *sblock)
1817 struct scrub_ctx *sctx = sblock->sctx;
1818 u8 csum[BTRFS_CSUM_SIZE];
1826 BUG_ON(sblock->page_count < 1);
1827 if (!sblock->pagev[0]->have_csum)
1830 on_disk_csum = sblock->pagev[0]->csum;
1831 page = sblock->pagev[0]->page;
1832 buffer = kmap_atomic(page);
1834 len = sctx->sectorsize;
1837 u64 l = min_t(u64, len, PAGE_SIZE);
1839 crc = btrfs_csum_data(buffer, crc, l);
1840 kunmap_atomic(buffer);
1845 BUG_ON(index >= sblock->page_count);
1846 BUG_ON(!sblock->pagev[index]->page);
1847 page = sblock->pagev[index]->page;
1848 buffer = kmap_atomic(page);
1851 btrfs_csum_final(crc, csum);
1852 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1853 sblock->checksum_error = 1;
1855 return sblock->checksum_error;
1858 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1860 struct scrub_ctx *sctx = sblock->sctx;
1861 struct btrfs_header *h;
1862 struct btrfs_root *root = sctx->dev_root;
1863 struct btrfs_fs_info *fs_info = root->fs_info;
1864 u8 calculated_csum[BTRFS_CSUM_SIZE];
1865 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1867 void *mapped_buffer;
1874 BUG_ON(sblock->page_count < 1);
1875 page = sblock->pagev[0]->page;
1876 mapped_buffer = kmap_atomic(page);
1877 h = (struct btrfs_header *)mapped_buffer;
1878 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1881 * we don't use the getter functions here, as we
1882 * a) don't have an extent buffer and
1883 * b) the page is already kmapped
1885 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1886 sblock->header_error = 1;
1888 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1889 sblock->header_error = 1;
1890 sblock->generation_error = 1;
1893 if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1894 sblock->header_error = 1;
1896 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1898 sblock->header_error = 1;
1900 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1901 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1902 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1905 u64 l = min_t(u64, len, mapped_size);
1907 crc = btrfs_csum_data(p, crc, l);
1908 kunmap_atomic(mapped_buffer);
1913 BUG_ON(index >= sblock->page_count);
1914 BUG_ON(!sblock->pagev[index]->page);
1915 page = sblock->pagev[index]->page;
1916 mapped_buffer = kmap_atomic(page);
1917 mapped_size = PAGE_SIZE;
1921 btrfs_csum_final(crc, calculated_csum);
1922 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1923 sblock->checksum_error = 1;
1925 return sblock->header_error || sblock->checksum_error;
1928 static int scrub_checksum_super(struct scrub_block *sblock)
1930 struct btrfs_super_block *s;
1931 struct scrub_ctx *sctx = sblock->sctx;
1932 u8 calculated_csum[BTRFS_CSUM_SIZE];
1933 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1935 void *mapped_buffer;
1944 BUG_ON(sblock->page_count < 1);
1945 page = sblock->pagev[0]->page;
1946 mapped_buffer = kmap_atomic(page);
1947 s = (struct btrfs_super_block *)mapped_buffer;
1948 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1950 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1953 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1956 if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1959 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1960 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1961 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1964 u64 l = min_t(u64, len, mapped_size);
1966 crc = btrfs_csum_data(p, crc, l);
1967 kunmap_atomic(mapped_buffer);
1972 BUG_ON(index >= sblock->page_count);
1973 BUG_ON(!sblock->pagev[index]->page);
1974 page = sblock->pagev[index]->page;
1975 mapped_buffer = kmap_atomic(page);
1976 mapped_size = PAGE_SIZE;
1980 btrfs_csum_final(crc, calculated_csum);
1981 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1984 if (fail_cor + fail_gen) {
1986 * if we find an error in a super block, we just report it.
1987 * They will get written with the next transaction commit
1990 spin_lock(&sctx->stat_lock);
1991 ++sctx->stat.super_errors;
1992 spin_unlock(&sctx->stat_lock);
1994 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1995 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1997 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1998 BTRFS_DEV_STAT_GENERATION_ERRS);
2001 return fail_cor + fail_gen;
2004 static void scrub_block_get(struct scrub_block *sblock)
2006 atomic_inc(&sblock->refs);
2009 static void scrub_block_put(struct scrub_block *sblock)
2011 if (atomic_dec_and_test(&sblock->refs)) {
2014 if (sblock->sparity)
2015 scrub_parity_put(sblock->sparity);
2017 for (i = 0; i < sblock->page_count; i++)
2018 scrub_page_put(sblock->pagev[i]);
2023 static void scrub_page_get(struct scrub_page *spage)
2025 atomic_inc(&spage->refs);
2028 static void scrub_page_put(struct scrub_page *spage)
2030 if (atomic_dec_and_test(&spage->refs)) {
2032 __free_page(spage->page);
2037 static void scrub_submit(struct scrub_ctx *sctx)
2039 struct scrub_bio *sbio;
2041 if (sctx->curr == -1)
2044 sbio = sctx->bios[sctx->curr];
2046 scrub_pending_bio_inc(sctx);
2047 btrfsic_submit_bio(sbio->bio);
2050 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2051 struct scrub_page *spage)
2053 struct scrub_block *sblock = spage->sblock;
2054 struct scrub_bio *sbio;
2059 * grab a fresh bio or wait for one to become available
2061 while (sctx->curr == -1) {
2062 spin_lock(&sctx->list_lock);
2063 sctx->curr = sctx->first_free;
2064 if (sctx->curr != -1) {
2065 sctx->first_free = sctx->bios[sctx->curr]->next_free;
2066 sctx->bios[sctx->curr]->next_free = -1;
2067 sctx->bios[sctx->curr]->page_count = 0;
2068 spin_unlock(&sctx->list_lock);
2070 spin_unlock(&sctx->list_lock);
2071 wait_event(sctx->list_wait, sctx->first_free != -1);
2074 sbio = sctx->bios[sctx->curr];
2075 if (sbio->page_count == 0) {
2078 sbio->physical = spage->physical;
2079 sbio->logical = spage->logical;
2080 sbio->dev = spage->dev;
2083 bio = btrfs_io_bio_alloc(GFP_KERNEL,
2084 sctx->pages_per_rd_bio);
2090 bio->bi_private = sbio;
2091 bio->bi_end_io = scrub_bio_end_io;
2092 bio->bi_bdev = sbio->dev->bdev;
2093 bio->bi_iter.bi_sector = sbio->physical >> 9;
2094 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2096 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2098 sbio->logical + sbio->page_count * PAGE_SIZE !=
2100 sbio->dev != spage->dev) {
2105 sbio->pagev[sbio->page_count] = spage;
2106 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2107 if (ret != PAGE_SIZE) {
2108 if (sbio->page_count < 1) {
2117 scrub_block_get(sblock); /* one for the page added to the bio */
2118 atomic_inc(&sblock->outstanding_pages);
2120 if (sbio->page_count == sctx->pages_per_rd_bio)
2126 static void scrub_missing_raid56_end_io(struct bio *bio)
2128 struct scrub_block *sblock = bio->bi_private;
2129 struct btrfs_fs_info *fs_info = sblock->sctx->dev_root->fs_info;
2132 sblock->no_io_error_seen = 0;
2136 btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2139 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2141 struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2142 struct scrub_ctx *sctx = sblock->sctx;
2144 struct btrfs_device *dev;
2146 logical = sblock->pagev[0]->logical;
2147 dev = sblock->pagev[0]->dev;
2149 if (sblock->no_io_error_seen)
2150 scrub_recheck_block_checksum(sblock);
2152 if (!sblock->no_io_error_seen) {
2153 spin_lock(&sctx->stat_lock);
2154 sctx->stat.read_errors++;
2155 spin_unlock(&sctx->stat_lock);
2156 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2157 "IO error rebuilding logical %llu for dev %s",
2158 logical, rcu_str_deref(dev->name));
2159 } else if (sblock->header_error || sblock->checksum_error) {
2160 spin_lock(&sctx->stat_lock);
2161 sctx->stat.uncorrectable_errors++;
2162 spin_unlock(&sctx->stat_lock);
2163 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2164 "failed to rebuild valid logical %llu for dev %s",
2165 logical, rcu_str_deref(dev->name));
2167 scrub_write_block_to_dev_replace(sblock);
2170 scrub_block_put(sblock);
2172 if (sctx->is_dev_replace &&
2173 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2174 mutex_lock(&sctx->wr_ctx.wr_lock);
2175 scrub_wr_submit(sctx);
2176 mutex_unlock(&sctx->wr_ctx.wr_lock);
2179 scrub_pending_bio_dec(sctx);
2182 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2184 struct scrub_ctx *sctx = sblock->sctx;
2185 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2186 u64 length = sblock->page_count * PAGE_SIZE;
2187 u64 logical = sblock->pagev[0]->logical;
2188 struct btrfs_bio *bbio = NULL;
2190 struct btrfs_raid_bio *rbio;
2194 ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
2196 if (ret || !bbio || !bbio->raid_map)
2199 if (WARN_ON(!sctx->is_dev_replace ||
2200 !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2202 * We shouldn't be scrubbing a missing device. Even for dev
2203 * replace, we should only get here for RAID 5/6. We either
2204 * managed to mount something with no mirrors remaining or
2205 * there's a bug in scrub_remap_extent()/btrfs_map_block().
2210 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2214 bio->bi_iter.bi_sector = logical >> 9;
2215 bio->bi_private = sblock;
2216 bio->bi_end_io = scrub_missing_raid56_end_io;
2218 rbio = raid56_alloc_missing_rbio(sctx->dev_root, bio, bbio, length);
2222 for (i = 0; i < sblock->page_count; i++) {
2223 struct scrub_page *spage = sblock->pagev[i];
2225 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2228 btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2229 scrub_missing_raid56_worker, NULL, NULL);
2230 scrub_block_get(sblock);
2231 scrub_pending_bio_inc(sctx);
2232 raid56_submit_missing_rbio(rbio);
2238 btrfs_put_bbio(bbio);
2239 spin_lock(&sctx->stat_lock);
2240 sctx->stat.malloc_errors++;
2241 spin_unlock(&sctx->stat_lock);
2244 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2245 u64 physical, struct btrfs_device *dev, u64 flags,
2246 u64 gen, int mirror_num, u8 *csum, int force,
2247 u64 physical_for_dev_replace)
2249 struct scrub_block *sblock;
2252 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2254 spin_lock(&sctx->stat_lock);
2255 sctx->stat.malloc_errors++;
2256 spin_unlock(&sctx->stat_lock);
2260 /* one ref inside this function, plus one for each page added to
2262 atomic_set(&sblock->refs, 1);
2263 sblock->sctx = sctx;
2264 sblock->no_io_error_seen = 1;
2266 for (index = 0; len > 0; index++) {
2267 struct scrub_page *spage;
2268 u64 l = min_t(u64, len, PAGE_SIZE);
2270 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2273 spin_lock(&sctx->stat_lock);
2274 sctx->stat.malloc_errors++;
2275 spin_unlock(&sctx->stat_lock);
2276 scrub_block_put(sblock);
2279 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2280 scrub_page_get(spage);
2281 sblock->pagev[index] = spage;
2282 spage->sblock = sblock;
2284 spage->flags = flags;
2285 spage->generation = gen;
2286 spage->logical = logical;
2287 spage->physical = physical;
2288 spage->physical_for_dev_replace = physical_for_dev_replace;
2289 spage->mirror_num = mirror_num;
2291 spage->have_csum = 1;
2292 memcpy(spage->csum, csum, sctx->csum_size);
2294 spage->have_csum = 0;
2296 sblock->page_count++;
2297 spage->page = alloc_page(GFP_KERNEL);
2303 physical_for_dev_replace += l;
2306 WARN_ON(sblock->page_count == 0);
2309 * This case should only be hit for RAID 5/6 device replace. See
2310 * the comment in scrub_missing_raid56_pages() for details.
2312 scrub_missing_raid56_pages(sblock);
2314 for (index = 0; index < sblock->page_count; index++) {
2315 struct scrub_page *spage = sblock->pagev[index];
2318 ret = scrub_add_page_to_rd_bio(sctx, spage);
2320 scrub_block_put(sblock);
2329 /* last one frees, either here or in bio completion for last page */
2330 scrub_block_put(sblock);
2334 static void scrub_bio_end_io(struct bio *bio)
2336 struct scrub_bio *sbio = bio->bi_private;
2337 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2339 sbio->err = bio->bi_error;
2342 btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2345 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2347 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2348 struct scrub_ctx *sctx = sbio->sctx;
2351 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2353 for (i = 0; i < sbio->page_count; i++) {
2354 struct scrub_page *spage = sbio->pagev[i];
2356 spage->io_error = 1;
2357 spage->sblock->no_io_error_seen = 0;
2361 /* now complete the scrub_block items that have all pages completed */
2362 for (i = 0; i < sbio->page_count; i++) {
2363 struct scrub_page *spage = sbio->pagev[i];
2364 struct scrub_block *sblock = spage->sblock;
2366 if (atomic_dec_and_test(&sblock->outstanding_pages))
2367 scrub_block_complete(sblock);
2368 scrub_block_put(sblock);
2373 spin_lock(&sctx->list_lock);
2374 sbio->next_free = sctx->first_free;
2375 sctx->first_free = sbio->index;
2376 spin_unlock(&sctx->list_lock);
2378 if (sctx->is_dev_replace &&
2379 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2380 mutex_lock(&sctx->wr_ctx.wr_lock);
2381 scrub_wr_submit(sctx);
2382 mutex_unlock(&sctx->wr_ctx.wr_lock);
2385 scrub_pending_bio_dec(sctx);
2388 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2389 unsigned long *bitmap,
2394 int sectorsize = sparity->sctx->dev_root->sectorsize;
2396 if (len >= sparity->stripe_len) {
2397 bitmap_set(bitmap, 0, sparity->nsectors);
2401 start -= sparity->logic_start;
2402 start = div_u64_rem(start, sparity->stripe_len, &offset);
2403 offset /= sectorsize;
2404 nsectors = (int)len / sectorsize;
2406 if (offset + nsectors <= sparity->nsectors) {
2407 bitmap_set(bitmap, offset, nsectors);
2411 bitmap_set(bitmap, offset, sparity->nsectors - offset);
2412 bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2415 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2418 __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2421 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2424 __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2427 static void scrub_block_complete(struct scrub_block *sblock)
2431 if (!sblock->no_io_error_seen) {
2433 scrub_handle_errored_block(sblock);
2436 * if has checksum error, write via repair mechanism in
2437 * dev replace case, otherwise write here in dev replace
2440 corrupted = scrub_checksum(sblock);
2441 if (!corrupted && sblock->sctx->is_dev_replace)
2442 scrub_write_block_to_dev_replace(sblock);
2445 if (sblock->sparity && corrupted && !sblock->data_corrected) {
2446 u64 start = sblock->pagev[0]->logical;
2447 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2450 scrub_parity_mark_sectors_error(sblock->sparity,
2451 start, end - start);
2455 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2457 struct btrfs_ordered_sum *sum = NULL;
2458 unsigned long index;
2459 unsigned long num_sectors;
2461 while (!list_empty(&sctx->csum_list)) {
2462 sum = list_first_entry(&sctx->csum_list,
2463 struct btrfs_ordered_sum, list);
2464 if (sum->bytenr > logical)
2466 if (sum->bytenr + sum->len > logical)
2469 ++sctx->stat.csum_discards;
2470 list_del(&sum->list);
2477 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2478 num_sectors = sum->len / sctx->sectorsize;
2479 memcpy(csum, sum->sums + index, sctx->csum_size);
2480 if (index == num_sectors - 1) {
2481 list_del(&sum->list);
2487 /* scrub extent tries to collect up to 64 kB for each bio */
2488 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2489 u64 physical, struct btrfs_device *dev, u64 flags,
2490 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2493 u8 csum[BTRFS_CSUM_SIZE];
2496 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2497 blocksize = sctx->sectorsize;
2498 spin_lock(&sctx->stat_lock);
2499 sctx->stat.data_extents_scrubbed++;
2500 sctx->stat.data_bytes_scrubbed += len;
2501 spin_unlock(&sctx->stat_lock);
2502 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2503 blocksize = sctx->nodesize;
2504 spin_lock(&sctx->stat_lock);
2505 sctx->stat.tree_extents_scrubbed++;
2506 sctx->stat.tree_bytes_scrubbed += len;
2507 spin_unlock(&sctx->stat_lock);
2509 blocksize = sctx->sectorsize;
2514 u64 l = min_t(u64, len, blocksize);
2517 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2518 /* push csums to sbio */
2519 have_csum = scrub_find_csum(sctx, logical, csum);
2521 ++sctx->stat.no_csum;
2522 if (sctx->is_dev_replace && !have_csum) {
2523 ret = copy_nocow_pages(sctx, logical, l,
2525 physical_for_dev_replace);
2526 goto behind_scrub_pages;
2529 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2530 mirror_num, have_csum ? csum : NULL, 0,
2531 physical_for_dev_replace);
2538 physical_for_dev_replace += l;
2543 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2544 u64 logical, u64 len,
2545 u64 physical, struct btrfs_device *dev,
2546 u64 flags, u64 gen, int mirror_num, u8 *csum)
2548 struct scrub_ctx *sctx = sparity->sctx;
2549 struct scrub_block *sblock;
2552 sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2554 spin_lock(&sctx->stat_lock);
2555 sctx->stat.malloc_errors++;
2556 spin_unlock(&sctx->stat_lock);
2560 /* one ref inside this function, plus one for each page added to
2562 atomic_set(&sblock->refs, 1);
2563 sblock->sctx = sctx;
2564 sblock->no_io_error_seen = 1;
2565 sblock->sparity = sparity;
2566 scrub_parity_get(sparity);
2568 for (index = 0; len > 0; index++) {
2569 struct scrub_page *spage;
2570 u64 l = min_t(u64, len, PAGE_SIZE);
2572 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2575 spin_lock(&sctx->stat_lock);
2576 sctx->stat.malloc_errors++;
2577 spin_unlock(&sctx->stat_lock);
2578 scrub_block_put(sblock);
2581 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2582 /* For scrub block */
2583 scrub_page_get(spage);
2584 sblock->pagev[index] = spage;
2585 /* For scrub parity */
2586 scrub_page_get(spage);
2587 list_add_tail(&spage->list, &sparity->spages);
2588 spage->sblock = sblock;
2590 spage->flags = flags;
2591 spage->generation = gen;
2592 spage->logical = logical;
2593 spage->physical = physical;
2594 spage->mirror_num = mirror_num;
2596 spage->have_csum = 1;
2597 memcpy(spage->csum, csum, sctx->csum_size);
2599 spage->have_csum = 0;
2601 sblock->page_count++;
2602 spage->page = alloc_page(GFP_KERNEL);
2610 WARN_ON(sblock->page_count == 0);
2611 for (index = 0; index < sblock->page_count; index++) {
2612 struct scrub_page *spage = sblock->pagev[index];
2615 ret = scrub_add_page_to_rd_bio(sctx, spage);
2617 scrub_block_put(sblock);
2622 /* last one frees, either here or in bio completion for last page */
2623 scrub_block_put(sblock);
2627 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2628 u64 logical, u64 len,
2629 u64 physical, struct btrfs_device *dev,
2630 u64 flags, u64 gen, int mirror_num)
2632 struct scrub_ctx *sctx = sparity->sctx;
2634 u8 csum[BTRFS_CSUM_SIZE];
2638 scrub_parity_mark_sectors_error(sparity, logical, len);
2642 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2643 blocksize = sctx->sectorsize;
2644 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2645 blocksize = sctx->nodesize;
2647 blocksize = sctx->sectorsize;
2652 u64 l = min_t(u64, len, blocksize);
2655 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2656 /* push csums to sbio */
2657 have_csum = scrub_find_csum(sctx, logical, csum);
2661 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2662 flags, gen, mirror_num,
2663 have_csum ? csum : NULL);
2675 * Given a physical address, this will calculate it's
2676 * logical offset. if this is a parity stripe, it will return
2677 * the most left data stripe's logical offset.
2679 * return 0 if it is a data stripe, 1 means parity stripe.
2681 static int get_raid56_logic_offset(u64 physical, int num,
2682 struct map_lookup *map, u64 *offset,
2692 last_offset = (physical - map->stripes[num].physical) *
2693 nr_data_stripes(map);
2695 *stripe_start = last_offset;
2697 *offset = last_offset;
2698 for (i = 0; i < nr_data_stripes(map); i++) {
2699 *offset = last_offset + i * map->stripe_len;
2701 stripe_nr = div_u64(*offset, map->stripe_len);
2702 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2704 /* Work out the disk rotation on this stripe-set */
2705 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2706 /* calculate which stripe this data locates */
2708 stripe_index = rot % map->num_stripes;
2709 if (stripe_index == num)
2711 if (stripe_index < num)
2714 *offset = last_offset + j * map->stripe_len;
2718 static void scrub_free_parity(struct scrub_parity *sparity)
2720 struct scrub_ctx *sctx = sparity->sctx;
2721 struct scrub_page *curr, *next;
2724 nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2726 spin_lock(&sctx->stat_lock);
2727 sctx->stat.read_errors += nbits;
2728 sctx->stat.uncorrectable_errors += nbits;
2729 spin_unlock(&sctx->stat_lock);
2732 list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2733 list_del_init(&curr->list);
2734 scrub_page_put(curr);
2740 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2742 struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2744 struct scrub_ctx *sctx = sparity->sctx;
2746 scrub_free_parity(sparity);
2747 scrub_pending_bio_dec(sctx);
2750 static void scrub_parity_bio_endio(struct bio *bio)
2752 struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2755 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2760 btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2761 scrub_parity_bio_endio_worker, NULL, NULL);
2762 btrfs_queue_work(sparity->sctx->dev_root->fs_info->scrub_parity_workers,
2766 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2768 struct scrub_ctx *sctx = sparity->sctx;
2770 struct btrfs_raid_bio *rbio;
2771 struct scrub_page *spage;
2772 struct btrfs_bio *bbio = NULL;
2776 if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2780 length = sparity->logic_end - sparity->logic_start;
2781 ret = btrfs_map_sblock(sctx->dev_root->fs_info, WRITE,
2782 sparity->logic_start,
2783 &length, &bbio, 0, 1);
2784 if (ret || !bbio || !bbio->raid_map)
2787 bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2791 bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2792 bio->bi_private = sparity;
2793 bio->bi_end_io = scrub_parity_bio_endio;
2795 rbio = raid56_parity_alloc_scrub_rbio(sctx->dev_root, bio, bbio,
2796 length, sparity->scrub_dev,
2802 list_for_each_entry(spage, &sparity->spages, list)
2803 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2805 scrub_pending_bio_inc(sctx);
2806 raid56_parity_submit_scrub_rbio(rbio);
2812 btrfs_put_bbio(bbio);
2813 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2815 spin_lock(&sctx->stat_lock);
2816 sctx->stat.malloc_errors++;
2817 spin_unlock(&sctx->stat_lock);
2819 scrub_free_parity(sparity);
2822 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2824 return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2827 static void scrub_parity_get(struct scrub_parity *sparity)
2829 atomic_inc(&sparity->refs);
2832 static void scrub_parity_put(struct scrub_parity *sparity)
2834 if (!atomic_dec_and_test(&sparity->refs))
2837 scrub_parity_check_and_repair(sparity);
2840 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2841 struct map_lookup *map,
2842 struct btrfs_device *sdev,
2843 struct btrfs_path *path,
2847 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2848 struct btrfs_root *root = fs_info->extent_root;
2849 struct btrfs_root *csum_root = fs_info->csum_root;
2850 struct btrfs_extent_item *extent;
2851 struct btrfs_bio *bbio = NULL;
2855 struct extent_buffer *l;
2856 struct btrfs_key key;
2859 u64 extent_physical;
2862 struct btrfs_device *extent_dev;
2863 struct scrub_parity *sparity;
2866 int extent_mirror_num;
2869 nsectors = div_u64(map->stripe_len, root->sectorsize);
2870 bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2871 sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2874 spin_lock(&sctx->stat_lock);
2875 sctx->stat.malloc_errors++;
2876 spin_unlock(&sctx->stat_lock);
2880 sparity->stripe_len = map->stripe_len;
2881 sparity->nsectors = nsectors;
2882 sparity->sctx = sctx;
2883 sparity->scrub_dev = sdev;
2884 sparity->logic_start = logic_start;
2885 sparity->logic_end = logic_end;
2886 atomic_set(&sparity->refs, 1);
2887 INIT_LIST_HEAD(&sparity->spages);
2888 sparity->dbitmap = sparity->bitmap;
2889 sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2892 while (logic_start < logic_end) {
2893 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2894 key.type = BTRFS_METADATA_ITEM_KEY;
2896 key.type = BTRFS_EXTENT_ITEM_KEY;
2897 key.objectid = logic_start;
2898 key.offset = (u64)-1;
2900 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2905 ret = btrfs_previous_extent_item(root, path, 0);
2909 btrfs_release_path(path);
2910 ret = btrfs_search_slot(NULL, root, &key,
2922 slot = path->slots[0];
2923 if (slot >= btrfs_header_nritems(l)) {
2924 ret = btrfs_next_leaf(root, path);
2933 btrfs_item_key_to_cpu(l, &key, slot);
2935 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2936 key.type != BTRFS_METADATA_ITEM_KEY)
2939 if (key.type == BTRFS_METADATA_ITEM_KEY)
2940 bytes = root->nodesize;
2944 if (key.objectid + bytes <= logic_start)
2947 if (key.objectid >= logic_end) {
2952 while (key.objectid >= logic_start + map->stripe_len)
2953 logic_start += map->stripe_len;
2955 extent = btrfs_item_ptr(l, slot,
2956 struct btrfs_extent_item);
2957 flags = btrfs_extent_flags(l, extent);
2958 generation = btrfs_extent_generation(l, extent);
2960 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2961 (key.objectid < logic_start ||
2962 key.objectid + bytes >
2963 logic_start + map->stripe_len)) {
2965 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2966 key.objectid, logic_start);
2967 spin_lock(&sctx->stat_lock);
2968 sctx->stat.uncorrectable_errors++;
2969 spin_unlock(&sctx->stat_lock);
2973 extent_logical = key.objectid;
2976 if (extent_logical < logic_start) {
2977 extent_len -= logic_start - extent_logical;
2978 extent_logical = logic_start;
2981 if (extent_logical + extent_len >
2982 logic_start + map->stripe_len)
2983 extent_len = logic_start + map->stripe_len -
2986 scrub_parity_mark_sectors_data(sparity, extent_logical,
2989 mapped_length = extent_len;
2991 ret = btrfs_map_block(fs_info, READ, extent_logical,
2992 &mapped_length, &bbio, 0);
2994 if (!bbio || mapped_length < extent_len)
2998 btrfs_put_bbio(bbio);
3001 extent_physical = bbio->stripes[0].physical;
3002 extent_mirror_num = bbio->mirror_num;
3003 extent_dev = bbio->stripes[0].dev;
3004 btrfs_put_bbio(bbio);
3006 ret = btrfs_lookup_csums_range(csum_root,
3008 extent_logical + extent_len - 1,
3009 &sctx->csum_list, 1);
3013 ret = scrub_extent_for_parity(sparity, extent_logical,
3020 scrub_free_csums(sctx);
3025 if (extent_logical + extent_len <
3026 key.objectid + bytes) {
3027 logic_start += map->stripe_len;
3029 if (logic_start >= logic_end) {
3034 if (logic_start < key.objectid + bytes) {
3043 btrfs_release_path(path);
3048 logic_start += map->stripe_len;
3052 scrub_parity_mark_sectors_error(sparity, logic_start,
3053 logic_end - logic_start);
3054 scrub_parity_put(sparity);
3056 mutex_lock(&sctx->wr_ctx.wr_lock);
3057 scrub_wr_submit(sctx);
3058 mutex_unlock(&sctx->wr_ctx.wr_lock);
3060 btrfs_release_path(path);
3061 return ret < 0 ? ret : 0;
3064 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3065 struct map_lookup *map,
3066 struct btrfs_device *scrub_dev,
3067 int num, u64 base, u64 length,
3070 struct btrfs_path *path, *ppath;
3071 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3072 struct btrfs_root *root = fs_info->extent_root;
3073 struct btrfs_root *csum_root = fs_info->csum_root;
3074 struct btrfs_extent_item *extent;
3075 struct blk_plug plug;
3080 struct extent_buffer *l;
3087 struct reada_control *reada1;
3088 struct reada_control *reada2;
3089 struct btrfs_key key;
3090 struct btrfs_key key_end;
3091 u64 increment = map->stripe_len;
3094 u64 extent_physical;
3098 struct btrfs_device *extent_dev;
3099 int extent_mirror_num;
3102 physical = map->stripes[num].physical;
3104 nstripes = div_u64(length, map->stripe_len);
3105 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3106 offset = map->stripe_len * num;
3107 increment = map->stripe_len * map->num_stripes;
3109 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3110 int factor = map->num_stripes / map->sub_stripes;
3111 offset = map->stripe_len * (num / map->sub_stripes);
3112 increment = map->stripe_len * factor;
3113 mirror_num = num % map->sub_stripes + 1;
3114 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3115 increment = map->stripe_len;
3116 mirror_num = num % map->num_stripes + 1;
3117 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3118 increment = map->stripe_len;
3119 mirror_num = num % map->num_stripes + 1;
3120 } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3121 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3122 increment = map->stripe_len * nr_data_stripes(map);
3125 increment = map->stripe_len;
3129 path = btrfs_alloc_path();
3133 ppath = btrfs_alloc_path();
3135 btrfs_free_path(path);
3140 * work on commit root. The related disk blocks are static as
3141 * long as COW is applied. This means, it is save to rewrite
3142 * them to repair disk errors without any race conditions
3144 path->search_commit_root = 1;
3145 path->skip_locking = 1;
3147 ppath->search_commit_root = 1;
3148 ppath->skip_locking = 1;
3150 * trigger the readahead for extent tree csum tree and wait for
3151 * completion. During readahead, the scrub is officially paused
3152 * to not hold off transaction commits
3154 logical = base + offset;
3155 physical_end = physical + nstripes * map->stripe_len;
3156 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3157 get_raid56_logic_offset(physical_end, num,
3158 map, &logic_end, NULL);
3161 logic_end = logical + increment * nstripes;
3163 wait_event(sctx->list_wait,
3164 atomic_read(&sctx->bios_in_flight) == 0);
3165 scrub_blocked_if_needed(fs_info);
3167 /* FIXME it might be better to start readahead at commit root */
3168 key.objectid = logical;
3169 key.type = BTRFS_EXTENT_ITEM_KEY;
3170 key.offset = (u64)0;
3171 key_end.objectid = logic_end;
3172 key_end.type = BTRFS_METADATA_ITEM_KEY;
3173 key_end.offset = (u64)-1;
3174 reada1 = btrfs_reada_add(root, &key, &key_end);
3176 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3177 key.type = BTRFS_EXTENT_CSUM_KEY;
3178 key.offset = logical;
3179 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3180 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3181 key_end.offset = logic_end;
3182 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3184 if (!IS_ERR(reada1))
3185 btrfs_reada_wait(reada1);
3186 if (!IS_ERR(reada2))
3187 btrfs_reada_wait(reada2);
3191 * collect all data csums for the stripe to avoid seeking during
3192 * the scrub. This might currently (crc32) end up to be about 1MB
3194 blk_start_plug(&plug);
3197 * now find all extents for each stripe and scrub them
3200 while (physical < physical_end) {
3204 if (atomic_read(&fs_info->scrub_cancel_req) ||
3205 atomic_read(&sctx->cancel_req)) {
3210 * check to see if we have to pause
3212 if (atomic_read(&fs_info->scrub_pause_req)) {
3213 /* push queued extents */
3214 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3216 mutex_lock(&sctx->wr_ctx.wr_lock);
3217 scrub_wr_submit(sctx);
3218 mutex_unlock(&sctx->wr_ctx.wr_lock);
3219 wait_event(sctx->list_wait,
3220 atomic_read(&sctx->bios_in_flight) == 0);
3221 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3222 scrub_blocked_if_needed(fs_info);
3225 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3226 ret = get_raid56_logic_offset(physical, num, map,
3231 /* it is parity strip */
3232 stripe_logical += base;
3233 stripe_end = stripe_logical + increment;
3234 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3235 ppath, stripe_logical,
3243 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3244 key.type = BTRFS_METADATA_ITEM_KEY;
3246 key.type = BTRFS_EXTENT_ITEM_KEY;
3247 key.objectid = logical;
3248 key.offset = (u64)-1;
3250 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3255 ret = btrfs_previous_extent_item(root, path, 0);
3259 /* there's no smaller item, so stick with the
3261 btrfs_release_path(path);
3262 ret = btrfs_search_slot(NULL, root, &key,
3274 slot = path->slots[0];
3275 if (slot >= btrfs_header_nritems(l)) {
3276 ret = btrfs_next_leaf(root, path);
3285 btrfs_item_key_to_cpu(l, &key, slot);
3287 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3288 key.type != BTRFS_METADATA_ITEM_KEY)
3291 if (key.type == BTRFS_METADATA_ITEM_KEY)
3292 bytes = root->nodesize;
3296 if (key.objectid + bytes <= logical)
3299 if (key.objectid >= logical + map->stripe_len) {
3300 /* out of this device extent */
3301 if (key.objectid >= logic_end)
3306 extent = btrfs_item_ptr(l, slot,
3307 struct btrfs_extent_item);
3308 flags = btrfs_extent_flags(l, extent);
3309 generation = btrfs_extent_generation(l, extent);
3311 if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3312 (key.objectid < logical ||
3313 key.objectid + bytes >
3314 logical + map->stripe_len)) {
3316 "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3317 key.objectid, logical);
3318 spin_lock(&sctx->stat_lock);
3319 sctx->stat.uncorrectable_errors++;
3320 spin_unlock(&sctx->stat_lock);
3325 extent_logical = key.objectid;
3329 * trim extent to this stripe
3331 if (extent_logical < logical) {
3332 extent_len -= logical - extent_logical;
3333 extent_logical = logical;
3335 if (extent_logical + extent_len >
3336 logical + map->stripe_len) {
3337 extent_len = logical + map->stripe_len -
3341 extent_physical = extent_logical - logical + physical;
3342 extent_dev = scrub_dev;
3343 extent_mirror_num = mirror_num;
3345 scrub_remap_extent(fs_info, extent_logical,
3346 extent_len, &extent_physical,
3348 &extent_mirror_num);
3350 ret = btrfs_lookup_csums_range(csum_root,
3354 &sctx->csum_list, 1);
3358 ret = scrub_extent(sctx, extent_logical, extent_len,
3359 extent_physical, extent_dev, flags,
3360 generation, extent_mirror_num,
3361 extent_logical - logical + physical);
3363 scrub_free_csums(sctx);
3368 if (extent_logical + extent_len <
3369 key.objectid + bytes) {
3370 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3372 * loop until we find next data stripe
3373 * or we have finished all stripes.
3376 physical += map->stripe_len;
3377 ret = get_raid56_logic_offset(physical,
3382 if (ret && physical < physical_end) {
3383 stripe_logical += base;
3384 stripe_end = stripe_logical +
3386 ret = scrub_raid56_parity(sctx,
3387 map, scrub_dev, ppath,
3395 physical += map->stripe_len;
3396 logical += increment;
3398 if (logical < key.objectid + bytes) {
3403 if (physical >= physical_end) {
3411 btrfs_release_path(path);
3413 logical += increment;
3414 physical += map->stripe_len;
3415 spin_lock(&sctx->stat_lock);
3417 sctx->stat.last_physical = map->stripes[num].physical +
3420 sctx->stat.last_physical = physical;
3421 spin_unlock(&sctx->stat_lock);
3426 /* push queued extents */
3428 mutex_lock(&sctx->wr_ctx.wr_lock);
3429 scrub_wr_submit(sctx);
3430 mutex_unlock(&sctx->wr_ctx.wr_lock);
3432 blk_finish_plug(&plug);
3433 btrfs_free_path(path);
3434 btrfs_free_path(ppath);
3435 return ret < 0 ? ret : 0;
3438 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3439 struct btrfs_device *scrub_dev,
3440 u64 chunk_offset, u64 length,
3442 struct btrfs_block_group_cache *cache,
3445 struct btrfs_mapping_tree *map_tree =
3446 &sctx->dev_root->fs_info->mapping_tree;
3447 struct map_lookup *map;
3448 struct extent_map *em;
3452 read_lock(&map_tree->map_tree.lock);
3453 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3454 read_unlock(&map_tree->map_tree.lock);
3458 * Might have been an unused block group deleted by the cleaner
3459 * kthread or relocation.
3461 spin_lock(&cache->lock);
3462 if (!cache->removed)
3464 spin_unlock(&cache->lock);
3469 map = em->map_lookup;
3470 if (em->start != chunk_offset)
3473 if (em->len < length)
3476 for (i = 0; i < map->num_stripes; ++i) {
3477 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3478 map->stripes[i].physical == dev_offset) {
3479 ret = scrub_stripe(sctx, map, scrub_dev, i,
3480 chunk_offset, length,
3487 free_extent_map(em);
3492 static noinline_for_stack
3493 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3494 struct btrfs_device *scrub_dev, u64 start, u64 end,
3497 struct btrfs_dev_extent *dev_extent = NULL;
3498 struct btrfs_path *path;
3499 struct btrfs_root *root = sctx->dev_root;
3500 struct btrfs_fs_info *fs_info = root->fs_info;
3506 struct extent_buffer *l;
3507 struct btrfs_key key;
3508 struct btrfs_key found_key;
3509 struct btrfs_block_group_cache *cache;
3510 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3512 path = btrfs_alloc_path();
3516 path->reada = READA_FORWARD;
3517 path->search_commit_root = 1;
3518 path->skip_locking = 1;
3520 key.objectid = scrub_dev->devid;
3522 key.type = BTRFS_DEV_EXTENT_KEY;
3525 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3529 if (path->slots[0] >=
3530 btrfs_header_nritems(path->nodes[0])) {
3531 ret = btrfs_next_leaf(root, path);
3544 slot = path->slots[0];
3546 btrfs_item_key_to_cpu(l, &found_key, slot);
3548 if (found_key.objectid != scrub_dev->devid)
3551 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3554 if (found_key.offset >= end)
3557 if (found_key.offset < key.offset)
3560 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3561 length = btrfs_dev_extent_length(l, dev_extent);
3563 if (found_key.offset + length <= start)
3566 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3569 * get a reference on the corresponding block group to prevent
3570 * the chunk from going away while we scrub it
3572 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3574 /* some chunks are removed but not committed to disk yet,
3575 * continue scrubbing */
3580 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3581 * to avoid deadlock caused by:
3582 * btrfs_inc_block_group_ro()
3583 * -> btrfs_wait_for_commit()
3584 * -> btrfs_commit_transaction()
3585 * -> btrfs_scrub_pause()
3587 scrub_pause_on(fs_info);
3588 ret = btrfs_inc_block_group_ro(root, cache);
3589 if (!ret && is_dev_replace) {
3591 * If we are doing a device replace wait for any tasks
3592 * that started dellaloc right before we set the block
3593 * group to RO mode, as they might have just allocated
3594 * an extent from it or decided they could do a nocow
3595 * write. And if any such tasks did that, wait for their
3596 * ordered extents to complete and then commit the
3597 * current transaction, so that we can later see the new
3598 * extent items in the extent tree - the ordered extents
3599 * create delayed data references (for cow writes) when
3600 * they complete, which will be run and insert the
3601 * corresponding extent items into the extent tree when
3602 * we commit the transaction they used when running
3603 * inode.c:btrfs_finish_ordered_io(). We later use
3604 * the commit root of the extent tree to find extents
3605 * to copy from the srcdev into the tgtdev, and we don't
3606 * want to miss any new extents.
3608 btrfs_wait_block_group_reservations(cache);
3609 btrfs_wait_nocow_writers(cache);
3610 ret = btrfs_wait_ordered_roots(fs_info, -1,
3611 cache->key.objectid,
3614 struct btrfs_trans_handle *trans;
3616 trans = btrfs_join_transaction(root);
3618 ret = PTR_ERR(trans);
3620 ret = btrfs_commit_transaction(trans,
3623 scrub_pause_off(fs_info);
3624 btrfs_put_block_group(cache);
3629 scrub_pause_off(fs_info);
3633 } else if (ret == -ENOSPC) {
3635 * btrfs_inc_block_group_ro return -ENOSPC when it
3636 * failed in creating new chunk for metadata.
3637 * It is not a problem for scrub/replace, because
3638 * metadata are always cowed, and our scrub paused
3639 * commit_transactions.
3644 "failed setting block group ro, ret=%d\n",
3646 btrfs_put_block_group(cache);
3650 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3651 dev_replace->cursor_right = found_key.offset + length;
3652 dev_replace->cursor_left = found_key.offset;
3653 dev_replace->item_needs_writeback = 1;
3654 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3655 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3656 found_key.offset, cache, is_dev_replace);
3659 * flush, submit all pending read and write bios, afterwards
3661 * Note that in the dev replace case, a read request causes
3662 * write requests that are submitted in the read completion
3663 * worker. Therefore in the current situation, it is required
3664 * that all write requests are flushed, so that all read and
3665 * write requests are really completed when bios_in_flight
3668 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3670 mutex_lock(&sctx->wr_ctx.wr_lock);
3671 scrub_wr_submit(sctx);
3672 mutex_unlock(&sctx->wr_ctx.wr_lock);
3674 wait_event(sctx->list_wait,
3675 atomic_read(&sctx->bios_in_flight) == 0);
3677 scrub_pause_on(fs_info);
3680 * must be called before we decrease @scrub_paused.
3681 * make sure we don't block transaction commit while
3682 * we are waiting pending workers finished.
3684 wait_event(sctx->list_wait,
3685 atomic_read(&sctx->workers_pending) == 0);
3686 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3688 scrub_pause_off(fs_info);
3690 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3691 dev_replace->cursor_left = dev_replace->cursor_right;
3692 dev_replace->item_needs_writeback = 1;
3693 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3696 btrfs_dec_block_group_ro(root, cache);
3699 * We might have prevented the cleaner kthread from deleting
3700 * this block group if it was already unused because we raced
3701 * and set it to RO mode first. So add it back to the unused
3702 * list, otherwise it might not ever be deleted unless a manual
3703 * balance is triggered or it becomes used and unused again.
3705 spin_lock(&cache->lock);
3706 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3707 btrfs_block_group_used(&cache->item) == 0) {
3708 spin_unlock(&cache->lock);
3709 spin_lock(&fs_info->unused_bgs_lock);
3710 if (list_empty(&cache->bg_list)) {
3711 btrfs_get_block_group(cache);
3712 list_add_tail(&cache->bg_list,
3713 &fs_info->unused_bgs);
3715 spin_unlock(&fs_info->unused_bgs_lock);
3717 spin_unlock(&cache->lock);
3720 btrfs_put_block_group(cache);
3723 if (is_dev_replace &&
3724 atomic64_read(&dev_replace->num_write_errors) > 0) {
3728 if (sctx->stat.malloc_errors > 0) {
3733 key.offset = found_key.offset + length;
3734 btrfs_release_path(path);
3737 btrfs_free_path(path);
3742 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3743 struct btrfs_device *scrub_dev)
3749 struct btrfs_root *root = sctx->dev_root;
3751 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
3754 /* Seed devices of a new filesystem has their own generation. */
3755 if (scrub_dev->fs_devices != root->fs_info->fs_devices)
3756 gen = scrub_dev->generation;
3758 gen = root->fs_info->last_trans_committed;
3760 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3761 bytenr = btrfs_sb_offset(i);
3762 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3763 scrub_dev->commit_total_bytes)
3766 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3767 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3772 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3778 * get a reference count on fs_info->scrub_workers. start worker if necessary
3780 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3783 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3784 int max_active = fs_info->thread_pool_size;
3786 if (fs_info->scrub_workers_refcnt == 0) {
3788 fs_info->scrub_workers =
3789 btrfs_alloc_workqueue(fs_info, "scrub", flags,
3792 fs_info->scrub_workers =
3793 btrfs_alloc_workqueue(fs_info, "scrub", flags,
3795 if (!fs_info->scrub_workers)
3796 goto fail_scrub_workers;
3798 fs_info->scrub_wr_completion_workers =
3799 btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3801 if (!fs_info->scrub_wr_completion_workers)
3802 goto fail_scrub_wr_completion_workers;
3804 fs_info->scrub_nocow_workers =
3805 btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
3806 if (!fs_info->scrub_nocow_workers)
3807 goto fail_scrub_nocow_workers;
3808 fs_info->scrub_parity_workers =
3809 btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3811 if (!fs_info->scrub_parity_workers)
3812 goto fail_scrub_parity_workers;
3814 ++fs_info->scrub_workers_refcnt;
3817 fail_scrub_parity_workers:
3818 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3819 fail_scrub_nocow_workers:
3820 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3821 fail_scrub_wr_completion_workers:
3822 btrfs_destroy_workqueue(fs_info->scrub_workers);
3827 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3829 if (--fs_info->scrub_workers_refcnt == 0) {
3830 btrfs_destroy_workqueue(fs_info->scrub_workers);
3831 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3832 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3833 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3835 WARN_ON(fs_info->scrub_workers_refcnt < 0);
3838 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3839 u64 end, struct btrfs_scrub_progress *progress,
3840 int readonly, int is_dev_replace)
3842 struct scrub_ctx *sctx;
3844 struct btrfs_device *dev;
3845 struct rcu_string *name;
3847 if (btrfs_fs_closing(fs_info))
3850 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
3852 * in this case scrub is unable to calculate the checksum
3853 * the way scrub is implemented. Do not handle this
3854 * situation at all because it won't ever happen.
3857 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3858 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
3862 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
3863 /* not supported for data w/o checksums */
3864 btrfs_err_rl(fs_info,
3865 "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3866 fs_info->chunk_root->sectorsize, PAGE_SIZE);
3870 if (fs_info->chunk_root->nodesize >
3871 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3872 fs_info->chunk_root->sectorsize >
3873 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3875 * would exhaust the array bounds of pagev member in
3876 * struct scrub_block
3879 "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3880 fs_info->chunk_root->nodesize,
3881 SCRUB_MAX_PAGES_PER_BLOCK,
3882 fs_info->chunk_root->sectorsize,
3883 SCRUB_MAX_PAGES_PER_BLOCK);
3888 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3889 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3890 if (!dev || (dev->missing && !is_dev_replace)) {
3891 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3895 if (!is_dev_replace && !readonly && !dev->writeable) {
3896 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3898 name = rcu_dereference(dev->name);
3899 btrfs_err(fs_info, "scrub: device %s is not writable",
3905 mutex_lock(&fs_info->scrub_lock);
3906 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
3907 mutex_unlock(&fs_info->scrub_lock);
3908 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3912 btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
3913 if (dev->scrub_device ||
3915 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3916 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3917 mutex_unlock(&fs_info->scrub_lock);
3918 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3919 return -EINPROGRESS;
3921 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
3923 ret = scrub_workers_get(fs_info, is_dev_replace);
3925 mutex_unlock(&fs_info->scrub_lock);
3926 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3930 sctx = scrub_setup_ctx(dev, is_dev_replace);
3932 mutex_unlock(&fs_info->scrub_lock);
3933 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3934 scrub_workers_put(fs_info);
3935 return PTR_ERR(sctx);
3937 sctx->readonly = readonly;
3938 dev->scrub_device = sctx;
3939 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3942 * checking @scrub_pause_req here, we can avoid
3943 * race between committing transaction and scrubbing.
3945 __scrub_blocked_if_needed(fs_info);
3946 atomic_inc(&fs_info->scrubs_running);
3947 mutex_unlock(&fs_info->scrub_lock);
3949 if (!is_dev_replace) {
3951 * by holding device list mutex, we can
3952 * kick off writing super in log tree sync.
3954 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3955 ret = scrub_supers(sctx, dev);
3956 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3960 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3963 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3964 atomic_dec(&fs_info->scrubs_running);
3965 wake_up(&fs_info->scrub_pause_wait);
3967 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3970 memcpy(progress, &sctx->stat, sizeof(*progress));
3972 mutex_lock(&fs_info->scrub_lock);
3973 dev->scrub_device = NULL;
3974 scrub_workers_put(fs_info);
3975 mutex_unlock(&fs_info->scrub_lock);
3977 scrub_put_ctx(sctx);
3982 void btrfs_scrub_pause(struct btrfs_root *root)
3984 struct btrfs_fs_info *fs_info = root->fs_info;
3986 mutex_lock(&fs_info->scrub_lock);
3987 atomic_inc(&fs_info->scrub_pause_req);
3988 while (atomic_read(&fs_info->scrubs_paused) !=
3989 atomic_read(&fs_info->scrubs_running)) {
3990 mutex_unlock(&fs_info->scrub_lock);
3991 wait_event(fs_info->scrub_pause_wait,
3992 atomic_read(&fs_info->scrubs_paused) ==
3993 atomic_read(&fs_info->scrubs_running));
3994 mutex_lock(&fs_info->scrub_lock);
3996 mutex_unlock(&fs_info->scrub_lock);
3999 void btrfs_scrub_continue(struct btrfs_root *root)
4001 struct btrfs_fs_info *fs_info = root->fs_info;
4003 atomic_dec(&fs_info->scrub_pause_req);
4004 wake_up(&fs_info->scrub_pause_wait);
4007 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4009 mutex_lock(&fs_info->scrub_lock);
4010 if (!atomic_read(&fs_info->scrubs_running)) {
4011 mutex_unlock(&fs_info->scrub_lock);
4015 atomic_inc(&fs_info->scrub_cancel_req);
4016 while (atomic_read(&fs_info->scrubs_running)) {
4017 mutex_unlock(&fs_info->scrub_lock);
4018 wait_event(fs_info->scrub_pause_wait,
4019 atomic_read(&fs_info->scrubs_running) == 0);
4020 mutex_lock(&fs_info->scrub_lock);
4022 atomic_dec(&fs_info->scrub_cancel_req);
4023 mutex_unlock(&fs_info->scrub_lock);
4028 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4029 struct btrfs_device *dev)
4031 struct scrub_ctx *sctx;
4033 mutex_lock(&fs_info->scrub_lock);
4034 sctx = dev->scrub_device;
4036 mutex_unlock(&fs_info->scrub_lock);
4039 atomic_inc(&sctx->cancel_req);
4040 while (dev->scrub_device) {
4041 mutex_unlock(&fs_info->scrub_lock);
4042 wait_event(fs_info->scrub_pause_wait,
4043 dev->scrub_device == NULL);
4044 mutex_lock(&fs_info->scrub_lock);
4046 mutex_unlock(&fs_info->scrub_lock);
4051 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
4052 struct btrfs_scrub_progress *progress)
4054 struct btrfs_device *dev;
4055 struct scrub_ctx *sctx = NULL;
4057 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
4058 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
4060 sctx = dev->scrub_device;
4062 memcpy(progress, &sctx->stat, sizeof(*progress));
4063 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
4065 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4068 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4069 u64 extent_logical, u64 extent_len,
4070 u64 *extent_physical,
4071 struct btrfs_device **extent_dev,
4072 int *extent_mirror_num)
4075 struct btrfs_bio *bbio = NULL;
4078 mapped_length = extent_len;
4079 ret = btrfs_map_block(fs_info, READ, extent_logical,
4080 &mapped_length, &bbio, 0);
4081 if (ret || !bbio || mapped_length < extent_len ||
4082 !bbio->stripes[0].dev->bdev) {
4083 btrfs_put_bbio(bbio);
4087 *extent_physical = bbio->stripes[0].physical;
4088 *extent_mirror_num = bbio->mirror_num;
4089 *extent_dev = bbio->stripes[0].dev;
4090 btrfs_put_bbio(bbio);
4093 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
4094 struct scrub_wr_ctx *wr_ctx,
4095 struct btrfs_fs_info *fs_info,
4096 struct btrfs_device *dev,
4099 WARN_ON(wr_ctx->wr_curr_bio != NULL);
4101 mutex_init(&wr_ctx->wr_lock);
4102 wr_ctx->wr_curr_bio = NULL;
4103 if (!is_dev_replace)
4106 WARN_ON(!dev->bdev);
4107 wr_ctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
4108 wr_ctx->tgtdev = dev;
4109 atomic_set(&wr_ctx->flush_all_writes, 0);
4113 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
4115 mutex_lock(&wr_ctx->wr_lock);
4116 kfree(wr_ctx->wr_curr_bio);
4117 wr_ctx->wr_curr_bio = NULL;
4118 mutex_unlock(&wr_ctx->wr_lock);
4121 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4122 int mirror_num, u64 physical_for_dev_replace)
4124 struct scrub_copy_nocow_ctx *nocow_ctx;
4125 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
4127 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4129 spin_lock(&sctx->stat_lock);
4130 sctx->stat.malloc_errors++;
4131 spin_unlock(&sctx->stat_lock);
4135 scrub_pending_trans_workers_inc(sctx);
4137 nocow_ctx->sctx = sctx;
4138 nocow_ctx->logical = logical;
4139 nocow_ctx->len = len;
4140 nocow_ctx->mirror_num = mirror_num;
4141 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4142 btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4143 copy_nocow_pages_worker, NULL, NULL);
4144 INIT_LIST_HEAD(&nocow_ctx->inodes);
4145 btrfs_queue_work(fs_info->scrub_nocow_workers,
4151 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4153 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4154 struct scrub_nocow_inode *nocow_inode;
4156 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4159 nocow_inode->inum = inum;
4160 nocow_inode->offset = offset;
4161 nocow_inode->root = root;
4162 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4166 #define COPY_COMPLETE 1
4168 static void copy_nocow_pages_worker(struct btrfs_work *work)
4170 struct scrub_copy_nocow_ctx *nocow_ctx =
4171 container_of(work, struct scrub_copy_nocow_ctx, work);
4172 struct scrub_ctx *sctx = nocow_ctx->sctx;
4173 u64 logical = nocow_ctx->logical;
4174 u64 len = nocow_ctx->len;
4175 int mirror_num = nocow_ctx->mirror_num;
4176 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4178 struct btrfs_trans_handle *trans = NULL;
4179 struct btrfs_fs_info *fs_info;
4180 struct btrfs_path *path;
4181 struct btrfs_root *root;
4182 int not_written = 0;
4184 fs_info = sctx->dev_root->fs_info;
4185 root = fs_info->extent_root;
4187 path = btrfs_alloc_path();
4189 spin_lock(&sctx->stat_lock);
4190 sctx->stat.malloc_errors++;
4191 spin_unlock(&sctx->stat_lock);
4196 trans = btrfs_join_transaction(root);
4197 if (IS_ERR(trans)) {
4202 ret = iterate_inodes_from_logical(logical, fs_info, path,
4203 record_inode_for_nocow, nocow_ctx);
4204 if (ret != 0 && ret != -ENOENT) {
4206 "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4207 logical, physical_for_dev_replace, len, mirror_num,
4213 btrfs_end_transaction(trans, root);
4215 while (!list_empty(&nocow_ctx->inodes)) {
4216 struct scrub_nocow_inode *entry;
4217 entry = list_first_entry(&nocow_ctx->inodes,
4218 struct scrub_nocow_inode,
4220 list_del_init(&entry->list);
4221 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4222 entry->root, nocow_ctx);
4224 if (ret == COPY_COMPLETE) {
4232 while (!list_empty(&nocow_ctx->inodes)) {
4233 struct scrub_nocow_inode *entry;
4234 entry = list_first_entry(&nocow_ctx->inodes,
4235 struct scrub_nocow_inode,
4237 list_del_init(&entry->list);
4240 if (trans && !IS_ERR(trans))
4241 btrfs_end_transaction(trans, root);
4243 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4244 num_uncorrectable_read_errors);
4246 btrfs_free_path(path);
4249 scrub_pending_trans_workers_dec(sctx);
4252 static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
4255 struct extent_state *cached_state = NULL;
4256 struct btrfs_ordered_extent *ordered;
4257 struct extent_io_tree *io_tree;
4258 struct extent_map *em;
4259 u64 lockstart = start, lockend = start + len - 1;
4262 io_tree = &BTRFS_I(inode)->io_tree;
4264 lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4265 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4267 btrfs_put_ordered_extent(ordered);
4272 em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4279 * This extent does not actually cover the logical extent anymore,
4280 * move on to the next inode.
4282 if (em->block_start > logical ||
4283 em->block_start + em->block_len < logical + len) {
4284 free_extent_map(em);
4288 free_extent_map(em);
4291 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4296 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4297 struct scrub_copy_nocow_ctx *nocow_ctx)
4299 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
4300 struct btrfs_key key;
4301 struct inode *inode;
4303 struct btrfs_root *local_root;
4304 struct extent_io_tree *io_tree;
4305 u64 physical_for_dev_replace;
4306 u64 nocow_ctx_logical;
4307 u64 len = nocow_ctx->len;
4308 unsigned long index;
4313 key.objectid = root;
4314 key.type = BTRFS_ROOT_ITEM_KEY;
4315 key.offset = (u64)-1;
4317 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4319 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4320 if (IS_ERR(local_root)) {
4321 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4322 return PTR_ERR(local_root);
4325 key.type = BTRFS_INODE_ITEM_KEY;
4326 key.objectid = inum;
4328 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4329 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4331 return PTR_ERR(inode);
4333 /* Avoid truncate/dio/punch hole.. */
4335 inode_dio_wait(inode);
4337 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4338 io_tree = &BTRFS_I(inode)->io_tree;
4339 nocow_ctx_logical = nocow_ctx->logical;
4341 ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
4343 ret = ret > 0 ? 0 : ret;
4347 while (len >= PAGE_SIZE) {
4348 index = offset >> PAGE_SHIFT;
4350 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4352 btrfs_err(fs_info, "find_or_create_page() failed");
4357 if (PageUptodate(page)) {
4358 if (PageDirty(page))
4361 ClearPageError(page);
4362 err = extent_read_full_page(io_tree, page,
4364 nocow_ctx->mirror_num);
4372 * If the page has been remove from the page cache,
4373 * the data on it is meaningless, because it may be
4374 * old one, the new data may be written into the new
4375 * page in the page cache.
4377 if (page->mapping != inode->i_mapping) {
4382 if (!PageUptodate(page)) {
4388 ret = check_extent_to_block(inode, offset, len,
4391 ret = ret > 0 ? 0 : ret;
4395 err = write_page_nocow(nocow_ctx->sctx,
4396 physical_for_dev_replace, page);
4406 offset += PAGE_SIZE;
4407 physical_for_dev_replace += PAGE_SIZE;
4408 nocow_ctx_logical += PAGE_SIZE;
4411 ret = COPY_COMPLETE;
4413 inode_unlock(inode);
4418 static int write_page_nocow(struct scrub_ctx *sctx,
4419 u64 physical_for_dev_replace, struct page *page)
4422 struct btrfs_device *dev;
4425 dev = sctx->wr_ctx.tgtdev;
4429 btrfs_warn_rl(dev->dev_root->fs_info,
4430 "scrub write_page_nocow(bdev == NULL) is unexpected");
4433 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4435 spin_lock(&sctx->stat_lock);
4436 sctx->stat.malloc_errors++;
4437 spin_unlock(&sctx->stat_lock);
4440 bio->bi_iter.bi_size = 0;
4441 bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4442 bio->bi_bdev = dev->bdev;
4443 bio_set_op_attrs(bio, REQ_OP_WRITE, WRITE_SYNC);
4444 ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4445 if (ret != PAGE_SIZE) {
4448 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4452 if (btrfsic_submit_bio_wait(bio))
4453 goto leave_with_eio;