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"
33 * This is only the first step towards a full-features scrub. It reads all
34 * extent and super block and verifies the checksums. In case a bad checksum
35 * is found or the extent cannot be read, good data will be written back if
38 * Future enhancements:
39 * - In case an unrepairable extent is encountered, track which files are
40 * affected and report them
41 * - track and record media errors, throw out bad devices
42 * - add a mode to also read unallocated space
49 * the following three values only influence the performance.
50 * The last one configures the number of parallel and outstanding I/O
51 * operations. The first two values configure an upper limit for the number
52 * of (dynamically allocated) pages that are added to a bio.
54 #define SCRUB_PAGES_PER_RD_BIO 32 /* 128k per bio */
55 #define SCRUB_PAGES_PER_WR_BIO 32 /* 128k per bio */
56 #define SCRUB_BIOS_PER_SCTX 64 /* 8MB per device in flight */
59 * the following value times PAGE_SIZE needs to be large enough to match the
60 * largest node/leaf/sector size that shall be supported.
61 * Values larger than BTRFS_STRIPE_LEN are not supported.
63 #define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */
66 struct scrub_block *sblock;
68 struct btrfs_device *dev;
69 u64 flags; /* extent flags */
73 u64 physical_for_dev_replace;
76 unsigned int mirror_num:8;
77 unsigned int have_csum:1;
78 unsigned int io_error:1;
80 u8 csum[BTRFS_CSUM_SIZE];
85 struct scrub_ctx *sctx;
86 struct btrfs_device *dev;
91 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
92 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
94 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
98 struct btrfs_work work;
102 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
104 atomic_t outstanding_pages;
105 atomic_t ref_count; /* free mem on transition to zero */
106 struct scrub_ctx *sctx;
108 unsigned int header_error:1;
109 unsigned int checksum_error:1;
110 unsigned int no_io_error_seen:1;
111 unsigned int generation_error:1; /* also sets header_error */
115 struct scrub_wr_ctx {
116 struct scrub_bio *wr_curr_bio;
117 struct btrfs_device *tgtdev;
118 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
119 atomic_t flush_all_writes;
120 struct mutex wr_lock;
124 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
125 struct btrfs_root *dev_root;
128 atomic_t bios_in_flight;
129 atomic_t workers_pending;
130 spinlock_t list_lock;
131 wait_queue_head_t list_wait;
133 struct list_head csum_list;
136 int pages_per_rd_bio;
142 struct scrub_wr_ctx wr_ctx;
147 struct btrfs_scrub_progress stat;
148 spinlock_t stat_lock;
151 struct scrub_fixup_nodatasum {
152 struct scrub_ctx *sctx;
153 struct btrfs_device *dev;
155 struct btrfs_root *root;
156 struct btrfs_work work;
160 struct scrub_copy_nocow_ctx {
161 struct scrub_ctx *sctx;
165 u64 physical_for_dev_replace;
166 struct btrfs_work work;
169 struct scrub_warning {
170 struct btrfs_path *path;
171 u64 extent_item_size;
177 struct btrfs_device *dev;
183 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
184 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
185 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
186 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
187 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
188 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
189 struct btrfs_fs_info *fs_info,
190 struct scrub_block *original_sblock,
191 u64 length, u64 logical,
192 struct scrub_block *sblocks_for_recheck);
193 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
194 struct scrub_block *sblock, int is_metadata,
195 int have_csum, u8 *csum, u64 generation,
197 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
198 struct scrub_block *sblock,
199 int is_metadata, int have_csum,
200 const u8 *csum, u64 generation,
202 static void scrub_complete_bio_end_io(struct bio *bio, int err);
203 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
204 struct scrub_block *sblock_good,
206 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
207 struct scrub_block *sblock_good,
208 int page_num, int force_write);
209 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
210 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
212 static int scrub_checksum_data(struct scrub_block *sblock);
213 static int scrub_checksum_tree_block(struct scrub_block *sblock);
214 static int scrub_checksum_super(struct scrub_block *sblock);
215 static void scrub_block_get(struct scrub_block *sblock);
216 static void scrub_block_put(struct scrub_block *sblock);
217 static void scrub_page_get(struct scrub_page *spage);
218 static void scrub_page_put(struct scrub_page *spage);
219 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
220 struct scrub_page *spage);
221 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
222 u64 physical, struct btrfs_device *dev, u64 flags,
223 u64 gen, int mirror_num, u8 *csum, int force,
224 u64 physical_for_dev_replace);
225 static void scrub_bio_end_io(struct bio *bio, int err);
226 static void scrub_bio_end_io_worker(struct btrfs_work *work);
227 static void scrub_block_complete(struct scrub_block *sblock);
228 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
229 u64 extent_logical, u64 extent_len,
230 u64 *extent_physical,
231 struct btrfs_device **extent_dev,
232 int *extent_mirror_num);
233 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
234 struct scrub_wr_ctx *wr_ctx,
235 struct btrfs_fs_info *fs_info,
236 struct btrfs_device *dev,
238 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
239 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
240 struct scrub_page *spage);
241 static void scrub_wr_submit(struct scrub_ctx *sctx);
242 static void scrub_wr_bio_end_io(struct bio *bio, int err);
243 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
244 static int write_page_nocow(struct scrub_ctx *sctx,
245 u64 physical_for_dev_replace, struct page *page);
246 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
248 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
249 int mirror_num, u64 physical_for_dev_replace);
250 static void copy_nocow_pages_worker(struct btrfs_work *work);
253 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
255 atomic_inc(&sctx->bios_in_flight);
258 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
260 atomic_dec(&sctx->bios_in_flight);
261 wake_up(&sctx->list_wait);
265 * used for workers that require transaction commits (i.e., for the
268 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
270 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
273 * increment scrubs_running to prevent cancel requests from
274 * completing as long as a worker is running. we must also
275 * increment scrubs_paused to prevent deadlocking on pause
276 * requests used for transactions commits (as the worker uses a
277 * transaction context). it is safe to regard the worker
278 * as paused for all matters practical. effectively, we only
279 * avoid cancellation requests from completing.
281 mutex_lock(&fs_info->scrub_lock);
282 atomic_inc(&fs_info->scrubs_running);
283 atomic_inc(&fs_info->scrubs_paused);
284 mutex_unlock(&fs_info->scrub_lock);
285 atomic_inc(&sctx->workers_pending);
288 /* used for workers that require transaction commits */
289 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
291 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
294 * see scrub_pending_trans_workers_inc() why we're pretending
295 * to be paused in the scrub counters
297 mutex_lock(&fs_info->scrub_lock);
298 atomic_dec(&fs_info->scrubs_running);
299 atomic_dec(&fs_info->scrubs_paused);
300 mutex_unlock(&fs_info->scrub_lock);
301 atomic_dec(&sctx->workers_pending);
302 wake_up(&fs_info->scrub_pause_wait);
303 wake_up(&sctx->list_wait);
306 static void scrub_free_csums(struct scrub_ctx *sctx)
308 while (!list_empty(&sctx->csum_list)) {
309 struct btrfs_ordered_sum *sum;
310 sum = list_first_entry(&sctx->csum_list,
311 struct btrfs_ordered_sum, list);
312 list_del(&sum->list);
317 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
324 scrub_free_wr_ctx(&sctx->wr_ctx);
326 /* this can happen when scrub is cancelled */
327 if (sctx->curr != -1) {
328 struct scrub_bio *sbio = sctx->bios[sctx->curr];
330 for (i = 0; i < sbio->page_count; i++) {
331 WARN_ON(!sbio->pagev[i]->page);
332 scrub_block_put(sbio->pagev[i]->sblock);
337 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
338 struct scrub_bio *sbio = sctx->bios[i];
345 scrub_free_csums(sctx);
349 static noinline_for_stack
350 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
352 struct scrub_ctx *sctx;
354 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
355 int pages_per_rd_bio;
359 * the setting of pages_per_rd_bio is correct for scrub but might
360 * be wrong for the dev_replace code where we might read from
361 * different devices in the initial huge bios. However, that
362 * code is able to correctly handle the case when adding a page
366 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
367 bio_get_nr_vecs(dev->bdev));
369 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
370 sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
373 sctx->is_dev_replace = is_dev_replace;
374 sctx->pages_per_rd_bio = pages_per_rd_bio;
376 sctx->dev_root = dev->dev_root;
377 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
378 struct scrub_bio *sbio;
380 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
383 sctx->bios[i] = sbio;
387 sbio->page_count = 0;
388 sbio->work.func = scrub_bio_end_io_worker;
390 if (i != SCRUB_BIOS_PER_SCTX - 1)
391 sctx->bios[i]->next_free = i + 1;
393 sctx->bios[i]->next_free = -1;
395 sctx->first_free = 0;
396 sctx->nodesize = dev->dev_root->nodesize;
397 sctx->leafsize = dev->dev_root->leafsize;
398 sctx->sectorsize = dev->dev_root->sectorsize;
399 atomic_set(&sctx->bios_in_flight, 0);
400 atomic_set(&sctx->workers_pending, 0);
401 atomic_set(&sctx->cancel_req, 0);
402 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
403 INIT_LIST_HEAD(&sctx->csum_list);
405 spin_lock_init(&sctx->list_lock);
406 spin_lock_init(&sctx->stat_lock);
407 init_waitqueue_head(&sctx->list_wait);
409 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
410 fs_info->dev_replace.tgtdev, is_dev_replace);
412 scrub_free_ctx(sctx);
418 scrub_free_ctx(sctx);
419 return ERR_PTR(-ENOMEM);
422 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
429 struct extent_buffer *eb;
430 struct btrfs_inode_item *inode_item;
431 struct scrub_warning *swarn = warn_ctx;
432 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
433 struct inode_fs_paths *ipath = NULL;
434 struct btrfs_root *local_root;
435 struct btrfs_key root_key;
437 root_key.objectid = root;
438 root_key.type = BTRFS_ROOT_ITEM_KEY;
439 root_key.offset = (u64)-1;
440 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
441 if (IS_ERR(local_root)) {
442 ret = PTR_ERR(local_root);
446 ret = inode_item_info(inum, 0, local_root, swarn->path);
448 btrfs_release_path(swarn->path);
452 eb = swarn->path->nodes[0];
453 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
454 struct btrfs_inode_item);
455 isize = btrfs_inode_size(eb, inode_item);
456 nlink = btrfs_inode_nlink(eb, inode_item);
457 btrfs_release_path(swarn->path);
459 ipath = init_ipath(4096, local_root, swarn->path);
461 ret = PTR_ERR(ipath);
465 ret = paths_from_inode(inum, ipath);
471 * we deliberately ignore the bit ipath might have been too small to
472 * hold all of the paths here
474 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
475 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
476 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
477 "length %llu, links %u (path: %s)\n", swarn->errstr,
478 swarn->logical, rcu_str_deref(swarn->dev->name),
479 (unsigned long long)swarn->sector, root, inum, offset,
480 min(isize - offset, (u64)PAGE_SIZE), nlink,
481 (char *)(unsigned long)ipath->fspath->val[i]);
487 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
488 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
489 "resolving failed with ret=%d\n", swarn->errstr,
490 swarn->logical, rcu_str_deref(swarn->dev->name),
491 (unsigned long long)swarn->sector, root, inum, offset, ret);
497 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
499 struct btrfs_device *dev;
500 struct btrfs_fs_info *fs_info;
501 struct btrfs_path *path;
502 struct btrfs_key found_key;
503 struct extent_buffer *eb;
504 struct btrfs_extent_item *ei;
505 struct scrub_warning swarn;
506 unsigned long ptr = 0;
512 const int bufsize = 4096;
515 WARN_ON(sblock->page_count < 1);
516 dev = sblock->pagev[0]->dev;
517 fs_info = sblock->sctx->dev_root->fs_info;
519 path = btrfs_alloc_path();
521 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
522 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
523 swarn.sector = (sblock->pagev[0]->physical) >> 9;
524 swarn.logical = sblock->pagev[0]->logical;
525 swarn.errstr = errstr;
527 swarn.msg_bufsize = bufsize;
528 swarn.scratch_bufsize = bufsize;
530 if (!path || !swarn.scratch_buf || !swarn.msg_buf)
533 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
538 extent_item_pos = swarn.logical - found_key.objectid;
539 swarn.extent_item_size = found_key.offset;
542 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
543 item_size = btrfs_item_size_nr(eb, path->slots[0]);
544 btrfs_release_path(path);
546 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
548 ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
549 &ref_root, &ref_level);
550 printk_in_rcu(KERN_WARNING
551 "btrfs: %s at logical %llu on dev %s, "
552 "sector %llu: metadata %s (level %d) in tree "
553 "%llu\n", errstr, swarn.logical,
554 rcu_str_deref(dev->name),
555 (unsigned long long)swarn.sector,
556 ref_level ? "node" : "leaf",
557 ret < 0 ? -1 : ref_level,
558 ret < 0 ? -1 : ref_root);
563 iterate_extent_inodes(fs_info, found_key.objectid,
565 scrub_print_warning_inode, &swarn);
569 btrfs_free_path(path);
570 kfree(swarn.scratch_buf);
571 kfree(swarn.msg_buf);
574 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
576 struct page *page = NULL;
578 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
581 struct btrfs_key key;
582 struct inode *inode = NULL;
583 u64 end = offset + PAGE_SIZE - 1;
584 struct btrfs_root *local_root;
587 key.type = BTRFS_ROOT_ITEM_KEY;
588 key.offset = (u64)-1;
589 local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key);
590 if (IS_ERR(local_root))
591 return PTR_ERR(local_root);
593 key.type = BTRFS_INODE_ITEM_KEY;
596 inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL);
598 return PTR_ERR(inode);
600 index = offset >> PAGE_CACHE_SHIFT;
602 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
608 if (PageUptodate(page)) {
609 struct btrfs_fs_info *fs_info;
610 if (PageDirty(page)) {
612 * we need to write the data to the defect sector. the
613 * data that was in that sector is not in memory,
614 * because the page was modified. we must not write the
615 * modified page to that sector.
617 * TODO: what could be done here: wait for the delalloc
618 * runner to write out that page (might involve
619 * COW) and see whether the sector is still
620 * referenced afterwards.
622 * For the meantime, we'll treat this error
623 * incorrectable, although there is a chance that a
624 * later scrub will find the bad sector again and that
625 * there's no dirty page in memory, then.
630 fs_info = BTRFS_I(inode)->root->fs_info;
631 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
632 fixup->logical, page,
638 * we need to get good data first. the general readpage path
639 * will call repair_io_failure for us, we just have to make
640 * sure we read the bad mirror.
642 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
643 EXTENT_DAMAGED, GFP_NOFS);
645 /* set_extent_bits should give proper error */
652 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
655 wait_on_page_locked(page);
657 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
658 end, EXTENT_DAMAGED, 0, NULL);
660 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
661 EXTENT_DAMAGED, GFP_NOFS);
673 if (ret == 0 && corrected) {
675 * we only need to call readpage for one of the inodes belonging
676 * to this extent. so make iterate_extent_inodes stop
684 static void scrub_fixup_nodatasum(struct btrfs_work *work)
687 struct scrub_fixup_nodatasum *fixup;
688 struct scrub_ctx *sctx;
689 struct btrfs_trans_handle *trans = NULL;
690 struct btrfs_fs_info *fs_info;
691 struct btrfs_path *path;
692 int uncorrectable = 0;
694 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
696 fs_info = fixup->root->fs_info;
698 path = btrfs_alloc_path();
700 spin_lock(&sctx->stat_lock);
701 ++sctx->stat.malloc_errors;
702 spin_unlock(&sctx->stat_lock);
707 trans = btrfs_join_transaction(fixup->root);
714 * the idea is to trigger a regular read through the standard path. we
715 * read a page from the (failed) logical address by specifying the
716 * corresponding copynum of the failed sector. thus, that readpage is
718 * that is the point where on-the-fly error correction will kick in
719 * (once it's finished) and rewrite the failed sector if a good copy
722 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
723 path, scrub_fixup_readpage,
731 spin_lock(&sctx->stat_lock);
732 ++sctx->stat.corrected_errors;
733 spin_unlock(&sctx->stat_lock);
736 if (trans && !IS_ERR(trans))
737 btrfs_end_transaction(trans, fixup->root);
739 spin_lock(&sctx->stat_lock);
740 ++sctx->stat.uncorrectable_errors;
741 spin_unlock(&sctx->stat_lock);
742 btrfs_dev_replace_stats_inc(
743 &sctx->dev_root->fs_info->dev_replace.
744 num_uncorrectable_read_errors);
745 printk_ratelimited_in_rcu(KERN_ERR
746 "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
747 (unsigned long long)fixup->logical,
748 rcu_str_deref(fixup->dev->name));
751 btrfs_free_path(path);
754 scrub_pending_trans_workers_dec(sctx);
758 * scrub_handle_errored_block gets called when either verification of the
759 * pages failed or the bio failed to read, e.g. with EIO. In the latter
760 * case, this function handles all pages in the bio, even though only one
762 * The goal of this function is to repair the errored block by using the
763 * contents of one of the mirrors.
765 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
767 struct scrub_ctx *sctx = sblock_to_check->sctx;
768 struct btrfs_device *dev;
769 struct btrfs_fs_info *fs_info;
773 unsigned int failed_mirror_index;
774 unsigned int is_metadata;
775 unsigned int have_csum;
777 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
778 struct scrub_block *sblock_bad;
783 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
784 DEFAULT_RATELIMIT_BURST);
786 BUG_ON(sblock_to_check->page_count < 1);
787 fs_info = sctx->dev_root->fs_info;
788 length = sblock_to_check->page_count * PAGE_SIZE;
789 logical = sblock_to_check->pagev[0]->logical;
790 generation = sblock_to_check->pagev[0]->generation;
791 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
792 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
793 is_metadata = !(sblock_to_check->pagev[0]->flags &
794 BTRFS_EXTENT_FLAG_DATA);
795 have_csum = sblock_to_check->pagev[0]->have_csum;
796 csum = sblock_to_check->pagev[0]->csum;
797 dev = sblock_to_check->pagev[0]->dev;
799 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
800 sblocks_for_recheck = NULL;
805 * read all mirrors one after the other. This includes to
806 * re-read the extent or metadata block that failed (that was
807 * the cause that this fixup code is called) another time,
808 * page by page this time in order to know which pages
809 * caused I/O errors and which ones are good (for all mirrors).
810 * It is the goal to handle the situation when more than one
811 * mirror contains I/O errors, but the errors do not
812 * overlap, i.e. the data can be repaired by selecting the
813 * pages from those mirrors without I/O error on the
814 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
815 * would be that mirror #1 has an I/O error on the first page,
816 * the second page is good, and mirror #2 has an I/O error on
817 * the second page, but the first page is good.
818 * Then the first page of the first mirror can be repaired by
819 * taking the first page of the second mirror, and the
820 * second page of the second mirror can be repaired by
821 * copying the contents of the 2nd page of the 1st mirror.
822 * One more note: if the pages of one mirror contain I/O
823 * errors, the checksum cannot be verified. In order to get
824 * the best data for repairing, the first attempt is to find
825 * a mirror without I/O errors and with a validated checksum.
826 * Only if this is not possible, the pages are picked from
827 * mirrors with I/O errors without considering the checksum.
828 * If the latter is the case, at the end, the checksum of the
829 * repaired area is verified in order to correctly maintain
833 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
834 sizeof(*sblocks_for_recheck),
836 if (!sblocks_for_recheck) {
837 spin_lock(&sctx->stat_lock);
838 sctx->stat.malloc_errors++;
839 sctx->stat.read_errors++;
840 sctx->stat.uncorrectable_errors++;
841 spin_unlock(&sctx->stat_lock);
842 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
846 /* setup the context, map the logical blocks and alloc the pages */
847 ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
848 logical, sblocks_for_recheck);
850 spin_lock(&sctx->stat_lock);
851 sctx->stat.read_errors++;
852 sctx->stat.uncorrectable_errors++;
853 spin_unlock(&sctx->stat_lock);
854 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
857 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
858 sblock_bad = sblocks_for_recheck + failed_mirror_index;
860 /* build and submit the bios for the failed mirror, check checksums */
861 scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
862 csum, generation, sctx->csum_size);
864 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
865 sblock_bad->no_io_error_seen) {
867 * the error disappeared after reading page by page, or
868 * the area was part of a huge bio and other parts of the
869 * bio caused I/O errors, or the block layer merged several
870 * read requests into one and the error is caused by a
871 * different bio (usually one of the two latter cases is
874 spin_lock(&sctx->stat_lock);
875 sctx->stat.unverified_errors++;
876 spin_unlock(&sctx->stat_lock);
878 if (sctx->is_dev_replace)
879 scrub_write_block_to_dev_replace(sblock_bad);
883 if (!sblock_bad->no_io_error_seen) {
884 spin_lock(&sctx->stat_lock);
885 sctx->stat.read_errors++;
886 spin_unlock(&sctx->stat_lock);
887 if (__ratelimit(&_rs))
888 scrub_print_warning("i/o error", sblock_to_check);
889 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
890 } else if (sblock_bad->checksum_error) {
891 spin_lock(&sctx->stat_lock);
892 sctx->stat.csum_errors++;
893 spin_unlock(&sctx->stat_lock);
894 if (__ratelimit(&_rs))
895 scrub_print_warning("checksum error", sblock_to_check);
896 btrfs_dev_stat_inc_and_print(dev,
897 BTRFS_DEV_STAT_CORRUPTION_ERRS);
898 } else if (sblock_bad->header_error) {
899 spin_lock(&sctx->stat_lock);
900 sctx->stat.verify_errors++;
901 spin_unlock(&sctx->stat_lock);
902 if (__ratelimit(&_rs))
903 scrub_print_warning("checksum/header error",
905 if (sblock_bad->generation_error)
906 btrfs_dev_stat_inc_and_print(dev,
907 BTRFS_DEV_STAT_GENERATION_ERRS);
909 btrfs_dev_stat_inc_and_print(dev,
910 BTRFS_DEV_STAT_CORRUPTION_ERRS);
913 if (sctx->readonly && !sctx->is_dev_replace)
914 goto did_not_correct_error;
916 if (!is_metadata && !have_csum) {
917 struct scrub_fixup_nodatasum *fixup_nodatasum;
920 WARN_ON(sctx->is_dev_replace);
923 * !is_metadata and !have_csum, this means that the data
924 * might not be COW'ed, that it might be modified
925 * concurrently. The general strategy to work on the
926 * commit root does not help in the case when COW is not
929 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
930 if (!fixup_nodatasum)
931 goto did_not_correct_error;
932 fixup_nodatasum->sctx = sctx;
933 fixup_nodatasum->dev = dev;
934 fixup_nodatasum->logical = logical;
935 fixup_nodatasum->root = fs_info->extent_root;
936 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
937 scrub_pending_trans_workers_inc(sctx);
938 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
939 btrfs_queue_worker(&fs_info->scrub_workers,
940 &fixup_nodatasum->work);
945 * now build and submit the bios for the other mirrors, check
947 * First try to pick the mirror which is completely without I/O
948 * errors and also does not have a checksum error.
949 * If one is found, and if a checksum is present, the full block
950 * that is known to contain an error is rewritten. Afterwards
951 * the block is known to be corrected.
952 * If a mirror is found which is completely correct, and no
953 * checksum is present, only those pages are rewritten that had
954 * an I/O error in the block to be repaired, since it cannot be
955 * determined, which copy of the other pages is better (and it
956 * could happen otherwise that a correct page would be
957 * overwritten by a bad one).
959 for (mirror_index = 0;
960 mirror_index < BTRFS_MAX_MIRRORS &&
961 sblocks_for_recheck[mirror_index].page_count > 0;
963 struct scrub_block *sblock_other;
965 if (mirror_index == failed_mirror_index)
967 sblock_other = sblocks_for_recheck + mirror_index;
969 /* build and submit the bios, check checksums */
970 scrub_recheck_block(fs_info, sblock_other, is_metadata,
971 have_csum, csum, generation,
974 if (!sblock_other->header_error &&
975 !sblock_other->checksum_error &&
976 sblock_other->no_io_error_seen) {
977 if (sctx->is_dev_replace) {
978 scrub_write_block_to_dev_replace(sblock_other);
980 int force_write = is_metadata || have_csum;
982 ret = scrub_repair_block_from_good_copy(
983 sblock_bad, sblock_other,
987 goto corrected_error;
992 * for dev_replace, pick good pages and write to the target device.
994 if (sctx->is_dev_replace) {
996 for (page_num = 0; page_num < sblock_bad->page_count;
1001 for (mirror_index = 0;
1002 mirror_index < BTRFS_MAX_MIRRORS &&
1003 sblocks_for_recheck[mirror_index].page_count > 0;
1005 struct scrub_block *sblock_other =
1006 sblocks_for_recheck + mirror_index;
1007 struct scrub_page *page_other =
1008 sblock_other->pagev[page_num];
1010 if (!page_other->io_error) {
1011 ret = scrub_write_page_to_dev_replace(
1012 sblock_other, page_num);
1014 /* succeeded for this page */
1018 btrfs_dev_replace_stats_inc(
1020 fs_info->dev_replace.
1028 * did not find a mirror to fetch the page
1029 * from. scrub_write_page_to_dev_replace()
1030 * handles this case (page->io_error), by
1031 * filling the block with zeros before
1032 * submitting the write request
1035 ret = scrub_write_page_to_dev_replace(
1036 sblock_bad, page_num);
1038 btrfs_dev_replace_stats_inc(
1039 &sctx->dev_root->fs_info->
1040 dev_replace.num_write_errors);
1048 * for regular scrub, repair those pages that are errored.
1049 * In case of I/O errors in the area that is supposed to be
1050 * repaired, continue by picking good copies of those pages.
1051 * Select the good pages from mirrors to rewrite bad pages from
1052 * the area to fix. Afterwards verify the checksum of the block
1053 * that is supposed to be repaired. This verification step is
1054 * only done for the purpose of statistic counting and for the
1055 * final scrub report, whether errors remain.
1056 * A perfect algorithm could make use of the checksum and try
1057 * all possible combinations of pages from the different mirrors
1058 * until the checksum verification succeeds. For example, when
1059 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1060 * of mirror #2 is readable but the final checksum test fails,
1061 * then the 2nd page of mirror #3 could be tried, whether now
1062 * the final checksum succeedes. But this would be a rare
1063 * exception and is therefore not implemented. At least it is
1064 * avoided that the good copy is overwritten.
1065 * A more useful improvement would be to pick the sectors
1066 * without I/O error based on sector sizes (512 bytes on legacy
1067 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1068 * mirror could be repaired by taking 512 byte of a different
1069 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1070 * area are unreadable.
1073 /* can only fix I/O errors from here on */
1074 if (sblock_bad->no_io_error_seen)
1075 goto did_not_correct_error;
1078 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1079 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1081 if (!page_bad->io_error)
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 = sblocks_for_recheck +
1090 struct scrub_page *page_other = sblock_other->pagev[
1093 if (!page_other->io_error) {
1094 ret = scrub_repair_page_from_good_copy(
1095 sblock_bad, sblock_other, page_num, 0);
1097 page_bad->io_error = 0;
1098 break; /* succeeded for this page */
1103 if (page_bad->io_error) {
1104 /* did not find a mirror to copy the page from */
1110 if (is_metadata || have_csum) {
1112 * need to verify the checksum now that all
1113 * sectors on disk are repaired (the write
1114 * request for data to be repaired is on its way).
1115 * Just be lazy and use scrub_recheck_block()
1116 * which re-reads the data before the checksum
1117 * is verified, but most likely the data comes out
1118 * of the page cache.
1120 scrub_recheck_block(fs_info, sblock_bad,
1121 is_metadata, have_csum, csum,
1122 generation, sctx->csum_size);
1123 if (!sblock_bad->header_error &&
1124 !sblock_bad->checksum_error &&
1125 sblock_bad->no_io_error_seen)
1126 goto corrected_error;
1128 goto did_not_correct_error;
1131 spin_lock(&sctx->stat_lock);
1132 sctx->stat.corrected_errors++;
1133 spin_unlock(&sctx->stat_lock);
1134 printk_ratelimited_in_rcu(KERN_ERR
1135 "btrfs: fixed up error at logical %llu on dev %s\n",
1136 (unsigned long long)logical,
1137 rcu_str_deref(dev->name));
1140 did_not_correct_error:
1141 spin_lock(&sctx->stat_lock);
1142 sctx->stat.uncorrectable_errors++;
1143 spin_unlock(&sctx->stat_lock);
1144 printk_ratelimited_in_rcu(KERN_ERR
1145 "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1146 (unsigned long long)logical,
1147 rcu_str_deref(dev->name));
1151 if (sblocks_for_recheck) {
1152 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1154 struct scrub_block *sblock = sblocks_for_recheck +
1158 for (page_index = 0; page_index < sblock->page_count;
1160 sblock->pagev[page_index]->sblock = NULL;
1161 scrub_page_put(sblock->pagev[page_index]);
1164 kfree(sblocks_for_recheck);
1170 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1171 struct btrfs_fs_info *fs_info,
1172 struct scrub_block *original_sblock,
1173 u64 length, u64 logical,
1174 struct scrub_block *sblocks_for_recheck)
1181 * note: the two members ref_count and outstanding_pages
1182 * are not used (and not set) in the blocks that are used for
1183 * the recheck procedure
1187 while (length > 0) {
1188 u64 sublen = min_t(u64, length, PAGE_SIZE);
1189 u64 mapped_length = sublen;
1190 struct btrfs_bio *bbio = NULL;
1193 * with a length of PAGE_SIZE, each returned stripe
1194 * represents one mirror
1196 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1197 &mapped_length, &bbio, 0);
1198 if (ret || !bbio || mapped_length < sublen) {
1203 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1204 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1206 struct scrub_block *sblock;
1207 struct scrub_page *page;
1209 if (mirror_index >= BTRFS_MAX_MIRRORS)
1212 sblock = sblocks_for_recheck + mirror_index;
1213 sblock->sctx = sctx;
1214 page = kzalloc(sizeof(*page), GFP_NOFS);
1217 spin_lock(&sctx->stat_lock);
1218 sctx->stat.malloc_errors++;
1219 spin_unlock(&sctx->stat_lock);
1223 scrub_page_get(page);
1224 sblock->pagev[page_index] = page;
1225 page->logical = logical;
1226 page->physical = bbio->stripes[mirror_index].physical;
1227 BUG_ON(page_index >= original_sblock->page_count);
1228 page->physical_for_dev_replace =
1229 original_sblock->pagev[page_index]->
1230 physical_for_dev_replace;
1231 /* for missing devices, dev->bdev is NULL */
1232 page->dev = bbio->stripes[mirror_index].dev;
1233 page->mirror_num = mirror_index + 1;
1234 sblock->page_count++;
1235 page->page = alloc_page(GFP_NOFS);
1249 * this function will check the on disk data for checksum errors, header
1250 * errors and read I/O errors. If any I/O errors happen, the exact pages
1251 * which are errored are marked as being bad. The goal is to enable scrub
1252 * to take those pages that are not errored from all the mirrors so that
1253 * the pages that are errored in the just handled mirror can be repaired.
1255 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1256 struct scrub_block *sblock, int is_metadata,
1257 int have_csum, u8 *csum, u64 generation,
1262 sblock->no_io_error_seen = 1;
1263 sblock->header_error = 0;
1264 sblock->checksum_error = 0;
1266 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1268 struct scrub_page *page = sblock->pagev[page_num];
1269 DECLARE_COMPLETION_ONSTACK(complete);
1271 if (page->dev->bdev == NULL) {
1273 sblock->no_io_error_seen = 0;
1277 WARN_ON(!page->page);
1278 bio = bio_alloc(GFP_NOFS, 1);
1281 sblock->no_io_error_seen = 0;
1284 bio->bi_bdev = page->dev->bdev;
1285 bio->bi_sector = page->physical >> 9;
1286 bio->bi_end_io = scrub_complete_bio_end_io;
1287 bio->bi_private = &complete;
1289 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1290 btrfsic_submit_bio(READ, bio);
1292 /* this will also unplug the queue */
1293 wait_for_completion(&complete);
1295 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1296 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1297 sblock->no_io_error_seen = 0;
1301 if (sblock->no_io_error_seen)
1302 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1303 have_csum, csum, generation,
1309 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1310 struct scrub_block *sblock,
1311 int is_metadata, int have_csum,
1312 const u8 *csum, u64 generation,
1316 u8 calculated_csum[BTRFS_CSUM_SIZE];
1318 struct btrfs_root *root = fs_info->extent_root;
1319 void *mapped_buffer;
1321 WARN_ON(!sblock->pagev[0]->page);
1323 struct btrfs_header *h;
1325 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1326 h = (struct btrfs_header *)mapped_buffer;
1328 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1329 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1330 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1332 sblock->header_error = 1;
1333 } else if (generation != le64_to_cpu(h->generation)) {
1334 sblock->header_error = 1;
1335 sblock->generation_error = 1;
1342 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1345 for (page_num = 0;;) {
1346 if (page_num == 0 && is_metadata)
1347 crc = btrfs_csum_data(root,
1348 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1349 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1351 crc = btrfs_csum_data(root, mapped_buffer, crc,
1354 kunmap_atomic(mapped_buffer);
1356 if (page_num >= sblock->page_count)
1358 WARN_ON(!sblock->pagev[page_num]->page);
1360 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1363 btrfs_csum_final(crc, calculated_csum);
1364 if (memcmp(calculated_csum, csum, csum_size))
1365 sblock->checksum_error = 1;
1368 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1370 complete((struct completion *)bio->bi_private);
1373 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1374 struct scrub_block *sblock_good,
1380 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1383 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1394 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1395 struct scrub_block *sblock_good,
1396 int page_num, int force_write)
1398 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1399 struct scrub_page *page_good = sblock_good->pagev[page_num];
1401 BUG_ON(page_bad->page == NULL);
1402 BUG_ON(page_good->page == NULL);
1403 if (force_write || sblock_bad->header_error ||
1404 sblock_bad->checksum_error || page_bad->io_error) {
1407 DECLARE_COMPLETION_ONSTACK(complete);
1409 if (!page_bad->dev->bdev) {
1410 printk_ratelimited(KERN_WARNING
1411 "btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1415 bio = bio_alloc(GFP_NOFS, 1);
1418 bio->bi_bdev = page_bad->dev->bdev;
1419 bio->bi_sector = page_bad->physical >> 9;
1420 bio->bi_end_io = scrub_complete_bio_end_io;
1421 bio->bi_private = &complete;
1423 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1424 if (PAGE_SIZE != ret) {
1428 btrfsic_submit_bio(WRITE, bio);
1430 /* this will also unplug the queue */
1431 wait_for_completion(&complete);
1432 if (!bio_flagged(bio, BIO_UPTODATE)) {
1433 btrfs_dev_stat_inc_and_print(page_bad->dev,
1434 BTRFS_DEV_STAT_WRITE_ERRS);
1435 btrfs_dev_replace_stats_inc(
1436 &sblock_bad->sctx->dev_root->fs_info->
1437 dev_replace.num_write_errors);
1447 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1451 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1454 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1456 btrfs_dev_replace_stats_inc(
1457 &sblock->sctx->dev_root->fs_info->dev_replace.
1462 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1465 struct scrub_page *spage = sblock->pagev[page_num];
1467 BUG_ON(spage->page == NULL);
1468 if (spage->io_error) {
1469 void *mapped_buffer = kmap_atomic(spage->page);
1471 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1472 flush_dcache_page(spage->page);
1473 kunmap_atomic(mapped_buffer);
1475 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1478 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1479 struct scrub_page *spage)
1481 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1482 struct scrub_bio *sbio;
1485 mutex_lock(&wr_ctx->wr_lock);
1487 if (!wr_ctx->wr_curr_bio) {
1488 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1490 if (!wr_ctx->wr_curr_bio) {
1491 mutex_unlock(&wr_ctx->wr_lock);
1494 wr_ctx->wr_curr_bio->sctx = sctx;
1495 wr_ctx->wr_curr_bio->page_count = 0;
1497 sbio = wr_ctx->wr_curr_bio;
1498 if (sbio->page_count == 0) {
1501 sbio->physical = spage->physical_for_dev_replace;
1502 sbio->logical = spage->logical;
1503 sbio->dev = wr_ctx->tgtdev;
1506 bio = bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1508 mutex_unlock(&wr_ctx->wr_lock);
1514 bio->bi_private = sbio;
1515 bio->bi_end_io = scrub_wr_bio_end_io;
1516 bio->bi_bdev = sbio->dev->bdev;
1517 bio->bi_sector = sbio->physical >> 9;
1519 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1520 spage->physical_for_dev_replace ||
1521 sbio->logical + sbio->page_count * PAGE_SIZE !=
1523 scrub_wr_submit(sctx);
1527 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1528 if (ret != PAGE_SIZE) {
1529 if (sbio->page_count < 1) {
1532 mutex_unlock(&wr_ctx->wr_lock);
1535 scrub_wr_submit(sctx);
1539 sbio->pagev[sbio->page_count] = spage;
1540 scrub_page_get(spage);
1542 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1543 scrub_wr_submit(sctx);
1544 mutex_unlock(&wr_ctx->wr_lock);
1549 static void scrub_wr_submit(struct scrub_ctx *sctx)
1551 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1552 struct scrub_bio *sbio;
1554 if (!wr_ctx->wr_curr_bio)
1557 sbio = wr_ctx->wr_curr_bio;
1558 wr_ctx->wr_curr_bio = NULL;
1559 WARN_ON(!sbio->bio->bi_bdev);
1560 scrub_pending_bio_inc(sctx);
1561 /* process all writes in a single worker thread. Then the block layer
1562 * orders the requests before sending them to the driver which
1563 * doubled the write performance on spinning disks when measured
1565 btrfsic_submit_bio(WRITE, sbio->bio);
1568 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1570 struct scrub_bio *sbio = bio->bi_private;
1571 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1576 sbio->work.func = scrub_wr_bio_end_io_worker;
1577 btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1580 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1582 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1583 struct scrub_ctx *sctx = sbio->sctx;
1586 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1588 struct btrfs_dev_replace *dev_replace =
1589 &sbio->sctx->dev_root->fs_info->dev_replace;
1591 for (i = 0; i < sbio->page_count; i++) {
1592 struct scrub_page *spage = sbio->pagev[i];
1594 spage->io_error = 1;
1595 btrfs_dev_replace_stats_inc(&dev_replace->
1600 for (i = 0; i < sbio->page_count; i++)
1601 scrub_page_put(sbio->pagev[i]);
1605 scrub_pending_bio_dec(sctx);
1608 static int scrub_checksum(struct scrub_block *sblock)
1613 WARN_ON(sblock->page_count < 1);
1614 flags = sblock->pagev[0]->flags;
1616 if (flags & BTRFS_EXTENT_FLAG_DATA)
1617 ret = scrub_checksum_data(sblock);
1618 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1619 ret = scrub_checksum_tree_block(sblock);
1620 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1621 (void)scrub_checksum_super(sblock);
1625 scrub_handle_errored_block(sblock);
1630 static int scrub_checksum_data(struct scrub_block *sblock)
1632 struct scrub_ctx *sctx = sblock->sctx;
1633 u8 csum[BTRFS_CSUM_SIZE];
1639 struct btrfs_root *root = sctx->dev_root;
1643 BUG_ON(sblock->page_count < 1);
1644 if (!sblock->pagev[0]->have_csum)
1647 on_disk_csum = sblock->pagev[0]->csum;
1648 page = sblock->pagev[0]->page;
1649 buffer = kmap_atomic(page);
1651 len = sctx->sectorsize;
1654 u64 l = min_t(u64, len, PAGE_SIZE);
1656 crc = btrfs_csum_data(root, buffer, crc, l);
1657 kunmap_atomic(buffer);
1662 BUG_ON(index >= sblock->page_count);
1663 BUG_ON(!sblock->pagev[index]->page);
1664 page = sblock->pagev[index]->page;
1665 buffer = kmap_atomic(page);
1668 btrfs_csum_final(crc, csum);
1669 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1675 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1677 struct scrub_ctx *sctx = sblock->sctx;
1678 struct btrfs_header *h;
1679 struct btrfs_root *root = sctx->dev_root;
1680 struct btrfs_fs_info *fs_info = root->fs_info;
1681 u8 calculated_csum[BTRFS_CSUM_SIZE];
1682 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1684 void *mapped_buffer;
1693 BUG_ON(sblock->page_count < 1);
1694 page = sblock->pagev[0]->page;
1695 mapped_buffer = kmap_atomic(page);
1696 h = (struct btrfs_header *)mapped_buffer;
1697 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1700 * we don't use the getter functions here, as we
1701 * a) don't have an extent buffer and
1702 * b) the page is already kmapped
1705 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1708 if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1711 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1714 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1718 WARN_ON(sctx->nodesize != sctx->leafsize);
1719 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1720 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1721 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1724 u64 l = min_t(u64, len, mapped_size);
1726 crc = btrfs_csum_data(root, p, crc, l);
1727 kunmap_atomic(mapped_buffer);
1732 BUG_ON(index >= sblock->page_count);
1733 BUG_ON(!sblock->pagev[index]->page);
1734 page = sblock->pagev[index]->page;
1735 mapped_buffer = kmap_atomic(page);
1736 mapped_size = PAGE_SIZE;
1740 btrfs_csum_final(crc, calculated_csum);
1741 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1744 return fail || crc_fail;
1747 static int scrub_checksum_super(struct scrub_block *sblock)
1749 struct btrfs_super_block *s;
1750 struct scrub_ctx *sctx = sblock->sctx;
1751 struct btrfs_root *root = sctx->dev_root;
1752 struct btrfs_fs_info *fs_info = root->fs_info;
1753 u8 calculated_csum[BTRFS_CSUM_SIZE];
1754 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1756 void *mapped_buffer;
1765 BUG_ON(sblock->page_count < 1);
1766 page = sblock->pagev[0]->page;
1767 mapped_buffer = kmap_atomic(page);
1768 s = (struct btrfs_super_block *)mapped_buffer;
1769 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1771 if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1774 if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1777 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1780 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1781 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1782 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1785 u64 l = min_t(u64, len, mapped_size);
1787 crc = btrfs_csum_data(root, p, crc, l);
1788 kunmap_atomic(mapped_buffer);
1793 BUG_ON(index >= sblock->page_count);
1794 BUG_ON(!sblock->pagev[index]->page);
1795 page = sblock->pagev[index]->page;
1796 mapped_buffer = kmap_atomic(page);
1797 mapped_size = PAGE_SIZE;
1801 btrfs_csum_final(crc, calculated_csum);
1802 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1805 if (fail_cor + fail_gen) {
1807 * if we find an error in a super block, we just report it.
1808 * They will get written with the next transaction commit
1811 spin_lock(&sctx->stat_lock);
1812 ++sctx->stat.super_errors;
1813 spin_unlock(&sctx->stat_lock);
1815 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1816 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1818 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1819 BTRFS_DEV_STAT_GENERATION_ERRS);
1822 return fail_cor + fail_gen;
1825 static void scrub_block_get(struct scrub_block *sblock)
1827 atomic_inc(&sblock->ref_count);
1830 static void scrub_block_put(struct scrub_block *sblock)
1832 if (atomic_dec_and_test(&sblock->ref_count)) {
1835 for (i = 0; i < sblock->page_count; i++)
1836 scrub_page_put(sblock->pagev[i]);
1841 static void scrub_page_get(struct scrub_page *spage)
1843 atomic_inc(&spage->ref_count);
1846 static void scrub_page_put(struct scrub_page *spage)
1848 if (atomic_dec_and_test(&spage->ref_count)) {
1850 __free_page(spage->page);
1855 static void scrub_submit(struct scrub_ctx *sctx)
1857 struct scrub_bio *sbio;
1859 if (sctx->curr == -1)
1862 sbio = sctx->bios[sctx->curr];
1864 scrub_pending_bio_inc(sctx);
1866 if (!sbio->bio->bi_bdev) {
1868 * this case should not happen. If btrfs_map_block() is
1869 * wrong, it could happen for dev-replace operations on
1870 * missing devices when no mirrors are available, but in
1871 * this case it should already fail the mount.
1872 * This case is handled correctly (but _very_ slowly).
1874 printk_ratelimited(KERN_WARNING
1875 "btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1876 bio_endio(sbio->bio, -EIO);
1878 btrfsic_submit_bio(READ, sbio->bio);
1882 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1883 struct scrub_page *spage)
1885 struct scrub_block *sblock = spage->sblock;
1886 struct scrub_bio *sbio;
1891 * grab a fresh bio or wait for one to become available
1893 while (sctx->curr == -1) {
1894 spin_lock(&sctx->list_lock);
1895 sctx->curr = sctx->first_free;
1896 if (sctx->curr != -1) {
1897 sctx->first_free = sctx->bios[sctx->curr]->next_free;
1898 sctx->bios[sctx->curr]->next_free = -1;
1899 sctx->bios[sctx->curr]->page_count = 0;
1900 spin_unlock(&sctx->list_lock);
1902 spin_unlock(&sctx->list_lock);
1903 wait_event(sctx->list_wait, sctx->first_free != -1);
1906 sbio = sctx->bios[sctx->curr];
1907 if (sbio->page_count == 0) {
1910 sbio->physical = spage->physical;
1911 sbio->logical = spage->logical;
1912 sbio->dev = spage->dev;
1915 bio = bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1921 bio->bi_private = sbio;
1922 bio->bi_end_io = scrub_bio_end_io;
1923 bio->bi_bdev = sbio->dev->bdev;
1924 bio->bi_sector = sbio->physical >> 9;
1926 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1928 sbio->logical + sbio->page_count * PAGE_SIZE !=
1930 sbio->dev != spage->dev) {
1935 sbio->pagev[sbio->page_count] = spage;
1936 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1937 if (ret != PAGE_SIZE) {
1938 if (sbio->page_count < 1) {
1947 scrub_block_get(sblock); /* one for the page added to the bio */
1948 atomic_inc(&sblock->outstanding_pages);
1950 if (sbio->page_count == sctx->pages_per_rd_bio)
1956 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1957 u64 physical, struct btrfs_device *dev, u64 flags,
1958 u64 gen, int mirror_num, u8 *csum, int force,
1959 u64 physical_for_dev_replace)
1961 struct scrub_block *sblock;
1964 sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1966 spin_lock(&sctx->stat_lock);
1967 sctx->stat.malloc_errors++;
1968 spin_unlock(&sctx->stat_lock);
1972 /* one ref inside this function, plus one for each page added to
1974 atomic_set(&sblock->ref_count, 1);
1975 sblock->sctx = sctx;
1976 sblock->no_io_error_seen = 1;
1978 for (index = 0; len > 0; index++) {
1979 struct scrub_page *spage;
1980 u64 l = min_t(u64, len, PAGE_SIZE);
1982 spage = kzalloc(sizeof(*spage), GFP_NOFS);
1985 spin_lock(&sctx->stat_lock);
1986 sctx->stat.malloc_errors++;
1987 spin_unlock(&sctx->stat_lock);
1988 scrub_block_put(sblock);
1991 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1992 scrub_page_get(spage);
1993 sblock->pagev[index] = spage;
1994 spage->sblock = sblock;
1996 spage->flags = flags;
1997 spage->generation = gen;
1998 spage->logical = logical;
1999 spage->physical = physical;
2000 spage->physical_for_dev_replace = physical_for_dev_replace;
2001 spage->mirror_num = mirror_num;
2003 spage->have_csum = 1;
2004 memcpy(spage->csum, csum, sctx->csum_size);
2006 spage->have_csum = 0;
2008 sblock->page_count++;
2009 spage->page = alloc_page(GFP_NOFS);
2015 physical_for_dev_replace += l;
2018 WARN_ON(sblock->page_count == 0);
2019 for (index = 0; index < sblock->page_count; index++) {
2020 struct scrub_page *spage = sblock->pagev[index];
2023 ret = scrub_add_page_to_rd_bio(sctx, spage);
2025 scrub_block_put(sblock);
2033 /* last one frees, either here or in bio completion for last page */
2034 scrub_block_put(sblock);
2038 static void scrub_bio_end_io(struct bio *bio, int err)
2040 struct scrub_bio *sbio = bio->bi_private;
2041 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2046 btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2049 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2051 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2052 struct scrub_ctx *sctx = sbio->sctx;
2055 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2057 for (i = 0; i < sbio->page_count; i++) {
2058 struct scrub_page *spage = sbio->pagev[i];
2060 spage->io_error = 1;
2061 spage->sblock->no_io_error_seen = 0;
2065 /* now complete the scrub_block items that have all pages completed */
2066 for (i = 0; i < sbio->page_count; i++) {
2067 struct scrub_page *spage = sbio->pagev[i];
2068 struct scrub_block *sblock = spage->sblock;
2070 if (atomic_dec_and_test(&sblock->outstanding_pages))
2071 scrub_block_complete(sblock);
2072 scrub_block_put(sblock);
2077 spin_lock(&sctx->list_lock);
2078 sbio->next_free = sctx->first_free;
2079 sctx->first_free = sbio->index;
2080 spin_unlock(&sctx->list_lock);
2082 if (sctx->is_dev_replace &&
2083 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2084 mutex_lock(&sctx->wr_ctx.wr_lock);
2085 scrub_wr_submit(sctx);
2086 mutex_unlock(&sctx->wr_ctx.wr_lock);
2089 scrub_pending_bio_dec(sctx);
2092 static void scrub_block_complete(struct scrub_block *sblock)
2094 if (!sblock->no_io_error_seen) {
2095 scrub_handle_errored_block(sblock);
2098 * if has checksum error, write via repair mechanism in
2099 * dev replace case, otherwise write here in dev replace
2102 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2103 scrub_write_block_to_dev_replace(sblock);
2107 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2110 struct btrfs_ordered_sum *sum = NULL;
2113 unsigned long num_sectors;
2115 while (!list_empty(&sctx->csum_list)) {
2116 sum = list_first_entry(&sctx->csum_list,
2117 struct btrfs_ordered_sum, list);
2118 if (sum->bytenr > logical)
2120 if (sum->bytenr + sum->len > logical)
2123 ++sctx->stat.csum_discards;
2124 list_del(&sum->list);
2131 num_sectors = sum->len / sctx->sectorsize;
2132 for (i = 0; i < num_sectors; ++i) {
2133 if (sum->sums[i].bytenr == logical) {
2134 memcpy(csum, &sum->sums[i].sum, sctx->csum_size);
2139 if (ret && i == num_sectors - 1) {
2140 list_del(&sum->list);
2146 /* scrub extent tries to collect up to 64 kB for each bio */
2147 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2148 u64 physical, struct btrfs_device *dev, u64 flags,
2149 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2152 u8 csum[BTRFS_CSUM_SIZE];
2155 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2156 blocksize = sctx->sectorsize;
2157 spin_lock(&sctx->stat_lock);
2158 sctx->stat.data_extents_scrubbed++;
2159 sctx->stat.data_bytes_scrubbed += len;
2160 spin_unlock(&sctx->stat_lock);
2161 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2162 WARN_ON(sctx->nodesize != sctx->leafsize);
2163 blocksize = sctx->nodesize;
2164 spin_lock(&sctx->stat_lock);
2165 sctx->stat.tree_extents_scrubbed++;
2166 sctx->stat.tree_bytes_scrubbed += len;
2167 spin_unlock(&sctx->stat_lock);
2169 blocksize = sctx->sectorsize;
2174 u64 l = min_t(u64, len, blocksize);
2177 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2178 /* push csums to sbio */
2179 have_csum = scrub_find_csum(sctx, logical, l, csum);
2181 ++sctx->stat.no_csum;
2182 if (sctx->is_dev_replace && !have_csum) {
2183 ret = copy_nocow_pages(sctx, logical, l,
2185 physical_for_dev_replace);
2186 goto behind_scrub_pages;
2189 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2190 mirror_num, have_csum ? csum : NULL, 0,
2191 physical_for_dev_replace);
2198 physical_for_dev_replace += l;
2203 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2204 struct map_lookup *map,
2205 struct btrfs_device *scrub_dev,
2206 int num, u64 base, u64 length,
2209 struct btrfs_path *path;
2210 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2211 struct btrfs_root *root = fs_info->extent_root;
2212 struct btrfs_root *csum_root = fs_info->csum_root;
2213 struct btrfs_extent_item *extent;
2214 struct blk_plug plug;
2220 struct extent_buffer *l;
2221 struct btrfs_key key;
2226 struct reada_control *reada1;
2227 struct reada_control *reada2;
2228 struct btrfs_key key_start;
2229 struct btrfs_key key_end;
2230 u64 increment = map->stripe_len;
2233 u64 extent_physical;
2235 struct btrfs_device *extent_dev;
2236 int extent_mirror_num;
2240 do_div(nstripes, map->stripe_len);
2241 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2242 offset = map->stripe_len * num;
2243 increment = map->stripe_len * map->num_stripes;
2245 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2246 int factor = map->num_stripes / map->sub_stripes;
2247 offset = map->stripe_len * (num / map->sub_stripes);
2248 increment = map->stripe_len * factor;
2249 mirror_num = num % map->sub_stripes + 1;
2250 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2251 increment = map->stripe_len;
2252 mirror_num = num % map->num_stripes + 1;
2253 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2254 increment = map->stripe_len;
2255 mirror_num = num % map->num_stripes + 1;
2257 increment = map->stripe_len;
2261 path = btrfs_alloc_path();
2266 * work on commit root. The related disk blocks are static as
2267 * long as COW is applied. This means, it is save to rewrite
2268 * them to repair disk errors without any race conditions
2270 path->search_commit_root = 1;
2271 path->skip_locking = 1;
2274 * trigger the readahead for extent tree csum tree and wait for
2275 * completion. During readahead, the scrub is officially paused
2276 * to not hold off transaction commits
2278 logical = base + offset;
2280 wait_event(sctx->list_wait,
2281 atomic_read(&sctx->bios_in_flight) == 0);
2282 atomic_inc(&fs_info->scrubs_paused);
2283 wake_up(&fs_info->scrub_pause_wait);
2285 /* FIXME it might be better to start readahead at commit root */
2286 key_start.objectid = logical;
2287 key_start.type = BTRFS_EXTENT_ITEM_KEY;
2288 key_start.offset = (u64)0;
2289 key_end.objectid = base + offset + nstripes * increment;
2290 key_end.type = BTRFS_EXTENT_ITEM_KEY;
2291 key_end.offset = (u64)0;
2292 reada1 = btrfs_reada_add(root, &key_start, &key_end);
2294 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2295 key_start.type = BTRFS_EXTENT_CSUM_KEY;
2296 key_start.offset = logical;
2297 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2298 key_end.type = BTRFS_EXTENT_CSUM_KEY;
2299 key_end.offset = base + offset + nstripes * increment;
2300 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2302 if (!IS_ERR(reada1))
2303 btrfs_reada_wait(reada1);
2304 if (!IS_ERR(reada2))
2305 btrfs_reada_wait(reada2);
2307 mutex_lock(&fs_info->scrub_lock);
2308 while (atomic_read(&fs_info->scrub_pause_req)) {
2309 mutex_unlock(&fs_info->scrub_lock);
2310 wait_event(fs_info->scrub_pause_wait,
2311 atomic_read(&fs_info->scrub_pause_req) == 0);
2312 mutex_lock(&fs_info->scrub_lock);
2314 atomic_dec(&fs_info->scrubs_paused);
2315 mutex_unlock(&fs_info->scrub_lock);
2316 wake_up(&fs_info->scrub_pause_wait);
2319 * collect all data csums for the stripe to avoid seeking during
2320 * the scrub. This might currently (crc32) end up to be about 1MB
2322 blk_start_plug(&plug);
2325 * now find all extents for each stripe and scrub them
2327 logical = base + offset;
2328 physical = map->stripes[num].physical;
2330 for (i = 0; i < nstripes; ++i) {
2334 if (atomic_read(&fs_info->scrub_cancel_req) ||
2335 atomic_read(&sctx->cancel_req)) {
2340 * check to see if we have to pause
2342 if (atomic_read(&fs_info->scrub_pause_req)) {
2343 /* push queued extents */
2344 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2346 mutex_lock(&sctx->wr_ctx.wr_lock);
2347 scrub_wr_submit(sctx);
2348 mutex_unlock(&sctx->wr_ctx.wr_lock);
2349 wait_event(sctx->list_wait,
2350 atomic_read(&sctx->bios_in_flight) == 0);
2351 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2352 atomic_inc(&fs_info->scrubs_paused);
2353 wake_up(&fs_info->scrub_pause_wait);
2354 mutex_lock(&fs_info->scrub_lock);
2355 while (atomic_read(&fs_info->scrub_pause_req)) {
2356 mutex_unlock(&fs_info->scrub_lock);
2357 wait_event(fs_info->scrub_pause_wait,
2358 atomic_read(&fs_info->scrub_pause_req) == 0);
2359 mutex_lock(&fs_info->scrub_lock);
2361 atomic_dec(&fs_info->scrubs_paused);
2362 mutex_unlock(&fs_info->scrub_lock);
2363 wake_up(&fs_info->scrub_pause_wait);
2366 ret = btrfs_lookup_csums_range(csum_root, logical,
2367 logical + map->stripe_len - 1,
2368 &sctx->csum_list, 1);
2372 key.objectid = logical;
2373 key.type = BTRFS_EXTENT_ITEM_KEY;
2374 key.offset = (u64)0;
2376 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2380 ret = btrfs_previous_item(root, path, 0,
2381 BTRFS_EXTENT_ITEM_KEY);
2385 /* there's no smaller item, so stick with the
2387 btrfs_release_path(path);
2388 ret = btrfs_search_slot(NULL, root, &key,
2397 slot = path->slots[0];
2398 if (slot >= btrfs_header_nritems(l)) {
2399 ret = btrfs_next_leaf(root, path);
2407 btrfs_item_key_to_cpu(l, &key, slot);
2409 if (key.objectid + key.offset <= logical)
2412 if (key.objectid >= logical + map->stripe_len)
2415 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
2418 extent = btrfs_item_ptr(l, slot,
2419 struct btrfs_extent_item);
2420 flags = btrfs_extent_flags(l, extent);
2421 generation = btrfs_extent_generation(l, extent);
2423 if (key.objectid < logical &&
2424 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2426 "btrfs scrub: tree block %llu spanning "
2427 "stripes, ignored. logical=%llu\n",
2428 (unsigned long long)key.objectid,
2429 (unsigned long long)logical);
2434 * trim extent to this stripe
2436 if (key.objectid < logical) {
2437 key.offset -= logical - key.objectid;
2438 key.objectid = logical;
2440 if (key.objectid + key.offset >
2441 logical + map->stripe_len) {
2442 key.offset = logical + map->stripe_len -
2446 extent_logical = key.objectid;
2447 extent_physical = key.objectid - logical + physical;
2448 extent_len = key.offset;
2449 extent_dev = scrub_dev;
2450 extent_mirror_num = mirror_num;
2452 scrub_remap_extent(fs_info, extent_logical,
2453 extent_len, &extent_physical,
2455 &extent_mirror_num);
2456 ret = scrub_extent(sctx, extent_logical, extent_len,
2457 extent_physical, extent_dev, flags,
2458 generation, extent_mirror_num,
2459 key.objectid - logical + physical);
2466 btrfs_release_path(path);
2467 logical += increment;
2468 physical += map->stripe_len;
2469 spin_lock(&sctx->stat_lock);
2470 sctx->stat.last_physical = physical;
2471 spin_unlock(&sctx->stat_lock);
2474 /* push queued extents */
2476 mutex_lock(&sctx->wr_ctx.wr_lock);
2477 scrub_wr_submit(sctx);
2478 mutex_unlock(&sctx->wr_ctx.wr_lock);
2480 blk_finish_plug(&plug);
2481 btrfs_free_path(path);
2482 return ret < 0 ? ret : 0;
2485 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2486 struct btrfs_device *scrub_dev,
2487 u64 chunk_tree, u64 chunk_objectid,
2488 u64 chunk_offset, u64 length,
2489 u64 dev_offset, int is_dev_replace)
2491 struct btrfs_mapping_tree *map_tree =
2492 &sctx->dev_root->fs_info->mapping_tree;
2493 struct map_lookup *map;
2494 struct extent_map *em;
2498 read_lock(&map_tree->map_tree.lock);
2499 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2500 read_unlock(&map_tree->map_tree.lock);
2505 map = (struct map_lookup *)em->bdev;
2506 if (em->start != chunk_offset)
2509 if (em->len < length)
2512 for (i = 0; i < map->num_stripes; ++i) {
2513 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2514 map->stripes[i].physical == dev_offset) {
2515 ret = scrub_stripe(sctx, map, scrub_dev, i,
2516 chunk_offset, length,
2523 free_extent_map(em);
2528 static noinline_for_stack
2529 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2530 struct btrfs_device *scrub_dev, u64 start, u64 end,
2533 struct btrfs_dev_extent *dev_extent = NULL;
2534 struct btrfs_path *path;
2535 struct btrfs_root *root = sctx->dev_root;
2536 struct btrfs_fs_info *fs_info = root->fs_info;
2543 struct extent_buffer *l;
2544 struct btrfs_key key;
2545 struct btrfs_key found_key;
2546 struct btrfs_block_group_cache *cache;
2547 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2549 path = btrfs_alloc_path();
2554 path->search_commit_root = 1;
2555 path->skip_locking = 1;
2557 key.objectid = scrub_dev->devid;
2559 key.type = BTRFS_DEV_EXTENT_KEY;
2562 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2566 if (path->slots[0] >=
2567 btrfs_header_nritems(path->nodes[0])) {
2568 ret = btrfs_next_leaf(root, path);
2575 slot = path->slots[0];
2577 btrfs_item_key_to_cpu(l, &found_key, slot);
2579 if (found_key.objectid != scrub_dev->devid)
2582 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2585 if (found_key.offset >= end)
2588 if (found_key.offset < key.offset)
2591 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2592 length = btrfs_dev_extent_length(l, dev_extent);
2594 if (found_key.offset + length <= start) {
2595 key.offset = found_key.offset + length;
2596 btrfs_release_path(path);
2600 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2601 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2602 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2605 * get a reference on the corresponding block group to prevent
2606 * the chunk from going away while we scrub it
2608 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2613 dev_replace->cursor_right = found_key.offset + length;
2614 dev_replace->cursor_left = found_key.offset;
2615 dev_replace->item_needs_writeback = 1;
2616 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2617 chunk_offset, length, found_key.offset,
2621 * flush, submit all pending read and write bios, afterwards
2623 * Note that in the dev replace case, a read request causes
2624 * write requests that are submitted in the read completion
2625 * worker. Therefore in the current situation, it is required
2626 * that all write requests are flushed, so that all read and
2627 * write requests are really completed when bios_in_flight
2630 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2632 mutex_lock(&sctx->wr_ctx.wr_lock);
2633 scrub_wr_submit(sctx);
2634 mutex_unlock(&sctx->wr_ctx.wr_lock);
2636 wait_event(sctx->list_wait,
2637 atomic_read(&sctx->bios_in_flight) == 0);
2638 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2639 atomic_inc(&fs_info->scrubs_paused);
2640 wake_up(&fs_info->scrub_pause_wait);
2641 wait_event(sctx->list_wait,
2642 atomic_read(&sctx->workers_pending) == 0);
2644 mutex_lock(&fs_info->scrub_lock);
2645 while (atomic_read(&fs_info->scrub_pause_req)) {
2646 mutex_unlock(&fs_info->scrub_lock);
2647 wait_event(fs_info->scrub_pause_wait,
2648 atomic_read(&fs_info->scrub_pause_req) == 0);
2649 mutex_lock(&fs_info->scrub_lock);
2651 atomic_dec(&fs_info->scrubs_paused);
2652 mutex_unlock(&fs_info->scrub_lock);
2653 wake_up(&fs_info->scrub_pause_wait);
2655 dev_replace->cursor_left = dev_replace->cursor_right;
2656 dev_replace->item_needs_writeback = 1;
2657 btrfs_put_block_group(cache);
2660 if (atomic64_read(&dev_replace->num_write_errors) > 0) {
2664 if (sctx->stat.malloc_errors > 0) {
2669 key.offset = found_key.offset + length;
2670 btrfs_release_path(path);
2673 btrfs_free_path(path);
2676 * ret can still be 1 from search_slot or next_leaf,
2677 * that's not an error
2679 return ret < 0 ? ret : 0;
2682 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2683 struct btrfs_device *scrub_dev)
2689 struct btrfs_root *root = sctx->dev_root;
2691 if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR)
2694 gen = root->fs_info->last_trans_committed;
2696 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2697 bytenr = btrfs_sb_offset(i);
2698 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2701 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2702 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2707 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2713 * get a reference count on fs_info->scrub_workers. start worker if necessary
2715 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2720 mutex_lock(&fs_info->scrub_lock);
2721 if (fs_info->scrub_workers_refcnt == 0) {
2723 btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2724 &fs_info->generic_worker);
2726 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2727 fs_info->thread_pool_size,
2728 &fs_info->generic_worker);
2729 fs_info->scrub_workers.idle_thresh = 4;
2730 ret = btrfs_start_workers(&fs_info->scrub_workers);
2733 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2735 fs_info->thread_pool_size,
2736 &fs_info->generic_worker);
2737 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2738 ret = btrfs_start_workers(
2739 &fs_info->scrub_wr_completion_workers);
2742 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2743 &fs_info->generic_worker);
2744 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2748 ++fs_info->scrub_workers_refcnt;
2750 mutex_unlock(&fs_info->scrub_lock);
2755 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2757 mutex_lock(&fs_info->scrub_lock);
2758 if (--fs_info->scrub_workers_refcnt == 0) {
2759 btrfs_stop_workers(&fs_info->scrub_workers);
2760 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2761 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2763 WARN_ON(fs_info->scrub_workers_refcnt < 0);
2764 mutex_unlock(&fs_info->scrub_lock);
2767 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2768 u64 end, struct btrfs_scrub_progress *progress,
2769 int readonly, int is_dev_replace)
2771 struct scrub_ctx *sctx;
2773 struct btrfs_device *dev;
2775 if (btrfs_fs_closing(fs_info))
2779 * check some assumptions
2781 if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2783 "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2784 fs_info->chunk_root->nodesize,
2785 fs_info->chunk_root->leafsize);
2789 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2791 * in this case scrub is unable to calculate the checksum
2792 * the way scrub is implemented. Do not handle this
2793 * situation at all because it won't ever happen.
2796 "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2797 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2801 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2802 /* not supported for data w/o checksums */
2804 "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2805 fs_info->chunk_root->sectorsize,
2806 (unsigned long long)PAGE_SIZE);
2810 if (fs_info->chunk_root->nodesize >
2811 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2812 fs_info->chunk_root->sectorsize >
2813 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2815 * would exhaust the array bounds of pagev member in
2816 * struct scrub_block
2818 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2819 fs_info->chunk_root->nodesize,
2820 SCRUB_MAX_PAGES_PER_BLOCK,
2821 fs_info->chunk_root->sectorsize,
2822 SCRUB_MAX_PAGES_PER_BLOCK);
2826 ret = scrub_workers_get(fs_info, is_dev_replace);
2830 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2831 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2832 if (!dev || (dev->missing && !is_dev_replace)) {
2833 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2834 scrub_workers_put(fs_info);
2837 mutex_lock(&fs_info->scrub_lock);
2839 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2840 mutex_unlock(&fs_info->scrub_lock);
2841 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2842 scrub_workers_put(fs_info);
2846 btrfs_dev_replace_lock(&fs_info->dev_replace);
2847 if (dev->scrub_device ||
2849 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2850 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2851 mutex_unlock(&fs_info->scrub_lock);
2852 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2853 scrub_workers_put(fs_info);
2854 return -EINPROGRESS;
2856 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2857 sctx = scrub_setup_ctx(dev, is_dev_replace);
2859 mutex_unlock(&fs_info->scrub_lock);
2860 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2861 scrub_workers_put(fs_info);
2862 return PTR_ERR(sctx);
2864 sctx->readonly = readonly;
2865 dev->scrub_device = sctx;
2867 atomic_inc(&fs_info->scrubs_running);
2868 mutex_unlock(&fs_info->scrub_lock);
2869 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2871 if (!is_dev_replace) {
2872 down_read(&fs_info->scrub_super_lock);
2873 ret = scrub_supers(sctx, dev);
2874 up_read(&fs_info->scrub_super_lock);
2878 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2881 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2882 atomic_dec(&fs_info->scrubs_running);
2883 wake_up(&fs_info->scrub_pause_wait);
2885 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2888 memcpy(progress, &sctx->stat, sizeof(*progress));
2890 mutex_lock(&fs_info->scrub_lock);
2891 dev->scrub_device = NULL;
2892 mutex_unlock(&fs_info->scrub_lock);
2894 scrub_free_ctx(sctx);
2895 scrub_workers_put(fs_info);
2900 void btrfs_scrub_pause(struct btrfs_root *root)
2902 struct btrfs_fs_info *fs_info = root->fs_info;
2904 mutex_lock(&fs_info->scrub_lock);
2905 atomic_inc(&fs_info->scrub_pause_req);
2906 while (atomic_read(&fs_info->scrubs_paused) !=
2907 atomic_read(&fs_info->scrubs_running)) {
2908 mutex_unlock(&fs_info->scrub_lock);
2909 wait_event(fs_info->scrub_pause_wait,
2910 atomic_read(&fs_info->scrubs_paused) ==
2911 atomic_read(&fs_info->scrubs_running));
2912 mutex_lock(&fs_info->scrub_lock);
2914 mutex_unlock(&fs_info->scrub_lock);
2917 void btrfs_scrub_continue(struct btrfs_root *root)
2919 struct btrfs_fs_info *fs_info = root->fs_info;
2921 atomic_dec(&fs_info->scrub_pause_req);
2922 wake_up(&fs_info->scrub_pause_wait);
2925 void btrfs_scrub_pause_super(struct btrfs_root *root)
2927 down_write(&root->fs_info->scrub_super_lock);
2930 void btrfs_scrub_continue_super(struct btrfs_root *root)
2932 up_write(&root->fs_info->scrub_super_lock);
2935 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2937 mutex_lock(&fs_info->scrub_lock);
2938 if (!atomic_read(&fs_info->scrubs_running)) {
2939 mutex_unlock(&fs_info->scrub_lock);
2943 atomic_inc(&fs_info->scrub_cancel_req);
2944 while (atomic_read(&fs_info->scrubs_running)) {
2945 mutex_unlock(&fs_info->scrub_lock);
2946 wait_event(fs_info->scrub_pause_wait,
2947 atomic_read(&fs_info->scrubs_running) == 0);
2948 mutex_lock(&fs_info->scrub_lock);
2950 atomic_dec(&fs_info->scrub_cancel_req);
2951 mutex_unlock(&fs_info->scrub_lock);
2956 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2957 struct btrfs_device *dev)
2959 struct scrub_ctx *sctx;
2961 mutex_lock(&fs_info->scrub_lock);
2962 sctx = dev->scrub_device;
2964 mutex_unlock(&fs_info->scrub_lock);
2967 atomic_inc(&sctx->cancel_req);
2968 while (dev->scrub_device) {
2969 mutex_unlock(&fs_info->scrub_lock);
2970 wait_event(fs_info->scrub_pause_wait,
2971 dev->scrub_device == NULL);
2972 mutex_lock(&fs_info->scrub_lock);
2974 mutex_unlock(&fs_info->scrub_lock);
2979 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
2981 struct btrfs_fs_info *fs_info = root->fs_info;
2982 struct btrfs_device *dev;
2986 * we have to hold the device_list_mutex here so the device
2987 * does not go away in cancel_dev. FIXME: find a better solution
2989 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2990 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2992 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2995 ret = btrfs_scrub_cancel_dev(fs_info, dev);
2996 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3001 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3002 struct btrfs_scrub_progress *progress)
3004 struct btrfs_device *dev;
3005 struct scrub_ctx *sctx = NULL;
3007 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3008 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3010 sctx = dev->scrub_device;
3012 memcpy(progress, &sctx->stat, sizeof(*progress));
3013 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3015 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3018 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3019 u64 extent_logical, u64 extent_len,
3020 u64 *extent_physical,
3021 struct btrfs_device **extent_dev,
3022 int *extent_mirror_num)
3025 struct btrfs_bio *bbio = NULL;
3028 mapped_length = extent_len;
3029 ret = btrfs_map_block(fs_info, READ, extent_logical,
3030 &mapped_length, &bbio, 0);
3031 if (ret || !bbio || mapped_length < extent_len ||
3032 !bbio->stripes[0].dev->bdev) {
3037 *extent_physical = bbio->stripes[0].physical;
3038 *extent_mirror_num = bbio->mirror_num;
3039 *extent_dev = bbio->stripes[0].dev;
3043 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3044 struct scrub_wr_ctx *wr_ctx,
3045 struct btrfs_fs_info *fs_info,
3046 struct btrfs_device *dev,
3049 WARN_ON(wr_ctx->wr_curr_bio != NULL);
3051 mutex_init(&wr_ctx->wr_lock);
3052 wr_ctx->wr_curr_bio = NULL;
3053 if (!is_dev_replace)
3056 WARN_ON(!dev->bdev);
3057 wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3058 bio_get_nr_vecs(dev->bdev));
3059 wr_ctx->tgtdev = dev;
3060 atomic_set(&wr_ctx->flush_all_writes, 0);
3064 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3066 mutex_lock(&wr_ctx->wr_lock);
3067 kfree(wr_ctx->wr_curr_bio);
3068 wr_ctx->wr_curr_bio = NULL;
3069 mutex_unlock(&wr_ctx->wr_lock);
3072 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3073 int mirror_num, u64 physical_for_dev_replace)
3075 struct scrub_copy_nocow_ctx *nocow_ctx;
3076 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3078 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3080 spin_lock(&sctx->stat_lock);
3081 sctx->stat.malloc_errors++;
3082 spin_unlock(&sctx->stat_lock);
3086 scrub_pending_trans_workers_inc(sctx);
3088 nocow_ctx->sctx = sctx;
3089 nocow_ctx->logical = logical;
3090 nocow_ctx->len = len;
3091 nocow_ctx->mirror_num = mirror_num;
3092 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3093 nocow_ctx->work.func = copy_nocow_pages_worker;
3094 btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3100 static void copy_nocow_pages_worker(struct btrfs_work *work)
3102 struct scrub_copy_nocow_ctx *nocow_ctx =
3103 container_of(work, struct scrub_copy_nocow_ctx, work);
3104 struct scrub_ctx *sctx = nocow_ctx->sctx;
3105 u64 logical = nocow_ctx->logical;
3106 u64 len = nocow_ctx->len;
3107 int mirror_num = nocow_ctx->mirror_num;
3108 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3110 struct btrfs_trans_handle *trans = NULL;
3111 struct btrfs_fs_info *fs_info;
3112 struct btrfs_path *path;
3113 struct btrfs_root *root;
3114 int not_written = 0;
3116 fs_info = sctx->dev_root->fs_info;
3117 root = fs_info->extent_root;
3119 path = btrfs_alloc_path();
3121 spin_lock(&sctx->stat_lock);
3122 sctx->stat.malloc_errors++;
3123 spin_unlock(&sctx->stat_lock);
3128 trans = btrfs_join_transaction(root);
3129 if (IS_ERR(trans)) {
3134 ret = iterate_inodes_from_logical(logical, fs_info, path,
3135 copy_nocow_pages_for_inode,
3137 if (ret != 0 && ret != -ENOENT) {
3138 pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3139 (unsigned long long)logical,
3140 (unsigned long long)physical_for_dev_replace,
3141 (unsigned long long)len,
3142 (unsigned long long)mirror_num, ret);
3148 if (trans && !IS_ERR(trans))
3149 btrfs_end_transaction(trans, root);
3151 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3152 num_uncorrectable_read_errors);
3154 btrfs_free_path(path);
3157 scrub_pending_trans_workers_dec(sctx);
3160 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3162 unsigned long index;
3163 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3165 struct btrfs_key key;
3166 struct inode *inode = NULL;
3167 struct btrfs_root *local_root;
3168 u64 physical_for_dev_replace;
3170 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3172 key.objectid = root;
3173 key.type = BTRFS_ROOT_ITEM_KEY;
3174 key.offset = (u64)-1;
3175 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3176 if (IS_ERR(local_root))
3177 return PTR_ERR(local_root);
3179 key.type = BTRFS_INODE_ITEM_KEY;
3180 key.objectid = inum;
3182 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3184 return PTR_ERR(inode);
3186 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3187 len = nocow_ctx->len;
3188 while (len >= PAGE_CACHE_SIZE) {
3189 struct page *page = NULL;
3192 index = offset >> PAGE_CACHE_SHIFT;
3194 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3196 pr_err("find_or_create_page() failed\n");
3201 if (PageUptodate(page)) {
3202 if (PageDirty(page))
3205 ClearPageError(page);
3206 ret_sub = extent_read_full_page(&BTRFS_I(inode)->
3208 page, btrfs_get_extent,
3209 nocow_ctx->mirror_num);
3214 wait_on_page_locked(page);
3215 if (!PageUptodate(page)) {
3220 ret_sub = write_page_nocow(nocow_ctx->sctx,
3221 physical_for_dev_replace, page);
3232 offset += PAGE_CACHE_SIZE;
3233 physical_for_dev_replace += PAGE_CACHE_SIZE;
3234 len -= PAGE_CACHE_SIZE;
3242 static int write_page_nocow(struct scrub_ctx *sctx,
3243 u64 physical_for_dev_replace, struct page *page)
3246 struct btrfs_device *dev;
3248 DECLARE_COMPLETION_ONSTACK(compl);
3250 dev = sctx->wr_ctx.tgtdev;
3254 printk_ratelimited(KERN_WARNING
3255 "btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3258 bio = bio_alloc(GFP_NOFS, 1);
3260 spin_lock(&sctx->stat_lock);
3261 sctx->stat.malloc_errors++;
3262 spin_unlock(&sctx->stat_lock);
3265 bio->bi_private = &compl;
3266 bio->bi_end_io = scrub_complete_bio_end_io;
3268 bio->bi_sector = physical_for_dev_replace >> 9;
3269 bio->bi_bdev = dev->bdev;
3270 ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3271 if (ret != PAGE_CACHE_SIZE) {
3274 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3277 btrfsic_submit_bio(WRITE_SYNC, bio);
3278 wait_for_completion(&compl);
3280 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3281 goto leave_with_eio;