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 */
67 struct scrub_block *sblock;
69 struct btrfs_device *dev;
70 u64 flags; /* extent flags */
74 u64 physical_for_dev_replace;
77 unsigned int mirror_num:8;
78 unsigned int have_csum:1;
79 unsigned int io_error:1;
81 u8 csum[BTRFS_CSUM_SIZE];
86 struct scrub_ctx *sctx;
87 struct btrfs_device *dev;
92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
93 struct scrub_page *pagev[SCRUB_PAGES_PER_WR_BIO];
95 struct scrub_page *pagev[SCRUB_PAGES_PER_RD_BIO];
99 struct btrfs_work work;
103 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
105 atomic_t outstanding_pages;
106 atomic_t ref_count; /* free mem on transition to zero */
107 struct scrub_ctx *sctx;
109 unsigned int header_error:1;
110 unsigned int checksum_error:1;
111 unsigned int no_io_error_seen:1;
112 unsigned int generation_error:1; /* also sets header_error */
116 struct scrub_wr_ctx {
117 struct scrub_bio *wr_curr_bio;
118 struct btrfs_device *tgtdev;
119 int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
120 atomic_t flush_all_writes;
121 struct mutex wr_lock;
125 struct scrub_bio *bios[SCRUB_BIOS_PER_SCTX];
126 struct btrfs_root *dev_root;
129 atomic_t bios_in_flight;
130 atomic_t workers_pending;
131 spinlock_t list_lock;
132 wait_queue_head_t list_wait;
134 struct list_head csum_list;
137 int pages_per_rd_bio;
143 struct scrub_wr_ctx wr_ctx;
148 struct btrfs_scrub_progress stat;
149 spinlock_t stat_lock;
152 struct scrub_fixup_nodatasum {
153 struct scrub_ctx *sctx;
154 struct btrfs_device *dev;
156 struct btrfs_root *root;
157 struct btrfs_work work;
161 struct scrub_nocow_inode {
165 struct list_head list;
168 struct scrub_copy_nocow_ctx {
169 struct scrub_ctx *sctx;
173 u64 physical_for_dev_replace;
174 struct list_head inodes;
175 struct btrfs_work work;
178 struct scrub_warning {
179 struct btrfs_path *path;
180 u64 extent_item_size;
186 struct btrfs_device *dev;
192 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
193 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
194 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
195 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
196 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
197 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
198 struct btrfs_fs_info *fs_info,
199 struct scrub_block *original_sblock,
200 u64 length, u64 logical,
201 struct scrub_block *sblocks_for_recheck);
202 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
203 struct scrub_block *sblock, int is_metadata,
204 int have_csum, u8 *csum, u64 generation,
206 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
207 struct scrub_block *sblock,
208 int is_metadata, int have_csum,
209 const u8 *csum, u64 generation,
211 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
212 struct scrub_block *sblock_good,
214 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
215 struct scrub_block *sblock_good,
216 int page_num, int force_write);
217 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
218 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
220 static int scrub_checksum_data(struct scrub_block *sblock);
221 static int scrub_checksum_tree_block(struct scrub_block *sblock);
222 static int scrub_checksum_super(struct scrub_block *sblock);
223 static void scrub_block_get(struct scrub_block *sblock);
224 static void scrub_block_put(struct scrub_block *sblock);
225 static void scrub_page_get(struct scrub_page *spage);
226 static void scrub_page_put(struct scrub_page *spage);
227 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
228 struct scrub_page *spage);
229 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
230 u64 physical, struct btrfs_device *dev, u64 flags,
231 u64 gen, int mirror_num, u8 *csum, int force,
232 u64 physical_for_dev_replace);
233 static void scrub_bio_end_io(struct bio *bio, int err);
234 static void scrub_bio_end_io_worker(struct btrfs_work *work);
235 static void scrub_block_complete(struct scrub_block *sblock);
236 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
237 u64 extent_logical, u64 extent_len,
238 u64 *extent_physical,
239 struct btrfs_device **extent_dev,
240 int *extent_mirror_num);
241 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
242 struct scrub_wr_ctx *wr_ctx,
243 struct btrfs_fs_info *fs_info,
244 struct btrfs_device *dev,
246 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
248 struct scrub_page *spage);
249 static void scrub_wr_submit(struct scrub_ctx *sctx);
250 static void scrub_wr_bio_end_io(struct bio *bio, int err);
251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
252 static int write_page_nocow(struct scrub_ctx *sctx,
253 u64 physical_for_dev_replace, struct page *page);
254 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
255 struct scrub_copy_nocow_ctx *ctx);
256 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
257 int mirror_num, u64 physical_for_dev_replace);
258 static void copy_nocow_pages_worker(struct btrfs_work *work);
259 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
260 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
263 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
265 atomic_inc(&sctx->bios_in_flight);
268 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
270 atomic_dec(&sctx->bios_in_flight);
271 wake_up(&sctx->list_wait);
274 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
276 while (atomic_read(&fs_info->scrub_pause_req)) {
277 mutex_unlock(&fs_info->scrub_lock);
278 wait_event(fs_info->scrub_pause_wait,
279 atomic_read(&fs_info->scrub_pause_req) == 0);
280 mutex_lock(&fs_info->scrub_lock);
284 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
286 atomic_inc(&fs_info->scrubs_paused);
287 wake_up(&fs_info->scrub_pause_wait);
289 mutex_lock(&fs_info->scrub_lock);
290 __scrub_blocked_if_needed(fs_info);
291 atomic_dec(&fs_info->scrubs_paused);
292 mutex_unlock(&fs_info->scrub_lock);
294 wake_up(&fs_info->scrub_pause_wait);
298 * used for workers that require transaction commits (i.e., for the
301 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
303 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
306 * increment scrubs_running to prevent cancel requests from
307 * completing as long as a worker is running. we must also
308 * increment scrubs_paused to prevent deadlocking on pause
309 * requests used for transactions commits (as the worker uses a
310 * transaction context). it is safe to regard the worker
311 * as paused for all matters practical. effectively, we only
312 * avoid cancellation requests from completing.
314 mutex_lock(&fs_info->scrub_lock);
315 atomic_inc(&fs_info->scrubs_running);
316 atomic_inc(&fs_info->scrubs_paused);
317 mutex_unlock(&fs_info->scrub_lock);
320 * check if @scrubs_running=@scrubs_paused condition
321 * inside wait_event() is not an atomic operation.
322 * which means we may inc/dec @scrub_running/paused
323 * at any time. Let's wake up @scrub_pause_wait as
324 * much as we can to let commit transaction blocked less.
326 wake_up(&fs_info->scrub_pause_wait);
328 atomic_inc(&sctx->workers_pending);
331 /* used for workers that require transaction commits */
332 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
334 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
337 * see scrub_pending_trans_workers_inc() why we're pretending
338 * to be paused in the scrub counters
340 mutex_lock(&fs_info->scrub_lock);
341 atomic_dec(&fs_info->scrubs_running);
342 atomic_dec(&fs_info->scrubs_paused);
343 mutex_unlock(&fs_info->scrub_lock);
344 atomic_dec(&sctx->workers_pending);
345 wake_up(&fs_info->scrub_pause_wait);
346 wake_up(&sctx->list_wait);
349 static void scrub_free_csums(struct scrub_ctx *sctx)
351 while (!list_empty(&sctx->csum_list)) {
352 struct btrfs_ordered_sum *sum;
353 sum = list_first_entry(&sctx->csum_list,
354 struct btrfs_ordered_sum, list);
355 list_del(&sum->list);
360 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
367 scrub_free_wr_ctx(&sctx->wr_ctx);
369 /* this can happen when scrub is cancelled */
370 if (sctx->curr != -1) {
371 struct scrub_bio *sbio = sctx->bios[sctx->curr];
373 for (i = 0; i < sbio->page_count; i++) {
374 WARN_ON(!sbio->pagev[i]->page);
375 scrub_block_put(sbio->pagev[i]->sblock);
380 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
381 struct scrub_bio *sbio = sctx->bios[i];
388 scrub_free_csums(sctx);
392 static noinline_for_stack
393 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
395 struct scrub_ctx *sctx;
397 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
398 int pages_per_rd_bio;
402 * the setting of pages_per_rd_bio is correct for scrub but might
403 * be wrong for the dev_replace code where we might read from
404 * different devices in the initial huge bios. However, that
405 * code is able to correctly handle the case when adding a page
409 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
410 bio_get_nr_vecs(dev->bdev));
412 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
413 sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
416 sctx->is_dev_replace = is_dev_replace;
417 sctx->pages_per_rd_bio = pages_per_rd_bio;
419 sctx->dev_root = dev->dev_root;
420 for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
421 struct scrub_bio *sbio;
423 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
426 sctx->bios[i] = sbio;
430 sbio->page_count = 0;
431 sbio->work.func = scrub_bio_end_io_worker;
433 if (i != SCRUB_BIOS_PER_SCTX - 1)
434 sctx->bios[i]->next_free = i + 1;
436 sctx->bios[i]->next_free = -1;
438 sctx->first_free = 0;
439 sctx->nodesize = dev->dev_root->nodesize;
440 sctx->leafsize = dev->dev_root->leafsize;
441 sctx->sectorsize = dev->dev_root->sectorsize;
442 atomic_set(&sctx->bios_in_flight, 0);
443 atomic_set(&sctx->workers_pending, 0);
444 atomic_set(&sctx->cancel_req, 0);
445 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
446 INIT_LIST_HEAD(&sctx->csum_list);
448 spin_lock_init(&sctx->list_lock);
449 spin_lock_init(&sctx->stat_lock);
450 init_waitqueue_head(&sctx->list_wait);
452 ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
453 fs_info->dev_replace.tgtdev, is_dev_replace);
455 scrub_free_ctx(sctx);
461 scrub_free_ctx(sctx);
462 return ERR_PTR(-ENOMEM);
465 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
472 struct extent_buffer *eb;
473 struct btrfs_inode_item *inode_item;
474 struct scrub_warning *swarn = warn_ctx;
475 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
476 struct inode_fs_paths *ipath = NULL;
477 struct btrfs_root *local_root;
478 struct btrfs_key root_key;
480 root_key.objectid = root;
481 root_key.type = BTRFS_ROOT_ITEM_KEY;
482 root_key.offset = (u64)-1;
483 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
484 if (IS_ERR(local_root)) {
485 ret = PTR_ERR(local_root);
489 ret = inode_item_info(inum, 0, local_root, swarn->path);
491 btrfs_release_path(swarn->path);
495 eb = swarn->path->nodes[0];
496 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
497 struct btrfs_inode_item);
498 isize = btrfs_inode_size(eb, inode_item);
499 nlink = btrfs_inode_nlink(eb, inode_item);
500 btrfs_release_path(swarn->path);
502 ipath = init_ipath(4096, local_root, swarn->path);
504 ret = PTR_ERR(ipath);
508 ret = paths_from_inode(inum, ipath);
514 * we deliberately ignore the bit ipath might have been too small to
515 * hold all of the paths here
517 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
518 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
519 "%s, sector %llu, root %llu, inode %llu, offset %llu, "
520 "length %llu, links %u (path: %s)\n", swarn->errstr,
521 swarn->logical, rcu_str_deref(swarn->dev->name),
522 (unsigned long long)swarn->sector, root, inum, offset,
523 min(isize - offset, (u64)PAGE_SIZE), nlink,
524 (char *)(unsigned long)ipath->fspath->val[i]);
530 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
531 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
532 "resolving failed with ret=%d\n", swarn->errstr,
533 swarn->logical, rcu_str_deref(swarn->dev->name),
534 (unsigned long long)swarn->sector, root, inum, offset, ret);
540 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
542 struct btrfs_device *dev;
543 struct btrfs_fs_info *fs_info;
544 struct btrfs_path *path;
545 struct btrfs_key found_key;
546 struct extent_buffer *eb;
547 struct btrfs_extent_item *ei;
548 struct scrub_warning swarn;
549 unsigned long ptr = 0;
555 const int bufsize = 4096;
558 WARN_ON(sblock->page_count < 1);
559 dev = sblock->pagev[0]->dev;
560 fs_info = sblock->sctx->dev_root->fs_info;
562 path = btrfs_alloc_path();
564 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
565 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
566 swarn.sector = (sblock->pagev[0]->physical) >> 9;
567 swarn.logical = sblock->pagev[0]->logical;
568 swarn.errstr = errstr;
570 swarn.msg_bufsize = bufsize;
571 swarn.scratch_bufsize = bufsize;
573 if (!path || !swarn.scratch_buf || !swarn.msg_buf)
576 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
581 extent_item_pos = swarn.logical - found_key.objectid;
582 swarn.extent_item_size = found_key.offset;
585 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
586 item_size = btrfs_item_size_nr(eb, path->slots[0]);
588 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
590 ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
591 &ref_root, &ref_level);
592 printk_in_rcu(KERN_WARNING
593 "BTRFS: %s at logical %llu on dev %s, "
594 "sector %llu: metadata %s (level %d) in tree "
595 "%llu\n", errstr, swarn.logical,
596 rcu_str_deref(dev->name),
597 (unsigned long long)swarn.sector,
598 ref_level ? "node" : "leaf",
599 ret < 0 ? -1 : ref_level,
600 ret < 0 ? -1 : ref_root);
602 btrfs_release_path(path);
604 btrfs_release_path(path);
607 iterate_extent_inodes(fs_info, found_key.objectid,
609 scrub_print_warning_inode, &swarn);
613 btrfs_free_path(path);
614 kfree(swarn.scratch_buf);
615 kfree(swarn.msg_buf);
618 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
620 struct page *page = NULL;
622 struct scrub_fixup_nodatasum *fixup = fixup_ctx;
625 struct btrfs_key key;
626 struct inode *inode = NULL;
627 struct btrfs_fs_info *fs_info;
628 u64 end = offset + PAGE_SIZE - 1;
629 struct btrfs_root *local_root;
633 key.type = BTRFS_ROOT_ITEM_KEY;
634 key.offset = (u64)-1;
636 fs_info = fixup->root->fs_info;
637 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
639 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
640 if (IS_ERR(local_root)) {
641 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
642 return PTR_ERR(local_root);
645 key.type = BTRFS_INODE_ITEM_KEY;
648 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
649 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
651 return PTR_ERR(inode);
653 index = offset >> PAGE_CACHE_SHIFT;
655 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
661 if (PageUptodate(page)) {
662 if (PageDirty(page)) {
664 * we need to write the data to the defect sector. the
665 * data that was in that sector is not in memory,
666 * because the page was modified. we must not write the
667 * modified page to that sector.
669 * TODO: what could be done here: wait for the delalloc
670 * runner to write out that page (might involve
671 * COW) and see whether the sector is still
672 * referenced afterwards.
674 * For the meantime, we'll treat this error
675 * incorrectable, although there is a chance that a
676 * later scrub will find the bad sector again and that
677 * there's no dirty page in memory, then.
682 fs_info = BTRFS_I(inode)->root->fs_info;
683 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
684 fixup->logical, page,
690 * we need to get good data first. the general readpage path
691 * will call repair_io_failure for us, we just have to make
692 * sure we read the bad mirror.
694 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
695 EXTENT_DAMAGED, GFP_NOFS);
697 /* set_extent_bits should give proper error */
704 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
707 wait_on_page_locked(page);
709 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
710 end, EXTENT_DAMAGED, 0, NULL);
712 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
713 EXTENT_DAMAGED, GFP_NOFS);
725 if (ret == 0 && corrected) {
727 * we only need to call readpage for one of the inodes belonging
728 * to this extent. so make iterate_extent_inodes stop
736 static void scrub_fixup_nodatasum(struct btrfs_work *work)
739 struct scrub_fixup_nodatasum *fixup;
740 struct scrub_ctx *sctx;
741 struct btrfs_trans_handle *trans = NULL;
742 struct btrfs_path *path;
743 int uncorrectable = 0;
745 fixup = container_of(work, struct scrub_fixup_nodatasum, work);
748 path = btrfs_alloc_path();
750 spin_lock(&sctx->stat_lock);
751 ++sctx->stat.malloc_errors;
752 spin_unlock(&sctx->stat_lock);
757 trans = btrfs_join_transaction(fixup->root);
764 * the idea is to trigger a regular read through the standard path. we
765 * read a page from the (failed) logical address by specifying the
766 * corresponding copynum of the failed sector. thus, that readpage is
768 * that is the point where on-the-fly error correction will kick in
769 * (once it's finished) and rewrite the failed sector if a good copy
772 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
773 path, scrub_fixup_readpage,
781 spin_lock(&sctx->stat_lock);
782 ++sctx->stat.corrected_errors;
783 spin_unlock(&sctx->stat_lock);
786 if (trans && !IS_ERR(trans))
787 btrfs_end_transaction(trans, fixup->root);
789 spin_lock(&sctx->stat_lock);
790 ++sctx->stat.uncorrectable_errors;
791 spin_unlock(&sctx->stat_lock);
792 btrfs_dev_replace_stats_inc(
793 &sctx->dev_root->fs_info->dev_replace.
794 num_uncorrectable_read_errors);
795 printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
796 "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
797 fixup->logical, rcu_str_deref(fixup->dev->name));
800 btrfs_free_path(path);
803 scrub_pending_trans_workers_dec(sctx);
807 * scrub_handle_errored_block gets called when either verification of the
808 * pages failed or the bio failed to read, e.g. with EIO. In the latter
809 * case, this function handles all pages in the bio, even though only one
811 * The goal of this function is to repair the errored block by using the
812 * contents of one of the mirrors.
814 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
816 struct scrub_ctx *sctx = sblock_to_check->sctx;
817 struct btrfs_device *dev;
818 struct btrfs_fs_info *fs_info;
822 unsigned int failed_mirror_index;
823 unsigned int is_metadata;
824 unsigned int have_csum;
826 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
827 struct scrub_block *sblock_bad;
832 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
833 DEFAULT_RATELIMIT_BURST);
835 BUG_ON(sblock_to_check->page_count < 1);
836 fs_info = sctx->dev_root->fs_info;
837 if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
839 * if we find an error in a super block, we just report it.
840 * They will get written with the next transaction commit
843 spin_lock(&sctx->stat_lock);
844 ++sctx->stat.super_errors;
845 spin_unlock(&sctx->stat_lock);
848 length = sblock_to_check->page_count * PAGE_SIZE;
849 logical = sblock_to_check->pagev[0]->logical;
850 generation = sblock_to_check->pagev[0]->generation;
851 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
852 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
853 is_metadata = !(sblock_to_check->pagev[0]->flags &
854 BTRFS_EXTENT_FLAG_DATA);
855 have_csum = sblock_to_check->pagev[0]->have_csum;
856 csum = sblock_to_check->pagev[0]->csum;
857 dev = sblock_to_check->pagev[0]->dev;
859 if (sctx->is_dev_replace && !is_metadata && !have_csum) {
860 sblocks_for_recheck = NULL;
865 * read all mirrors one after the other. This includes to
866 * re-read the extent or metadata block that failed (that was
867 * the cause that this fixup code is called) another time,
868 * page by page this time in order to know which pages
869 * caused I/O errors and which ones are good (for all mirrors).
870 * It is the goal to handle the situation when more than one
871 * mirror contains I/O errors, but the errors do not
872 * overlap, i.e. the data can be repaired by selecting the
873 * pages from those mirrors without I/O error on the
874 * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
875 * would be that mirror #1 has an I/O error on the first page,
876 * the second page is good, and mirror #2 has an I/O error on
877 * the second page, but the first page is good.
878 * Then the first page of the first mirror can be repaired by
879 * taking the first page of the second mirror, and the
880 * second page of the second mirror can be repaired by
881 * copying the contents of the 2nd page of the 1st mirror.
882 * One more note: if the pages of one mirror contain I/O
883 * errors, the checksum cannot be verified. In order to get
884 * the best data for repairing, the first attempt is to find
885 * a mirror without I/O errors and with a validated checksum.
886 * Only if this is not possible, the pages are picked from
887 * mirrors with I/O errors without considering the checksum.
888 * If the latter is the case, at the end, the checksum of the
889 * repaired area is verified in order to correctly maintain
893 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
894 sizeof(*sblocks_for_recheck),
896 if (!sblocks_for_recheck) {
897 spin_lock(&sctx->stat_lock);
898 sctx->stat.malloc_errors++;
899 sctx->stat.read_errors++;
900 sctx->stat.uncorrectable_errors++;
901 spin_unlock(&sctx->stat_lock);
902 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
906 /* setup the context, map the logical blocks and alloc the pages */
907 ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
908 logical, sblocks_for_recheck);
910 spin_lock(&sctx->stat_lock);
911 sctx->stat.read_errors++;
912 sctx->stat.uncorrectable_errors++;
913 spin_unlock(&sctx->stat_lock);
914 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
917 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
918 sblock_bad = sblocks_for_recheck + failed_mirror_index;
920 /* build and submit the bios for the failed mirror, check checksums */
921 scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
922 csum, generation, sctx->csum_size);
924 if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
925 sblock_bad->no_io_error_seen) {
927 * the error disappeared after reading page by page, or
928 * the area was part of a huge bio and other parts of the
929 * bio caused I/O errors, or the block layer merged several
930 * read requests into one and the error is caused by a
931 * different bio (usually one of the two latter cases is
934 spin_lock(&sctx->stat_lock);
935 sctx->stat.unverified_errors++;
936 spin_unlock(&sctx->stat_lock);
938 if (sctx->is_dev_replace)
939 scrub_write_block_to_dev_replace(sblock_bad);
943 if (!sblock_bad->no_io_error_seen) {
944 spin_lock(&sctx->stat_lock);
945 sctx->stat.read_errors++;
946 spin_unlock(&sctx->stat_lock);
947 if (__ratelimit(&_rs))
948 scrub_print_warning("i/o error", sblock_to_check);
949 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
950 } else if (sblock_bad->checksum_error) {
951 spin_lock(&sctx->stat_lock);
952 sctx->stat.csum_errors++;
953 spin_unlock(&sctx->stat_lock);
954 if (__ratelimit(&_rs))
955 scrub_print_warning("checksum error", sblock_to_check);
956 btrfs_dev_stat_inc_and_print(dev,
957 BTRFS_DEV_STAT_CORRUPTION_ERRS);
958 } else if (sblock_bad->header_error) {
959 spin_lock(&sctx->stat_lock);
960 sctx->stat.verify_errors++;
961 spin_unlock(&sctx->stat_lock);
962 if (__ratelimit(&_rs))
963 scrub_print_warning("checksum/header error",
965 if (sblock_bad->generation_error)
966 btrfs_dev_stat_inc_and_print(dev,
967 BTRFS_DEV_STAT_GENERATION_ERRS);
969 btrfs_dev_stat_inc_and_print(dev,
970 BTRFS_DEV_STAT_CORRUPTION_ERRS);
973 if (sctx->readonly) {
974 ASSERT(!sctx->is_dev_replace);
978 if (!is_metadata && !have_csum) {
979 struct scrub_fixup_nodatasum *fixup_nodatasum;
982 WARN_ON(sctx->is_dev_replace);
985 * !is_metadata and !have_csum, this means that the data
986 * might not be COW'ed, that it might be modified
987 * concurrently. The general strategy to work on the
988 * commit root does not help in the case when COW is not
991 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
992 if (!fixup_nodatasum)
993 goto did_not_correct_error;
994 fixup_nodatasum->sctx = sctx;
995 fixup_nodatasum->dev = dev;
996 fixup_nodatasum->logical = logical;
997 fixup_nodatasum->root = fs_info->extent_root;
998 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
999 scrub_pending_trans_workers_inc(sctx);
1000 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
1001 btrfs_queue_worker(&fs_info->scrub_workers,
1002 &fixup_nodatasum->work);
1007 * now build and submit the bios for the other mirrors, check
1009 * First try to pick the mirror which is completely without I/O
1010 * errors and also does not have a checksum error.
1011 * If one is found, and if a checksum is present, the full block
1012 * that is known to contain an error is rewritten. Afterwards
1013 * the block is known to be corrected.
1014 * If a mirror is found which is completely correct, and no
1015 * checksum is present, only those pages are rewritten that had
1016 * an I/O error in the block to be repaired, since it cannot be
1017 * determined, which copy of the other pages is better (and it
1018 * could happen otherwise that a correct page would be
1019 * overwritten by a bad one).
1021 for (mirror_index = 0;
1022 mirror_index < BTRFS_MAX_MIRRORS &&
1023 sblocks_for_recheck[mirror_index].page_count > 0;
1025 struct scrub_block *sblock_other;
1027 if (mirror_index == failed_mirror_index)
1029 sblock_other = sblocks_for_recheck + mirror_index;
1031 /* build and submit the bios, check checksums */
1032 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1033 have_csum, csum, generation,
1036 if (!sblock_other->header_error &&
1037 !sblock_other->checksum_error &&
1038 sblock_other->no_io_error_seen) {
1039 if (sctx->is_dev_replace) {
1040 scrub_write_block_to_dev_replace(sblock_other);
1042 int force_write = is_metadata || have_csum;
1044 ret = scrub_repair_block_from_good_copy(
1045 sblock_bad, sblock_other,
1049 goto corrected_error;
1054 * for dev_replace, pick good pages and write to the target device.
1056 if (sctx->is_dev_replace) {
1058 for (page_num = 0; page_num < sblock_bad->page_count;
1063 for (mirror_index = 0;
1064 mirror_index < BTRFS_MAX_MIRRORS &&
1065 sblocks_for_recheck[mirror_index].page_count > 0;
1067 struct scrub_block *sblock_other =
1068 sblocks_for_recheck + mirror_index;
1069 struct scrub_page *page_other =
1070 sblock_other->pagev[page_num];
1072 if (!page_other->io_error) {
1073 ret = scrub_write_page_to_dev_replace(
1074 sblock_other, page_num);
1076 /* succeeded for this page */
1080 btrfs_dev_replace_stats_inc(
1082 fs_info->dev_replace.
1090 * did not find a mirror to fetch the page
1091 * from. scrub_write_page_to_dev_replace()
1092 * handles this case (page->io_error), by
1093 * filling the block with zeros before
1094 * submitting the write request
1097 ret = scrub_write_page_to_dev_replace(
1098 sblock_bad, page_num);
1100 btrfs_dev_replace_stats_inc(
1101 &sctx->dev_root->fs_info->
1102 dev_replace.num_write_errors);
1110 * for regular scrub, repair those pages that are errored.
1111 * In case of I/O errors in the area that is supposed to be
1112 * repaired, continue by picking good copies of those pages.
1113 * Select the good pages from mirrors to rewrite bad pages from
1114 * the area to fix. Afterwards verify the checksum of the block
1115 * that is supposed to be repaired. This verification step is
1116 * only done for the purpose of statistic counting and for the
1117 * final scrub report, whether errors remain.
1118 * A perfect algorithm could make use of the checksum and try
1119 * all possible combinations of pages from the different mirrors
1120 * until the checksum verification succeeds. For example, when
1121 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1122 * of mirror #2 is readable but the final checksum test fails,
1123 * then the 2nd page of mirror #3 could be tried, whether now
1124 * the final checksum succeedes. But this would be a rare
1125 * exception and is therefore not implemented. At least it is
1126 * avoided that the good copy is overwritten.
1127 * A more useful improvement would be to pick the sectors
1128 * without I/O error based on sector sizes (512 bytes on legacy
1129 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1130 * mirror could be repaired by taking 512 byte of a different
1131 * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1132 * area are unreadable.
1135 /* can only fix I/O errors from here on */
1136 if (sblock_bad->no_io_error_seen)
1137 goto did_not_correct_error;
1140 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1141 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1143 if (!page_bad->io_error)
1146 for (mirror_index = 0;
1147 mirror_index < BTRFS_MAX_MIRRORS &&
1148 sblocks_for_recheck[mirror_index].page_count > 0;
1150 struct scrub_block *sblock_other = sblocks_for_recheck +
1152 struct scrub_page *page_other = sblock_other->pagev[
1155 if (!page_other->io_error) {
1156 ret = scrub_repair_page_from_good_copy(
1157 sblock_bad, sblock_other, page_num, 0);
1159 page_bad->io_error = 0;
1160 break; /* succeeded for this page */
1165 if (page_bad->io_error) {
1166 /* did not find a mirror to copy the page from */
1172 if (is_metadata || have_csum) {
1174 * need to verify the checksum now that all
1175 * sectors on disk are repaired (the write
1176 * request for data to be repaired is on its way).
1177 * Just be lazy and use scrub_recheck_block()
1178 * which re-reads the data before the checksum
1179 * is verified, but most likely the data comes out
1180 * of the page cache.
1182 scrub_recheck_block(fs_info, sblock_bad,
1183 is_metadata, have_csum, csum,
1184 generation, sctx->csum_size);
1185 if (!sblock_bad->header_error &&
1186 !sblock_bad->checksum_error &&
1187 sblock_bad->no_io_error_seen)
1188 goto corrected_error;
1190 goto did_not_correct_error;
1193 spin_lock(&sctx->stat_lock);
1194 sctx->stat.corrected_errors++;
1195 spin_unlock(&sctx->stat_lock);
1196 printk_ratelimited_in_rcu(KERN_ERR
1197 "BTRFS: fixed up error at logical %llu on dev %s\n",
1198 logical, rcu_str_deref(dev->name));
1201 did_not_correct_error:
1202 spin_lock(&sctx->stat_lock);
1203 sctx->stat.uncorrectable_errors++;
1204 spin_unlock(&sctx->stat_lock);
1205 printk_ratelimited_in_rcu(KERN_ERR
1206 "BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1207 logical, rcu_str_deref(dev->name));
1211 if (sblocks_for_recheck) {
1212 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1214 struct scrub_block *sblock = sblocks_for_recheck +
1218 for (page_index = 0; page_index < sblock->page_count;
1220 sblock->pagev[page_index]->sblock = NULL;
1221 scrub_page_put(sblock->pagev[page_index]);
1224 kfree(sblocks_for_recheck);
1230 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1231 struct btrfs_fs_info *fs_info,
1232 struct scrub_block *original_sblock,
1233 u64 length, u64 logical,
1234 struct scrub_block *sblocks_for_recheck)
1241 * note: the two members ref_count and outstanding_pages
1242 * are not used (and not set) in the blocks that are used for
1243 * the recheck procedure
1247 while (length > 0) {
1248 u64 sublen = min_t(u64, length, PAGE_SIZE);
1249 u64 mapped_length = sublen;
1250 struct btrfs_bio *bbio = NULL;
1253 * with a length of PAGE_SIZE, each returned stripe
1254 * represents one mirror
1256 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1257 &mapped_length, &bbio, 0);
1258 if (ret || !bbio || mapped_length < sublen) {
1263 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1264 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1266 struct scrub_block *sblock;
1267 struct scrub_page *page;
1269 if (mirror_index >= BTRFS_MAX_MIRRORS)
1272 sblock = sblocks_for_recheck + mirror_index;
1273 sblock->sctx = sctx;
1274 page = kzalloc(sizeof(*page), GFP_NOFS);
1277 spin_lock(&sctx->stat_lock);
1278 sctx->stat.malloc_errors++;
1279 spin_unlock(&sctx->stat_lock);
1283 scrub_page_get(page);
1284 sblock->pagev[page_index] = page;
1285 page->logical = logical;
1286 page->physical = bbio->stripes[mirror_index].physical;
1287 BUG_ON(page_index >= original_sblock->page_count);
1288 page->physical_for_dev_replace =
1289 original_sblock->pagev[page_index]->
1290 physical_for_dev_replace;
1291 /* for missing devices, dev->bdev is NULL */
1292 page->dev = bbio->stripes[mirror_index].dev;
1293 page->mirror_num = mirror_index + 1;
1294 sblock->page_count++;
1295 page->page = alloc_page(GFP_NOFS);
1309 * this function will check the on disk data for checksum errors, header
1310 * errors and read I/O errors. If any I/O errors happen, the exact pages
1311 * which are errored are marked as being bad. The goal is to enable scrub
1312 * to take those pages that are not errored from all the mirrors so that
1313 * the pages that are errored in the just handled mirror can be repaired.
1315 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1316 struct scrub_block *sblock, int is_metadata,
1317 int have_csum, u8 *csum, u64 generation,
1322 sblock->no_io_error_seen = 1;
1323 sblock->header_error = 0;
1324 sblock->checksum_error = 0;
1326 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1328 struct scrub_page *page = sblock->pagev[page_num];
1330 if (page->dev->bdev == NULL) {
1332 sblock->no_io_error_seen = 0;
1336 WARN_ON(!page->page);
1337 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1340 sblock->no_io_error_seen = 0;
1343 bio->bi_bdev = page->dev->bdev;
1344 bio->bi_sector = page->physical >> 9;
1346 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1347 if (btrfsic_submit_bio_wait(READ, bio))
1348 sblock->no_io_error_seen = 0;
1353 if (sblock->no_io_error_seen)
1354 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1355 have_csum, csum, generation,
1361 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1362 struct scrub_block *sblock,
1363 int is_metadata, int have_csum,
1364 const u8 *csum, u64 generation,
1368 u8 calculated_csum[BTRFS_CSUM_SIZE];
1370 void *mapped_buffer;
1372 WARN_ON(!sblock->pagev[0]->page);
1374 struct btrfs_header *h;
1376 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1377 h = (struct btrfs_header *)mapped_buffer;
1379 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1380 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1381 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1383 sblock->header_error = 1;
1384 } else if (generation != btrfs_stack_header_generation(h)) {
1385 sblock->header_error = 1;
1386 sblock->generation_error = 1;
1393 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1396 for (page_num = 0;;) {
1397 if (page_num == 0 && is_metadata)
1398 crc = btrfs_csum_data(
1399 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1400 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1402 crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1404 kunmap_atomic(mapped_buffer);
1406 if (page_num >= sblock->page_count)
1408 WARN_ON(!sblock->pagev[page_num]->page);
1410 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1413 btrfs_csum_final(crc, calculated_csum);
1414 if (memcmp(calculated_csum, csum, csum_size))
1415 sblock->checksum_error = 1;
1418 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1419 struct scrub_block *sblock_good,
1425 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1428 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1439 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1440 struct scrub_block *sblock_good,
1441 int page_num, int force_write)
1443 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1444 struct scrub_page *page_good = sblock_good->pagev[page_num];
1446 BUG_ON(page_bad->page == NULL);
1447 BUG_ON(page_good->page == NULL);
1448 if (force_write || sblock_bad->header_error ||
1449 sblock_bad->checksum_error || page_bad->io_error) {
1453 if (!page_bad->dev->bdev) {
1454 printk_ratelimited(KERN_WARNING "BTRFS: "
1455 "scrub_repair_page_from_good_copy(bdev == NULL) "
1456 "is unexpected!\n");
1460 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1463 bio->bi_bdev = page_bad->dev->bdev;
1464 bio->bi_sector = page_bad->physical >> 9;
1466 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1467 if (PAGE_SIZE != ret) {
1472 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1473 btrfs_dev_stat_inc_and_print(page_bad->dev,
1474 BTRFS_DEV_STAT_WRITE_ERRS);
1475 btrfs_dev_replace_stats_inc(
1476 &sblock_bad->sctx->dev_root->fs_info->
1477 dev_replace.num_write_errors);
1487 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1491 for (page_num = 0; page_num < sblock->page_count; page_num++) {
1494 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1496 btrfs_dev_replace_stats_inc(
1497 &sblock->sctx->dev_root->fs_info->dev_replace.
1502 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1505 struct scrub_page *spage = sblock->pagev[page_num];
1507 BUG_ON(spage->page == NULL);
1508 if (spage->io_error) {
1509 void *mapped_buffer = kmap_atomic(spage->page);
1511 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1512 flush_dcache_page(spage->page);
1513 kunmap_atomic(mapped_buffer);
1515 return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1518 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1519 struct scrub_page *spage)
1521 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1522 struct scrub_bio *sbio;
1525 mutex_lock(&wr_ctx->wr_lock);
1527 if (!wr_ctx->wr_curr_bio) {
1528 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1530 if (!wr_ctx->wr_curr_bio) {
1531 mutex_unlock(&wr_ctx->wr_lock);
1534 wr_ctx->wr_curr_bio->sctx = sctx;
1535 wr_ctx->wr_curr_bio->page_count = 0;
1537 sbio = wr_ctx->wr_curr_bio;
1538 if (sbio->page_count == 0) {
1541 sbio->physical = spage->physical_for_dev_replace;
1542 sbio->logical = spage->logical;
1543 sbio->dev = wr_ctx->tgtdev;
1546 bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1548 mutex_unlock(&wr_ctx->wr_lock);
1554 bio->bi_private = sbio;
1555 bio->bi_end_io = scrub_wr_bio_end_io;
1556 bio->bi_bdev = sbio->dev->bdev;
1557 bio->bi_sector = sbio->physical >> 9;
1559 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1560 spage->physical_for_dev_replace ||
1561 sbio->logical + sbio->page_count * PAGE_SIZE !=
1563 scrub_wr_submit(sctx);
1567 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1568 if (ret != PAGE_SIZE) {
1569 if (sbio->page_count < 1) {
1572 mutex_unlock(&wr_ctx->wr_lock);
1575 scrub_wr_submit(sctx);
1579 sbio->pagev[sbio->page_count] = spage;
1580 scrub_page_get(spage);
1582 if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1583 scrub_wr_submit(sctx);
1584 mutex_unlock(&wr_ctx->wr_lock);
1589 static void scrub_wr_submit(struct scrub_ctx *sctx)
1591 struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1592 struct scrub_bio *sbio;
1594 if (!wr_ctx->wr_curr_bio)
1597 sbio = wr_ctx->wr_curr_bio;
1598 wr_ctx->wr_curr_bio = NULL;
1599 WARN_ON(!sbio->bio->bi_bdev);
1600 scrub_pending_bio_inc(sctx);
1601 /* process all writes in a single worker thread. Then the block layer
1602 * orders the requests before sending them to the driver which
1603 * doubled the write performance on spinning disks when measured
1605 btrfsic_submit_bio(WRITE, sbio->bio);
1608 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1610 struct scrub_bio *sbio = bio->bi_private;
1611 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1616 sbio->work.func = scrub_wr_bio_end_io_worker;
1617 btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1620 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1622 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1623 struct scrub_ctx *sctx = sbio->sctx;
1626 WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1628 struct btrfs_dev_replace *dev_replace =
1629 &sbio->sctx->dev_root->fs_info->dev_replace;
1631 for (i = 0; i < sbio->page_count; i++) {
1632 struct scrub_page *spage = sbio->pagev[i];
1634 spage->io_error = 1;
1635 btrfs_dev_replace_stats_inc(&dev_replace->
1640 for (i = 0; i < sbio->page_count; i++)
1641 scrub_page_put(sbio->pagev[i]);
1645 scrub_pending_bio_dec(sctx);
1648 static int scrub_checksum(struct scrub_block *sblock)
1653 WARN_ON(sblock->page_count < 1);
1654 flags = sblock->pagev[0]->flags;
1656 if (flags & BTRFS_EXTENT_FLAG_DATA)
1657 ret = scrub_checksum_data(sblock);
1658 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1659 ret = scrub_checksum_tree_block(sblock);
1660 else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1661 (void)scrub_checksum_super(sblock);
1665 scrub_handle_errored_block(sblock);
1670 static int scrub_checksum_data(struct scrub_block *sblock)
1672 struct scrub_ctx *sctx = sblock->sctx;
1673 u8 csum[BTRFS_CSUM_SIZE];
1682 BUG_ON(sblock->page_count < 1);
1683 if (!sblock->pagev[0]->have_csum)
1686 on_disk_csum = sblock->pagev[0]->csum;
1687 page = sblock->pagev[0]->page;
1688 buffer = kmap_atomic(page);
1690 len = sctx->sectorsize;
1693 u64 l = min_t(u64, len, PAGE_SIZE);
1695 crc = btrfs_csum_data(buffer, crc, l);
1696 kunmap_atomic(buffer);
1701 BUG_ON(index >= sblock->page_count);
1702 BUG_ON(!sblock->pagev[index]->page);
1703 page = sblock->pagev[index]->page;
1704 buffer = kmap_atomic(page);
1707 btrfs_csum_final(crc, csum);
1708 if (memcmp(csum, on_disk_csum, sctx->csum_size))
1714 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1716 struct scrub_ctx *sctx = sblock->sctx;
1717 struct btrfs_header *h;
1718 struct btrfs_root *root = sctx->dev_root;
1719 struct btrfs_fs_info *fs_info = root->fs_info;
1720 u8 calculated_csum[BTRFS_CSUM_SIZE];
1721 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1723 void *mapped_buffer;
1732 BUG_ON(sblock->page_count < 1);
1733 page = sblock->pagev[0]->page;
1734 mapped_buffer = kmap_atomic(page);
1735 h = (struct btrfs_header *)mapped_buffer;
1736 memcpy(on_disk_csum, h->csum, sctx->csum_size);
1739 * we don't use the getter functions here, as we
1740 * a) don't have an extent buffer and
1741 * b) the page is already kmapped
1744 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1747 if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1750 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1753 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1757 WARN_ON(sctx->nodesize != sctx->leafsize);
1758 len = sctx->nodesize - BTRFS_CSUM_SIZE;
1759 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1760 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1763 u64 l = min_t(u64, len, mapped_size);
1765 crc = btrfs_csum_data(p, crc, l);
1766 kunmap_atomic(mapped_buffer);
1771 BUG_ON(index >= sblock->page_count);
1772 BUG_ON(!sblock->pagev[index]->page);
1773 page = sblock->pagev[index]->page;
1774 mapped_buffer = kmap_atomic(page);
1775 mapped_size = PAGE_SIZE;
1779 btrfs_csum_final(crc, calculated_csum);
1780 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1783 return fail || crc_fail;
1786 static int scrub_checksum_super(struct scrub_block *sblock)
1788 struct btrfs_super_block *s;
1789 struct scrub_ctx *sctx = sblock->sctx;
1790 struct btrfs_root *root = sctx->dev_root;
1791 struct btrfs_fs_info *fs_info = root->fs_info;
1792 u8 calculated_csum[BTRFS_CSUM_SIZE];
1793 u8 on_disk_csum[BTRFS_CSUM_SIZE];
1795 void *mapped_buffer;
1804 BUG_ON(sblock->page_count < 1);
1805 page = sblock->pagev[0]->page;
1806 mapped_buffer = kmap_atomic(page);
1807 s = (struct btrfs_super_block *)mapped_buffer;
1808 memcpy(on_disk_csum, s->csum, sctx->csum_size);
1810 if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1813 if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1816 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1819 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1820 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1821 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1824 u64 l = min_t(u64, len, mapped_size);
1826 crc = btrfs_csum_data(p, crc, l);
1827 kunmap_atomic(mapped_buffer);
1832 BUG_ON(index >= sblock->page_count);
1833 BUG_ON(!sblock->pagev[index]->page);
1834 page = sblock->pagev[index]->page;
1835 mapped_buffer = kmap_atomic(page);
1836 mapped_size = PAGE_SIZE;
1840 btrfs_csum_final(crc, calculated_csum);
1841 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1844 if (fail_cor + fail_gen) {
1846 * if we find an error in a super block, we just report it.
1847 * They will get written with the next transaction commit
1850 spin_lock(&sctx->stat_lock);
1851 ++sctx->stat.super_errors;
1852 spin_unlock(&sctx->stat_lock);
1854 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1855 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1857 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1858 BTRFS_DEV_STAT_GENERATION_ERRS);
1861 return fail_cor + fail_gen;
1864 static void scrub_block_get(struct scrub_block *sblock)
1866 atomic_inc(&sblock->ref_count);
1869 static void scrub_block_put(struct scrub_block *sblock)
1871 if (atomic_dec_and_test(&sblock->ref_count)) {
1874 for (i = 0; i < sblock->page_count; i++)
1875 scrub_page_put(sblock->pagev[i]);
1880 static void scrub_page_get(struct scrub_page *spage)
1882 atomic_inc(&spage->ref_count);
1885 static void scrub_page_put(struct scrub_page *spage)
1887 if (atomic_dec_and_test(&spage->ref_count)) {
1889 __free_page(spage->page);
1894 static void scrub_submit(struct scrub_ctx *sctx)
1896 struct scrub_bio *sbio;
1898 if (sctx->curr == -1)
1901 sbio = sctx->bios[sctx->curr];
1903 scrub_pending_bio_inc(sctx);
1905 if (!sbio->bio->bi_bdev) {
1907 * this case should not happen. If btrfs_map_block() is
1908 * wrong, it could happen for dev-replace operations on
1909 * missing devices when no mirrors are available, but in
1910 * this case it should already fail the mount.
1911 * This case is handled correctly (but _very_ slowly).
1913 printk_ratelimited(KERN_WARNING
1914 "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1915 bio_endio(sbio->bio, -EIO);
1917 btrfsic_submit_bio(READ, sbio->bio);
1921 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1922 struct scrub_page *spage)
1924 struct scrub_block *sblock = spage->sblock;
1925 struct scrub_bio *sbio;
1930 * grab a fresh bio or wait for one to become available
1932 while (sctx->curr == -1) {
1933 spin_lock(&sctx->list_lock);
1934 sctx->curr = sctx->first_free;
1935 if (sctx->curr != -1) {
1936 sctx->first_free = sctx->bios[sctx->curr]->next_free;
1937 sctx->bios[sctx->curr]->next_free = -1;
1938 sctx->bios[sctx->curr]->page_count = 0;
1939 spin_unlock(&sctx->list_lock);
1941 spin_unlock(&sctx->list_lock);
1942 wait_event(sctx->list_wait, sctx->first_free != -1);
1945 sbio = sctx->bios[sctx->curr];
1946 if (sbio->page_count == 0) {
1949 sbio->physical = spage->physical;
1950 sbio->logical = spage->logical;
1951 sbio->dev = spage->dev;
1954 bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1960 bio->bi_private = sbio;
1961 bio->bi_end_io = scrub_bio_end_io;
1962 bio->bi_bdev = sbio->dev->bdev;
1963 bio->bi_sector = sbio->physical >> 9;
1965 } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1967 sbio->logical + sbio->page_count * PAGE_SIZE !=
1969 sbio->dev != spage->dev) {
1974 sbio->pagev[sbio->page_count] = spage;
1975 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1976 if (ret != PAGE_SIZE) {
1977 if (sbio->page_count < 1) {
1986 scrub_block_get(sblock); /* one for the page added to the bio */
1987 atomic_inc(&sblock->outstanding_pages);
1989 if (sbio->page_count == sctx->pages_per_rd_bio)
1995 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1996 u64 physical, struct btrfs_device *dev, u64 flags,
1997 u64 gen, int mirror_num, u8 *csum, int force,
1998 u64 physical_for_dev_replace)
2000 struct scrub_block *sblock;
2003 sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2005 spin_lock(&sctx->stat_lock);
2006 sctx->stat.malloc_errors++;
2007 spin_unlock(&sctx->stat_lock);
2011 /* one ref inside this function, plus one for each page added to
2013 atomic_set(&sblock->ref_count, 1);
2014 sblock->sctx = sctx;
2015 sblock->no_io_error_seen = 1;
2017 for (index = 0; len > 0; index++) {
2018 struct scrub_page *spage;
2019 u64 l = min_t(u64, len, PAGE_SIZE);
2021 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2024 spin_lock(&sctx->stat_lock);
2025 sctx->stat.malloc_errors++;
2026 spin_unlock(&sctx->stat_lock);
2027 scrub_block_put(sblock);
2030 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2031 scrub_page_get(spage);
2032 sblock->pagev[index] = spage;
2033 spage->sblock = sblock;
2035 spage->flags = flags;
2036 spage->generation = gen;
2037 spage->logical = logical;
2038 spage->physical = physical;
2039 spage->physical_for_dev_replace = physical_for_dev_replace;
2040 spage->mirror_num = mirror_num;
2042 spage->have_csum = 1;
2043 memcpy(spage->csum, csum, sctx->csum_size);
2045 spage->have_csum = 0;
2047 sblock->page_count++;
2048 spage->page = alloc_page(GFP_NOFS);
2054 physical_for_dev_replace += l;
2057 WARN_ON(sblock->page_count == 0);
2058 for (index = 0; index < sblock->page_count; index++) {
2059 struct scrub_page *spage = sblock->pagev[index];
2062 ret = scrub_add_page_to_rd_bio(sctx, spage);
2064 scrub_block_put(sblock);
2072 /* last one frees, either here or in bio completion for last page */
2073 scrub_block_put(sblock);
2077 static void scrub_bio_end_io(struct bio *bio, int err)
2079 struct scrub_bio *sbio = bio->bi_private;
2080 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2085 btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2088 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2090 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2091 struct scrub_ctx *sctx = sbio->sctx;
2094 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2096 for (i = 0; i < sbio->page_count; i++) {
2097 struct scrub_page *spage = sbio->pagev[i];
2099 spage->io_error = 1;
2100 spage->sblock->no_io_error_seen = 0;
2104 /* now complete the scrub_block items that have all pages completed */
2105 for (i = 0; i < sbio->page_count; i++) {
2106 struct scrub_page *spage = sbio->pagev[i];
2107 struct scrub_block *sblock = spage->sblock;
2109 if (atomic_dec_and_test(&sblock->outstanding_pages))
2110 scrub_block_complete(sblock);
2111 scrub_block_put(sblock);
2116 spin_lock(&sctx->list_lock);
2117 sbio->next_free = sctx->first_free;
2118 sctx->first_free = sbio->index;
2119 spin_unlock(&sctx->list_lock);
2121 if (sctx->is_dev_replace &&
2122 atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2123 mutex_lock(&sctx->wr_ctx.wr_lock);
2124 scrub_wr_submit(sctx);
2125 mutex_unlock(&sctx->wr_ctx.wr_lock);
2128 scrub_pending_bio_dec(sctx);
2131 static void scrub_block_complete(struct scrub_block *sblock)
2133 if (!sblock->no_io_error_seen) {
2134 scrub_handle_errored_block(sblock);
2137 * if has checksum error, write via repair mechanism in
2138 * dev replace case, otherwise write here in dev replace
2141 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2142 scrub_write_block_to_dev_replace(sblock);
2146 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2149 struct btrfs_ordered_sum *sum = NULL;
2150 unsigned long index;
2151 unsigned long num_sectors;
2153 while (!list_empty(&sctx->csum_list)) {
2154 sum = list_first_entry(&sctx->csum_list,
2155 struct btrfs_ordered_sum, list);
2156 if (sum->bytenr > logical)
2158 if (sum->bytenr + sum->len > logical)
2161 ++sctx->stat.csum_discards;
2162 list_del(&sum->list);
2169 index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2170 num_sectors = sum->len / sctx->sectorsize;
2171 memcpy(csum, sum->sums + index, sctx->csum_size);
2172 if (index == num_sectors - 1) {
2173 list_del(&sum->list);
2179 /* scrub extent tries to collect up to 64 kB for each bio */
2180 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2181 u64 physical, struct btrfs_device *dev, u64 flags,
2182 u64 gen, int mirror_num, u64 physical_for_dev_replace)
2185 u8 csum[BTRFS_CSUM_SIZE];
2188 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2189 blocksize = sctx->sectorsize;
2190 spin_lock(&sctx->stat_lock);
2191 sctx->stat.data_extents_scrubbed++;
2192 sctx->stat.data_bytes_scrubbed += len;
2193 spin_unlock(&sctx->stat_lock);
2194 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2195 WARN_ON(sctx->nodesize != sctx->leafsize);
2196 blocksize = sctx->nodesize;
2197 spin_lock(&sctx->stat_lock);
2198 sctx->stat.tree_extents_scrubbed++;
2199 sctx->stat.tree_bytes_scrubbed += len;
2200 spin_unlock(&sctx->stat_lock);
2202 blocksize = sctx->sectorsize;
2207 u64 l = min_t(u64, len, blocksize);
2210 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2211 /* push csums to sbio */
2212 have_csum = scrub_find_csum(sctx, logical, l, csum);
2214 ++sctx->stat.no_csum;
2215 if (sctx->is_dev_replace && !have_csum) {
2216 ret = copy_nocow_pages(sctx, logical, l,
2218 physical_for_dev_replace);
2219 goto behind_scrub_pages;
2222 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2223 mirror_num, have_csum ? csum : NULL, 0,
2224 physical_for_dev_replace);
2231 physical_for_dev_replace += l;
2236 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2237 struct map_lookup *map,
2238 struct btrfs_device *scrub_dev,
2239 int num, u64 base, u64 length,
2242 struct btrfs_path *path;
2243 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2244 struct btrfs_root *root = fs_info->extent_root;
2245 struct btrfs_root *csum_root = fs_info->csum_root;
2246 struct btrfs_extent_item *extent;
2247 struct blk_plug plug;
2252 struct extent_buffer *l;
2253 struct btrfs_key key;
2259 struct reada_control *reada1;
2260 struct reada_control *reada2;
2261 struct btrfs_key key_start;
2262 struct btrfs_key key_end;
2263 u64 increment = map->stripe_len;
2266 u64 extent_physical;
2268 struct btrfs_device *extent_dev;
2269 int extent_mirror_num;
2272 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2273 BTRFS_BLOCK_GROUP_RAID6)) {
2274 if (num >= nr_data_stripes(map)) {
2281 do_div(nstripes, map->stripe_len);
2282 if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2283 offset = map->stripe_len * num;
2284 increment = map->stripe_len * map->num_stripes;
2286 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2287 int factor = map->num_stripes / map->sub_stripes;
2288 offset = map->stripe_len * (num / map->sub_stripes);
2289 increment = map->stripe_len * factor;
2290 mirror_num = num % map->sub_stripes + 1;
2291 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2292 increment = map->stripe_len;
2293 mirror_num = num % map->num_stripes + 1;
2294 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2295 increment = map->stripe_len;
2296 mirror_num = num % map->num_stripes + 1;
2298 increment = map->stripe_len;
2302 path = btrfs_alloc_path();
2307 * work on commit root. The related disk blocks are static as
2308 * long as COW is applied. This means, it is save to rewrite
2309 * them to repair disk errors without any race conditions
2311 path->search_commit_root = 1;
2312 path->skip_locking = 1;
2315 * trigger the readahead for extent tree csum tree and wait for
2316 * completion. During readahead, the scrub is officially paused
2317 * to not hold off transaction commits
2319 logical = base + offset;
2321 wait_event(sctx->list_wait,
2322 atomic_read(&sctx->bios_in_flight) == 0);
2323 scrub_blocked_if_needed(fs_info);
2325 /* FIXME it might be better to start readahead at commit root */
2326 key_start.objectid = logical;
2327 key_start.type = BTRFS_EXTENT_ITEM_KEY;
2328 key_start.offset = (u64)0;
2329 key_end.objectid = base + offset + nstripes * increment;
2330 key_end.type = BTRFS_METADATA_ITEM_KEY;
2331 key_end.offset = (u64)-1;
2332 reada1 = btrfs_reada_add(root, &key_start, &key_end);
2334 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2335 key_start.type = BTRFS_EXTENT_CSUM_KEY;
2336 key_start.offset = logical;
2337 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2338 key_end.type = BTRFS_EXTENT_CSUM_KEY;
2339 key_end.offset = base + offset + nstripes * increment;
2340 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2342 if (!IS_ERR(reada1))
2343 btrfs_reada_wait(reada1);
2344 if (!IS_ERR(reada2))
2345 btrfs_reada_wait(reada2);
2349 * collect all data csums for the stripe to avoid seeking during
2350 * the scrub. This might currently (crc32) end up to be about 1MB
2352 blk_start_plug(&plug);
2355 * now find all extents for each stripe and scrub them
2357 logical = base + offset;
2358 physical = map->stripes[num].physical;
2359 logic_end = logical + increment * nstripes;
2361 while (logical < logic_end) {
2365 if (atomic_read(&fs_info->scrub_cancel_req) ||
2366 atomic_read(&sctx->cancel_req)) {
2371 * check to see if we have to pause
2373 if (atomic_read(&fs_info->scrub_pause_req)) {
2374 /* push queued extents */
2375 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2377 mutex_lock(&sctx->wr_ctx.wr_lock);
2378 scrub_wr_submit(sctx);
2379 mutex_unlock(&sctx->wr_ctx.wr_lock);
2380 wait_event(sctx->list_wait,
2381 atomic_read(&sctx->bios_in_flight) == 0);
2382 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2383 scrub_blocked_if_needed(fs_info);
2386 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2387 key.type = BTRFS_METADATA_ITEM_KEY;
2389 key.type = BTRFS_EXTENT_ITEM_KEY;
2390 key.objectid = logical;
2391 key.offset = (u64)-1;
2393 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2398 ret = btrfs_previous_extent_item(root, path, 0);
2402 /* there's no smaller item, so stick with the
2404 btrfs_release_path(path);
2405 ret = btrfs_search_slot(NULL, root, &key,
2417 slot = path->slots[0];
2418 if (slot >= btrfs_header_nritems(l)) {
2419 ret = btrfs_next_leaf(root, path);
2428 btrfs_item_key_to_cpu(l, &key, slot);
2430 if (key.type == BTRFS_METADATA_ITEM_KEY)
2431 bytes = root->leafsize;
2435 if (key.objectid + bytes <= logical)
2438 if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2439 key.type != BTRFS_METADATA_ITEM_KEY)
2442 if (key.objectid >= logical + map->stripe_len) {
2443 /* out of this device extent */
2444 if (key.objectid >= logic_end)
2449 extent = btrfs_item_ptr(l, slot,
2450 struct btrfs_extent_item);
2451 flags = btrfs_extent_flags(l, extent);
2452 generation = btrfs_extent_generation(l, extent);
2454 if (key.objectid < logical &&
2455 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2457 "scrub: tree block %llu spanning "
2458 "stripes, ignored. logical=%llu",
2459 key.objectid, logical);
2464 extent_logical = key.objectid;
2468 * trim extent to this stripe
2470 if (extent_logical < logical) {
2471 extent_len -= logical - extent_logical;
2472 extent_logical = logical;
2474 if (extent_logical + extent_len >
2475 logical + map->stripe_len) {
2476 extent_len = logical + map->stripe_len -
2480 extent_physical = extent_logical - logical + physical;
2481 extent_dev = scrub_dev;
2482 extent_mirror_num = mirror_num;
2484 scrub_remap_extent(fs_info, extent_logical,
2485 extent_len, &extent_physical,
2487 &extent_mirror_num);
2489 ret = btrfs_lookup_csums_range(csum_root, logical,
2490 logical + map->stripe_len - 1,
2491 &sctx->csum_list, 1);
2495 ret = scrub_extent(sctx, extent_logical, extent_len,
2496 extent_physical, extent_dev, flags,
2497 generation, extent_mirror_num,
2498 extent_logical - logical + physical);
2502 scrub_free_csums(sctx);
2503 if (extent_logical + extent_len <
2504 key.objectid + bytes) {
2505 logical += increment;
2506 physical += map->stripe_len;
2508 if (logical < key.objectid + bytes) {
2513 if (logical >= logic_end) {
2521 btrfs_release_path(path);
2522 logical += increment;
2523 physical += map->stripe_len;
2524 spin_lock(&sctx->stat_lock);
2526 sctx->stat.last_physical = map->stripes[num].physical +
2529 sctx->stat.last_physical = physical;
2530 spin_unlock(&sctx->stat_lock);
2535 /* push queued extents */
2537 mutex_lock(&sctx->wr_ctx.wr_lock);
2538 scrub_wr_submit(sctx);
2539 mutex_unlock(&sctx->wr_ctx.wr_lock);
2541 blk_finish_plug(&plug);
2542 btrfs_free_path(path);
2543 return ret < 0 ? ret : 0;
2546 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2547 struct btrfs_device *scrub_dev,
2548 u64 chunk_tree, u64 chunk_objectid,
2549 u64 chunk_offset, u64 length,
2550 u64 dev_offset, int is_dev_replace)
2552 struct btrfs_mapping_tree *map_tree =
2553 &sctx->dev_root->fs_info->mapping_tree;
2554 struct map_lookup *map;
2555 struct extent_map *em;
2559 read_lock(&map_tree->map_tree.lock);
2560 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2561 read_unlock(&map_tree->map_tree.lock);
2566 map = (struct map_lookup *)em->bdev;
2567 if (em->start != chunk_offset)
2570 if (em->len < length)
2573 for (i = 0; i < map->num_stripes; ++i) {
2574 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2575 map->stripes[i].physical == dev_offset) {
2576 ret = scrub_stripe(sctx, map, scrub_dev, i,
2577 chunk_offset, length,
2584 free_extent_map(em);
2589 static noinline_for_stack
2590 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2591 struct btrfs_device *scrub_dev, u64 start, u64 end,
2594 struct btrfs_dev_extent *dev_extent = NULL;
2595 struct btrfs_path *path;
2596 struct btrfs_root *root = sctx->dev_root;
2597 struct btrfs_fs_info *fs_info = root->fs_info;
2604 struct extent_buffer *l;
2605 struct btrfs_key key;
2606 struct btrfs_key found_key;
2607 struct btrfs_block_group_cache *cache;
2608 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2610 path = btrfs_alloc_path();
2615 path->search_commit_root = 1;
2616 path->skip_locking = 1;
2618 key.objectid = scrub_dev->devid;
2620 key.type = BTRFS_DEV_EXTENT_KEY;
2623 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2627 if (path->slots[0] >=
2628 btrfs_header_nritems(path->nodes[0])) {
2629 ret = btrfs_next_leaf(root, path);
2636 slot = path->slots[0];
2638 btrfs_item_key_to_cpu(l, &found_key, slot);
2640 if (found_key.objectid != scrub_dev->devid)
2643 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2646 if (found_key.offset >= end)
2649 if (found_key.offset < key.offset)
2652 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2653 length = btrfs_dev_extent_length(l, dev_extent);
2655 if (found_key.offset + length <= start) {
2656 key.offset = found_key.offset + length;
2657 btrfs_release_path(path);
2661 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2662 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2663 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2666 * get a reference on the corresponding block group to prevent
2667 * the chunk from going away while we scrub it
2669 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2674 dev_replace->cursor_right = found_key.offset + length;
2675 dev_replace->cursor_left = found_key.offset;
2676 dev_replace->item_needs_writeback = 1;
2677 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2678 chunk_offset, length, found_key.offset,
2682 * flush, submit all pending read and write bios, afterwards
2684 * Note that in the dev replace case, a read request causes
2685 * write requests that are submitted in the read completion
2686 * worker. Therefore in the current situation, it is required
2687 * that all write requests are flushed, so that all read and
2688 * write requests are really completed when bios_in_flight
2691 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2693 mutex_lock(&sctx->wr_ctx.wr_lock);
2694 scrub_wr_submit(sctx);
2695 mutex_unlock(&sctx->wr_ctx.wr_lock);
2697 wait_event(sctx->list_wait,
2698 atomic_read(&sctx->bios_in_flight) == 0);
2699 atomic_inc(&fs_info->scrubs_paused);
2700 wake_up(&fs_info->scrub_pause_wait);
2703 * must be called before we decrease @scrub_paused.
2704 * make sure we don't block transaction commit while
2705 * we are waiting pending workers finished.
2707 wait_event(sctx->list_wait,
2708 atomic_read(&sctx->workers_pending) == 0);
2709 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2711 mutex_lock(&fs_info->scrub_lock);
2712 __scrub_blocked_if_needed(fs_info);
2713 atomic_dec(&fs_info->scrubs_paused);
2714 mutex_unlock(&fs_info->scrub_lock);
2715 wake_up(&fs_info->scrub_pause_wait);
2717 btrfs_put_block_group(cache);
2720 if (is_dev_replace &&
2721 atomic64_read(&dev_replace->num_write_errors) > 0) {
2725 if (sctx->stat.malloc_errors > 0) {
2730 dev_replace->cursor_left = dev_replace->cursor_right;
2731 dev_replace->item_needs_writeback = 1;
2733 key.offset = found_key.offset + length;
2734 btrfs_release_path(path);
2737 btrfs_free_path(path);
2740 * ret can still be 1 from search_slot or next_leaf,
2741 * that's not an error
2743 return ret < 0 ? ret : 0;
2746 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2747 struct btrfs_device *scrub_dev)
2753 struct btrfs_root *root = sctx->dev_root;
2755 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2758 gen = root->fs_info->last_trans_committed;
2760 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2761 bytenr = btrfs_sb_offset(i);
2762 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2765 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2766 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2771 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2777 * get a reference count on fs_info->scrub_workers. start worker if necessary
2779 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2784 if (fs_info->scrub_workers_refcnt == 0) {
2786 btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2787 &fs_info->generic_worker);
2789 btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2790 fs_info->thread_pool_size,
2791 &fs_info->generic_worker);
2792 fs_info->scrub_workers.idle_thresh = 4;
2793 ret = btrfs_start_workers(&fs_info->scrub_workers);
2796 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2798 fs_info->thread_pool_size,
2799 &fs_info->generic_worker);
2800 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2801 ret = btrfs_start_workers(
2802 &fs_info->scrub_wr_completion_workers);
2805 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2806 &fs_info->generic_worker);
2807 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2811 ++fs_info->scrub_workers_refcnt;
2816 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2818 if (--fs_info->scrub_workers_refcnt == 0) {
2819 btrfs_stop_workers(&fs_info->scrub_workers);
2820 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2821 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2823 WARN_ON(fs_info->scrub_workers_refcnt < 0);
2826 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2827 u64 end, struct btrfs_scrub_progress *progress,
2828 int readonly, int is_dev_replace)
2830 struct scrub_ctx *sctx;
2832 struct btrfs_device *dev;
2834 if (btrfs_fs_closing(fs_info))
2838 * check some assumptions
2840 if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2842 "scrub: size assumption nodesize == leafsize (%d == %d) fails",
2843 fs_info->chunk_root->nodesize,
2844 fs_info->chunk_root->leafsize);
2848 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2850 * in this case scrub is unable to calculate the checksum
2851 * the way scrub is implemented. Do not handle this
2852 * situation at all because it won't ever happen.
2855 "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2856 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2860 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2861 /* not supported for data w/o checksums */
2863 "scrub: size assumption sectorsize != PAGE_SIZE "
2864 "(%d != %lu) fails",
2865 fs_info->chunk_root->sectorsize, PAGE_SIZE);
2869 if (fs_info->chunk_root->nodesize >
2870 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2871 fs_info->chunk_root->sectorsize >
2872 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2874 * would exhaust the array bounds of pagev member in
2875 * struct scrub_block
2877 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2878 "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2879 fs_info->chunk_root->nodesize,
2880 SCRUB_MAX_PAGES_PER_BLOCK,
2881 fs_info->chunk_root->sectorsize,
2882 SCRUB_MAX_PAGES_PER_BLOCK);
2887 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2888 dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2889 if (!dev || (dev->missing && !is_dev_replace)) {
2890 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2894 mutex_lock(&fs_info->scrub_lock);
2895 if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2896 mutex_unlock(&fs_info->scrub_lock);
2897 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2901 btrfs_dev_replace_lock(&fs_info->dev_replace);
2902 if (dev->scrub_device ||
2904 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2905 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2906 mutex_unlock(&fs_info->scrub_lock);
2907 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2908 return -EINPROGRESS;
2910 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2912 ret = scrub_workers_get(fs_info, is_dev_replace);
2914 mutex_unlock(&fs_info->scrub_lock);
2915 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2919 sctx = scrub_setup_ctx(dev, is_dev_replace);
2921 mutex_unlock(&fs_info->scrub_lock);
2922 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2923 scrub_workers_put(fs_info);
2924 return PTR_ERR(sctx);
2926 sctx->readonly = readonly;
2927 dev->scrub_device = sctx;
2928 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2931 * checking @scrub_pause_req here, we can avoid
2932 * race between committing transaction and scrubbing.
2934 __scrub_blocked_if_needed(fs_info);
2935 atomic_inc(&fs_info->scrubs_running);
2936 mutex_unlock(&fs_info->scrub_lock);
2938 if (!is_dev_replace) {
2940 * by holding device list mutex, we can
2941 * kick off writing super in log tree sync.
2943 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2944 ret = scrub_supers(sctx, dev);
2945 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2949 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2952 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2953 atomic_dec(&fs_info->scrubs_running);
2954 wake_up(&fs_info->scrub_pause_wait);
2956 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2959 memcpy(progress, &sctx->stat, sizeof(*progress));
2961 mutex_lock(&fs_info->scrub_lock);
2962 dev->scrub_device = NULL;
2963 scrub_workers_put(fs_info);
2964 mutex_unlock(&fs_info->scrub_lock);
2966 scrub_free_ctx(sctx);
2971 void btrfs_scrub_pause(struct btrfs_root *root)
2973 struct btrfs_fs_info *fs_info = root->fs_info;
2975 mutex_lock(&fs_info->scrub_lock);
2976 atomic_inc(&fs_info->scrub_pause_req);
2977 while (atomic_read(&fs_info->scrubs_paused) !=
2978 atomic_read(&fs_info->scrubs_running)) {
2979 mutex_unlock(&fs_info->scrub_lock);
2980 wait_event(fs_info->scrub_pause_wait,
2981 atomic_read(&fs_info->scrubs_paused) ==
2982 atomic_read(&fs_info->scrubs_running));
2983 mutex_lock(&fs_info->scrub_lock);
2985 mutex_unlock(&fs_info->scrub_lock);
2988 void btrfs_scrub_continue(struct btrfs_root *root)
2990 struct btrfs_fs_info *fs_info = root->fs_info;
2992 atomic_dec(&fs_info->scrub_pause_req);
2993 wake_up(&fs_info->scrub_pause_wait);
2996 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2998 mutex_lock(&fs_info->scrub_lock);
2999 if (!atomic_read(&fs_info->scrubs_running)) {
3000 mutex_unlock(&fs_info->scrub_lock);
3004 atomic_inc(&fs_info->scrub_cancel_req);
3005 while (atomic_read(&fs_info->scrubs_running)) {
3006 mutex_unlock(&fs_info->scrub_lock);
3007 wait_event(fs_info->scrub_pause_wait,
3008 atomic_read(&fs_info->scrubs_running) == 0);
3009 mutex_lock(&fs_info->scrub_lock);
3011 atomic_dec(&fs_info->scrub_cancel_req);
3012 mutex_unlock(&fs_info->scrub_lock);
3017 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3018 struct btrfs_device *dev)
3020 struct scrub_ctx *sctx;
3022 mutex_lock(&fs_info->scrub_lock);
3023 sctx = dev->scrub_device;
3025 mutex_unlock(&fs_info->scrub_lock);
3028 atomic_inc(&sctx->cancel_req);
3029 while (dev->scrub_device) {
3030 mutex_unlock(&fs_info->scrub_lock);
3031 wait_event(fs_info->scrub_pause_wait,
3032 dev->scrub_device == NULL);
3033 mutex_lock(&fs_info->scrub_lock);
3035 mutex_unlock(&fs_info->scrub_lock);
3040 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3041 struct btrfs_scrub_progress *progress)
3043 struct btrfs_device *dev;
3044 struct scrub_ctx *sctx = NULL;
3046 mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3047 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3049 sctx = dev->scrub_device;
3051 memcpy(progress, &sctx->stat, sizeof(*progress));
3052 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3054 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3057 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3058 u64 extent_logical, u64 extent_len,
3059 u64 *extent_physical,
3060 struct btrfs_device **extent_dev,
3061 int *extent_mirror_num)
3064 struct btrfs_bio *bbio = NULL;
3067 mapped_length = extent_len;
3068 ret = btrfs_map_block(fs_info, READ, extent_logical,
3069 &mapped_length, &bbio, 0);
3070 if (ret || !bbio || mapped_length < extent_len ||
3071 !bbio->stripes[0].dev->bdev) {
3076 *extent_physical = bbio->stripes[0].physical;
3077 *extent_mirror_num = bbio->mirror_num;
3078 *extent_dev = bbio->stripes[0].dev;
3082 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3083 struct scrub_wr_ctx *wr_ctx,
3084 struct btrfs_fs_info *fs_info,
3085 struct btrfs_device *dev,
3088 WARN_ON(wr_ctx->wr_curr_bio != NULL);
3090 mutex_init(&wr_ctx->wr_lock);
3091 wr_ctx->wr_curr_bio = NULL;
3092 if (!is_dev_replace)
3095 WARN_ON(!dev->bdev);
3096 wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3097 bio_get_nr_vecs(dev->bdev));
3098 wr_ctx->tgtdev = dev;
3099 atomic_set(&wr_ctx->flush_all_writes, 0);
3103 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3105 mutex_lock(&wr_ctx->wr_lock);
3106 kfree(wr_ctx->wr_curr_bio);
3107 wr_ctx->wr_curr_bio = NULL;
3108 mutex_unlock(&wr_ctx->wr_lock);
3111 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3112 int mirror_num, u64 physical_for_dev_replace)
3114 struct scrub_copy_nocow_ctx *nocow_ctx;
3115 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3117 nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3119 spin_lock(&sctx->stat_lock);
3120 sctx->stat.malloc_errors++;
3121 spin_unlock(&sctx->stat_lock);
3125 scrub_pending_trans_workers_inc(sctx);
3127 nocow_ctx->sctx = sctx;
3128 nocow_ctx->logical = logical;
3129 nocow_ctx->len = len;
3130 nocow_ctx->mirror_num = mirror_num;
3131 nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3132 nocow_ctx->work.func = copy_nocow_pages_worker;
3133 INIT_LIST_HEAD(&nocow_ctx->inodes);
3134 btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3140 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3142 struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3143 struct scrub_nocow_inode *nocow_inode;
3145 nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3148 nocow_inode->inum = inum;
3149 nocow_inode->offset = offset;
3150 nocow_inode->root = root;
3151 list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3155 #define COPY_COMPLETE 1
3157 static void copy_nocow_pages_worker(struct btrfs_work *work)
3159 struct scrub_copy_nocow_ctx *nocow_ctx =
3160 container_of(work, struct scrub_copy_nocow_ctx, work);
3161 struct scrub_ctx *sctx = nocow_ctx->sctx;
3162 u64 logical = nocow_ctx->logical;
3163 u64 len = nocow_ctx->len;
3164 int mirror_num = nocow_ctx->mirror_num;
3165 u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3167 struct btrfs_trans_handle *trans = NULL;
3168 struct btrfs_fs_info *fs_info;
3169 struct btrfs_path *path;
3170 struct btrfs_root *root;
3171 int not_written = 0;
3173 fs_info = sctx->dev_root->fs_info;
3174 root = fs_info->extent_root;
3176 path = btrfs_alloc_path();
3178 spin_lock(&sctx->stat_lock);
3179 sctx->stat.malloc_errors++;
3180 spin_unlock(&sctx->stat_lock);
3185 trans = btrfs_join_transaction(root);
3186 if (IS_ERR(trans)) {
3191 ret = iterate_inodes_from_logical(logical, fs_info, path,
3192 record_inode_for_nocow, nocow_ctx);
3193 if (ret != 0 && ret != -ENOENT) {
3194 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3195 "phys %llu, len %llu, mir %u, ret %d",
3196 logical, physical_for_dev_replace, len, mirror_num,
3202 btrfs_end_transaction(trans, root);
3204 while (!list_empty(&nocow_ctx->inodes)) {
3205 struct scrub_nocow_inode *entry;
3206 entry = list_first_entry(&nocow_ctx->inodes,
3207 struct scrub_nocow_inode,
3209 list_del_init(&entry->list);
3210 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3211 entry->root, nocow_ctx);
3213 if (ret == COPY_COMPLETE) {
3221 while (!list_empty(&nocow_ctx->inodes)) {
3222 struct scrub_nocow_inode *entry;
3223 entry = list_first_entry(&nocow_ctx->inodes,
3224 struct scrub_nocow_inode,
3226 list_del_init(&entry->list);
3229 if (trans && !IS_ERR(trans))
3230 btrfs_end_transaction(trans, root);
3232 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3233 num_uncorrectable_read_errors);
3235 btrfs_free_path(path);
3238 scrub_pending_trans_workers_dec(sctx);
3241 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3242 struct scrub_copy_nocow_ctx *nocow_ctx)
3244 struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3245 struct btrfs_key key;
3246 struct inode *inode;
3248 struct btrfs_root *local_root;
3249 struct btrfs_ordered_extent *ordered;
3250 struct extent_map *em;
3251 struct extent_state *cached_state = NULL;
3252 struct extent_io_tree *io_tree;
3253 u64 physical_for_dev_replace;
3254 u64 len = nocow_ctx->len;
3255 u64 lockstart = offset, lockend = offset + len - 1;
3256 unsigned long index;
3261 key.objectid = root;
3262 key.type = BTRFS_ROOT_ITEM_KEY;
3263 key.offset = (u64)-1;
3265 srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3267 local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3268 if (IS_ERR(local_root)) {
3269 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3270 return PTR_ERR(local_root);
3273 key.type = BTRFS_INODE_ITEM_KEY;
3274 key.objectid = inum;
3276 inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3277 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3279 return PTR_ERR(inode);
3281 /* Avoid truncate/dio/punch hole.. */
3282 mutex_lock(&inode->i_mutex);
3283 inode_dio_wait(inode);
3285 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3286 io_tree = &BTRFS_I(inode)->io_tree;
3288 lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3289 ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3291 btrfs_put_ordered_extent(ordered);
3295 em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3302 * This extent does not actually cover the logical extent anymore,
3303 * move on to the next inode.
3305 if (em->block_start > nocow_ctx->logical ||
3306 em->block_start + em->block_len < nocow_ctx->logical + len) {
3307 free_extent_map(em);
3310 free_extent_map(em);
3312 while (len >= PAGE_CACHE_SIZE) {
3313 index = offset >> PAGE_CACHE_SHIFT;
3315 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3317 btrfs_err(fs_info, "find_or_create_page() failed");
3322 if (PageUptodate(page)) {
3323 if (PageDirty(page))
3326 ClearPageError(page);
3327 err = extent_read_full_page_nolock(io_tree, page,
3329 nocow_ctx->mirror_num);
3337 * If the page has been remove from the page cache,
3338 * the data on it is meaningless, because it may be
3339 * old one, the new data may be written into the new
3340 * page in the page cache.
3342 if (page->mapping != inode->i_mapping) {
3344 page_cache_release(page);
3347 if (!PageUptodate(page)) {
3352 err = write_page_nocow(nocow_ctx->sctx,
3353 physical_for_dev_replace, page);
3358 page_cache_release(page);
3363 offset += PAGE_CACHE_SIZE;
3364 physical_for_dev_replace += PAGE_CACHE_SIZE;
3365 len -= PAGE_CACHE_SIZE;
3367 ret = COPY_COMPLETE;
3369 unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3372 mutex_unlock(&inode->i_mutex);
3377 static int write_page_nocow(struct scrub_ctx *sctx,
3378 u64 physical_for_dev_replace, struct page *page)
3381 struct btrfs_device *dev;
3384 dev = sctx->wr_ctx.tgtdev;
3388 printk_ratelimited(KERN_WARNING
3389 "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3392 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3394 spin_lock(&sctx->stat_lock);
3395 sctx->stat.malloc_errors++;
3396 spin_unlock(&sctx->stat_lock);
3400 bio->bi_sector = physical_for_dev_replace >> 9;
3401 bio->bi_bdev = dev->bdev;
3402 ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3403 if (ret != PAGE_CACHE_SIZE) {
3406 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3410 if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3411 goto leave_with_eio;