1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2011, 2012 STRATO. All rights reserved.
6 #include <linux/blkdev.h>
7 #include <linux/ratelimit.h>
8 #include <linux/sched/mm.h>
9 #include <crypto/hash.h>
14 #include "ordered-data.h"
15 #include "transaction.h"
17 #include "extent_io.h"
18 #include "dev-replace.h"
19 #include "check-integrity.h"
21 #include "block-group.h"
24 #include "accessors.h"
25 #include "file-item.h"
29 * This is only the first step towards a full-features scrub. It reads all
30 * extent and super block and verifies the checksums. In case a bad checksum
31 * is found or the extent cannot be read, good data will be written back if
34 * Future enhancements:
35 * - In case an unrepairable extent is encountered, track which files are
36 * affected and report them
37 * - track and record media errors, throw out bad devices
38 * - add a mode to also read unallocated space
44 * The following value only influences the performance.
46 * This determines the batch size for stripe submitted in one go.
48 #define SCRUB_STRIPES_PER_SCTX 8 /* That would be 8 64K stripe per-device. */
51 * The following value times PAGE_SIZE needs to be large enough to match the
52 * largest node/leaf/sector size that shall be supported.
54 #define SCRUB_MAX_SECTORS_PER_BLOCK (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
56 /* Represent one sector and its needed info to verify the content. */
57 struct scrub_sector_verification {
62 * Csum pointer for data csum verification. Should point to a
63 * sector csum inside scrub_stripe::csums.
65 * NULL if this data sector has no csum.
70 * Extra info for metadata verification. All sectors inside a
71 * tree block share the same generation.
77 enum scrub_stripe_flags {
78 /* Set when @mirror_num, @dev, @physical and @logical are set. */
79 SCRUB_STRIPE_FLAG_INITIALIZED,
81 /* Set when the read-repair is finished. */
82 SCRUB_STRIPE_FLAG_REPAIR_DONE,
85 * Set for data stripes if it's triggered from P/Q stripe.
86 * During such scrub, we should not report errors in data stripes, nor
87 * update the accounting.
89 SCRUB_STRIPE_FLAG_NO_REPORT,
92 #define SCRUB_STRIPE_PAGES (BTRFS_STRIPE_LEN / PAGE_SIZE)
95 * Represent one contiguous range with a length of BTRFS_STRIPE_LEN.
98 struct scrub_ctx *sctx;
99 struct btrfs_block_group *bg;
101 struct page *pages[SCRUB_STRIPE_PAGES];
102 struct scrub_sector_verification *sectors;
104 struct btrfs_device *dev;
110 /* Should be BTRFS_STRIPE_LEN / sectorsize. */
114 * How many data/meta extents are in this stripe. Only for scrub status
115 * reporting purposes.
121 wait_queue_head_t io_wait;
122 wait_queue_head_t repair_wait;
125 * Indicate the states of the stripe. Bits are defined in
126 * scrub_stripe_flags enum.
130 /* Indicate which sectors are covered by extent items. */
131 unsigned long extent_sector_bitmap;
134 * The errors hit during the initial read of the stripe.
136 * Would be utilized for error reporting and repair.
138 * The remaining init_nr_* records the number of errors hit, only used
139 * by error reporting.
141 unsigned long init_error_bitmap;
142 unsigned int init_nr_io_errors;
143 unsigned int init_nr_csum_errors;
144 unsigned int init_nr_meta_errors;
147 * The following error bitmaps are all for the current status.
148 * Every time we submit a new read, these bitmaps may be updated.
150 * error_bitmap = io_error_bitmap | csum_error_bitmap | meta_error_bitmap;
152 * IO and csum errors can happen for both metadata and data.
154 unsigned long error_bitmap;
155 unsigned long io_error_bitmap;
156 unsigned long csum_error_bitmap;
157 unsigned long meta_error_bitmap;
159 /* For writeback (repair or replace) error reporting. */
160 unsigned long write_error_bitmap;
162 /* Writeback can be concurrent, thus we need to protect the bitmap. */
163 spinlock_t write_error_lock;
166 * Checksum for the whole stripe if this stripe is inside a data block
171 struct work_struct work;
175 struct scrub_stripe stripes[SCRUB_STRIPES_PER_SCTX];
176 struct scrub_stripe *raid56_data_stripes;
177 struct btrfs_fs_info *fs_info;
184 /* State of IO submission throttling affecting the associated device */
185 ktime_t throttle_deadline;
191 struct mutex wr_lock;
192 struct btrfs_device *wr_tgtdev;
197 struct btrfs_scrub_progress stat;
198 spinlock_t stat_lock;
201 * Use a ref counter to avoid use-after-free issues. Scrub workers
202 * decrement bios_in_flight and workers_pending and then do a wakeup
203 * on the list_wait wait queue. We must ensure the main scrub task
204 * doesn't free the scrub context before or while the workers are
205 * doing the wakeup() call.
210 struct scrub_warning {
211 struct btrfs_path *path;
212 u64 extent_item_size;
216 struct btrfs_device *dev;
219 static void release_scrub_stripe(struct scrub_stripe *stripe)
224 for (int i = 0; i < SCRUB_STRIPE_PAGES; i++) {
225 if (stripe->pages[i])
226 __free_page(stripe->pages[i]);
227 stripe->pages[i] = NULL;
229 kfree(stripe->sectors);
230 kfree(stripe->csums);
231 stripe->sectors = NULL;
232 stripe->csums = NULL;
237 static int init_scrub_stripe(struct btrfs_fs_info *fs_info,
238 struct scrub_stripe *stripe)
242 memset(stripe, 0, sizeof(*stripe));
244 stripe->nr_sectors = BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
247 init_waitqueue_head(&stripe->io_wait);
248 init_waitqueue_head(&stripe->repair_wait);
249 atomic_set(&stripe->pending_io, 0);
250 spin_lock_init(&stripe->write_error_lock);
252 ret = btrfs_alloc_page_array(SCRUB_STRIPE_PAGES, stripe->pages);
256 stripe->sectors = kcalloc(stripe->nr_sectors,
257 sizeof(struct scrub_sector_verification),
259 if (!stripe->sectors)
262 stripe->csums = kcalloc(BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits,
263 fs_info->csum_size, GFP_KERNEL);
268 release_scrub_stripe(stripe);
272 static void wait_scrub_stripe_io(struct scrub_stripe *stripe)
274 wait_event(stripe->io_wait, atomic_read(&stripe->pending_io) == 0);
277 static void scrub_put_ctx(struct scrub_ctx *sctx);
279 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
281 while (atomic_read(&fs_info->scrub_pause_req)) {
282 mutex_unlock(&fs_info->scrub_lock);
283 wait_event(fs_info->scrub_pause_wait,
284 atomic_read(&fs_info->scrub_pause_req) == 0);
285 mutex_lock(&fs_info->scrub_lock);
289 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
291 atomic_inc(&fs_info->scrubs_paused);
292 wake_up(&fs_info->scrub_pause_wait);
295 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
297 mutex_lock(&fs_info->scrub_lock);
298 __scrub_blocked_if_needed(fs_info);
299 atomic_dec(&fs_info->scrubs_paused);
300 mutex_unlock(&fs_info->scrub_lock);
302 wake_up(&fs_info->scrub_pause_wait);
305 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
307 scrub_pause_on(fs_info);
308 scrub_pause_off(fs_info);
311 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
318 for (i = 0; i < SCRUB_STRIPES_PER_SCTX; i++)
319 release_scrub_stripe(&sctx->stripes[i]);
324 static void scrub_put_ctx(struct scrub_ctx *sctx)
326 if (refcount_dec_and_test(&sctx->refs))
327 scrub_free_ctx(sctx);
330 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
331 struct btrfs_fs_info *fs_info, int is_dev_replace)
333 struct scrub_ctx *sctx;
336 sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
339 refcount_set(&sctx->refs, 1);
340 sctx->is_dev_replace = is_dev_replace;
341 sctx->fs_info = fs_info;
342 for (i = 0; i < SCRUB_STRIPES_PER_SCTX; i++) {
345 ret = init_scrub_stripe(fs_info, &sctx->stripes[i]);
348 sctx->stripes[i].sctx = sctx;
350 sctx->first_free = 0;
351 atomic_set(&sctx->cancel_req, 0);
353 spin_lock_init(&sctx->stat_lock);
354 sctx->throttle_deadline = 0;
356 mutex_init(&sctx->wr_lock);
357 if (is_dev_replace) {
358 WARN_ON(!fs_info->dev_replace.tgtdev);
359 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
365 scrub_free_ctx(sctx);
366 return ERR_PTR(-ENOMEM);
369 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
370 u64 root, void *warn_ctx)
376 struct extent_buffer *eb;
377 struct btrfs_inode_item *inode_item;
378 struct scrub_warning *swarn = warn_ctx;
379 struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
380 struct inode_fs_paths *ipath = NULL;
381 struct btrfs_root *local_root;
382 struct btrfs_key key;
384 local_root = btrfs_get_fs_root(fs_info, root, true);
385 if (IS_ERR(local_root)) {
386 ret = PTR_ERR(local_root);
391 * this makes the path point to (inum INODE_ITEM ioff)
394 key.type = BTRFS_INODE_ITEM_KEY;
397 ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
399 btrfs_put_root(local_root);
400 btrfs_release_path(swarn->path);
404 eb = swarn->path->nodes[0];
405 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
406 struct btrfs_inode_item);
407 nlink = btrfs_inode_nlink(eb, inode_item);
408 btrfs_release_path(swarn->path);
411 * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
412 * uses GFP_NOFS in this context, so we keep it consistent but it does
413 * not seem to be strictly necessary.
415 nofs_flag = memalloc_nofs_save();
416 ipath = init_ipath(4096, local_root, swarn->path);
417 memalloc_nofs_restore(nofs_flag);
419 btrfs_put_root(local_root);
420 ret = PTR_ERR(ipath);
424 ret = paths_from_inode(inum, ipath);
430 * we deliberately ignore the bit ipath might have been too small to
431 * hold all of the paths here
433 for (i = 0; i < ipath->fspath->elem_cnt; ++i)
434 btrfs_warn_in_rcu(fs_info,
435 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
436 swarn->errstr, swarn->logical,
437 btrfs_dev_name(swarn->dev),
440 fs_info->sectorsize, nlink,
441 (char *)(unsigned long)ipath->fspath->val[i]);
443 btrfs_put_root(local_root);
448 btrfs_warn_in_rcu(fs_info,
449 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
450 swarn->errstr, swarn->logical,
451 btrfs_dev_name(swarn->dev),
453 root, inum, offset, ret);
459 static void scrub_print_common_warning(const char *errstr, struct btrfs_device *dev,
460 bool is_super, u64 logical, u64 physical)
462 struct btrfs_fs_info *fs_info = dev->fs_info;
463 struct btrfs_path *path;
464 struct btrfs_key found_key;
465 struct extent_buffer *eb;
466 struct btrfs_extent_item *ei;
467 struct scrub_warning swarn;
472 /* Super block error, no need to search extent tree. */
474 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
475 errstr, btrfs_dev_name(dev), physical);
478 path = btrfs_alloc_path();
482 swarn.physical = physical;
483 swarn.logical = logical;
484 swarn.errstr = errstr;
487 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
492 swarn.extent_item_size = found_key.offset;
495 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
496 item_size = btrfs_item_size(eb, path->slots[0]);
498 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
499 unsigned long ptr = 0;
504 ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
505 item_size, &ref_root,
509 "failed to resolve tree backref for logical %llu: %d",
515 btrfs_warn_in_rcu(fs_info,
516 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
517 errstr, swarn.logical, btrfs_dev_name(dev),
518 swarn.physical, (ref_level ? "node" : "leaf"),
519 ref_level, ref_root);
521 btrfs_release_path(path);
523 struct btrfs_backref_walk_ctx ctx = { 0 };
525 btrfs_release_path(path);
527 ctx.bytenr = found_key.objectid;
528 ctx.extent_item_pos = swarn.logical - found_key.objectid;
529 ctx.fs_info = fs_info;
534 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
538 btrfs_free_path(path);
541 static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
546 if (!btrfs_is_zoned(sctx->fs_info))
549 if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
552 if (sctx->write_pointer < physical) {
553 length = physical - sctx->write_pointer;
555 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
556 sctx->write_pointer, length);
558 sctx->write_pointer = physical;
563 static struct page *scrub_stripe_get_page(struct scrub_stripe *stripe, int sector_nr)
565 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
566 int page_index = (sector_nr << fs_info->sectorsize_bits) >> PAGE_SHIFT;
568 return stripe->pages[page_index];
571 static unsigned int scrub_stripe_get_page_offset(struct scrub_stripe *stripe,
574 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
576 return offset_in_page(sector_nr << fs_info->sectorsize_bits);
579 static void scrub_verify_one_metadata(struct scrub_stripe *stripe, int sector_nr)
581 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
582 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
583 const u64 logical = stripe->logical + (sector_nr << fs_info->sectorsize_bits);
584 const struct page *first_page = scrub_stripe_get_page(stripe, sector_nr);
585 const unsigned int first_off = scrub_stripe_get_page_offset(stripe, sector_nr);
586 SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
587 u8 on_disk_csum[BTRFS_CSUM_SIZE];
588 u8 calculated_csum[BTRFS_CSUM_SIZE];
589 struct btrfs_header *header;
592 * Here we don't have a good way to attach the pages (and subpages)
593 * to a dummy extent buffer, thus we have to directly grab the members
596 header = (struct btrfs_header *)(page_address(first_page) + first_off);
597 memcpy(on_disk_csum, header->csum, fs_info->csum_size);
599 if (logical != btrfs_stack_header_bytenr(header)) {
600 bitmap_set(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
601 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
602 btrfs_warn_rl(fs_info,
603 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
604 logical, stripe->mirror_num,
605 btrfs_stack_header_bytenr(header), logical);
608 if (memcmp(header->fsid, fs_info->fs_devices->fsid, BTRFS_FSID_SIZE) != 0) {
609 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
610 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
611 btrfs_warn_rl(fs_info,
612 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
613 logical, stripe->mirror_num,
614 header->fsid, fs_info->fs_devices->fsid);
617 if (memcmp(header->chunk_tree_uuid, fs_info->chunk_tree_uuid,
618 BTRFS_UUID_SIZE) != 0) {
619 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
620 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
621 btrfs_warn_rl(fs_info,
622 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
623 logical, stripe->mirror_num,
624 header->chunk_tree_uuid, fs_info->chunk_tree_uuid);
628 /* Now check tree block csum. */
629 shash->tfm = fs_info->csum_shash;
630 crypto_shash_init(shash);
631 crypto_shash_update(shash, page_address(first_page) + first_off +
632 BTRFS_CSUM_SIZE, fs_info->sectorsize - BTRFS_CSUM_SIZE);
634 for (int i = sector_nr + 1; i < sector_nr + sectors_per_tree; i++) {
635 struct page *page = scrub_stripe_get_page(stripe, i);
636 unsigned int page_off = scrub_stripe_get_page_offset(stripe, i);
638 crypto_shash_update(shash, page_address(page) + page_off,
639 fs_info->sectorsize);
642 crypto_shash_final(shash, calculated_csum);
643 if (memcmp(calculated_csum, on_disk_csum, fs_info->csum_size) != 0) {
644 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
645 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
646 btrfs_warn_rl(fs_info,
647 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
648 logical, stripe->mirror_num,
649 CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
650 CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
653 if (stripe->sectors[sector_nr].generation !=
654 btrfs_stack_header_generation(header)) {
655 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
656 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
657 btrfs_warn_rl(fs_info,
658 "tree block %llu mirror %u has bad generation, has %llu want %llu",
659 logical, stripe->mirror_num,
660 btrfs_stack_header_generation(header),
661 stripe->sectors[sector_nr].generation);
664 bitmap_clear(&stripe->error_bitmap, sector_nr, sectors_per_tree);
665 bitmap_clear(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
666 bitmap_clear(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
669 static void scrub_verify_one_sector(struct scrub_stripe *stripe, int sector_nr)
671 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
672 struct scrub_sector_verification *sector = &stripe->sectors[sector_nr];
673 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
674 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
675 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
676 u8 csum_buf[BTRFS_CSUM_SIZE];
679 ASSERT(sector_nr >= 0 && sector_nr < stripe->nr_sectors);
681 /* Sector not utilized, skip it. */
682 if (!test_bit(sector_nr, &stripe->extent_sector_bitmap))
685 /* IO error, no need to check. */
686 if (test_bit(sector_nr, &stripe->io_error_bitmap))
689 /* Metadata, verify the full tree block. */
690 if (sector->is_metadata) {
692 * Check if the tree block crosses the stripe boudary. If
693 * crossed the boundary, we cannot verify it but only give a
696 * This can only happen on a very old filesystem where chunks
697 * are not ensured to be stripe aligned.
699 if (unlikely(sector_nr + sectors_per_tree > stripe->nr_sectors)) {
700 btrfs_warn_rl(fs_info,
701 "tree block at %llu crosses stripe boundary %llu",
703 (sector_nr << fs_info->sectorsize_bits),
707 scrub_verify_one_metadata(stripe, sector_nr);
712 * Data is easier, we just verify the data csum (if we have it). For
713 * cases without csum, we have no other choice but to trust it.
716 clear_bit(sector_nr, &stripe->error_bitmap);
720 ret = btrfs_check_sector_csum(fs_info, page, pgoff, csum_buf, sector->csum);
722 set_bit(sector_nr, &stripe->csum_error_bitmap);
723 set_bit(sector_nr, &stripe->error_bitmap);
725 clear_bit(sector_nr, &stripe->csum_error_bitmap);
726 clear_bit(sector_nr, &stripe->error_bitmap);
730 /* Verify specified sectors of a stripe. */
731 static void scrub_verify_one_stripe(struct scrub_stripe *stripe, unsigned long bitmap)
733 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
734 const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
737 for_each_set_bit(sector_nr, &bitmap, stripe->nr_sectors) {
738 scrub_verify_one_sector(stripe, sector_nr);
739 if (stripe->sectors[sector_nr].is_metadata)
740 sector_nr += sectors_per_tree - 1;
744 static int calc_sector_number(struct scrub_stripe *stripe, struct bio_vec *first_bvec)
748 for (i = 0; i < stripe->nr_sectors; i++) {
749 if (scrub_stripe_get_page(stripe, i) == first_bvec->bv_page &&
750 scrub_stripe_get_page_offset(stripe, i) == first_bvec->bv_offset)
753 ASSERT(i < stripe->nr_sectors);
758 * Repair read is different to the regular read:
760 * - Only reads the failed sectors
761 * - May have extra blocksize limits
763 static void scrub_repair_read_endio(struct btrfs_bio *bbio)
765 struct scrub_stripe *stripe = bbio->private;
766 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
767 struct bio_vec *bvec;
768 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
772 ASSERT(sector_nr < stripe->nr_sectors);
774 bio_for_each_bvec_all(bvec, &bbio->bio, i)
775 bio_size += bvec->bv_len;
777 if (bbio->bio.bi_status) {
778 bitmap_set(&stripe->io_error_bitmap, sector_nr,
779 bio_size >> fs_info->sectorsize_bits);
780 bitmap_set(&stripe->error_bitmap, sector_nr,
781 bio_size >> fs_info->sectorsize_bits);
783 bitmap_clear(&stripe->io_error_bitmap, sector_nr,
784 bio_size >> fs_info->sectorsize_bits);
787 if (atomic_dec_and_test(&stripe->pending_io))
788 wake_up(&stripe->io_wait);
791 static int calc_next_mirror(int mirror, int num_copies)
793 ASSERT(mirror <= num_copies);
794 return (mirror + 1 > num_copies) ? 1 : mirror + 1;
797 static void scrub_stripe_submit_repair_read(struct scrub_stripe *stripe,
798 int mirror, int blocksize, bool wait)
800 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
801 struct btrfs_bio *bbio = NULL;
802 const unsigned long old_error_bitmap = stripe->error_bitmap;
805 ASSERT(stripe->mirror_num >= 1);
806 ASSERT(atomic_read(&stripe->pending_io) == 0);
808 for_each_set_bit(i, &old_error_bitmap, stripe->nr_sectors) {
813 page = scrub_stripe_get_page(stripe, i);
814 pgoff = scrub_stripe_get_page_offset(stripe, i);
816 /* The current sector cannot be merged, submit the bio. */
817 if (bbio && ((i > 0 && !test_bit(i - 1, &stripe->error_bitmap)) ||
818 bbio->bio.bi_iter.bi_size >= blocksize)) {
819 ASSERT(bbio->bio.bi_iter.bi_size);
820 atomic_inc(&stripe->pending_io);
821 btrfs_submit_bio(bbio, mirror);
823 wait_scrub_stripe_io(stripe);
828 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
829 fs_info, scrub_repair_read_endio, stripe);
830 bbio->bio.bi_iter.bi_sector = (stripe->logical +
831 (i << fs_info->sectorsize_bits)) >> SECTOR_SHIFT;
834 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
835 ASSERT(ret == fs_info->sectorsize);
838 ASSERT(bbio->bio.bi_iter.bi_size);
839 atomic_inc(&stripe->pending_io);
840 btrfs_submit_bio(bbio, mirror);
842 wait_scrub_stripe_io(stripe);
846 static void scrub_stripe_report_errors(struct scrub_ctx *sctx,
847 struct scrub_stripe *stripe)
849 static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
850 DEFAULT_RATELIMIT_BURST);
851 struct btrfs_fs_info *fs_info = sctx->fs_info;
852 struct btrfs_device *dev = NULL;
854 int nr_data_sectors = 0;
855 int nr_meta_sectors = 0;
856 int nr_nodatacsum_sectors = 0;
857 int nr_repaired_sectors = 0;
860 if (test_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state))
864 * Init needed infos for error reporting.
866 * Although our scrub_stripe infrastucture is mostly based on btrfs_submit_bio()
867 * thus no need for dev/physical, error reporting still needs dev and physical.
869 if (!bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors)) {
870 u64 mapped_len = fs_info->sectorsize;
871 struct btrfs_io_context *bioc = NULL;
872 int stripe_index = stripe->mirror_num - 1;
875 /* For scrub, our mirror_num should always start at 1. */
876 ASSERT(stripe->mirror_num >= 1);
877 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
878 stripe->logical, &mapped_len, &bioc,
881 * If we failed, dev will be NULL, and later detailed reports
882 * will just be skipped.
886 physical = bioc->stripes[stripe_index].physical;
887 dev = bioc->stripes[stripe_index].dev;
888 btrfs_put_bioc(bioc);
892 for_each_set_bit(sector_nr, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
893 bool repaired = false;
895 if (stripe->sectors[sector_nr].is_metadata) {
899 if (!stripe->sectors[sector_nr].csum)
900 nr_nodatacsum_sectors++;
903 if (test_bit(sector_nr, &stripe->init_error_bitmap) &&
904 !test_bit(sector_nr, &stripe->error_bitmap)) {
905 nr_repaired_sectors++;
909 /* Good sector from the beginning, nothing need to be done. */
910 if (!test_bit(sector_nr, &stripe->init_error_bitmap))
914 * Report error for the corrupted sectors. If repaired, just
915 * output the message of repaired message.
919 btrfs_err_rl_in_rcu(fs_info,
920 "fixed up error at logical %llu on dev %s physical %llu",
921 stripe->logical, btrfs_dev_name(dev),
924 btrfs_err_rl_in_rcu(fs_info,
925 "fixed up error at logical %llu on mirror %u",
926 stripe->logical, stripe->mirror_num);
931 /* The remaining are all for unrepaired. */
933 btrfs_err_rl_in_rcu(fs_info,
934 "unable to fixup (regular) error at logical %llu on dev %s physical %llu",
935 stripe->logical, btrfs_dev_name(dev),
938 btrfs_err_rl_in_rcu(fs_info,
939 "unable to fixup (regular) error at logical %llu on mirror %u",
940 stripe->logical, stripe->mirror_num);
943 if (test_bit(sector_nr, &stripe->io_error_bitmap))
944 if (__ratelimit(&rs) && dev)
945 scrub_print_common_warning("i/o error", dev, false,
946 stripe->logical, physical);
947 if (test_bit(sector_nr, &stripe->csum_error_bitmap))
948 if (__ratelimit(&rs) && dev)
949 scrub_print_common_warning("checksum error", dev, false,
950 stripe->logical, physical);
951 if (test_bit(sector_nr, &stripe->meta_error_bitmap))
952 if (__ratelimit(&rs) && dev)
953 scrub_print_common_warning("header error", dev, false,
954 stripe->logical, physical);
957 spin_lock(&sctx->stat_lock);
958 sctx->stat.data_extents_scrubbed += stripe->nr_data_extents;
959 sctx->stat.tree_extents_scrubbed += stripe->nr_meta_extents;
960 sctx->stat.data_bytes_scrubbed += nr_data_sectors << fs_info->sectorsize_bits;
961 sctx->stat.tree_bytes_scrubbed += nr_meta_sectors << fs_info->sectorsize_bits;
962 sctx->stat.no_csum += nr_nodatacsum_sectors;
963 sctx->stat.read_errors += stripe->init_nr_io_errors;
964 sctx->stat.csum_errors += stripe->init_nr_csum_errors;
965 sctx->stat.verify_errors += stripe->init_nr_meta_errors;
966 sctx->stat.uncorrectable_errors +=
967 bitmap_weight(&stripe->error_bitmap, stripe->nr_sectors);
968 sctx->stat.corrected_errors += nr_repaired_sectors;
969 spin_unlock(&sctx->stat_lock);
973 * The main entrance for all read related scrub work, including:
975 * - Wait for the initial read to finish
976 * - Verify and locate any bad sectors
977 * - Go through the remaining mirrors and try to read as large blocksize as
979 * - Go through all mirrors (including the failed mirror) sector-by-sector
981 * Writeback does not happen here, it needs extra synchronization.
983 static void scrub_stripe_read_repair_worker(struct work_struct *work)
985 struct scrub_stripe *stripe = container_of(work, struct scrub_stripe, work);
986 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
987 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
992 ASSERT(stripe->mirror_num > 0);
994 wait_scrub_stripe_io(stripe);
995 scrub_verify_one_stripe(stripe, stripe->extent_sector_bitmap);
996 /* Save the initial failed bitmap for later repair and report usage. */
997 stripe->init_error_bitmap = stripe->error_bitmap;
998 stripe->init_nr_io_errors = bitmap_weight(&stripe->io_error_bitmap,
1000 stripe->init_nr_csum_errors = bitmap_weight(&stripe->csum_error_bitmap,
1001 stripe->nr_sectors);
1002 stripe->init_nr_meta_errors = bitmap_weight(&stripe->meta_error_bitmap,
1003 stripe->nr_sectors);
1005 if (bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors))
1009 * Try all remaining mirrors.
1011 * Here we still try to read as large block as possible, as this is
1012 * faster and we have extra safety nets to rely on.
1014 for (mirror = calc_next_mirror(stripe->mirror_num, num_copies);
1015 mirror != stripe->mirror_num;
1016 mirror = calc_next_mirror(mirror, num_copies)) {
1017 const unsigned long old_error_bitmap = stripe->error_bitmap;
1019 scrub_stripe_submit_repair_read(stripe, mirror,
1020 BTRFS_STRIPE_LEN, false);
1021 wait_scrub_stripe_io(stripe);
1022 scrub_verify_one_stripe(stripe, old_error_bitmap);
1023 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1028 * Last safety net, try re-checking all mirrors, including the failed
1029 * one, sector-by-sector.
1031 * As if one sector failed the drive's internal csum, the whole read
1032 * containing the offending sector would be marked as error.
1033 * Thus here we do sector-by-sector read.
1035 * This can be slow, thus we only try it as the last resort.
1038 for (i = 0, mirror = stripe->mirror_num;
1040 i++, mirror = calc_next_mirror(mirror, num_copies)) {
1041 const unsigned long old_error_bitmap = stripe->error_bitmap;
1043 scrub_stripe_submit_repair_read(stripe, mirror,
1044 fs_info->sectorsize, true);
1045 wait_scrub_stripe_io(stripe);
1046 scrub_verify_one_stripe(stripe, old_error_bitmap);
1047 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1051 scrub_stripe_report_errors(stripe->sctx, stripe);
1052 set_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state);
1053 wake_up(&stripe->repair_wait);
1056 static void scrub_read_endio(struct btrfs_bio *bbio)
1058 struct scrub_stripe *stripe = bbio->private;
1060 if (bbio->bio.bi_status) {
1061 bitmap_set(&stripe->io_error_bitmap, 0, stripe->nr_sectors);
1062 bitmap_set(&stripe->error_bitmap, 0, stripe->nr_sectors);
1064 bitmap_clear(&stripe->io_error_bitmap, 0, stripe->nr_sectors);
1066 bio_put(&bbio->bio);
1067 if (atomic_dec_and_test(&stripe->pending_io)) {
1068 wake_up(&stripe->io_wait);
1069 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1070 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1074 static void scrub_write_endio(struct btrfs_bio *bbio)
1076 struct scrub_stripe *stripe = bbio->private;
1077 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1078 struct bio_vec *bvec;
1079 int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1083 bio_for_each_bvec_all(bvec, &bbio->bio, i)
1084 bio_size += bvec->bv_len;
1086 if (bbio->bio.bi_status) {
1087 unsigned long flags;
1089 spin_lock_irqsave(&stripe->write_error_lock, flags);
1090 bitmap_set(&stripe->write_error_bitmap, sector_nr,
1091 bio_size >> fs_info->sectorsize_bits);
1092 spin_unlock_irqrestore(&stripe->write_error_lock, flags);
1094 bio_put(&bbio->bio);
1096 if (atomic_dec_and_test(&stripe->pending_io))
1097 wake_up(&stripe->io_wait);
1100 static void scrub_submit_write_bio(struct scrub_ctx *sctx,
1101 struct scrub_stripe *stripe,
1102 struct btrfs_bio *bbio, bool dev_replace)
1104 struct btrfs_fs_info *fs_info = sctx->fs_info;
1105 u32 bio_len = bbio->bio.bi_iter.bi_size;
1106 u32 bio_off = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT) -
1109 fill_writer_pointer_gap(sctx, stripe->physical + bio_off);
1110 atomic_inc(&stripe->pending_io);
1111 btrfs_submit_repair_write(bbio, stripe->mirror_num, dev_replace);
1112 if (!btrfs_is_zoned(fs_info))
1115 * For zoned writeback, queue depth must be 1, thus we must wait for
1116 * the write to finish before the next write.
1118 wait_scrub_stripe_io(stripe);
1121 * And also need to update the write pointer if write finished
1124 if (!test_bit(bio_off >> fs_info->sectorsize_bits,
1125 &stripe->write_error_bitmap))
1126 sctx->write_pointer += bio_len;
1130 * Submit the write bio(s) for the sectors specified by @write_bitmap.
1132 * Here we utilize btrfs_submit_repair_write(), which has some extra benefits:
1134 * - Only needs logical bytenr and mirror_num
1135 * Just like the scrub read path
1137 * - Would only result in writes to the specified mirror
1138 * Unlike the regular writeback path, which would write back to all stripes
1140 * - Handle dev-replace and read-repair writeback differently
1142 static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
1143 unsigned long write_bitmap, bool dev_replace)
1145 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1146 struct btrfs_bio *bbio = NULL;
1149 for_each_set_bit(sector_nr, &write_bitmap, stripe->nr_sectors) {
1150 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
1151 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
1154 /* We should only writeback sectors covered by an extent. */
1155 ASSERT(test_bit(sector_nr, &stripe->extent_sector_bitmap));
1157 /* Cannot merge with previous sector, submit the current one. */
1158 if (bbio && sector_nr && !test_bit(sector_nr - 1, &write_bitmap)) {
1159 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1163 bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_WRITE,
1164 fs_info, scrub_write_endio, stripe);
1165 bbio->bio.bi_iter.bi_sector = (stripe->logical +
1166 (sector_nr << fs_info->sectorsize_bits)) >>
1169 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1170 ASSERT(ret == fs_info->sectorsize);
1173 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1177 * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1178 * second. Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
1180 static void scrub_throttle_dev_io(struct scrub_ctx *sctx, struct btrfs_device *device,
1181 unsigned int bio_size)
1183 const int time_slice = 1000;
1189 bwlimit = READ_ONCE(device->scrub_speed_max);
1194 * Slice is divided into intervals when the IO is submitted, adjust by
1195 * bwlimit and maximum of 64 intervals.
1197 div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
1198 div = min_t(u32, 64, div);
1200 /* Start new epoch, set deadline */
1202 if (sctx->throttle_deadline == 0) {
1203 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
1204 sctx->throttle_sent = 0;
1207 /* Still in the time to send? */
1208 if (ktime_before(now, sctx->throttle_deadline)) {
1209 /* If current bio is within the limit, send it */
1210 sctx->throttle_sent += bio_size;
1211 if (sctx->throttle_sent <= div_u64(bwlimit, div))
1214 /* We're over the limit, sleep until the rest of the slice */
1215 delta = ktime_ms_delta(sctx->throttle_deadline, now);
1217 /* New request after deadline, start new epoch */
1224 timeout = div_u64(delta * HZ, 1000);
1225 schedule_timeout_interruptible(timeout);
1228 /* Next call will start the deadline period */
1229 sctx->throttle_deadline = 0;
1233 * Given a physical address, this will calculate it's
1234 * logical offset. if this is a parity stripe, it will return
1235 * the most left data stripe's logical offset.
1237 * return 0 if it is a data stripe, 1 means parity stripe.
1239 static int get_raid56_logic_offset(u64 physical, int num,
1240 struct map_lookup *map, u64 *offset,
1246 const int data_stripes = nr_data_stripes(map);
1248 last_offset = (physical - map->stripes[num].physical) * data_stripes;
1250 *stripe_start = last_offset;
1252 *offset = last_offset;
1253 for (i = 0; i < data_stripes; i++) {
1258 *offset = last_offset + btrfs_stripe_nr_to_offset(i);
1260 stripe_nr = (u32)(*offset >> BTRFS_STRIPE_LEN_SHIFT) / data_stripes;
1262 /* Work out the disk rotation on this stripe-set */
1263 rot = stripe_nr % map->num_stripes;
1264 stripe_nr /= map->num_stripes;
1265 /* calculate which stripe this data locates */
1267 stripe_index = rot % map->num_stripes;
1268 if (stripe_index == num)
1270 if (stripe_index < num)
1273 *offset = last_offset + btrfs_stripe_nr_to_offset(j);
1278 * Return 0 if the extent item range covers any byte of the range.
1279 * Return <0 if the extent item is before @search_start.
1280 * Return >0 if the extent item is after @start_start + @search_len.
1282 static int compare_extent_item_range(struct btrfs_path *path,
1283 u64 search_start, u64 search_len)
1285 struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
1287 struct btrfs_key key;
1289 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1290 ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
1291 key.type == BTRFS_METADATA_ITEM_KEY);
1292 if (key.type == BTRFS_METADATA_ITEM_KEY)
1293 len = fs_info->nodesize;
1297 if (key.objectid + len <= search_start)
1299 if (key.objectid >= search_start + search_len)
1305 * Locate one extent item which covers any byte in range
1306 * [@search_start, @search_start + @search_length)
1308 * If the path is not initialized, we will initialize the search by doing
1309 * a btrfs_search_slot().
1310 * If the path is already initialized, we will use the path as the initial
1311 * slot, to avoid duplicated btrfs_search_slot() calls.
1313 * NOTE: If an extent item starts before @search_start, we will still
1314 * return the extent item. This is for data extent crossing stripe boundary.
1316 * Return 0 if we found such extent item, and @path will point to the extent item.
1317 * Return >0 if no such extent item can be found, and @path will be released.
1318 * Return <0 if hit fatal error, and @path will be released.
1320 static int find_first_extent_item(struct btrfs_root *extent_root,
1321 struct btrfs_path *path,
1322 u64 search_start, u64 search_len)
1324 struct btrfs_fs_info *fs_info = extent_root->fs_info;
1325 struct btrfs_key key;
1328 /* Continue using the existing path */
1330 goto search_forward;
1332 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1333 key.type = BTRFS_METADATA_ITEM_KEY;
1335 key.type = BTRFS_EXTENT_ITEM_KEY;
1336 key.objectid = search_start;
1337 key.offset = (u64)-1;
1339 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1345 * Here we intentionally pass 0 as @min_objectid, as there could be
1346 * an extent item starting before @search_start.
1348 ret = btrfs_previous_extent_item(extent_root, path, 0);
1352 * No matter whether we have found an extent item, the next loop will
1353 * properly do every check on the key.
1357 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1358 if (key.objectid >= search_start + search_len)
1360 if (key.type != BTRFS_METADATA_ITEM_KEY &&
1361 key.type != BTRFS_EXTENT_ITEM_KEY)
1364 ret = compare_extent_item_range(path, search_start, search_len);
1371 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
1372 ret = btrfs_next_leaf(extent_root, path);
1374 /* Either no more item or fatal error */
1375 btrfs_release_path(path);
1380 btrfs_release_path(path);
1384 static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
1385 u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
1387 struct btrfs_key key;
1388 struct btrfs_extent_item *ei;
1390 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1391 ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
1392 key.type == BTRFS_EXTENT_ITEM_KEY);
1393 *extent_start_ret = key.objectid;
1394 if (key.type == BTRFS_METADATA_ITEM_KEY)
1395 *size_ret = path->nodes[0]->fs_info->nodesize;
1397 *size_ret = key.offset;
1398 ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
1399 *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
1400 *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
1403 static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
1404 u64 physical, u64 physical_end)
1406 struct btrfs_fs_info *fs_info = sctx->fs_info;
1409 if (!btrfs_is_zoned(fs_info))
1412 mutex_lock(&sctx->wr_lock);
1413 if (sctx->write_pointer < physical_end) {
1414 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
1416 sctx->write_pointer);
1419 "zoned: failed to recover write pointer");
1421 mutex_unlock(&sctx->wr_lock);
1422 btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
1427 static void fill_one_extent_info(struct btrfs_fs_info *fs_info,
1428 struct scrub_stripe *stripe,
1429 u64 extent_start, u64 extent_len,
1430 u64 extent_flags, u64 extent_gen)
1432 for (u64 cur_logical = max(stripe->logical, extent_start);
1433 cur_logical < min(stripe->logical + BTRFS_STRIPE_LEN,
1434 extent_start + extent_len);
1435 cur_logical += fs_info->sectorsize) {
1436 const int nr_sector = (cur_logical - stripe->logical) >>
1437 fs_info->sectorsize_bits;
1438 struct scrub_sector_verification *sector =
1439 &stripe->sectors[nr_sector];
1441 set_bit(nr_sector, &stripe->extent_sector_bitmap);
1442 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1443 sector->is_metadata = true;
1444 sector->generation = extent_gen;
1449 static void scrub_stripe_reset_bitmaps(struct scrub_stripe *stripe)
1451 stripe->extent_sector_bitmap = 0;
1452 stripe->init_error_bitmap = 0;
1453 stripe->init_nr_io_errors = 0;
1454 stripe->init_nr_csum_errors = 0;
1455 stripe->init_nr_meta_errors = 0;
1456 stripe->error_bitmap = 0;
1457 stripe->io_error_bitmap = 0;
1458 stripe->csum_error_bitmap = 0;
1459 stripe->meta_error_bitmap = 0;
1463 * Locate one stripe which has at least one extent in its range.
1465 * Return 0 if found such stripe, and store its info into @stripe.
1466 * Return >0 if there is no such stripe in the specified range.
1467 * Return <0 for error.
1469 static int scrub_find_fill_first_stripe(struct btrfs_block_group *bg,
1470 struct btrfs_device *dev, u64 physical,
1471 int mirror_num, u64 logical_start,
1473 struct scrub_stripe *stripe)
1475 struct btrfs_fs_info *fs_info = bg->fs_info;
1476 struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bg->start);
1477 struct btrfs_root *csum_root = btrfs_csum_root(fs_info, bg->start);
1478 const u64 logical_end = logical_start + logical_len;
1479 struct btrfs_path path = { 0 };
1480 u64 cur_logical = logical_start;
1488 memset(stripe->sectors, 0, sizeof(struct scrub_sector_verification) *
1489 stripe->nr_sectors);
1490 scrub_stripe_reset_bitmaps(stripe);
1492 /* The range must be inside the bg. */
1493 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1495 path.search_commit_root = 1;
1496 path.skip_locking = 1;
1498 ret = find_first_extent_item(extent_root, &path, logical_start, logical_len);
1499 /* Either error or not found. */
1502 get_extent_info(&path, &extent_start, &extent_len, &extent_flags, &extent_gen);
1503 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1504 stripe->nr_meta_extents++;
1505 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1506 stripe->nr_data_extents++;
1507 cur_logical = max(extent_start, cur_logical);
1510 * Round down to stripe boundary.
1512 * The extra calculation against bg->start is to handle block groups
1513 * whose logical bytenr is not BTRFS_STRIPE_LEN aligned.
1515 stripe->logical = round_down(cur_logical - bg->start, BTRFS_STRIPE_LEN) +
1517 stripe->physical = physical + stripe->logical - logical_start;
1520 stripe->mirror_num = mirror_num;
1521 stripe_end = stripe->logical + BTRFS_STRIPE_LEN - 1;
1523 /* Fill the first extent info into stripe->sectors[] array. */
1524 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1525 extent_flags, extent_gen);
1526 cur_logical = extent_start + extent_len;
1528 /* Fill the extent info for the remaining sectors. */
1529 while (cur_logical <= stripe_end) {
1530 ret = find_first_extent_item(extent_root, &path, cur_logical,
1531 stripe_end - cur_logical + 1);
1538 get_extent_info(&path, &extent_start, &extent_len,
1539 &extent_flags, &extent_gen);
1540 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1541 stripe->nr_meta_extents++;
1542 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1543 stripe->nr_data_extents++;
1544 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1545 extent_flags, extent_gen);
1546 cur_logical = extent_start + extent_len;
1549 /* Now fill the data csum. */
1550 if (bg->flags & BTRFS_BLOCK_GROUP_DATA) {
1552 unsigned long csum_bitmap = 0;
1554 /* Csum space should have already been allocated. */
1555 ASSERT(stripe->csums);
1558 * Our csum bitmap should be large enough, as BTRFS_STRIPE_LEN
1559 * should contain at most 16 sectors.
1561 ASSERT(BITS_PER_LONG >= BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1563 ret = btrfs_lookup_csums_bitmap(csum_root, stripe->logical,
1564 stripe_end, stripe->csums,
1565 &csum_bitmap, true);
1571 for_each_set_bit(sector_nr, &csum_bitmap, stripe->nr_sectors) {
1572 stripe->sectors[sector_nr].csum = stripe->csums +
1573 sector_nr * fs_info->csum_size;
1576 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1578 btrfs_release_path(&path);
1582 static void scrub_reset_stripe(struct scrub_stripe *stripe)
1584 scrub_stripe_reset_bitmaps(stripe);
1586 stripe->nr_meta_extents = 0;
1587 stripe->nr_data_extents = 0;
1590 for (int i = 0; i < stripe->nr_sectors; i++) {
1591 stripe->sectors[i].is_metadata = false;
1592 stripe->sectors[i].csum = NULL;
1593 stripe->sectors[i].generation = 0;
1597 static void scrub_submit_initial_read(struct scrub_ctx *sctx,
1598 struct scrub_stripe *stripe)
1600 struct btrfs_fs_info *fs_info = sctx->fs_info;
1601 struct btrfs_bio *bbio;
1602 int mirror = stripe->mirror_num;
1605 ASSERT(stripe->mirror_num > 0);
1606 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1608 bbio = btrfs_bio_alloc(SCRUB_STRIPE_PAGES, REQ_OP_READ, fs_info,
1609 scrub_read_endio, stripe);
1611 /* Read the whole stripe. */
1612 bbio->bio.bi_iter.bi_sector = stripe->logical >> SECTOR_SHIFT;
1613 for (int i = 0; i < BTRFS_STRIPE_LEN >> PAGE_SHIFT; i++) {
1616 ret = bio_add_page(&bbio->bio, stripe->pages[i], PAGE_SIZE, 0);
1617 /* We should have allocated enough bio vectors. */
1618 ASSERT(ret == PAGE_SIZE);
1620 atomic_inc(&stripe->pending_io);
1623 * For dev-replace, either user asks to avoid the source dev, or
1624 * the device is missing, we try the next mirror instead.
1626 if (sctx->is_dev_replace &&
1627 (fs_info->dev_replace.cont_reading_from_srcdev_mode ==
1628 BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID ||
1629 !stripe->dev->bdev)) {
1630 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1631 stripe->bg->length);
1633 mirror = calc_next_mirror(mirror, num_copies);
1635 btrfs_submit_bio(bbio, mirror);
1638 static bool stripe_has_metadata_error(struct scrub_stripe *stripe)
1642 for_each_set_bit(i, &stripe->error_bitmap, stripe->nr_sectors) {
1643 if (stripe->sectors[i].is_metadata) {
1644 struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1647 "stripe %llu has unrepaired metadata sector at %llu",
1649 stripe->logical + (i << fs_info->sectorsize_bits));
1656 static int flush_scrub_stripes(struct scrub_ctx *sctx)
1658 struct btrfs_fs_info *fs_info = sctx->fs_info;
1659 struct scrub_stripe *stripe;
1660 const int nr_stripes = sctx->cur_stripe;
1666 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &sctx->stripes[0].state));
1668 scrub_throttle_dev_io(sctx, sctx->stripes[0].dev,
1669 btrfs_stripe_nr_to_offset(nr_stripes));
1670 for (int i = 0; i < nr_stripes; i++) {
1671 stripe = &sctx->stripes[i];
1672 scrub_submit_initial_read(sctx, stripe);
1675 for (int i = 0; i < nr_stripes; i++) {
1676 stripe = &sctx->stripes[i];
1678 wait_event(stripe->repair_wait,
1679 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1683 * Submit the repaired sectors. For zoned case, we cannot do repair
1684 * in-place, but queue the bg to be relocated.
1686 if (btrfs_is_zoned(fs_info)) {
1687 for (int i = 0; i < nr_stripes; i++) {
1688 stripe = &sctx->stripes[i];
1690 if (!bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors)) {
1691 btrfs_repair_one_zone(fs_info,
1692 sctx->stripes[0].bg->start);
1696 } else if (!sctx->readonly) {
1697 for (int i = 0; i < nr_stripes; i++) {
1698 unsigned long repaired;
1700 stripe = &sctx->stripes[i];
1702 bitmap_andnot(&repaired, &stripe->init_error_bitmap,
1703 &stripe->error_bitmap, stripe->nr_sectors);
1704 scrub_write_sectors(sctx, stripe, repaired, false);
1708 /* Submit for dev-replace. */
1709 if (sctx->is_dev_replace) {
1711 * For dev-replace, if we know there is something wrong with
1712 * metadata, we should immedately abort.
1714 for (int i = 0; i < nr_stripes; i++) {
1715 if (stripe_has_metadata_error(&sctx->stripes[i])) {
1720 for (int i = 0; i < nr_stripes; i++) {
1723 stripe = &sctx->stripes[i];
1725 ASSERT(stripe->dev == fs_info->dev_replace.srcdev);
1727 bitmap_andnot(&good, &stripe->extent_sector_bitmap,
1728 &stripe->error_bitmap, stripe->nr_sectors);
1729 scrub_write_sectors(sctx, stripe, good, true);
1733 /* Wait for the above writebacks to finish. */
1734 for (int i = 0; i < nr_stripes; i++) {
1735 stripe = &sctx->stripes[i];
1737 wait_scrub_stripe_io(stripe);
1738 scrub_reset_stripe(stripe);
1741 sctx->cur_stripe = 0;
1745 static void raid56_scrub_wait_endio(struct bio *bio)
1747 complete(bio->bi_private);
1750 static int queue_scrub_stripe(struct scrub_ctx *sctx, struct btrfs_block_group *bg,
1751 struct btrfs_device *dev, int mirror_num,
1752 u64 logical, u32 length, u64 physical)
1754 struct scrub_stripe *stripe;
1757 /* No available slot, submit all stripes and wait for them. */
1758 if (sctx->cur_stripe >= SCRUB_STRIPES_PER_SCTX) {
1759 ret = flush_scrub_stripes(sctx);
1764 stripe = &sctx->stripes[sctx->cur_stripe];
1766 /* We can queue one stripe using the remaining slot. */
1767 scrub_reset_stripe(stripe);
1768 ret = scrub_find_fill_first_stripe(bg, dev, physical, mirror_num,
1769 logical, length, stripe);
1770 /* Either >0 as no more extents or <0 for error. */
1777 static int scrub_raid56_parity_stripe(struct scrub_ctx *sctx,
1778 struct btrfs_device *scrub_dev,
1779 struct btrfs_block_group *bg,
1780 struct map_lookup *map,
1781 u64 full_stripe_start)
1783 DECLARE_COMPLETION_ONSTACK(io_done);
1784 struct btrfs_fs_info *fs_info = sctx->fs_info;
1785 struct btrfs_raid_bio *rbio;
1786 struct btrfs_io_context *bioc = NULL;
1788 struct scrub_stripe *stripe;
1789 bool all_empty = true;
1790 const int data_stripes = nr_data_stripes(map);
1791 unsigned long extent_bitmap = 0;
1792 u64 length = btrfs_stripe_nr_to_offset(data_stripes);
1795 ASSERT(sctx->raid56_data_stripes);
1797 for (int i = 0; i < data_stripes; i++) {
1802 stripe = &sctx->raid56_data_stripes[i];
1803 rot = div_u64(full_stripe_start - bg->start,
1804 data_stripes) >> BTRFS_STRIPE_LEN_SHIFT;
1805 stripe_index = (i + rot) % map->num_stripes;
1806 physical = map->stripes[stripe_index].physical +
1807 btrfs_stripe_nr_to_offset(rot);
1809 scrub_reset_stripe(stripe);
1810 set_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state);
1811 ret = scrub_find_fill_first_stripe(bg,
1812 map->stripes[stripe_index].dev, physical, 1,
1813 full_stripe_start + btrfs_stripe_nr_to_offset(i),
1814 BTRFS_STRIPE_LEN, stripe);
1818 * No extent in this data stripe, need to manually mark them
1819 * initialized to make later read submission happy.
1822 stripe->logical = full_stripe_start +
1823 btrfs_stripe_nr_to_offset(i);
1824 stripe->dev = map->stripes[stripe_index].dev;
1825 stripe->mirror_num = 1;
1826 set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1830 /* Check if all data stripes are empty. */
1831 for (int i = 0; i < data_stripes; i++) {
1832 stripe = &sctx->raid56_data_stripes[i];
1833 if (!bitmap_empty(&stripe->extent_sector_bitmap, stripe->nr_sectors)) {
1843 for (int i = 0; i < data_stripes; i++) {
1844 stripe = &sctx->raid56_data_stripes[i];
1845 scrub_submit_initial_read(sctx, stripe);
1847 for (int i = 0; i < data_stripes; i++) {
1848 stripe = &sctx->raid56_data_stripes[i];
1850 wait_event(stripe->repair_wait,
1851 test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1853 /* For now, no zoned support for RAID56. */
1854 ASSERT(!btrfs_is_zoned(sctx->fs_info));
1856 /* Writeback for the repaired sectors. */
1857 for (int i = 0; i < data_stripes; i++) {
1858 unsigned long repaired;
1860 stripe = &sctx->raid56_data_stripes[i];
1862 bitmap_andnot(&repaired, &stripe->init_error_bitmap,
1863 &stripe->error_bitmap, stripe->nr_sectors);
1864 scrub_write_sectors(sctx, stripe, repaired, false);
1867 /* Wait for the above writebacks to finish. */
1868 for (int i = 0; i < data_stripes; i++) {
1869 stripe = &sctx->raid56_data_stripes[i];
1871 wait_scrub_stripe_io(stripe);
1875 * Now all data stripes are properly verified. Check if we have any
1876 * unrepaired, if so abort immediately or we could further corrupt the
1879 * During the loop, also populate extent_bitmap.
1881 for (int i = 0; i < data_stripes; i++) {
1882 unsigned long error;
1884 stripe = &sctx->raid56_data_stripes[i];
1887 * We should only check the errors where there is an extent.
1888 * As we may hit an empty data stripe while it's missing.
1890 bitmap_and(&error, &stripe->error_bitmap,
1891 &stripe->extent_sector_bitmap, stripe->nr_sectors);
1892 if (!bitmap_empty(&error, stripe->nr_sectors)) {
1894 "unrepaired sectors detected, full stripe %llu data stripe %u errors %*pbl",
1895 full_stripe_start, i, stripe->nr_sectors,
1900 bitmap_or(&extent_bitmap, &extent_bitmap,
1901 &stripe->extent_sector_bitmap, stripe->nr_sectors);
1904 /* Now we can check and regenerate the P/Q stripe. */
1905 bio = bio_alloc(NULL, 1, REQ_OP_READ, GFP_NOFS);
1906 bio->bi_iter.bi_sector = full_stripe_start >> SECTOR_SHIFT;
1907 bio->bi_private = &io_done;
1908 bio->bi_end_io = raid56_scrub_wait_endio;
1910 btrfs_bio_counter_inc_blocked(fs_info);
1911 ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, full_stripe_start,
1912 &length, &bioc, NULL, NULL, 1);
1914 btrfs_put_bioc(bioc);
1915 btrfs_bio_counter_dec(fs_info);
1918 rbio = raid56_parity_alloc_scrub_rbio(bio, bioc, scrub_dev, &extent_bitmap,
1919 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1920 btrfs_put_bioc(bioc);
1923 btrfs_bio_counter_dec(fs_info);
1926 /* Use the recovered stripes as cache to avoid read them from disk again. */
1927 for (int i = 0; i < data_stripes; i++) {
1928 stripe = &sctx->raid56_data_stripes[i];
1930 raid56_parity_cache_data_pages(rbio, stripe->pages,
1931 full_stripe_start + (i << BTRFS_STRIPE_LEN_SHIFT));
1933 raid56_parity_submit_scrub_rbio(rbio);
1934 wait_for_completion_io(&io_done);
1935 ret = blk_status_to_errno(bio->bi_status);
1937 btrfs_bio_counter_dec(fs_info);
1944 * Scrub one range which can only has simple mirror based profile.
1945 * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
1948 * Since we may need to handle a subset of block group, we need @logical_start
1949 * and @logical_length parameter.
1951 static int scrub_simple_mirror(struct scrub_ctx *sctx,
1952 struct btrfs_block_group *bg,
1953 struct map_lookup *map,
1954 u64 logical_start, u64 logical_length,
1955 struct btrfs_device *device,
1956 u64 physical, int mirror_num)
1958 struct btrfs_fs_info *fs_info = sctx->fs_info;
1959 const u64 logical_end = logical_start + logical_length;
1960 /* An artificial limit, inherit from old scrub behavior */
1961 struct btrfs_path path = { 0 };
1962 u64 cur_logical = logical_start;
1965 /* The range must be inside the bg */
1966 ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1968 path.search_commit_root = 1;
1969 path.skip_locking = 1;
1970 /* Go through each extent items inside the logical range */
1971 while (cur_logical < logical_end) {
1972 u64 cur_physical = physical + cur_logical - logical_start;
1975 if (atomic_read(&fs_info->scrub_cancel_req) ||
1976 atomic_read(&sctx->cancel_req)) {
1981 if (atomic_read(&fs_info->scrub_pause_req)) {
1982 /* Push queued extents */
1983 scrub_blocked_if_needed(fs_info);
1985 /* Block group removed? */
1986 spin_lock(&bg->lock);
1987 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
1988 spin_unlock(&bg->lock);
1992 spin_unlock(&bg->lock);
1994 ret = queue_scrub_stripe(sctx, bg, device, mirror_num,
1995 cur_logical, logical_end - cur_logical,
1998 /* No more extent, just update the accounting */
1999 sctx->stat.last_physical = physical + logical_length;
2006 ASSERT(sctx->cur_stripe > 0);
2007 cur_logical = sctx->stripes[sctx->cur_stripe - 1].logical
2010 /* Don't hold CPU for too long time */
2013 btrfs_release_path(&path);
2017 /* Calculate the full stripe length for simple stripe based profiles */
2018 static u64 simple_stripe_full_stripe_len(const struct map_lookup *map)
2020 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2021 BTRFS_BLOCK_GROUP_RAID10));
2023 return btrfs_stripe_nr_to_offset(map->num_stripes / map->sub_stripes);
2026 /* Get the logical bytenr for the stripe */
2027 static u64 simple_stripe_get_logical(struct map_lookup *map,
2028 struct btrfs_block_group *bg,
2031 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2032 BTRFS_BLOCK_GROUP_RAID10));
2033 ASSERT(stripe_index < map->num_stripes);
2036 * (stripe_index / sub_stripes) gives how many data stripes we need to
2039 return btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes) +
2043 /* Get the mirror number for the stripe */
2044 static int simple_stripe_mirror_num(struct map_lookup *map, int stripe_index)
2046 ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2047 BTRFS_BLOCK_GROUP_RAID10));
2048 ASSERT(stripe_index < map->num_stripes);
2050 /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
2051 return stripe_index % map->sub_stripes + 1;
2054 static int scrub_simple_stripe(struct scrub_ctx *sctx,
2055 struct btrfs_block_group *bg,
2056 struct map_lookup *map,
2057 struct btrfs_device *device,
2060 const u64 logical_increment = simple_stripe_full_stripe_len(map);
2061 const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
2062 const u64 orig_physical = map->stripes[stripe_index].physical;
2063 const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
2064 u64 cur_logical = orig_logical;
2065 u64 cur_physical = orig_physical;
2068 while (cur_logical < bg->start + bg->length) {
2070 * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
2071 * just RAID1, so we can reuse scrub_simple_mirror() to scrub
2074 ret = scrub_simple_mirror(sctx, bg, map, cur_logical,
2075 BTRFS_STRIPE_LEN, device, cur_physical,
2079 /* Skip to next stripe which belongs to the target device */
2080 cur_logical += logical_increment;
2081 /* For physical offset, we just go to next stripe */
2082 cur_physical += BTRFS_STRIPE_LEN;
2087 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2088 struct btrfs_block_group *bg,
2089 struct extent_map *em,
2090 struct btrfs_device *scrub_dev,
2093 struct btrfs_fs_info *fs_info = sctx->fs_info;
2094 struct map_lookup *map = em->map_lookup;
2095 const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
2096 const u64 chunk_logical = bg->start;
2099 u64 physical = map->stripes[stripe_index].physical;
2100 const u64 dev_stripe_len = btrfs_calc_stripe_length(em);
2101 const u64 physical_end = physical + dev_stripe_len;
2104 /* The logical increment after finishing one stripe */
2106 /* Offset inside the chunk */
2111 scrub_blocked_if_needed(fs_info);
2113 if (sctx->is_dev_replace &&
2114 btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
2115 mutex_lock(&sctx->wr_lock);
2116 sctx->write_pointer = physical;
2117 mutex_unlock(&sctx->wr_lock);
2120 /* Prepare the extra data stripes used by RAID56. */
2121 if (profile & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2122 ASSERT(sctx->raid56_data_stripes == NULL);
2124 sctx->raid56_data_stripes = kcalloc(nr_data_stripes(map),
2125 sizeof(struct scrub_stripe),
2127 if (!sctx->raid56_data_stripes) {
2131 for (int i = 0; i < nr_data_stripes(map); i++) {
2132 ret = init_scrub_stripe(fs_info,
2133 &sctx->raid56_data_stripes[i]);
2136 sctx->raid56_data_stripes[i].bg = bg;
2137 sctx->raid56_data_stripes[i].sctx = sctx;
2141 * There used to be a big double loop to handle all profiles using the
2142 * same routine, which grows larger and more gross over time.
2144 * So here we handle each profile differently, so simpler profiles
2145 * have simpler scrubbing function.
2147 if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
2148 BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2150 * Above check rules out all complex profile, the remaining
2151 * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
2152 * mirrored duplication without stripe.
2154 * Only @physical and @mirror_num needs to calculated using
2157 ret = scrub_simple_mirror(sctx, bg, map, bg->start, bg->length,
2158 scrub_dev, map->stripes[stripe_index].physical,
2163 if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
2164 ret = scrub_simple_stripe(sctx, bg, map, scrub_dev, stripe_index);
2165 offset = btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes);
2169 /* Only RAID56 goes through the old code */
2170 ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
2173 /* Calculate the logical end of the stripe */
2174 get_raid56_logic_offset(physical_end, stripe_index,
2175 map, &logic_end, NULL);
2176 logic_end += chunk_logical;
2178 /* Initialize @offset in case we need to go to out: label */
2179 get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
2180 increment = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2183 * Due to the rotation, for RAID56 it's better to iterate each stripe
2184 * using their physical offset.
2186 while (physical < physical_end) {
2187 ret = get_raid56_logic_offset(physical, stripe_index, map,
2188 &logical, &stripe_logical);
2189 logical += chunk_logical;
2191 /* it is parity strip */
2192 stripe_logical += chunk_logical;
2193 ret = scrub_raid56_parity_stripe(sctx, scrub_dev, bg,
2194 map, stripe_logical);
2201 * Now we're at a data stripe, scrub each extents in the range.
2203 * At this stage, if we ignore the repair part, inside each data
2204 * stripe it is no different than SINGLE profile.
2205 * We can reuse scrub_simple_mirror() here, as the repair part
2206 * is still based on @mirror_num.
2208 ret = scrub_simple_mirror(sctx, bg, map, logical, BTRFS_STRIPE_LEN,
2209 scrub_dev, physical, 1);
2213 logical += increment;
2214 physical += BTRFS_STRIPE_LEN;
2215 spin_lock(&sctx->stat_lock);
2217 sctx->stat.last_physical =
2218 map->stripes[stripe_index].physical + dev_stripe_len;
2220 sctx->stat.last_physical = physical;
2221 spin_unlock(&sctx->stat_lock);
2226 ret2 = flush_scrub_stripes(sctx);
2229 if (sctx->raid56_data_stripes) {
2230 for (int i = 0; i < nr_data_stripes(map); i++)
2231 release_scrub_stripe(&sctx->raid56_data_stripes[i]);
2232 kfree(sctx->raid56_data_stripes);
2233 sctx->raid56_data_stripes = NULL;
2236 if (sctx->is_dev_replace && ret >= 0) {
2239 ret2 = sync_write_pointer_for_zoned(sctx,
2240 chunk_logical + offset,
2241 map->stripes[stripe_index].physical,
2247 return ret < 0 ? ret : 0;
2250 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2251 struct btrfs_block_group *bg,
2252 struct btrfs_device *scrub_dev,
2256 struct btrfs_fs_info *fs_info = sctx->fs_info;
2257 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2258 struct map_lookup *map;
2259 struct extent_map *em;
2263 read_lock(&map_tree->lock);
2264 em = lookup_extent_mapping(map_tree, bg->start, bg->length);
2265 read_unlock(&map_tree->lock);
2269 * Might have been an unused block group deleted by the cleaner
2270 * kthread or relocation.
2272 spin_lock(&bg->lock);
2273 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
2275 spin_unlock(&bg->lock);
2279 if (em->start != bg->start)
2281 if (em->len < dev_extent_len)
2284 map = em->map_lookup;
2285 for (i = 0; i < map->num_stripes; ++i) {
2286 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2287 map->stripes[i].physical == dev_offset) {
2288 ret = scrub_stripe(sctx, bg, em, scrub_dev, i);
2294 free_extent_map(em);
2299 static int finish_extent_writes_for_zoned(struct btrfs_root *root,
2300 struct btrfs_block_group *cache)
2302 struct btrfs_fs_info *fs_info = cache->fs_info;
2303 struct btrfs_trans_handle *trans;
2305 if (!btrfs_is_zoned(fs_info))
2308 btrfs_wait_block_group_reservations(cache);
2309 btrfs_wait_nocow_writers(cache);
2310 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
2312 trans = btrfs_join_transaction(root);
2314 return PTR_ERR(trans);
2315 return btrfs_commit_transaction(trans);
2318 static noinline_for_stack
2319 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2320 struct btrfs_device *scrub_dev, u64 start, u64 end)
2322 struct btrfs_dev_extent *dev_extent = NULL;
2323 struct btrfs_path *path;
2324 struct btrfs_fs_info *fs_info = sctx->fs_info;
2325 struct btrfs_root *root = fs_info->dev_root;
2330 struct extent_buffer *l;
2331 struct btrfs_key key;
2332 struct btrfs_key found_key;
2333 struct btrfs_block_group *cache;
2334 struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2336 path = btrfs_alloc_path();
2340 path->reada = READA_FORWARD;
2341 path->search_commit_root = 1;
2342 path->skip_locking = 1;
2344 key.objectid = scrub_dev->devid;
2346 key.type = BTRFS_DEV_EXTENT_KEY;
2351 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2355 if (path->slots[0] >=
2356 btrfs_header_nritems(path->nodes[0])) {
2357 ret = btrfs_next_leaf(root, path);
2370 slot = path->slots[0];
2372 btrfs_item_key_to_cpu(l, &found_key, slot);
2374 if (found_key.objectid != scrub_dev->devid)
2377 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2380 if (found_key.offset >= end)
2383 if (found_key.offset < key.offset)
2386 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2387 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
2389 if (found_key.offset + dev_extent_len <= start)
2392 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2395 * get a reference on the corresponding block group to prevent
2396 * the chunk from going away while we scrub it
2398 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2400 /* some chunks are removed but not committed to disk yet,
2401 * continue scrubbing */
2405 ASSERT(cache->start <= chunk_offset);
2407 * We are using the commit root to search for device extents, so
2408 * that means we could have found a device extent item from a
2409 * block group that was deleted in the current transaction. The
2410 * logical start offset of the deleted block group, stored at
2411 * @chunk_offset, might be part of the logical address range of
2412 * a new block group (which uses different physical extents).
2413 * In this case btrfs_lookup_block_group() has returned the new
2414 * block group, and its start address is less than @chunk_offset.
2416 * We skip such new block groups, because it's pointless to
2417 * process them, as we won't find their extents because we search
2418 * for them using the commit root of the extent tree. For a device
2419 * replace it's also fine to skip it, we won't miss copying them
2420 * to the target device because we have the write duplication
2421 * setup through the regular write path (by btrfs_map_block()),
2422 * and we have committed a transaction when we started the device
2423 * replace, right after setting up the device replace state.
2425 if (cache->start < chunk_offset) {
2426 btrfs_put_block_group(cache);
2430 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
2431 if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
2432 btrfs_put_block_group(cache);
2438 * Make sure that while we are scrubbing the corresponding block
2439 * group doesn't get its logical address and its device extents
2440 * reused for another block group, which can possibly be of a
2441 * different type and different profile. We do this to prevent
2442 * false error detections and crashes due to bogus attempts to
2445 spin_lock(&cache->lock);
2446 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
2447 spin_unlock(&cache->lock);
2448 btrfs_put_block_group(cache);
2451 btrfs_freeze_block_group(cache);
2452 spin_unlock(&cache->lock);
2455 * we need call btrfs_inc_block_group_ro() with scrubs_paused,
2456 * to avoid deadlock caused by:
2457 * btrfs_inc_block_group_ro()
2458 * -> btrfs_wait_for_commit()
2459 * -> btrfs_commit_transaction()
2460 * -> btrfs_scrub_pause()
2462 scrub_pause_on(fs_info);
2465 * Don't do chunk preallocation for scrub.
2467 * This is especially important for SYSTEM bgs, or we can hit
2468 * -EFBIG from btrfs_finish_chunk_alloc() like:
2469 * 1. The only SYSTEM bg is marked RO.
2470 * Since SYSTEM bg is small, that's pretty common.
2471 * 2. New SYSTEM bg will be allocated
2472 * Due to regular version will allocate new chunk.
2473 * 3. New SYSTEM bg is empty and will get cleaned up
2474 * Before cleanup really happens, it's marked RO again.
2475 * 4. Empty SYSTEM bg get scrubbed
2478 * This can easily boost the amount of SYSTEM chunks if cleaner
2479 * thread can't be triggered fast enough, and use up all space
2480 * of btrfs_super_block::sys_chunk_array
2482 * While for dev replace, we need to try our best to mark block
2483 * group RO, to prevent race between:
2484 * - Write duplication
2485 * Contains latest data
2487 * Contains data from commit tree
2489 * If target block group is not marked RO, nocow writes can
2490 * be overwritten by scrub copy, causing data corruption.
2491 * So for dev-replace, it's not allowed to continue if a block
2494 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
2495 if (!ret && sctx->is_dev_replace) {
2496 ret = finish_extent_writes_for_zoned(root, cache);
2498 btrfs_dec_block_group_ro(cache);
2499 scrub_pause_off(fs_info);
2500 btrfs_put_block_group(cache);
2507 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
2508 !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
2510 * btrfs_inc_block_group_ro return -ENOSPC when it
2511 * failed in creating new chunk for metadata.
2512 * It is not a problem for scrub, because
2513 * metadata are always cowed, and our scrub paused
2514 * commit_transactions.
2516 * For RAID56 chunks, we have to mark them read-only
2517 * for scrub, as later we would use our own cache
2518 * out of RAID56 realm.
2519 * Thus we want the RAID56 bg to be marked RO to
2520 * prevent RMW from screwing up out cache.
2523 } else if (ret == -ETXTBSY) {
2525 "skipping scrub of block group %llu due to active swapfile",
2527 scrub_pause_off(fs_info);
2532 "failed setting block group ro: %d", ret);
2533 btrfs_unfreeze_block_group(cache);
2534 btrfs_put_block_group(cache);
2535 scrub_pause_off(fs_info);
2540 * Now the target block is marked RO, wait for nocow writes to
2541 * finish before dev-replace.
2542 * COW is fine, as COW never overwrites extents in commit tree.
2544 if (sctx->is_dev_replace) {
2545 btrfs_wait_nocow_writers(cache);
2546 btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
2550 scrub_pause_off(fs_info);
2551 down_write(&dev_replace->rwsem);
2552 dev_replace->cursor_right = found_key.offset + dev_extent_len;
2553 dev_replace->cursor_left = found_key.offset;
2554 dev_replace->item_needs_writeback = 1;
2555 up_write(&dev_replace->rwsem);
2557 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
2559 if (sctx->is_dev_replace &&
2560 !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
2561 cache, found_key.offset))
2564 down_write(&dev_replace->rwsem);
2565 dev_replace->cursor_left = dev_replace->cursor_right;
2566 dev_replace->item_needs_writeback = 1;
2567 up_write(&dev_replace->rwsem);
2570 btrfs_dec_block_group_ro(cache);
2573 * We might have prevented the cleaner kthread from deleting
2574 * this block group if it was already unused because we raced
2575 * and set it to RO mode first. So add it back to the unused
2576 * list, otherwise it might not ever be deleted unless a manual
2577 * balance is triggered or it becomes used and unused again.
2579 spin_lock(&cache->lock);
2580 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
2581 !cache->ro && cache->reserved == 0 && cache->used == 0) {
2582 spin_unlock(&cache->lock);
2583 if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
2584 btrfs_discard_queue_work(&fs_info->discard_ctl,
2587 btrfs_mark_bg_unused(cache);
2589 spin_unlock(&cache->lock);
2592 btrfs_unfreeze_block_group(cache);
2593 btrfs_put_block_group(cache);
2596 if (sctx->is_dev_replace &&
2597 atomic64_read(&dev_replace->num_write_errors) > 0) {
2601 if (sctx->stat.malloc_errors > 0) {
2606 key.offset = found_key.offset + dev_extent_len;
2607 btrfs_release_path(path);
2610 btrfs_free_path(path);
2615 static int scrub_one_super(struct scrub_ctx *sctx, struct btrfs_device *dev,
2616 struct page *page, u64 physical, u64 generation)
2618 struct btrfs_fs_info *fs_info = sctx->fs_info;
2619 struct bio_vec bvec;
2621 struct btrfs_super_block *sb = page_address(page);
2624 bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_READ);
2625 bio.bi_iter.bi_sector = physical >> SECTOR_SHIFT;
2626 __bio_add_page(&bio, page, BTRFS_SUPER_INFO_SIZE, 0);
2627 ret = submit_bio_wait(&bio);
2632 ret = btrfs_check_super_csum(fs_info, sb);
2634 btrfs_err_rl(fs_info,
2635 "super block at physical %llu devid %llu has bad csum",
2636 physical, dev->devid);
2639 if (btrfs_super_generation(sb) != generation) {
2640 btrfs_err_rl(fs_info,
2641 "super block at physical %llu devid %llu has bad generation %llu expect %llu",
2642 physical, dev->devid,
2643 btrfs_super_generation(sb), generation);
2647 return btrfs_validate_super(fs_info, sb, -1);
2650 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2651 struct btrfs_device *scrub_dev)
2658 struct btrfs_fs_info *fs_info = sctx->fs_info;
2660 if (BTRFS_FS_ERROR(fs_info))
2663 page = alloc_page(GFP_KERNEL);
2665 spin_lock(&sctx->stat_lock);
2666 sctx->stat.malloc_errors++;
2667 spin_unlock(&sctx->stat_lock);
2671 /* Seed devices of a new filesystem has their own generation. */
2672 if (scrub_dev->fs_devices != fs_info->fs_devices)
2673 gen = scrub_dev->generation;
2675 gen = fs_info->last_trans_committed;
2677 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2678 bytenr = btrfs_sb_offset(i);
2679 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2680 scrub_dev->commit_total_bytes)
2682 if (!btrfs_check_super_location(scrub_dev, bytenr))
2685 ret = scrub_one_super(sctx, scrub_dev, page, bytenr, gen);
2687 spin_lock(&sctx->stat_lock);
2688 sctx->stat.super_errors++;
2689 spin_unlock(&sctx->stat_lock);
2696 static void scrub_workers_put(struct btrfs_fs_info *fs_info)
2698 if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
2699 &fs_info->scrub_lock)) {
2700 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
2702 fs_info->scrub_workers = NULL;
2703 mutex_unlock(&fs_info->scrub_lock);
2706 destroy_workqueue(scrub_workers);
2711 * get a reference count on fs_info->scrub_workers. start worker if necessary
2713 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2716 struct workqueue_struct *scrub_workers = NULL;
2717 unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
2718 int max_active = fs_info->thread_pool_size;
2721 if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
2725 scrub_workers = alloc_ordered_workqueue("btrfs-scrub", flags);
2727 scrub_workers = alloc_workqueue("btrfs-scrub", flags, max_active);
2731 mutex_lock(&fs_info->scrub_lock);
2732 if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
2733 ASSERT(fs_info->scrub_workers == NULL);
2734 fs_info->scrub_workers = scrub_workers;
2735 refcount_set(&fs_info->scrub_workers_refcnt, 1);
2736 mutex_unlock(&fs_info->scrub_lock);
2739 /* Other thread raced in and created the workers for us */
2740 refcount_inc(&fs_info->scrub_workers_refcnt);
2741 mutex_unlock(&fs_info->scrub_lock);
2745 destroy_workqueue(scrub_workers);
2749 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2750 u64 end, struct btrfs_scrub_progress *progress,
2751 int readonly, int is_dev_replace)
2753 struct btrfs_dev_lookup_args args = { .devid = devid };
2754 struct scrub_ctx *sctx;
2756 struct btrfs_device *dev;
2757 unsigned int nofs_flag;
2758 bool need_commit = false;
2760 if (btrfs_fs_closing(fs_info))
2763 /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
2764 ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
2767 * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
2768 * value (max nodesize / min sectorsize), thus nodesize should always
2771 ASSERT(fs_info->nodesize <=
2772 SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
2774 /* Allocate outside of device_list_mutex */
2775 sctx = scrub_setup_ctx(fs_info, is_dev_replace);
2777 return PTR_ERR(sctx);
2779 ret = scrub_workers_get(fs_info, is_dev_replace);
2783 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2784 dev = btrfs_find_device(fs_info->fs_devices, &args);
2785 if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
2787 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2792 if (!is_dev_replace && !readonly &&
2793 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2794 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2795 btrfs_err_in_rcu(fs_info,
2796 "scrub on devid %llu: filesystem on %s is not writable",
2797 devid, btrfs_dev_name(dev));
2802 mutex_lock(&fs_info->scrub_lock);
2803 if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
2804 test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
2805 mutex_unlock(&fs_info->scrub_lock);
2806 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2811 down_read(&fs_info->dev_replace.rwsem);
2812 if (dev->scrub_ctx ||
2814 btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2815 up_read(&fs_info->dev_replace.rwsem);
2816 mutex_unlock(&fs_info->scrub_lock);
2817 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2821 up_read(&fs_info->dev_replace.rwsem);
2823 sctx->readonly = readonly;
2824 dev->scrub_ctx = sctx;
2825 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2828 * checking @scrub_pause_req here, we can avoid
2829 * race between committing transaction and scrubbing.
2831 __scrub_blocked_if_needed(fs_info);
2832 atomic_inc(&fs_info->scrubs_running);
2833 mutex_unlock(&fs_info->scrub_lock);
2836 * In order to avoid deadlock with reclaim when there is a transaction
2837 * trying to pause scrub, make sure we use GFP_NOFS for all the
2838 * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
2839 * invoked by our callees. The pausing request is done when the
2840 * transaction commit starts, and it blocks the transaction until scrub
2841 * is paused (done at specific points at scrub_stripe() or right above
2842 * before incrementing fs_info->scrubs_running).
2844 nofs_flag = memalloc_nofs_save();
2845 if (!is_dev_replace) {
2846 u64 old_super_errors;
2848 spin_lock(&sctx->stat_lock);
2849 old_super_errors = sctx->stat.super_errors;
2850 spin_unlock(&sctx->stat_lock);
2852 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
2854 * by holding device list mutex, we can
2855 * kick off writing super in log tree sync.
2857 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2858 ret = scrub_supers(sctx, dev);
2859 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2861 spin_lock(&sctx->stat_lock);
2863 * Super block errors found, but we can not commit transaction
2864 * at current context, since btrfs_commit_transaction() needs
2865 * to pause the current running scrub (hold by ourselves).
2867 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
2869 spin_unlock(&sctx->stat_lock);
2873 ret = scrub_enumerate_chunks(sctx, dev, start, end);
2874 memalloc_nofs_restore(nofs_flag);
2876 atomic_dec(&fs_info->scrubs_running);
2877 wake_up(&fs_info->scrub_pause_wait);
2880 memcpy(progress, &sctx->stat, sizeof(*progress));
2882 if (!is_dev_replace)
2883 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
2884 ret ? "not finished" : "finished", devid, ret);
2886 mutex_lock(&fs_info->scrub_lock);
2887 dev->scrub_ctx = NULL;
2888 mutex_unlock(&fs_info->scrub_lock);
2890 scrub_workers_put(fs_info);
2891 scrub_put_ctx(sctx);
2894 * We found some super block errors before, now try to force a
2895 * transaction commit, as scrub has finished.
2898 struct btrfs_trans_handle *trans;
2900 trans = btrfs_start_transaction(fs_info->tree_root, 0);
2901 if (IS_ERR(trans)) {
2902 ret = PTR_ERR(trans);
2904 "scrub: failed to start transaction to fix super block errors: %d", ret);
2907 ret = btrfs_commit_transaction(trans);
2910 "scrub: failed to commit transaction to fix super block errors: %d", ret);
2914 scrub_workers_put(fs_info);
2916 scrub_free_ctx(sctx);
2921 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
2923 mutex_lock(&fs_info->scrub_lock);
2924 atomic_inc(&fs_info->scrub_pause_req);
2925 while (atomic_read(&fs_info->scrubs_paused) !=
2926 atomic_read(&fs_info->scrubs_running)) {
2927 mutex_unlock(&fs_info->scrub_lock);
2928 wait_event(fs_info->scrub_pause_wait,
2929 atomic_read(&fs_info->scrubs_paused) ==
2930 atomic_read(&fs_info->scrubs_running));
2931 mutex_lock(&fs_info->scrub_lock);
2933 mutex_unlock(&fs_info->scrub_lock);
2936 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
2938 atomic_dec(&fs_info->scrub_pause_req);
2939 wake_up(&fs_info->scrub_pause_wait);
2942 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2944 mutex_lock(&fs_info->scrub_lock);
2945 if (!atomic_read(&fs_info->scrubs_running)) {
2946 mutex_unlock(&fs_info->scrub_lock);
2950 atomic_inc(&fs_info->scrub_cancel_req);
2951 while (atomic_read(&fs_info->scrubs_running)) {
2952 mutex_unlock(&fs_info->scrub_lock);
2953 wait_event(fs_info->scrub_pause_wait,
2954 atomic_read(&fs_info->scrubs_running) == 0);
2955 mutex_lock(&fs_info->scrub_lock);
2957 atomic_dec(&fs_info->scrub_cancel_req);
2958 mutex_unlock(&fs_info->scrub_lock);
2963 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
2965 struct btrfs_fs_info *fs_info = dev->fs_info;
2966 struct scrub_ctx *sctx;
2968 mutex_lock(&fs_info->scrub_lock);
2969 sctx = dev->scrub_ctx;
2971 mutex_unlock(&fs_info->scrub_lock);
2974 atomic_inc(&sctx->cancel_req);
2975 while (dev->scrub_ctx) {
2976 mutex_unlock(&fs_info->scrub_lock);
2977 wait_event(fs_info->scrub_pause_wait,
2978 dev->scrub_ctx == NULL);
2979 mutex_lock(&fs_info->scrub_lock);
2981 mutex_unlock(&fs_info->scrub_lock);
2986 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
2987 struct btrfs_scrub_progress *progress)
2989 struct btrfs_dev_lookup_args args = { .devid = devid };
2990 struct btrfs_device *dev;
2991 struct scrub_ctx *sctx = NULL;
2993 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2994 dev = btrfs_find_device(fs_info->fs_devices, &args);
2996 sctx = dev->scrub_ctx;
2998 memcpy(progress, &sctx->stat, sizeof(*progress));
2999 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3001 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;