[PATCH 4/4] btrfs-progs: disk-io: Try to find a best copy when reading tree blocks

[PATCH 4/4] btrfs-progs: disk-io: Try to find a best copy when reading tree blocks

[Qu Wenruo]

[BUG]
If the first copy of a tree block has a bad key order, but the second
copy is completely good, then "btrfs ins dump-tree -b <bytenr>" fails to
print anything past the bad key:
  leaf 29786112 items 47 free space 983 generation 20 owner EXTENT_TREE
  leaf 29786112 flags 0x1(WRITTEN) backref revision 1
  fs uuid 3381d111-94a3-4ac7-8f39-611bbbdab7e6
  chunk uuid 9af1c3c7-2af5-488b-8553-530bd515f14c
  [snip]
  	item 9 key (20975616 METADATA_ITEM 0) itemoff 3543 itemsize 33
  		refs 1 gen 16 flags TREE_BLOCK
  		tree block skinny level 0
  		tree block backref root CHUNK_TREE
  	item 10 key (29360128 BLOCK_GROUP_ITEM 33554432) itemoff 3519 itemsize 24
  		block group used 94208 chunk_objectid 256 flags METADATA|DUP
  ERROR: leaf 29786112 slot 11 pointer invalid, offset 1245184 size 0 leaf data limit 3995
  ERROR: skip remaining slots

While kernel can locate the good copy and acts just like nothing
happened.

[CAUSE]
btrfs-progs uses read_tree_block() to try each copy. But it only uses
less strict check_tree_block(), which has less sanity check than
btrfs_check_node/leaf().

Some error like bad key order is ignored to allow btrfs check to fix it.

This leads to above problem.

[FIX]
Introduce a new member, @candidate_mirror in read_tree_block(), which
records the copy passes check_tree_block() but fails
btrfs_check_leaf/node() as last chance.

Only if no better copy found, then use @candidate_mirror.

So btrfs-progs can act just like kernel to use best copy.

Reported-by: Yoon Jungyeon <jungyeon@gatech.edu>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=202691
[Inspired by that image, not to fix any bug of that bugzilla]
Signed-off-by: Qu Wenruo <wqu@suse.com>
---
 disk-io.c | 30 +++++++++++++++++++++++++++---
 1 file changed, 27 insertions(+), 3 deletions(-)

diff --git a/disk-io.c b/disk-io.c
index 1736d4a1f94b..737f1cdf6f4c 100644
--- a/disk-io.c
+++ b/disk-io.c
@@ -327,6 +327,7 @@ struct extent_buffer* read_tree_block(struct btrfs_fs_info *fs_info, u64 bytenr,
 	u32 sectorsize = fs_info->sectorsize;
 	int mirror_num = 1;
 	int good_mirror = 0;
+	int candidate_mirror = 0;
 	int num_copies;
 	int ignore = 0;
 
@@ -362,10 +363,30 @@ struct extent_buffer* read_tree_block(struct btrfs_fs_info *fs_info, u64 bytenr,
 					      &fs_info->recow_ebs);
 				eb->refs++;
 			}
-			btrfs_set_buffer_uptodate(eb);
-			return eb;
+
+			/*
+			 * check_tree_block() is less strict to allow btrfs
+			 * check to get raw eb with bad key order and fix it.
+			 * But we still need to try to get a good copy if
+			 * possible, or bad key order can go into tools like
+			 * btrfs ins dump-tree.
+			 */
+			if (btrfs_header_level(eb))
+				ret = btrfs_check_node(fs_info, NULL, eb);
+			else
+				ret = btrfs_check_leaf(fs_info, NULL, eb);
+			if (!ret || candidate_mirror == mirror_num) {
+				btrfs_set_buffer_uptodate(eb);
+				return eb;
+			}
+			if (candidate_mirror <= 0)
+				candidate_mirror = mirror_num;
 		}
 		if (ignore) {
+			if (candidate_mirror > 0) {
+				mirror_num = candidate_mirror;
+				continue;
+			}
 			if (check_tree_block(fs_info, eb)) {
 				if (!fs_info->suppress_check_block_errors)
 					print_tree_block_error(fs_info, eb,
@@ -387,7 +408,10 @@ struct extent_buffer* read_tree_block(struct btrfs_fs_info *fs_info, u64 bytenr,
 		}
 		mirror_num++;
 		if (mirror_num > num_copies) {
-			mirror_num = good_mirror;
+			if (candidate_mirror > 0)
+				mirror_num = candidate_mirror;
+			else
+				mirror_num = good_mirror;
 			ignore = 1;
 			continue;
 		}
-- 
2.21.0