[Cluster-devel] [gfs2-utils PATCH 3/7] fsck.gfs2: Fix directory link on relocated directory dirents

[Bob Peterson]

There was a problem whereby fsck.gfs2 would discover a directory entry
that's on the wrong leaf block (the dirent's hash value doesn't fit
the acceptable range for that leaf block). In that case, fsck would
relocate the dirent to an appropriate leaf block. The problem is, if
that dirent was for a directory, the directory link count was adjusted
appropriately, but it was not properly added to the directory linkage
due to the fact that it took an error path in order to delete the
dirent that was relocated. This resulted in a problem when pass3 went
to verify the directory linkage. Despite the intact directory linkage,
it flagged the directory as not linked, then it improperly added it
to lost+found. Subsequent fsck.gfs2 would discover the double-linkage
problems.

rhbz#984085
---
 gfs2/fsck/pass2.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/gfs2/fsck/pass2.c b/gfs2/fsck/pass2.c
index fb15eae..dabfc13 100644
--- a/gfs2/fsck/pass2.c
+++ b/gfs2/fsck/pass2.c
@@ -391,7 +391,18 @@ static int wrong_leaf(struct gfs2_inode *ip, struct gfs2_inum *entry,
 		   leaf, but that leaf has already been processed. So we have
 		   to nuke the dent from this leaf when we return, but we
 		   still need to do the "good dent" accounting. */
-		error = incr_link_count(*entry, ip, _("valid reference"));
+		if (de->de_type == (sdp->gfs1 ? GFS_FILE_DIR : DT_DIR)) {
+			error = set_parent_dir(sdp, de->de_inum,
+					       ip->i_di.di_num);
+			if (error > 0)
+				/* This is a bit of a kludge, but returning 0
+				   in this case causes the caller to go through
+				   function set_parent_dir a second time and
+				   deal properly with the hard link. */
+				return 0;
+		}
+		error = incr_link_count(*entry, ip,
+					_("moved valid reference"));
 		if (error > 0 &&
 		    bad_formal_ino(ip, dent, *entry, tmp_name, q, de, bh) == 1)
 			return 1; /* nuke it */
-- 
1.8.3.1