From: Bob Peterson <rpete...@redhat.com>
The job of pass1b is to resolve duplicate references to the same block.
Eventually it does a fair job of determining the rightful owner of the
block, and then it has to deal with the other dinode(s) that referenced
the block improperly. If another dinode improperly referenced the block
as data or metadata, it's obvious file corruption and the dinode should
be deleted. However, if the other dinode improperly referenced the
block as an extended attribute, it can fix the situation by removing
the extended attributes from the dinode. Prior to this patch, there
was a check in the code for this situation so that the dinode was only
deleted if the bad block reference was as data or metadata. However,
regardless of the situation, the code removed the inode from the
inode rbtree. That resulted in the dinode being considered unlinked,
so it would get improperly tossed into lost+found and left in a
indeterminate state. Subsequent runs of fsck.gfs2 could find the
discrepancy and flag it as unlinked again. This patch adds another
check so that the inode is not removed from the inode rbtree, so it
is linked properly during pass2.
rhbz#902920
---
gfs2/fsck/pass1b.c | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/gfs2/fsck/pass1b.c b/gfs2/fsck/pass1b.c
index f3f90ef..bd60d84 100644
--- a/gfs2/fsck/pass1b.c
+++ b/gfs2/fsck/pass1b.c
@@ -523,9 +523,12 @@ static int resolve_dup_references(struct gfs2_sbd *sdp,
struct duptree *b,
(unsigned long long)id->block_no);
ip = fsck_load_inode(sdp, id->block_no);
- ii = inodetree_find(ip->i_di.di_num.no_addr);
- if (ii)
- inodetree_delete(ii);
+ if (id->reftypecount[ref_as_data] ||
+ id->reftypecount[ref_as_meta]) {
+ ii = inodetree_find(ip->i_di.di_num.no_addr);
+ if (ii)
+ inodetree_delete(ii);
+ }
clear_dup_fxns.private = (void *) dh;
/* Clear the EAs for the inode first */
check_inode_eattr(ip, &clear_dup_fxns);
--
1.7.11.7
