diff --git a/[refs] b/[refs]
index 11480cd5bf60..233bd9ca819d 100644
--- a/[refs]
+++ b/[refs]
@@ -1,2 +1,2 @@
 ---
-refs/heads/master: da353b0d64e070ae7c5342a0d56ec20ae9ef5cfb
+refs/heads/master: b93bd20cd59eb7ec172f95d08b100fea688d8bcf
diff --git a/trunk/fs/xfs/xfs_vfsops.c b/trunk/fs/xfs/xfs_vfsops.c
index 1644be14a144..b4db8a208076 100644
--- a/trunk/fs/xfs/xfs_vfsops.c
+++ b/trunk/fs/xfs/xfs_vfsops.c
@@ -1720,7 +1720,18 @@ xfs_parseargs(
 	int			dsunit, dswidth, vol_dsunit, vol_dswidth;
 	int			iosize;
 
-	args->flags |= XFSMNT_IDELETE;
+	/*
+	 * Applications using DMI filesystems often expect the
+	 * inode generation number to be monotonically increasing.
+	 * If we delete inode chunks we break this assumption, so
+	 * keep unused inode chunks on disk for DMI filesystems
+	 * until we come up with a better solution.
+	 * Note that if "ikeep" or "noikeep" mount options are
+	 * supplied, then they are honored.
+	 */
+	if (!(args->flags & XFSMNT_DMAPI))
+		args->flags |= XFSMNT_IDELETE;
+
 	args->flags |= XFSMNT_BARRIER;
 	args->flags2 |= XFSMNT2_COMPAT_IOSIZE;