ceph_fill_inode() seems to be mishandling I_NEW. It only check I_NEW when setting i_mode. It then goes on to clobber a bunch of things in the inode struct and ceph_inode_info struct (granted in some cases it's overwriting with the same thing), irrespective of whether the inode is already set up (i.e. if I_NEW isn't set). It looks like I_NEW has been interpreted as to indicating that the inode is being created as a filesystem object (e.g. by mkdir) whereas it's actually merely about allocation and initialisation of struct inode in memory. David
