Ceph

The kclient is smart about osd disconnect: if there are outstanding requests, it reopens the connection. Objecter does not appear to be:

void Objecter::ms_handle_reset(Connection *con)
{
  if (con->get_peer_type() == CEPH_ENTITY_TYPE_OSD)
    maybe_request_map();
}

This is mitigated by the fact that we send a ping to any OSD with outstanding requests, but it does mean there is a built-in delay before we "notice" a disconnect.

While we're at it, double-check the locking is correct (is ms_handle_reset called by dispatcher, or manually invoked by the parent's ms_handle_reset method?).