I think I found a bug about the clientreply of mds. (different from #4742) After failover, the standby mds begin to recover and enter the clientreply state and never moves to the next state(Active). I gdbed the mds process by gcore and found that the main thread (dispatch thread) is idle and mdcache->active_request is empty, but mds->replay_queue still has one element, that is strange. >From the code, replay_queue has all requests which need to be replayed. When the mds enters clientreplay state, MDS::queue_one_replay will be called to pick a requect from the replay_queue and put the request into finished_queue. So the replay operation begins to work. After the first replay request has finished, MDS::queue_one_replay should be called again to deal with the next replay request. There are three paths to do this: 1) Server::journal_and_reply 2) MDCache::reqeust_cleanup 3) Server::handle_client_request But it seems that no path called the MDS::queue_one_replay method. As a result, the mds stuck in clientreplay state. Maybe there is a request process path which will never use the above three methed. But I can't find the previous request while it seems to completed and cleaned up from the MDCache. There is any one has some idea about these problem? I can give more details if needed. I have the core dump but it is too big (300MB+) to upload. Thanks for any help. -- Dong Yuan Email:yuandong1222@xxxxxxxxx -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
