Revisit Namenode locking
------------------------
Key: HDFS-2184
URL: https://issues.apache.org/jira/browse/HDFS-2184
Project: Hadoop HDFS
Issue Type: Improvement
Components: name-node
Reporter: Eli Collins
While working on HDFS-988 I noticed that the locking in FSNamesystem and
FSDirectory could be improved. Some observations:
The namesystem lock (fsLock) is always taken before acquiring the directory
lock (dirLock). Therefore the only time when the directory lock is needed is
when the fsLock is taken for reading and the directory lock is taken for
writing, but I don't think that ever happens. Therefore we can probably get rid
of the directory lock.
In HDFS-988 I modified handleHeartbeat to take the read lock so it's
synchronized with register datanode. I also added a missing synchronization of
datanodeMap to wipeDatanode because handleHeartbeat calls getDatanode() while
only holding locks on heartbeats and datanodeMap, but registerDatanode mutates
datanodeMap without locking either. We should revisit which
locks/synchronization protect which data structures, there may be other similar
bugs and also opportunities to increase parallelism.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
