Hello, I found possible deadlock in AbstractCommitLogSegmentManager. The root cause is incorrect use of LockSupport.park/unpark pair. Unpark should be invoked only if caller is sure that thread was parked in appropriate place. Otherwice permission given by calling unpark can be consumed by other structures (for example - inside ReadWriteLock).
Jira: https://issues.apache.org/jira/browse/CASSANDRA-13652 I suggest simplest solution: change LockSupport to Semaphore. PR: https://github.com/apache/cassandra/pull/127 Also I suggest another solution with SynchronousQueue-like structure to move available segment from Manager Thread to consumers. With theese changes code became more clear and straightforward. PR https://github.com/apache/cassandra/pull/129 We can not use j.u.c.SynchronousQueue because we need to support shutdown and there is only way to terminate SynchronousQueue.put is to call Thread.interrupt(). But C* uses nio and it does not expect ClosedByInterruptException during IO operations. Thus we can not interrupt Manager Thread. I implemented o.a.c.u.c.Transferer that supports shutdown and restart (needed for tests). https://github.com/Fuud/cassandra/blob/e1a695874dc24e532ae21ef627e852bf999a75f3/src/java/org/apache/cassandra/utils/concurrent/Transferer.java Also I modified o.a.c.d.c.SimpleCachedBufferPool to support waiting for free space. Please feel free to ask any questions. Thank you. Feodor Bobin fuudtorrent...@gmail.com
