Details
Major Description
I have been debugging some threading issues in NMS, and I have attached a patch file with some proposed changes to correct the following issues:
- Some threading issues in Connection, Session, and DispatchingThread
- it is insufficient to use ArrayList.Synchronized or Hashtable.Synchronized since enumeration is not synchronized. Instead, use regular collections and explicit locking using a private object, myLock
- instead of Connection.connected, use a volatile single-state-change bool called triedConnect, to add synchronization to CheckConnected without adversely-affecting performance
- mark as volatile certain bools which have a single state change (e.g., Connection.closed, which starts at false and changes once to true) and then double-check for minimal locking
- some uses of AtomicBoolean (e.g., Connection.started) offered insufficient code synchronization. Instead, use regular bool and locking of the aforementioned myLock
- remove Connection.closing field, and instead always accommodate that Connection.RemoveSession modifies the sessions list.
- instead of (bool) DispatchingThread.m_bStopFlag, use a ManualResetEvent (in addition to the existing AutoResetEvent) to signal the worker thread
- in TcpTransport, recognize that ShutdownCommand may cause broker to close connection, to avoid a spurious error message
- accommodate potential ThreadAbortExceptions which can occur from explicitly-aborted worker threads
- increase the wait for stopping async delivery, from 5 seconds to 30
- use Interlocked.Increment for thread-safe int increments. (Interlocked.Increment wraps Int32.MaxValue to Int32.MinValue; will this be a problem?)
- Tweak implementation of IDisposeable in TcpTransport, MutexTransport, TransportFilter, WireFormatNegotiator so finalizers call Dispose
Thank you,
Chris
