Affects Version/s: 2.0.8
Fix Version/s: 2.0.14
Multiple OSes / Java configurations:
* Mac OS X El Capitan on Java 8 (1.8.0_60)
* CentOS 6.4 on Java 8 (1.8.0_60)
* CentOS 6.5 on Java 8 (1.8.0_20-b26).
MINA SSHD isn't removing sessions when using the MINA/NIO backend if an exception as received as the session is closing (such as a connection reset is received with data still in the write buffer). When this case happens, it seems that NioProcessor.getState returns the state as CLOSING (I'm assuming since the underlying channel is now closed), which means that the AbstractPollingIoProcessor.removeSessions() won't ever prune the session, since a CLOSING state is simply ignored. The result is a resource leak over time, since these sessions are never pruned (it's a slow leak, since entering this condition is racy – on my workstation, I can produce it through randomly interrupting connections anywhere from 1/6 to 1/10th of the time). (This may either be major or critical; reprioritize as necessary.)
I specifically see this error with Gerrit 2.10.4 and Gerrit 2.11.5 (using mina-sshd 0.14.0 / mina-core 2.0.8), and it looks like the code path is unchanged in mina-sshd 1.0.0 / mina-core 2.0.9. I was unsure if this is specifically a bug in mina-core or, if it's something unique to mina-sshd. My local development system runs Mac OS X El Capitan on Java 8 (1.8.0_60), but I've also seen this on Linux (CentOS 6.4, again Java 1.8.0_60 and CentOS 6.5 on Java 1.8.0_20-b26).
The fix may be as simple as attempting to remove the session if OPENED or CLOSING, but I'm unsure what side-effects this may have with other backends. I'll be happy to test it locally, but I'm fairly ignorant when it comes to MINA's code.
The attached patch (to mina-core) seems to resolve the issue by following the reproduction case I have on the Gerrit issue tracker.