[jira] [Created] (SSHD-855) CachingPublicKeyAuthenticator not clearing up closed sessions

["jpalacios (JIRA)" <ji...@apache.org>]

jpalacios created SSHD-855:
------------------------------

             Summary: CachingPublicKeyAuthenticator not clearing up closed sessions
                 Key: SSHD-855
                 URL: https://issues.apache.org/jira/browse/SSHD-855
             Project: MINA SSHD
          Issue Type: Bug
    Affects Versions: 1.7.0
            Reporter: jpalacios


I'm currently investigating an incident in a production environment which resulted in an {{OutOfMemoryError}}.

According to the heap dump we are seeing massive object graphs made up of nested MINA's {{NioSocketSessions}}:
{code:java}
Thread -> ServerSessionImpl -> SshServer -> JohnsonAwarePublicKeyAuthenticator -> CachingPublicKeyAuthenticator -> ConcurrentHashMap -> ConcurrentHashMap$Node[] -> ConcurrentHashMap$Node -> ServerSessionImpl -> MinaSession -> NioSocketSession -> SelectionKeyImpl -> EpollSelectorImpl -> HashMap -> SelectionKeyImpl -> NioSocketSession -> ...
{code}
As you can see pass the first {{NioSocketSession}} the graph starts to repeat itself. These are not loops however. These are all different instances with a small shallow size but which add up to massive retained sizes. The {{CachingPublicKeyAuthenticator}} in particular has a retained size of ~6GB.

Another interesting thing is that AFAICT out the 94K {{NioSocketSession}} instances in the heap, 93983 have a closed channel and the {{ServerSessionImpl}} that directly reference them is closed as well. 59 of these {{ServerSessionImpl}} instances are directly referenced from the {{cache}} map in {{CachingPublicKeyAuthenticator}}.

I'm still gathering information to understand the load the system was in at the time of the incident, but I would appreciate your help in trying to understand how this nested graph could be put together, and if there's a scenario in which the {{CachingPublicKeyAuthenticator}} may fail to clear it's cached correctly.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)