The issue is that our SSL session cache is by both host and port number. Using PASV to initiate the download will give a different port number than the control connection on port 21 and therefore uses a new SSL session context.

Thanks for your feedback. Haven't looked at the code myself, but couldn't we retrieve and clone the session key from the cache and create a new related entry in the session cache instead of doing a renegotiation? I guess at the time of creating the new connection for the download it's still known what the related control connection is.

Replying to abrax5:

Thanks for your feedback. Haven't looked at the code myself, but couldn't we retrieve and clone the session key from the cache and create a new related entry in the session cache instead of doing a renegotiation? I guess at the time of creating the new connection for the download it's still known what the related control connection is.

The session caching is in a private class in the sun.security package I cannot access.

Issue is documented in the wiki.

Cross referencing to bug reports in thirdparty software:

• ​FireFTP issue
• ​Original ProFTPd bug report

I looked at GNU classpath, and there's a file called ClientHandshake.java:

            case WRITE_CLIENT_HELLO:
            {
              ClientHelloBuilder hello = new ClientHelloBuilder();
              AbstractSessionContext ctx = (AbstractSessionContext)
                engine.contextImpl.engineGetClientSessionContext();
              continued = (SessionImpl) ctx.getSession(engine.getPeerHost(),
                                                       engine.getPeerPort());

Maybe Cyberduck could use a modified version of GNU classpath for this use case and modify the AbstractSessionContext implementation so that it returns a previously used session for the same host even though the port is different?

I'm not really that familiar with Java, but looking at this: ​http://download.oracle.com/javase/6/docs/api/javax/net/ssl/SSLContext.html

The protected constructor for an SSLContext allows you to specify an SSLContextSpi. Maybe we could override SSLContext implementation with one that uses an SSLContextSpi derived from the default one, that just differs in what it would return for engineGetClientSessionContext(). This method returns an SSLSessionContext which allows enumerating SessionIDs.

These are just some rough ideas, but may through this path we could tell the TLS engine to try reusing a session. We don't really need access to the masterSecret of the session, I think. We just need to make sure that the TLS engine can find the old ID and advertise the reuse ID to the server in the CLIENT_HELLO msg.

Replying to abrax5:

I looked at GNU classpath, and there's a file called ClientHandshake.java:

            case WRITE_CLIENT_HELLO:
            {
              ClientHelloBuilder hello = new ClientHelloBuilder();
              AbstractSessionContext ctx = (AbstractSessionContext)
                engine.contextImpl.engineGetClientSessionContext();
              continued = (SessionImpl) ctx.getSession(engine.getPeerHost(),
                                                       engine.getPeerPort());

Maybe Cyberduck could use a modified version of GNU classpath for this use case and modify the AbstractSessionContext implementation so that it returns a previously used session for the same host even though the port is different?

Yes, the peer port is what is causing a new session being created. The solution must somehow overwrite the peer port.

Replying to abrax5:

We have a problem delegating to the default context because everything has protected access. Also a new SSLContext is created using the factory method getInstance. Have to dig further.

In OpenJDK, this code is in ClientHandshaker.java: (from ​http://download.java.net/openjdk/jdk6/)

    HandshakeMessage getKickstartMessage() throws SSLException {
        ClientHello mesg = new ClientHello(sslContext.getSecureRandom(),
                                        protocolVersion);
        maxProtocolVersion = protocolVersion;

        clnt_random = mesg.clnt_random;

        //
        // Try to resume an existing session.  This might be mandatory,
        // given certain API options.
        //
        session = ((SSLSessionContextImpl)sslContext
                        .engineGetClientSessionContext())
                        .get(getHostSE(), getPortSE());

So it'd be really useful if we can hook in there to replace the SSLSessionContext with our own implementation.

Replying to abrax5:

In OpenJDK, this code is in ClientHandshaker.java: (from ​http://download.java.net/openjdk/jdk6/)

Yes, that is what I meant above that the relevant code is in sun.security. We would need to replace a very large part of the JDK with custom implementations if we want to replace that implementation.

Because of the proposal to switch to a secure connection if the server advertises AUTH TLS in its features, this has become a major issue for many. See #5168.

Added configuration note to wiki.

#5285 closed as duplicate.

A workaround if the server allows would be to open the data socket without TLS. Uncheck Try to use TLS for data channel in Preferences → FTP → Data Channel Security.

Replying to abrax5:

In OpenJDK, this code is in ClientHandshaker.java

Let me know if you have found any resolution for this. I am clueless how this could be best achieved.

#5484 closed as duplicate.

#5658 closed as duplicate.

Blocked by ​NET-408.

#6027 closed as duplicate.

In r10760.

Addendum in r10836.

Addendum for running with latest runtime in r50436.