JDK-8219568 extended master secret performance problems
Xuelei Fan
xuelei.fan at oracle.com
Mon Apr 8 19:00:00 UTC 2019
Hi Daniel,
Thanks for the quick feedback. It helps me a lot.
On 4/8/2019 9:59 AM, Daniel Jeliński wrote:
> Hi Xuelei,
> Thanks for your response!
> My understanding is that legacy resumption = resumption of a session
> that was established without extended master secret extension.
>
> Our Java application is a web server that is communicating with a large
> number of clients, majority of which are built on top of OpenSSL 1.0.2,
> which does not implement extended master secret. The clients send data
> to server using frequent short-lived connections.
>
> When we use Java pre-8u161 or disable extended master secret
> (/jdk/./tls/.useExtendedMasterSecret=false), the usual workflow is as
> follows:
> - client connects to server for the first time
> - Full handshake happens, server creates a session ID and caches it
> - session is established, data is transferred, connection is closed.
> Later:
> - subsequent client connection sends the cached session ID
> - server resumes session using abbreviated handshake
> - data is transferred, connection is closed.
>
> The same workflow with extended master secret enabled is as follows:
> - client connects to server for the first time
> - Full handshake happens, server creates a session ID and caches it
> - session is established, data is transferred, connection is closed.
> Later:
> - subsequent client connection sends the cached session ID
> - server checks that the session ID was established without extended
> master secret and rejects it. Full handshake happens, server creates a
> session ID and caches it
> - session is established, data is transferred, connection is closed.
>
It sounds like a reasonable use case if applications want to take the
risks. I will think more about if we can make an enhancement to allow
legacy resumption again if the extended master secret extension is not used.
> Full handshake is much more expensive than abbreviated handshake, and
> caching thousands of session IDs that are never reused creates a burden
> on GC.
>
> My understanding of RFC 7627 is that rejecting abbreviated handshake
> when extended master secret is not used makes sense only when we are
> using client certificates for authentication. We are not using client
> certificates in our communication, so we would prefer to resume sessions
> whether extended master secret is used or not.
>
> TLS specification does not require the server to assign a session ID
> when it knows it will not allow the client to resume session. We should
> take advantage of that and not assign a session ID when the user does
> not want to resume legacy sessions.
>
Good idea!
Thanks,
Xuelei
> Let me know if that makes sense now.
> Thanks,
> Daniel
>
>
> pon., 8 kwi 2019 o 17:43 Xuelei Fan <xuelei.fan at oracle.com
> <mailto:xuelei.fan at oracle.com>> napisał(a):
>
> Hi Daniel,
>
> Was extended master secret extension used when legacy resumption is
> expected? I did not get the point from JDK-8219568 and this
> description. It would be helpful if there is a test code to reproduce
> the behavior.
>
> Thanks,
> Xuelei
>
> On 4/6/2019 11:36 AM, Daniel Jeliński wrote:
> > Hi all,
> > Ever since upgrading to Java 8u161 we are running into performance
> > problems that were caused by the implementation of extended
> master secret.
> >
> > First problem was described in 8219568; server does not allow
> resuming
> > legacy sessions even when jdk.tls.allowLegacyResumption is set to
> true.
> > Based on the mail archives of the original discussion [1] and the
> > release notes [2] I think this was not what was intended. Should the
> > setting (jdk.tls.allowLegacyResumption) on the server side work like
> > this instead?
> > allow = true -> proceed with abbreviated handshake
> > allow = false -> proceed with full handshake
> >
> > Documentation is ambiguous enough that we would probably not even
> need
> > to change it. Today it states that setting allowLegacyResumption to
> > false rejects abbreviated handshakes, without clarifying what the
> > default does.
> >
> > Second problem is that while the server rejects the abbreviated
> > handshake, it generates and caches a new session ID on every client
> > reconnect, effectively thrashing the session cache. These IDs are
> never
> > used. Should we stop generating session IDs when legacy
> resumption is
> > disabled?
> >
> > Regards,
> > Daniel
> >
> >
> > [1]
> >
> http://openjdk.5641.n7.nabble.com/Code-Review-Request-JDK-8148421-Extended-Master-Secret-TLS-extension-td311192.html
> > [2] https://bugs.openjdk.java.net/browse/JDK-8192045
>
More information about the security-dev
mailing list