TLS NPN support

[crazyproger]

On my production system found that clients with TLS NPN can not connect to envoy. This patch fixes that, tested on my production system.

[RomanDzhabarov]

@PiotrSikora do you mind taking a look? :-)

[PiotrSikora]

I'll test this in a bit, but at the very least, it looks that the NPN support is added only on the downstream side and not the upstream. I'm not sure if we want to have this asymmetry.

[crazyproger]

You are absolutely right - I've only tested downstream part, since my infrastructure does not use TLS to upstreams. If you suggest some code point where it should be implemented for upstreams - I will try to fix it tomorrow.

[crazyproger]

After some investigation I think that process should be "mirrored" in ClientContextImpl. So in constructor callback should be installed through SSL_CTX_set_next_proto_select_cb() which will be very similar to alpnSelectCallback(). But one thing that I've still didn't found in OpenSSL documentation is - how to switch between ALPN and NPN on-the-fly(and ideally prefer ALPN?), since by RFCs negotiation variant is initiated by client in first ClientHello message.

[mattklein123]

@crazyproger @PiotrSikora no objection from me for supporting NPN (though it's been deprecated). But we should support both server and client. Also, you will need to add tests. Testing will be a little complicated insofar as without a config option, it will be difficult to differentiate NPN from ALPN. @PiotrSikora thoughts on what to do about tests?

[PiotrSikora]

@crazyproger for the client side, you need to install callback using SSL_CTX_set_next_proto_select_cb() in ClientContextImpl::ClientContextImpl(). In the callback, you can use SSL_select_next_proto() to find the first shared protocol.

As for the choice between ALPN & NPN - BoringSSL will ignore NPN if ALPN was also announced by the client, and client will reject connection if server incorrectly negotiated both, so we should be covered there.

@mattklein123 We can't really test it without #186 (soon...), not without adding knobs to the config, which we shouldn't do.

[mattklein123]

OK fair enough, I'm willing to let tests slide on this one once we add the client side code.

[mattklein123]

Going to close this for now. We can reopen when ready to finish it up.