Re: [quicwg/base-drafts] loss of only two packets can lead to an unrecoverable situation (#2267)

[David Schinazi <[e-mail]>]

Having thought about this some more, I think that the HANDSHAKE_DONE proposal might be our best way forward.

Knowing when to stop sending ClientFinished/Ack is a variant of the two generals problem, meaning it's actually impossible to know when to stop sending over a lossy network without out-of-band information.

There are two ways to solve this. One is timeouts, similar to what TCP does on graceful connection close. However timeouts aren't great, because they're by definition always too short or too long. In this case, we can do better by using an explicit event that would trigger key discarding. We even have a suitable stream of out-of-band information: our 1-RTT protected packets.

One could consider having the client treat the receipt of any valid 1-RTT packet as confirmation that the ClientFinished was received, but that doesn't work because the server is allowed to send non-sensitive application data before validating the ClientFinished.

Having the server send a retransmittable HANDSHAKE_DONE frame as soon as it receives ClientFinished solves this issue. The server discards its handshake keys when it sends HANDSHAKE_DONE and the client discards them when it receives it.

I personally find this preferable to a short timeout (which can cause an infinite send) and a long timeout (which costs resources).

-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/issues/2267#issuecomment-455723789