Re: [TLS] What does it mean to not include 0-RTT message in the handshake hash?

[Christian Huitema <[e-mail]>]

On Monday, December 21, 2015 6:30 PM, Martin Thomson wrote:
> 
> On 22 December 2015 at 13:25, Christian Huitema <[e-mail]>
> wrote:
> >> Unless I'm confused (which is possible given the time of night),
> >> the intention, as you say, is to separate out the 0-RTT handshake
> >> messages i.e., (cert, cert verify, finished) from the 1-RTT computations.
> >
> > OK. That does not simplify implementations using running hashes...
> 
> It does if you consider the possibility of having to drop the 0-RTT data.

That's right. In fact, it may be a good idea to add to the spec a description of a "Failed 0-RTT handshake." If I understand correctly, the following will happen:

* Server will receive the client hello, ignore the Early Data Indication extension, and proceed as in 1-RTT.
* Server will indicate that by not adding an Early Data Indication to the server hello.
* Server will receive a series of 0-RTT messages that it cannot decipher, and just drop the messages.
* Client will receive server hello, and proceed as per 1-RTT. Client API will signal that 0-RTT data was lost, application may decide to retransmit.
* Server may send client authentication requests. Client will have to repeat the authentication messages, even if it already sent them as 0-RTT.

In that scenario, the handshake hash cannot include the 0-RTT messages, since the server does not in fact receive them, and they do not contribute to the state of the connection. 

We can of course debate whether the 0-RTT messages should also not be included in the hash if the 0-RTT exchange was successful, the messages were received, and they contributed to the state of the connection. If they are not included, then the "Finished" HMAC does not offer a protection against tampering. This may open the possibility of some kind of substitution or replay attack.

The failed 0-RTT handshake scenario also has interesting consequences on the Record layer. We have a legitimate scenario in which received records cannot be decrypted. This should not trigger alarms. And the numbering scheme should be robust against these missing records.

-- Christian Huitema