Re: [TLS] RFC 8449 – DTLS 1.2 considerations

[Martin Thomson <[e-mail]>]

On Tue, Jul 14, 2020, at 22:41, Achim Kraus wrote:
> For me it seems, that an agreement about this message buffer size is
> still missing.

This is a question we dealt with in QUIC by limiting UDP datagram size rather than the record size (or packet size in QUIC terminology).

In this case, the spec doesn't really say the most helpful thing about UDP datagram size in DTLS.  Indeed, you might say that pointing to the ability to include multiple records per datagrams is an invitation to do bad things.  After all, a receiver that receives a datagram that is too large for it likely has to drop it or only process part of it.  The partial processing is possible if there are small records, but it does lead to lots of losses.

The way I would approach this is to send one record per datagram in DTLS, no matter what the spec says.  

(NSS implements this spec, but does not assume any limits, it only respects a limit set by a peer.  For DTLS it sends one record per datagram, outside of the handshake.)

Assuming that you do have a constrained implementation that can't receive large datagrams, then we probably need different signaling.  But there are workarounds.

If you start receiving large UDP datagrams, then it is possible to drop those and hope that your peer has some sort of PMTU discovery (NSS has something, but it is crude).  At that point, the size of datagrams is limited as though it were a path limitation.  That likely gets you down to the minimum MTU for the IP version you are using.  For IPv6, that's not so different from the 1400-odd that you describe, but IPv4 - and therefore some implementations - might go as low as the IPv4 minimum MTU.

(FWIW, NSS will go as low as a 228 byte MTU, which is pretty extreme.  And it takes a lot of failed retransmissions to get that low.  I wouldn't expect that small an MTU from other implementations.)

All that said, if you have a size constraint smaller than a typical UDP MTU, then you have a very constrained implementation.  Not being able to spare ~1k but expecting to do TLS is something I struggle to comprehend.