Re: [BEHAVE] TURN Issue: Preserving bits in the IP header

Le Wednesday 09 January 2008 11:46:17 Lars Eggert, vous avez écrit :
> On 2008-1-8, at 11:23, ext Rémi Denis-Courmont wrote:
> > I understand why TURN should preserve TOS and ECN bits, but we have
> > to know that if this were required, this would require in-kernel
> > implementation for most typical operating systems. And because
> > people will anyway try to make userland implementations, they'll
> > get it wrong. 
>
> Understood, doing things correctly is hard with existing APIs. But
> having TURN break ECN and DiffServ isn't the right approach, either.

Of course. I am rather wondering if this would be a SHOULD, or a MUST. That 
would make some a whole class implementations non-compliant, though they'd 
still mostly work.

For UDP (and UDP-Lite if we ever define if for TURN), we certainly want TOS 
and ECN preservation. That works fine because each client-to-relay packet 
corresponds to one or zero relay-to-peer packet, and each peer-to-relay 
packet corresponds to one or zero relay-to-client packet.

I am a whole lot more suspicious when it comes to TURN-TCP though. It is 
fortunate, we have split that out of the TURN baseline spec, because I can 
already imagine the headaches. With TCP, there is no obvious matching of IP 
packets on each sides of the relay. TCP does not preserve datagram 
boundaries. Hence a single packets from one side could turn into multiples 
packets on the other side for various reasons. Worse yet, a single packet on 
one side could be "coming from" different packets from the original side.

-- 
Rémi Denis-Courmont

_______________________________________________
Behave mailing list
Behave@ietf.org
https://www1.ietf.org/mailman/listinfo/behave