> >
> > Just to be sure that we aren't talking past each other, let's ensure
> > that we are talking about the same scenario. The GRE egress router
> > receives complete, non-fragmented packets, containing fragmented
> > payloads.
>
> [Linda] The scenario I am talking about is:
> The packet arriving at the Ingress node is has packet size less than
> the MTU. When the ingress node adds the GRE header, the new packet size
> exceeds the MTU. When this happens, the Ingress node has to break the
> received the packet to two (half) frames, encapsulate each half frame
> with proper GRE header, and send two separate encapsulated frames to
> Egress node.
>
>
> Is this same as your scenario? From your description below, it doesn't
> sound like the same.
> If it not the same, this scenario should be included in the draft. I
> can provide a section to describe this scenario and proper processing
> in Ingress node and Egress node.
>
> Under my scenario, the Egress node needs to reassemble the two half
> frames before sending to the destination because each half frame is not
> a complete IP packet.
>
>
Hi Linda,
The scenarios are nearly the same. In the scenario that I describe, packet
(size > TMTU). In the scenario that you describe (MTU > packet size > TMTU).
The difference between the two scenarios isn't significant. You should get the
same behavior either way.
Given this scenario, the GRE ingress has the following choices:
1) fragment the payload and forward each payload fragment through the tunnel.
The egress receives N complete, non-fragmented packets. Each of those packets
contains a fragment of the payload. The egress forwards the fragments to the
payload destination for reassembly.
2) do not fragment the payload. Alternatively, forward it through the tunnel,
causing the delivery (outer) packet to be fragmented. The tunnel egress router
reassembles the fragments of the delivery packet and forwards the payload to
its destination.
Today, most implementations do #1, because reassembly is computationally
expensive and they want to push computationally expensive activities out to the
endpoint.
One might argue that there are advantages to pros and cons of each strategy.
While I am not averse to documenting those, the purpose of this document
existing practice. It is not to change how people implement GRE.
Ron
_______________________________________________
Int-area mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/int-area
