The MPLS WG Archive

[Giles Heron <giles@level3.net>]

I would agree with you on this one.

Your second point is key.  How can you know that the destination address
in the packet is an IGP prefix?  I suspect that many networks will want
to run BGP edge-to-edge, but to have only IGP routes in the core.

In addition to this you have the issue of MPLS VPN and Layer 2 over MPLS
traffic.  Either one of these will generate a label stack in which the
innermost label is assigned edge-to-edge and is not visible to the
intermediate devices.

Giles

David Charlap wrote:
> 
> Here's a hypothetical network:
> 
>         A ----- B ----- C
> 
> The network is running LDP and is configured for unordered, downstream
> unsolicited label distribution.
> 
> Assume that the IGP knows about one prefix, which is reachable through
> router C.
> 
> Suppose now, that B sends out a label mapping for this prefix, and C has
> not yet sent out its label mapping.
> 
> A, since it has received a label mapping, decides to start labelling
> traffic for this prefix, and forwards it on to B.
> 
> B now receives this traffic, but does not have a label mapping for the
> FEC.  What should it do?
> 
> I would assume, based on section 3.22 of RFC 3031 that router B should
> discard these packets until it receives a label mapping from C.
> 
> A coworker here believes that router B should pop the label stack and
> forward the unlabeled packet based on its routing tables.  He claims
> that section 3.22 doesn't really apply.  His rationale is that:
> 
>         - We can assume no routing loops, because the LSP from ingress
>           to this point is following the same IGP shortest path that
>           it would be forwarding along.  If there's a loop in the IGP,
>           it will still be there after receiving a label mapping.
> 
>         - We know we'll be able to forwrad the packet based on its IP
>           header, because router A used that same header against the
>           same IGP tables when it determined the initial label for the
>           packet.
> 
>         - C may (for some reason) choose to never send a label mapping
>           for the prefix.  Since the network is downstream unsolicited,
>           it may never receive a request.  It wouldn't make sense for
>           this legal scenario to result in a black hole.
> 
> I counter his arguments by claiming that:
> 
>         - We don't know there won't be routing loops.  The IGP may not
>           yet have converged to a stable state.  When the last label
>           mapping message is sent, the loop detection features of LDP
>           (if turned on, of course) will detect it and prevent any
>           traffic from flowing until the IGP stabilizes.
> 
>         - We can't assume that B has the same routing table that A and
>           C has.  If A and C are border routers, they may be running an
>           EGP (like BGP) and know about far more prefixes than a core
>           router, which would only be running an IGP (like OSPF).
> 
>           The FEC for the LSP may be one of C's addresses, and the
>           packet destination may be an address many hops beyond C.  B
>           does not necessarily have any clue how to get to the
>           destination address in the absence of the LSP.
> 
>         - C will eventually send a label mapping for everything in its
>           routing table.  It would be a broken implementation for it to
>           arbitrarily choose to not map a prefix without deliberate
>           operator intervention.  If an operator chooses to configure
>           this, then he probably intends to create a black hole.  If
>           he really wants the label to be popped, he can send a NULL
>           label mapping an explicitly request it.
> 
> So, which of us is correct here?
> 
> -- David

-- 
===================================================
Giles Heron - IP Architect - Level 3 Communications
phone: +44 20 7864 0719     mobile: +44 7880 506185
===================================================