On 15/11/2012 10:19, Mattia Rossi wrote:
> On Thu, Nov 15, 2012 at 9:16 AM, Brian E Carpenter <
> [email protected]> wrote:
> 
>> On 14/11/2012 22:44, james woodyatt wrote:
>>> On Nov 14, 2012, at 13:34 , Mikael Abrahamsson <[email protected]> wrote:
>>>> I've always seen it to be solved via some kind of source based routing
>> automatically discovered between the ISP routers.
>>>
>>> My point is that it isn't sufficient to handle this problem at just the
>> routers.  At a minimum, the *hosts* need to be told which default router to
>> use with each source prefix.
>>
>> Of course. I suggested this when MIF started out, and it's a MIF
>> issue still.
>>
> 
> The hosts do already select the correct router depending on the prefix.
> They're tied together. An RA contains a prefix and router address. That's
> what the host keeps in memory.

Where is that specified? And what if the network is using DHCPv6
to provide such information?

The strongest statement I can find in RFC 6724 about this is a
MAY in section 7, which doesn't make this predictable.

> If it's two RA's one router becomes default, the other more specific. 

The two might be of equal standing and both offer a route to ::/0.

> The
> host/applications will also only use one prefix at a time, thus always send
> the packets down the correct path.

I don't believe that is fully specified either. Each new socket
might make a new choice.

> In my experience it was always the same prefix, the one that got registered
> first (if no preference was setup in the advertising router).
> If the host is connected via two or more interfaces (so we're in MIF area
> now), there will always be one preferred prefix, and interface, and the
> outgoing routing will work.

Then why is MIF still arguing?

> If applications are able to chose a specific prefix (e.g. VPN) they usually
> implement some source routing on the host anyways, and send the packets to
> the router registered with that prefix.

Sure, but we are worried here about zero-conf defaults.

> If you have an application listening for incoming connections, then it's
> important that the application is smart enough to use the address on which
> the incoming connection arrived as source address, and not the default one
> which might be on a different prefix to avoid asynchronous communication.
> As far as I know this is already common practice. 

Not good enough - it needs to be specified.

> So I don't really see a
> problem here, unless you want to do some fancy load sharing and play ping
> pong with source addresses.
> If there's only one interface which connects to a multihomed router, the
> principle is the same, but the multihomed router must be able to perform
> source address routing, and forward the packets down the right path. And
> this is something I'm not too sure about routers can currently do.

Indeed. Again, this needs specifying, as does behaviour when
there are multiple routers and one of them receives a packet
that should exit via another one.

https://datatracker.ietf.org/doc/draft-ietf-v6ops-ipv6-multihoming-without-ipv6nat/

   Brian

> Mat
> 
_______________________________________________
homenet mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/homenet

Reply via email to