On Thu, Jul 22, 1999 at 03:22:53AM +0900, itojun@iijlab.net wrote:
> 
> 	Now I see what you meant.  You are using same prefix for both
> 	tunnel and the ethernet (I should have asked you to draw a diagram).
> 
> 	Sorry this does not work at this moment.  This complicates many
> 	things, for example scope of link-local addresses.  Steve Deering
> 	likes this kind of "subnet over multiple link, connected by router"
> 	configuration, but I believe specification is not ready for this.
> 	(if "link-local" scope is just one link, like tunnel, what happens to
> 	DAD, RA, or ripng packets between ethernet 3ffe:2a00:100:3002::/64 and
> 	tunnel 3ffe:2a00:100:3002::/64?)
> 
> ==+=== 3ffe:2a00:100:3001::/64
>   | vx0
> pastaws0 ---------------------------------+ tunnel
>   | xl0					  |
> ==+=== 3ffe:2a00:100:3002::/64		  |3ffe:2a00:100:3002::/64
> 					spam

Yes, as this is similar to what you can do in IPv4 with proxy-arp, I made
an uninformed guess that  it would work in IPv6 with ND too. I see now
this is not the case. 

I think the most elegant solution would be to specify `link-local' scope
independent from the underlying physical network, and only specify it
in terms of addressing. I.e. link-local scope is e.g. all hosts with
network-prefix 3ffe:2a00:100:3002::/64. I think this would be rather
elegant, but I cannot oversee all the implementation difficulties
that might arrise from it. But it makes sense to me, especially
the gained flexibility in defining link-local hosts for underlying
networks that are not-quite-entirely-like-Ethernet, i.e. non-shared
medium ones.

The alternative would be to implement something similar as proxy-arp
for IPv6. I think that would be something we really would like to 
avoid as it has the taste of an ugly hack to it. Or a little less
ugly might be something like a meta-tunnel interface, i.e. a single
IPv6-interface representing all `real' tunnels with the same prefix,
and multiplexing traffic between the tunnels as if it was a shared
medium. This meta-tunnel would be similar to an Ethernet switch in
some ways. Not really elegant, but it would not require modification
of link-local specs and such. Hmmm, maybe I should try to implement
such a thing and see how well or how bad it works... Or... you tell
me now it is a very silly idea (if you think so, please do ;-}.

Not providing a solution is the final alternative, but wouldn't that
in the current v6 addressing scheme imply you'd need to assign a 64 bit
network prefix to each and every tunnel for everything to work
nicely? For a large ISP handing out v6 over v4 tunnels that might
eat-up address-space at a rather alarming rate I think.

Feico.