On Mon, Nov 12, 2001 at 09:42:55PM +0000, Ben Clifford wrote:
> On Mon, 12 Nov 2001, Bernd Walter wrote:
> 
> > Link local address are automaticaly co-ordinated.
> > 
> > If you receive a connect - how would you know on which interface to
> > send the answer back?
> > OK - you can remember the interface from which you got the request, but
> > this requires symetric routes.
> 
> Yes, this is basically what you would do - but you wouldn't need to store
> the specific interface, just which site.
> 
> The routes don't need to be completely symmetric - if I have interfaces
> A and B connected to site Z, if a packet comes in through interface A,
> then I can send the response out through interface B.
> 
> I just have to make sure it doesn't go out of, for example, interface C
> attached to site X.
> 
> This requires some additional kernel configuration, I think, to allow you
> to tell the kernel that links A,B are in one site and C is in another.

That way you are left with a manual configuration.
Not very differend to static address translation.

> > With link local you know for shure its symetric because it came via a
> > direct connection.
> > With site local you never know.
> 
> Whatever indicator I get to tell me that a packet came from a link-local
> address on link A, I can also use to tell me that it came from a
> site-local address in site Z.
> 
> > > > IP Packets should never leave their area of validity which is what
> > > > you are doing in your example.
> > > 
> > > I am not suggesting routing packets from one site to another - I am just
> > > saying that a particular machine may be connected to multiple sites
> > > (without routing packets between those sites).
> > 
> > Well site local addresses are defined to be fec0::/10.
> 
> > Lets say each site has enough with /48 and the leaving 38 bits are
> > filled with a site specific random value.
> > If you are ever in need to connect to another site you have a
> > (2^38)-1 : 1 chance that you don't collide.
> > If you don't do you shouldn't be surprised some day.
> 
> > After all you can always renumber.
> 
> No, absolutely not - I as an individual may have a workstation connected
> to a major corporation and a major university. I cannot tell the
> University of California to to renumber just because IBM is using the same
> random value.

OK - now I asume I understand your problem.
But think of site local addresses as beeing private addresses.
If the University of California would allow you to be connected to both
the University network and the IBM network - why would they use private
addresses.

> That is what the globally routable addresses are for. The point of
> site-local addresses are so that the site administrators can assign
> numbers with *no* interaction with the rest of the world.

That's more a situation you have when companies with their own private
addresses merge and then need to connect their private networks.

> > If you are in need for such a hack it's a good sign for a bad network
> > design.
> > The correct answer is to fix the bad design instead of working around.
> 
> I disagree.
> 
> A host may be connected to multiple sites quite legitimately. I may be
> connected to my local university, the company I work for and my local
> community wireless network. 

Right and if they would support this scenario they would use global
unique addresses.
Either you are using a network for which it is not designed for, or it
is badly designed.

> Each one of those three is an independent administration and should not be
> required to co-ordinate their site addressing, just because one out of
> 20000 users wants to connect to all three.
> 
> What is the correct network design for this scenario? Do I just take site
> local addresses from one of my connections? 

Whenever you use site local addresses you don't need to connect other
networks.

-- 
B.Walter              COSMO-Project         http://www.cosmo-project.de
[EMAIL PROTECTED]         Usergroup           [EMAIL PROTECTED]


---------------------------------------------------------------------
The IPv6 Users Mailing List
Unsubscribe by sending "unsubscribe users" to [EMAIL PROTECTED]

Reply via email to