reachability issue with 3ffe:80a::/32 (PAIX IX segment)

[Jun-ichiro itojun Hagino]

*This message was transferred with a trial version of CommuniGate(tm) Pro*
# not sure why it did not get through.

	at PAIX, we use 3ffe:80a::/32 (acutally 3ffe:80a::/64) for the
	peering segment between ISPs.  3ffe:800::/24 is assigned to ISI,
	and as ISI and PAIX are not directly connected, 3ffe:800::/24 has
	two (or more) disconnected networks.

	we did not expect to receive/propagate prefixes longer than
	sTLA/pTLA prefixes (*), our EBGP routers filter out logner prefixes
	(as suggested in 6bone operation RFC) and internal routers do not have
	the route for 3ffe:80a::/32.  packets to 3ffe:80a::/32 get routed to
	ISI (instead of PAIX) and get dropped.

	(*) currently our rule is as follows:
	- prefixes that match 3ffe:0000::/17 and prefixlen = 24
	- prefixes that match 3ffe:8000::/17 and prefixlen = 28
	- prefixes that match 2001::/16 and prefixlen = 29 to 35
	- prefixes that match 2000::/3 and prefixlen = 16
	- prefixes specifically agreed with other peers

	what should we do?  if we need to receive/propagate 3ffe:80a::/32 or
	/64, we may just need to do that and then the particular problem will
	be solved.  however, i have some worry here... if we add more and more
	practice like this, we eventually get more external routes.

	i can think of couple of solutions:
	- an IX (say PAIX) gets an sTLA/pTLA.  the IX announces it to the world
	  based on normal prefix length (like /28 for pTLA).  the IX will
	  ensure connectivity between IX segments (so there will be no
	  reachability issue).
	- define an address range for IX segments (like 3ffe:fff0::/28).
	  assign /48 out of it to IX (who assigns it is another question).
	  ask everyone to accept /48 (not just /28) for the prefix.
	- do not use global address on the IX segment.  I know some routers
	  cannot establish BGP4+ peers using linklocal address.
	- get 1 sTLA/pTLA prefix for each of the IX segment.  it is infeasible.

	what do people think?

itojun