Ole, > -----Original Message----- > From: Ole Troan [mailto:[email protected]] On Behalf Of Ole Troan > Sent: Monday, February 22, 2010 2:47 PM > To: Templin, Fred L > Cc: softwires; DHC WG > Subject: Re: [Softwires] SOFTWIRE working group last call on 6rd > > Fred, > > >>> As a first comment, the document is lacking a discussion of the > >>> implications of having the BR use its anycast address as the source > >>> address for the packets it relays. > >>> > >>> Take for example the case of two BRs A and B that configure the > >>> same IP anycast address, and a CE router C within the 6rd domain. > >>> If BR A forwards packets toward C, but the packets are lost in the SP > >>> network, any resulting ICMPs could just as easily flow back through > >>> BR B instead of A. Then, if the ICMPs don't contain enough information > >>> for translation, BR B has no way to send a translated message back > >>> to the original source and a black hole can result. > >> > >> BRs are stateless. it does not matter which one the ICMP comes back to. if > >> the ICMP message > doesn't > >> contain enough of the IPv6 packet there is nothing it can do, regardless > >> of which one got the ICMP > >> message. the use of the anycast address doesn't change that. > > > > Then, there needs to be an analysis of the domain of > > applicability. For example, a totally stateless 6rd > > solution may be incompatible with SP networks in which > > there may be ICMP messages that do not return enough > > information for translation. > > > > If the BRs were allowed to keep a small amount of state, > > however, then they would be able to return appropriate > > ICMPs to the IPv6 host on the outside even if the ICMPs > > coming from the SP network on the inside did not include > > enough information. Provided, that is, that the ICMPs > > within the SP network are returned to the correct BR. > > That is where a unicast source address instead of an > > anycast one would steer the ICMPs to the correct BR. > > you are suggesting that a router should store a copy of every IPv6 header > plus IP tunnel header for > _every_ packet it forwards? and it should do this at wire-speed? > how much do you want to pay for this box? ;-)
No; that's not what I'm suggesting. I am suggesting that for only those CEs for which ICMP unreachables are being received that the BR store an MTU value and a reachable flag. Then, when new packets come in from an outside IPv6 host, the BR can drop the packet and return an ICMP if the packet would violate the MTU or if the CE has recently been known to be unreachable. > are you talking about pre-RFC1812 IPv4 routers? in that case you wouldn't > have enough information to > correlate that accurately with the stored IPv6 packet in any case. The BR does not need to know anything beyond the minimum 8 bytes in the ICMP; all it needs to know is whether the CE is reachable, whether there is an MTU limitation on the path, etc. Then, it can return suitable unreachables or PTBs to the source host. Fred [email protected] > cheers, > Ole _______________________________________________ Softwires mailing list [email protected] https://www.ietf.org/mailman/listinfo/softwires
