[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Detecting router id change in p2p-over-lan connection

Acee Lindem said the following on 02/21/2006 06:04 AM: 
Naiming Shen wrote:

Agree with Dave's argument, and this is the same idea as in the
current draft of p2p-over-lan (section 4.5):

                                     ... If the system ID or the
   router ID of incoming hello packet does not match the system ID or
   the router ID of already established adjacency over this p2p-over-lan
   circuit, it MUST discard the packet. The implementation should offer
   logging and debugging information of the above events.

Do we need to change anything here? 


Hi Naiming,

Given that this situation can happen in the case of misconfiguration or
router-ID change, I wouldn't optimize for the former. Especially
considering Kalyan's point that if hello's are suppressed, the router-id change
will either not be recognized or require additional logic. I'd vote to either
leave it unspecified (consistent with RFC 2328's omission of router-id
change P2P link handling) or discard the old neighbor and bring up
an adjacency with the new (which is what at least one implementation does
today). 

Hello Acee,

The original idea in the draft was to prevent multiple routers joinning
onto the same LAN rather than handing the router-ID change from the
same router. Even through there is a side effect to it.

For the router-ID change issue. If one is willing to change router-ID
or system-ID during normal operation, there are more things broken
and there is nothing wrong to wait until the current adjacent timeout.
If the hello is suppressed over the link, then the behaviour is not
bounded by this paragraph since it does not say the purpose is to
detect neighbor router-id change. Also there is little reason to
suppress a hello over p2p-over-lan links. Thus I don't think
router-id is an issue with this sentence.

Since p2p-over-lan is different from normal p2p circuit that multiple
routers can easily join the same LAN unintentionally. If we ignore
the new station to join, then we can prevent the adjacency constant
flapping(this does not happen to normal p2p link), the worst it can
happen is slow to handle router-id change as mentioned above, which
is a non issue. Thus I think it's a good thing to ignore the extra
adjacency joinning. This sentence does not govern the normal p2p link
behaviour so the implementation is ok(even prefered) to reset the
existing adjacency on a normal p2p link. Thus I don't think we
should say we must reset the current adjacency for p2p-over-lan.

My preference of this sentence in the draft is with this order:

 - leave it as is
 - change the MUST to SHOULD
 - remove the router-id related sentence

thanks.
- Naiming


Thanks,
Acee

Thanks,
Acee


thanks.
- Naiming

Dave Katz said the following on 02/20/2006 08:53 PM:

RFC2328 is silent on what to do if you see a different router ID on a P2P link; if you take it literally, you will create two adjacencies (temporarily), advertise both in your router LSA, and your network will most likely black hole traffic until one of them times out.

On real P2P links, this will suffice (if you don't mind a network outage for the dead time of the old adjacency) and things will eventually stabilize since it is "impossible" (cough) two get two neighbors in 2Way. The fact that it is underspecified doesn't matter that much.

In the P2P-over-LAN case, there needs to be a mechanism to try to do something useful if there turn out to be more than two routers on the link. I don't think it's reasonable to say "gee it's misconfigured, we don't care what happens to the network."


It seems to me that there are two possible cases, namely, use the old adjacency until it times out (effectively ignoring the hellos from the new guy) or to form an adjacency with the new guy (and blow away or otherwise ignore the old guy.)

The first case has the advantage of stability and simplicity--the third router that shows up for the party will be ignored by both parties. There is a potential for a circle of one-way adjacencies to stabilize, but this will still be stable (there won't be a usable adjacency across the subnetwork.)

The second case has the advantage that it adapts more quickly when something changes. Some implementations already do this for the pure P2P case. However, it is destabilizing in the misconfiguration case, as the router will flap its router LSA each time another Hello is received. Probably the only reasonable way out would be to detect the flap and then choose one neighbor and stick to it, ignoring the other (which degenerates into the first case.)

Given all of this, I'd propose that the right thing to do is the first case, and suffer through the dead time (or use BFD and suffer through the detection time.)

(A third possible case would be to actually bring up the adjacency and end up running full-mesh multipoint, which would probably even work, but would be butt ugly.)

--Dave


On Feb 20, 2006, at 3:25 PM, Acee Lindem wrote:

Kalyan Bade wrote:

Acee,


This might not work if the interface is defined as demand circuit.



Good point - We're in the process of discussing whether this behavior
should be in the
draft. Possibly, handling detection of a router-id change on a P2P

over

LAN link
differently than a normal P2P link may be a bad idea.


How about bringing down the neighbor (rather than ignoring) when a third
neighbor on a P2P over LAN sends a new hello? This creates unnecessary
flaps, but again having more than 2 routers on a p2p LAN is a misconfig.
Atleast, the processing will be similar to a regular p2p interface.

That would be fine and consistent with what most routers do on normal P2P links. I think the
P2P over LAN specific router-ID change processing should be removed from the draft.

Thanks,
Acee

Thanks,
Kalyan.