[ofa-general] Re: [openib-general] Fw: [PATCH] enable IPoIB only if broadcast join finish

[Jason Gunthorpe]

On Tue, Mar 06, 2007 at 08:00:23PM -0800, Shirley Ma wrote:
 
> Comparing to the IPoIB accessibility(let IPv4 working) with playing trick to
> carrier on to avoid IPv6 link local DAD in a small possibility, this patch is a
> better choice for switches with limited MGCs resouce today in a large cluster.
> Then IPoIB will have the same behavior as IPv4 and IPv6 over ethernet:
> interface will be up and running no matter whether there is any duplicated
> address or not.

Don't get me wrong, I think your patch is ultimately the right way to
go, but it needs another part to address the problem IPv6 has - or at
least a plan on how to address it. I don't think ignoring
the synchronizing problem is the way to go.

Also, in my view, the problem you are seeing with MLID exhaustion is
purely a SM problem and has nothing to do with IPoIB and switch
limits. SMs need to treat MLIDs as a precious resource and share them
agressively. Especially IPv6 solicited node multicast addresses.

How about this as a nice simple alternative..

If IPoIB is asked to send to a multicast address that is not yet
joined (join is pending to the SA, or whatever) then it uses the
broadcast MLID for the packet. This is similar to how ethernet works
and would let DAD and RS work correctly in all cases.

Jason