[openib-general] re the core network changes to support network event notification

Steve Wise swise at opengridcomputing.com
Tue Aug 8 07:31:28 PDT 2006


On Tue, 2006-08-08 at 17:15 +0300, Or Gerlitz wrote:
> Steve Wise wrote:
> > On Tue, 2006-08-08 at 13:54 +0300, Or Gerlitz wrote:
> 
> Steve,
> 
> OK, thanks for all the clarifications and information re the rtnetlink 
> and why its an overkill to use it (vs netevents) within the kernel.
> 
> >> I guess IPoIB code would be able to take advantage of RTM_NEIGHUPD netlink event
> >> or a netevents based filter to remove the data path memcmp on the neigh HA address.
> 
> > I haven't looked at the IPoIB code.  What is it doing exactly (I am
> > somewhat familiar with the IPoIB specification and how it works)?
> 
> http://kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=8a7f752125a930a83f4d8dfe37fa5a081ab19d31
> 
> Take a look on this "IB/ipoib: Fix packet loss after hardware address 
> update" patch, it overcomes a situation where the neighbor structure is 
> **updated** (for example the HA address is changed, as of gratuitous 
> ARP) but as the neighbor destructor is not called, IPoIB is not aware to 
> that and hence does not update the AH (IB Address-Handle) associated 
> with the HA.
> 
> So the current solution was to memcmp the neigh info with what IPoIB 
> knows about this nieghbor on the data path.
> 
> With your patch, i guess this memcmp can be eliminated.
> 

I think so. With netevents, I think you'll update your AH entries when
the HA changes as opposed to doing the memcmp() on every xmit.  






More information about the general mailing list