[ofa-general] Re: A question about tx lock in ipoib_flush_paths

[akepner at sgi.com]

On Fri, Jul 10, 2009 at 01:44:14PM -0700, Roland Dreier wrote:
> ...
> Sorry for not replying to this sooner.  I'd prefer not to take that
> approach, because all this code is complex enough that it is hard to
> maintain already; adding RCU on top of that just seems like it will make
> things even harder to fix in the future.
> 

Sigh. But I can understand that preference.

> I wish someone would come up with a way to slice through the gordian
> knot and simplify the neighbour handling -- maybe the whole strategy of
> stashing things in the network stack's neighbour structure is wrong?

Yes, I think it is basically wrong. It was a good hack, but it subtly 
broke when CM was added.

> Maybe the core network stack could be tweaked to make things easier for
> ipoib to deal with?
> 

I can investigate a different approach (but I'd certainly appreciate 
any good ideas!)

In the meantime, we have found this patch to be a very effective 
workaround in practice:

http://lists.openfabrics.org/pipermail/general/2009-June/060205.html

It's obviously not a complete fix, but it improves the situation, and 
is very simple. Could we add that as a stopgap measure for now?

-- 
Arthur