[ofa-general] [RFC] ipoib: avoid using stale ipoib_neigh* in ipoib_neigh_cleanup()

[Or Gerlitz]

akepner at sgi.com wrote:
> On Sun, May 31, 2009 at 02:34:56PM +0300, Or Gerlitz wrote:
>> do we agree that accessing the neigbour from the cm tx completion flow is buggy?
> But the fundamental problem is that the ipoib_neigh structures aren't read and written using consitent locking (particularly in ipoib_neigh_cleanup()). 
I am still not with you as for the root cause of the problem... if by 
the time the neigh cleanup code is called, no locking would be required 
(since on one should have refs in that step) --> no consistency is needed.

 From what I understand the only code that can race with the neigh 
cleanup is the cm tx completion, do you agree to this assertion? do you 
see the race happening when working in datagram mode?

Or.