From: Patrick McHardy <kaber@xxxxxxxxx>
Date: Wed, 15 Apr 2009 12:41:01 +0200
Herbert (I think) suggested to make address list updates in softirq
context a two-step process, where addresses would first be added to
a temporary list and the final change would be done in process context
while holding the RTNL.
Given the complicated mess we currently have, this would be a very
worthwhile change IMO.
This would break the IPV6 TAHI tests if you think we could use
such an idea for that.
When IPV6 packets arrive that influence multicast and unicast
address lists, the effect must be essentially immediate. Such
that a subsequent packet will cause the kernel the behave
with the necessary side effects, no matter how quickly that
next packet arrives.