On (06/01/17 00:41), Julian Anastasov wrote:

	So, we do not hold reference to neigh while accessing
its fields. I suspect we need to move the table lock from
neigh_remove_one here, for example:

good point, let me think over your suggestion carefully (it sounds
right, I want to make sure I dont miss any other race-windows)
and post patch v4 tomorrow..

	Another solution to cause faster removal would be
to cancel the gc_work and to schedule it after 1 jiffie.
It helps when many entries are deleted at once but the
work prefers to just sleep when gc_thresh1 is not reached,
so such solution is not good enough.

Right the other drawback of relying on gc for cleanup is
that it doesn't give higher preference to cleaning up FAILED
entries first- so it can end up cleaning up other useful entries
(as I was pointing out to David Ahern)

--Sowmini