Re: [RFC] NETDEV_UNREGISTER_BATCH seems unused nowaday ?

On Fri, 2012-08-10 at 07:45 -0700, Eric W. Biederman wrote:

Can someone help bring me up to speed.  What has changed in the
dst ref counting that has invalidated our previous solutions?

In fact your patch (850a545bd8a416484 net: Move rcu_barrier from
rollback_registered_many to netdev_run_todo.) reinstated the problem
again. You didnt notice, but other people can see the problem.

As for the idea of putting an rcu_barrier inside of the rtnl_lock.  I
really don't like it. You are trading off a 1000ms singled threaded wait
without locks for extending the hold times of rtnl lock by 12ms or so.

We probably can keep the rcu_barrier() in netdev_run_todo() and kick the
UNREGISTER_FINAL in netdev_run_todo() too.

I'll try that.

We already have an rcu_barrier on that path in netdev_run_todo,
so we can reorganize things to use that barrier I would be much
happer.  Furthermore I talked to Paul McKenney a while ago
about creating an rcu_barrier expedited and he really did not
like the idea.

Reading through the code we really should get dst_rcu_free
out of the header and make it non-line.  dst_rcu_free can't
possibly be called from a location where it can be inlined.

Trying to understand your analysis I have stared at the code for
a while and I am definitely not seeing any rcu callbacks that
result in calling rt_free.  So one of us is missing something.

Recent commits add the rt_free() calls.

Of course, if you unregister a dummy device you wont see the problem.

If you unregister a real device, you definitely hit the problem.