Re: [net-next,1/3] bonding: fix vlan 0 addition and removal

[Nikolay Aleksandrov <nikolay@redhat.com>]

On 08/06/2013 10:59 AM, Veaceslav Falico wrote:
> On Tue, Aug 06, 2013 at 10:39:22AM +0200, Nikolay Aleksandrov wrote:
>>> From 1c89abefebe90568ed52d2df59fcfdd650bc4696 Mon Sep 17 00:00:00 2001
>>> From: Veaceslav Falico <vfalico@redhat.com>
>>> Date: Mon, 5 Aug 2013 23:29:12 +0200
>>> Subject: [PATCH] bonding: add vlan_uses_dev_rcu() and make bond_vlan_used()
>>> use it
>>>
>>> Currently, bond_vlan_used() looks for any vlan, including the pseudo-vlan
>>> id 0, and always returns true if 8021q is loaded. This creates several bad
>>> situations - some warnings in __bond_release_one() because it thinks that
>>> we still have vlans while removing, sending LB packets with vlan id 0 and,
>>> possibly, other caused by vlan id 0.
>>>
>>> Fix it by adding a new call, vlan_uses_dev_rcu(), which is the same as
>>> vlan_uses_dev(), but uses rcu_dereference() instead of rtnl, and thus we
>>> can use it in bond_vlan_used() wrapped in rcu_read_lock().
>>>
>>> Also, use the pure vlan_uses_dev() in __bond_release_one() cause the rtnl
>>> lock is held there.
>>>
>> Just 1 more note, you can't trust nr_vlan_devs under RCU.
> 
> Yes, you're right, however we actually don't care anyway if we race with
> (un)register_vlan_dev() - we'll end up either in using the (un)registered
> vlan or not, and in both cases it's ok. So I don't see a real problem here,
> tbh, though I'll look into this also.
You might have stale value in the cache, the implications don't stop there.
I'd like to avoid inconsistent behaviour if there's a way.
A solution that can be relied on and works always would be much more
preferable.