Re: [PATCH] net: dev: fix the incorrect hold of net namespace's lo device

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2012-08-22 at 19:00 +0800, Gao feng wrote:

> Hi Eric
> 
> I saw your patch and think this patch is clear and doesn't change too much logic.
> 
> I test your patch, it not fix this problem.
> 
> In my test case,when moving a net device to another net namespace,
> Because you patch delete NETDEV_UNREGISTER event from dst_dev_event,
> we will just put dst entries into the dst garbage list in event
> NETDEV_DOWN,without call dst_ifdown to change these dst entries' device
> to the lo device,and now this net device belongs to the new net namespace.
> 

Then fix the "moving a net device to another net namespace", instead
of slowing down other common operations.

dev_change_net_namespace() is probably a better place to put your patch

> After the net device beeing moved to another net namespace, I rmmod this
> net device's driver,this will trigger the new added event NETDEV_UNREGISTER_FINISH,
> so in dst_dev_event,we will change these dst entries's device to the new net
> namespace's lo device,and this will make the referenct count of the new net namespace's
> lo device incorrect. when we exit the new net namespace,this emg message is still exist.
> 
> Message from syslogd@Donkey at Aug 22 18:50:13 ...
>  kernel:[ 1161.979036] unregister_netdevice: waiting for lo to become free. Usage count = 1
> 
> And because net_mutex is locked here,so we can't create new net namespace.
> 
> > rcu_barrier() at this place will kill some workloads.
> > 
> 
> I think this will only add some workloads when unregister a net device.
> Do I miss something?

Yes, rcu_barrier() at this point is killing performance, because we hold
RTNL.

We worked hard to batch things, your patch is a huge step backward.