Re: tun: Use netif_receive_skb instead of netif_rx

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

Le mercredi 19 mai 2010 à 08:05 -0400, Neil Horman a écrit :
> On Wed, May 19, 2010 at 10:18:09AM +0200, Eric Dumazet wrote:
> > Le mercredi 19 mai 2010 à 10:09 +0200, Eric Dumazet a écrit :
> > 
> > > Another concern I have is about RPS.
> > > 
> > > netif_receive_skb() must be called from process_backlog() context, or
> > > there is no guarantee the IPI will be sent if this skb is enqueued for
> > > another cpu.
> > 
> > Hmm, I just checked again, and this is wrong.
> > 
> > In case we enqueue skb on a remote cpu backlog, we also
> > do __raise_softirq_irqoff(NET_RX_SOFTIRQ); so the IPI will be done
> > later.
> > 
> But if this happens, then we loose the connection between the packet being
> received and the process doing the reception, so the network cgroup classifier
> breaks again.
> 
> Performance gains are still a big advantage here of course.

RPS is enabled on a per device (or more precisely per subqueue) basis, and disabled
by default, so if cgroup classifier is needed, it should work as is.

Speaking of net/sched/cls_cgroup.c, I am contemplating following
sequence :

rcu_read_lock();
classid = task_cls_state(current)->classid;
rcu_read_unlock();

RCU is definitly so special (should I say magic ?), that we use it even
if not needed. It makes us happy...