Re: [PATCH RFC] act_cpu: packet distributing

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

Le vendredi 16 juillet 2010 à 07:30 +0800, Changli Gao a écrit :
> On Fri, Jul 16, 2010 at 12:54 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:

> > Even if we solve this locking problem, using percpu variables, act_cpu
> > hits another problem :
> >
> > The socket refcount, taken by the 'master' cpu, and released by the
> > consumer cpu.
> This is why I asked if I can assign sk to skb.
> 

This assignement doesnt change the cache line ping pong problem.

Let me explain for you :

You lookup the TCP or UDP socket. This automatically takes a refcount on
it (atomic operation on sk->sk_refcnt).

Then you assign sk to skb.  (skb->sk = sk)

Then you transfert skb handling to another cpu (IPI cost already in RPS
as we all know)

This remote cpu finds skb->sk (and avoids the lookup/refcnt) and handles
the packet through TCP/UDP stack.

This remote cpu _releases_ the socket refcount.

Conclusion :

"The socket refcount, taken by the 'master' cpu, and released by the
consumer cpu."