On Wed, Jan 04, 2017 at 02:26:09PM +0100, Steffen Klassert wrote:
> On Wed, Jan 04, 2017 at 04:34:15AM -0800, Eric Dumazet wrote:
> > >
> > > @@ -4843,7 +4853,12 @@ static int process_backlog(struct napi_struct
> > > *napi, int quota)
> > >
> > > while ((skb = __skb_dequeue(&sd->process_queue))) {
> > > rcu_read_lock();
> > > - __netif_receive_skb(skb);
> > > +
> > > + if (skb_xfrm_gro(skb))
> > > + napi_gro_receive(napi, skb);
> > > + else
> > > + __netif_receive_skb(skb);
> > > +
> >
> >
> > But napi here is device independent. It is a fake NAPI, per cpu.
> >
> > I am not sure of the various implications of using it at this point,
> > this looks quite dangerous/invasive to me, compared to the gro_cells
> > infra which was designed to have no impact on core code paths.
> >
> > To me, the caller should call napi_gro_receive() skb, instead of letting
> > core networking stack add this extra skb bit and extra conditional.
>
> I had a quick look at the gro_cells, it looks like I could avoid
> at least the extra codition with this.
On a second view, it does not look so promising. I don't
have a netdevice where I can hang this off. Looks like I
need such a fake per cpu NAPI as the backlog is. I could
create my own one, then I would not have to add this
condition to core networking.
