On Sun, 2004-10-17 at 16:14, Denis Vlasenko wrote:
> > Your patch:
> >
> > + preempt_disable();
> > netif_rx_ni(skb);
> > + preempt_enable();
> >
> > just wraps this code in preempt_disable/enable:
> >
> > static inline int netif_rx_ni(struct sk_buff *skb)
> > {
> > int err = netif_rx(skb);
> > if (softirq_pending(smp_processor_id()))
> > do_softirq();
> > return err;
> > }
> >
> > Isn't this considered an incorrect use of preempt_disable/enable? My
> > reasoning is that if this was correct we would see preempt_dis/enable
> > sprinkled all over the code which it isn't.
> >
> > Why do you have to call do_softirq like that? I was under the
> > impression that you raise a softirq and it gets run later.
>
> There is a possibility that this guy just wanted to fix his
> small problem.
>
Yes, that is what I thought. The question was more directed at the
list. I added netdev to the cc:.
I looked at Robert Love's book and I am still unclear on the use of
do_softirq above. To reiterate the question: why does netif_rx_ni have
to manually flush any pending softirqs on the current proccessor after
doing the rx? Is this just a performance hack?
Lee
|