> > +#ifdef CONFIG_NET_POLL_CONTROLLER
> > +static void pxa168_eth_netpoll(struct net_device *dev)
> > +{
> > + struct pxa168_eth_private *pep = netdev_priv(dev);
> > + napi_schedule(&pep->napi);
> > +}
> > +#endif
>
> This definitely is not sufficient.
>
> Look at what other drivers do.
Sorry, I did that, and got confused because some drivers bracket the interrupt
handler with disable_irq/enable_irq, some other drivers use local_irq_save/
local_irq_restore (which seems like a no-op because netconsole already uses
spin_lock_irqsave and netpoll_send_udp checks irqs_disabled), and a few drivers
call bare napi_schedule.
> You have to invoke the interrupt handler with the device's interrupts
> disabled,
Is this required for correctness? I have to admit I'm unsure why.
> collect the events, and most importantly mask chip interrupts before
> scheduling
> the NAPI instance.
And is this a matter of efficiency (not calling napi_schedule when there's
nothing to do and not keeping interrupts enabled for its duration), or also
a matter of correctness?
Sorry I'm not sending a revised patch yet, but without a firm understanding
of why changes are needed that would be a bit of a sin.
Thanks.
Alexander
