[PATCH 0/5] Fix regression introduced by set_irq_flags() removal

[tglx@linutronix.de (Thomas Gleixner)]

Thomas,

On Sat, 5 Dec 2015, Thomas Gleixner wrote:
> On Fri, 4 Dec 2015, Thomas Petazzoni wrote:
> > Well, the problem is that IRQ_NOAUTOEN is a global flag, which is OK
> > for global interrupts, but not good for per-CPU interrupts, since you
> > don't have the information on a per-CPU basis of which interrupt was
> > enabled before suspend, and therefore should be re-enabled after resume.
> > 
> > Until now, we don't have the problem since the only per-CPU interrupt
> > we were using was the local timer interrupt, and the local timers on
> > secondary CPUs are switched off during suspend and re-enabled during
> > resume. So re-enabling the interrupt on the boot CPU on resume is
> > sufficient.
> > 
> > However, our network driver recently switched to using per-CPU
> > interrupts as well, and in this case, it is really important to be able
> > to re-enable the per-CPU interrupts and the appropriate CPUs at resume
> > time. Since our HW registers are made so that it is not possible to
> > read out at suspend time which interrupts are enabled, we have to ask
> > the Linux kernel at resume time which interrupts should be re-enabled
> > at the HW level. Which is what my more complicated series was doing.
> > 
> > Do you have other suggestions to allow us to know which per-CPU
> > interrupts should be re-enabled on the different CPUs at resume time ?
> 
> Ok. That makes sense. So I'm going to pick up the core change.

Second thoughts. That network driver example does not make sense.

You have a suspend/resume mechanism and a cpu hotplug machinery in
that driver, right? So that should be responsible for
disabling/enabling the per cpu interrupts. I don't think it's the
proper way to do that in the irq chip driver at some random point
during resume as you'd reenable interrupts on cpus which are not
online yet.

Thanks,

	tglx