[PATCHv2 02/10] ARM: vic: MULTI_IRQ_HANDLER handler

[jamie@jamieiles.com (Jamie Iles)]

On Wed, Nov 02, 2011 at 01:40:24PM +0000, Russell King - ARM Linux wrote:
> On Thu, Sep 29, 2011 at 10:30:09AM +0100, Jamie Iles wrote:
> > +#ifdef CONFIG_MULTI_IRQ_HANDLER
> > +static void vic_single_handle_irq(struct vic_device *vic, struct pt_regs *regs)
> > +{
> > +	u32 stat, irq;
> > +
> > +	stat = readl_relaxed(vic->base + VIC_IRQ_STATUS);
> > +	while (stat) {
> > +		irq = ffs(stat) - 1;
> > +		handle_IRQ(irq_domain_to_irq(&vic->domain, irq), regs);
> > +		stat &= ~(1 << irq);
> > +	}
> > +}
> > +
> > +asmlinkage void __exception_irq_entry vic_handle_irq(struct pt_regs *regs)
> > +{
> > +	int i;
> > +
> > +	for (i = 0; i < vic_id; ++i)
> > +		vic_single_handle_irq(&vic_devices[i], regs);
> > +}
> 
> And if we receive another interrupt after the read of the register, we'll
> have to exit all the way back (possibly to userspace) before re-entering
> the IRQ handling paths back to this point to process it.

OK, so how about something like this instead:

static int vic_single_handle_irq(struct vic_device *vic,
				 struct pt_regs *regs)
{
	u32 stat, irq;
	int handled = 0;

	stat = readl_relaxed(vic->base + VIC_IRQ_STATUS);
	while (stat) {
		irq = ffs(stat) - 1;
		handle_IRQ(irq_domain_to_irq(&vic->domain, irq), regs);
		stat &= ~(1 << irq);
		handled = 1;
	}

	return handled;
}

asmlinkage void __exception_irq_entry vic_handle_irq(struct pt_regs *regs)
{
	int i, handled;

	do {
		handled = 0;
		for (i = 0; i < vic_id; ++i)
			if (vic_single_handle_irq(&vic_devices[i], regs))
				handled = 1;
	} while (handled);
}

which I think should keep handling IRQ's until no VIC has them pending 
(or as best can be determined).

> Is there any particular reason folk are destroying the built-in efficiency
> of the IRQ handling which is common-place in the existing assembly
> approach?

Well this approach makes a single image kernel a bit easier.  The other 
thing is that it plays a lot nicer with dynamic irq_desc assignment.  
Grant's IRQ domain patches make this quite easy here, but I can't see an 
obvious way to do that with the assembly method.

Jamie