Re: [PATCH 1/8] SH: intc: Add support OF for INTC

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 09 January 2013, Mark Rutland wrote:
> Hi,
> 
> Thanks for updating the text, this is far easier to read than previously.
> 
> However, I'm still concerned by how complex the binding seems. As I don't have
> any familiarity with the device, I don't know whether that's just an artifact
> of the hardware or something that can be cleared up.
> 
> I think the approach used by the binding needs some serious review before this
> should be merged. It seems far more complex than any existing interrupt
> controller binding. Without a dts example for a complete board (complete with
> devices wired up to the interrupt controller), it's difficult to judge how this
> will work in practice.
> 
> I've added Arnd to Cc in case he has any thoughts on the matter.

Sorry for having been absent from this discussion for so long. I didn't
realize there were 9 versions of this patch set.

I tend to agree with your interpretation above, but I may be missing
important facts from the previous review rounds.

For all I can tell, the binding is an attempt to describe the
entire drivers/sh/intc capabilities, which is probably not the
best way to approach things. The sh intc driver is not just an
irqchip driver, but rather a framework to describe arbitrary
irqchips, which is what makes this so hard.

When I first looked at the situation last year, I suggested doing
a new irqchip driver with a much simpler binding that can only
handle the irq chips from shmobile, rather than the whole thing.

I am not sure if the binding in the current form is already the
"simplified" version, or if it actually implements all the
capabilities of the intc driver.

> > +       intca: interrupt-controller@0 {
> > +               compatible = "renesas,sh_intc";
> > +               interrupt-controller;
> > +               #address-cells = <1>;
> > +               #size-cells = <1>;
> > +               #interrupt-cells = <1>;
> > +               ranges;
> > +
> > +               reg = <0xe6940000 0x200>, <0xe6950000 0x200>;
> > +               group_size = <19>;
> > +
> > +               DIRC: intsrc1 { vector = <0x0560>; };
> > +               ATAPI: intsrc2 { vector = <0x05E0>; };
> 
> This looks suspiciously like a way of encoding a device's interrupt information
> into the interrupt controller's device node. That strikes me as being the wrong
> way round.

Agreed, it would be simpler to have e.g. #interrupt-cells = <4>, to describe
the various offsets when needed (I forgot how many are actually required
in practice, rather than being computable from the other numbers), and
possibly a global interrupt-map/interrupt-map-mask property pair to map
this into a flat number space.

I need to take some more time to understand the actual requirements again,
but IIRC it would be possible to do something much simpler than the
proposed binding.

	Arnd