Re: [PATCH v6 1/2] Documentation: bindings: add dt doc for Rockchip PCIe controller

[Brian Norris <briannorris@chromium.org>]

Hi Shawn,

On Wed, Jul 13, 2016 at 09:10:15AM +0800, Shawn Lin wrote:
> 在 2016/7/7 8:39, Brian Norris 写道:
> >On Wed, Jul 06, 2016 at 03:16:37PM +0800, Shawn Lin wrote:
> >>+	#interrupt-cells = <1>;
> >>+	interrupt-map-mask = <0 0 0 7>;
> >>+	interrupt-map = <0 0 0 1 &pcie0_intc 1>,
> >>+			<0 0 0 2 &pcie0_intc 2>,
> >>+			<0 0 0 3 &pcie0_intc 3>,
> >>+			<0 0 0 4 &pcie0_intc 4>;
> >
> >I'm a little lost on this one, so forgive my ignorance; how did you
> >determine the last value in each entry (i.e., the 1, 2, 3, and 4 IRQ
> >numbers for pcie0_intc)? IIUC, those are supposed to represent indeces
> >into the IRQ status register found in the PCIe interrupt status
> >register, and so they should be 0-based (i.e., 0, 1, 2, 3). And then
> >you'd have:
> >
> >	interrupt-map = <0 0 0 1 &pcie0_intc 0>,
> >			<0 0 0 2 &pcie0_intc 1>,
> >			<0 0 0 3 &pcie0_intc 2>,
> >			<0 0 0 4 &pcie0_intc 3>;
> >
> >But then, I never got this sub-node binding to work quite right, so I
> >may be missing something.
> >
> >EDIT: ooh, I see what's going on! I'll comment on the driver as well,
> >but it looks like you're translating the register status to a HW IRQ
> >number with 'ffs(reg)', which yields a 1-based index. I think it is most
> >sensible to use a 0-based index (i.e., 'ffs(reg) - 1'). Now, that only
> >will work if you get the whole interrupt-map + interrupt-controller
> >thing right (i.e., using a subnode for the interrupt controller) --
> >otherwise, IRQ mapping might not work right. I suspect that's one reason
> >the original driver writer might have used 1-based indexing in the first
> >place.
> 
> yes, I got it but.....what's the difference?

At some level, it's a matter of preference. But when you're talking
about the rk3399 PCIe "interrupt controller" domain, it seems that you
should be talking about HW bits in the controller -- i.e., you have a
4-bit interrupt status bitfield, that we typically call [0:3]. If you
use [1:4], then you have to remember to subtract 1 mentally when mapping
to the actual HW bit. I believe that confusion (since bitfields normally
count from 0) might have helped cause the infinite loop bug I noticed
too. And I also think that counting from 0 helps clarify the fact that
your interrupt controller indexing is an independent numbering from the
PCI interrupt numbering, even though they happen to map 1:1.

But then, PCI INTx numbering is kinda weird already, as it starts from
1. So maybe it's just as valid to say our domain starts from 1 as well.

> You still need to get the whole interrupt-map + interrupt-controller
> things right and the code(ffs(reg) - 1)if applied your suggestion.

Yes, of course. And I already sent you patches that do that.

> Look at most of the docs for pcie bindings, I saw they also take
> 0-base index, how about?

I don't know which ones you're referring to. I see that altera-pcie.txt
supports interrupt indeces counting from 1, but that's probably because
they're using the same broken binding that was in your ~v3 patches
(where the pcie node has both 'interrupt-controller' and
'interrupt-map', with phandles to itself), so they had no other choice.

If you still think it makes more sense to count from 1, then I won't
stop you.

Regards,
Brian