On Fri, Aug 19, 2011 at 01:16:09PM +0900, Magnus Damm wrote:

On Fri, Aug 19, 2011 at 10:10 AM, Simon Horman [off-list ref] wrote:

quoted

This is intended to make it easier to correctly order IRQs.

As suggested by Guennadi Liakhovetski.

quoted

--- a/arch/arm/mach-shmobile/board-ag5evm.c
+++ b/arch/arm/mach-shmobile/board-ag5evm.c

@@ -352,15 +352,15 @@ static struct resource sdhi0_resources[] = {

               .end    = 0xee1000ff,
               .flags  = IORESOURCE_MEM,
       },
-       [1] = {
+       [1 + SH_MOBILE_SDHI_IRQ_CARD_DETECT] = {
               .start  = gic_spi(83),
               .flags  = IORESOURCE_IRQ,
       },
-       [2] = {
+       [1 + SH_MOBILE_SDHI_IRQ_SDCARD] = {
               .start  = gic_spi(84),
               .flags  = IORESOURCE_IRQ,
       },
-       [3] = {
+       [1 + SH_MOBILE_SDHI_IRQ_SDIO] = {
               .start  = gic_spi(85),
               .flags  = IORESOURCE_IRQ,
       },

Hm...

So I know you guys have been discussing this back and forth, so I'm
not sure if jumping in to this thread makes things any better. But
anyhow, here are my opinions. Feel free to ignore them. =)

First of all, having some kind of association with each IRQ is a good
thing. I am however not convinced that using the index number of the
platform device resource irq is the best option. Consider the case
when someone modifies the SDHI resource in the code above to only
include this interrupt:

        [1 + SH_MOBILE_SDHI_IRQ_SDCARD] = {
                .start  = gic_spi(84),
                .flags  = IORESOURCE_IRQ,
        },

quoted

From my point of view, the common sense for this would be that only

the SDCARD interrupt would be enabled and the rest would be disabled
since they are unspecified. However, with the current code the
behavior would be something else, and since the index number of SDCARD
is not matching it will be detected as CARD_DETECT.

So isn't it really ugly to depend on the number of IRQs when they are
supposed to be used as an index? I've been toying around with this
driver for a few years now, and when I have a hard time creating
correct platform data then it's _probably_ a sign that there must be
better ways to implement this.

I would propose just adding interrupts in struct resource [] as usual,
and then have thee separate flags in the platform data for each
interrupt type. If the number of IRQ bits set in the platform data
flags doesn't match the number of interrupt resources then return
error. If they match then simply go through each flag set in the
platform data flags and assign next available interrupt resource. And
if no flags are set then go for the combined interrupt handler for all
available interrupt resources.

That's what I would do at least. Any other ideas? Perhaps just keep an
array of interrupt numbers in the platform data as the sh-sci driver
does and use the fixed indexes there if non-zero?

Hi Magnus,

unfortunately during the course of the review of this series a number
of changes have crept in which I regard as being tangential to the
goal of the series - which is to introduce broken-out handlers (I have
already introduce support for broken-out IRQ sources).

One area where such changes have occurred is in the subtle altering of the
list of IRQ sources that it is valid to supply (again, support for
broken-out IRQ sources is not introduced by this series).

With regards to your comment and example above. I believe that your
understanding of my code is incorrect. The configuration you suggest will
be rejected because CARD_DETECT is not supplied, not because of an index
miss-match. It could be made acceptable within the current framework by
simply loosening up the logic a little (specifically to allow CARD_DETECT
to be omitted even if only one other IRQ source is supplied).  Incidently,
I think that would make sense but Guennadi specifically asked for that
combination to be regarded as invalid.

You suggestion to not rely on the array offset is an interesting one
and perhaps it would make things easier (to understand). But I don't
believe it would entirely eliminate the possibility of creating bogus
platform data. And in any case I believe it is tangential to the aim
of this series - to introduce broken-out handlers.

Moreover, I think that there needs to be some agreement between Guennadi
and yourself with regards to what IRQ combinations are valid before I (or
anyone else) can implement something that is acceptable to you both. And
I think that in the mean time the current implementation is reasonable
given that it works with all current use-cases (and several non-existent
ones too).

In short, can we have this discussion in the context of further enhancements?