Re: [PATCH 5/5] ARM: gic: add OF based initialization

["Cousson, Benoit" <b-cousson-l0cyMroinI0@public.gmane.org>]

On 9/19/2011 2:07 PM, Dave Martin wrote:
> On Sun, Sep 18, 2011 at 7:21 AM, Grant Likely<grant.likely-s3s/WqlpOiPyB63q8FvJNQ@public.gmane.org>  wrote:
>> On Fri, Sep 16, 2011 at 05:09:39PM +0100, Dave Martin wrote:
>>> For now, we express the mapping by putting an interrupt-map in the
>>> core-tile DT, but this feels inelegant as well as wasteful -- expressing
>>> "+ 32" using a table which is about 1K in size and duplicates that
>>> information 43 times.
>>>
>>> Using a dedicated irq domain or a fake interrupt controller node to
>>> encapsulate the motherboard interrupts feels like a cleaner approach,
>>> but for now I'm not clear on the best way to do it.
>>
>> An irq nexus node would indeed be something to investigate for your
>> particular case.  Look for examples of interrupt-map.  It is most
>> often used for handling IRQ swizzling on PCI busses.
>
> That's what I currently have -- this is logically correct, and it
> works; it just feels like a sledgehammer for cracking this particular
> nut, because we don't really have 43 independent interrupt mappings
> and types.  We have one offset and one type which is applied to all
> the interrupts, and this situation is expected to be the same for all
> variations of this board, except that offset may change.  I suspect
> this situation applies to many platforms -- the number of interrupts
> may in general be much larger than the number of independent mappings.
>
> Another way of looking at the problem is that it's difficult to come
> up with a one-size-fits-all description of interrupt mappings which is
> also efficient.  Requiring a single set of mapping semantics requires
> the mappings to be both rather overspecified for most cases, and
> flattened into a large, structureless table which may become pretty
> large and unwieldy.  If there were 100+ interrupts instead of 43, we'd
> really have to be generating the device tree using a script in order
> for it to be maintainable (which is not necessarily a problem, but can
> be a warning sign)

Yep, this is exactly the issue I was facing when I tried to map the 128 
interrupts lines of an OMAP4 to the GIC. Adding 128 entries to an 
interrupt map just to handle a simple linear translation with a constant 
value of 32 is clearly overkill.

> An alternative approach is to introduce a special interrupt controller
> node which serves as the interrupt-parent for the child domain and
> maps the interrupts with flexible semantics defined by the node's
> bindings; or different kinds of interrupt-map/interrupt-map-mask
> properties could be defined.  Bindings could be added as needed to
> support different cases -- though really we should only expect to have
> a small number at most.  When the interrupt mapping is significantly
> complex, using interrupt-map will probably be the best approach
> anyway.

Maybe we can just extend or add a new type of interrupt nexus to handle 
simple translation. The actual one was done for complex PCI mapping but 
with a small number of lines to handle. In our case, the mapping is 
simple, but the number of lines is huge.

Regards,
Benoit