Re: [RFC PATCH v1 17/40] metag: IRQ handling

[James Hogan <james.hogan@imgtec.com>]

On 20/11/12 16:15, Arnd Bergmann wrote:
> On Tuesday 20 November 2012, James Hogan wrote:
>> On 09/11/12 14:12, Arnd Bergmann wrote:
>>> On Wednesday 31 October 2012, James Hogan wrote:
>>>
>>>> diff --git a/arch/metag/include/asm/irq.h b/arch/metag/include/asm/irq.h
>>>> new file mode 100644
>>>> index 0000000..da24886
>>>> --- /dev/null
>>>> +++ b/arch/metag/include/asm/irq.h
>>>> +#ifndef HW_IRQS
>>>> +#define HW_IRQS  (HWSTATMETA_OFFSET_MAX + HWSTATEXT_OFFSET_MAX + \
>>>> +		HWSTATEXT2_OFFSET_MAX + HWSTATEXT4_OFFSET_MAX + \
>>>> +		HWSTATEXT6_OFFSET_MAX)
>>>> +#endif
>>>> +
>>>> +#define META_IRQS 32
>>>> +
>>>> +#define NR_IRQS (META_IRQS + HW_IRQS)
>>>
>>> I think you should use sparse IRQs right from the start and avoid
>>> hardcoding interrupt numbers from the start.
>>
>> I presume core irq numbers need hard coding (e.g. those required by core
>> code like cross-hardware-thread irqs and perf counters)? Or would the
>> recommended way be to have a function which is given the core hw irq
>> number and converts/maps a virtual irq for it using the irq domain?
> 
> Anything that is probed as a platform device can get the interrupt numbers
> using the device resources that are filled in by the device tree probe.
> 
> If you need something earlier, you can call irq_of_parse_and_map()
> manually.

So does that leave any way for a non-GPL driver to simply get an IRQ
number? Reading the code I can't find a way other than using legacy irq
domains (are they acceptable for new architecture ports?).
* I think the usual way is to specify an IRQ in a module parameter, but
that no longer makes any sense as it would need mapping.
irq_create_mapping could have worked (maybe taking advantage of the
irq_default_domain), but it's EXPORT_SYMBOL_GPL.
* Using a platform driver isn't possible because
platform_driver_register and friends are EXPORT_SYMBOL_GPL.
* Similarly irq_of_parse_and_map is EXPORT_SYMBOL_GPL (except for sparc
curiously enough).

So what would be the expected way to do that with linear irq domains?

Disclaimer: I don't maintain or work on such drivers, but like other
architectures in the kernel the ability needs to be supported one way or
another.

Thanks
James