[RFC PATCH 0/6] DRIVERS: IRQCHIP: Add support for crossbar IP

[Santosh Shilimkar]

On Tuesday 01 October 2013 10:53 AM, Rob Herring wrote:
> On 10/01/2013 08:57 AM, Santosh Shilimkar wrote:
>> On Tuesday 01 October 2013 09:48 AM, Rob Herring wrote:
>>> On 10/01/2013 06:13 AM, Sricharan R wrote:
>>>> Hi,
>>>>
>>>> On Monday 30 September 2013 08:39 PM, Rob Herring wrote:
>>>>> On 09/30/2013 08:59 AM, Sricharan R wrote:
>>>>>> Some socs have a large number of interrupts requests to service
>>>>>> the needs of its many peripherals and subsystems. All of the interrupt
>>>>>> requests lines from the subsystems are not needed at the same
>>>>>> time, so they have to be muxed to the controllers appropriately.
>>>>>> In such places a interrupt controllers are preceded by an
>>>>>> IRQ CROSSBAR that provides flexibility in muxing the device interrupt
>>>>>> requests to the controller inputs.
>>>>>>
>>>>>> This series models the peripheral interrupts that can be routed through
>>>>>> the crossbar to the GIC as 'routable-irqs'. The routable irqs are added
>>>>>> in a separate linear domain inside the GIC. The registered routable domain's
>>>>>> callback are invoked as a part of the GIC's callback, which in turn should
>>>>>> allocate a free irq line and configure the IP accordingly. So every peripheral
>>>>>> in the dts files mentions the fixed crossbar number as its interrupt. A free
>>>>>> gic line for that gets allocated and configured when the peripheral's interrupt
>>>>>> is mapped.
>>>>>>
>>>>>> The minimal crossbar driver to track and allocate free GIC lines and configure the
>>>>>> crossbar is added here, along with the DT bindings.
>>>>> Seems like interrupt-map property is what you need here.
>>>>>
>>>>> http://devicetree.org/Device_Tree_Usage#Advanced_Interrupt_Mapping
>>>>>
>>>>> Versatile Express also has an example.
>>>>    OK, but the idea was not to tie up the crossbar<->interrupt numbers at the
>>>>    DTS level, but to assign it dynamically during runtime. This was one of the
>>>>   comments that came up with first crossbar support patches, which was assigning a
>>>>   interrupt line to crossbar number in the DTS and setting it up in crossbar probe.
>>>
>>> Is there an actual usecase on a single h/w design that you run out of
>>> interrupts and it is a user decision which interrupts to use?
>>>
>> Yes. There are 240 peripheral interrupts connected out of which 160 can
>> be used in this specific case. 
> 
> Yes, I understand the SOC connections. That does not answer my question.
> The 240 interrupts are likely to be limited to fewer by board design,
> pinmuxing, etc.
> 
yes limited by different board designs ...

> How do you handle the 161st interrupt request? Will never happen? Just
> rely on the random driver probe ordering?
>
Well the board dts are expected to provide the peripheral support info to optimise it.
If a board actually has more than 160 peripherals available then in that
case the 161 interrupt will not be mapped. 
 
>>> You could fill in the interrupt-map at run-time. It would have to be
>>> early (bootloader or early kernel init) and can't be at request_irq time.
>>>
>> Well all options are tried before coming up to the $subject solution.
>> It was suggested by Thomas in the last review.
>>  
>>>>     https://lkml.org/lkml/2013/7/18/416
>>>>
>>>>    Since this approach of assigning in DTS was opposed, we moved to IRQCHIP and
>>>>    that did not go as well. Finally was asked to handle this as a part of GIC driver with
>>>>    a separate domain.
>>>>    
>>>>   http://www.spinics.net/lists/linux-omap/msg97085.html
>>>
>>> This has nothing to do with the GIC, so it does not belong there.
>>>
>> Well the router makes connections from peripheral to GIC. Thomas can
>> better explain it but I think since its doing irq routing for GIC on
>> a given hardware, I don't see any issue having some generic map/unmap
>> function in GIC. The actual implementation is still outside of GIC.
> 
> I read Thomas' reply as don't put this crap in his code.
>
That was for the IRQCHIP based approach and as part of that review
Thomas suggested why not irqdomain and suggested a prototype code
as well.
 
> You can call it generic, but it is not. It is specific to the GIC and
> looks like an abuse of irqdomains to me. Look where the function
> declaration for register_routable_domain_ops is.
> 
I am not sure why you call it abuse of irqdomain since the map/unmap
are exactly the interfaces where the logical to physical irq
connections are made. Look at existing GIC code as well. I still
let Thomas give his expert comment whether it is abusive because it
it was, am sure he wouldn't have suggested that.

Now if your concern is the register_routable_domain_ops() then
we are open to hear if there is any better way to do that. Thats
why the series is still RFC.

Regards,
Santosh