Re: [PATCH 2/4] ARM: OMAP2/3: intc: Add DT support for TI interrupt controller

["Cousson, Benoit" <b-cousson@ti.com>]

On 12/9/2011 4:22 PM, Rob Herring wrote:
> On 12/09/2011 08:52 AM, Cousson, Benoit wrote:
>> On 12/9/2011 2:20 PM, Rob Herring wrote:
>>> On 12/08/2011 08:59 AM, Cousson, Benoit wrote:
>>>> On 12/7/2011 10:20 PM, Rob Herring wrote:
>>>>> On 12/07/2011 02:50 PM, Benoit Cousson wrote:
>>>>
>>>> [...]
>>>>
>>>>>> +#ifdef CONFIG_OF
>>>>>> +int __init intc_of_init(struct device_node *node, struct device_node
>>>>>> *parent)
>>>>>> +{
>>>>>> +    struct resource res;
>>>>>> +    u32 nr_irqs;
>>>>>> +
>>>>>> +    if (WARN_ON(!node))
>>>>>> +        return -ENODEV;
>>>>>> +
>>>>>> +    if (of_address_to_resource(node, 0,&res)) {
>>>>>> +        WARN(1, "unable to get intc registers\n");
>>>>>> +        return -EINVAL;
>>>>>> +    }
>>>>>> +
>>>>>> +    if (of_property_read_u32(node, "ti,intc-size",&nr_irqs)) {
>>>>>> +        WARN(1, "unable to get intc-size\n");
>>>>>> +        return -EINVAL;
>>>>>
>>>>> There is no default value that makes sense?
>>>>
>>>> So far we have 96 or 128, so I can put 96 as a default and just keep the
>>>> warning without returning an error.
>>>>
>>>>>> +    }
>>>>>> +
>>>>>> +    omap_init_irq(res.start, nr_irqs);
>>>>>> +    irq_domain_add_simple(node, 0);
>>>>>
>>>>> Have you read the NO_IRQ thread...
>>>>
>>>> Yeah, I tried, but that's a long email thread with some unclear
>>>> conclusion...
>>>> The point is that the few users of that API today are using 0 as a base
>>>> as well, so I thought it was still valid.
>>>>
>>>>> Is 0 ever a valid interrupt for a driver? If so, you must not use 0 for
>>>>> the base. I would pick 16 to skip over legacy ISA irqs.
>>>>
>>>> I do not have any 0 interrupts right now, my concern is that I still
>>>> have some legacy non-DT drivers that expect the previous hwirq = irq
>>>> mapping.
>>>
>>> I guess it depends how easily you can shift all the irq defines. You
>>> allow specifying the base so that you can set it to 0 for non-DT and -1
>>> (dynamic allocation) for DT.
>>
>> The issue, is that the IRQs are not defined anymore but hard coded in
>> the hwmod database. And the idea is that this is reflecting exactly the
>> hwirq from the spec, so I do not want to add any artificial offset for
>> the domain in the original data.
>>
>>>>> irqdomains should always be enabled regardless of CONFIG_OF. So either
>>>>> you can leave it as is if OF is always enabled for OMAP, or you should
>>>>> move domain setup into omap_init_irq.
>>>>
>>>> OK, but it looks like I cannot really modify the current INTC to DT
>>>> without having fully adapted the OMAP drivers to use
>>>> irq_of_parse_and_map. Or did I miss something?
>>>>
>>>
>>> No, the drivers should not need to be modified as long as they get the
>>> irq's from platform device resources. You just want to make sure the
>>> INTC has no knowledge of it's irq base so it can change.
>>
>> OK, the driver will not have to change but the IRQ value will not be the
>> same in the case of DT since it will use the irq_create_of_mapping.
>>
>> Currently, the driver IRQ resource is 7 for example for the twl, this is
>> the hwirq (= irq). If I use a domain with a base of 16, the resource
>> will still be 7 except if the resource is created from OF, then it will
>> be 23.
>>
>> The only way I have today to maintain the legacy drivers to work without
>> hacking the OMAP resource code is to keep the base_irq at 0.
>>
>
> For non-DT case, that is fine. It should not be hard to support 0 for
> legacy and !0 for DT.

Mmm, I'm probably missing something here.

My point is that even in the DT case I do have some devices that are 
initialized without DT for the moment and thus cannot get access to the 
interrupt-controller node and then cannot retrieve the domain information.

How can I ensure the proper hwirq -> irq translation then for such devices?
Only the one created by DT will have the correct irq number.

Adding some hacks to add a +16 is clearly possible, but that will be a 
dirty hacks in the OMAP core code that will have to be removed later. So 
I'd rather keep the domain with a 0 offset to avoid all these temporary 
hacks.

>> Since I do not have any hwirq at 0 so far, is it a big deal to keep the
>> base_irq at 0 for the moment?
>> It will be easier to transition to DT that way without breaking the
>> existing drivers.
>
> As long as it is trivial to change later. I'm afraid that if it is not
> changed, then we won't know if it is trivial.

As soon as every devices are initialized from DT, the irq_base will 
become transparent, since both the IRQ handler and the IRQ resource will 
use the irq_domain_to_irq to get the irq from the hwirq. At that time 
adding 16 will then be straightforward.
That's why I prefer that approach which is the one that generate the 
minimal temporary effort.

Regards,
Benoit