[PATCH] ARM: EXYNOS: Clear OF_POPULATED flag from PMU node in IRQ init callback

[Rob Herring]

On Wed, Aug 24, 2016 at 12:44 PM, Krzysztof Kozlowski <krzk at kernel.org> wrote:
> On Sun, Aug 21, 2016 at 03:27:45AM -0400, Javier Martinez Canillas wrote:
>> The Exynos PMU node is an interrupt, clock and PMU (Power Management Unit)
>> controller, and these functionalities are supported by different drivers
>> that matches the same compatible strings.
>>
>> Since commit 15cc2ed6dcf9 ("of/irq: Mark initialised interrupt controllers
>> as populated") the OF core flags interrupt controllers registered with the
>> IRQCHIP_DECLARE() macro as OF_POPULATED, so platform devices with the same
>> compatible string as the interrupt controller will not be registered.
>>
>> This prevents the PMU platform device to be registered so the Exynos PMU
>> driver is never probed. This breaks (among other things) Suspend-to-RAM.
>>
>> Fix this by clearing the OF_POPULATED flag in the PMU IRQ init callback,
>> to allow the Exynos PMU platform driver to be probed. The patch is based
>> on Philipp Zabel's "ARM: imx6: mark GPC node as not populated after irq
>> init to probe pm domain driver".
>>
>> Fixes: 15cc2ed6dcf9 ("of/irq: Mark initialised interrupt controllers as populated")
>> Signed-off-by: Javier Martinez Canillas <javier at osg.samsung.com>
>>
>> ---
>>
>> This was tested on an Exynos5800 Peach Pi Chromebook with v4.8-rc2 +
>> Philipp's patch: https://patchwork.kernel.org/patch/9271361/
>>
>>  arch/arm/mach-exynos/suspend.c | 6 ++++++
>>  1 file changed, 6 insertions(+)
>>
>> diff --git a/arch/arm/mach-exynos/suspend.c b/arch/arm/mach-exynos/suspend.c
>> index 3750575c73c5..06332f626565 100644
>> --- a/arch/arm/mach-exynos/suspend.c
>> +++ b/arch/arm/mach-exynos/suspend.c
>> @@ -255,6 +255,12 @@ static int __init exynos_pmu_irq_init(struct device_node *node,
>>               return -ENOMEM;
>>       }
>>
>> +     /*
>> +      * Clear the OF_POPULATED flag set in of_irq_init so that
>> +      * later the Exynos PMU platform device won't be skipped.
>> +      */
>> +     of_node_clear_flag(node, OF_POPULATED);
>> +
>
> Looks like a proper workaround (at least till of_irq_init sets this flag
> before calling irq_init_cb()) however I have some more doubts:
>
> 1. The commit 15cc2ed6dcf9 ("of/irq: Mark initialised interrupt
> controllers as populated") might be also applied to clock CLK_OF_DECLARE
> and the problem will appear again.

We now know what to look for...

> 2. We are reusing PMU compatible for irqcip, clock provider and a
> platform driver. This is one PMU block with many features thus many
> drivers were created. What happens if the Exynos platform driver
> (drivers/soc/samsung/exynos-pmu.c) binds before irqchip or clk provider?

It can't. OF_DECLARE calls happen before initcalls*.

*I did find that IOMMU's are during initcalls which in my opinion is
wrong. But I think there is some plan to fix them to support deferred
probe and that should go away.

> Probably we should not reuse the compatible but create a new one for
> each type of driver? How does it match DeviceTree?

No, compatible strings are for h/w blocks, not drivers. Now, if you
need multiple platform drivers, then you should have 1 parent
device/driver that's bound from DT, then it can create any child
devices for specific functions.

Rob