[PATCH 06/11] ACPI / gsi: Add gsi_mutex to synchronize acpi_register_gsi()/acpi_unregister_gsi()

[Marc Zyngier]

On 19/06/15 08:31, Hanjun Guo wrote:
> On 06/11/2015 09:16 PM, Hanjun Guo wrote:
>> On 06/10/2015 11:58 PM, Marc Zyngier wrote:
>>> On 18/05/15 13:59, Hanjun Guo wrote:
>>>> Add a mutex for acpi_register_gsi()/acpi_unregister_gsi() to avoid
>>>> concurrency issues.
>>>>
>>>> Signed-off-by: Hanjun Guo <hanjun.guo at linaro.org>
>>>> ---
>>>>   drivers/acpi/gsi.c | 17 +++++++++++++----
>>>>   1 file changed, 13 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/acpi/gsi.c b/drivers/acpi/gsi.c
>>>> index 55b5f31..ab0dcb4 100644
>>>> --- a/drivers/acpi/gsi.c
>>>> +++ b/drivers/acpi/gsi.c
>>>> @@ -16,6 +16,7 @@
>>>>   enum acpi_irq_model_id acpi_irq_model;
>>>>   /* ACPI core domian pointing to GICv2/3 core domain */
>>>>   struct irq_domain *acpi_irq_domain __read_mostly;
>>>> +static DEFINE_MUTEX(gsi_mutex);
>>>>
>>>>   static unsigned int acpi_gsi_get_irq_type(int trigger, int polarity)
>>>>   {
>>>> @@ -73,20 +74,24 @@ int acpi_register_gsi(struct device *dev, u32
>>>> gsi, int trigger,
>>>>       int irq;
>>>>       unsigned int irq_type = acpi_gsi_get_irq_type(trigger, polarity);
>>>>
>>>> +    mutex_lock(&gsi_mutex);
>>>>       irq = irq_find_mapping(acpi_irq_domain, gsi);
>>>>       if (irq > 0)
>>>> -        return irq;
>>>> +        goto out;
>>>>
>>>>       irq = irq_domain_alloc_irqs(acpi_irq_domain, 1, dev_to_node(dev),
>>>>                       &gsi);
>>>>       if (irq <= 0)
>>>> -        return -EINVAL;
>>>> +        goto out;
>>>>
>>>>       /* Set irq type if specified and different than the current one */
>>>>       if (irq_type != IRQ_TYPE_NONE &&
>>>>           irq_type != irq_get_trigger_type(irq))
>>>>           irq_set_irq_type(irq, irq_type);
>>>> -    return irq;
>>>> +
>>>> +out:
>>>> +    mutex_unlock(&gsi_mutex);
>>>> +    return irq > 0 ? irq : -EINVAL;
>>>>   }
>>>>   EXPORT_SYMBOL_GPL(acpi_register_gsi);
>>>>
>>>> @@ -96,8 +101,12 @@ EXPORT_SYMBOL_GPL(acpi_register_gsi);
>>>>    */
>>>>   void acpi_unregister_gsi(u32 gsi)
>>>>   {
>>>> -    int irq = irq_find_mapping(acpi_irq_domain, gsi);
>>>> +    int irq;
>>>> +
>>>> +    mutex_lock(&gsi_mutex);
>>>> +    irq = irq_find_mapping(acpi_irq_domain, gsi);
>>>>
>>>>       irq_dispose_mapping(irq);
>>>> +    mutex_unlock(&gsi_mutex);
>>>>   }
>>>>   EXPORT_SYMBOL_GPL(acpi_unregister_gsi);
>>>>
>>>
>>> Can you point out why we need this locking? The rest of the kernel seems
>>> to live without it pretty well. And if we really have an issue, I'd
>>
>> Hmm, I'm not so sure, I will look deep into that and come back later.
> 
> Sorry for the late reply, I finally recalled that why I introduced
> the lock here.
> 
> The lock introduced here is because of I assume ACPI based container
> (such as CPU socket) hotplug will be available on ARM64 in the future,
> then we may online/offline a irqchip at the same time (similar way of
> IOAPIC hotplug which introduced by Jiang Liu), it turns out that only
> one GICD for now so it's not the case for ARM64 in the near future,

Well, there can only be one GICD, or at least only one that is visible.
If the HW uses multiple GICDs (one per socket?), it must make ensure
that they are all synchronized transparently. That's a strong
requirement from the architecture. Otherwise, we enter the FrankeinGIC
realm, and ACPI has no business with it.

> I will drop this patch in next version.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...