[PATCH] arm: use cpu_online_mask when using forced irq_set_affinity

[Sudeep Holla]

On 23/05/14 13:10, Russell King - ARM Linux wrote:
> On Fri, May 09, 2014 at 05:40:40PM +0100, Sudeep Holla wrote:
>> From: Sudeep Holla <sudeep.holla at arm.com>
>>
>> Commit 01f8fa4f01d8("genirq: Allow forcing cpu affinity of interrupts")
>> enabled the forced irq_set_affinity which previously refused to route an
>> interrupt to an offline cpu.
>>
>> Commit ffde1de64012("irqchip: Gic: Support forced affinity setting")
>> implements this force logic and disables the cpu online check for GIC
>> interrupt controller.
>>
>> When __cpu_disable calls migrate_irqs, it disables the current cpu in
>> cpu_online_mask and uses forced irq_set_affinity to migrate the IRQs
>> away from the cpu but passes affinity mask with the cpu being offlined
>> also included in it.
>>
>> When calling irq_set_affinity with force == true in a cpu hotplug path,
>> the caller must ensure that the cpu being offlined is not present in the
>> affinity mask or it may be selected as the target CPU, leading to the
>> interrupt not being migrated.
>>
>> This patch uses cpu_online_mask when using forced irq_set_affinity so
>> that the IRQs are properly migrated away.
>>
>> Tested on TC2 hotpluging CPU0 in and out. Without this patch the system
>> locks up as the IRQs are not migrated away from CPU0.
>
> You don't explain /how/ this happens, and I'm not convinced that you've
> properly diagnosed this bug.
>

Sorry for not being elaborate enough.
- On boot by default all the irqs have cpu_online_mask as affinity
- Now if CPU0 is being hotplugged out, CPU0 is removed from cpu_online_mask
   and migrate_irqs is called
- In migrate_one_irq, when affinity is read from the irq_desc, it still contains
   CPU0 which is expected.
- irq_set_affinity is called with affinity with CPU0 set and force = true,
   which chooses CPU0 resulting in not migrating the IRQ.

>> @@ -155,11 +155,15 @@ static bool migrate_one_irq(struct irq_desc *desc)
>>   	if (irqd_is_per_cpu(d) || !cpumask_test_cpu(smp_processor_id(), affinity))
>>   		return false;
>>
>> -	if (cpumask_any_and(affinity, cpu_online_mask) >= nr_cpu_ids) {
>> -		affinity = cpu_online_mask;
>> +	if (cpumask_any_and(affinity, cpu_online_mask) >= nr_cpu_ids)
>>   		ret = true;
>> -	}
>
> The idea here with the original code is:
>
> - if the current CPU (which is the one being offlined) is not in the
>    affinity mask, do nothing.
> - if "affinity & cpu_online_mask" indicates that there's no CPUs in the
>    new set (cpu_online_mask must have been updated to indicate that the
>    current CPU is offline) then re-set the affinity mask and report that
>    we forced a change.
> - otherwise, re-set the existing affinity (which will force the IRQ
>    controller to re-evaluate it's routing given the affinity and online
>    CPUs.)
>

I completely understand the above idea, except that the new feature added
to allow forced affinity setting(as mentioned in the commit log by 2 commits),
changes the behaviour of last step.

IRQ controller now re-evaluates it's routing based on the given affinity alone
and doesn't consider online CPUs when force = true is set. This will result in
the CPU being offlined chosen as the target if it happens to be the first in the
affinity mask.

> This code is correct.  In fact, changing it as you have, you /always/
> reset the affinity mask whether or not the CPU being offlined is the
> last CPU in the affinity set.
>
> If you are finding that CPU0 is left with interrupts afterwards, the
> bug lies elsewhere - probably in the IRQ controller code.
>

Since the IRQ controller code is changed to provide that feature, either
- we have to choose not to use forced option, or
- we need to make sure we pass valid affinity mask with force = true option

I chose latter in this patch. Let me know your opinion.

Regards,
Sudeep