Re: [PATCH 2/6] irqchip/ls-scfg-msi: Fix MSI affinity handling

[Marc Zyngier]

On 19/06/18 13:00, Thomas Gleixner wrote:
> On Tue, 19 Jun 2018, Marc Zyngier wrote:
> 
>> The ls-scfs-msi driver is not dealing with the effective affinity
>> as it should. Let's fix that, and make it clear that the effective
>> affinity is restricted to a single CPU.
>>
>> Reported-by: Alexandre Belloni <alexandre.belloni@xxxxxxxxxxx>
>> Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
>> ---
>>  drivers/irqchip/irq-ls-scfg-msi.c | 3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/irqchip/irq-ls-scfg-msi.c b/drivers/irqchip/irq-ls-scfg-msi.c
>> index 1ec3bfe56693..2811cc53e425 100644
>> --- a/drivers/irqchip/irq-ls-scfg-msi.c
>> +++ b/drivers/irqchip/irq-ls-scfg-msi.c
>> @@ -121,7 +121,8 @@ static int ls_scfg_msi_set_affinity(struct irq_data *irq_data,
>>  		return -EINVAL;
>>  	}
>>  
>> -	cpumask_copy(irq_data->common->affinity, mask);
>> +	cpumask_copy(irq_data->common->affinity, cpumask_of(cpu));
> 
> That's wrong. The driver should not fiddle with the affinity mask at all
> and just store the effective mask.

Ah, very good point. The ls_scfg_msi_compose_msg callback is also pretty
dodgy in that respect...

> 
>> +	irq_data_update_effective_affinity(irq_data, cpumask_of(cpu));
>>  
>>  	return IRQ_SET_MASK_OK;
> 
> If yuo look at the call site, then you'll notice that it will do:
> 
>         case IRQ_SET_MASK_OK:
>         case IRQ_SET_MASK_OK_DONE:
>                 cpumask_copy(desc->irq_common_data.affinity, mask);
> 
> So the cpumask_copy() can just go away.

Thanks for noticing that. I've fixed it up as follow:

diff --git a/drivers/irqchip/irq-ls-scfg-msi.c b/drivers/irqchip/irq-ls-scfg-msi.c
index 2811cc53e425..c671b3212010 100644
--- a/drivers/irqchip/irq-ls-scfg-msi.c
+++ b/drivers/irqchip/irq-ls-scfg-msi.c
@@ -93,8 +93,12 @@ static void ls_scfg_msi_compose_msg(struct irq_data *data, struct msi_msg *msg)
 	msg->address_lo = lower_32_bits(msi_data->msiir_addr);
 	msg->data = data->hwirq;
 
-	if (msi_affinity_flag)
-		msg->data |= cpumask_first(data->common->affinity);
+	if (msi_affinity_flag) {
+		const struct cpumask *mask;
+
+		mask = irq_data_get_effective_affinity_mask(data);
+		msg->data |= cpumask_first(mask);
+	}
 
 	iommu_dma_map_msi_msg(data->irq, msg);
 }
@@ -121,7 +125,6 @@ static int ls_scfg_msi_set_affinity(struct irq_data *irq_data,
 		return -EINVAL;
 	}
 
-	cpumask_copy(irq_data->common->affinity, cpumask_of(cpu));
 	irq_data_update_effective_affinity(irq_data, cpumask_of(cpu));
 
 	return IRQ_SET_MASK_OK;

Thanks,

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