[PATCH 1/1] perf: hisi: Make irq shared

[PATCH 1/1] perf: hisi: Make irq shared

[Chen Jun]

On some platforms, there are some error:
genirq: Flags mismatch irq 23. 00010804 (xxx) vs. 00010804 (xxx)

The reason is that there are more than one pmu nodes using the same
irq number.

Add IROF_SHARED when devm_request_irq.

Signed-off-by: Chen Jun <chenjun102@huawei.com>
---
 drivers/perf/hisilicon/hisi_uncore_pmu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/perf/hisilicon/hisi_uncore_pmu.c b/drivers/perf/hisilicon/hisi_uncore_pmu.c
index 358e4e284a62..bcbd3b467f34 100644
--- a/drivers/perf/hisilicon/hisi_uncore_pmu.c
+++ b/drivers/perf/hisilicon/hisi_uncore_pmu.c
@@ -168,7 +168,7 @@ int hisi_uncore_pmu_init_irq(struct hisi_pmu *hisi_pmu,
 		return irq;
 
 	ret = devm_request_irq(&pdev->dev, irq, hisi_uncore_pmu_isr,
-			       IRQF_NOBALANCING | IRQF_NO_THREAD,
+			       IRQF_NOBALANCING | IRQF_NO_THREAD | IRQF_SHARED,
 			       dev_name(&pdev->dev), hisi_pmu);
 	if (ret < 0) {
 		dev_err(&pdev->dev,
-- 
2.17.1

Re: [PATCH 1/1] perf: hisi: Make irq shared

[Robin Murphy]

On 2022-05-16 14:05, Chen Jun wrote:
> On some platforms, there are some error:
> genirq: Flags mismatch irq 23. 00010804 (xxx) vs. 00010804 (xxx)
> 
> The reason is that there are more than one pmu nodes using the same
> irq number.
> 
> Add IROF_SHARED when devm_request_irq.

You really need more than this to share an interrupt between multiple 
PMUs correctly. Perf relies on IRQs being handled on event->cpu to 
provide mutual exclusion - if another PMU comes along and moves the 
interrupt affinity to a different CPU, the interrupt handler may update 
the event state from that CPU simultaneously with a regular API call on 
event->cpu also updating it, with potential corruption as a result.

For general solutions, one option is to add a layer of explicit 
interrupt management like arm_dmc620_pmu (FWIW the original idea there 
was to factor that out into common helpers anyway). It's possible that 
IRQ affinity notifiers might work as an alternative approach too, but I 
never got round to looking into that in detail.

Thanks,
Robin.

> Signed-off-by: Chen Jun <chenjun102@huawei.com>
> ---
>   drivers/perf/hisilicon/hisi_uncore_pmu.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/perf/hisilicon/hisi_uncore_pmu.c b/drivers/perf/hisilicon/hisi_uncore_pmu.c
> index 358e4e284a62..bcbd3b467f34 100644
> --- a/drivers/perf/hisilicon/hisi_uncore_pmu.c
> +++ b/drivers/perf/hisilicon/hisi_uncore_pmu.c
> @@ -168,7 +168,7 @@ int hisi_uncore_pmu_init_irq(struct hisi_pmu *hisi_pmu,
>   		return irq;
>   
>   	ret = devm_request_irq(&pdev->dev, irq, hisi_uncore_pmu_isr,
> -			       IRQF_NOBALANCING | IRQF_NO_THREAD,
> +			       IRQF_NOBALANCING | IRQF_NO_THREAD | IRQF_SHARED,
>   			       dev_name(&pdev->dev), hisi_pmu);
>   	if (ret < 0) {
>   		dev_err(&pdev->dev,