[Linux-Xtensa] [PATCH 7/8] xtensa: implement counting and sampling perf events

Max Filippov jcmvbkbc at gmail.com
Mon Jul 6 14:22:33 UTC 2015

On Mon, Jul 6, 2015 at 5:05 PM, Peter Zijlstra <peterz at infradead.org> wrote:
> On Mon, Jul 06, 2015 at 04:56:09PM +0300, Max Filippov wrote:
>> On Mon, Jul 6, 2015 at 4:47 PM, Peter Zijlstra <peterz at infradead.org> wrote:
>> > On Mon, Jul 06, 2015 at 04:32:48PM +0300, Max Filippov wrote:
>> >> +static int __init xtensa_pmu_init(void)
>> >> +{
>> >> +     int ret;
>> >> +     int irq = irq_create_mapping(NULL, XCHAL_PROFILING_INTERRUPT);
>> >
>> > Does this platform have interrupt priorities which you can partially
>> > mask in order to create NMI like behaviour?
>> Not sure what you mean by "NMI like".
> There's a number of archs where we implement NMIs by having
> local_irq_disable() only disable part of the interrupt priority range
> and making sure all 'normal' IRQs are mapped in that priority range.
> We then map our NMI handlers to a priority above the 'normal' range,
> such that these interrupts can indeed happen when interrupts are
> 'disabled.
> See for example:
> b4f4372f96e0 ("sparc64: Make %pil level 15 a pseudo-NMI.")
> 0c25e9e6cbe7 ("sparc64: Adjust __raw_local_irq_save() to cooperate in NMIs.")
> c011f80ba091 ("sparc64: Add some more commentary to __raw_local_irq_save()")

Ok, I see. I guess I can change IRQ disabling logic to not mask perf IRQ
in case it's configured as the only interrupt on its level and it's the highest
medium-level IRQ.

>> Interrupt priorities are fixed in the current xtensa architecture, and
>> we can in theory mask certain level and below, but practically we
>> always mask all low- and medium- level interrupts.
>> Also we currently can't have handlers for high priority interrupts written in C.
> Why not? Surely this can be cured with an assembly stub?

IIUC that was a deliberate architecture design choice and working around
it penalizes all interrupt handlers. But let me take another close look.

> The advantage of having NMIs is that profiling information for the
> kernel becomes much more useful. Without this local_irq_enable() will be
> a very 'hot' function.

I haven't noticed that in my testing.

-- Max

More information about the linux-xtensa mailing list