Re: [PATCH] perf: arm_cspmu: Don't touch interrupt registers if no interrupt was assigned

[Ilkka Koskinen <ilkka@os.amperecomputing.com>]

On Mon, 8 Apr 2024, Robin Murphy wrote:
> On 2024-04-05 11:33 pm, Ilkka Koskinen wrote:
>> 
>> On Fri, 5 Apr 2024, Robin Murphy wrote:
>>> On 2024-03-07 7:31 pm, Ilkka Koskinen wrote:
>>>> The driver enabled and disabled interrupts even if no interrupt was
>>>> assigned to the device.
>>> 
>>> Why's that a concern - if the interrupt isn't routed anywhere, surely it 
>>> makes no difference what happens at the source end?
>> 
>> The issue is that we have two PMUs attached to the same interrupt line.
>> Unfortunately, I just don't seem to find time to add support for shared 
>> interrupts to the cspmu driver. Meanwhile, I assigned the interrupt to one 
>> of the PMUs while the other one has zero in the APMT table.
>
> I suspected something like that ;)
>
>> Without the patch, I can trigger "ghost interrupt" in the latter PMU.
>
> An occasional spurious interrupt should be no big deal. If it ends up as a 
> screaming spurious interrupt because we never handle the overflow condition 
> on the "other" PMU, then what matters most is that we never handle the 
> overflow, thus the "other" PMU is still useless since you can't assume the 
> user is going to read it frequently enough to avoid losing information and 
> getting nonsense counts back. So this hack really isn't a viable solution for 
> anything.

IIRC, what happens is that kernel will disable the interrupt eventually 
due to unhandled spurious interrupts making the "working" PMU also 
useless.

Cheers, Ilkka

>
> Thanks,
> Robin.
>