Re: [PATCH] drivers: perf: arm_pmuv3: Update 'pmc_width' based on actual HW event width

[Robin Murphy]

On 09/10/2023 5:37 am, Anshuman Khandual wrote:
> This updates 'perf_event_mmap_page->pmc_width' based on actual HW event's
> width that are currently missing i.e ARMPMU_EVT_63BIT and ARMPMU_EVT_47BIT.
>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Cc: Mark Rutland <mark.rutland@xxxxxxx>
> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
> Cc: linux-kernel@xxxxxxxxxxxxxxx
> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
> ---
> This applies on v6.6-rc5.
>
>   drivers/perf/arm_pmuv3.c | 4 ++++
>   1 file changed, 4 insertions(+)
>
> diff --git a/drivers/perf/arm_pmuv3.c b/drivers/perf/arm_pmuv3.c
> index fe4db1831662..94723d00548e 100644
> --- a/drivers/perf/arm_pmuv3.c
> +++ b/drivers/perf/arm_pmuv3.c
> @@ -1375,6 +1375,10 @@ void arch_perf_update_userpage(struct perf_event *event,
>   	if (userpg->cap_user_rdpmc) {
>   		if (event->hw.flags & ARMPMU_EVT_64BIT)
>   			userpg->pmc_width = 64;
> +		else if (event->hw.flags & ARMPMU_EVT_63BIT)
> +			userpg->pmc_width = 63;
> +		else if (event->hw.flags & ARMPMU_EVT_47BIT)
> +			userpg->pmc_width = 47;

I think this will give the wrong behaviour in at least some cases - if 
the user has not requested a "long" event which would lead to 
ARMPMU_EVT_64BIT being set, we will bias the values to only count 32 
effective bits regardless of how wide the physical counters are. I 
believe this is what "the valid width of the counter" in the 
documentation is trying to refer to.

Thanks,
Robin.

>   		else
>   			userpg->pmc_width = 32;
>   	}