Re: [PATCH 1/3] perf: arm_spe: Add barrier before enabling profiling buffer

[Will Deacon <will@kernel.org>]

On Tue, Jul 01, 2025 at 04:31:57PM +0100, James Clark wrote:
> DEN0154 states that PMBPTR_EL1 must not be modified while the profiling
> buffer is enabled. Ensure that enabling the buffer comes after setting
> PMBPTR_EL1 by inserting an isb().
> 
> This only applies to guests for now, but in future versions of the
> architecture the PE will be allowed to behave in the same way.
> 
> Fixes: d5d9696b0380 ("drivers/perf: Add support for ARMv8.2 Statistical Profiling Extension")
> Signed-off-by: James Clark <james.clark@linaro.org>
> ---
>  drivers/perf/arm_spe_pmu.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> index 3efed8839a4e..6235ca7ecd48 100644
> --- a/drivers/perf/arm_spe_pmu.c
> +++ b/drivers/perf/arm_spe_pmu.c
> @@ -537,6 +537,7 @@ static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
>  	limit += (u64)buf->base;
>  	base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
>  	write_sysreg_s(base, SYS_PMBPTR_EL1);
> +	isb();


Hmm.

arm_spe_perf_aux_output_begin() is only called in two places:

1. From arm_spe_pmu_start()
2. From arm_spe_pmu_irq_handler()

For (1), we know that profiling is disabled by PMSCR_EL1.ExSPE.
For (2), we know that profiling is disabled by PMBSR_EL1.S.

In both cases, we already have an isb() before enabling profiling again
so I don't understand what this additional isb() is achieving.

Will