Re: [PATCH] drivers/perf: arm_pmu: save/restore cpu cycle counter in cpu_pm_pmu_notify

From: Jia He
Date: Mon Nov 20 2017 - 08:50:29 EST

Hi Mark

Thanks for your review.

On 11/20/2017 8:32 PM, Mark Rutland Wrote:

On Thu, Nov 16, 2017 at 06:27:28AM +0000, Jia He wrote:
Sometimes userspace need a high resolution cycle counter by reading

In commit da4e4f18afe0 ("drivers/perf: arm_pmu: implement CPU_PM
notifier"), it resets all the counters even when the pmcr_el0.E and
pmcntenset_el0.C are both 1 . That is incorrect.
I appreciate that you may wish to make use of the cycle counter from
userspace, but this is the intended behaviour kernel-side. Direct
userspace counter acceess is not supported.
Sorry for my previous description. Not only user space, but also any pmccntr_el0 reading
in kernel space is 0 except perf infrastructure.

In power states where context is lost, any perf events are
saved/restored by cpu_pm_pmu_setup(). So we certainly shouldn't be
modifying the counter registers in any other PM code.

We *could* expose counters to userspace on homogeneous systems, so long
as users stuck to the usual perf data page interface. However, this
comes with a number of subtle problems, and no-one's done the work to
enable this.

Even then, perf may modify counters at any point in time, and
monotonicity (and/or presence) of counters is not guaranteed.

Yes, the cycle counter register pmccntr_el0 will be impacted by perf usage.But do you think
pmccntr_el0 is intented for perf only? Currently, many user space/kernel space programs will use
pmccntr_el0 as a timestamp( just likethe rdtsc() on X86). After commit da4e4f18afe0, all the cycle counter readingis 0 because of CPU PM enter/exit state changes in armv8pmu_reset.