[PATCH] ARM: perf: add PMU hotplug notifier

[Ashwin Chaugule]

Hi Will,

On Thu, Feb 23, 2012 at 10:41 AM, Will Deacon <will.deacon at arm.com> wrote:
> From: Lorenzo Pieralisi <Lorenzo.Pieralisi at arm.com>
>
> When a CPU is taken out of reset, either cold booted or hotplugged in,
> some of its PMU registers can contain UNKNOWN values.
>
> This patch adds a hotplug notifier to ARM core perf code so that upon
> CPU restart the PMU unit is reset and becomes ready to use again.
>
> Signed-off-by: Lorenzo Pieralisi <lorenzo.pieralisi at arm.com>
> Signed-off-by: Will Deacon <will.deacon at arm.com>

I was playing around with this stuff recently, but used the CPU PM code from

http://lists.infradead.org/pipermail/linux-arm-kernel/2011-June/052829.html

> ---
>  arch/arm/kernel/perf_event.c |   23 +++++++++++++++++++++++
>  1 files changed, 23 insertions(+), 0 deletions(-)
>
> diff --git a/arch/arm/kernel/perf_event.c b/arch/arm/kernel/perf_event.c
> index ab59c3b..fa30ff9 100644
> --- a/arch/arm/kernel/perf_event.c
> +++ b/arch/arm/kernel/perf_event.c
> @@ -680,6 +680,28 @@ static void __init cpu_pmu_init(struct arm_pmu *armpmu)
>  }
>
>  /*
> + * PMU hardware loses all context when a CPU goes offline.
> + * When a CPU is hotplugged back in, since some hardware registers are
> + * UNKNOWN at reset, the PMU must be explicitly reset to avoid reading
> + * junk values out of them.
> + */
> +static int __cpuinit pmu_cpu_notify(struct notifier_block *b,
> +                                       unsigned long action, void *hcpu)
> +{
> +       if (action != CPU_UP_PREPARE)
> +               return NOTIFY_DONE;
> +
> +       if (cpu_pmu && cpu_pmu->reset)
> +               cpu_pmu->reset(NULL);
> +
> +       return NOTIFY_OK;
> +}
> +

I realized I needed to save the counters before going to sleep. In my
case, the notifiers were being called from the idle thread, so I
didn't know which "sleep state" I was going to go into. In the deepest
idle state, the counters are completely reset. When we come out of
sleep, unless there had been an explicit "read" earlier, we will have
lost the cached value(event->count) of the counters. This manifests as
counts reduced by an order of magnitude.

I'll send a link to my code once the git server is up. It's a bit flaky today.


Cheers,
Ashwin