Re: [PATCH v2 3/7] perf, x86: use the PEBS auto reload mechanism when possible

[Peter Zijlstra]

On Tue, Jul 15, 2014 at 04:58:55PM +0800, Yan, Zheng wrote:
> When a fixed period is specified, this patch make perf use the PEBS
> auto reload mechanism. This makes normal profiling faster, because
> it avoids one costly MSR write in the PMI handler.
> 
> Signef-off-by: Yan, Zheng <zheng.z.yan@xxxxxxxxx>
> ---
>  arch/x86/kernel/cpu/perf_event.c          | 15 +++++++++------
>  arch/x86/kernel/cpu/perf_event_intel_ds.c |  7 +++++++
>  include/linux/perf_event.h                |  1 +
>  3 files changed, 17 insertions(+), 6 deletions(-)
> 
> diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
> index 8868e9b..ae723c8 100644
> --- a/arch/x86/kernel/cpu/perf_event.c
> +++ b/arch/x86/kernel/cpu/perf_event.c
> @@ -979,13 +979,16 @@ int x86_perf_event_set_period(struct perf_event *event)
>  
>  	per_cpu(pmc_prev_left[idx], smp_processor_id()) = left;
>  
> -	/*
> -	 * The hw event starts counting from this event offset,
> -	 * mark it to be able to extra future deltas:
> -	 */
> -	local64_set(&hwc->prev_count, (u64)-left);
> +	if (!hwc->autoreload ||
> +	    local64_read(&hwc->prev_count) != (u64)-left) {

Do you really need that line break? I suspect it comes it at or below 80
if you concat.

> +		/*
> +		 * The hw event starts counting from this event offset,
> +		 * mark it to be able to extra future deltas:
> +		 */
> +		local64_set(&hwc->prev_count, (u64)-left);
>  
> -	wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
> +		wrmsrl(hwc->event_base, (u64)(-left) & x86_pmu.cntval_mask);
> +	}
>  
>  	/*
>  	 * Due to erratum on certan cpu we need
> diff --git a/arch/x86/kernel/cpu/perf_event_intel_ds.c b/arch/x86/kernel/cpu/perf_event_intel_ds.c
> index 980970c..1db4ce5 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel_ds.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel_ds.c
> @@ -714,6 +714,7 @@ void intel_pmu_pebs_enable(struct perf_event *event)
>  	struct hw_perf_event *hwc = &event->hw;
>  
>  	hwc->config &= ~ARCH_PERFMON_EVENTSEL_INT;
> +	hwc->autoreload = !event->attr.freq;
>  
>  	cpuc->pebs_enabled |= 1ULL << hwc->idx;
>  
> @@ -721,6 +722,11 @@ void intel_pmu_pebs_enable(struct perf_event *event)
>  		cpuc->pebs_enabled |= 1ULL << (hwc->idx + 32);
>  	else if (event->hw.flags & PERF_X86_EVENT_PEBS_ST)
>  		cpuc->pebs_enabled |= 1ULL << 63;
> +
> +	/* Use auto-reload if possible to save a MSR write in the PMI */
> +	if (hwc->autoreload)
> +		ds->pebs_event_reset[hwc->idx] =
> +			(u64)-hwc->sample_period & x86_pmu.cntval_mask;

The rule is to add { } for any multi-line block, not the strict single
stmt as per the C language.

>  }
>  
>  void intel_pmu_pebs_disable(struct perf_event *event)
> @@ -739,6 +745,7 @@ void intel_pmu_pebs_disable(struct perf_event *event)
>  		wrmsrl(MSR_IA32_PEBS_ENABLE, cpuc->pebs_enabled);
>  
>  	hwc->config |= ARCH_PERFMON_EVENTSEL_INT;
> +	hwc->autoreload = false;
>  }
>  
>  void intel_pmu_pebs_enable_all(void)
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 5d665e8..37a2b70 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -149,6 +149,7 @@ struct hw_perf_event {
>  
>  	u64				freq_time_stamp;
>  	u64				freq_count_stamp;
> +	bool				autoreload;

No, that's not the right place. If you'd looked at the structure you'd
have seen that this is the place for generic members, ones that apply to
all events, this is PEBS only, PEBS is limited to hardware events,
there's a section for that.

When you've found that, you'll note we have a flags field in there,
there's spare bits in there, use one.

Attachment: pgpgNv_jqFTJc.pgp
Description: PGP signature