2012-05-22 11:21:55

by Anshuman Khandual

Subject: [PATCH] perf, x86: Optimal way of reading LBR entries inside Intel PMU interrupt handler

perf, x86: Optimal way of reading LBR entries inside Intel PMU interrupt handler

We read LBR entries even if no event has requested for it either explicitly
through branch sampling or implicitly through precise IP. This patch would
fix this potential unnecessary read by moving the intel_pmu_lbr_read function.

Signed-off-by: Anshuman Khandual <[email protected]>
---
arch/x86/kernel/cpu/perf_event_intel.c | 12 +++++++++---
1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
index 26b3e2f..c50da47 100644
--- a/arch/x86/kernel/cpu/perf_event_intel.c
+++ b/arch/x86/kernel/cpu/perf_event_intel.c
@@ -1026,6 +1026,7 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
int bit, loops;
u64 status;
int handled;
+ int lbr_read;

perf_sample_data_init(&data, 0);

@@ -1061,16 +1062,16 @@ again:

inc_irq_stat(apic_perf_irqs);

- intel_pmu_lbr_read();
-
/*
* PEBS overflow sets bit 62 in the global status register
*/
if (__test_and_clear_bit(62, (unsigned long *)&status)) {
handled++;
+ intel_pmu_lbr_read();
x86_pmu.drain_pebs(regs);
}

+ lbr_read = 0;
for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
struct perf_event *event = cpuc->events[bit];

@@ -1084,8 +1085,13 @@ again:

data.period = event->hw.last_period;

- if (has_branch_stack(event))
+ if (has_branch_stack(event)) {
+ if(lbr_read == 0) {
+ intel_pmu_lbr_read();
+ lbr_read = 1;
+ }
data.br_stack = &cpuc->lbr_stack;
+ }

if (perf_event_overflow(event, &data, regs))
x86_pmu_stop(event, 0);

2012-05-22 11:49:58

by Stephane Eranian

Subject: Re: [PATCH] perf, x86: Optimal way of reading LBR entries inside Intel PMU interrupt handler

On Tue, May 22, 2012 at 1:10 PM, Anshuman Khandual
<[email protected]> wrote:
> perf, x86: Optimal way of reading LBR entries inside Intel PMU interrupt handler
>
>        We read LBR entries even if no event has requested for it either explicitly
>        through branch sampling or implicitly through precise IP. This patch would
>        fix this potential unnecessary read by moving the intel_pmu_lbr_read function.
>
But that's not quite the case because intel_pmu_lbr_read() checks:

if (!cpuc->lbr_users)
return;

> Signed-off-by: Anshuman Khandual <[email protected]>
> ---
>  arch/x86/kernel/cpu/perf_event_intel.c |   12 +++++++++---
>  1 file changed, 9 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/kernel/cpu/perf_event_intel.c b/arch/x86/kernel/cpu/perf_event_intel.c
> index 26b3e2f..c50da47 100644
> --- a/arch/x86/kernel/cpu/perf_event_intel.c
> +++ b/arch/x86/kernel/cpu/perf_event_intel.c
> @@ -1026,6 +1026,7 @@ static int intel_pmu_handle_irq(struct pt_regs *regs)
>        int bit, loops;
>        u64 status;
>        int handled;
> +       int lbr_read;
>
>        perf_sample_data_init(&data, 0);
>
> @@ -1061,16 +1062,16 @@ again:
>
>        inc_irq_stat(apic_perf_irqs);
>
> -       intel_pmu_lbr_read();
> -
>        /*
>         * PEBS overflow sets bit 62 in the global status register
>         */
>        if (__test_and_clear_bit(62, (unsigned long *)&status)) {
>                handled++;
> +               intel_pmu_lbr_read();
>                x86_pmu.drain_pebs(regs);
>        }
>
> +       lbr_read = 0;
>        for_each_set_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
>                struct perf_event *event = cpuc->events[bit];
>
> @@ -1084,8 +1085,13 @@ again:
>
>                data.period = event->hw.last_period;
>
> -               if (has_branch_stack(event))
> +               if (has_branch_stack(event)) {
> +                       if(lbr_read == 0) {
> +                               intel_pmu_lbr_read();
> +                               lbr_read = 1;
> +                       }
>                        data.br_stack = &cpuc->lbr_stack;
> +               }
>
>                if (perf_event_overflow(event, &data, regs))
>                        x86_pmu_stop(event, 0);
>