[PATCH] perf/core: Avoid context switch overheads

[PATCH] perf/core: Avoid context switch overheads

[石祤]

From: "leilei.lin" <leilei.lin@alibaba-inc.com>

A performance issue caused by less strickly check in task
sched when these tasks were once attached by per-task perf_event.

A task will alloc task->perf_event_ctxp[ctxn] when it was called
by perf_event_open, and task->perf_event_ctxp[ctxn] would not
ever be freed to NULL.

__perf_event_task_sched_in()
	if (task->perf_event_ctxp[ctxn]) //  here is always true
		perf_event_context_sched_in() // operate pmu

50% at most performance overhead was observed under some extreme
test case. Therefor, add a more strick check as to ctx->nr_events,
when ctx->nr_events == 0, it's no need to continue.

Signed-off-by: leilei.lin <leilei.lin@alibaba-inc.com>
---
 kernel/events/core.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 426c2ff..f071013 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3179,6 +3179,9 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
 	if (cpuctx->task_ctx == ctx)
 		return;
 
+	if (!cpuctx->task_ctx && !ctx->nr_events)
+		return;
+
 	perf_ctx_lock(cpuctx, ctx);
 	perf_pmu_disable(ctx->pmu);
 	/*
-- 
2.8.4.31.g9ed660f

Re: [PATCH] perf/core: Avoid context switch overheads

[Peter Zijlstra]

On Fri, Aug 04, 2017 at 04:50:54PM +0800, 石祤 wrote:
> From: "leilei.lin" <leilei.lin@alibaba-inc.com>
> 
> A performance issue caused by less strickly check in task
> sched when these tasks were once attached by per-task perf_event.
> 
> A task will alloc task->perf_event_ctxp[ctxn] when it was called
> by perf_event_open, and task->perf_event_ctxp[ctxn] would not
> ever be freed to NULL.
> 
> __perf_event_task_sched_in()
> 	if (task->perf_event_ctxp[ctxn]) //  here is always true
> 		perf_event_context_sched_in() // operate pmu
> 
> 50% at most performance overhead was observed under some extreme
> test case. Therefor, add a more strick check as to ctx->nr_events,
> when ctx->nr_events == 0, it's no need to continue.
> 
> Signed-off-by: leilei.lin <leilei.lin@alibaba-inc.com>
> ---
>  kernel/events/core.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 426c2ff..f071013 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -3179,6 +3179,9 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
>  	if (cpuctx->task_ctx == ctx)
>  		return;
>  
> +	if (!cpuctx->task_ctx && !ctx->nr_events)
> +		return;
> +
>  	perf_ctx_lock(cpuctx, ctx);
>  	perf_pmu_disable(ctx->pmu);
>  	/*

I _think_ we must do this check after acquiring ctx->lock, because of
how perf_install_in_context() works. See commit:

  63cae12bce98 ("perf/core: Fix sys_perf_event_open() vs. hotplug")

That is a giant bag of tricky, but the gist of it is that if
perf_event_ctxp[] is !NULL (as is the case in your scenario) then we
must acquire ctx->lock.

So I think we want something like the below.


---
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 426c2ffba16d..bf40b7633cb0 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3180,6 +3180,13 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
 		return;
 
 	perf_ctx_lock(cpuctx, ctx);
+	/*
+	 * We must check ctx->nr_events while holding ctx->lock, such
+	 * that we serialize against perf_install_in_context().
+	 */
+	if (!ctx->nr_events)
+		goto unlock;
+
 	perf_pmu_disable(ctx->pmu);
 	/*
 	 * We want to keep the following priority order:
@@ -3193,6 +3200,7 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
 		cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 	perf_event_sched_in(cpuctx, ctx, task);
 	perf_pmu_enable(ctx->pmu);
+unlock:
 	perf_ctx_unlock(cpuctx, ctx);
 }
 

Re: [PATCH] perf/core: Avoid context switch overheads

[Peter Zijlstra]

On Mon, Aug 07, 2017 at 02:15:51PM +0200, Peter Zijlstra wrote:
> On Fri, Aug 04, 2017 at 04:50:54PM +0800, 石祤 wrote:
> > From: "leilei.lin" <leilei.lin@alibaba-inc.com>
> > 
> > A performance issue caused by less strickly check in task
> > sched when these tasks were once attached by per-task perf_event.
> > 
> > A task will alloc task->perf_event_ctxp[ctxn] when it was called
> > by perf_event_open, and task->perf_event_ctxp[ctxn] would not
> > ever be freed to NULL.
> > 
> > __perf_event_task_sched_in()
> > 	if (task->perf_event_ctxp[ctxn]) //  here is always true
> > 		perf_event_context_sched_in() // operate pmu
> > 
> > 50% at most performance overhead was observed under some extreme
> > test case. Therefor, add a more strick check as to ctx->nr_events,
> > when ctx->nr_events == 0, it's no need to continue.
> > 
> > Signed-off-by: leilei.lin <leilei.lin@alibaba-inc.com>
> > ---
> >  kernel/events/core.c | 3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 426c2ff..f071013 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -3179,6 +3179,9 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
> >  	if (cpuctx->task_ctx == ctx)
> >  		return;
> >  
> > +	if (!cpuctx->task_ctx && !ctx->nr_events)
> > +		return;
> > +
> >  	perf_ctx_lock(cpuctx, ctx);
> >  	perf_pmu_disable(ctx->pmu);
> >  	/*
> 
> I _think_ we must do this check after acquiring ctx->lock, because of
> how perf_install_in_context() works. See commit:
> 
>   63cae12bce98 ("perf/core: Fix sys_perf_event_open() vs. hotplug")
> 
> That is a giant bag of tricky, but the gist of it is that if
> perf_event_ctxp[] is !NULL (as is the case in your scenario) then we
> must acquire ctx->lock.

The alternative is of course, trying to free the task ctx. I've not gone
through what would be required to make that happen.