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, çç wrote:
> > From: "leilei.lin" <leilei.lin@xxxxxxxxxxxxxxx>
> > 
> > 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@xxxxxxxxxxxxxxx>
> > ---
> >  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.