Re: [PATCH v3 4/4] perf/core,x86: synchronize PMU task contexts on optimized context switches

[Peter Zijlstra]

On Mon, Oct 21, 2019 at 09:59:42AM +0200, Ingo Molnar wrote:
> 
> * Alexey Budankov <alexey.budankov@xxxxxxxxxxxxxxx> wrote:
> 
> > +			/*
> > +			 * PMU specific parts of task perf context may require
> > +			 * additional synchronization, at least for proper Intel
> > +			 * LBR callstack data profiling;
> > +			 */
> > +			pmu->sync_task_ctx(ctx->task_ctx_data,
> > +					   next_ctx->task_ctx_data);
> 
> Firstly, I'm pretty sure you never run this on a CPU where 
> pmu->sync_task_ctx is NULL, right? ;-)
> 
> Secondly, even on Intel CPUs in many cases we'll just call into a ~2 deep 
> function pointer based call hierarchy, just to find that nothing needs to 

See prototype here for getting rid of at least one layer of indirect
calls:

  https://lkml.kernel.org/r/20191007083831.26880701.6@xxxxxxxxxxxxx

> be done, because there's no LBR call stack maintained:
> 
> +       if (!one || !another)
> +               return;
> 
> So while it's technically a layering violation, it might make sense to 
> elevate this check to the generic layer and say that synchronization 
> calls by the core layer will always provide two valid pointers?

Alternatively we can write the thing like:

	if (pmu->swap_task_ctx)
		pmu->swap_task_ctx(ctx, next_ctx)
	else
		swap(ctx->task_ctx_data, next_ctx->task_ctx_data);