Re: [RFC v2] perf: Rewrite core context handling

[Peter Zijlstra]

On Tue, Aug 02, 2022 at 11:47:24AM +0530, Ravi Bangoria wrote:
> [...]
> 
> >  static void
> > -ctx_sched_in(struct perf_event_context *ctx,
> > -	     struct perf_cpu_context *cpuctx,
> > -	     enum event_type_t event_type,
> > +ctx_sched_in(struct perf_event_context *ctx, enum event_type_t event_type,
> >  	     struct task_struct *task)
> >  {
> > +	struct perf_cpu_context *cpuctx = this_cpu_ptr(&cpu_context);
> >  	int is_active = ctx->is_active;
> >  	u64 now;
> >  
> > @@ -3818,6 +3905,7 @@ ctx_sched_in(struct perf_event_context *ctx,
> >  		/* start ctx time */
> >  		now = perf_clock();
> >  		ctx->timestamp = now;
> > +		// XXX ctx->task =? task
> 
> Couldn't get this XXX, it's from your original patch. If you can recall, it
> would be helpful.

No memories at all; but looking at it; it seems to worry if ctx->task is
up-to-date; in this context the only thing that relies on the task is
the cgroup for which we update the timestamp in the next statement:

> >  		perf_cgroup_set_timestamp(task, ctx);

I suppose I should really write less cryptic notes; then again, I never
imagined this would take that many years to complete :/

> >  	}
> 
> Also, this hunk is under if (is_active ^ EVENT_TIME), which effectively is
> (is_active != EVENT_TIME). I'm assuming it should be (is_active & EVENT_TIME)?

So that code is identical to what it currently is upstream; but yes that
looks somewhat dodgy.

So the code itself (does as the comment says) starts time. This should
only be done if EVENT_TIME is not set. That is, I'm thinking it should
be something like:

	!(is_active & EVENT_TIME)

which happens to be the same as:

	is_active ^ EVENT_TIME

under the assumption is_active contains no other bits -- which I don't
think is a valid assumption.