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

From: Peter Zijlstra
Date: Mon Aug 22 2022 - 12:44:50 EST


On Tue, Aug 02, 2022 at 11:40:34AM +0530, Ravi Bangoria wrote:
> On 13-Jun-22 8:25 PM, Peter Zijlstra wrote:
> > On Mon, Jun 13, 2022 at 04:35:11PM +0200, Peter Zijlstra wrote:
> >> @@ -12125,6 +12232,8 @@ SYSCALL_DEFINE5(perf_event_open,
> >> goto err_task;
> >> }
> >>
> >> + // XXX premature; what if this is allowed, but we get moved to a PMU
> >> + // that doesn't have this.
> >> if (is_sampling_event(event)) {
> >> if (event->pmu->capabilities & PERF_PMU_CAP_NO_INTERRUPT) {
> >> err = -EOPNOTSUPP;
> >
> > No; this really should be against the event's native PMU. If the event
> > can't natively sample, it can't sample when placed in another group
> > either.
>
> Right. But IIUC, the question was, would there be any issue if we allow
> grouping of perf_sw_context sampling event as group leader and
> perf_{hw|invalid}_context counting event as group member. I think no. It
> should just work fine. And, there could be real usecases of it as you
> described in one old thread[1].

Like you I need to bend my brain around this again, but I'm not seeing a
contradiction. The use-case from [1] is a software sampler with a bunch
of non-sampling uncore events.

The uncore events aren't sampling, the are simply read by the software
event (SAMPLE_READ). And moving the sampling software event to the
non-sample capable uncore PMU shouldn't matter.

That is; the code as it stands here seems right, we should check
is_sampling_event() against an event's native pmu->capabilities.

Or am I misunderstanding things?