On Mon, Oct 26, 2020 at 10:28:07AM -0400, Mathieu Desnoyers wrote:
> ----- On Oct 26, 2020, at 4:20 AM, Peter Zijlstra peterz@xxxxxxxxxxxxx wrote:
>
> > On Fri, Oct 23, 2020 at 05:13:59PM -0400, Joel Fernandes wrote:
> >> On Fri, Oct 23, 2020 at 03:53:52PM -0400, Michael Jeanson wrote:
> >> > From: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
> >> >
> >> > Considering that tracer callbacks expect RCU to be watching (for
> >> > instance, perf uses rcu_read_lock), we need rcuidle tracepoints to issue
> >> > rcu_irq_{enter,exit}_irqson around calls to the callbacks. So there is
> >> > no point in using SRCU anymore given that rcuidle tracepoints need to
> >> > ensure RCU is watching. Therefore, simply use sched-RCU like normal
> >> > tracepoints for rcuidle tracepoints.
> >>
> >> High level question:
> >>
> >> IIRC, doing this increases overhead for general tracing that does not use
> >> perf, for 'rcuidle' tracepoints such as the preempt/irq enable/disable
> >> tracepoints. I remember adding SRCU because of this reason.
> >>
> >> Can the 'rcuidle' information not be pushed down further, such that perf does
> >> it because it requires RCU to be watching, so that it does not effect, say,
> >> trace events?
> >
> > There's very few trace_.*_rcuidle() users left. We should eradicate them
> > and remove the option. It's bugs to begin with.
>
> I agree with Peter. Removing the trace_.*_rcuidle weirdness from the tracepoint
> API and fixing all callers to ensure they trace from a context where RCU is
> watching would simplify instrumentation of the Linux kernel, thus making it harder
> for subtle bugs to hide and be unearthed only when tracing is enabled. This is
> AFAIU the general approach Thomas Gleixner has been aiming for recently, and I
> think it is a good thing.
>
> So if we consider this our target, and that the current state of things is that
> we need to have RCU watching around callback invocation, then removing the
> dependency on SRCU seems like an overall simplification which does not regress
> feature-wise nor speed-wise compared with what we have upstream today. The next
> steps would then be to audit all rcuidle tracepoints and make sure the context
> where they are placed has RCU watching already, so we can remove the tracepoint
> rcuidle API. That would effectively remove the calls to rcu_irq_{enter,exit}_irqson
> from the tracepoint code.
>
> This is however beyond the scope of the proposed patch set.
You are right, it doesn't regress speedwise - I got confused since the code
was modified to call rcu_enter_irqson() even for the rcuidle case (which I
had avoided when I added SRCU). So in current code, SRCU is kind of
pointless. I think keep the patch in the series.
thanks,
- Joel
