Re: [PATCH v2] tracing/perf: Move rcu_irq_enter/exit_irqson() to perf trace point hook

[Steven Rostedt]

On Tue, 11 Feb 2020 11:27:36 -0500 (EST)
Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:

> ----- On Feb 11, 2020, at 11:18 AM, rostedt rostedt@xxxxxxxxxxx wrote:
> 
> > On Tue, 11 Feb 2020 16:34:52 +0100
> > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >   
> >> > +		if (unlikely(in_nmi()))
> >> > +			goto out;  
> >> 
> >> unless I'm mistaken, we can simply do rcu_nmi_enter() in this case, and
> >> rcu_nmi_exit() on the other end.
> >>   
> >> > +		rcu_irq_enter_irqson();  
> > 
> > The thing is, I don't think this can ever happen. We've had in the
> > tracepoint.h:
> > 
> >		/* srcu can't be used from NMI */			\
> >		WARN_ON_ONCE(rcuidle && in_nmi());			\
> > 
> > And this has yet to trigger.  
> 
> But that "rcuidle" state is defined on a per-tracepoint basis, whereas
> "!rcu_is_watching()" is a state which depends on the current execution
> context. I don't follow how the fact that this WARN_ON_ONCE() never
> triggered allows us to infer anything about (!rcu_is_watching() && in_nmi()).
>

The "_rcuidle()" version of the tracepoint was to be used in places
that RCU may not be watching, otherwise you would get a lockdep splat.

As that "rcuidle" variable is a hardcoded constant, it would be
compiled out when rcuidle is zero. But, in all purposes, rcuidle is
basically equivalent to rcu_is_watching(), because if it wasn't you
would have lockdep splats.

-- Steve