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

[Paul E. McKenney]

On Tue, Feb 11, 2020 at 09:05:03AM -0500, Steven Rostedt wrote:
> On Tue, 11 Feb 2020 12:49:54 +0100
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Mon, Feb 10, 2020 at 05:06:43PM -0500, Steven Rostedt wrote:
> > > +	if (!rcu_watching) {						\
> > > +		/* Can not use RCU if rcu is not watching and in NMI */	\
> > > +		if (in_nmi())						\
> > > +			return;						\
> > > +		rcu_irq_enter_irqson();					\
> > > +	}								\  
> > 
> > I saw the same weirdness in __trace_stack(), and I'm confused by it.
> > 
> > How can we ever get to: in_nmi() && !rcu_watching() ? That should be a
> > BUG.  In particular, nmi_enter() has rcu_nmi_enter().
> > 
> > Paul, can that really happen?
> 
> The stack tracer connects to the function tracer and is called at all
> the places that function tracing can be called from. As I like being
> able to trace RCU internal functions (especially as they are complex),
> I don't want to set them all to notrace. But, for callbacks that
> require RCU to be watching, we need this check, because there's some
> internal state that we can be in an NMI and RCU is not watching (as
> there's some places in nmi_enter that can be traced!).
> 
> And if we are tracing preempt_enable and preempt_disable (as Joel added
> trace events there), it may be the case for trace events too.

Ah, thank you for the reminder!

Should Documentation/RCU/Design/Requirements/Requirements.rst be
updated to include this?

And I have to ask...  What happens if we are very early in from-idle
NMI entry (or very late in NMI exit), such that both in_nmi() and
rcu_is_watching() are returning false?  Or did I miss a turn somewhere?

							Thanx, Paul