Re: [PATCH v4 16/27] tracing: Remove regular RCU context for _rcuidle tracepoints (again)

[Joel Fernandes]

On Fri, Mar 06, 2020 at 01:59:25PM -0500, Steven Rostedt wrote:
[snip]
> > > -			rcu_irq_enter_irqson();				\
> > > -		}							\
> > >  									\
> > >  		it_func_ptr = rcu_dereference_raw((tp)->funcs);		\
> > >  									\
> > >  		if (it_func_ptr) {					\
> > >  			do {						\
> > > +				int rcu_flags;				\
> > >  				it_func = (it_func_ptr)->func;		\
> > > +				if (rcuidle &&				\
> > > +				    (it_func_ptr)->requires_rcu)	\
> > > +					rcu_flags = trace_rcu_enter();	\
> > >  				__data = (it_func_ptr)->data;		\
> > >  				((void(*)(proto))(it_func))(args);	\
> > > +				if (rcuidle &&				\
> > > +				    (it_func_ptr)->requires_rcu)	\
> > > +					trace_rcu_exit(rcu_flags);	\  
> > 
> > Nit: If we have incurred the cost of trace_rcu_enter() once, we can call
> > it only once and then call trace_rcu_exit() after the do-while loop. That way
> > we pay the price only once.
> >
> 
> I thought about that, but the common case is only one callback attached at
> a time. To make the code complex for the non common case seemed too much
> of an overkill. If we find that it does help, it's best to do that as a
> separate patch because then if something goes wrong we know where it
> happened.
> 
> Currently, this provides the same overhead as if each callback did it
> themselves like we were proposing (but without the added need to do it for
> all instances of the callback).

That's ok, it could be a separate patch.

thanks,

 - Joel