Re: [PATCH RFC idle 2/3] arm: Avoid invoking RCU when CPU is idle

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Fri, Feb 03, 2012 at 01:55:09PM -0500, Steven Rostedt wrote:
> On Thu, 2012-02-02 at 22:04 -0800, Paul E. McKenney wrote:
> > On Thu, Feb 02, 2012 at 09:45:31PM -0500, Steven Rostedt wrote:
> 
> > It is an atomic instruction or two, plus some memory barriers.  Entering
> > idle is more heavyweight for RCU_FAST_NO_HZ.  But as you say, it is
> > entering and exiting idle.
> > 
> > But should I make an empty definition of RCU_NONIDLE() for some #define
> > or another?
> > 
> > 	#ifdef CONFIG_YOU_TELL_ME
> > 	#define RCU_NONIDLE(a) \
> > 		do { \
> > 			rcu_idle_exit(); \
> > 			do { a; } while (0); \
> > 			rcu_idle_enter(); \
> > 		} while (0)
> > 	#else /* #ifdef CONFIG_YOU_TELL_ME */
> > 	#define RCU_NONIDLE(a) do { } while (0);
> > 	#endif /* #else #ifdef CONFIG_YOU_TELL_ME */
> > 
> > Or is event tracing unconditional these days?
> 
> I don't like it. As it binds the RCU_NONIDLE to tracepoints only without
> any annotation that they are bound. Still doesn't help when tracepoints
> are configured but not enabled.
> 
> I have no problem in making a special TRACE_EVENT_IDLE() that does this
> inside the jump label. Basically what we have today is:
> 
> 
> 	if (static_branch(tracepoint_key)) {
> 		rcu_read_lock_sched_notrace();
> 		for (all attached tracepoints) {
> 			[...]
> 		}
> 		rcu_read_unlock_sched_notrace();
> 	}
> 
> Ideally we want the enter/exit idle inside that static_branch()
> condition:
> 
> 	if (static_branch(tracepoint_key)) {
> 		rcu_idle_exit();
> 		rcu_read_lock_sched_notrace();
> 		for (all attached tracepoints) {
> 			[...]
> 		}
> 		rcu_read_unlock_sched_notrace();
> 		rcu_idle_enter();
> 	}
> 
> The static_branch() is the jump label code when it's a nop when disabled
> and a jump to the tracing code when enabled:
> 
> 	nop; /* or jmp 2f */  <<--- jump label
> 1:	[ normal code ]
> 	ret;
> 
> 2:	[trace code]
> 	jmp 1b
> 
> 
> The jump label when disabled is just a nop that ignores the trace code
> (although current gcc has a bug that it currently doesn't do it this
> elegantly). When tracing is enabled the nop is converted to a jump to
> the tracing code. This makes tracepoints very light weight in hot paths.
> 
> Ideally, we want the exit/enter rcu idle with in the [trace code], which
> makes it not used when not needed.

So the idea is that if you have a trace event that is to be used in idle,
you use TRACE_EVENT_IDLE() rather than TRACE_EVENT() to declare that
trace event?  That would work for me, and might make for fewer changes
for the architecture guys.  Also, this should address the code-size
concerns we discussed yesterday.

So sounds good!

Is a DEFINE_EVENT_IDLE() also needed?  Or prehaps a
DECLARE_EVENT_CLASS_IDLE()?  My guess is "yes" for at least one of the
two based on include/trace/events/power.h.

I will keep RCU_NONIDLE() for at least a little while (reworking comments
to point out TRACE_EVENT_IDLE() and friends) in case there turn out to
be non-tracepoint uses of RCU in the idle loop.

							Thanx, Paul