Re: [PATCH] sched: Support current clocksource handling in fallback sched_clock().

[Paul Mundt <lethal@linux-sh.org>]

On Wed, May 27, 2009 at 09:25:25AM -0700, Daniel Walker wrote:
> On Wed, 2009-05-27 at 09:15 +0900, Paul Mundt wrote:
> 
> >  unsigned long long __attribute__((weak)) sched_clock(void)
> >  {
> > +	/*
> > +	 * Use the current clocksource when it becomes available later in
> > +	 * the boot process, and ensure that it is usable for sched_clock().
> > +	 */
> > +	if (clock && (clock->flags & CLOCK_SOURCE_USE_FOR_SCHED_CLOCK))
> > +		return cyc2ns(clock, clocksource_read(clock));
> > +
> > +	/* Otherwise just fall back on jiffies */
> >  	return (unsigned long long)(jiffies - INITIAL_JIFFIES)
> 
> Do we really need all this complexity in a fast path? the jiffies
> clocksource is static and always ready. Could we instead remove the
> usage of "clock" and create a new pointer called "sched_clocksource" and
> initialize it to the jiffies clock, and allow the clocksource management
> code to update that new pointer based on the
> CLOCK_SOURCE_USE_FOR_SCHED_CLOCK flag when new clocksources are
> registered/removed/marked unstable etc..
> 
> that would eliminate all the code in sched_clock except one line,
> 
> unsigned long long __attribute__((weak)) sched_clock(void)
> {
> 	return cyc2ns(sched_clocksource, clocksource_read(sched_clocksource));
> }
> 
It would, but then what would that do to the sched_clock() value? It will
reset once a CLOCK_SOURCE_USE_FOR_SCHED_CLOCK clocksource comes along,
which basically means that sched_clock() then rolls back, which is worse
than simply being delayed a bit. So even if we were to use a special
sched_clocksource, it still could not be used until it was set up at
registration time.

Now we could of course add a sched_clocksource assignment in
select_clocksource(), but that's really no different than just testing
'clock' as we do today. All it does is move the flag test in to a
different path.

So, I think the current v3 is still the best solution.