Re: [RFC][PATCH] use cycle_t instead of u64 in struct time_interpolator

[john stultz <johnstul@us.ibm.com>]

On Wed, 2007-01-03 at 21:23 +0100, Helge Deller wrote:
> On Wed Jan 3 2007, john stultz wrote:
> > On Wed, 2007-01-03 at 19:36 +0100, Helge Deller wrote:
> > > On Wed Jan 3 2007, Christoph Lameter wrote:
> > > > On Tue, 2 Jan 2007, Helge Deller wrote:
> > > >
> > > > > As far as I could see, this patch does not change anything for the
> > > > > existing architectures which use this framework (IA64 and SPARC64),
> > > > > since "cycles_t" is defined there as unsigned 64bit-integer anyway
> > > > > (which then makes this patch a no-change for them).
> > > >
> > > > The 64bit nature of some entities was so far necessary to get the
> > > > proper accuracy of interpolation. Maybe it can be made to work with 32 bit
> > > > entities. The macro GET_TI_SECS must work correctly and the less bits are
> > > > specified in shift the less self-tuning accuracy you will get.
> > >
> > > Yes, it was easily possible to make it 32bit-ready without loosing the accuracy.
> > >
> > > Nevertheless, in the meantime John Stultz pointed me to the CONFIG_GENERIC_TIME framework,  and I implemented it that way:
> > > http://git.parisc-linux.org/?p=linux-2.6.git;a=commit;h=b6de83b58b8b07f057deacdef8a95b6c32d1c4e6
> >
> > This looks pretty good, although setting the rating to 200 for a
> > clocksource you don't want to use seems a bit high (there's a rough
> > rating scale in clocksource.h). Zero is probably what you want to use
> > there.
> >
> > > http://git.parisc-linux.org/?p=linux-2.6.git;a=commit;h=f70a979c843e4610edfb2a316648fe8ae8718f69
> >
> > This looks to be correct, although as the clocksource infrastructure
> > evolves it looks like we'll be removing the update_callback code in the
> > future. So this is fine for now, but will probably need a reevaluation
> > at some point.
> >
> > Also to avoid jumping between clocksources, I'd keep the initial
> > disqualification that occurs before you register the clocksource
> > (otherwise it will be used for one tick, then be disqualified and you're
> > back to jiffies).
> 
> That's true, but James Bottomley noticed, that time_init() is called before
> we've done system inventory (which detects if we have a SMP box with multiple CPUs),
> so num_online_cpus() would always be one. The update_callback function enables
> us to switch back to jiffies if we actually run on a SMP box.

Ah, yes, right now most clocksources currently register themselves at
module_init time to avoid this sort of thing. Although we might move
that to be a bit earlier, since there is some need for high precision
timeofday earlier at bootup.

> That said, it would be nice to keep the update_callback() functionality or provide another
> nice solution around that problem...

Yea. There will be an alternate solution, hopefully something more
direct that just removes the clocksource and moves on, rather then
polling each tick for change.

thanks
-john


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/