Re: [PATCH] sched: Support current clocksource handling in fallback sched_clock().
From: Paul Mundt
Date: Tue May 26 2009 - 19:44:52 EST
On Tue, May 26, 2009 at 04:25:03PM -0700, john stultz wrote:
> On Wed, 2009-05-27 at 08:08 +0900, Paul Mundt wrote:
> > On Tue, May 26, 2009 at 10:17:02PM +0200, Thomas Gleixner wrote:
> > > On Tue, 26 May 2009, Peter Zijlstra wrote:
> > > > On Tue, 2009-05-26 at 16:31 +0200, Linus Walleij wrote:
> > > > > The definition of "rating" from the kerneldoc does not
> > > > > seem to imply that, it's a subjective measure AFAICT.
> > >
> > > Right, there is no rating threshold defined, which allows to deduce
> > > that. The TSC on x86 which might be unreliable, but usable as
> > > sched_clock has an initial rating of 300 which can be changed later
> > > on to 0 when the TSC is unusable as a time of day source. In that
> > > case clock is replaced by HPET which has a rating > 100 but is
> > > definitely not a good choice for sched_clock
> > >
> > > > > Else you might want an additional criteria, like
> > > > > cyc2ns(1) (much less than) jiffies_to_usecs(1)*1000
> > > > > (however you do that the best way)
> > > > > so you don't pick something
> > > > > that isn't substantially faster than the jiffy counter atleast?
> > >
> > > What we can do is add another flag to the clocksource e.g.
> > > CLOCK_SOURCE_USE_FOR_SCHED_CLOCK and check this instead of the
> > > rating.
> > >
> > Ok, so based on this and John's locking concerns, how about something
> > like this? It doesn't handle the wrapping cases, but I wonder if we
> > really want to add that amount of logic to sched_clock() in the first
> > place. Clocksources that wrap frequently could either leave the flag
> > unset, or do something similar to the TSC code where the cyc2ns shift is
> > used. If this is something we want to handle generically, then I'll have
> > a go at generalizing the TSC cyc2ns scaling bits for the next spin.
> Yea. So this is a little better. There's still a few other issues to
> 1) What if a clocksource is registered that has the _SCHED_CLOCK bit
> set, but is not selected for timekeeping due it being unstable like the
See, this is what I thought the rating information was useful for, as the
rating is subsequently dropped if it is not usable. But perhaps it makes
more sense to just clear the bit at the same time that the rating is
lowered once it turns out to be unstable.
> 2) Conditionally returning jiffies if the lock is held seems troubling.
> Might get some crazy values that way.
What would you recommend instead? We do not want to spin here, and if we
are in the middle of changing clocksources and returning jiffies anyways,
then this same issue pops up in the current sched_clock() implementation
regardless of whether we are testing for lock contention or not.
Likewise, even if we were to spin, the same situation exists if the new
clocksource does not have the _SCHED_CLOCK bit set and we have to fall
back on jiffies anyways, doesn't it?
Put another way, and unless I'm missing something obvious, if we ignore
my changes to sched_clock(), how is your concern not applicable to case
where we are changing clocksources and using generic sched_clock() as it
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/