Re: clocksource change of behavior in 2.6.22 compared to 2.6.20 causes massive system clock slowdown

["Alessandro Suardi" <alessandro.suardi@gmail.com>]

On 7/18/07, john stultz <johnstul@us.ibm.com> wrote:
> On Wed, 2007-07-18 at 00:31 +0200, Alessandro Suardi wrote:
> > On 7/11/07, john stultz <johnstul@us.ibm.com> wrote:
> > > On Wed, 2007-07-11 at 01:31 +0200, Alessandro Suardi wrote:
> > > > On 7/10/07, john stultz <johnstul@us.ibm.com> wrote:
> > > > > On Tue, 2007-07-10 at 00:29 -0700, Andrew Morton wrote:
> > > > > > On Mon, 9 Jul 2007 16:27:59 +0200 "Alessandro Suardi" <alessandro.suardi@gmail.com> wrote:
> > > > > >
> > > > > > > My oldish AMD K7-800's clock began falling behind after
> > > > > > >  rebooting from 2.6.20 (and 109 days uptime with a spotless
> > > > > > >  clock) into 2.6.22; time lost is about four minutes each hour.
> > > > > > >
> > > > > > > Turns out that 2.6.22 marks my TSC as unstable and starts
> > > > > > >  using PIT instead. Rebooting 2.6.22 with clocksource=tsc
> > > > > > >  gets the original stable system time back.
> > > > >
> > > > > Alessandro,
> > > > > Can you send me dmesg output for 2.6.20 and 2.6.22 (without
> > > > > clocksource=tsc)?
> > > >
> > > > Actually, I lied a little bit - it was 2.6.22 with clock=tsc (which
> > > >  warns on boot about clock= being deprecated in favor of
> > > >  clocksource= ). I assume behavior is identical for now.
> > > >
> > > > Please find attached the dmesg ring (incomplete, as the
> > > >  kernel ring size I have seems too small to hold the full
> > > >  buffer, but it seems to have all the interesting stuff) of
> > > >  2.6.20, 2.6.22, 2.6.22 with clock=tsc.
> > > >
> > > > If you need more info, just ask. I'll be out of the country
> > > >  from July 12 to the morning of July 16, and again from
> > > >  July 17 to July 20, so if you'd rather get at this later on,
> > > >  it's okay for me ;)
> > >
> > > You're dmesg output got chopped at the top. Please increase the kernel
> > > log buffer size.
> >
> > Done, please find attached both 2620 and 2622 full dmesg output,
> >  with no clock= or clocksource= parameters.
> >
> > > Sounds like your PIT frequency is out of whack, and I'm guessing the
> > > generic clocksource watchdog blames the TSC and disqualifies it.
> > >
> > > Few things to check:
> > > 1) Make sure you're running the latest BIOS.
> >
> > I'm probably not, though I'm not sure I'd flash anything on such
> >  an old machine - been running it since RedHat 9 with the
> >  current BIOS. Plus, I removed the floppy drive to make room
> >  for an extra IDE disk... it'd take me a little bit to find out whether
> >  I still have the floppy drive around.
> >
> > > 2) See if booting w/ noapic changes anything
> >
> > Nope, 2622+noapic => PIT is still used and clock very quickly
> >  lags behind.
> >
> > For the moment being I'll keep booting with clocksource=tsc.
> > Let me know whether you want me to test anything more...
>
> Hmm. One other thing to check: If you boot 2.6.20 w/ clocksource=pit, is
> it consistent w/ 2.6.22 and same slow timekeeping issue shows up?

Yes, 2620+clocksource=pit is noticeably slow in timekeeping as well.

> From my laptop (Dell D610 running 2.6.22-git10 on top of Fedora7)
I ssh'd into my K7-800 and with two terminal windows I ran the same
simple bash loop, starting more or less at the same time:
while :; do let i=i+1; echo $i; sleep 1; done

When the laptop reached 60, the K7-800 was at 57. In order to
eliminate any possible doubt about how I started the two sessions,
I let the loop run; when the laptop reached 120, the K7-800 was
displaying 114. So it's indeed a 3-second per minute loss with
PIT in either 2620 or 2622.

--alessandro

"Did you get married but forgot to get divorced ?"

    (Danny and Dusty, 'The Good Old Days')
-
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/