Re: [patch 00/21] high resolution timers / dynamic ticks - V2

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

On Mon, 2006-10-02 at 15:08 -0400, Valdis.Kletnieks@vt.edu wrote:
> On Mon, 02 Oct 2006 11:38:36 PDT, john stultz said:
> 
> > Hmmm. So w/ -mm2 we're seeing the TSC get detected as running too slowly
> > (and its replaced w/ the ACPI PM), but for some reason that doesn't
> > happen w/ the dynticks patch.
> 
> It's been switching to ACPI PM for somewhere near forever, I never bothered
> to check into that because the PM timer provides a reasonably stable clock
> source (it drifts at about 24 ppm and NTP is happy with it, and I haven't
> gotten annoyed at the fact the PM timer is slow to read...)
> 
> I wonder if the TSC has been broken for forever on this box, and I'm just
> seeing it because dynticks doesn't fall over to PM timer..

This is what I suspect is the issue. Probably due to the new jiffies
accounting being now time based, and one of the TSC unstable checks (the
one you're tripping) being jiffies based. A tad bit circular :). I'm
working w/ tglx to see what we can do here.

> > Now, how is cpuspeed changing the cpufreq? Is it using the /sys
> > interface? I've got hooks in so when the cpufreq changes we should mark
> > it unstable and fall back to ACPI PM, but maybe I missed whatever hook
> > cpuspeed is using.
> 
> Looking at the source, it appears to do this:
> 
> const char SYSFS_CURRENT_SPEED_FILE[] =
>      "/sys/devices/system/cpu/cpu%u/cpufreq/scaling_setspeed";
> 
> // set the current CPU speed
> void set_speed(unsigned value)
> {
> #ifdef DEBUG
>     fprintf(stderr, "[cpu%u] Setting speed to: %uKHz\n", cpu, value);
> #endif
>     write_line(CURRENT_SPEED_FILE, "%u\n", value);
>     // give CPU / chipset voltage time to settle down
>     usleep(10000);
> }

I'll also take a peek there and see if I cannot add an extra hook, so we
don't have to rely on the jiffies stability check.

thanks
-john