Re: Regression in 2.6.27 caused by commit bfc0f59

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 1 Sep 2008, Larry Finger wrote:
> Thomas Gleixner wrote:
> > On Mon, 1 Sep 2008, Larry Finger wrote:
> > > The timed sleep is as accurate as I can measure.
> > > 
> > > I put in some test prints. The value of pm2 is zero when the else branch
> > > of
> > > the "if (hpet)" is entered; however, pm1 is 15768471. When we reach the
> > > do_div(tsc2, tsc1) statement, tsc2 is zero, which I think means that the
> > > two
> > > calls to tsc_read_refs() are returning the same junk value.
> > 
> > Ok, so the pmtimer is probably detected later as unusable and disabled.
> > Please check your logs for: "PM-Timer had inconsistent results:"
> 
> Booting 2.6.26, the dmesg output has a line that says:
> 
> PM-Timer running at invalid rate: 200% of normal - aborting.
> 
> Amazing that it should be exactly 200%. Why is the CPU running at half speed
> when the PM-Timer rate is measured?

The kernel assumes that the PM timer frequency is normal, so it does:

    read pm-timer start value, read TSC start value
    wait for a some time
    read pm-timer end value, read TSC end value

And the TSC frequency is calculated via:

                  TSC-End - TSC-Start
TSC-Frequency =  -------------------- * PM-Frequency
                   PM-End - PM-Start

So if your PM-Timer runs at the double frequency for reasons only
known to the Chip Manufacturer the kernel miscalculates the TSC
frequency by factor 0.5.

Simple rule of three.

Thanks,

	tglx