Re: Linux 2.6.29-rc6

[Ingo Molnar <mingo@elte.hu>]

* john stultz <johnstul@us.ibm.com> wrote:

> On Thu, 2009-03-05 at 19:13 -0800, john stultz wrote:
> > On Thu, 2009-03-05 at 09:43 +0100, Ingo Molnar wrote:
> > > * john stultz <johnstul@us.ibm.com> wrote:
> > > 
> > > > > Ingo, Thomas: On the hardware I'm testing the fast-pit 
> > > > > calibration only triggers probably 80-90% of the time. About 
> > > > > 10-20% of the time, the initial check to 
> > > > > pit_expect_msb(0xff) fails (count=0), so we may need to look 
> > > > > more at this approach.
> > > 
> > > We definitely need to improve calibration quality.
> > > 
> > > The question is - why does fast-calibration fail 10-20% of the 
> > > time on your test-system? Also, why exactly do we miscalibrate? 
> > > Could you please have a look at that?
> > 
> > Working on it, I just wanted to let you know I was seeing some different
> > odd behavior then Jesper.
> > 
> > > One theory would be that the PIT readout is unreliable. Windows 
> > > does not make use of it, so it's not the most tested aspect of 
> > > the PIT. Is that what happens on your box?
> > 
> > Still looking into it, but from my initial debugging it seems that by
> > reading the PIT very quickly after setting it, we may be getting junk
> > values. If I re-read the PIT again, I see the expected 0xff value. 
> > 
> > Its been somewhat of a heisenbug, as if I add any printk's or even just
> > a mb() after the outb it seems to make the problem go away (or just rare
> > enough I don't have the patience to reproduce it :)
> > 
> > So I don't know if a small delay is appropriate here (seems counter
> > productive to the whole fast-pit calibration ;) or if we should just try
> > to catch these bad reads and try again before failing?
> 
> Maybe something like the following? (Not tested heavily yet!)
> 
> Again, just for clarity, as we've mixed a few issues here, this patch is
> for a side issue and not related to the original regression reported by
> Jesper. I'm still waiting on debug output from Jesper to further
> diagnose whats going wrong with his TSC calibration.
> 
> thanks
> -john
> 
> 
> Apparently some hardware may occasionally return junk values if you try
> to read the pit immediately after setting it. This causes the
> pit_expect_msb() to occasionally fail (~10% of the time).
> 
> This patch tries to work around this issue by not failing if the first
> read right after setting the PIT is not what we expect.
> 
> NOT FOR INCLUSION (yet!)
> 
> Signed-off-by: John Stultz <johnstul@us.ibm.com>
> 
> diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
> index 599e581..2ca5ba4 100644
> --- a/arch/x86/kernel/tsc.c
> +++ b/arch/x86/kernel/tsc.c
> @@ -280,8 +280,17 @@ static inline int pit_expect_msb(unsigned char val)
>  	for (count = 0; count < 50000; count++) {
>  		/* Ignore LSB */
>  		inb(0x42);
> -		if (inb(0x42) != val)
> +		if (inb(0x42) != val) {
> +			/*
> +			 * If we're too fast, we may read
> +			 * junk values right after we set
> +			 * the PIT. So if this is the first
> +			 * read, try again
> +			 */
> +			if (val == 0xff && count == 0)
> +				continue;
>  			break;

We could do something like that if it helps the end result. But 
this special thing inside the loop should just be an 
unconditional inb(0x42) outside the loop. It does not hurt 
performance there, and we'll get simpler code that way.

But ... i really dont like how we rely on PIT readouts and how 
we work around PIT readout artifacts. Only Linux does PIT 
readouts while Windows does not - so we rely on a under-tested 
aspect of PC hardware.

I think we should think about a fundamentally different, IRQ 
driven way of calibration. For example we could program a 27 
milliseconds PIT periodic interrupt with the maximum count and 
measure its arrival timestamp in two subsequent interrupts.

We could do that with about 1-2 usecs precision realistically 
(this early during bootup we are really quiescent) - and over a 
27,000 usecs period that gives us an accuracy of 1:13500, or 
about 75 ppm. That's still only about 50 milliseconds spent 
calibrating, so very fast.

We can re-write the IRQ#0 vector with a special temporary 
calibration interrupt handler to make this really single-purpose 
and precise.

Hm?

	Ingo