Re: [PATCH] Fix TSC calibration issues

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, 4 Sep 2008, Willy Tarreau wrote:
>
> Basically, I would do this :
> 
>     pit1 = readpit();
>     while (readpit() == pit1);
>     t1 = rdtsc(); // precise beginning of tick 0
>     while (readpit() != pit1 - 5000);
>     t2 = rdtsc(); // precise beginning of tick 5000

There's a few caveats here:

 - the "readpit()" has to read without actually latching the value

   latching the PIT value will stop counting.

 - and all the docs say that you have to be careful about reading the PIT 
   without latching it because the two 8-bit accesses aren't atomic.

so the above will work in practice, but there are dangers.

The best way to fix most of the dangers is probably to only care about the 
*high* byte, so that it doesn't matter if the low byte doesn't match the 
high byte.

So you could probably change your version to wait for 4096 cycles (a 
change of 16 in the high byte):

	static unsigned char read_pit_msb(void)
	{
		/* Read but throw away the LSB */
		inb(0x42);
		return inb(0x42);
	}

	..
	/* PIT ch2: square wave, full 16-bit count */
	outb(0xb6, 0x43);
	outb(0, 0x42);
	outb(0, 0x42);
	..

	unsigned char pit = read_pit_msb();
	/* Wait until the MSB changes */
	while (read_pit_msb() == pit1);
	t1 = rdtsc();
	while ((unsigned char) (pit - read_pit_msb()) < 9);
	t2 = rdtsc();

and it might work out ok without explicit latching, and without having to 
worry about low/high bytes being out of sync.

> If someone wants to test this, I'd be interested in the number of
> ticks required to get a good accuracy, I bet that even with a few
> hundred ones it's already precise by a few ppm (about the precision
> of the input clock in fact).

I actually tested a patch with a counter value of just 1024, and I got the 
right answer. 

But if the busy loops aren't busy (due to MSI or virtualization), then all 
those things fly out the window.

			Linus