Re: time patches by Roman Zippel
From: Roman Zippel
Date: Tue Feb 14 2006 - 09:31:39 EST
On Tue, 14 Feb 2006, Ulrich Windl wrote:
> > > 15_time_offset and 18_time_freq change some well-known constants (like MAXPHASE)
> > > by three orders of magnitude.
> --- linux-2.6-mm.orig/include/linux/timex.h 2005-12-21 12:12:00.000000000 +0100
> +++ linux-2.6-mm/include/linux/timex.h 2005-12-21 12:12:08.000000000 +0100
> @@ -95,11 +95,11 @@
> #define SHIFT_USEC 16 /* frequency offset scale (shift) */
> #define FINENSEC (1L << SHIFT_SCALE) /* ~1 ns in phase units */
> -#define MAXPHASE 512000L /* max phase error (us) */
> +#define MAXPHASE 500000000L /* max phase error (ns) */
> #define MAXFREQ (512L << SHIFT_USEC) /* max frequency error (ppm) */
> #define MINSEC 16L /* min interval between updates (s) */
> #define MAXSEC 1200L /* max interval between updates (s) */
> -#define NTP_PHASE_LIMIT (MAXPHASE << 5) /* beyond max. dispersion */
> +#define NTP_PHASE_LIMIT ((MAXPHASE / 1000) << 5) /* beyond max. dispersion */
The reference timex.h has the same change for MAXPHASE and NTP_PHASE_LIMIT
is Linux specific. Where is the problem?
> > > the new adjtime() (16_time_adjust, 12_time_adj) changes the semantics: Since about
> > > Linux 0.99, adjtime() had the adjtime_is_accurate property, i.e. on the long term
> > > it behaved like an addition.
> > I disagree, could you please explain how you come to this conclusion?
> + tick_nsec_curr += time_adjust * 1000 / HZ;
> Assuming 1024Hz interrupt frequency:
> (1µs * 1000) / 1024 == 0ns; 0 * 1024 == 0µs, not 1µs
> (2µs * 1000) / 1024 == 1ns; 1 * 1024 == 1.024µs, not 2µs
Ok, I didn't put much effort into optimizing it for uncommon HZ values.
Why is it so important? It's currently unused on any Linux machine
synchronized via NTP.
> > The patches don't change the behaviour beyond that they increase
> > resolution and precision. Only the final patch changes the ntp code to
> > match the behaviour of ntp reference code without including all its mess.
> It's quite hard to tell: The code is very different what I've ever seen.
Actually it's not that hard, under http://www.xs4all.nl/~zippel/ntp/ you
can also find the user space test code I used to verify it.
kernel.tar.Z is the old reference code, which the current Linux code is
based on, under patches-kernel you can find a number of patches to convert
it to the new model and which match the new kernel implementation.
I updated the kern.dat and added a nano.sh script with matching test
parameters for nanokernel, so you can compare the output of both test