> On Tue, 30 Apr 1996, Wolfram Gloger wrote:
>
> > There was some talk about the pentium do_fast_gettimeoffset() routine
> > lately, and claimed that the timer would progress monotonously. Well,
> > it does not, for me. With the following program
>
> [ test program that checks for gettimeofday() monotonity ]
>
> i could reproduce it on an i486 system too, so it's not Pentium timer
> related. Stock kernel, IDE with irqs masked (prone to loosing jiffies?),
> no NTP, no fancy stuff.
>
> Maybe it's the NTP fix from Ulrich Windl that came in lately? [sorry if not]
I'd say, "no". It seems the little changes improved several points,
but I'm still seeking the "big point". My measurements of interrupt
delays (no IDE disk) showed about 99% under 5ms. at 10ms we would be
very close to loosing a tick. I've not investigated how the several
microtime routines and delayed interrupts cooperate.
For the NTP changes: Unless you play with the NTP settings, the
kernel clock should be totally unaffected by the changes. The
pentium fasttimer was not affected. A test yesterday over 4 hours of
clock synchronization showed no missed interrupts, but still the
"sawtooth" (up to one tick high, but very periodically) that I'm
fighting so much. A current theory is whether the fast time offsets
are applied (added) at every necessary part of the kernel routines.
I'll have to dig in the sources again.
But still, if anybody found a real problem, I'm not the pope, and I
make mistakes.
>
> -- mingo
Ulrich