Re: [PATCH 0/3] timekeeping: Improved NOHZ frequency steering (v2)

[Miroslav Lichvar <mlichvar@redhat.com>]

On Tue, Jul 15, 2014 at 09:02:38PM -0700, John Stultz wrote:
> On 07/08/2014 04:08 AM, Miroslav Lichvar wrote:
> > I spent some time trying to figure out a workaround for the nanosecond
> > rounding, but I didn't find anything that wouldn't complicate the mult
> > adjustment logic and bring back the problems which the direct division
> > approach is supposed to solve.
> >
> > It seems it may be a while before the old vsyscalls are fixed. How
> > about including only the first two patches from this set for now?

> So thanks for the ping here. If you're happy with the first two as an
> initial step, I'd be willing to try to push those in. The only trouble
> is there's a whole lot of timekeeping churn headed for 3.17 that Thomas
> has cooked up. While there isn't likely to be direct conflicts in the
> changes, I get nervous about mixing too many changes in subtle code at once.

I'm sorry for digging up this skeleton. Are we any closer to being
able to remove CONFIG_GENERIC_TIME_VSYSCALL_OLD, which prevented
simplifying the steering logic of the internal clock?

The two patches added in 3.17 helped, but there is still the problem
that it takes too long for the kernel clock to converge to the
internal NTP time when a large error was accumulated, e.g. a large
frequency adjustment was made to correct the initial offset of the
clock.

It seems the error can reach milliseconds and take hours or even days
to be corrected. As the rate at which the error is decreasing is not
constant (due to random clock updates), it doesn't seem to be possible
to synchronize the clock with better stability than few tens or
hundreds of nanoseconds.

With the new PTP KVM clock the problem can be easily observed. Here is
a graph of the offset and frequency as measured by chronyd when
configured to synchronize the guest's clock to the host using the
virtual PHC. In the middle is the point when the NTP error reached
zero. The apparent frequency jumped by about 50 ppb and the offset
improved by an order of magnitude.

https://mlichvar.fedorapeople.org/tmp/kvm_phc.png

I see this with real PHCs and PTP/NTP synchronization too. It's very
confusing when the timekeeping changes so much for no apparent reason.
If we can't remove the old vsyscalls yet, I was thinking maybe a new
flag could be added to adjtimex to report the error, so applications
can at least detect this problem and consider stepping the clock in
order to reset the error?

Thoughts?

-- 
Miroslav Lichvar