freebsd 10 kqueue timer regression

[John Baldwin]

On Thursday, October 02, 2014 06:49:49 PM Ian Lepore wrote:
> On Thu, 2014-10-02 at 16:15 -0600, Ian Lepore wrote:
> > On Thu, 2014-10-02 at 16:00 -0400, John Baldwin wrote:
> > > On Thursday, October 02, 2014 3:53:28 pm Ian Lepore wrote:
> > > > On Thu, 2014-10-02 at 12:47 -0700, Adrian Chadd wrote:
> > > > > I'm confused; it's doing 50 loops of a 20msec timer, right? So
> > > > > that's
> > > 
> > > 1000ms.
> > > 
> > > > Yes, so the entire loop should take 1000ms maybe + 1ms.  Instead it
> > > > takes 1070.  When I run it on an armv6 system running -current it
> > > > takes
> > > > 1050.  When I run it on my 8.4 desktop (pre-eventtimers) it takes
> > > > 1013.
> > > > 
> > > > -- Ian
> > > 
> > > What if you set kern.eventtimer.periodic=1?
> > 
> > Some interesting results...
> > 
> >            HZ   100   500    1000
> > 
> > ---------------------------------
> > periodic=0     1050  1050    1080
> > periodic=1     1110  1012    1049
> > 
> > 
> > The 1080 number was +/- 3ms, all the other numbers were +/- 1ms (except
> > for one outlier of 24363 at 100Hz non-periodic which I'm going to
> > pretend didn't happen).
> > 
> > The 1050 numbers are probably each 20ms sleep actually taking 21ms, but
> > the old tvtohz code with -1 adjustments from the old email thread isn't
> > in play anymore.  I don't know how to account for the other numbers at
> > all.  There's all kinds of stuff I don't understand in the new code
> > involving tick thresholds and such.
> > 
> > -- Ian
> 
> The attached patch seems to fix the problem in what I think is the most
> correct way: scheduling the callout with absolute times based on the
> time the current event was scheduled for plus the requested interval.
> The net effect should be metronomic events that do not drift (or phase
> shift if you prefer) over time, regardless of any latency involved in
> processing the events.
> 
> This makes all the numbers in the tests I ran above come out 1000.
> 
> It doesn't make me understand the strange results from the prior tests
> any better.
> 
> -- Ian

Are you running ntpd or ptpd?  If so, perhaps try the original tests without 
the patch.

That said, I think one of the reasons the old code worked was that the 
previous callout had the equivalent of the C_HARDCLOCK flag set.  Thus, when 
the timer interrupt fires and we rescheuled for N ticks, it was actually N 
ticks - <time since the timer interrupt that triggered this callout> (if that 
makes sense?).  Now I think what is happening is that that the 
callout_reset_sbt() is doing 'delta time t + <time since the timer interrupt>' 
and that that accounts for your numbers.  In that case, I wonder if the reason 
you are seeing such large gaps is that you have Cx states enabled (including 
C1E) and that those are making that second factor large enough to account for 
the skew.  Try setting machdep.idle=spin as a test (this is a better test than 
ntpd/ptpd I think).

The reason I don't like using C_ABSOLUTE is that in theory that will result in 
longer or shorter intervals if time is adjusted via ntp/ptp, whereas 
historically EVFILT_TIMER has been CLOCK_UPTIME based instead of 
CLOCK_REALTIME.

-- 
John Baldwin