Re: [RFC - 0/9] Generic timekeeping subsystem (v. B5)

[john stultz <johnstul@us.ibm.com>]

On Wed, 2005-08-24 at 16:46 -0700, George Anzinger wrote:
> john stultz wrote:
> > On Tue, 2005-08-23 at 17:29 -0700, George Anzinger wrote:
> > 
> >>Roman Zippel wrote:
> >>
> >>>Hi,
> >>>
> >>>On Tue, 23 Aug 2005, john stultz wrote:
> >>>
> >>>
> >>>
> >>>>I'm assuming gettimeofday()/clock_gettime() looks something like:
> >>>>  xtime + (get_cycles()-last_update)*(mult+ntp_adj)>>shift
> >>>
> >>>
> >>>Where did you get the ntp_adj from? It's not in my example.
> >>>gettimeofday() was in the previous mail: "xtime + (cycle_offset * mult +
> >>>error) >> shift". The difference between system time and reference 
> >>>time is really important. gettimeofday() returns the system time, NTP 
> >>>controls the reference time and these two are synchronized regularly.
> >>>I didn't see that anywhere in your example.
> >>>
> > 
> > 
> >>If I read your example right, the problem is when the NTP adjustment 
> >>changes while the two clocks are out of sync (because of a late tick). 
> > 
> > 
> > Not quite. The issue that I'm trying to describe is that if, we
> > inconsistently calculate time intervals in gettimeofday and the timer
> > interrupt, we have the possibility for time inconsistencies.
> > 
> > The trivial example using the current code would be something like:
> > 
> > Again with my 2 cyc per tick clock, HZ=1000.
> > 
> > gettimeofday():
> > 	xtime + offset_ns
> > 
> > timer_interrupt:
> > 	xtime += tick_length + ntp_adj
> > 	offset_ns = 0
> > 
> > 0:  gettimeofday:  0 + 0 = 0 ns
> > 1:  gettimeofday:  0 + 500k ns = 500k ns
> > 2:  gettimeofday:  0 + 1M ns = 1M ns
> > 2:  timer_interrupt:  
> > 2:  gettimeofday:  1M ns + 0 ns = 1M ns
> > 3:  gettimeofday:  1M ns + 500k ns = 1.5M ns
> > 4:  gettimeofday:  1M ns + 1M ns = 2 ns
> > 4:  timer_interrupt (using -500ppm adjustment)
> > 4:  gettimeofday:  1,999,500 ns + 0 ns = 1,999,500 ns
> > 
> At point 4 you are introducing a NEW ntp adjustment.  This, I submit, 
> needs to actually have been introduced to the system prior to the 
> interrupt at point 2 with the first xtime change at point 4.  However, 
> gettimeofday() should be aware of it from the interrupt at point 2 and 
> be doing corrections from that time forward.  Thus when the point 4 
> interrutp happens xtime will be the same at the gettimeofday a ns earlier.

Yes, clearly a forward knowledge of the NTP adjustment is necessary for
gettimeofday(), because after the NTP adjustment has been accumulated
into xtime, there's nothing left for gettimeofday to adjust (its already
been applied). :)


> Likewise, gettimeofday() needs to know when to stop apply the correction 
> so that if a tick is late, it will apply the correction only for those 
> times that it was needed.  This, could be done by figuring the offset 
> thusly:
> 
> offset = (offset from last tick to end of ntp period * ntp_adj1) + 
> (offset from end of ntp period to now)

Well, in my example, the ntp_adjustment is a fixed nanosecond offset, so
it would be added to the nanosecond offset from the last tick (which is
how the current code works). If you are doing scaling (as you have in
the equation above), then the problem goes away, since you can apply the
adjustment consistently through any interval.

> I suppose it is possible that the latter part of the offset is also 
> under a different ntp correction which would mean a "* ntp_adj2" is 
> needed.  

Ok, so your forcing gettimeofday to be interval aware, so its applying
different fixed NTP adjustments to different chunks of the current
interval. The issue of course is if you're using fixed adjustments, is
that you have to have n ntp adjustments for n intervals, or you have to
apply the same ntp adjustment to multiple intervals. 


> I would argue that only two terms are needed here regardless of 
> how late a tick is.  This is because, I would expect the ntp system call 
> to sync the two clocks.  This means in your example, the ntp call would 
> have been made at, or prior to the timer interrupt at 2 and this is the 
> same edge that gettimeofday is to used to start applying the correction.

If you argue that we only need two adjustments, why not argue for only
one? You're saying have one adjustment that you apply for the first
tick's worth of time, and a second adjustment that you apply for the
following N ticks' worth of time in the interval. Why the odd base
case? 


> >>It would appear that gettimeofday would need to know that the NTP 
> >>adjustment is changing  (and to what).  It would also appear that this 
> >>is known by the ntp code and could be made available to gettimeofday. 
> >>If it is changing due to an NTP call, that system call, itself, 
> >>should/must force synchronization.  So the only case gettimeofday needs 
> >>to worry/know about is that an adjustment is to change at time X to 
> >>value Y.  Also, me thinks there is only one such change that can be 
> >>present at any given time.
> > 
> > 
> > Well, in many arches gettimeofday() works around the above issue by
> > capping the offset_ns value as such:
> 
> I think this may have been done with only usec gettimeofday.  Now that 
> we have clock_gettime() returning nsec, we need to be a bit more careful.

Indeed.

thanks
-john

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/