On Fri, Apr 12, 2019 at 01:16:01PM +1000, Nicholas Piggin wrote:
> Paul E. McKenney's on April 12, 2019 1:42 am:
> > On Tue, Apr 09, 2019 at 07:21:54PM +1000, Nicholas Piggin wrote:
> >> Thomas Gleixner's on April 6, 2019 3:54 am:
> >> > On Fri, 5 Apr 2019, Nicholas Piggin wrote:
> >> >> Thomas Gleixner's on April 5, 2019 12:36 am:
> >> >> > On Thu, 4 Apr 2019, Nicholas Piggin wrote:
> >> >> >
> >> >> >> I've been looking at ways to fix suspend breakage with CPU0 as a
> >> >> >> nohz CPU. I started looking at various things like allowing CPU0
> >> >> >> to take over do_timer again temporarily or allowing nohz full
> >> >> >> to be stopped at runtime (that is quite a significant change for
> >> >> >> little real benefit). The problem then was having the housekeeping
> >> >> >> CPU go offline.
> >> >> >>
> >> >> >> So I decided to try just allowing the freeze to occur on non-zero
> >> >> >> CPU. This seems to be a lot simpler to get working, but I guess
> >> >> >> some archs won't be able to deal with this? Would it be okay to
> >> >> >> make it opt-in per arch?
> >> >> >
> >> >> > It needs to be opt in. x86 will fall on its nose with that.
> >> >>
> >> >> Okay I can add that.
> >> >>
> >> >> > Now the real interesting question is WHY do we need that at all?
> >> >>
> >> >> Why full nohz for CPU0? Basically this is how their job system was
> >> >> written and used, testing nohz full was a change that came much later
> >> >> as an optimisation.
> >> >>
> >> >> I don't think there is a fundamental reason an equivalent system
> >> >> could not be made that uses a different CPU for housekeeping, but I
> >> >> was assured the change would be quite difficult for them.
> >> >>
> >> >> If we can support it, it seems nice if you can take a particular
> >> >> configuration and just apply nohz_full to your application processors
> >> >> without any other changes.
> >> >
> >> > This wants an explanation in the patches.
> >>
> >> Okay.
> >>
> >> > And patch 4 has in the changelog:
> >> >
> >> > nohz_full has been successful at significantly reducing jitter for a
> >> > large supercomputer customer, but their job control system requires
> >> > CPU0
> >> > to be for housekeeping.
> >> >
> >> > which just makes me dazed and confused :)
> >> >
> >> > Other than some coherent explanation and making it opt in, I don't think
> >> > there is a fundamental issue with that.
> >>
> >> I will try to make the changelogs less jibberish then :)
> >
> > Maybe this is all taken care of now, but do the various clocks stay
> > synchronized with wall-clock time if all CPUs are in nohz_full mode?
> > At one time, at least one CPU needed to keep its scheduler-clock
> > interrupt going in order to keep things in sync.
>
> Ah, may not have been clear in the changelog -- the series still
> requires at least one CPU present at boot time to be a housekeeper
> that keeps things running. So conceptually this doesn't change
> anything about runtime behaviour, the main change is the boot-time
> handoff from CPU0.
I did miss that, thank you for the update.
> > The ppc timebase register might make it possible to do this without any
> > scheduler-clock interrupts, but figured I should check. ;-)
>
> I dont know all this code too well, but if we really wanted to push
> things, I think nohz-full could be more aggressive in shutting down
> the tick and possibly even avoiding a housekeeping CPU completely, but
> you would have to do that work on user->kernel switch too. Likely the
> complexity and overhead is not worthwhile.
There was some RCU functionality that detected when all the
non-housekeeping CPUs went idle, but it went unused for some years, so
I reverted it. This revert commit is at tag sysidle.2017.05.11a in my
-rcu tree. If it is actually going to be used, I could of course add
it back. ;-)
> Other thing is you might be able to avoid the jiffies tick completely
> and change jiffies to read from timebase register. Lot of interesting
> things we could try.
Or make userspace use the timebase register to avoid the need for
in-kernel time adjustments, though the connection to NTP and similar
would still need to be maintained. I supposed that the jiffies
counter could be fixed up on entry to the kernel?
Thanx, Paul
