Re: [PATCH 1/4] sched: Don't account tickless CPU load on tick

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Jan 19, 2016 at 07:56:47PM +0100, Peter Zijlstra wrote:
> On Tue, Jan 19, 2016 at 05:22:11PM +0100, Frederic Weisbecker wrote:
> > On Tue, Jan 19, 2016 at 02:08:57PM +0100, Peter Zijlstra wrote:
> > > On Wed, Jan 13, 2016 at 05:01:28PM +0100, Frederic Weisbecker wrote:
> > > > The cpu load update on tick doesn't care about dynticks and as such is
> > > > buggy when occuring on nohz ticks (including idle ticks) as it resets
> > > > the jiffies snapshot that was recorded on nohz entry. We eventually
> > > > ignore the potentially long tickless load that happened before the
> > > > tick.
> > > 
> > > I don't get it, how can we call scheduler_tick() while
> > > tick_nohz_tick_stopped() ?
> > 
> > tick_nohz_tick_stopped() (which is ts->tick_stopped == 1) doesn't actually
> > mean that the tick is really stopped. It just means that the tick fires only
> > when it's really needed (timer list expired, RCU stuff, irq_work, ...).
> 
> That's insane and broken. Fix _that_.
> 
> If RCU, irq_work etc.. needs the tick, do not stop the tick.

This is not the first time we have this conversation :-)

RCU/irq_work/foo_needs_tick() are treated just like any timer that expire in one
tick, although RCU is some more tunable there.

And timers that expire in 1 jiffy can be treated in two ways:

* If the tick is still periodic (ts->tick_stopped = 0), we don't stop the
  tick: we don't enter dynticks mode.

* If the tick is already stopped (or rather in dynticks mode to be more exact:
  ts->tick_stopped == 1) we just program the tick one jiffy ahead. This is
  an optimization and a simplification, if we were to restart the tick everytime
  we see a tick one jiffy ahead in the middle of a dynticks frame, we would have
  to perform all the accounting in tick_nohz_idle_exit() as well, including
  update_cpu_load_nohz() that locks rq->lock. Having a bunch of jiffies subsequently
  ticking in the middle of a dynticks frame is a common and frequent scenario and
  removing that optimization would have a bad visible impact.

Certainly the issue here is that "tick_stopped" can be misunderstood. ts->dynticks_active
would be better but we already have ts->nohz_active which reflects something very
different. I guess we need a cascading rename.

Anyway whether the next tick is one jiffy ahead or more doesn't really matter here.
The issue is that ticks can fire while dynticks-idle or dyntick-buzy and
update_cpu_load_active() treats them in a broken way.