Re: vmstat: On demand vmstat workers V4

[Frederic Weisbecker]

On Sat, May 10, 2014 at 06:17:08PM -0700, Paul E. McKenney wrote:
> On Sat, May 10, 2014 at 03:14:25PM +0200, Frederic Weisbecker wrote:
> > On Sat, May 10, 2014 at 02:31:28PM +0200, Thomas Gleixner wrote:
> > > On Sat, 10 May 2014, Frederic Weisbecker wrote:
> > > > But I still have the plan to make the timekeeper use the full sysidle
> > > > facility in order to adaptively get to dynticks idle.
> > > > 
> > > > Reminder for others: in NO_HZ_FULL, the timekeeper (always CPU 0) stays
> > > > completely periodic. It can't enter in dynticks idle mode because it
> > > > must maintain timekeeping on behalf of full dynticks CPUs. So that's
> > > > a power issue.
> > > > 
> > > > But Paul has a feature in RCU that lets us know when all CPUs are idle
> > > > and the timekeeper can finally sleep. Then when a full nohz CPU wakes
> > > > up from idle, it sends an IPI to the timekeeper if needed so the latter
> > > > restarts timekeeping maintainance.
> > > >
> > > > It's not complicated to add to the timer code.
> > > > Most of the code is already there, in RCU, for a while already.
> > > > 
> > > > Are we keeping that direction? 
> > > 
> > > So the idea is that the timekeeper stays on cpu0, but if everything is
> > > idle it is allowed to take a long nap as well. So if some other cpu
> > > wakes up it updates timekeeping without taking over the time keeper
> > > duty and if it has work to do, it kicks cpu0 into gear. If it just
> > > goes back to sleep, then nothing to do.
> 
> Hmmm...  If RCU is supposed to ignore the fact that one of the other
> CPUs woke up momentarily, we will need to adjust things a bit.

Maybe not that much actually.

> 
> > Exactly! Except perhaps the last sentence "If it just goes back to sleep,
> > then nothing to do.", I didn't think about that although this special case
> > is quite frequent indeed when an interrupt fires on idle but no task is woken up.
> > 
> > Maybe I should move the code that fires the IPI to cpu0, if it is sleeping,
> > on irq exit (the plan was to do it right away on irq enter) and fire it
> > only if need_resched().
> 
> And of course if that code path contains any RCU read-side critical
> sections, RCU absolutely cannot ignore that CPU's momentary wakeup.

Sure the core RCU still needs to know that the CPU went out of dynticks the
time of the irq, so we keep the rcu_irq_enter/rcu_irq_exit calls.

But if the CPU only wakes up to serve an IRQ, it doesn't need to tell the RCU
sysidle detection about it. The irq entry fixup jiffies on dynticks idle mode,
this should be enough.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/