Re: Question about nohz and sysidle

[Paul E. McKenney]

On Mon, Feb 14, 2022 at 11:57:44AM +0100, Frederic Weisbecker wrote:
> On Mon, Feb 14, 2022 at 05:40:55PM +0800, Yu Liao wrote:
> 
> Hi Yu Liao,
> 
> > 
> > On 2022/2/14 16:28, Yu Liao wrote:
> > > Hi Frederic,
> > > 
> > > I'm working on an issue about nohz. When NO_HZ_FULL is enabled, CPU 0
> > > handles the timekeeping duty on behalf of all other CPUs, which means
> > > CPU 0 never stop tick even in sysidle state. This is a powersaving
> > > issue.
> > > 
> > > I found your patchset (nohz: Support sysidle) in the below link.
> > > https://lore.kernel.org/all/1406569056-30217-1-git-send-email-fweisbec@xxxxxxxxx/
> > > 
> > > But these patches haven't been merged into mainline yet and sysidle
> > > state detection has been removed by commit fe5ac724d81a (rcu: Remove
> > > nohz_full full-system-idle state machine) as well.
> > > 
> > > I tried your patches and it does work, why are we no longer working on
> > > stopping timekeeping duty when all full dynticks CPUs are idle?
> 
> Because it was not a priority at that time. There were so many things to handle
> first (and we are not even done yet) that we postponed that feature until
> someone ever comes up with powersaving issues on nohz_full. We were waiting for
> you :)
> 
> It's possible to unearth this. I think the first step will be to merge the
> RCU dynticks counters into context tracking, something that was on my queue
> anyway, and then revive this:
> 
> 	https://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git/commit/?h=sysidle.2017.05.11a&id=fe5ac724d81a3c7803e60c2232718f212f3f38d4

Given a valid use case, getting that functionality back would be fine
by me.  But yes, properly merging RCU dynticks into context tracking
would reduce the resulting idle-entry/exit performance penalty, so it
would be good to do that first.

							Thanx, Paul