Re: [RFC][PATCH 1/3] cpuidle: Inject tick boundary state

[Rafael J. Wysocki]

On Sat, Jul 29, 2023 at 10:44 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Fri, Jul 28, 2023 at 05:36:55PM +0200, Rafael J. Wysocki wrote:
> > On Fri, Jul 28, 2023 at 5:01 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > In order to facilitate governors that track history in idle-state
> > > buckets (TEO) making a useful decision about NOHZ, make sure we have a
> > > bucket that counts tick-and-longer.
> > >
> > > In order to be inclusive of the tick itself -- after all, if we do not
> > > disable NOHZ we'll sleep for a full tick, the actual boundary should
> > > be just short of a full tick.
> > >
> > > IOW, when registering the idle-states, add one that is always
> > > disabled, just to have a bucket.
> >
> > This extra bucket can be created in the governor itself, can't it?
>
> I couldn't find a nice spot for the governor to add idle-states.

Well, I've thought this through and recalled a couple of things and my
conclusion is that the decision whether or not to stop the tick really
depends on the idle state choice.

There are three cases:

1. The selected idle state is shallow (that is, its target residency
is below the tick period length), but it is not the deepest one.
2. The selected idle state is shallow, but it is the deepest one (or
at least the deepest enabled one).
3. The selected idle state is deep (that is, its target residency is
above the tick length period).

In case 1, the tick should not be stopped so as to prevent the CPU
from spending too much time in a suboptimal idle state.

In case 3, the tick needs to be stopped, because otherwise the target
residency of the selected state would not be met.

Case 2 is somewhat a mixed bag, but generally speaking stopping the
tick doesn't hurt if the selected idle state is the deepest one,
because in that case the governor kind of expects a significant exit
latency anyway.  If it is not the deepest one (which is disabled),
it's better to let the tick tick.