Re: [RFC/RFT][PATCH 4/7] cpuidle: menu: Split idle duration prediction from state selection

[Rafael J. Wysocki]

On Mon, Mar 5, 2018 at 1:50 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Mon, Mar 05, 2018 at 12:47:23PM +0100, Rafael J. Wysocki wrote:
>> On Mon, Mar 5, 2018 at 12:38 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > We really should be predicting state not duration. Yes the duration
>> > thing is an intermediate value, but I don't think it makes any sense
>> > what so ever to preserve that in the predictor. The end result is the
>> > idle state, we should aim for that.
>> >
>> > As per:
>> >
>> >   https://lkml.org/lkml/2017/7/18/615
>> >
>> > there are definite advantages to _not_ preserving duration information
>> > beyond the state boundaries.
>>
>> Well, OK
>>
>> The reason why I need the predicted idle duration is because the
>> target residency of the selected state may be below the tick period
>> duration and if this is the deepest state available, we still want to
>> stop the tick if the predicted idle duration is long.
>
> Right, so in that case we'd split the deepest state and mark the
> resulting smaller state as not disabling the tick and the resulting
> larger state as disabling the tick.
>
> So suppose your deepest state is < TICK_USEC, then we introduce a copy
> of that state, modify the boundary to be TICK_USEC and set the 'disable
> tick for this state' thing to true.
>
>> IOW, the target residency of the selected state doesn't tell you how
>> much time you should expect to be idle in general.
>
> Right, but I think that measure isn't of primary relevance. What we want
> to know is: 'should I stop the tick' and 'what C state do I go to'.
>
> In order to answer those questions we need durations as input, but I
> don't think we should preserve durations throughout. The scheme from the
> above link reduces to N states in order to deal with arbitrary
> distributions, only the actual states -- ie boundaries where our answers
> changes -- are relevant, anything inside those boundaries would lead to
> the exact same answer anyway.

I generally agree here, but I'm not convinced about flagging the
states, splitting them and so on.

Maybe just return a "nohz" indicator from cpuidle_select() in addition
to the state index and make the decision in the governor?