Re: [RFC/RFT][PATCH 6/7] sched: idle: Predict idle duration before stopping the tick

[Rafael J. Wysocki]

On Mon, Mar 5, 2018 at 1:42 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Mon, Mar 05, 2018 at 01:07:07PM +0100, Rafael J. Wysocki wrote:
>> On Mon, Mar 5, 2018 at 12:50 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>> > On Mon, Mar 5, 2018 at 12:45 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
>> >> So I think this is entirely wrong, I would much rather see something
>> >> like:
>> >>
>> >>         tick_nohz_idle_go_idle(next_state->nohz);
>> >>
>> >> Where the selected state itself has the nohz property or not.
>> >
>> > Can you elaborate here, I'm not following?
>> >
>> >> We can always insert an extra state at whatever the right boundary point
>> >> is for nohz if it doesn't line up with an existing point.
>>
>> OK, I guess I know what you mean: to add a state flag meaning "stop
>> the tick if this state is selected".
>
> Yes, that.
>
>> That could work, but I see problems, like having to go through all of
>> the already defined states and deciding what to do with them.
>
> Shouldn't be too hard, upon registering a cpuidle driver to the cpuidle
> core, the core could go through the provided states and flag all those <
> TICK_USEC as not stopping, all those > TICK_USEC as stopping and
> splitting the state we'd select for TICK_NSEC sleeps, stopping it for <
> and disabling it for >.
>

Well, on Intel everything below C8 has target residencies below 1 ms. :-)

I think that we want C6 to be "nohz" too, though, at least in some cases.

And what about C3 if C6 is disabled?