Re: [PATCH 2/2] sched: update runqueue clock before migrations away

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Dec 10, 2013 at 03:55:43PM +0000, Chris Redpath wrote:
> >That's guestimating the last_runnable_update based on decay_count, and
> >per the previous the decay count can get slightly out of sync.
> 
> The guesstimation works fine, the issue is only that we can't tell at
> this point how much time that entity was asleep when the CPU it ran on
> has no tick and since it is too expensive to synchronize the clocks,
> there isn't (currently) a way to find out without updating the rq we
> came from.
> 
> I can't see anything handy lying around in struct rq, but is there a
> copy of the jiffies held anywhere per-cpu presuming it would stop being
> updated when the tick stops? If not, I could store one somewhere as part
> of turning the tick off and then we could use the difference between
> that and the current jiffies count to estimate the amount of time in
> limbo. That would almost certainly be accurate enough for me - a few ms
> won't hurt but when we lose seconds it does.

Would pre_schedule_idle() -> rq_last_tick_reset() -> rq->last_sched_tick
be useful?

I suppose we could easily lift that to NO_HZ_COMMON.

Which raises another point; I dislike this idle pre/post business. Why
can't the post_schedule_idle() calls be done from pick_next_task_idle()
and pre_schedule_idle() done from put_prev_task_idle() ?

That would avoid the post_schedule() rq->lock dance.

Vince?