Re: [PATCH v2] sched/pelt: fix false running accounting

[Peter Zijlstra]

On Tue, Jul 04, 2017 at 11:12:34AM +0200, Vincent Guittot wrote:
> On 4 July 2017 at 10:34, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Tue, Jul 04, 2017 at 09:27:07AM +0200, Peter Zijlstra wrote:
> >> On Sat, Jul 01, 2017 at 07:06:13AM +0200, Vincent Guittot wrote:
> >> > The running state is a subset of runnable state which means that running
> >> > can't be set if runnable (weight) is cleared. There are corner cases
> >> > where the current sched_entity has been already dequeued but cfs_rq->curr
> >> > has not been updated yet and still points to the dequeued sched_entity.
> >> > If ___update_load_avg is called at that time, weight will be 0 and running
> >> > will be set which is not possible.
> >> >
> >> > This case happens during pick_next_task_fair() when a cfs_rq becomes idles.
> >> > The current sched_entity has been dequeued so se->on_rq is cleared and
> >> > cfs_rq->weight is null. But cfs_rq->curr still points to se (it will be
> >> > cleared when picking the idle thread). Because the cfs_rq becomes idle,
> >> > idle_balance() is called and ends up to call update_blocked_averages()
> >> > with these wrong running and runnable states.
> >> >
> >> > Add a test in ___update_load_avg to correct the running state in this case.
> >>
> >> Cute, however did you find that ?
> >
> > Hmm,.. could you give a little more detail?
> >
> > Because if ->on_rq=0, we'll have done dequeue_task() which will have
> > done update_curr() with ->on_rq, weight and ->running consistently.
> >
> > Then the above, inconsistent update should not happen, because delta=0.
> 
> In fact, the delta between dequeue_entity_load_avg() and
> update_blocked_averages() is not 0 on my platform (hikey) but can be
> longer than 60us (at lowest frequency with only 1 task group level)

But but but, how can that happen? Should it not all be under the same
rq->lock and thus have only a single update_rq_clock() and thus be at
the same 'instant' ?