Re: [PATCH] sched/fair: vruntime should normalize when switching from fair

[Dietmar Eggemann]

On 08/28/2018 03:53 PM, Dietmar Eggemann wrote:
> On 08/27/2018 12:14 PM, Peter Zijlstra wrote:
> > On Fri, Aug 24, 2018 at 02:24:48PM -0700, Steve Muckle wrote:
> > > On 08/24/2018 02:47 AM, Peter Zijlstra wrote:
> > > > > > On 08/17/2018 11:27 AM, Steve Muckle wrote:
> > > >
> > > > > > > When rt_mutex_setprio changes a task's scheduling class to RT,
> > > > > > > we're seeing cases where the task's vruntime is not updated
> > > > > > > correctly upon return to the fair class.
> > > >
> > > > > > > Specifically, the following is being observed:
> > > > > > > - task is deactivated while still in the fair class
> > > > > > > - task is boosted to RT via rt_mutex_setprio, which changes
> > > > > > >       the task to RT and calls check_class_changed.
> > > > > > > - check_class_changed leads to detach_task_cfs_rq, at which point
> > > > > > >       the vruntime_normalized check sees that the task's state is TASK_WAKING,
> > > > > > >       which results in skipping the subtraction of the rq's min_vruntime
> > > > > > >       from the task's vruntime
> > > > > > > - later, when the prio is deboosted and the task is moved back
> > > > > > >       to the fair class, the fair rq's min_vruntime is added to
> > > > > > >       the task's vruntime, even though it wasn't subtracted earlier.
> > > >
> > > > I'm thinking that is an incomplete scenario; where do we get to
> > > > TASK_WAKING.
> > >
> > > Yes there's a missing bit of context here at the beginning that the task to
> > > be boosted had already been put into TASK_WAKING.
> >
> > See, I'm confused...
> >
> > The only time TASK_WAKING is visible, is if we've done a remote wakeup
> > and it's 'stuck' on the remote wake_list. And in that case we've done
> > migrate_task_rq_fair() on it.
> >
> > So by the time either rt_mutex_setprio() or __sched_setscheduler() get
> > to calling check_class_changed(), under both pi_lock and rq->lock, the
> > vruntime_normalized() thing should be right.
> >
> > So please detail the exact scenario. Because I'm not seeing it.
>
> Using Steve's test program (https://lkml.org/lkml/2018/8/24/686) I see the
> issue but only if the two tasks (rt_task, fair_task) run on 2 cpus which
> don't share LLC (e.g. CPU0 and CPU4 on hikey960).
>
> So the wakeup goes the TTWU_QUEUE && !share_cache (ttwu_queue_remote) path.

I forgot to mention that since fair_task's cpu affinity is restricted to 
CPU4, there is no call to set_task_cpu()->migrate_task_rq_fair() since 
if (task_cpu(p) != cpu) fails.

I think the combination of cpu affinity of the fair_task to CPU4 and the 
fact that the scheduler runs on CPU1 when waking fair_task (with the two 
cpus not sharing LLC) while TTWU_QUEUE is enabled is the situation in 
which this vruntime issue can happen.

[...]