[Devel] [PATCH rh7 2/2] sched: Fix double put_prev_task_fair() because of trigger_cpulimit_balance()

[Vladimir Davydov]

On Thu, Aug 20, 2015 at 06:22:40PM +0300, Kirill Tkhai wrote:
> On 20.08.2015 17:55, Vladimir Davydov wrote:
> > On Thu, Aug 20, 2015 at 05:22:31PM +0300, Kirill Tkhai wrote:
> >>
> >>
> >> On 20.08.2015 15:59, Vladimir Davydov wrote:
> >>> On Thu, Aug 20, 2015 at 02:27:05PM +0300, Kirill Tkhai wrote:
> >>> ...
> >>>> +static void pre_schedule_fair(struct rq *rq, struct task_struct *prev)
> >>>> +{
> >>>> +	trigger_cpulimit_balance(prev);
> >>>> +}
> >>>> +
> >>>
> >>> You don't remove trigger_cpulimit_balance call from put_prev_task_fair.
> >>> Do we still need it there? [The same question for PCS6 version of this
> >>> patch]
> >>
> >> I simulated the same behaviour as it used to be before. If we remove the
> >> trigger from put_prev_task_fair(), there will be a later cpulimit balancing
> >> in some cases.
> > 
> > I don't see any point in it. You call pre_schedule ->
> > trigger_cpulimit_balance, then immediately after it put_prev_task ->
> > trigger_cpulimit_balance again.
> > 
> 
> ->put_prev_task() is called from several places, not only from __schedule().
> I mean that. If you think we may skip these unlikely cases, I'll remove it
> from put_prev_task_fair(). Should we?

These cases are very rare, and we are not that precise anyway. That
said, it's better remove it in order not to clutter hot schedule().