On Mon, Aug 06, 2018 at 05:25:28PM +0200, Peter Zijlstra wrote:
> On Mon, Aug 06, 2018 at 11:05:50AM -0400, Johannes Weiner wrote:
> > Argh, that's right. This needs an explicit count if we want to access
> > it locklessly. And you already said you didn't like that this is the
> > only state not derived purely from the task counters, so maybe this is
> > the way to go after all.
> >
> > How about something like this (untested)?
>
>
> > +static inline void psi_switch(struct rq *rq, struct task_struct *prev,
> > + struct task_struct *next)
> > +{
> > + if (psi_disabled)
> > + return;
> > +
> > + if (unlikely(prev->flags & PF_MEMSTALL))
> > + psi_task_change(prev, rq_clock(rq), TSK_RECLAIMING, 0);
> > + if (unlikely(next->flags & PF_MEMSTALL))
> > + psi_task_change(next, rq_clock(rq), 0, TSK_RECLAIMING);
> > +}
>
>
> Urgh... can't say I really like that.
>
> I would really rather do that scheduler_tick() thing to avoid the remote
> update. The tick is a lot less hot than the switch path and esp.
> next->flags might be a cold line (prev->flags is typically the same line
> as prev->state so we already have that, but I don't think anybody now
> looks at next->flags or its line, so that'd be cold load).
Okay, the tick updater sounds like a much better option then. HZ
frequency should produce more than recent enough data.
That means we will retain the not-so-nice PF_MEMSTALL flag test under
rq lock, but it'll eliminate most of that memory ordering headache.
I'll do that. Thanks!
