On Mon, Feb 12, 2018 at 05:47:29PM +0000, Mel Gorman wrote:
> On Mon, Feb 12, 2018 at 06:31:47PM +0100, Peter Zijlstra wrote:
> > On Mon, Feb 12, 2018 at 02:58:57PM +0000, Mel Gorman wrote:
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 28c8d9c91955..50442697b455 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -5710,8 +5710,14 @@ wake_affine_idle(int this_cpu, int prev_cpu, int 
> > > sync)
> > >   if (idle_cpu(this_cpu) && cpus_share_cache(this_cpu, prev_cpu))
> > >           return idle_cpu(prev_cpu) ? prev_cpu : this_cpu;
> > >  
> > > - if (sync && cpu_rq(this_cpu)->nr_running == 1)
> > > + if (sync && cpu_rq(this_cpu)->nr_running == 1) {
> > > +         /* Avoid tasks exiting pulling parents to new nodes */
> > > +         if ((current->flags & PF_EXITING) &&
> > > +             !cpus_share_cache(this_cpu, prev_cpu))
> > > +                 return prev_cpu;
> > > +
> > 
> > Cute, but should we not kill @sync right at the source in this case?
> > 
> > Something a little like this?
> > 
> 
> That works too but note that it has a slightly disadvantage in that we
> add weight to the select_task_rq_fair fast path for all wakeups, not just
> those that are affine. It's also not 100% functionally equivalent as I was
> still allowing parents to migrate to the CPU if it shared cache although I
> expect any advantage there is marginal. I can replace my patch with yours
> if you really prefer it as the overhead in the fast path should be marginal.
> 

Clearly my brains leaked out my ears at the end of a work day. The overhead
only applies to sync wakeups and is a very small amount of overhead in the
case where this_cpu or prev_cpu was idle so I'm happy with either version.

-- 
Mel Gorman
SUSE Labs

Reply via email to