Re: [PATCH] sched/core: fix pick_next_task 'max' tracking

[Tao Zhou]

Hi,

On Wed, Aug 18, 2021 at 12:35:13PM +0800, Tao Zhou wrote:
> Hi Josh,
> 
> On Tue, Aug 17, 2021 at 05:56:15PM -0700, Josh Don wrote:
> > For core-sched, pick_next_task will update the 'max' task if there is a
> > cookie mismatch (since in this case the new task must be of higher
> > priority than the current max). However, we fail to update 'max' if
> > we've found a task with a matching cookie and higher priority than
> > 'max'.
> > 
> > This can result in extra iterations on SMT-X machines, where X > 2.
> > 
> > As an example, on a machine with SMT=3, on core 0, SMT-0 might pick
> > the following, in order:
> > 
> > - SMT-0: p1, with cookie A, and priority 10 (max = p1)
> > - SMT-1: p2, with cookie A, and priority 30 (max not updated here)
> 
> Thanks for your illustration. Good catch.
> The guilty is 'cookie_equals(class_pick, cookie))' condition in pick_task()
> 
> > - SMT-2: p3, with cookie B, and priority 20 (max = p2)
> > 	> invalidate the other picks and retry
> > 
> > Here, we should have instead updated 'max' when picking for SMT-1. Note
> > that this code would eventually have righted itself, since the retry
> > loop would re-pick p2, and update 'max' accordingly. However, this patch
> > avoids the extra round-trip.
> 
> This is correct then may increase the chance to retry. That means it is
> more possible to filter the max first(not sure).
> 
> > Signed-off-by: Josh Don <joshdon@xxxxxxxxxx>
> > ---
> >  kernel/sched/core.c | 2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index 3431939699dc..110ea7582a33 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -5623,6 +5623,8 @@ pick_next_task(struct rq *rq, struct task_struct *prev, struct rq_flags *rf)
> >  					occ = 1;
> >  					goto again;
> >  				}
> > +			} else if (prio_less(max, p, fi_before)) {
> > +				max = p;
> >  			}
> >  		}
> >  	}
> > -- 
> > 2.33.0.rc1.237.g0d66db33f3-goog
> > 

My ugly patch need to continue, not compiled.
Rebase on the previous patch.