Re: [PATCH v2] sched: fair: Prevent negative lag increase during delayed dequeue

[Vincent Guittot]

On Thu, 23 Apr 2026 at 11:41, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Apr 23, 2026 at 09:28:22AM +0200, Vincent Guittot wrote:
> > On Thu, 23 Apr 2026 at 00:20, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Wed, Apr 22, 2026 at 04:28:28PM +0200, Peter Zijlstra wrote:
> > >
> > > > Let me ponder this a bit...
> > >
> > > Like this? Or am I still making a mess of things? AFAICT this is the
> > > exact same as your initial version.
> > >
> > > ---
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 69361c63353a..24e8c78b110a 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -847,13 +847,13 @@ static s64 entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 avrunt
> > >   * Similarly, check that the entity didn't gain positive lag when DELAY_ZERO
> > >   * is set.
> > >   *
> > > - * Return true if the lag has been adjusted.
> > > + * Return true if the lag of a delayed entity has been adjusted.
> > >   */
> > >  static __always_inline
> > >  bool update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
> > >  {
> > >         s64 vlag = entity_lag(cfs_rq, se, avg_vruntime(cfs_rq));
> > > -       bool ret;
> > > +       bool ret = false;
> > >
> > >         WARN_ON_ONCE(!se->on_rq);
> > >
> > > @@ -862,8 +862,9 @@ bool update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
> > >                 vlag = max(vlag, se->vlag);
> > >                 if (sched_feat(DELAY_ZERO))
> > >                         vlag = min(vlag, 0);
> > > +
> > > +               ret = (vlag != se->vlag);
> >
> > No this is not enough.
>
> Argh yes. I think I finally see. How about this then?

Yes, this looks good.

I have also launched more tests in case I missed something.

>
> ---
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 69361c63353a..f4d1457d1837 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -847,13 +847,19 @@ static s64 entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 avrunt
>   * Similarly, check that the entity didn't gain positive lag when DELAY_ZERO
>   * is set.
>   *
> - * Return true if the lag has been adjusted.
> + * Return true if the vlag has been modified. Specifically:
> + *
> + *   se->vlag != avg_vruntime() - se->vruntime
> + *
> + * This can be due to clamping in entity_lag() or clamping due to
> + * sched_delayed. Either way, when vlag is modified and the entity is
> + * retained, the tree needs to be adjusted.
>   */
>  static __always_inline
>  bool update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
>  {
> -       s64 vlag = entity_lag(cfs_rq, se, avg_vruntime(cfs_rq));
> -       bool ret;
> +       u64 avruntime = avg_vruntime(cfs_rq);
> +       s64 vlag = entity_lag(cfs_rq, se, avruntime);
>
>         WARN_ON_ONCE(!se->on_rq);
>
> @@ -863,10 +869,9 @@ bool update_entity_lag(struct cfs_rq *cfs_rq, struct sched_entity *se)
>                 if (sched_feat(DELAY_ZERO))
>                         vlag = min(vlag, 0);
>         }
> -       ret = (vlag == se->vlag);
>         se->vlag = vlag;
>
> -       return ret;
> +       return avruntime - vlag != se->vruntime;
>  }
>
>  /*