Re: [PATCH v2] sched/fair: Fix enqueue_task_fair warning some more

[Vincent Guittot]

Hi Tao,

On Fri, 8 May 2020 at 18:58, Tao Zhou <zohooouoto@xxxxxxxxxxx> wrote:
>
> On Fri, May 08, 2020 at 05:27:44PM +0200, Vincent Guittot wrote:
> > On Fri, 8 May 2020 at 17:12, Tao Zhou <zohooouoto@xxxxxxxxxxx> wrote:
> > >
> > > Hi Phil,
> > >
> > > On Thu, May 07, 2020 at 04:36:12PM -0400, Phil Auld wrote:
> > > > sched/fair: Fix enqueue_task_fair warning some more
> > > >
> > > > The recent patch, fe61468b2cb (sched/fair: Fix enqueue_task_fair warning)
> > > > did not fully resolve the issues with the rq->tmp_alone_branch !=
> > > > &rq->leaf_cfs_rq_list warning in enqueue_task_fair. There is a case where
> > > > the first for_each_sched_entity loop exits due to on_rq, having incompletely
> > > > updated the list.  In this case the second for_each_sched_entity loop can
> > > > further modify se. The later code to fix up the list management fails to do
> > > > what is needed because se no longer points to the sched_entity which broke
> > > > out of the first loop.
> > > >
> > >
> > > > Address this by calling leaf_add_rq_list if there are throttled parents while
> > > > doing the second for_each_sched_entity loop.
> > >
> > > Thanks for your trace imformation and explanation. I
> > > truely have learned from this and that.
> > >
> > > s/leaf_add_rq_list/list_add_leaf_cfs_rq/
> > >
> > > >
> > > > Suggested-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > > > Signed-off-by: Phil Auld <pauld@xxxxxxxxxx>
> > > > Cc: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > > > Cc: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > > > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > > > Cc: Juri Lelli <juri.lelli@xxxxxxxxxx>
> > > > ---
> > > >  kernel/sched/fair.c | 7 +++++++
> > > >  1 file changed, 7 insertions(+)
> > > >
> > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > index 02f323b85b6d..c6d57c334d51 100644
> > > > --- a/kernel/sched/fair.c
> > > > +++ b/kernel/sched/fair.c
> > > > @@ -5479,6 +5479,13 @@ enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
> > > >               /* end evaluation on encountering a throttled cfs_rq */
> > > >               if (cfs_rq_throttled(cfs_rq))
> > > >                       goto enqueue_throttle;
> > > > +
> > > > +               /*
> > > > +                * One parent has been throttled and cfs_rq removed from the
> > > > +                * list. Add it back to not break the leaf list.
> > > > +                */
> > > > +               if (throttled_hierarchy(cfs_rq))
> > > > +                       list_add_leaf_cfs_rq(cfs_rq);
> > > >       }
> > >
> > > I was confused by why the throttled cfs rq can be on list.
> > > It is possible when enqueue a task and thanks to the 'threads'.
> > > But I think the above comment does not truely put the right
> > > intention, right ?
> > > If throttled parent is onlist, the child cfs_rq is ignored
> > > to be added to the leaf cfs_rq list me think.
> > >
> > > unthrottle_cfs_rq() follows the same logic if i am not wrong.
> > > Is it necessary to add the above to it ?
> >
> > When a cfs_rq is throttled, its sched group is dequeued and all child
> > cfs_rq are removed from  leaf_cfs_rq list. But the sched group of the
> > child cfs_rq stay enqueued in the throttled cfs_rq so child sched
> > group->on_rq might be still set.
>
> If there is a throttle of throttle, and unthrottle the child throttled
> cfs_rq(ugly):
>                                ...
>                                 |
>                       cfs_rq throttled (parent A)
>                                 |
>                                 |
>                       cfs_rq in hierarchy (B)
>                                 |
>                                 |
>                       cfs_rq throttled (C)
>                                 |
>                                ...
>
> Then unthrottle the child throttled cfs_rq C, now the A is on the
> leaf_cfs_rq list. sched_group entity of C is enqueued to B, and
> sched_group entity of B is on_rq and is ignored by enqueue but in
> the throttled hierarchy and not add to leaf_cfs_rq list.
> The above may be absolutely wrong that I miss something.

several things:

your example above is safe IMO because when C is unthrottle, It's
group se will be enqueued on B which will be added to leaf_cfs_rq
list.
Then the group se of B is already on_rq but A is throttled and the 1st
loop break.  The 2nd loop will ensure that A is added to leaf_cfs_rq
list

Now, if we add one more level between C and A, we have a problem and
we should add something similar in the else

Finally, while checking the unthrottle_cfs_rq, the test if
(!cfs_rq->load.weight) return"  skips all the for_each_entity loop and
can break the leaf_cfs_rq

We need to jump to the last loop in such case

>
> Another thing :
> In enqueue_task_fair():
>
>         for_each_sched_entity(se) {
>                 cfs_rq = cfs_rq_of(se);
>
>                 if (list_add_leaf_cfs_rq(cfs_rq))
>                         break;
>         }
>
> In unthrottle_cfs_rq():
>
>         for_each_sched_entity(se) {
>                 cfs_rq = cfs_rq_of(se);
>
>                 list_add_leaf_cfs_rq(cfs_rq);
>         }
>
> The difference between them is that if condition, add if
> condition to unthrottle_cfs_rq() may be an optimization and
> keep the same.

Yes we can do the same kind of optimization

>
> > >
> > > Thanks,
> > > Tau
> > >
> > > >
> > > >  enqueue_throttle:
> > > > --
> > > > 2.18.0
> > > >
> > > > V2 rework the fix based on Vincent's suggestion. Thanks Vincent.
> > > >
> > > >
> > > > Cheers,
> > > > Phil
> > > >
> > > > --
> > > >