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

From: Vincent Guittot
Date: Mon May 11 2020 - 08:12:51 EST


On Mon, 11 May 2020 at 12:39, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>
> On 11/05/2020 11:36, Vincent Guittot wrote:
> > On Mon, 11 May 2020 at 10:40, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
> >>
> >> On 08/05/2020 19:02, Tao Zhou 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
>
> [...]
>
> >> I'm not 100% sure if this is exactly what Tao pointed out here but I
> >> also had difficulties understanding understanding how this patch works:
> >>
> >> p.se
> >> |
> >> __________________|
> >> |
> >> V
> >> cfs_c -> tg_c -> se_c (se->on_rq = 1)
> >> |
> >> __________________|
> >> |
> >> v
> >> cfs_b -> tg_b -> se_b
> >> |
> >> __________________|
> >> |
> >> V
> >> cfs_a -> tg_a -> se_a
> >> |
> >> __________________|
> >> |
> >> V
> >> cfs_r -> tg_r
> >> |
> >> V
> >> rq
> >>
> >
> > In your example, which cfs_ rq has been throttled ? cfs_a ?
>
> Yes, cfs_a. 0xffffa085e48ce000 in Phil's trace.
>
> >
> >> (1) The incomplete update happens with cfs_c at the end of
> >> enqueue_entity() in the first loop because of 'if ( .... ||
> >> cfs_bandwidth_used())' (cfs_b->on_list=0 since cfs_a is throttled)
> >
> > so cfs_c is added with the 1st loop
>
> Yes.
>
> >> (2) se_c breaks out of the first loop (se_c->on_rq = 1)
> >>
> >> (3) With the patch cfs_b is added back to the list.
> >> But only because cfs_a->on_list=1.
> >
> > hmm I don't understand the link between cfs_b been added and cfs_a->on_list=1
>
> cfs_b, 0xffffa085e48ce000 is the one which is now added in the 2. loop.
>
> Isn't the link between cfs_b and cfs_a the first if condition in

on_list is only there to say if the cfs_rq is already in the list but
there is not dependency with the child

> list_add_leaf_cfs_rq():
>
> if (cfs_rq->tg->parent &&
> cfs_rq->tg->parent->cfs_rq[cpu]->on_list)
>
> to 'connect the branch' or not (default, returning false case)?
>

In your example above if the parent is already on the list then we
know where to insert the child.

> > cfs_b is added with 2nd loop because its throttle_count > 0 due to
> > cfs_a been throttled (purpose of this patch)
> >
> >>
> >> But since cfs_a is throttled it should be cfs_a->on_list=0 as well.
> >
> > So 2nd loop breaks because cfs_a is throttled
>
> Yes.
>
> > The 3rd loop will add cfs_a
>
> Yes, but in the example, cfs_a->on_list=1, so we bail out of
> list_add_leaf_cfs_rq() early.

Because the cfs_rq is on the list already so we don't have to add it

>
> I don't grasp how can cfs_a->on_list=1, when cfs_a is throttled and
> cfs_b, cfs_c are in a throttled hierarchy?
>
> >> throttle_cfs_rq()->walk_tg_tree_from(..., tg_throttle_down, ...) should
> >> include cfs_a when calling list_del_leaf_cfs_rq().
> >>
> >> IMHO, throttle_cfs_rq() calls tg_throttle_down() for the throttled
> >> cfs_rq too.
> >>
> >>
> >> Another thing: Why don't we use throttled_hierarchy(cfs_rq) instead of
> >> cfs_bandwidth_used() in enqueue_entity() as well?
> >
> > Mainly to be conservative because as this patch demonstrates, there
> > are a lot of possible use cases and combinations and I can't ensure
> > that it is always safe to use the throttled_hierarchy.
>
> Maybe this deserves a comment then.