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

[Vincent Guittot]

On Mon, 11 May 2020 at 19:02, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>
> On 11/05/2020 14:12, Vincent Guittot wrote:
> > 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
>
> Yes, I agree. But coming back to what the patch does in the example:
>
> W/ the patch, list_add_leaf_cfs_rq() is now called for cfs_b and since
> cfs_b->tg->parent->cfs_a and cfs_a->on_list=1 the 'branch is now
> connected' which means 'rq->tmp_alone_branch = &rq->leaf_cfs_rq_list'.
>
> I.e. assert_list_leaf_cfs_rq() at the end of enqueue_task_fair() is not
> barfing anymore.
>
> W/o the patch, list_add_leaf_cfs_rq() called w/ cfs_c left the 'branch
> open', it's not called on cfs_b and since cfs_a->on_list=1, the 3rd
> for_each_sched_entity() in enqueue_task_fair() doesn't 'connect the
> branch' so the assert fires.
>
> What I don't immediately see is how can cfs_a be throttled (which causes
> cfs_b -> cfs_c being a throttled hierarchy) and be on the list
> (cfs_a->on_list=1) at the same time.
>
> So the only thing how this could happen is when there was a task enqueue
> in a parallel cfs_b' (another child of cfs_a) sub hierarchy just before
> the example.

Yes. A task has been enqueued on another child (cfs_b') and cfs_a has
been be added back to ensure that cfs are correctly ordered

>
> >> 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.
>
> True, we go the 2nd if() condition in list_add_leaf_cfs_rq().
>
> >>> 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
>
> Yes.
>
> [...]