Re: [PATCH v3] sched/topology: fix the issue groups don't span domain->span for NUMA diameter > 2

[Vincent Guittot]

On Tue, 9 Feb 2021 at 12:46, Valentin Schneider
<valentin.schneider@xxxxxxx> wrote:
>
> On 09/02/21 10:46, Vincent Guittot wrote:
> > On Tue, 9 Feb 2021 at 09:27, Barry Song <song.bao.hua@xxxxxxxxxxxxx> wrote:
> >> Real servers which suffer from this problem include Kunpeng920 and 8-node
> >> Sun Fire X4600-M2, at least.
> >>
> >> Here we move to use the *child* domain of the *child* domain of node2's
> >> domain2 as the new added sched_group. At the same, we re-use the lower
> >> level sgc directly.
> >
> > Have you evaluated the impact on the imbalance and next_update fields ?
> >
>
> sgc->next_update is safe since it's only touched by CPUs that have the
> group span as local group (which is never the case for CPUs where we do
> this "grandchildren" trick).

It would be good to explain this in the commit message

>
> I'm a bit less clear about sgc->imbalance. I think it can be set by remote
> CPUs, but it should only be cleared when running load_balance() by CPUs
> that have that group span as local group, as per:
>
>   int *group_imbalance = &sd_parent->groups->sgc->imbalance;

We are also safe because sd_parent remains the same as the beg of
load_balance and LB only tries other CPUs from the local group