On Tue, 8 Oct 2019 at 19:39, Peter Zijlstra <pet...@infradead.org> wrote:
>
> On Tue, Oct 08, 2019 at 05:30:02PM +0200, Vincent Guittot wrote:
>
> > This is how I plan to get ride of the problem:
> > + if (busiest->group_weight == 1 || sds->prefer_sibling) {
> > + unsigned int nr_diff = busiest->sum_h_nr_running;
> > + /*
> > + * When prefer sibling, evenly spread running tasks on
> > + * groups.
> > + */
> > + env->migration_type = migrate_task;
> > + lsub_positive(&nr_diff, local->sum_h_nr_running);
> > + env->imbalance = nr_diff >> 1;
> > + return;
> > + }
>
> I'm thinking the max_t(long, 0, ...); variant reads a lot simpler and
> really _should_ work given that -fno-strict-overflow / -fwrapv mandates
> 2s complement.
Another point that I have overlooked is that sum_h_nr_running is
unsigned int whereas imbalance is long
In fact, (long) (unsigned long A - unsigned long B) >> 1 works correctly
but
(long) (unsigned int A - unsigned int B) >> 1 doesn't
