* Peter Zijlstra <pet...@infradead.org> [2018-06-04 21:28:21]:
> > if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) {
> > - spin_lock(&pgdat->numabalancing_migrate_lock);
> > - pgdat->numabalancing_migrate_nr_pages = 0;
> > - pgdat->numabalancing_migrate_next_window = jiffies +
> > - msecs_to_jiffies(migrate_interval_millisecs);
> > - spin_unlock(&pgdat->numabalancing_migrate_lock);
> > + if (xchg(&pgdat->numabalancing_migrate_nr_pages, 0))
> > + pgdat->numabalancing_migrate_next_window = jiffies +
> > + msecs_to_jiffies(migrate_interval_millisecs);
>
> Note that both are in fact wrong. That wants to be something like:
>
> pgdat->numabalancing_migrate_next_window += interval;
>
> Otherwise you stretch every interval by 'jiffies -
> numabalancing_migrate_next_window'.
Okay, I get your point.
>
> Also, that all wants READ_ONCE/WRITE_ONCE, irrespective of the
> spinlock/xchg.
>
> I suppose the problem here is that PPC has a very nasty test-and-set
> spinlock with fwd progress issues while xchg maps to a fairly simple
> ll/sc that (hopefully) has some hardware fairness.
>
> And pgdata being a rather course data structure (per node?) there could
> be a lot of CPUs stomping on this here thing.
>
> So simpler not really, but better for PPC.
>
unsigned long interval = READ_ONCE(pgdat->numabalancing_migrate_next_window);
if (time_after(jiffies, interval)) {
interval += msecs_to_jiffies(migrate_interval_millisecs));
if (xchg(&pgdat->numabalancing_migrate_nr_pages, 0))
WRITE_ONCE(pgdat->numabalancing_migrate_next_window, interval);
}
Something like this?
