On Fri 07-12-18 10:54:50, Vlastimil Babka wrote:
> On 12/5/18 8:12 PM, Michal Hocko wrote:
> > [Cc Vlastimil]
> >
> > On Tue 04-12-18 13:43:31, osalvador@xxxxxxx wrote:
> >> On 2018-12-04 12:31, David Hildenbrand wrote:
> >> > If I am not wrong, zone_contiguous is a pure mean for performance
> >>> improvement, right? So leaving zone_contiguous unset is always save. I
> >>> always disliked the whole clear/set_zone_contiguous thingy. I wonder if
> >>> we can find a different way to boost performance there (in the general
> >>> case). Or is this (zone_contiguous) even worth keeping around at all for
> >>> now? (do we have performance numbers?)
> >>
> >> It looks like it was introduced by 7cf91a98e607
> >> ("mm/compaction: speed up pageblock_pfn_to_page() when zone is contiguous").
> >>
> >> The improve numbers are in the commit.
> >> So I would say that we need to keep it around.
> >
> > Is that still the case though?
>
> Well, __pageblock_pfn_to_page() has to be called for each pageblock in
> compaction, when zone_contiguous is false. And that's unchanged since
> the introduction of zone_contiguous, so the numbers should still hold.
OK, this means that we have to carefully re-evaluate zone_contiguous for
each offline operation.
Thanks!
--
Michal Hocko
SUSE Labs
