Re: [patch 4/5]thp: correct order in lru list for split huge page

[Andrea Arcangeli <aarcange@redhat.com>]

Hi Minchan and Shaohua,

On Fri, Oct 28, 2011 at 04:21:25PM +0900, Minchan Kim wrote:
> On Fri, Oct 28, 2011 at 01:08:58PM +0800, Shaohua Li wrote:
> > On Fri, 2011-10-28 at 07:19 +0800, Minchan Kim wrote:
> > > On Tue, Oct 25, 2011 at 10:59:37AM +0800, Shaohua Li wrote:
> > > > If a huge page is split, all the subpages should live in lru list adjacently
> > > > because they should be taken as a whole.
> > > > In page split, with current code:
> > > > a. if huge page is in lru list, the order is: page, page+HPAGE_PMD_NR-1,
> > > > page + HPAGE_PMD_NR-2, ..., page + 1(in lru page reclaim order)
> > > > b. otherwise, the order is: page, ..other pages.., page + 1, page + 2, ...(in
> > > > lru page reclaim order). page + 1 ... page + HPAGE_PMD_NR - 1 are in the lru
> > > > reclaim tail.
> > > > 
> > > > In case a, the order is wrong. In case b, page is isolated (to be reclaimed),
> > > > but other tail pages will not soon.
> > > > 
> > > > With below patch:
> > > > in case a, the order is: page, page + 1, ... page + HPAGE_PMD_NR-1(in lru page
> > > > reclaim order).
> > > > in case b, the order is: page + 1, ... page + HPAGE_PMD_NR-1 (in lru page reclaim
> > > > order). The tail pages are in the lru reclaim head.
> > > > 
> > > > Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> > > 
> > > In case of a, it doesn't matter ordering of subpages.
> > > As a huge page, age of sub pages are same.
> > It does matter. Hugepage is split first and then reclaim. if page, page
> > +HPAGE_PMD_NR-1 is reclaimed, you can't get an order-1 page. but if
> > page, page+1 is reclaimed, you can.
> 
> Right you are. I didn't catch up it.
> It would be better to add it in description.
> It's most important part in this patch.

Actually the way the buddy allocator works it will compact the
hugepage identically, regardless of the order of the freeing of the
subpages. It might be slightly more efficient because of CPU cache
effects to do it in order for the buddy algorithm so we may touch one
less cacheline by finishing building an entire 1m page before jumping
to the second half, but from a practical standpoint it's irrelevant.

Case b only can materialize if the splitted page is under VM
isolation, which is a fairly uncommon case, and the page being
isolated while I splitted it looked a tricky enough that I guess I
didn't attempt to optimize it for the lru ordering and I was happy
enough it could work safe too :). Now seeing this optimization it's
strightforward, so it's certainly good idea to apply. We can apply
both but for case a it's purely theoretical and no better runtime
change is possible out of it.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>