Re: [PATCH] mm: compaction: Abort compaction if too many pages areisolated and caller is asynchronous

[Mel Gorman]

On Fri, Jun 03, 2011 at 08:01:44AM +0900, Minchan Kim wrote:
> On Fri, Jun 3, 2011 at 7:32 AM, Andrea Arcangeli <aarcange@xxxxxxxxxx> wrote:
> > On Fri, Jun 03, 2011 at 07:23:48AM +0900, Minchan Kim wrote:
> >> I mean we have more tail pages than head pages. So I think we are likely to
> >> meet tail pages. Of course, compared to all pages(page cache, anon and
> >> so on), compound pages would be very small percentage.
> >
> > Yes that's my point, that being a small percentage it's no big deal to
> > break the loop early.
> 
> Indeed.
> 
> >
> >> > isolated the head and it's useless to insist on more tail pages (at
> >> > least for large page size like on x86). Plus we've compaction so
> >>
> >> I can't understand your point. Could you elaborate it?
> >
> > What I meant is that if we already isolated the head page of the THP,
> > we don't need to try to free the tail pages and breaking the loop
> > early, will still give us a chance to free a whole 2m because we
> > isolated the head page (it'll involve some work and swapping but if it
> > was a compoundtranspage we're ok to break the loop and we're not
> > making the logic any worse). Provided the PMD_SIZE is quite large like
> > 2/4m...
> 
> Do you want this? (it's almost pseudo-code)
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 7a4469b..9d7609f 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1017,7 +1017,7 @@ static unsigned long isolate_lru_pages(unsigned
> long nr_to_scan,
>         for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
>                 struct page *page;
>                 unsigned long pfn;
> -               unsigned long end_pfn;
> +               unsigned long start_pfn, end_pfn;
>                 unsigned long page_pfn;
>                 int zone_id;
> 
> @@ -1057,9 +1057,9 @@ static unsigned long isolate_lru_pages(unsigned
> long nr_to_scan,
>                  */
>                 zone_id = page_zone_id(page);
>                 page_pfn = page_to_pfn(page);
> -               pfn = page_pfn & ~((1 << order) - 1);
> +               start_pfn = pfn = page_pfn & ~((1 << order) - 1);
>                 end_pfn = pfn + (1 << order);
> -               for (; pfn < end_pfn; pfn++) {
> +               while (pfn < end_pfn) {
>                         struct page *cursor_page;
> 
>                         /* The target page is in the block, ignore it. */
> @@ -1086,17 +1086,25 @@ static unsigned long
> isolate_lru_pages(unsigned long nr_to_scan,
>                                 break;
> 
>                         if (__isolate_lru_page(cursor_page, mode, file) == 0) {
> +                               int isolated_pages;
>                                 list_move(&cursor_page->lru, dst);
>                                 mem_cgroup_del_lru(cursor_page);
> -                               nr_taken += hpage_nr_pages(page);
> +                               isolated_pages = hpage_nr_pages(page);
> +                               nr_taken += isolated_pages;
> +                               /* if we isolated pages enough, let's
> break early */
> +                               if (nr_taken > end_pfn - start_pfn)
> +                                       break;
> +                               pfn += isolated_pages;

I think this condition is somewhat unlikely. We are scanning within
aligned blocks in this linear scanner. Huge pages are always aligned
so the only situation where we'll encounter a hugepage in the middle
of this linear scan is when the requested order is larger than a huge
page. This is exceptionally rare.

Did I miss something?

>                                 nr_lumpy_taken++;
>                                 if (PageDirty(cursor_page))
>                                         nr_lumpy_dirty++;
>                                 scan++;
>                         } else {
>                                 /* the page is freed already. */
> -                               if (!page_count(cursor_page))
> +                               if (!page_count(cursor_page)) {
> +                                       pfn++;
>                                         continue;
> +                               }
>                                 break;
>                         }
>                 }
> 
> >
> > The only way this patch makes things worse is for slub order 3 in the
> > process of being freed. But tail pages aren't generally free anyway so
> > I doubt this really makes any difference plus the tail is getting
> > cleared as soon as the page reaches the buddy so it's probably
> 
> Okay. Considering getting clear PG_tail as soon as slub order 3 is
> freed, it would be very rare case.
> 
> > unnoticeable as this then makes a difference only during a race (plus
> > the tail page can't be isolated, only head page can be part of lrus
> > and only if they're THP).
> >
> >> > insisting and screwing lru ordering isn't worth it, better to be
> >> > permissive and abort... in fact I wouldn't dislike to remove the
> >> > entire lumpy logic when COMPACTION_BUILD is true, but that alters the
> >> > trace too...
> >>
> >> AFAIK, it's final destination to go as compaction will not break lru
> >> ordering if my patch(inorder-putback) is merged.
> >
> > Agreed. I like your patchset, sorry for not having reviewed it in
> > detail yet but there were other issues popping up in the last few
> > days.
> 
> No problem. it's urgent than mine. :)
> 

I'm going to take the opportunity to apologise for not reviewing that
series yet. I've been kept too busy with other bugs to set side the
few hours I need to review the series. I'm hoping to get to it this
week if everything goes well.

-- 
Mel Gorman
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/