Re: [PATCH] hugetlb: Do not account hugetlb pages as NR_FILE_PAGES

[Michal Hocko]

On Tue 02-06-15 11:33:05, Vlastimil Babka wrote:
> On 06/02/2015 11:25 AM, Michal Hocko wrote:
> >On Mon 25-05-15 17:24:28, Vlastimil Babka wrote:
> >>On 05/22/2015 04:35 PM, Mel Gorman wrote:
> >>>>
> >>>>Thanks!
> >>>>
> >>>>>This makes a lot of sense to me.  The only thing I worry about is the
> >>>>>proliferation of PageHuge(), a function call, in relatively hot paths.
> >>>>
> >>>>I've tried that (see the patch below) but it enlarged the code by almost
> >>>>1k
> >>>>    text    data     bss     dec     hex filename
> >>>>  510323   74273   44440  629036   9992c mm/built-in.o.before
> >>>>  511248   74273   44440  629961   99cc9 mm/built-in.o.after
> >>>>
> >>>>I am not sure the code size increase is worth it. Maybe we can reduce
> >>>>the check to only PageCompound(page) as huge pages are no in the page
> >>>>cache (yet).
> >>>>
> >>>
> >>>That would be a more sensible route because it also avoids exposing the
> >>>hugetlbfs destructor unnecessarily.
> >>
> >>You could maybe do test such as (PageCompound(page) && PageHuge(page)) to
> >>short-circuit the call while remaining future-proof.
> >
> >How about this?
> 
> Yeah (see below)
> 
> >---
> >diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> >index 91b7f9b2b774..bb8a70e8fc77 100644
> >--- a/include/linux/page-flags.h
> >+++ b/include/linux/page-flags.h
> >@@ -547,7 +547,13 @@ static inline void ClearPageCompound(struct page *page)
> >  #endif /* !PAGEFLAGS_EXTENDED */
> >
> >  #ifdef CONFIG_HUGETLB_PAGE
> >-int PageHuge(struct page *page);
> >+int __PageHuge(struct page *page);
> >+static inline int PageHuge(struct page *page)
> >+{
> >+	if (!PageCompound(page))
> 
> Perhaps the above as likely()?

I have added it already when writing the changelog.

> [...]
> 
> >-EXPORT_SYMBOL_GPL(PageHuge);
> >+EXPORT_SYMBOL_GPL(__PageHuge);
> >
> >  /*
> >   * PageHeadHuge() only returns true for hugetlbfs head page, but not for
> >
> 
> Do the same thing here by inlining the PageHead() test?
> I guess the page_to_pgoff and __compound_tail_refcounted callers are rather
> hot?

Yes, that sounds like a good idea.

-- 
Michal Hocko
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/