Re: [PATCH] MADVISE_FREE, THP: Fix madvise_free_huge_pmd return value after splitting

From: Minchan Kim
Date: Sun Jun 19 2016 - 21:07:34 EST


On Fri, Jun 17, 2016 at 12:45:43PM -0700, Huang, Ying wrote:
> Minchan Kim <minchan@xxxxxxxxxx> writes:
>
> > Hi,
> >
> > On Thu, Jun 16, 2016 at 08:03:54PM -0700, Huang, Ying wrote:
> >> From: Huang Ying <ying.huang@xxxxxxxxx>
> >>
> >> madvise_free_huge_pmd should return 0 if the fallback PTE operations are
> >> required. In madvise_free_huge_pmd, if part pages of THP are discarded,
> >> the THP will be split and fallback PTE operations should be used if
> >> splitting succeeds. But the original code will make fallback PTE
> >> operations skipped, after splitting succeeds. Fix that via make
> >> madvise_free_huge_pmd return 0 after splitting successfully, so that the
> >> fallback PTE operations will be done.
> >
> > You're right. Thanks!
> >
> >>
> >> Know issues: if my understanding were correct, return 1 from
> >> madvise_free_huge_pmd means the following processing for the PMD should
> >> be skipped, while return 0 means the following processing is still
> >> needed. So the function should return 0 only if the THP is split
> >> successfully or the PMD is not trans huge. But the pmd_trans_unstable
> >> after madvise_free_huge_pmd guarantee the following processing will be
> >> skipped for huge PMD. So current code can run properly. But if my
> >> understanding were correct, we can clean up return code of
> >> madvise_free_huge_pmd accordingly.
> >
> > I like your clean up. Just a minor comment below.
> >
> >>
> >> Signed-off-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
> >> ---
> >> mm/huge_memory.c | 7 +------
> >> 1 file changed, 1 insertion(+), 6 deletions(-)
> >>
> >> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >> index 2ad52d5..64dc95d 100644
> >> --- a/mm/huge_memory.c
> >> +++ b/mm/huge_memory.c
> >
> > First of all, let's change ret from int to bool.
> > And then, add description in the function entry.
>
> Yes. bool looks better than int.
>
> > /*
> > * Return true if we do MADV_FREE successfully on entire pmd page.
> > * Otherwise, return false.
> > */
>
> This way, we need to return false if we failed to split huge page, this
> will cause unnecessary pmd_trans_unstable check. How about to change
> the comments to

I focused the function name "madvise_free_huge_pmd". IOW, the function
should free huge pmd page. If it is successful, done. Otherwise, next
routines should handle it.

If it fail to split, pmd_trans_unstable will check it and return.
I don't think it's heavy operation to affect performance so rather
than making function fast, I wanted to make it simple by following
function name.

>
> /*
> * Return true if we finished processing entire pmd page and needn't
> * fall back pte processing. Otherwise, return false.
> */
>
> Best Regards,
> Huang, Ying
>
> > And do not set to 1 if it is huge_zero_pmd but just goto out to
> > return false.
> >
> > Thanks!
> >
> >> @@ -1655,14 +1655,9 @@ int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
> >> if (next - addr != HPAGE_PMD_SIZE) {
> >> get_page(page);
> >> spin_unlock(ptl);
> >> - if (split_huge_page(page)) {
> >> - put_page(page);
> >> - unlock_page(page);
> >> - goto out_unlocked;
> >> - }
> >> + split_huge_page(page);
> >> put_page(page);
> >> unlock_page(page);
> >> - ret = 1;
> >> goto out_unlocked;
> >> }
> >>
> >> --
> >> 2.8.1
> >>
> >> --
> >> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> >> the body to majordomo@xxxxxxxxxx For more info on Linux MM,
> >> see: http://www.linux-mm.org/ .
> >> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>