Re: [PATCH] mm: thp: don't have to lock page anymore when splitting PMD

From: Yang Shi
Date: Thu Mar 03 2022 - 22:13:14 EST


On Thu, Mar 3, 2022 at 6:25 PM Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:
>
> On 2022/3/4 6:20, Yang Shi wrote:
> > The commit c444eb564fb1 ("mm: thp: make the THP mapcount atomic against
> > __split_huge_pmd_locked()") locked the page for PMD split to make
> > mapcount stable for reuse_swap_page(), then commit 1c2f67308af4 ("mm:
> > thp: fix MADV_REMOVE deadlock on shmem THP") reduce the scope to
> > anonymous page only.
> >
> > However COW has not used mapcount to determine if the page is shared or
> > not anymore due to the COW fixes [1] from David Hildenbrand and the
> > reuse_swap_page() was removed as well. So PMD split doesn't have to
> > lock the page anymore. This patch basically reverted the above two
> > commits.
> >
>
> Sounds reasonable. Since mapcount is not used to determine if we need to COW
> the page, PMD split doesn't have to lock the page anymore.
>
> > [1] https://lore.kernel.org/linux-mm/20220131162940.210846-1-david@xxxxxxxxxx/
> >
> > Cc: David Hildenbrand <david@xxxxxxxxxx>
> > Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> > Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> > Cc: "Kirill A . Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx>
> > Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> > ---
> > mm/huge_memory.c | 44 +++++---------------------------------------
> > 1 file changed, 5 insertions(+), 39 deletions(-)
> >
> > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > index b49e1a11df2e..daaa698bd273 100644
> > --- a/mm/huge_memory.c
> > +++ b/mm/huge_memory.c
> > @@ -2134,8 +2134,6 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> > {
> > spinlock_t *ptl;
> > struct mmu_notifier_range range;
> > - bool do_unlock_folio = false;
> > - pmd_t _pmd;
> >
> > mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
> > address & HPAGE_PMD_MASK,
> > @@ -2148,48 +2146,16 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
> > * pmd against. Otherwise we can end up replacing wrong folio.
> > */
> > VM_BUG_ON(freeze && !folio);
> > - if (folio) {
> > - VM_WARN_ON_ONCE(!folio_test_locked(folio));
> > - if (folio != page_folio(pmd_page(*pmd)))
> > - goto out;
> > - }
> > + if (folio && folio != page_folio(pmd_page(*pmd)))
> > + goto out;
> >
> > -repeat:
> > - if (pmd_trans_huge(*pmd)) {
> > - if (!folio) {
> > - folio = page_folio(pmd_page(*pmd));
> > - /*
> > - * An anonymous page must be locked, to ensure that a
> > - * concurrent reuse_swap_page() sees stable mapcount;
> > - * but reuse_swap_page() is not used on shmem or file,
> > - * and page lock must not be taken when zap_pmd_range()
> > - * calls __split_huge_pmd() while i_mmap_lock is held.
> > - */
> > - if (folio_test_anon(folio)) {
> > - if (unlikely(!folio_trylock(folio))) {
> > - folio_get(folio);
> > - _pmd = *pmd;
> > - spin_unlock(ptl);
> > - folio_lock(folio);
> > - spin_lock(ptl);
> > - if (unlikely(!pmd_same(*pmd, _pmd))) {
> > - folio_unlock(folio);
> > - folio_put(folio);
> > - folio = NULL;
> > - goto repeat;
> > - }
> > - folio_put(folio);
> > - }
> > - do_unlock_folio = true;
> > - }
> > - }
> > - } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
> > + if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
> > goto out;
> > +
> > __split_huge_pmd_locked(vma, pmd, range.start, freeze);
>
> IUUC, here should be something like below:
> if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd))
> __split_huge_pmd_locked(vma, pmd, range.start, freeze);
>
> i.e. the pmd_trans_huge case is missing in the above change. Or am I miss something ?

Yes, you are definitely right. Must have pmd_trans_huge(*pmd).

>
> Thanks for the patch. This really simplify the code and avoid the unneeded overhead.
>
> > out:
> > spin_unlock(ptl);
> > - if (do_unlock_folio)
> > - folio_unlock(folio);
> > +
> > /*
> > * No need to double call mmu_notifier->invalidate_range() callback.
> > * They are 3 cases to consider inside __split_huge_pmd_locked():
> >
>