On Fri, Mar 27, 2020 at 5:40 PM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
>
> On Fri, Mar 27, 2020 at 01:45:55PM -0700, Yang Shi wrote:
> > On Fri, Mar 27, 2020 at 10:06 AM Kirill A. Shutemov
> > <kirill@xxxxxxxxxxxxx> wrote:
> > >
> > > We can collapse PTE-mapped compound pages. We only need to avoid
> > > handling them more than once: lock/unlock page only once if it's present
> > > in the PMD range multiple times as it handled on compound level. The
> > > same goes for LRU isolation and putpack.
> > >
> > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > > ---
> > > mm/khugepaged.c | 41 +++++++++++++++++++++++++++++++----------
> > > 1 file changed, 31 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > > index b47edfe57f7b..c8c2c463095c 100644
> > > --- a/mm/khugepaged.c
> > > +++ b/mm/khugepaged.c
> > > @@ -515,6 +515,17 @@ void __khugepaged_exit(struct mm_struct *mm)
> > >
> > > static void release_pte_page(struct page *page)
> > > {
> > > + /*
> > > + * We need to unlock and put compound page on LRU only once.
> > > + * The rest of the pages have to be locked and not on LRU here.
> > > + */
> > > + VM_BUG_ON_PAGE(!PageCompound(page) &&
> > > + (!PageLocked(page) && PageLRU(page)), page);
> > > +
> > > + if (!PageLocked(page))
> > > + return;
> > > +
> > > + page = compound_head(page);
> > > dec_node_page_state(page, NR_ISOLATED_ANON + page_is_file_cache(page));
> > > unlock_page(page);
> > > putback_lru_page(page);
> >
> > BTW, wouldn't this unlock the whole THP and put it back to LRU?
>
> It is the intention.
Yes, understood. Considering the below case:
Subpages 0, 1, 2, 3 are PTE mapped. Once subpage 0 is copied
release_pte_page() would be called then the whole THP would be
unlocked and putback to lru, then the loop would iterate to subpage 1,
2 and 3, but the page is not locked and on lru already. Is it
intentional?
>
> > Then we may copy the following PTE mapped pages with page unlocked and
> > on LRU. I don't see critical problem, just the pages might be on and off
> > LRU by others, i.e. vmscan, compaction, migration, etc. But no one could
> > take the page away since try_to_unmap() would fail, but not very
> > productive.
> >
> >
> > > @@ -537,6 +548,7 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
> > > pte_t *_pte;
> > > int none_or_zero = 0, result = 0, referenced = 0;
> > > bool writable = false;
> > > + LIST_HEAD(compound_pagelist);
> > >
> > > for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
> > > _pte++, address += PAGE_SIZE) {
> > > @@ -561,13 +573,23 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
> > > goto out;
> > > }
> > >
> > > - /* TODO: teach khugepaged to collapse THP mapped with pte */
> > > + VM_BUG_ON_PAGE(!PageAnon(page), page);
> > > +
> > > if (PageCompound(page)) {
> > > - result = SCAN_PAGE_COMPOUND;
> > > - goto out;
> > > - }
> > > + struct page *p;
> > > + page = compound_head(page);
> > >
> > > - VM_BUG_ON_PAGE(!PageAnon(page), page);
> > > + /*
> > > + * Check if we have dealt with the compount page
> > > + * already
> > > + */
> > > + list_for_each_entry(p, &compound_pagelist, lru) {
> > > + if (page == p)
> > > + break;
> > > + }
> > > + if (page == p)
> > > + continue;
> > > + }
> > >
> > > /*
> > > * We can do it before isolate_lru_page because the
> > > @@ -640,6 +662,9 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
> > > page_is_young(page) || PageReferenced(page) ||
> > > mmu_notifier_test_young(vma->vm_mm, address))
> > > referenced++;
> > > +
> > > + if (PageCompound(page))
> > > + list_add_tail(&page->lru, &compound_pagelist);
> > > }
> > > if (likely(writable)) {
> > > if (likely(referenced)) {
> > > @@ -1185,11 +1210,7 @@ static int khugepaged_scan_pmd(struct mm_struct *mm,
> > > goto out_unmap;
> > > }
> > >
> > > - /* TODO: teach khugepaged to collapse THP mapped with pte */
> > > - if (PageCompound(page)) {
> > > - result = SCAN_PAGE_COMPOUND;
> > > - goto out_unmap;
> > > - }
> > > + page = compound_head(page);
> > >
> > > /*
> > > * Record which node the original page is from and save this
> > > --
> > > 2.26.0
> > >
> > >
>
> --
> Kirill A. Shutemov
