Re: [PATCH 2/3] mm: free zapped tail pages when splitting isolated thp

[Yu Zhao]

()
On Fri, Aug 13, 2021 at 5:24 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> On Wed, Aug 11, 2021 at 4:12 PM Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> >
> > On Wed, Aug 11, 2021 at 4:25 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
> > >
> > > On Sun, Aug 8, 2021 at 10:49 AM Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Aug 4, 2021 at 6:13 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Jul 30, 2021 at 11:39 PM Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> > > > > >
> > > > > > If a tail page has only two references left, one inherited from the
> > > > > > isolation of its head and the other from lru_add_page_tail() which we
> > > > > > are about to drop, it means this tail page was concurrently zapped.
> > > > > > Then we can safely free it and save page reclaim or migration the
> > > > > > trouble of trying it.
> > > > > >
> > > > > > Signed-off-by: Yu Zhao <yuzhao@xxxxxxxxxx>
> > > > > > Tested-by: Shuang Zhai <zhais@xxxxxxxxxx>
> > > > > > ---
> > > > > >  include/linux/vm_event_item.h |  1 +
> > > > > >  mm/huge_memory.c              | 28 ++++++++++++++++++++++++++++
> > > > > >  mm/vmstat.c                   |  1 +
> > > > > >  3 files changed, 30 insertions(+)
> > > > > >
> > > > > > diff --git a/include/linux/vm_event_item.h b/include/linux/vm_event_item.h
> > > > > > index ae0dd1948c2b..829eeac84094 100644
> > > > > > --- a/include/linux/vm_event_item.h
> > > > > > +++ b/include/linux/vm_event_item.h
> > > > > > @@ -99,6 +99,7 @@ enum vm_event_item { PGPGIN, PGPGOUT, PSWPIN, PSWPOUT,
> > > > > >  #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> > > > > >                 THP_SPLIT_PUD,
> > > > > >  #endif
> > > > > > +               THP_SPLIT_FREE,
> > > > > >                 THP_ZERO_PAGE_ALLOC,
> > > > > >                 THP_ZERO_PAGE_ALLOC_FAILED,
> > > > > >                 THP_SWPOUT,
> > > > > > diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> > > > > > index d8b655856e79..5120478bca41 100644
> > > > > > --- a/mm/huge_memory.c
> > > > > > +++ b/mm/huge_memory.c
> > > > > > @@ -2432,6 +2432,8 @@ static void __split_huge_page(struct page *page, struct list_head *list,
> > > > > >         struct address_space *swap_cache = NULL;
> > > > > >         unsigned long offset = 0;
> > > > > >         unsigned int nr = thp_nr_pages(head);
> > > > > > +       LIST_HEAD(pages_to_free);
> > > > > > +       int nr_pages_to_free = 0;
> > > > > >         int i;
> > > > > >
> > > > > >         VM_BUG_ON_PAGE(list && PageLRU(head), head);
> > > > > > @@ -2506,6 +2508,25 @@ static void __split_huge_page(struct page *page, struct list_head *list,
> > > > > >                         continue;
> > > > > >                 unlock_page(subpage);
> > > > > >
> > > > > > +               /*
> > > > > > +                * If a tail page has only two references left, one inherited
> > > > > > +                * from the isolation of its head and the other from
> > > > > > +                * lru_add_page_tail() which we are about to drop, it means this
> > > > > > +                * tail page was concurrently zapped. Then we can safely free it
> > > > > > +                * and save page reclaim or migration the trouble of trying it.
> > > > > > +                */
> > > > > > +               if (list && page_ref_freeze(subpage, 2)) {
> > > > > > +                       VM_BUG_ON_PAGE(PageLRU(subpage), subpage);
> > > > > > +                       VM_BUG_ON_PAGE(PageCompound(subpage), subpage);
> > > > > > +                       VM_BUG_ON_PAGE(page_mapped(subpage), subpage);
> > > > > > +
> > > > > > +                       ClearPageActive(subpage);
> > > > > > +                       ClearPageUnevictable(subpage);
> > > > > > +                       list_move(&subpage->lru, &pages_to_free);
> > > > > > +                       nr_pages_to_free++;
> > > > > > +                       continue;
> > > > > > +               }
> > > > >
> > > > > Yes, such page could be freed instead of swapping out. But I'm
> > > > > wondering if we could have some simpler implementation. Since such
> > > > > pages will be re-added to page list, so we should be able to check
> > > > > their refcount in shrink_page_list(). If the refcount is 1, the
> > > > > refcount inc'ed by lru_add_page_tail() has been put by later
> > > > > put_page(), we know it is freed under us since the only refcount comes
> > > > > from isolation, we could just jump to "keep" (the label in
> > > > > shrink_page_list()), then such page will be freed later by
> > > > > shrink_inactive_list().
> > > > >
> > > > > For MADV_PAGEOUT, I think we could add some logic to handle such page
> > > > > after shrink_page_list(), just like what shrink_inactive_list() does.
> > > > >
> > > > > Migration already handles refcount == 1 page, so should not need any change.
> > > > >
> > > > > Is this idea feasible?
> > > >
> > > > Yes, but then we would have to loop over the tail pages twice, here
> > > > and in shrink_page_list(), right?
> > >
> > > I don't quite get what you mean "loop over the tail pages twice". Once
> > > THP is isolated then get split, all the tail pages will be put on the
> > > list (local list for isolated pages), then the reclaimer would deal
> > > with the head page, then continue to iterate the list to deal with
> > > tail pages. Your patch could free the tail pages earlier. But it
> > > should not make too much difference to free the tail pages a little
> > > bit later IMHO.
> >
> > We are in a (the first) loop here. If we free the tail pages later,
> > then we will need to loop over them again (the second).
> >
> > IOW,
> > 1) __split_huge_page(): for each of the 511 tail pages (first loop).
> > 2) shrink_page_list(): for each of the 511 tail pages (second loop).
> >
> > > > In addition, if we try to freeze the refcount of a page in
> > > > shrink_page_list(), we couldn't be certain whether this page used to
> > > > be a tail page. So we would have to test every page. If a page wasn't
> > > > a tail page, it's unlikely for its refcount to drop unless there is a
> > > > race. But this patch isn't really intended to optimize such a race.
> > > > It's mainly for the next, i.e., we know there is a good chance to drop
> > > > tail pages (~10% on our systems). Sounds reasonable? Thanks.
> > >
> > > I'm not sure what is the main source of the partial mapped THPs from
> > > your fleets. But if most of them are generated by MADV_DONTNEED (this
> > > is used by some userspace memory allocator libs), they should be on
> > > deferred split list too. Currently deferred split shrinker just
> > > shrinks those THPs (simply split them and free unmapped sub pages)
> > > proportionally, we definitely could shrink them more aggressively, for
> > > example, by setting shrinker->seeks to 0. I'm wondering if this will
> > > achieve a similar effect or not.
> >
> > Not partially mapped but internal fragmentation.
> >
> > IOW, some of the 4KB pages within a THP were never written into, which
> > can be common depending on the implementations of userspace memory
> > allocators.
>
> OK, this is actually what the patch #3 does. The patch #3 just doesn't
> remap the "all zero" page when splitting the THP IIUC. But the page
> has refcount from isolation so it can't be simply freed by put_page().
>
> Actually this makes me think my suggestion is better. It doesn't make
> too much sense to me to have page free logic (manipulate flags,
> uncharge memcg, etc) in THP split.
>
> There have been a couple of places to handle such cases:
> - deferred split shrinker: the unmapped subpage is just freed by
> put_page() since there is no extra refcount
> - migration: check page refcount then free the refcount == 1 one
>
> Here you add the third case in the page reclaim path, so why not just
> let the page reclaim handle all the work for freeing page?

As I have explained previously:

1) We would have to loop over tail pages twice. Not much overhead but
unnecessary.
2) We would have to try to freeze the refcount on _every_ page in
shrink_page_list() -- shrink_page_list() takes all pages, not just
relevant ones (previously being tail). Attempting to freeze refcount
on an irrelevant page will likely fail. Again, not a significant
overhead but better to avoid.

I'm not against your idea. But I'd like to hear some clarifications
about the points above. That is whether you think it's still a good
idea to do what you suggested after taking these into account.

> > > I really don't have any objection to free such pages, but just
> > > wondering if we could have something simpler or not.
> >
> > Thanks.
> >
> > > > > > +
> > > > > >                 /*
> > > > > >                  * Subpages may be freed if there wasn't any mapping
> > > > > >                  * like if add_to_swap() is running on a lru page that
> > > > > > @@ -2515,6 +2536,13 @@ static void __split_huge_page(struct page *page, struct list_head *list,
> > > > > >                  */
> > > > > >                 put_page(subpage);
> > > > > >         }
> > > > > > +
> > > > > > +       if (!nr_pages_to_free)
> > > > > > +               return;
> > > > > > +
> > > > > > +       mem_cgroup_uncharge_list(&pages_to_free);
> > > > > > +       free_unref_page_list(&pages_to_free);
> > > > > > +       count_vm_events(THP_SPLIT_FREE, nr_pages_to_free);
> > > > > >  }
> > > > > >
> > > > > >  int total_mapcount(struct page *page)
> > > > > > diff --git a/mm/vmstat.c b/mm/vmstat.c
> > > > > > index b0534e068166..f486e5d98d96 100644
> > > > > > --- a/mm/vmstat.c
> > > > > > +++ b/mm/vmstat.c
> > > > > > @@ -1300,6 +1300,7 @@ const char * const vmstat_text[] = {
> > > > > >  #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> > > > > >         "thp_split_pud",
> > > > > >  #endif
> > > > > > +       "thp_split_free",
> > > > > >         "thp_zero_page_alloc",
> > > > > >         "thp_zero_page_alloc_failed",
> > > > > >         "thp_swpout",
> > > > > > --
> > > > > > 2.32.0.554.ge1b32706d8-goog
> > > > > >