Received-SPF: pass (google.com: domain of linux-kernel+bounces-159463-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Precedence: bulk
MIME-Version: 1.0
References: <20240425211136.486184-1-zi.yan@sent.com> <CAGsJ_4wa0LskQkoZf9r5bG5+wEkyfCYveMBSTbuDe0=t1QetTg@mail.gmail.com>
 <6C31DF81-94FB-4D09-A3B8-0CED2AD8EDDB@nvidia.com> <CAGsJ_4xzb8RrEuPEbnvR4GbDWuoGCYL4FsC3TObOifAZ4CHGOA@mail.gmail.com>
 <730660D2-E1BA-4A2E-B99C-2F160F9D9A9B@nvidia.com>
In-Reply-To: <730660D2-E1BA-4A2E-B99C-2F160F9D9A9B@nvidia.com>
From: Barry Song <21cnbao@gmail.com>
Date: Fri, 26 Apr 2024 11:28:23 +0800
Message-ID: <CAGsJ_4yNSKdft-G=4X2Dr5HJ-axh-8SohGW7nkUfTYrsNFr9_Q@mail.gmail.com>
Subject: Re: [PATCH v4] mm/rmap: do not add fully unmapped large folio to
 deferred split list
To: Zi Yan <ziy@nvidia.com>
Cc: Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org, 
	"Matthew Wilcox (Oracle)" <willy@infradead.org>, Yang Shi <shy828301@gmail.com>, 
	Ryan Roberts <ryan.roberts@arm.com>, David Hildenbrand <david@redhat.com>, 
	Lance Yang <ioworker0@gmail.com>, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Fri, Apr 26, 2024 at 10:50=E2=80=AFAM Zi Yan <ziy@nvidia.com> wrote:
>
> On 25 Apr 2024, at 22:23, Barry Song wrote:
>
> > On Fri, Apr 26, 2024 at 9:55=E2=80=AFAM Zi Yan <ziy@nvidia.com> wrote:
> >>
> >> On 25 Apr 2024, at 21:45, Barry Song wrote:
> >>
> >>> On Fri, Apr 26, 2024 at 5:11=E2=80=AFAM Zi Yan <zi.yan@sent.com> wrot=
e:
> >>>>
> >>>> From: Zi Yan <ziy@nvidia.com>
> >>>>
> >>>> In __folio_remove_rmap(), a large folio is added to deferred split l=
ist
> >>>> if any page in a folio loses its final mapping. But it is possible t=
hat
> >>>> the folio is fully unmapped and adding it to deferred split list is
> >>>> unnecessary.
> >>>>
> >>>> For PMD-mapped THPs, that was not really an issue, because removing =
the
> >>>> last PMD mapping in the absence of PTE mappings would not have added=
 the
> >>>> folio to the deferred split queue.
> >>>>
> >>>> However, for PTE-mapped THPs, which are now more prominent due to mT=
HP,
> >>>> they are always added to the deferred split queue. One side effect
> >>>> is that the THP_DEFERRED_SPLIT_PAGE stat for a PTE-mapped folio can =
be
> >>>> unintentionally increased, making it look like there are many partia=
lly
> >>>> mapped folios -- although the whole folio is fully unmapped stepwise=
.
> >>>>
> >>>> Core-mm now tries batch-unmapping consecutive PTEs of PTE-mapped THP=
s
> >>>> where possible starting from commit b06dc281aa99 ("mm/rmap: introduc=
e
> >>>> folio_remove_rmap_[pte|ptes|pmd]()"). When it happens, a whole PTE-m=
apped
> >>>> folio is unmapped in one go and can avoid being added to deferred sp=
lit
> >>>> list, reducing the THP_DEFERRED_SPLIT_PAGE noise. But there will sti=
ll be
> >>>> noise when we cannot batch-unmap a complete PTE-mapped folio in one =
go
> >>>> -- or where this type of batching is not implemented yet, e.g., migr=
ation.
> >>>>
> >>>> To avoid the unnecessary addition, folio->_nr_pages_mapped is checke=
d
> >>>> to tell if the whole folio is unmapped. If the folio is already on
> >>>> deferred split list, it will be skipped, too.
> >>>>
> >>>> Note: commit 98046944a159 ("mm: huge_memory: add the missing
> >>>> folio_test_pmd_mappable() for THP split statistics") tried to exclud=
e
> >>>> mTHP deferred split stats from THP_DEFERRED_SPLIT_PAGE, but it does =
not
> >>>> fix the above issue. A fully unmapped PTE-mapped order-9 THP was sti=
ll
> >>>> added to deferred split list and counted as THP_DEFERRED_SPLIT_PAGE,
> >>>> since nr is 512 (non zero), level is RMAP_LEVEL_PTE, and inside
> >>>> deferred_split_folio() the order-9 folio is folio_test_pmd_mappable(=
).
> >>>>
> >>>> Signed-off-by: Zi Yan <ziy@nvidia.com>
> >>>> Reviewed-by: Yang Shi <shy828301@gmail.com>
> >>>> ---
> >>>>  mm/rmap.c | 8 +++++---
> >>>>  1 file changed, 5 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/mm/rmap.c b/mm/rmap.c
> >>>> index a7913a454028..220ad8a83589 100644
> >>>> --- a/mm/rmap.c
> >>>> +++ b/mm/rmap.c
> >>>> @@ -1553,9 +1553,11 @@ static __always_inline void __folio_remove_rm=
ap(struct folio *folio,
> >>>>                  * page of the folio is unmapped and at least one pa=
ge
> >>>>                  * is still mapped.
> >>>>                  */
> >>>> -               if (folio_test_large(folio) && folio_test_anon(folio=
))
> >>>> -                       if (level =3D=3D RMAP_LEVEL_PTE || nr < nr_p=
mdmapped)
> >>>> -                               deferred_split_folio(folio);
> >>>> +               if (folio_test_large(folio) && folio_test_anon(folio=
) &&
> >>>> +                   list_empty(&folio->_deferred_list) &&
> >>>> +                   ((level =3D=3D RMAP_LEVEL_PTE && atomic_read(map=
ped)) ||
> >>>> +                    (level =3D=3D RMAP_LEVEL_PMD && nr < nr_pmdmapp=
ed)))
> >>>> +                       deferred_split_folio(folio);
> >>>
> >>> Hi Zi Yan,
> >>> in case a mTHP is mapped by two processed (forked but not CoW yet), i=
f we
> >>> unmap the whole folio by pte level in one process only, are we still =
adding this
> >>> folio into deferred list?
> >>
> >> No. Because the mTHP is still fully mapped by the other process. In te=
rms of code,
> >> nr will be 0 in that case and this if condition is skipped. nr is only=
 increased
> >> from 0 when one of the subpages in the mTHP has no mapping, namely pag=
e->_mapcount
> >> becomes negative and last is true in the case RMAP_LEVEL_PTE.
> >
> > Ok. i see, so "last" won't be true?
> >
> > case RMAP_LEVEL_PTE:
> > do {
> > last =3D atomic_add_negative(-1, &page->_mapcount);
> >    if (last && folio_test_large(folio)) {
> >        last =3D atomic_dec_return_relaxed(mapped);
> >        last =3D (last < ENTIRELY_MAPPED);
> > }
> >
> > if (last)
> >      nr++;
> > } while (page++, --nr_pages > 0);
> > break;
>
> Right, because for every subpage its corresponding
> last =3D atomic_add_negative(-1, &page->_mapcount); is not true after the=
 unmapping.2

if a mTHP is mapped only by one process, and we unmap it entirely, we will
get nr > 0, then we are executing adding it into deferred_list? so it seems
atomic_read(mapped) is preventing this case from adding deferred_list?

I wonder if  it is possible to fixup nr to 0 from the first place?
for example
/* we are doing an entire unmapping */
if (page=3D=3D&folio->page && nr_pages =3D=3D  folio_nr_pages(folio))
..

>
>
> --
> Best Regards,
> Yan, Zi