Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:8 as permitted sender) client-ip=2620:137:e000::3:8;
MIME-Version: 1.0
References: <20230918211608.3580629-1-surenb@google.com> <ZQlhaO1C5kf8zGjf@dhcp22.suse.cz>
In-Reply-To: <ZQlhaO1C5kf8zGjf@dhcp22.suse.cz>
From:   Yang Shi <shy828301@gmail.com>
Date:   Tue, 19 Sep 2023 14:09:43 -0700
Message-ID: <CAHbLzkpV8+0Bn_mpGODDbRsAOmDexG_JofUKQEVW-tGPJB-iyw@mail.gmail.com>
Subject: Re: [PATCH 1/1] mm: lock VMAs skipped by a failed queue_pages_range()
To:     Michal Hocko <mhocko@suse.com>
Cc:     Suren Baghdasaryan <surenb@google.com>, akpm@linux-foundation.org,
        willy@infradead.org, hughd@google.com, vbabka@suse.cz,
        syzkaller-bugs@googlegroups.com, linux-mm@kvack.org,
        linux-kernel@vger.kernel.org,
        syzbot+b591856e0f0139f83023@syzkaller.appspotmail.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Tue, Sep 19, 2023 at 1:53=E2=80=AFAM Michal Hocko <mhocko@suse.com> wrot=
e:
>
> On Mon 18-09-23 14:16:08, Suren Baghdasaryan wrote:
> > When queue_pages_range() encounters an unmovable page, it terminates
> > its page walk. This walk, among other things, locks the VMAs in the ran=
ge.
> > This termination might result in some VMAs being left unlock after
> > queue_pages_range() completes. Since do_mbind() continues to operate on
> > these VMAs despite the failure from queue_pages_range(), it will encoun=
ter
> > an unlocked VMA.
> > This mbind() behavior has been modified several times before and might
> > need some changes to either finish the page walk even in the presence
> > of unmovable pages or to error out immediately after the failure to
> > queue_pages_range(). However that requires more discussions, so to
> > fix the immediate issue, explicitly lock the VMAs in the range if
> > queue_pages_range() failed. The added condition does not save much
> > but is added for documentation purposes to understand when this extra
> > locking is needed.
>
> The semantic of the walk in this case is really clear as mud. I was
> trying to reconstruct the whole picture and it really hurts... Then I
> found http://lkml.kernel.org/r/CAHbLzkrmTaqBRmHVdE2kyW57Uoghqd_E+jAXC9cB5=
ofkhL-uvw@mail.gmail.com
> and that helped a lot. Let's keep it a reference at least in the email
> thread here for future.

FYI, I'm working on a fix for the regression mentioned in that series,
and Hugh has some clean up and enhancement for that too.

>
> > Fixes: 49b0638502da ("mm: enable page walking API to lock vmas during t=
he walk")
> > Reported-by: syzbot+b591856e0f0139f83023@syzkaller.appspotmail.com
> > Closes: https://lore.kernel.org/all/000000000000f392a60604a65085@google=
.com/
> > Signed-off-by: Suren Baghdasaryan <surenb@google.com>
>
> I cannot say I like the patch (it looks like a potential double locking
> unless you realize this lock is special) but considering this might be ju=
st
> temporal I do not mind.
>
> Acked-by: Michal Hocko <mhocko@suse.com>
>
> Thanks!
>
> > ---
> >  mm/mempolicy.c | 3 +++
> >  1 file changed, 3 insertions(+)
> >
> > diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> > index 42b5567e3773..cbc584e9b6ca 100644
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -1342,6 +1342,9 @@ static long do_mbind(unsigned long start, unsigne=
d long len,
> >       vma_iter_init(&vmi, mm, start);
> >       prev =3D vma_prev(&vmi);
> >       for_each_vma_range(vmi, vma, end) {
> > +             /* If queue_pages_range failed then not all VMAs might be=
 locked */
> > +             if (ret)
> > +                     vma_start_write(vma);
> >               err =3D mbind_range(&vmi, vma, &prev, start, end, new);
> >               if (err)
> >                       break;
> > --
> > 2.42.0.459.ge4e396fd5e-goog
>
> --
> Michal Hocko
> SUSE Labs