Received-SPF: pass (google.com: domain of linux-kernel+bounces-132121-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <20240404171726.2302435-1-lokeshgidra@google.com>
 <Zg7hrt5HudXLBUn_@casper.infradead.org> <Zg8OYYV7DDo7S2Yf@x1n>
 <CAJuCfpGHe2=noJomL0XonT4dVGvZmVujRMEbgpYgVg_d5wo-+g@mail.gmail.com> <Zg8V5OycPYYk7-Qp@x1n>
In-Reply-To: <Zg8V5OycPYYk7-Qp@x1n>
From: Suren Baghdasaryan <surenb@google.com>
Date: Thu, 4 Apr 2024 14:07:45 -0700
Message-ID: <CAJuCfpGj5xk-NxSwW6Mt8NGZcV9N__8zVPMGXDPAYKMcN9=Oig@mail.gmail.com>
Subject: Re: [PATCH] userfaultfd: change src_folio after ensuring it's
 unpinned in UFFDIO_MOVE
To: Peter Xu <peterx@redhat.com>
Cc: Matthew Wilcox <willy@infradead.org>, Lokesh Gidra <lokeshgidra@google.com>, 
	akpm@linux-foundation.org, linux-mm@kvack.org, linux-kernel@vger.kernel.org, 
	kernel-team@android.com, aarcange@redhat.com, david@redhat.com, 
	zhengqi.arch@bytedance.com, kaleshsingh@google.com, ngeoffray@google.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Thu, Apr 4, 2024 at 2:04=E2=80=AFPM Peter Xu <peterx@redhat.com> wrote:
>
> On Thu, Apr 04, 2024 at 01:55:07PM -0700, Suren Baghdasaryan wrote:
> > On Thu, Apr 4, 2024 at 1:32=E2=80=AFPM Peter Xu <peterx@redhat.com> wro=
te:
> > >
> > > On Thu, Apr 04, 2024 at 06:21:50PM +0100, Matthew Wilcox wrote:
> > > > On Thu, Apr 04, 2024 at 10:17:26AM -0700, Lokesh Gidra wrote:
> > > > > -           folio_move_anon_rmap(src_folio, dst_vma);
> > > > > -           WRITE_ONCE(src_folio->index, linear_page_index(dst_vm=
a, dst_addr));
> > > > > -
> > > > >             src_pmdval =3D pmdp_huge_clear_flush(src_vma, src_add=
r, src_pmd);
> > > > >             /* Folio got pinned from under us. Put it back and fa=
il the move. */
> > > > >             if (folio_maybe_dma_pinned(src_folio)) {
> > > > > @@ -2270,6 +2267,9 @@ int move_pages_huge_pmd(struct mm_struct *m=
m, pmd_t *dst_pmd, pmd_t *src_pmd, pm
> > > > >                     goto unlock_ptls;
> > > > >             }
> > > > >
> > > > > +           folio_move_anon_rmap(src_folio, dst_vma);
> > > > > +           WRITE_ONCE(src_folio->index, linear_page_index(dst_vm=
a, dst_addr));
> > > > > +
> > > >
> > > > This use of WRITE_ONCE scares me.  We hold the folio locked.  Why d=
o
> > > > we need to use WRITE_ONCE?  Who's looking at folio->index without
> > > > holding the folio lock?
> > >
> > > Seems true, but maybe suitable for a separate patch to clean it even =
so?
> > > We also have the other pte level which has the same WRITE_ONCE(), so =
if we
> > > want to drop we may want to drop both.
> >
> > Yes, I'll do that separately and will remove WRITE_ONCE() in both place=
s.
>
> Thanks, Suren.  Besides, any comment on below?
>
> It's definely a generic per-vma question too (besides my willingness to
> remove that userfault specific code..), so comments welcomed.

Yes, I was typing my reply :)
This might have happened simply because lock_vma_under_rcu() was
originally developed to handle only anonymous page faults and then got
expanded to cover file-backed cases as well. Your suggestion seems
fine to me but I would feel much more comfortable after Matthew (who
added file-backed support) reviewed it.

>
> >
> > >
> > > I just got to start reading some the new move codes (Lokesh, apologie=
s on
> > > not be able to provide feedbacks previously..), but then I found one =
thing
> > > unclear, on special handling of private file mappings only in userfau=
lt
> > > context, and I didn't know why:
> > >
> > > lock_vma():
> > >         if (vma) {
> > >                 /*
> > >                  * lock_vma_under_rcu() only checks anon_vma for priv=
ate
> > >                  * anonymous mappings. But we need to ensure it is as=
signed in
> > >                  * private file-backed vmas as well.
> > >                  */
> > >                 if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->an=
on_vma))
> > >                         vma_end_read(vma);
> > >                 else
> > >                         return vma;
> > >         }
> > >
> > > AFAIU even for generic users of lock_vma_under_rcu(), anon_vma must b=
e
> > > stable to be used.  Here it's weird to become an userfault specific
> > > operation to me.
> > >
> > > I was surprised how it worked for private file maps on faults, then I=
 had a
> > > check and it seems we postponed such check until vmf_anon_prepare(), =
which
> > > is the CoW path already, so we do as I expected, but seems unnecessar=
y to
> > > that point?
> > >
> > > Would something like below make it much cleaner for us?  As I just do=
n't
> > > yet see why userfault is special here.
> > >
> > > Thanks,
> > >
> > > =3D=3D=3D8<=3D=3D=3D
> > > diff --git a/mm/memory.c b/mm/memory.c
> > > index 984b138f85b4..d5cf1d31c671 100644
> > > --- a/mm/memory.c
> > > +++ b/mm/memory.c
> > > @@ -3213,10 +3213,8 @@ vm_fault_t vmf_anon_prepare(struct vm_fault *v=
mf)
> > >
> > >         if (likely(vma->anon_vma))
> > >                 return 0;
> > > -       if (vmf->flags & FAULT_FLAG_VMA_LOCK) {
> > > -               vma_end_read(vma);
> > > -               return VM_FAULT_RETRY;
> > > -       }
> > > +       /* We shouldn't try a per-vma fault at all if anon_vma isn't =
solid */
> > > +       WARN_ON_ONCE(vmf->flags & FAULT_FLAG_VMA_LOCK);
> > >         if (__anon_vma_prepare(vma))
> > >                 return VM_FAULT_OOM;
> > >         return 0;
> > > @@ -5817,9 +5815,9 @@ struct vm_area_struct *lock_vma_under_rcu(struc=
t mm_struct *mm,
> > >          * find_mergeable_anon_vma uses adjacent vmas which are not l=
ocked.
> > >          * This check must happen after vma_start_read(); otherwise, =
a
> > >          * concurrent mremap() with MREMAP_DONTUNMAP could dissociate=
 the VMA
> > > -        * from its anon_vma.
> > > +        * from its anon_vma.  This applies to both anon or private f=
ile maps.
> > >          */
> > > -       if (unlikely(vma_is_anonymous(vma) && !vma->anon_vma))
> > > +       if (unlikely(!(vma->vm_flags & VM_SHARED) && !vma->anon_vma))
> > >                 goto inval_end_read;
> > >
> > >         /* Check since vm_start/vm_end might change before we lock th=
e VMA */
> > > diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> > > index f6267afe65d1..61f21da77dcd 100644
> > > --- a/mm/userfaultfd.c
> > > +++ b/mm/userfaultfd.c
> > > @@ -72,17 +72,8 @@ static struct vm_area_struct *lock_vma(struct mm_s=
truct *mm,
> > >         struct vm_area_struct *vma;
> > >
> > >         vma =3D lock_vma_under_rcu(mm, address);
> > > -       if (vma) {
> > > -               /*
> > > -                * lock_vma_under_rcu() only checks anon_vma for priv=
ate
> > > -                * anonymous mappings. But we need to ensure it is as=
signed in
> > > -                * private file-backed vmas as well.
> > > -                */
> > > -               if (!(vma->vm_flags & VM_SHARED) && unlikely(!vma->an=
on_vma))
> > > -                       vma_end_read(vma);
> > > -               else
> > > -                       return vma;
> > > -       }
> > > +       if (vma)
> > > +               return vma;
> > >
> > >         mmap_read_lock(mm);
> > >         vma =3D find_vma_and_prepare_anon(mm, address);
> > > --
> > > 2.44.0
> > >
> > >
> > > --
> > > Peter Xu
> > >
> >
>
> --
> Peter Xu
>