Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
MIME-Version: 1.0
References: <20230923013148.1390521-1-surenb@google.com> <20230923013148.1390521-3-surenb@google.com>
 <CAG48ez1N2kryy08eo0dcJ5a9O-3xMT8aOrgrcD+CqBN=cBfdDw@mail.gmail.com> <03f95e90-82bd-6ee2-7c0d-d4dc5d3e15ee@redhat.com>
In-Reply-To: <03f95e90-82bd-6ee2-7c0d-d4dc5d3e15ee@redhat.com>
From:   Suren Baghdasaryan <surenb@google.com>
Date:   Wed, 27 Sep 2023 11:25:22 -0700
Message-ID: <CAJuCfpHf6BWaf_k5dBx7mAz49kF5BwBhW_mUxu4E_p2iAy9-iA@mail.gmail.com>
Subject: Re: [PATCH v2 2/3] userfaultfd: UFFDIO_REMAP uABI
To:     David Hildenbrand <david@redhat.com>
Cc:     Jann Horn <jannh@google.com>, akpm@linux-foundation.org,
        viro@zeniv.linux.org.uk, brauner@kernel.org, shuah@kernel.org,
        aarcange@redhat.com, lokeshgidra@google.com, peterx@redhat.com,
        hughd@google.com, mhocko@suse.com, axelrasmussen@google.com,
        rppt@kernel.org, willy@infradead.org, Liam.Howlett@oracle.com,
        zhangpeng362@huawei.com, bgeffon@google.com,
        kaleshsingh@google.com, ngeoffray@google.com, jdduke@google.com,
        linux-mm@kvack.org, linux-fsdevel@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-kselftest@vger.kernel.org,
        kernel-team@android.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Wed, Sep 27, 2023 at 6:29=E2=80=AFAM David Hildenbrand <david@redhat.com=
> wrote:
>
> >> +static int remap_anon_pte(struct mm_struct *dst_mm, struct mm_struct =
*src_mm,
> >> +                         struct vm_area_struct *dst_vma,
> >> +                         struct vm_area_struct *src_vma,
> >> +                         unsigned long dst_addr, unsigned long src_ad=
dr,
> >> +                         pte_t *dst_pte, pte_t *src_pte,
> >> +                         pte_t orig_dst_pte, pte_t orig_src_pte,
> >> +                         spinlock_t *dst_ptl, spinlock_t *src_ptl,
> >> +                         struct folio *src_folio)
> >> +{
> >> +       struct anon_vma *dst_anon_vma;
> >> +
> >> +       double_pt_lock(dst_ptl, src_ptl);
> >> +
> >> +       if (!pte_same(*src_pte, orig_src_pte) ||
> >> +           !pte_same(*dst_pte, orig_dst_pte) ||
> >> +           folio_test_large(src_folio) ||
> >> +           folio_estimated_sharers(src_folio) !=3D 1) {
>
> ^ here you should check PageAnonExclusive. Please get rid of any
> implicit explicit/implcit mapcount checks.

Ack.

>
> >> +               double_pt_unlock(dst_ptl, src_ptl);
> >> +               return -EAGAIN;
> >> +       }
> >> +
> >> +       BUG_ON(!folio_test_anon(src_folio));
> >> +
> >> +       dst_anon_vma =3D (void *)dst_vma->anon_vma + PAGE_MAPPING_ANON=
;
> >> +       WRITE_ONCE(src_folio->mapping,
> >> +                  (struct address_space *) dst_anon_vma);
>
> I have some cleanups pending for page_move_anon_rmap(), that moves the
> SetPageAnonExclusive hunk out. Here we should be using
> page_move_anon_rmap() [or rather, folio_move_anon_rmap() after my cleanup=
s]
>
> I'll send them out soonish.

Should I keep this as is in my next version until you post the
cleanups? I can add a TODO comment to convert it to
folio_move_anon_rmap() once it's ready.

>
> >> +       WRITE_ONCE(src_folio->index, linear_page_index(dst_vma,
> >> +                                                     dst_addr)); >> +
> >> +       orig_src_pte =3D ptep_clear_flush(src_vma, src_addr, src_pte);
> >> +       orig_dst_pte =3D mk_pte(&src_folio->page, dst_vma->vm_page_pro=
t);
> >> +       orig_dst_pte =3D maybe_mkwrite(pte_mkdirty(orig_dst_pte),
> >> +                                    dst_vma);
> >
> > I think there's still a theoretical issue here that you could fix by
> > checking for the AnonExclusive flag, similar to the huge page case.
> >
> > Consider the following scenario:
> >
> > 1. process P1 does a write fault in a private anonymous VMA, creating
> > and mapping a new anonymous page A1
> > 2. process P1 forks and creates two children P2 and P3. afterwards, A1
> > is mapped in P1, P2 and P3 as a COW page, with mapcount 3.
> > 3. process P1 removes its mapping of A1, dropping its mapcount to 2.
> > 4. process P2 uses vmsplice() to grab a reference to A1 with get_user_p=
ages()
> > 5. process P2 removes its mapping of A1, dropping its mapcount to 1.
> >
> > If at this point P3 does a write fault on its mapping of A1, it will
> > still trigger copy-on-write thanks to the AnonExclusive mechanism; and
> > this is necessary to avoid P3 mapping A1 as writable and writing data
> > into it that will become visible to P2, if P2 and P3 are in different
> > security contexts.
> >
> > But if P3 instead moves its mapping of A1 to another address with
> > remap_anon_pte() which only does a page mapcount check, the
> > maybe_mkwrite() will directly make the mapping writable, circumventing
> > the AnonExclusive mechanism.
> >
>
> Yes, can_change_pte_writable() contains the exact logic when we can turn
> something easily writable even if it wasn't writable before. which
> includes that PageAnonExclusive is set. (but with uffd-wp or softdirty
> tracking, there is more to consider)

For uffd_remap can_change_pte_writable() would fail it VM_WRITE is not
set, but we want remapping to work for RO memory as well. Are you
saying that a PageAnonExclusive() check alone would not be enough
here?

Thanks,
Suren.

>
> --
> Cheers,
>
> David / dhildenb
>