Re: [PATCH 1/1] userfaultfd: fix a crash when UFFDIO_MOVE handles a THP hole

[Lokesh Gidra]

On Thu, Jul 31, 2025 at 7:24 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Thu, Jul 31, 2025 at 12:35 AM Lokesh Gidra <lokeshgidra@xxxxxxxxxx> wrote:
> >
> > On Wed, Jul 30, 2025 at 6:58 PM Hillf Danton <hdanton@xxxxxxxx> wrote:
> > >
> > > #syz test
> > >
> > > When UFFDIO_MOVE is used with UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES and it
> > > encounters a non-present THP, it fails to properly recognize an unmapped
> > > hole and tries to access a non-existent folio, resulting in
> > > a crash. Add a check to skip non-present THPs.
> > >
> > Thanks Suren for promptly addressing this issue.
> >
> > > Fixes: adef440691ba ("userfaultfd: UFFDIO_MOVE uABI")
> > > Reported-by: syzbot+b446dbe27035ef6bd6c2@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > Closes: https://lore.kernel.org/all/68794b5c.a70a0220.693ce.0050.GAE@xxxxxxxxxx/
> > > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > ---
> > >  mm/userfaultfd.c | 38 +++++++++++++++++++++++---------------
> > >  1 file changed, 23 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/mm/userfaultfd.c b/mm/userfaultfd.c
> > > index cbed91b09640..60be8080ddd0 100644
> > > --- a/mm/userfaultfd.c
> > > +++ b/mm/userfaultfd.c
> > > @@ -1818,27 +1818,35 @@ ssize_t move_pages(struct userfaultfd_ctx *ctx, unsigned long dst_start,
> > >
> > >                 ptl = pmd_trans_huge_lock(src_pmd, src_vma);
> > >                 if (ptl) {
> > > -                       /* Check if we can move the pmd without splitting it. */
> > > -                       if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
> > > -                           !pmd_none(dst_pmdval)) {
> > > -                               struct folio *folio = pmd_folio(*src_pmd);
> > > +                       if (pmd_present(*src_pmd) || is_pmd_migration_entry(*src_pmd)) {
> > > +                               /* Check if we can move the pmd without splitting it. */
> > > +                               if (move_splits_huge_pmd(dst_addr, src_addr, src_start + len) ||
> > > +                                   !pmd_none(dst_pmdval)) {
> > > +                                       if (pmd_present(*src_pmd)) {
> > > +                                               struct folio *folio = pmd_folio(*src_pmd);
> > > +
> > > +                                               if (!folio || (!is_huge_zero_folio(folio) &&
> > > +                                                              !PageAnonExclusive(&folio->page))) {
> > > +                                                       spin_unlock(ptl);
> > > +                                                       err = -EBUSY;
> > > +                                                       break;
> > > +                                               }
> > > +                                       }
> > >
> > > -                               if (!folio || (!is_huge_zero_folio(folio) &&
> > > -                                              !PageAnonExclusive(&folio->page))) {
> > >                                         spin_unlock(ptl);
> > > -                                       err = -EBUSY;
> > > -                                       break;
> > > +                                       split_huge_pmd(src_vma, src_pmd, src_addr);
> > > +                                       /* The folio will be split by move_pages_pte() */
> > > +                                       continue;
> > >                                 }
> > >
> > > +                               err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
> > > +                                                         dst_pmdval, dst_vma, src_vma,
> > > +                                                         dst_addr, src_addr);
> > > +                       } else {
> > > +                               /* nothing to do to move a hole */
> > >                                 spin_unlock(ptl);
> > > -                               split_huge_pmd(src_vma, src_pmd, src_addr);
> > > -                               /* The folio will be split by move_pages_pte() */
> > > -                               continue;
> > > +                               err = 0;
> > I think we need to act here depending on whether
> > UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES is set or not.
>
> Hmm, yes, I think you are right. I thought we would bail out earlier
> if !UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES but I think it's possible to get
> here if the PMD was established earlier but then unmapped.

That makes sense too. My thinking was that the
!UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES check above is only when !src_pmd,
which means the src pmd-page itself isn't present. However, the case
where pmd-page is present, but the pmd entry is not; continuing
skipping the hole even when user cannot tolerate src-holes would break
the userspace logic.
>
> >
> >            err = (mode & UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES) ? 0 : -ENOENT;
> >
> > Also, IMO, the step_size in this case should be the minimum of
> > remaining length and HPAGE_PMD_SIZE.
>
> Ah, ok. I think it matters only for incrementing "moved" correctly
> because otherwise the functionality is the same.
>
Right. Returning an incorrect "moved" value to userspace can break the
logic, possibly causing data corruption.

> > >                         }
> > > -
> > > -                       err = move_pages_huge_pmd(mm, dst_pmd, src_pmd,
> > > -                                                 dst_pmdval, dst_vma, src_vma,
> > > -                                                 dst_addr, src_addr);
> > >                         step_size = HPAGE_PMD_SIZE;
> > >                 } else {
> > >                         if (pmd_none(*src_pmd)) {
> > I have a related question/doubt: why do we populate the page-table
> > hierarchy on the src side [1] (and then also at line 1857) when a hole
> > is found? IMHO, it shouldn't be needed. Depending on whether
> > UFFDIO_MOVE_MODE_ALLOW_SRC_HOLES is set or not, it should either
> > return -ENOENT, or continue past the hole. Please correct me if I'm
> > wrong.
>
> I thought about that too. I think it's done to simplify the logic.
> This way we can treat the cases when PMD was never allocated and when
> PMD was allocated, mapped and then unmapped the same way.
>
Makes sense. Thanks for clarifying.
> >
> > [1] https://elixir.bootlin.com/linux/v6.16/source/mm/userfaultfd.c#L1797
> >
> > >
> > > base-commit: 01da54f10fddf3b01c5a3b80f6b16bbad390c302
> > > --
> > > 2.50.1.552.g942d659e1b-goog