Re: [PATCH 2/3] userfaultfd: prevent unregistering VMAs through a different userfaultfd

[Tal Zussman <tz2294@columbia.edu>]

On Tue, Jun 3, 2025 at 8:52 PM James Houghton <jthoughton@google.com> wrote:
>
> On Tue, Jun 3, 2025 at 3:15 PM Tal Zussman <tz2294@columbia.edu> wrote:
> >
> > Currently, a VMA registered with a uffd can be unregistered through a
> > different uffd asssociated with the same mm_struct.
> >
> > Change this behavior to be stricter by requiring VMAs to be unregistered
> > through the same uffd they were registered with.
> >
> > While at it, correct the comment for the no userfaultfd case. This seems
> > to be a copy-paste artifact from the analagous userfaultfd_register()
> > check.
> >
> > Fixes: 86039bd3b4e6 ("userfaultfd: add new syscall to provide memory externalization")
> > Signed-off-by: Tal Zussman <tz2294@columbia.edu>
>
> Thanks, Tal! I like this patch, but I can't really meaningfully
> comment on if it's worth it to change the UAPI.
>
> > ---
> >  fs/userfaultfd.c | 15 +++++++++++++--
> >  1 file changed, 13 insertions(+), 2 deletions(-)
> >
> > diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
> > index 22f4bf956ba1..9289e30b24c4 100644
> > --- a/fs/userfaultfd.c
> > +++ b/fs/userfaultfd.c
> > @@ -1477,6 +1477,16 @@ static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
> >                 if (!vma_can_userfault(cur, cur->vm_flags, wp_async))
> >                         goto out_unlock;
> >
> > +               /*
> > +                * Check that this vma isn't already owned by a different
> > +                * userfaultfd. This provides for more strict behavior by
> > +                * preventing a VMA registered with a userfaultfd from being
> > +                * unregistered through a different userfaultfd.
> > +                */
> > +               if (cur->vm_userfaultfd_ctx.ctx &&
> > +                   cur->vm_userfaultfd_ctx.ctx != ctx)
> > +                       goto out_unlock;
> > +
>
> Very minor nitpick: I think this check should go above the
> !vma_can_userfault() check above, as `wp_async` was derived from
> `ctx`, not `cur->vm_userfaultfd_ctx.ctx`.

Thanks, this is a good point! I'll change it for v2.

This also seems to indicate that the current behavior is broken and may reject
unregistering some VMAs incorrectly. For example, a file-backed VMA registered
with `wp_async` and UFFD_WP cannot be unregistered through a VMA that does not
have `wp_async` set.

> >                 found = true;
> >         } for_each_vma_range(vmi, cur, end);
>
> I don't really like this for_each_vma_range() for loop, but I guess it
> is meaningful to the user: invalid unregistration attempts will fail
> quickly instead of potentially making some progress. So unfortunately,
> without a good reason, I suppose we can't get rid of it. :(
>
> >         BUG_ON(!found);
> > @@ -1491,10 +1501,11 @@ static int userfaultfd_unregister(struct userfaultfd_ctx *ctx,
> >                 cond_resched();
> >
> >                 BUG_ON(!vma_can_userfault(vma, vma->vm_flags, wp_async));
> > +               BUG_ON(vma->vm_userfaultfd_ctx.ctx &&
> > +                      vma->vm_userfaultfd_ctx.ctx != ctx);
>
> IMO, this new BUG_ON should either be
> (1) moved and should not be a BUG_ON. See the WARN_ON_ONCE() below,
> OR
> (2) removed.
>
> Perhaps the older BUG_ON() should be removed/changed too.

I added this mainly to maintain symmetry with the userfaulfd_register()
implementation. I'm happy to leave it out, so  I'll either convert it, and
the other one, to a VM_WARN_ON_ONCE(), as per David, or remove it.

> >
> >                 /*
> > -                * Nothing to do: this vma is already registered into this
> > -                * userfaultfd and with the right tracking mode too.
> > +                * Nothing to do: this vma is not registered with userfaultfd.
> >                  */
> >                 if (!vma->vm_userfaultfd_ctx.ctx)
> >                         goto skip;
>
> if (WARN_ON_ONCE(vmx->vm_userfaultfd_ctx.ctx != ctx)) {
>     ret = -EINVAL;
>     break;
> }
>
> where the WARN_ON_ONCE() indicates that the VMA should have been
> filtered out earlier. The WARN_ON_ONCE() isn't even really necessary.
>
>
> >
> > --
> > 2.39.5
> >
> >