Re: [PATCH v2 3/3] userfaultfd: use per-vma locks in userfaultfd operations

[Lokesh Gidra <lokeshgidra@google.com>]

On Mon, Feb 5, 2024 at 2:00 PM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
>
> * Lokesh Gidra <lokeshgidra@google.com> [240205 16:55]:
> ...
>
> > > > > We can take care of anon_vma as well here right? I can take a bool
> > > > > parameter ('prepare_anon' or something) and then:
> > > > >
> > > > >            if (vma) {
> > > > >                     if (prepare_anon && vma_is_anonymous(vma)) &&
> > > > > !anon_vma_prepare(vma)) {
> > > > >                                       vma = ERR_PTR(-ENOMEM);
> > > > >                                       goto out_unlock;
> > > > >                    }
> > > > > >                 vma_aquire_read_lock(vma);
> > > > >            }
> > > > > out_unlock:
> > > > > >         mmap_read_unlock(mm);
> > > > > >         return vma;
> > > > > > }
> > > >
> > > > Do you need this?  I didn't think this was happening in the code as
> > > > written?  If you need it I would suggest making it happen always and
> > > > ditch the flag until a user needs this variant, but document what's
> > > > going on in here or even have a better name.
> > >
> > > I think yes, you do need this. I can see calls to anon_vma_prepare()
> > > under mmap_read_lock() protection in both mfill_atomic_hugetlb() and
> > > in mfill_atomic(). This means, just like in the pagefault path, we
> > > modify vma->anon_vma under mmap_read_lock protection which guarantees
> > > that adjacent VMAs won't change. This is important because
> > > __anon_vma_prepare() uses find_mergeable_anon_vma() that needs the
> > > neighboring VMAs to be stable. Per-VMA lock guarantees stability of
> > > the VMA we locked but not of its neighbors, therefore holding per-VMA
> > > lock while calling anon_vma_prepare() is not enough. The solution
> > > Lokesh suggests would call anon_vma_prepare() under mmap_read_lock and
> > > therefore would avoid the issue.
> > >
>
> ...
>
> > anon_vma_prepare() is also called in validate_move_areas() via move_pages().
>
> Probably worth doing it unconditionally and have a comment as to why it
> is necessary.
>
The src_vma (in case of move_pages()) doesn't need to have it.

The only reason I'm not inclined to make it unconditional is what if
some future user of lock_vma() doesn't need it for their purpose? Why
allocate anon_vma in that case.

> Does this avoid your locking workaround?

Not sure which workaround you are referring to. I am almost done
implementing your suggestion. Very soon will share the next version of
the patch-set.
>
> Thanks,
> Liam