Re: [PATCH v13 2/8] mm: rust: add vm_area_struct methods that require read access

[Alice Ryhl <aliceryhl@google.com>]

On Wed, Feb 5, 2025 at 8:22 PM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
>
> * Alice Ryhl <aliceryhl@google.com> [250205 14:13]:
> > On Wed, Feb 5, 2025 at 8:10 PM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
> > >
> > > * Alice Ryhl <aliceryhl@google.com> [250205 10:24]:
> > > > On Wed, Feb 5, 2025 at 3:38 PM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
> > > > >
> > > > > * Alice Ryhl <aliceryhl@google.com> [250205 07:10]:
> > > > > > On Tue, Feb 4, 2025 at 4:46 PM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
> > > > > > > > > > > > +        let vma = unsafe { bindings::vma_lookup(self.mm.as_raw(), vma_addr) };
> > > > > > > > > > > > +
> > > > > > > > > > > > +        if vma.is_null() {
> > > > > > > > > > > > +            None
> > > > > > > > > > > > +        } else {
> > > > > > > > > > > > +            // SAFETY: We just checked that a vma was found, so the pointer is valid. Furthermore,
> > > > > > > > > > > > +            // the returned area will borrow from this read lock guard, so it can only be used
> > > > > > > > > > > > +            // while the mmap read lock is still held.
> > > > > > > > > > >
> > > > > > > > > > > So We have complicated the locking of the vmas with rcu and per-vma
> > > > > > > > > > > locking recently.  We are now able to look up and use a vma under the
> > > > > > > > > > > rcu read lock.  Does this translate to rust model?
> > > > > > > > > > >
> > > > > > > > > > > I believe this is true in recent version of binder as well?
> > > > > > > > > >
> > > > > > > > > > Yes. The safety requirements of VmAreaRef is that you must hold the
> > > > > > > > > > mmap read lock *or* the vma read lock while you have a VmAreaRef
> > > > > > > > > > reference. This particular method achieves that requirement by holding
> > > > > > > > > > the mmap read lock. But there is also a Rust lock_vma_under_rcu(), see
> > > > > > > > > > patch 4 for that.
> > > > > > > > >
> > > > > > > > > Right, okay.  Thanks.  You can get the reference by only holding the rcu
> > > > > > > > > read lock, but you should hold the vma lock to ensure that the vma
> > > > > > > > > itself (and not just the pointer) is safe to use.
> > > > > > > >
> > > > > > > > Hmm... To modify the vma, you must hold the mmap *and* vma write lock,
> > > > > > > > so holding the mmap read lock prevents mutations?
> > > > > > >
> > > > > > > Sorry, I think I confused things with my answer.  Your code is fine.
> > > > > > > The phrasing of the "only be used while the mmap read lock is still
> > > > > > > held" made me wonder about further clarification on the locking here
> > > > > > > (because the locking is confusing).
> > > > > > >
> > > > > > > Yes, mmap read lock means there are no writers that can modify the vma.
> > > > > > > Essentially, you are using the lock to ensure the entire vma space isn't
> > > > > > > changed during your operation - which is heavier than just locking one
> > > > > > > vma.
> > > > > >
> > > > > > I could extend the safety comment like this:
> > > > > >
> > > > > > SAFETY: We just checked that a vma was found, so the pointer is valid.
> > > > > > Furthermore, the returned area will borrow from this read lock guard,
> > > > > > so it can only be used while the mmap read lock is still held. This
> > > > > > ensures that there are no writers because writers must hold both the
> > > > > > mmap and vma write lock.
> > > > >
> > > > > How about just changing the last part to:
> > > > >
> > > > > Furthermore, the returned vma is still under the protection of the read
> > > > > lock guard and can be used while the mmap read lock is still held.
> > > >
> > > > Well, the important part here is that you can't do this:
> > > >
> > > > let guard = mm.mmap_read_lock();
> > > > let vma = guard.vma_lookup(...)?;
> > > > drop(guard);
> > > > vma.foo();
> > > >
> > > > since that would use the vma after the lock has been released. The
> > > > reason that the above is prevented is because `vma` borrows from
> > > > `guard`, so you can only use `vma` while `guard` is still valid.
> > > >
> > >
> > > But it implies that this isn't valid:
> > >
> > >  let guard = mm.mmap_read_lock();
> > >  let vma = guard.vma_lookup(...)?;
> > >
> > >  vma_lock(vma);
> > >
> > >  drop(guard);
> > >  vma.foo();
> > >
> > >  vma_unlock(vma);
> > >
> > > See mm/userfaultfd.c:uffd_lock_vma(), which falls back to mmap read lock
> > > to do this if rcu lock + lock_vma_under_rcu() is unable to lock the vma.
> >
> > This patchset does not have the functionality for doing that, but it's
> > definitely possible to add.
> >
>
> I don't think that's necessary right now.  It's just that I read that
> comment and it seemed to imply something that isn't strictly true with
> the "only valid" part.  I think?  Is rust doing something that makes it
> true?

The code would look like:

let guard = mm.mmap_read_lock();
let vma = guard.vma_lookup(...)?;
let vma_guard = vma.vma_lock();
drop(guard); // this is mmap_read_unlock()
vma_guard.foo();
drop(vma_guard); // this is vma_unlock()

So you couldn't use vma after drop(guard), but vma_guard would be
usable. (And of course the drop calls also happen automatically at the
end of the scope if you do not drop it explicitly.)

Alice