Re: [PATCH 3/4] rust: Add dma_fence abstractions

[Alice Ryhl <aliceryhl@google.com>]

On Tue, Jun 02, 2026 at 01:52:29PM +0200, Philipp Stanner wrote:
> On Tue, 2026-06-02 at 11:44 +0000, Alice Ryhl wrote:
> > On Tue, Jun 02, 2026 at 01:31:52PM +0200, Philipp Stanner wrote:
> > > On Mon, 2026-06-01 at 13:44 +0000, Alice Ryhl wrote:
> > > > On Mon, Jun 01, 2026 at 03:14:05PM +0200, Philipp Stanner wrote:
> > > > > +Cc Boris
> > > > > 
> > > > > On Mon, 2026-06-01 at 14:55 +0200, Alice Ryhl wrote:
> > > > > > On Mon, Jun 1, 2026 at 2:34 PM Philipp Stanner <phasta@mailbox.org> wrote:
> > > > > > > 
> > > > > > > On Mon, 2026-06-01 at 10:20 +0000, Alice Ryhl wrote:
> > > > > > > > On Sat, May 30, 2026 at 03:06:21PM +0000, sashiko-bot@kernel.org wrote:
> > > > > > > > > > +impl<F: Send + Sync, C: Send + Sync> DriverFence<F, C> {
> > > > > > > > > > +    fn as_raw(&self) -> *mut bindings::dma_fence {
> > > > > > > > > > +        // SAFETY: Valid because `self` is valid.
> > > > > > > > > > +        let fence_data = unsafe { &mut *self.data.as_ptr() };
> > > > > > > > > 
> > > > > > > > > [Severity: High]
> > > > > > > > > Does this create an exclusive mutable reference to actively shared memory?
> > > > > > > > > 
> > > > > > > > > DriverFenceData can be accessed concurrently by other threads holding Fence
> > > > > > > > > references (for instance, when checking if the fence is signaled). Creating
> > > > > > > > > a mutable reference (&mut) in Rust asserts exclusive access and violates
> > > > > > > > > aliasing rules, which allows the compiler to make invalid optimization
> > > > > > > > > assumptions.
> > > > > > > > > 
> > > > > > > > > Could this use an immutable reference &*self.data.as_ptr() instead?
> > > > > > > > 
> > > > > > > > Yes, please use an immutable reference here.
> > > > > > > > 
> > > > > > > > > > +
> > > > > > > > > > +        fence_data.inner.inner.get()
> > > > > > > > > > +    }
> > > > > > > > > 
> > > > > > > > > [ ... ]
> > > > > > > > > 
> > > > > > > > > > +impl<F: Send + Sync, C: Send + Sync> DriverFenceBorrow<F, C> {
> > > > > > > > > > +    fn as_raw(&self) -> *mut bindings::dma_fence {
> > > > > > > > > > +        // SAFETY: Valid because `self` is valid.
> > > > > > > > > > +        let fence_data = unsafe { &mut *self.data.as_ptr() };
> > > > > > > > > 
> > > > > > > > > [Severity: High]
> > > > > > > > > Similar to DriverFence::as_raw(), does this also incorrectly create a
> > > > > > > > > mutable reference to shared data?
> > > > > > > > 
> > > > > > > > Here as well.
> > > > > > > 
> > > > > > > `data` is not shared. By design there is only ever one DriverFence, and
> > > > > > > the driver's data (`data.data`) is `Sync`.
> > > > > > > 
> > > > > > > But I guess an immutable one should do the trick, too.
> > > > > > 
> > > > > > There's only one DriverFence, but I can perform shared access to that
> > > > > > one DriverFence from two threads in parallel. You made the type Sync,
> > > > > > and this is what you are allowing when you do so.
> > > > > 
> > > > > Nope, DriverFence is just Send, not Sync.
> > > > > 
> > > > > data.data is Sync, but `data` in the code above is not the actual user
> > > > > data, but the memory backing it up.
> > > > 
> > > > Ok, well, it probably should be Sync. I don't see any &self methods that
> > > > can't be called from multiple threads in parallel.
> > > 
> > > No. Why?
> > > 
> > > There can only be one owner of DriverFence.
> > 
> > Values normally have only one owner, and they are also normally Sync.
> > 
> > > Regarding the backend_ops that might access DriverFence::data, we take
> > > care to guard that.
> > > 
> > > And DriverFence itself can only be owned by one party. That can only be
> > > accessed by many threads if the driver stuffs that fence into something
> > > that implements Sync.
> > 
> > If you don't implement Sync, then DriverFence cannot be stored in an
> > Arc. I wouldn't take away that ability unless you have to, and I don't
> > see anything in the DriverFence API that would mean you can't do that.
> 
> Nope. We explicitly agreed on this design.
> 
> Just 1 DriverFence. Just 1 party that can signal it.
> Note that we also agreed upon the Driverfence disappearing with
> .signal(), which certainly prevents several from existing, unless you
> do an Option.take()

I would like to clarify that I'm not suggesting any changes to the
design. Implementing Sync is not the same as having multiple driver
fences.

> > > > > >  so even though
> > > > > > the fence context may be valid for another grace period, the *pointer*
> > > > > > to the fence context is not. The pointer could have been zeroed by the
> > > > > > destructor.
> > > > > 
> > > > > That particular pointer to the DriverFenceData could have been zeroed.
> > > > > But potential other accessors have already crafted themselves a new
> > > > > pointer to the, by the power of RCU, still valid data. That new pointer
> > > > > is container-of-ed from struct dma_fence *f.
> > > > 
> > > > I'm not talking about the pointer to DriverFenceData, I'm talking about
> > > > the pointer to the FenceCtx, or the pointer to the data (if F is
> > > > RcuBox).
> > > 
> > > Yeah, but the backing memory is still alive. And new pointers to that
> > > memory get crafted by the accessors. If a callback accesses the data
> > > through `container_of(Fence)`, it gets a new pointer.
> > > 
> > > So what's the problem?
> > > 
> > > Where is the invalid pointer that someone is accessing?
> > > 
> > > > 
> > > > The Arc type is not a type that opts-out of &mut == exclusive, so the
> > > > second drop_in_place() above is assumed exclusive access to the
> > > > Arc<FenceCtx<F,C>> field.
> > > 
> > > OK, so I think I see the problem. So the invalid pointer is
> > > Arc<FenceCtx…>? And potentially the <F> pointer (although we don't have
> > > a picture yet as to how that would be accessed through other callbacks.
> > > 
> > > >  If another thread obtains a pointer to the
> > > > FenceCtx via reading the fctx field of the DriverFence in parallel with
> > > > this, then that's not allowed because the drop_in_place() call has
> > > > exclusive access to that field.
> > > 
> > > I think I have been asking in several of our meetings in the past
> > > whether it is actually a problem to access data that has been dropped()
> > > IF we know that drop does not cause UAF and the answer was kind of like
> > > a "well if it does not actually get freed…"
> > 
> > Ok, well, IMO the simplest approach is to say you can't. There may be
> > roundabout ways to do it, but I would suggest that we just ... don't.
> 
> Ack.
> 
> > 
> > > Anyways.
> > > 
> > > It would seem the way to get this right is then
> > > 
> > > synchronize_rcu();
> > > drop_in_palace(data);
> > > 
> > > 
> > > Agreed?
> > > 
> > > This would then mean, however, that every time a fence drops, you have
> > > to wait a grace period.
> > > 
> > > Or maybe stuff DriverFenceData into an RcuBox, too, and defer its
> > > dropping.
> > 
> > That would work, but I think we can do better and avoid the
> > synchronize_rcu() along these lines:
> > 
> > unsafe trait RcuRevocable {
> >     unsafe fn rcu_revoke_in_place(ptr: *mut Self);
> > }
> > 
> > This trait provides a method that's like drop_in_place(), except that
> > when you use this destructor, the value remains usable for one grace
> > period. You could implement it for RcuBox, and for any Copy type, and
> > for ARef<T> when T is cleaned up with rcu, and probably also other
> > stuff.
> 
> I mean, this cannot be magic. It also boils down to executing one RCU
> callback per DriverFence dropping.
> 
> Is there a significant difference to stuffing DriverFenceData into an
> RcuBox?

Do you mean hard-coding that the user-data of a driver fence is always
stored in an RcuBox?

Alice