On 6/29/26 9:56 AM, Philipp Stanner wrote:
The commit mentioned in the fixes tag below introduced a mechanism through which fence producers can fully decouple from fence consumers. This, desirable, mechanism is based on the fence's signaled-bit as the "decoupling point".
A sophisticated interaction between RCU and atomic instructions attempts to ensure that fence consumers can still interact with fence producers through the dma_fence_ops (callback pointers into the producer).
This is the desired behavior: to check for decoupling, the signaled-bit is first checked. If it's not yet signaled, RCU ensures that the ops pointer cannot yet be NULL.
Hereby, dma_fence_signal_timestamp_locked() first sets the signaled-bit, and then sets the ops pointer to NULL. Readers first load the ops pointer, and then check through the signaled-bit whether the pointer can legally be accessed.
These set and load operations could occur out of order on weakly ordered platforms. This problem can be solved very elegantly by using the ops pointer itself as the synchronization point. The pointer is either NULL, or cannot become NULL while it is being used thanks to RCU.
Replace the signaled-bit check in dma_fence_timeline_name() and dma_fence_driver_name().
Cc: stable@vger.kernel.org Fixes: f4cc3ab824d6 ("dma-buf: protected fence ops by RCU v8") Signed-off-by: Philipp Stanner phasta@kernel.org
Reviewed-by: Danilo Krummrich dakr@kernel.org