OffscreenSurfaceHelper accesses AbstractRenderer which (I believe) lives in a rendering thread. Surface creation is done via a queued slot connection but there is no synchronisation to whether the renderer is still valid. It seems safer to queue the parameters instead.

The proof-of-concept patch is attached. I've verified it fixes the crash for us. It's not particularly formatted according to guidelines (I can do it if the patch approach is approved as-is).

Unfortunately I don't have a minimal example (it only happens under a heavy load, only on certain machines, and only when there are plenty of other business logic threads).