Bug report: 455.23.04 - Kernel Panic due to NULL pointer dereference

Most likely because it was introduced in that version.

Because “not using hardware acceleration” option in one userspace program does not reliably prevent any particular piece of driver’s functionality from being used, especially in a modern desktop that uses compositing for everything. Also the problem is probably in the implementation of some basic functionality, most likely a race condition in something very common. The number of calls may affect the likeliness of a crash, however it can’t be eliminated entirely.

If a guess made by @generix is correct, and preemption is either the necessary condition or it greatly increases the probability of a crash being triggered, it would strongly indicate a race condition.