Having same issue. Seems to be a bug.

it seems to me spinlocks are fundamentally dangerous in user space because you cannot disable preemption in user space as you can in kernel space.

how slow was that then?

compare-and-swap instruction itself may not be privileged operation but user space spin-lock seldom work in user-space (it's dangerous at least), because a user thread could be preempted.

What's even more dangerous, on a single core machine or one that is lightly loaded without any aggressive contention for the lock it may even seem to be working.

Ah yes, good point. I had forgotten about that issue.