On Linux you can still easily go extremely slow with plain sched_yield() as there's no guarantee the thread who holds the lock will get to run, especially in nu

On Linux you can still easily go extremely slow with plain sched_yield() as there's no guarantee the thread who holds the lock will get to run, especially in numa situations due to a strong bias against migrating threads across numa boundaries. You have to use the e.g. futex API to tell the kernel what you're waiting for. Bare spinlocks are very bad in this way, unless you have dedicated hardware threads (with SMT you can pin all but one of the threads with the remaining thread being general purpose, though you may want a way to inform the scheduler which cores to currently prefer due to their pinned tasks being more idle).