> My patch calls generic unregister() and register() of the base
> classes, but it only uses one syscall per modify() call.

Doesn't that mean doing 3 operations (unregister(), register(), modify()) instead of the current 2 (unregister(), register())? I don't see how it can be faster than a single modify() syscall.

> Oh by the way, it seems like my patch changes KqueueSelector, 
> but your change doesn't. Am I right?

You are right but it looks like you end up doing the same thing as unregister() and register(). kqueue() has no modify() method so I don't think it can benefit from this change.