However, recursive locking is a bad idea overall, and the tracking needed to make it work on rwlock contexts isn't worth the overhead it imposes in the common case of non-recursive rwlocks.
It would be better if the spec simply said it was disallowed.
And no, I don't want to "high performance" lock implementations that regularly completely deadlock the whole process (deadlocked processes are not very performant) unless wrap every single one of the several hundred uses of them in a test with dubiously-predictable branches, or worse, just completely break the lock invariants (e.g., the lock is now permanently unlocked, even after a lock() on it succeeds) ― it really is not that important how fast you can do the wrong thing.