So I wasted considerably more than the warranted amount of time on this, and 
what I found, at least in one case, is a deadlock.  It goes this like this:

1) One thread takes a page fault in pthread_create().  When this happens, it 
holds pthread__deadqueue_lock.

2) A sequence of upcalls is pushed on the upcall stacks.  The first one to get 
into pthread__sched_idle2() wedges there, waiting for 
pthread__deadqueue_lock.

3) The upcall chaining code spins forever trying to get the lock, preventing 
the thread that's in pthread_create() from running and unlocking the queue.

"Oopsie."