The key point is that I don't understand why we cannot get the effect
we are looking for with the following in sched.h (or wherever):

static inline int cond_resched_rcu(void)
{
#if defined(CONFIG_DEBUG_ATOMIC_SLEEP) || !defined(CONFIG_PREEMPT_RCU)
	rcu_read_unlock();
	cond_resched();
	rcu_read_lock();
#endif
}

This adds absolutely no overhead in non-debug builds of CONFIG_PREEMPT_RCU,
does the checking in debug builds, and allows voluntary preemption in
!CONFIG_PREEMPT_RCU builds.  CONFIG_PROVE_RCU builds will check for an
(illegal) outer rcu_read_lock() in CONFIG_PREEMPT_RCU builds, and you
will get "scheduling while atomic" in response to an outer rcu_read_lock()
in !CONFIG_PREEMPT_RCU builds.

It also seems to me a lot simpler.

Does this work, or am I still missing something?

It can do quite a number of superfluous rcu_read_unlock()/lock() pairs for
voluntary preemption kernels?