On Fri, May 03, 2013 at 07:04:47PM +0200, Peter Zijlstra wrote:

quoted

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?

This happens in only two cases:

1.	CONFIG_PREEMPT_RCU=n kernels.  But in this case, rcu_read_unlock()
	and rcu_read_lock() are free, at least for CONFIG_PROVE_LOCKING=n
	kernels.  And if you have CONFIG_PROVE_LOCKING=y, any contribution
	from rcu_read_unlock() and rcu_read_lock() will be in the noise.

2.	CONFIG_DEBUG_ATOMIC_SLEEP=y kernels -- but in this case, you
	-want- the debugging.

So either the overhead is non-existent, or you explicitly asked for the
resulting debugging.

In particular, CONFIG_PREEMPT_RCU=y kernels have an empty static inline
function, which is free -- unless CONFIG_DEBUG_ATOMIC_SLEEP=y, in which
case you again explicitly asked for the debugging.

So I do not believe that the extra rcu_read_unlock()/lock() pairs are a
problem in any of the possible combinations of configurations.

							Thanx, Paul