On Wed, 2018-07-18 at 20:11 -0700, Paul E. McKenney wrote:
> 
> > That is interesting. As I replied to Paul, we are already calling
> > rcu_user_enter/exit() on guest_enter/exit_irqsoff(). So I'm wondering why
> > you're seeing such an optimization by repeating those calls.
> > 
> > Perhaps the rcu_user_* somehow aren't actually called from
> > __context_tracking_enter()...? Some bug in context tracking?
> > Otherwise it's a curious side effect.
> 
> David is working with v4.15.  Is this maybe something that has changed
> since then?

To clarify: in 4.15 without CONFIG_PREEMPT and without NO_HZ_FULL I was
seeing RCU stalls because a thread in vcpu_run() was *never* seen to go
through a quiescent state. Hence the change to need_resched() in the
first patch in this thread, which fixed the problem at hand and seemed
to address the general case.

It then seemed by *inspection* that the NO_HZ_FULL case was probably
broken, because we'd failed to spot the rcu_user_* calls. But
rcu_user_enter() does nothing in the !NO_HZ_FULL case, so wouldn't have
helped in the testing that we were doing anyway.