On Thu, Jul 12, 2018 at 05:53:51AM -0700, Paul E. McKenney wrote:
> On Thu, Jul 12, 2018 at 01:00:42PM +0100, David Woodhouse wrote:
> > On Wed, 2018-07-11 at 14:08 -0700, Paul E. McKenney wrote:
> > >
> > > > Also... why in $DEITY's name was the existing
> > > > rcu_virt_note_context_switch() not actually sufficient? If we had that
> > > > there, why did we need an additional explicit calls to rcu_all_qs() in
> > > > the KVM loop, or the more complex fixes to need_resched() which
> > > > ultimately had the same effect, to avoid ten-second latencies?
> > >
> > > My guess is that this was because control passed through the
> > > rcu_virt_note_context_switch() only once, and then subsequent
> > > scheduling-clock interrupts bypassed this code.
> Gah! My guess was instead that the code did a rcu_kvm_enter() going in,
> but somehow managed to miss the rcu_kvm_exit() going out. But that makes
> absolutely no sense -- had that happened, rcutorture would likely have
> screamed bloody murder, loudly and often. No mere near misses!
> And besides, thus far, -ENOREPRODUCE. :-/
OK, one close call in 63 hours of rcutorture, this one on scenario TREE03
(yesterday hit TREE01 and TREE03). Time for probabilitistic long-runtime
bisection. Plus thought about how to get more information out of the near
misses. Fun! ;-)
> Which indicates that I have an opportunity to improve rcutorture and
> that this patch was with high probability an innocent bystander.
> > > But that is just a guess.
> > > I need to defer to someone who understands the KVM code better than I do.
> > I think it's more likely that we just never happened at all. It's
> > conditional. From the latest patch iteration (see it being removed):
> > @@ -118,12 +118,12 @@ static inline void guest_enter_irqoff(void)
> > * one time slice). Lets treat guest mode as quiescent state, just
> > like
> > * we do with user-mode execution.
> > */
> > - if (!context_tracking_cpu_is_enabled())
> > - rcu_virt_note_context_switch(smp_processor_id());
> > + rcu_kvm_enter();
> > }
> > Given the vmexit overhead, I don't think we can do the currently-
> > proposed rcu_kvm_enter() thing except for CONFIG_NO_HZ_FULL where it's
> > really necessary. I'll make that conditional, but probably on the RCU
> > side.
> > Without CONFIG_NO_HZ_FULL, rcu_kvm_exit() can do nothing, and
> > rcu_kvm_enter() can do rcu_virt_note_context_switch().
> > OK?
> Makes sense to me! And a big "thank you!" to Christian for testing
> and analyzing this in a timely fashion!!!
> Thanx, Paul