Re: [RFC] Make need_resched() return true when rcu_urgent_qs requested

[Frederic Weisbecker]

On Thu, Jul 19, 2018 at 09:20:33AM +0200, David Woodhouse wrote:
> On Thu, 2018-07-19 at 08:45 +0200, Christian Borntraeger wrote:
> > 
> > > My thought would be something like this:
> > >Â
> > >ÂÂÂÂÂÂÂif (context_tracking_cpu_is_enabled())
> > >ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂrcu_kvm_enter();
> > >ÂÂÂÂÂÂÂelse
> > >ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂrcu_virt_note_context_switch(smp_processor_id());
> > 
> > In the past we needed that (when we introduced that). At least with every
> > host interrupt we called this making an rcu event at least every HZ.
> > Will the changes in need_resched make this part unnecessary?
> 
> Yes, the change in need_resched() should make this part unnecessary.
> Unless your architecture's version of the vcpu_run() loop just loops
> forever even when TIF_NEED_RESCHED is set? :)
> 
> I'm not sure about the context tracking condition in the code snippet
> cited above, though. I think that's what caused my problem in the first
> place â I have CONTEXT_TRACKING_FORCE && !NO_HZ_FULL. So in 4.15, that
> means rcu_user_enter() did nothing and rcu_virt_note_context_switch()
> wasn't called. Hence the observed stalls.
> 
> Should rcu_user_enter() itself be conditional on CONTEXT_TRACKING not
> NO_HZ_FULL?Â

Ah, CONTEXT_TRACKING_FORCE is only for testing purpose, you should not select
it, it's going to introduce overhead. Actually I should remove that. Although
since we have removed CONFIG_NOHZ_FULL_ALL it's the last way we have to test
NOHZ_FULL from config alone.