Re: There is a Tasks RCU stall warning

[Steven Rostedt]

On Wed, 12 Apr 2017 09:27:54 -0700
"Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> wrote:

> On Wed, Apr 12, 2017 at 10:59:37AM -0400, Steven Rostedt wrote:
> > On Wed, 12 Apr 2017 07:48:00 -0700
> > "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> wrote:
> >   
> > > > > > Like this?  (Untested, but builds at least some of the time.)    
> > > > > 
> > > > > Not like that...  :-/  Update on its way.    
> > > > 
> > > > Perhaps more like this.  Started rcutorture on it, will see how it goes.    
> > 
> > I just love the above discussion with yourself ;-)  
> 
> Talking to oneself used to cause passersby to get really bent out of
> shape.  But one big benefit of ubiquitous cellphones is that now
> people just assume that you are talking on a cellphone that they
> cannot see.  ;-)
> 
> > > Do you need this patch?  If so, I should do some more work on it to
> > > eliminate the extra common-case branch on the scheduler fastpath.  
> > 
> > Do I still need this patch? Maybe. :-)
> > 
> > I changed my benchmark test to call cond_resched_rcu_qs() instead and
> > that appears to fix the issue. But I'm not sure if there's any other
> > kthread out there that just calls cond_resched() or schedule().
> > 
> > Actually, I think it is still a good idea to have it. I believe that it
> > will still allow synchronize_rcu_tasks() to progress even if there's a
> > kthread task that is constantly being woken up, and never sleeps when
> > it calls schedule(), as it may always have the R state.  
> 
> OK, will optimize it a bit.  When are you planning to get this in?
> 

Well, I added the use case for synchronize_rcu_tasks() in my current
for-next. I'll have to make sure I get the schedule_idle() in as well
as my update to the event benchmark thread as well.

I don't think anything will truly break without it yet. But that's
assuming there's not another kernel thread somewhere that just spins
calling schedule.

And this patch will still speed up those that do call
synchronize_rcu_tasks(). But that's an optimization and not really a
fix.

-- Steve