Re: How to turn scheduler tick on for current nohz_full CPU?

[Paul E. McKenney]

On Wed, Jul 24, 2019 at 04:30:13PM +0200, Frederic Weisbecker wrote:
> On Wed, Jul 24, 2019 at 06:52:19AM -0700, Paul E. McKenney wrote:
> > On Wed, Jul 24, 2019 at 03:22:59PM +0200, Frederic Weisbecker wrote:
> > > On Wed, Jul 24, 2019 at 04:53:31AM -0700, Paul E. McKenney wrote:
> > > > Hello!
> > > > 
> > > > One of the callback-invocation forward-progress issues turns out to
> > > > be nohz_full CPUs not turning their scheduling-clock interrupt back on
> > > > when running in kernel mode.  Given that callback floods can cause RCU's
> > > > callback-invocation loop to run for some time, it would be good for this
> > > > loop to re-enable this interrupt.  Of course, this problem applies to
> > > > pretty much any kernel code that might loop for an extended time period,
> > > > not just RCU.
> > > > 
> > > > I took a quick look at kernel/time/tick-sched.c and the closest thing
> > > > I found was tick_nohz_full_kick_cpu(), except that (1) it isn't clear
> > > > that this does much when invoked on the current CPU and (2) it doesn't
> > > > help in rcutorture TREE04.  In contrast, disabling NO_HZ_FULL and using
> > > > RCU_NOCB_CPU instead works quite well.
> > > > 
> > > > So what should I be calling instead of tick_nohz_full_kick_cpu() to
> > > > re-enable the current CPU's scheduling-clock interrupt?
> > > 
> > > Indeed, kernel code is assumed to be quick enough (between two extended grace
> > > periods) to avoid running the tick for RCU. But some long lasting kernel code
> > > may require to tick temporarily.
> > > 
> > > You can use tick_nohz_dep_set_cpu(cpu, TICK_DEP_MASK_RCU) with the
> > > following:
> > > 
> > > diff --git a/include/linux/tick.h b/include/linux/tick.h
> > > index f92a10b5e112..3f476e2a4bf7 100644
> > > --- a/include/linux/tick.h
> > > +++ b/include/linux/tick.h
> > > @@ -108,7 +108,8 @@ enum tick_dep_bits {
> > >  	TICK_DEP_BIT_POSIX_TIMER	= 0,
> > >  	TICK_DEP_BIT_PERF_EVENTS	= 1,
> > >  	TICK_DEP_BIT_SCHED		= 2,
> > > -	TICK_DEP_BIT_CLOCK_UNSTABLE	= 3
> > > +	TICK_DEP_BIT_CLOCK_UNSTABLE	= 3,
> > > +	TICK_DEP_BIT_RCU		= 4
> > >  };
> > >  
> > >  #define TICK_DEP_MASK_NONE		0
> > > @@ -116,6 +117,7 @@ enum tick_dep_bits {
> > >  #define TICK_DEP_MASK_PERF_EVENTS	(1 << TICK_DEP_BIT_PERF_EVENTS)
> > >  #define TICK_DEP_MASK_SCHED		(1 << TICK_DEP_BIT_SCHED)
> > >  #define TICK_DEP_MASK_CLOCK_UNSTABLE	(1 << TICK_DEP_BIT_CLOCK_UNSTABLE)
> > > +#define TICK_DEP_MASK_RCU		(1 << TICK_DEP_BIT_RCU)
> > >  
> > >  #ifdef CONFIG_NO_HZ_COMMON
> > >  extern bool tick_nohz_enabled;
> > 
> > I will give this a try, thank you!  (Testing will take a few days.)
> 
> Sure!
> 
> For the background: expect a self-IPI to fire which then restart the tick on IRQ exit.
> Once you're later done with the work, don't forget to remove the tick dependency:
> 
>      tick_nohz_dep_clear_cpu(cpu, TICK_DEP_MASK_RCU);

Does this get cleared automatically if preempted, or should I clear
this is one of RCU's scheduler hooks?  (My guess is the latter, but
I figured I should ask.)

							Thanx, Paul