Re: need_heavy_qs flag for PREEMPT=y kernels

[Joel Fernandes <joel@joelfernandes.org>]

On Thu, Aug 15, 2019 at 02:45:42PM -0700, Paul E. McKenney wrote:
> On Thu, Aug 15, 2019 at 05:22:16PM -0400, Joel Fernandes wrote:
> > On Thu, Aug 15, 2019 at 01:31:07PM -0700, Paul E. McKenney wrote:
> > > On Thu, Aug 15, 2019 at 04:04:32PM -0400, Joel Fernandes wrote:
> > > > On Thu, Aug 15, 2019 at 10:17:14AM -0700, Paul E. McKenney wrote:
> > > > > On Mon, Aug 12, 2019 at 09:02:49PM -0400, Joel Fernandes wrote:
> > > > > > On Mon, Aug 12, 2019 at 04:01:38PM -0700, Paul E. McKenney wrote:
> > > > > 
> > > > > [ . . . ]
> > > > > 
> > > > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > > > > index 8c494a692728..ad906d6a74fb 100644
> > > > > > > --- a/kernel/rcu/tree.c
> > > > > > > +++ b/kernel/rcu/tree.c
> > > > > > > @@ -651,6 +651,12 @@ static __always_inline void rcu_nmi_exit_common(bool irq)
> > > > > > >  	 */
> > > > > > >  	if (rdp->dynticks_nmi_nesting != 1) {
> > > > > > >  		trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2, rdp->dynticks);
> > > > > > > +		if (tick_nohz_full_cpu(rdp->cpu) &&
> > > > > > > +		    rdp->dynticks_nmi_nesting == 2 &&
> > > > > > > +		    rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
> > > > > > > +			rdp->rcu_forced_tick = true;
> > > > > > > +			tick_dep_set_cpu(rdp->cpu, TICK_DEP_MASK_RCU);
> > > > > > > +		}
> > > > > > 
> > > > > > 
> > > > > > Instead of checking dynticks_nmi_nesting == 2 in rcu_nmi_exit_common(), can
> > > > > > we do the tick_dep_set_cpu(rdp->cpu, TICK_DEP_MASK_RCU)  from
> > > > > > rcu_nmi_enter_common() ? We could add this code there, under the "if
> > > > > > (rcu_dynticks_curr_cpu_in_eqs())".
> > > > > 
> > > > > This would need to go in an "else" clause, correct?  But there would still
> > > > > want to be a check for interrupt from base level (which would admittedly
> > > > > be an equality comparison with zero) and we would also still need to check
> > > > > for rdp->rcu_urgent_qs && !rdp->rcu_forced_tick.
> > > > 
> > > > True, agreed. I replied to this before saying it should be
> > > > !rcu_dynticks_curr_cpu_in_eqs() in the "if" ;) But it seems I could also be
> > > > missing the check for TICK_DEP_MASK_RCU in my tree so I think we need this as
> > > > well which is below as diff. Testing it more now!
> > > > 
> > > > And, with this I do get many more ticks during the test. But there are
> > > > intervals where the tick is not seen. Still it is much better than before:
> > > > 
> > > > diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> > > > index be9707f68024..e697c7a2ce67 100644
> > > > --- a/kernel/time/tick-sched.c
> > > > +++ b/kernel/time/tick-sched.c
> > > > @@ -198,6 +198,10 @@ static bool check_tick_dependency(atomic_t *dep)
> > > >  		return true;
> > > >  	}
> > > >  
> > > > +	if (val & TICK_DEP_MASK_CLOCK_RCU) {
> > > > +		return true;
> > > > +	}
> > > > +
> > > >  	return false;
> > > >  }
> > > 
> > > That one is not in my tree, either.  Frederic, should I add this to
> > > your patch?  For that matter, may I add your Signed-off-by as well?
> > > Your original is in my -rcu tree at:
> > > 
> > > 0cb41806c799 ("EXP nohz: Add TICK_DEP_BIT_RCU")
> > > 
> > > I am testing Joel's suggested addition now.
> > 
> > Actually there's more addition needed! I found another thing missing:
> > 
> > There's a per-cpu &tick_dep_mask and a per-cpu ts->tick_dep_mask. It seems
> > RCU is setting the latter.
> 
> As I understand it, tick_dep_mask forces the tick on globally,
> ts->tick_dep_mask forces it on for a specific CPU (which RCU uses when it
> needs a quiescent state from that CPU), current->tick_dep_mask forces
> it on for a specific task (which RCU uses for callback invocation
> and certain rcutorture kthreads), and I don't pretend to understand
> current->signal->tick_dep_mask (the comment says something about POSIX
> CPU timers).

Right. I am pretty new to all of these so I could have something incorrect in
a hurry. But thanks for the explanation of your understanding of these.

Yes this commit talks about timers as well for the signal->tick_dep_mask:
d027d45d8a17 ("nohz: New tick dependency mask")

> But it looks to me that can_stop_full_tick() and check_tick_dependency()
> already cover all of these.  What am I missing?

As you mentioned, all of these are covered.

> > So I added a check for both, below is the diff:
> > 
> > However, I see in some cases that the tick_dep_mask is just 0 but I have to
> > debug that tomorrow if that's an issue on the RCU side of things. For now,
> > below should be the completed Frederick patch which you could squash into his
> > if he's Ok with it:
> > 
> > ---8<-----------------------
> > 
> > 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;
> > diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> > index be9707f68024..a613916cc3f0 100644
> > --- a/kernel/time/tick-sched.c
> > +++ b/kernel/time/tick-sched.c
> > @@ -198,6 +198,11 @@ static bool check_tick_dependency(atomic_t *dep)
> >  		return true;
> >  	}
> >  
> > +	if (val & TICK_DEP_MASK_RCU) {
> > +		trace_tick_stop(0, TICK_DEP_MASK_RCU);
> > +		return true;
> > +	}
> > +
> >  	return false;
> >  }
> >  
> > @@ -208,8 +213,13 @@ static bool can_stop_full_tick(int cpu, struct tick_sched *ts)
> >  	if (unlikely(!cpu_online(cpu)))
> >  		return false;
> >  
> > -	if (check_tick_dependency(&tick_dep_mask))
> > +	if (check_tick_dependency(&ts->tick_dep_mask)) {
> >  		return false;
> > +	}
> > +
> > +	if (check_tick_dependency(&tick_dep_mask)) {
> > +		return false;
> > +	}
> >  
> >  	if (check_tick_dependency(&ts->tick_dep_mask))
> >  		return false;
> 
> You lost me on this one.  How does it help to check ts->tick_dep_mask
> twice?  And why is it important to check it before checking tick_dep_mask,
> especially given that the common case of all-zero masks will cause
> all to be checked anyway?

You are right. In a later reply I had mentioned to you to drop this hunk. It
is not needed.

I will pull your -rcu dev branch now to get the latest and will test the RCU
dyntick code to see if I can make that work with the tick-sched code.

Keep you posted!

thanks,

 - Joel