Re: EXP rcu: Revert expedited GP parallelization cleverness

[Paul E. McKenney]

On Thu, Nov 08, 2018 at 05:58:45PM +0100, Sebastian Andrzej Siewior wrote:
> On 2018-11-01 16:30:31 [-0700], Paul E. McKenney wrote:
> > > (Commit 258ba8e089db23f760139266c232f01bad73f85c from linux-rcu)
> > > 
> > > This commit reverts a series of commits starting with fcc635436501 ("rcu:
> > > Make expedited GPs handle CPU 0 being offline") and its successors, thus
> > > queueing each rcu_node structure's expedited grace-period initialization
> > > work on the first CPU of that rcu_node structure.
> > > 
> > > Suggested-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxx>
> > > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > > 
> > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > > index 0b2c2ad69629..a0486414edb4 100644
> > > --- a/kernel/rcu/tree_exp.h
> > > +++ b/kernel/rcu/tree_exp.h
> > > @@ -472,7 +472,6 @@ static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
> > >  static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
> > >  				     smp_call_func_t func)
> > >  {
> > > -	int cpu;
> > >  	struct rcu_node *rnp;
> > >  
> > >  	trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("reset"));
> > > @@ -494,13 +493,7 @@ static void sync_rcu_exp_select_cpus(struct rcu_state *rsp,
> > >  			continue;
> > >  		}
> > >  		INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
> > > -		preempt_disable();
> > > -		cpu = cpumask_next(rnp->grplo - 1, cpu_online_mask);
> > > -		/* If all offline, queue the work on an unbound CPU. */
> > > -		if (unlikely(cpu > rnp->grphi))
> > > -			cpu = WORK_CPU_UNBOUND;
> > > -		queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
> > > -		preempt_enable();
> > > +		queue_work_on(rnp->grplo, rcu_par_gp_wq, &rnp->rew.rew_work);
> > >  		rnp->exp_need_flush = true;
> > >  	}
> > 
> > How about instead changing the earlier "if" statement to read as follows?
> > 
> > 		if (!READ_ONCE(rcu_par_gp_wq) ||
> > 		    rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
> > 		    rcu_is_last_leaf_node(rnp) ||
> > 		    IS_ENABLED(CONFIG_PREEMPT_RT_FULL)) {
> > 			/* No workqueues yet or last leaf, do direct call. */
> > 			sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
> > 			continue;
> > 		}
> > 
> > This just adds the "|| IS_ENABLED(CONFIG_PREEMPT_RT_FULL)" to the "if"
> > condition.
> > 
> > The advantage of this approach is that it leaves the parallelization
> > alone for mainline, and avoids the overhead of the workqueues for -rt.
> 
> I don't oppose to the workqueue approach. It is just preempt_disable() +
> workqueue don't work on -RT. And if I remember correctly, we can't take
> CPU hotplug lock for other reasons (which woould make the
> preempt_disable() go away). Also the original argument why that patch
> went in was not solid so I though removing the extra complexity would be
> a good thing.

>From what I can see, always using the unbound workqueue can serialize
things on some platforms, which kind of defeats the whole purpose of
using the workqueues in the first place.

> However using sync_rcu_exp_select_node_cpus() (based von v4.20-rc1)
> should work on -RT from what I can see. And performance wise it should
> not matter for -RT because the whole synchronize_.*_expedited() is
> disabled on -RT anyway. So it should be used only during boot-up.

Agreed, which was why I proposed making -RT use the boot-time code path,
given that -RT only uses this code during boot.  ;-)

							Thanx, Paul