Re: Making rcu_normal=1 in RT

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Oct 31, 2016 at 07:52:40PM -0700, Paul E. McKenney wrote:
> On Tue, Nov 01, 2016 at 04:36:54AM +0200, Michael S. Tsirkin wrote:
> > On Mon, Oct 31, 2016 at 07:19:41PM -0700, Paul E. McKenney wrote:
> > > On Tue, Nov 01, 2016 at 12:37:04AM +0200, Michael S. Tsirkin wrote:
> > > > On Mon, Oct 31, 2016 at 11:15:43AM -0700, Paul E. McKenney wrote:
> > > > > On Mon, Oct 31, 2016 at 06:38:52PM +0100, Sebastian Andrzej Siewior wrote:
> > > > > > On 2016-10-16 04:28:46 [-0700], Paul E. McKenney wrote:
> > > > > > > If the relevant maintainers are OK with that, I am OK with it as long
> > > > > > > as it is non-default (at least to begin with) and does not introduce
> > > > > > > additional Kconfig questions.  My guess is that a boot parameter would
> > > > > > > work best, but something to discuss.
> > > > > > 
> > > > > > Okay. For me to summary:
> > > > > > - we want rcu_normal_after_boot=1 on -RT via CONFIG_PREEMPT_RT_FULL
> > > > > 
> > > > > That would be good.
> > > > > 
> > > > > > - this makes synchronize_rcu_expedited() behave like
> > > > > >   synchronize_rcu() and therefore I can drop all patches replacing
> > > > > >   synchronize_rcu_expedited() with synchronize_rcu().
> > > > > 
> > > > > And this would be a benefit.  ;-)
> > > > > 
> > > > > > - optionally it has been requested to make synchronize_rcu() behave like
> > > > > >   synchronize_rcu_expedited() on shutdown and kexec().
> > > > > 
> > > > > You can invoke rcu_unexpedite_gp() to force all subsequennt grace periods
> > > > > to be expedited, but you do need to clear rcu_normal for this to have
> > > > > effect.  Right now, that means "WRITE_ONCE(rcu_normal, 0)", but I could
> > > > > easily supply a formal API if you would prefer.  Which you probably
> > > > > would given  that TINY_RCU doesn't have rcu_normal...
> > > > > 
> > > > > > Did I miss understood / forgot something?
> > > > > 
> > > > > It would not hurt to boot with rcu_expedited if boot speed is critical,
> > > > > but I don't know whether or not this should be enabled by default.
> > > > > 
> > > > > 							Thanx, Paul
> > > > 
> > > > I don't know either but I'd like to know ATM it's expedited
> > > > by default.
> > > 
> > > In current mainline, no.  That is, rcu_expedited=0 by default.
> > > 
> > > Or am I missing the point of your question?
> > > 
> > > 							Thanx, Paul
> > 
> > That's true - what I meant is that adding and removing network devices
> > happens on boot and calls synchronize_rcu_expedited,
> > with the commit log explaining that this is done for speeding up boot
> > and shutdown.
> 
> The rcu_normal boot parameter will indeed nullify this optimization.
> The boot parameters are normally dumped into dmesg, which should allow
> determining when this has happened.  But I do still feel like I am not
> fully understanding your concern.
> 
> 							Thanx, Paul

I thought you asked whether people that care about boot time
should just boot with rcu_expedited=1.

I was trying to say that the fact net core calls expedited sync
implies that many people care about boot time.

Should rcu_normal_after_boot override rcu_expedited?
I'm not sure what happens if you specify both ATM.

-- 
MST