Re: [PATCH RFC tip/core/rcu] Eliminate deadlock between CPU hotplug and expedited grace periods

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Sep 02, 2014 at 09:36:56AM -0700, Paul E. McKenney wrote:
> On Mon, Sep 01, 2014 at 06:17:35PM +0200, Peter Zijlstra wrote:
> > On Mon, Sep 01, 2014 at 09:05:50AM -0700, Paul E. McKenney wrote:
> > > > URGH.. I really hate that. The hotplug interface is already too
> > > > horrible, we should not add such hacks to it.
> > > 
> > > We do have try_ interfaces to a number of other subsystems, so I don't
> > > believe that it qualifies as such a hack.
> > 
> > We do indeed, but I'm not sure about adding this to the hotplug stuff.
> 
> Looks pretty straightforward to me.
> 
> > Also; not really understanding the problem doesn't help. 
> 
> The current implementation of synchronize_sched_expedited()
> calls get_online_cpus().  Some of the ACPI code needs to hold the
> acpi_ioremap_lock mutex across synchronize_sched_expedited(), and
> also needs to acquire this same mutex from a CPU hotplug notifier.
> This results in deadlock between the cpu_hotplug.lock mutex and the
> acpi_ioremap_lock mutex.


  acpi_ioremap_lock			cpu_hotplug_begin()
    synchronize_sched()			  acpi_ioremap_lock
      get_online_cpus()

So yes, AB-BA.

> Normal RCU grace periods avoid this by synchronizing on a lock acquired by
> the RCU CPU-hotplug notifiers, but this does not work for the expedited
> grace periods because the outgoing CPU can be running random tasks for
> quite some time after RCU's notifier executes.  So the fix is just to
> drop back to a normal grace period when there is a CPU-hotplug operation
> in progress.

So why are we 'normally' doing an expedited call here anyhow? 

> > > > How about ripping that rcu_expedited stuff out instead? That's all
> > > > conditional anyhow, so might as well not do it.
> > > 
> > > In what way is the expedited stuff conditional?
> > 
> > synchronize_sched() conditionally calls synchronize_sched_expedited()
> > and its condition: rcu_expedited, gets set/cleared on pm notifiers and
> > nr_cpu_ids.
> 
> There are also direct calls to both synchronize_sched_expedited() and
> synchronize_rcu_expedited().

But those are not within hotplug bits. Also weren't we removing them? I
thought we didn't appreciate spraying IPIs like they do?