Re: [PATCH tip/core/rcu 55/55] powerpc: Work around tracing fromdyntick-idle mode

[Frederic Weisbecker]

On Wed, Sep 07, 2011 at 06:44:00AM -0700, Paul E. McKenney wrote:
> On Wed, Sep 07, 2011 at 07:00:22AM -0300, Benjamin Herrenschmidt wrote:
> > On Tue, 2011-09-06 at 11:00 -0700, Paul E. McKenney wrote:
> > > PowerPC LPAR's __trace_hcall_exit() can invoke event tracing at a
> > > point where RCU has been told that the CPU is in dyntick-idle mode.
> > > Because event tracing uses RCU, this can result in failures.
> > > 
> > > A correct fix would arrange for RCU to be told about dyntick-idle
> > > mode after tracing had completed, however, this will require some care
> > > because it appears that __trace_hcall_exit() can also be called from
> > > non-dyntick-idle mode.
> > 
> > This obviously needs to be fixed properly. hcall tracing is very useful
> > and if I understand your patch properly, it just comments it out :-)
> 
> That is exactly what it does, and I completely agree that this patch
> is nothing but a short-term work-around to allow my RCU tests to find
> other bugs.
> 
> > I'm not sure what the best approach is, maybe have the hcall tracing
> > test for the dyntick-idle mode and skip tracing in that case ?
> 
> Another approach would be to update Frederic Weisbecker's patch at:
> 
> 	https://lkml.org/lkml/2011/8/20/83
> 
> so that powerpc does tick_nohz_enter_idle(false), and then uses
> rcu_enter_nohz() explicitly just after doing the hcall tracing.
> If pseries is the only powerpc architecture requiring this, then
> the argument to tick_nohz_enter_idle() could depend on the powerpc
> sub-architecture.

I'm trying to fix this but I need a bit of help to understand the
pseries cpu sleeping.

In pseries_dedicated_idle_sleep(), what is the function that does
the real sleeping? Is it cede_processor()?

> 
> The same thing would be needed for tick_nohz_exit_idle() and
> rcu_exit_nohz(): powerpc would need to invoke rcu_exit_nohz() after
> gaining control from the hypervisor but before doing its first tracing,
> and then it would need the idle loop to to tick_nohz_exit_idle(false).
> Again, if pseries is the only powerpc architecture requiring this,
> the argument to tick_nohz_exit_idle() could depend on the architecture.
> 
> Would this approach work?

Sounds like we really need that.

Thanks.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/