Re: [GIT PULL rcu/next] rcu commits for 2.6.40

[Frederic Weisbecker <fweisbec@gmail.com>]

On Thu, May 19, 2011 at 02:45:52PM -0700, Yinghai Lu wrote:
> On 05/19/2011 02:15 PM, Frederic Weisbecker wrote:
> > On Thu, May 19, 2011 at 12:51:26PM -0700, Yinghai Lu wrote:
> >> [   85.194797]   <idle>-0       0d... 85152953us : rcu_irq_enter: 1 49
> >> [   85.194808]   <idle>-0       0dN.. 85153081us : <stack trace>
> >> [   85.194809]  => rcu_irq_exit
> >> [   85.194810]  => irq_exit
> >> [   85.194811]  => smp_apic_timer_interrupt
> >> [   85.194812]  => apic_timer_interrupt
> >> [   85.194813]  => cpu_idle
> >> [   85.194813]  => rest_init
> >> [   85.194814]  => start_kernel
> >> [   85.194815]  => x86_64_start_reservations
> > 
> > So it's the dynticks_nesting going crazy. As if we had rcu_irq_enter()
> > without rcu_irq_exit().
> > 
> > I see you have CONFIG_DEBUG_LOCKING_API_SELFTESTS=y
> > and the locking api selftest is doing a strange thing: if we look
> > at HARDIRQ_ENTER(), it calls irq_enter(), but HARDIRQ_EXIT() only
> > calls __irq_exit(). Which means it lacks the rcu_irq_exit().
> > 
> > So, if I understood correctly the thing there, some selftests simulating
> > the hardirq context are unbalancing the rcu state.
> > 
> > Does that help if you unset CONFIG_DEBUG_LOCKING_API_SELFTESTS ?
> 
> 
> yes, after unset that, no warning and delay...

Thanks a lot for all your testing! I've just proposed a fix, let's see
how that goes.