Re: linux-next 20111025: warnings in rcu_idle_exit_common()/rcu_idle_enter_common()

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Tue, Nov 01, 2011 at 06:34:29PM +0100, Frederic Weisbecker wrote:
> On Mon, Oct 31, 2011 at 04:26:34PM +0800, Wu Fengguang wrote:
> > Hi Paul,
> > 
> > I got two warnings in rcutree.c. The last working kernels are
> > linux-next 20111014 and linux v3.1.
> > 
> > [    0.194593] ------------[ cut here ]------------
> > [    0.194707] lockdep: fixing up alternatives.
> > [    0.194730]  #2
> > [    0.194731] smpboot cpu 2: start_ip = 97000
> > [    0.195737] WARNING: at /c/wfg/linux-next/kernel/rcutree.c:444 rcu_idle_exit_common+0xd2/0x117()
> > [    0.196325] Hardware name:         
> > [    0.196603] Modules linked in:
> > [    0.196899] Pid: 0, comm: kworker/0:0 Not tainted 3.1.0-ioless-full-next-20111025+ #881
> > [    0.197459] Call Trace:
> > [    0.197699]  <IRQ>  [<ffffffff81074534>] warn_slowpath_common+0x85/0x9d
> > [    0.201075]  [<ffffffff81074566>] warn_slowpath_null+0x1a/0x1c
> > [    0.201438]  [<ffffffff810d5afd>] rcu_idle_exit_common+0xd2/0x117
> > [    0.201812]  [<ffffffff810d5fff>] rcu_irq_enter+0x75/0xa2
> > [    0.202160]  [<ffffffff8107ac7f>] irq_enter+0x1b/0x74
> > [    0.202496]  [<ffffffff8106f29e>] scheduler_ipi+0x5e/0xd5
> > [    0.202845]  [<ffffffff8104ce6b>] smp_reschedule_interrupt+0x2a/0x2c
> > [    0.203229]  [<ffffffff8198bb73>] reschedule_interrupt+0x73/0x80
> > [    0.203598]  <EOI>  [<ffffffff8198661f>] ? notifier_call_chain+0x63/0x63
> > [    0.204030]  [<ffffffff8103ce2b>] ? mwait_idle+0xef/0x175
> > [    0.204378]  [<ffffffff8103ce22>] ? mwait_idle+0xe6/0x175
> > [    0.204727]  [<ffffffff810351bb>] cpu_idle+0x91/0xb8
> > [    0.205068]  [<ffffffff81978bd5>] start_secondary+0x1de/0x1e2
> > [    0.205454] ---[ end trace 4eaa2a86a8e2da22 ]---
> 
> I'm seeing something similar but on my boot CPU.
> 
> The problem is that idle_cpu() gives a false negative due to the following
> check:
> 
> 
>         if (!llist_empty(&rq->wake_list))
>                 return 0;
> 
> When a task gets enqueued for waking, we call the scheduler
> IPI, but since we call irq_enter() -> rcu_irq_enter() before
> that wakee gets processed and flushed from the wake_list,
> this is not a right condition to look at in order to know if
> we are idle.

OK, that could explain the otherwise-mystifying results Wu Fengguang
just sent -- "No, this is not the idle task, but it has the same
PID and command line!"  ;-)

And idle_cpu() does seem to have grown a bit recently.  Hmmm...

Perhaps I should add something like the following and call it from
RCU's dyntick-idle code path?  Thomas, Peter, seem reasonable?

/**
 * cpu_is_running_idle_task - is a given cpu running its idle task?
 * @cpu: the processor in question.
 */
int cpu_is_running_idle_task(int cpu)
{
	return cpu_curr(cpu) == cpu_rq(cpu)->idle;
}