Re: Regression: WARNINGS and lockdep spews in 2.6.38-rc3+ (bisected).

[Peter Zijlstra <peterz@infradead.org>]

On Thu, 2011-02-03 at 19:42 +0800, Yong Zhang wrote:
> My patch is inspired by this http://lkml.org/lkml/2010/8/16/291

Right, so that splat warns about:

spin_lock_irqsave(&shost->host_lock);
del_timer_sync()
  spin_lock_irqsave(&t->split_timeout_timer);

vs

<SOFTIRQ>
  call_timer_fn()
    spin_lock(&t->split_timeout_timer);
    <IRQ>
      spin_lock_irqsave(&shost->host_lock);

Which Steven just pointed out is an actual deadlock because:


CPU0					CPU1
spin_lock_irqsave(&shost->host_lock);
					<SOFTIRQ>
					  call_timer_fn();
					    base->running_timer = timer;
					    func();
					      <IRQ>
						spin_lock(&shost->host_lock);
del_timer_sync()
  while(base->running_timer == timer);

And the world stops turning..


Also note that your patch changes the above to:

spin_lock_irqsave(&shost->host_lock);
del_timer_sync()
  spin_lock_bh(&t->split_timeout_timer);

Which wouldn't make one bit of difference.