Re: [debug patch] printk: Add a printk killswitch to robustify NMI watchdog messages

[Ingo Molnar <mingo@elte.hu>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Sun, 2011-06-05 at 22:15 +0200, Arne Jansen wrote:
> > 
> > Can lockdep just get confused by the lockdep_off/on calls in printk
> > while scheduling is allowed? There aren't many users of lockdep_off().
> 
> Yes!, in that case lock_is_held() returns false, triggering the warning.
> I guess there's an argument to be made in favour of the below..
> 
> ---
>  kernel/lockdep.c |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/kernel/lockdep.c b/kernel/lockdep.c
> index 53a6895..e4129cf 100644
> --- a/kernel/lockdep.c
> +++ b/kernel/lockdep.c
> @@ -3242,7 +3242,7 @@ int lock_is_held(struct lockdep_map *lock)
>  	int ret = 0;
>  
>  	if (unlikely(current->lockdep_recursion))
> -		return ret;
> +		return 1; /* avoid false negative lockdep_assert_held */
>  
>  	raw_local_irq_save(flags);
>  	check_flags(flags);

Oh, this explains the full bug i think.

lockdep_off() causes us to not track pi_lock, and thus the assert 
inside printk() called try_to_wake_up() triggers incorrectly.

The reason why Arne triggered it is probably because console_lock 
*wakeups* from printk are very, very rare: almost nothing actually 
locks the console. His remote system probably has some VT-intense 
application (screen?) that hits console_lock more intensely.

Arne, do you use some vt-intense application there?

The real fix might be to remove the lockdep_off()/on() call from 
printk(), that looks actively evil ... we had to hack through several 
layers of side-effects before we found the real bug - so it's not 
like the off()/on() made things more robust!

So i think what we want to apply is the lockdep_off()/on() removal, 
once Arne has it tested.

Thanks,

	Ingo