Re: [PATCH 6/8] netpoll: Allow netpoll_setup/cleanup recursion

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 25 Jun 2010 10:08:56 +0200 Peter Zijlstra <peterz@infradead.org> wrote:

> On Thu, 2010-06-24 at 21:42 -0700, Andrew Morton wrote:
> > That being said, I wonder why Herbert didn't hit this in his testing. 
> > I suspect that he'd enabled lockdep, which hid the bug.  I haven't
> > worked out _why_ lockdep hides the double-mutex_unlock bug, but it's a
> > pretty bad thing to do. 
> 
> Most weird indeed, lockdep is supposed so shout its lungs out when
> someone wants to unlock a lock that isn't actually owned by him (and it
> not being locked at all certainly implies you're not the owner).
> 
> In fact, the below patch results in the below splat -- its also
> something that's tested by the locking self-test:

When I enabled lockdep, the bug actually went away.  Is it possible
that when lockdep detects this bug, it prevents mutex.count from going
from 1 to 2?

It could be that lockdep _did_ detect (and correct!) the bug.  But
because I had no usable console output at the time, I didn't see it.

I did notice that the taint output was "G W".  So something warned
about something, but I don't know what.  But that was happening with
lockdep disabled.


> @@ -1344,6 +1346,10 @@ SYSCALL_DEFINE0(getppid)
>  {
>  	int pid;
>  
> +	mutex_lock(&foo);
> +	mutex_unlock(&foo);
> +	mutex_unlock(&foo);
> +
>  	rcu_read_lock();
>  	pid = task_tgid_vnr(current->real_parent);
>  	rcu_read_unlock();

It'd be interesting to add

	printk("%d:%d\n", __LINE__, atomic_read(&foo.count));

after the mutex_unlock()s.