Re: processes hung after sys_renameat, and 'missing' processes

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2012-06-08 at 10:46 -0400, J. Bruce Fields wrote:
> > > > Sadly, if you get that annotation wrong you can annotate an actual
> > > > deadlock away.
> 
> What's a (contrived as you want) example where that happens?

spinlock_t lock_array[10];

void init_array(void)
{
	int i;

	for (i = 0; i < ARRAY_SIZE(lock_array); i++)
		spin_lock_init(array + i);
}

void double_lock(int a, int b)
{
	spin_lock(lock_array + a);
	spin_lock_nested(lock_array + b, SINGLE_DEPTH_NESTING);
}

The above places all locks in the array in the same class, it then does
a double lock without order, but tells lockdep the nesting is ok.

A correct version of the double_lock() function would look like:

void double_lock(int a, int b)
{
	if (b < a)
		swap(a, b);

	spin_lock(lock_array + a);
	spin_lock_nested(lock_array + b, SINGLE_DEPTH_NESTING);
}

This orders the locks in array order.

> > > > This the reason you have to be very careful when
> > > > annotating stuff.
> 
> Or alternatively--what do I need to check before I call
> mutex_lock_nested? 

That the lock order you tell lockdep is ok, is indeed correct.