Re: [RFC][PATCH 9/7] mm: fix mm_take_all_locks() locking order

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Thu, 2008-08-07 at 23:46 +0200, Andrea Arcangeli wrote:

> As for 8/7 you know my opinion from somebody who's way beyond the
> point: check_deadlock should be dropped 

I'll try again one more time, don't feel obliged to reply or
anything :-)

Suppose you have objects initialized from a single site:

struct my_obj *create_obj()
{
	...
	spin_lock_init(&obj->lock);
	...

	return obj;
}


So that all these object's their locks are in a single class.

Now, put these objects into two lists without fixed order.

L1: A B C D
L2: B A D C

If you were to lock-couple your way through these lists there is a
deadlock potential.

The check_noncircular stuff will not catch this due to it all being of a
single class. The only thing we have to indicate you need to pay
attention is check_deadlock.

Yes, check_deadlock is a tad too sensitive in the amount of false
positives, but without it there are gaping holes in 'proving' lock
correctness.

Currently - if you get 100% code coverage and lockdep doesn't shout,
you're good. If we were to drop check_deadlock we can't say that
anymore.