Re: check && lockdep_no_validate (Was: lockdep: Introduce wait-type checks)

[Oleg Nesterov <oleg@redhat.com>]

On 01/16, Alan Stern wrote:
>
> On Thu, 16 Jan 2014, Peter Zijlstra wrote:
>
> > On Thu, Jan 16, 2014 at 06:43:48PM +0100, Oleg Nesterov wrote:
>
> > > Perhaps we should change the meaning of lockdep_set_novalidate_class?
> > > (perhaps with rename). What do you think about the patch below?
> > >
> > > With this patch __lockdep_no_validate__ means "automatically nested",
> >
> > Yes, I suppose that might work, it would allow some validation.
>
> I haven't seen the patch, but I'm not so sure it will work.  Suppose we
> have two devices, D1 and D2, and some other mutex, M.  Then the locking
> pattern:
>
> 	lock(D1);
> 	lock(M);
> 	unlock(M);
> 	unlock(D1);
>
> generally should not conflict with:
>
> 	lock(M);
> 	lock(D2);
> 	unlock(D2);
> 	unlock(M);

Yes, sure. This change assumes that the only problem in drivers/base is
dev->parent->mutex / dev->mutex dependency. If the locking is even more
"broken" (wrt lockdep), we can't replace lockdep_set_novalidate_class()
with lockdep_set_auto_nested().

And, otoh, with this change lockdep can miss the real problems too, for
example:

	func1(dev)
	{
		device_lock(dev->parent);
		mutex_lock(MUTEX);
		device_lock(dev);
		...
	}

	func2(dev)
	{
		device_lock(dev);
		mutex_lock(MUTEX);
		...
	}

lockdep will only notice dev -> MUTEX dependency.

I booted the kernel (under kvm) with this change and there is nothing
in dmesg, but of course this is not the real testing.

So do you think that dev->mutex should not be validated at all ?


Just in case... Of course, if we actually add auto_nested we should not
use a single class unless dev->mutex will be the only user.

Oleg.