Re: [RFC][PATCH 7/7] lockdep: spin_lock_nest_lock()

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Thursday 09 October 2008 02:27, Steven Rostedt wrote:
> On Mon, Aug 04, 2008 at 12:31:22PM -0700, Linus Torvalds wrote:
> > On Mon, 4 Aug 2008, Jeremy Fitzhardinge wrote:
> > > OK.  I don't actually need to do this, but I was asking for
> > > completeness.  But to clarify, you only need to do the reverse unlock
> > > if you do it after unlocking the outer lock?  If you're still holding
> > > the outer lock, you can unlock in any order?
> >
> > Release order should always be totally irrelevant, whether you hold outer
> > locks or not. Only the order of _getting_ locks matter.
>
> Technically, you are 100% correct.
>
> > And yes, if there is an outer lock, even the order of getting locks is
> > irrelevant, as long as anybody who gets more than one inner lock always
> > holds the outer one.
>
> But I need to disagree on a programming practice style.  Unlocking locks
> in a non nested order is just bad programming practice. Unless there is
> a good reason to do so, one should never release locks in a non reverse
> order they were taken.

An outer one might be more likely to be contended, so you might want
to release it asap.

Other times, you have lock A and lock B held (like scheduler rqs).
You can say unlock(A); unlock(B);
or if (A < B) unlock(B); unlock(A); if (A > B) unlock (B);


> This can be a source of bugs, where people might notice an outer lock
> being released and think the inner locks were too.
>
> Lately the kernel has been going through a lot of clean ups that have
> been making the kernel a much more maintainable beast. I feel we should
> enforce the rule of unlocking order (again, unless there is a good
> reason not to). Not for a technical reason, but just for a more
> maintainable one.

I don't really think it would make things more maintainable, FWIW.