Re: [PATCH] debug work struct cancel deadlocks with lockdep

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

On Thu, 2007-07-05 at 10:58 +0200, Johannes Berg wrote:
> On Thu, 2007-07-05 at 10:53 +0200, Ingo Molnar wrote:
> 
> > > +#ifdef CONFIG_LOCKDEP
> > > +/*
> > > + * HACK! This really should call lockdep_init_map() but can't
> > > + * because there's no requirement to initialise work structs
> > > + * at runtime. This works because subclass == 0.
> > > + *
> > > + * NB: because we have to copy the lockdep_map, setting .key
> > > + * here is required!
> > > + */
> > 
> > why do you consider this a hack? A static object is a static object, and 
> > its own address is its key. That's what we have for like 80% of all the 
> > spinlocks in the kernel. Static initialization is not as flexible as 
> > dynamic initialization, but the lockdep engine handles it. Am i missing 
> > something?
> 
> Well, there's nothing in lockdep that guarantees that. I'd be much more
> comfortable doing that when lockdep had a STATIC_LOCKDEP_MAP_INIT()
> macro that looks like my __WORK_INIT_LOCKDEP_MAP() macro because then
> people changing lockdep would see that they cannot rely on
> lockdep_init_map() having been called (unless subclass != 0)

You could of course make this STATIC_LOCKDEP_MAP_INIT() and place it
near lockdep_init_map() :-)

That way it would be clear that changes to either ought to be reflected
in the other.