Re: [PATCH] cpumask: fix lg_lock/br_lock.

[Ingo Molnar <mingo@elte.hu>]

* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> On Wed, 2012-02-29 at 09:29 +0100, Ingo Molnar wrote:
>
> > As part of any cleanup they should first be converted from 
> > arch_spinlock_t to regular spinlock_t - I bet if that is 
> > done then that not only simplifies the wrappers massively, 
> > it also turns the above soft lockup report into a nice, 
> > actionable lockdep splat.
> 
> It might help if you'd actually read the code.. that will 
> simply not work.

It cannot find all bugs - such as the CPU hotplug race that is 
still present in the code.

Still there's no excuse to go outside regular spinlock debug 
primitives via arch_spinlock_t.

If lockdep blows up in br_write_lock() due to holding up to 4096 
individual locks then we should add the exceptions to this 
particular write lock when the CPU count is too high - but:

 - do not disable the checking on saner configs

 - not disable all the *OTHER* lock debugging checks such as:

     - spin-lockup detection [this works even without ->held_locks]

     - allocate/free failure detection:

         The percpu code could be extended to run the equivalent 
         of debug_check_no_locks_freed() over the percpu area 
         that is going away, to make sure no held locks are 
         freed.

etc.

Thanks,

	Ingo