Re: [patch 2/3] NUMA slab locking fixes - move irq disabling from cahep->spinlock to l3 lock

[Ravikiran G Thirumalai <kiran@scalex86.org>]

On Tue, Feb 07, 2006 at 09:36:40AM +0200, Pekka J Enberg wrote:
> On Mon, 6 Feb 2006, Andrew Morton wrote:
> > > This is getting scary.  Manfred, Christoph, Pekka: have you guys taken a
> > > close look at what's going on in here?
> 
> On Mon, 6 Feb 2006, Christoph Lameter wrote:
> > I looked at his patch and he seems to be right. Most of the kmem_cache 
> > structure is established at slab creation. Updates are to the debug 
> > counters and to nodelists[] during node online/offline and to array[] 
> > during cpu online/offline. The chain mutex is used to protect the 
> > setting of the tuning parameters. I still need to have a look at the 
> > details though.
> 
> The patch looks correct but I am wondering if we should keep the spinlock 
> around for clarity? The chain mutex doesn't really have anything to do 
> with the tunables, it's there to protect the cache chain. I am worried 
> that this patch makes code restructuring harder. Hmm?

IMHO, if you keep something around which is not needed, it might later get
abused/misused.  And what would you add in as comments for the
cachep->spinlock?  

Instead,  bold comments on cachep structure stating what all members are 
protected by which lock/mutex should be sufficient no?

Thanks,
Kiran