Re: [GIT PULL] SLAB changes for v3.10

[Tony Lindgren <tony@atomide.com>]

* Christoph Lameter <cl@linux.com> [130508 12:06]:
> On Wed, 8 May 2013, Chris Mason wrote:
> 
> > > You correctly moved the checks out of the if (!kmalloc_cacheS())
> > > condition so that the caches are created properly.
> >
> > But if the ordering is required at all, why is it ok to create cache 2
> > after cache 6 instead of after cache 7?
> 
> The power of two caches are 2^x beginning with KMALLOC_MIN_SHIFT. The non
> power of two caches were folded into number 1 + 2 since they do not fit
> into the scheme and they are special cased throughout. This works since
> the minimal slab cache size is 8 bytes.
> 
> > IOW if we can safely do cache 2 after cache 6, why can't we just do both
> > cache 1 and cache 2 after the loop?
> 
> Because the cache creation in SLAB can cause the use of a fractional slab
> size if kmem_cache_create() thinks its better to put the metadata on a
> different slab cache (OFF_SLAB type) because data will align better that
> way. Its weird I know but its due to the way that SLAB aligns data in the
> page frame.

Hmm OK so kmalloc_caches[7] got created earlier with INDEX_AC != INDEX_NODE,
and those are defined as:

#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))

So the different sizes for the structs can trigger it like Pekka was
speculating earlier.

Regards,

Tony