Christoph Lameter wrote:

>The -1 is optimized away for the non NUMA case. In the NUMA case its an
>additional parameter that is passed to kmem_cache_alloc. So its one
>additional register load that allows us to not have an additional function
>for the case non node specific allocations.
>  
>
Correct, I was thinking about the NUMA case.
You've decided to add one register load to every call of kmalloc. On 
i386, kmalloc_node() is a 24-byte function. I'd bet that adding the node 
parameter to every call of kmalloc causes a .text increase larger than 
240 bytes. And I have not yet considered that you have increased the 
number of conditional branches in every kmalloc(32,GFP_KERNEL) call by 
33%, i.e. from 3 to 4 conditional branch instructions.
I'd add an explicit kmalloc_node function. Attached is a prototype 
patch. You'd have to reintroduce the flags field to 
kmem_cache_alloc_node() and update kmalloc_node.
The patch was manually edited, I hope it applies to a recent tree ;-)

What do you think?
--
    Manfred