Badari Pulavarty wrote:

>Am I missing something fundamental here ?
>
>  
>
No, you are right. The current implementation is just wrong(tm). 
Attached is a patch - partially tested.

Description:
kmem_cache_alloc_node allocates memory from a particular node. The patch 
fixes two problems with the current implementation:
- for !CONFIG_NUMA, kmem_cache_alloc_node is identical to kmalloc. The 
patch implements kmem_cache_alloc_node as an alias to kmalloc for 
!CONFIG_NUMA. Right now, the special node aware code runs even on 
non-NUMA systems.
- checks the slab lists instead of allocating a new slab for every 
allocation. This reduces the internal fragmentation.

Badri - could you test the patch?

Andrew, please do not merge the patch yet: it contains a severe bug: if 
a node doesn't contain any memory, then it livelocks because the loop 
never finds a suitable slab. I must think about that case.

--
    Manfred