Re: [PATCH] kmalloc_percpu

[Martin J. Bligh]

>> As just pointed out by dipankar the only issue is NUMA...
>> so it has to be something more sophisticated than simply
>> kmalloc()[smp_processor_id];
> 
> The proposed patch doesn't do anything about that either.
> 
> +	ptr = alloc_bootmem(PERCPU_POOL_SIZE * NR_CPUS);
> 
> So yes, we need an api which could be extended to use node-affine memory at
> some time in the future.  I think we have that.

You can just call alloc_bootmem_node for each CPU instead. It doesn't 
work on i386 at the moment (well, it'll work but come out of node 0), 
but it probably ought to.

M.