Christoph Lameter wrote:
> On Wed, 18 May 2005, Christoph Lameter wrote:
> 
> 
>>Fixes to the slab allocator in 2.6.12-rc4-mm2
>>- Remove MAX_NUMNODES check
>>- use for_each_node/cpu
>>- Fix determination of INDEX_AC
> 
> 
> Rats! The whole thing with cpu online and node online is not as easy as I 
> thought. There may be bugs in V3 of the numa slab allocator 
> because offline cpus and offline are not properly handled. Maybe 
> that also contributed to the ppc64 issues. 
>  
> The earlier patch fails if I boot an x86_64 NUMA kernel on a x86_64 single 
> processor system.
> 
> Here is a revised patch. Would be good if someone could review my use 
> of online_cpu / online_node etc. Is there some way to bring cpus 
> online and offline to test if this really works? Seems that the code in 
> alloc_percpu is suspect even in the old allocator because it may have
> to allocate memory for non present cpus.

I ran this patch on top of rc4-mm2 w/ my topology fix (to get it to
compile) and booting it resulted in this log.  Looks like it's hitting a
bug in mempolicy code?  My intuition tells me this is because Andrew
changed PPC64 to have a NODES_SHIFT of 0 (and hence a MAX_NUMNODES of 1),
so I'm retrying w/ that patch reverted.  I'll poke around to see it
anything else comes up, but this seems promising, since I got a completely
different panic with your old patch (and the NODES_SHIFT set to 0).

-Matt