Re: [discuss] Re: [patch 1/3] x86_64: Node local PDA -- early cpu_to_node

[Ravikiran G Thirumalai <kiran@scalex86.org>]

On Sat, Dec 03, 2005 at 12:02:06AM +0100, Andi Kleen wrote:
> On Fri, Dec 02, 2005 at 02:51:56PM -0800, Ravikiran G Thirumalai wrote:
> > On Fri, Dec 02, 2005 at 12:43:49PM +0100, Andi Kleen wrote:
> > > And are you sure it will work with k8topology.c. Doesn't look like
> > > that to me.
> > 
> > I don't have a K8 box yet :(, so I cannot confirm either ways.  
> > But I thought newer opterons need to use  ACPI_NUMA instead...
> 
> k8topology still needs to work - e.g. for LinuxBios and users which use
> acpi=off and as a fallback for broken SRAT tables. You can't break it right now.
>

I don't think this breaks K8 per-se, because x86_cpu_to_apicid[] is setup if
acpi is compiled in and k8topology sets up apicid_to_node[] at
k8_scan_nodes. That said, I don't know for sure as I don't have a K8 yet. If
someone can test this patch on a opteron, compiled with
ACPI_NUMA as well as K8, (but which falls back to K8 at boot), 
it will be helpful.
 
> > 
> > Even if K8 detection is used, cpu_pda will have memory allocated from node0
> > which is not different from the current state.  So this patch helps Opterons
> > and EM64t boxes which use ACPI_NUMA, right?  Also the newer opteron boxes
> > and em64t NUMA boxes can now get node local memory for static per-cpu areas.
> 
> Hmm good point. However i would prefer if there was no performance regression
> between the two options. However i guess it can be kept like this now.
> Just make sure to comment it well.

Sure.

Thanks,
Kiran