Re: [PATCH 3/3] x86_64: offset apicid_to_node before use it before init_cpu_to_node

[Andi Kleen <ak@suse.de>]

On Sunday 22 July 2007 02:49:41 Yinghai Lu wrote:
> [PATCH 3/3] x86_64: offset apicid_to_node before use it before init_cpu_to_node
> 
> When acpi=off or there is no SRAT defined, apicid_to_node is got from K8
> Northbridge PCI configuration space in k8_scan_nodes() in
> arch/x86_64/mm/k8toplogy.c.
> The problem is that it assumes bsp apic id is 0 at that point.
> For four socket system with Quad core cpus installed, all cpus apic id
> is offset by 4, and bsp apic id is 4.
> For eight socket system with dual core cpus installed, all cpus apic id
> is offset by 2, and bsp apic id is 2.
> We need offset apicid_to_node array according to boot_cpu_id.--- bsp apic id.
> before we use apicid_to_node array.
> boot_cpu_id is only valid init_apic_mappings.

<rant>
This thing is getting more and more messy. If it gets any more complicated
I promise I'll rip out the non ACPI support for quad core NUMA completely
and let it require ACPI.  Even the people who have a religious problem
with ACPI will need to eventually get over it and LinuxBIOS just has
to create proper tables, not pile hacks over hacks. It probably was a mistake in 
the first place to add it.
</rant>

I don't think you can mess with apicid_to_node[] unconditionally here. 
e.g. for the ACPI case or for the Intel NUMA case you'll just break everything.

What you should do is split init_apic_mappings() up and do a early
call that just checks if the CPU has an APIC and maps it using the 
fixmap and reads boot_cpu_id.  Then you can use that information
in k8topology.c to create correct tables.

-Andi