Re: 48GB NUMA-Q boots, with major IO-APIC hassles

["Martin J. Bligh" <mbligh@aracnet.com>]

>> (3) setup_ioapic_ids_from_mpc() panic()'s.
>> -- the clustered_apic_mode check and/or its current equivalent
>> -- no longer suffices with 16 IO-APIC's. Turn off all the
>> -- renumbering logic and hardcode the numbers to alternate
>> -- between 13 and 14, where they belong.
>> -- The real issue here is that the phys_id_present_map is not
>> -- properly per- APIC bus. The physid's of IO-APIC's are
>> -- irrelevant from the standpoint of the rest of the kernel,
>> -- but are inexplicably used to identify them throughout the
>> -- rest of arch/i386/ when physids are nothing resembling
>> -- unique identifiers in multiple APIC bus systems. This
> 
> I also have a problem with setup_ioapic_ids_from_mpc(). I opt for 0xFF as
> max io_apic phys_id (and leave it alone!), because even though we have fewer
> IO-APICs than that, I'd like to keep the actual numbers from MP table or
> ACPI, because all APIC and IO-APIC id's on ES7000 are 8 bit, unique, and
> meaningful (used as a bitmaps) when I have to implement CPU, PCI hot plug
> and dynamic partitioning (I hate to think of possible confusing tables and
> dependencies I will have to maintain otherwise...). 
> 
> Could this routine be made with alternative architecturally private path (as
> a hook or with a hook inside)?

I don't think changing the Linux data structures is a problem, but you
need to be really careful not to change anything for normal machines
when writing out to / reading from the IO-APIC - that stuff is too fragile, 
and breaks on strange machines in wierd ways. 

If you can find a clean way to change the internal stuff, and just wrap 
the in/out interfaces, that would seem best to me ...

M.