Re: AMD topology broken on various 754/AM2+/AM3/AM3+ systems causes NB/EDAC/GART regression since 6.14

[Yazen Ghannam]

On Mon, Nov 03, 2025 at 06:12:45PM +0100, Michal Pecio wrote:
> On Mon, 3 Nov 2025 09:38:51 -0500, Yazen Ghannam wrote:
> > > I have this AM4 system with some proprietary HP BIOS:
> > > 
> > > [02Fh 0047 001h]               Local Apic ID : 10
> > > [037h 0055 001h]               Local Apic ID : 11
> > > [03Fh 0063 001h]               Local Apic ID : 12
> > > [047h 0071 001h]               Local Apic ID : 13
> > > 
> > > domain: Thread     shift: 0 dom_size:     1 max_threads:     1
> > > domain: Core       shift: 4 dom_size:    16 max_threads:    16
> > > domain: Module     shift: 4 dom_size:     1 max_threads:    16
> > > domain: Tile       shift: 4 dom_size:     1 max_threads:    16
> > > domain: Die        shift: 4 dom_size:     1 max_threads:    16
> > > domain: DieGrp     shift: 4 dom_size:     1 max_threads:    16
> > > domain: Package    shift: 4 dom_size:     1 max_threads:    16
> > > 
> > > It seems that pkgid is 0x1 here, which is not a problem because
> > > it's single socket, but dunno if HP or somebody else couldn't do
> > > similar things in an 8-socket system and end up with pkgid > 8.
> > >   
> > 
> > So is this another bogus case?
> 
> No, it isn't bogus. It's a quad-core Carrizo APU with its four LAPICs,
> but their numbers start from 0x10 rather than 0x00. And AFAIU, the
> calculated pkgid vaule will be 1.
> 

Oh I see. That's interesting.

Though something still feels off about it, because these systems never
had 16 threads/cores.

> If HP or other vendor would do similar thing on an 8-socket system,
> the assumption that (pkgid < 8) could no longer hold, even if the CPUs
> are completely real.

The proposed check only applies to older systems. Were there any
8-socket AMD systems in production? I'm only aware of systems up to 4
sockets from those days. Hence the arbitrary value check.

Now there are systems with 8 "AMD nodes". However, before Zen these
would be 2 AMD nodes per socket. With 4 sockets you'd get 8 AMD nodes.
Naples had 4 AMD nodes per socket, but the max system had 2 sockets.

But I think I see your point that this check would break if the pkgid
values can start from an arbitrary number. If such a case is reported,
then this could be a new quirk or rework of the existing quirks.

Thanks,
Yazen