Re: [PATCH 2/4] x86/amd_nb: add support for newer PCI topologies

[Bjorn Helgaas <helgaas@kernel.org>]

[+cc Takashi, Andy, Colin, Myron for potential distro impact]

[Beginning of thread:
https://lore.kernel.org/linux-pci/20181102181055.130531-1-brian.woods@amd.com/]

On Sat, Nov 03, 2018 at 12:29:48AM +0100, Borislav Petkov wrote:
> On Fri, Nov 02, 2018 at 02:59:25PM -0500, Bjorn Helgaas wrote:
> > This isn't my code, and I'm not really objecting to these changes, but
> > from where I sit, the fact that you need this sort of vendor-specific
> > topology discovery is a little bit ugly and seems like something of a
> > maintenance issue.  

I think this is the most important part, and I should have elaborated
on it instead of getting into the driver structure details below.

It is a major goal of ACPI and PCI that an old kernel should work
unchanged on a new platform unless it needs to use new functionality
introduced in the new platform.

amd_nb.c prevents us from achieving that goal.  These patches don't
add new functionality; they merely describe minor topographical
differences in new hardware.  We usually try to do that in a more
generic way, e.g., via an ACPI method, so the new platform can update
the ACPI method and use an old, already-qualified, already-shipped
kernel.

I'm not strenuously objecting to these because this isn't a *huge*
deal, but I suspect it is a source of friction for distros that don't
want to update and requalify their software for every new platform.

> > You could argue that this is sort of an "AMD CPU
> > driver", which is entitled to be device-specific, and that does make
> > some sense.
> 
> It is a bunch of glue code which enumerates the PCI devices a CPU
> has and other in-kernel users can use that instead of doing the
> discovery/enumeration themselves.
> 
> > But device-specific code is typically packaged as a driver that uses
> > driver registration interfaces like acpi_bus_register_driver(),
> > pci_register_driver(), etc.  That gives you a consistent structure
> > and, more importantly, a framework for dealing with hotplug.  It
> > doesn't look like amd_nb.c would deal well with hot-add of CPUs.
> 
> If you mean physical hotadd, then that's a non-issue as, AFAIK, AMD
> doesn't support that.
> 
> Now, TBH I've never tried soft-offlining the cores of a node and then
> check whether using the PCI devices of that node would work.
> 
> Now, I don't mind this getting converted to a proper PCI driver as long
> as it is not a module as it has to be present at all times. Other than
> that, I'm a happy camper.

amd_nb.c uses pci_get_device(), which is incompatible with hotplug and
subverts the usual driver/device ownership model.  We could pursue
this part of the conversation, but I think it's more fruitful to
approach this from the "new machine, old kernel" angle above.