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

[Bjorn Helgaas]

On Wed, Nov 07, 2018 at 10:18:38AM +0100, Borislav Petkov wrote:
> On Tue, Nov 06, 2018 at 05:20:41PM -0600, Bjorn Helgaas wrote:
> > Or maybe even drivers/acpi/thermal.c, which claims every Thermal Zone
> > (ACPI 6.2, sec 11), would be sufficient.  I don't know what the
> > relationship between hwmon and other thermal stuff, e.g.,
> > Documentation/thermal/sysfs-api.txt is.  acpi/thermal.c looks tied
> > into the drivers/thermal stuff (it registers "thermal_zone" devices),
> > but not to hwmon.
> 
> Err, I still don't think I'm catching your drift but let me stop you
> right there: amd_nb is not there only for hwmon/k10temp. It is a small
> interface glue if you will, which exports the CPU functionality in PCI
> config space to other consumers.
> 
> So it is not really a driver - it is used by drivers to talk/query CPU
> settings through it.
> 
> With that said, I don't think I understand all that talk about PNP IDs
> and ACPI methods. But maybe I'm missing something...

Firmware supplies ACPI namespace.  The namespace contains an abstract
description of the platform, including devices.  Devices are
identified by PNP IDs, which are analogous to PCI vendor/device IDs,
except that a device may have several generic "compatible device IDs"
in addition to an ID unique to the device.  Devices may also contain
methods (supplied by firmware as part of the namespace), which are
essentially bytecode that can be executed by the ACPI interpreter in
the kernel.  Linux drivers claim ACPI devices based on PNP ID and
operate them using either ACPI methods (which can decouple the driver
from device specifics) or the usual direct MMIO/IO port/MSR style.

Here's an outline of how it *could* work:

  - AMD defines "AMD0001" device ID for the CPU temp sensor
  - BIOS supplies AMD0001 devices in the ACPI namespace
  - Each AMD0001 device has a _TMP method (supplied by BIOS and
    specific to the CPU)
  - Linux driver claims AMD0001 devices
  - Driver reads temp sensors by executing _TMP methods (Linux ACPI
    interpreter runs the bytecode)

That way when you release a new platform with different temp sensors,
you update the BIOS AMD0001 devices and _TMP methods to know about
them, and the old Linux driver works unchanged.

Bjorn