[net-next: PATCH 0/8] Armada 7k/8k PP2 ACPI support

[Ard Biesheuvel]

On 21 January 2018 at 16:13, Andrew Lunn <andrew at lunn.ch> wrote:
>> Right. So if you need to have some additional "parameters" with the
>> connection, then I suppose you may want to go with the GenericSerialBus
>> route. However, looking at the sample device tree description:
>>
>>         davinci_mdio: ethernet at 5c030000 {
>>                 compatible = "ti,davinci_mdio";
>>                 reg = <0x5c030000 0x1000>;
>>                 #address-cells = <1>;
>>                 #size-cells = <0>;
>>
>>                 reset-gpios = <&gpio2 5 GPIO_ACTIVE_LOW>;
>>                 reset-delay-us = <2>;
>>
>>                 ethphy0: ethernet-phy at 1 {
>>                         reg = <1>;
>>                 };
>>
>>                 ethphy1: ethernet-phy at 3 {
>>                         reg = <3>;
>>                 };
>>         };
>>
>> would pretty much translate directly to this in ACPI if you don't need
>> any additional attributes:
>>
>>       Device (ETH0) {
>>               Name (_ADR, /* PCI address of the NIC */)
>>
>>               Device (PHY0) {
>>                       Name (_ADR, 1)
>>                       ...
>>               }
>>
>>               Device (PHY1) {
>>                       Name (_ADR, 3)
>>                       ...
>>               }
>>       }
>>
>> which looks pretty simple to me. You can also use _DSM and _DSD here to
>> pass information (like the protocol number) for the PHY devices to Linux.
>
> I'm not particularly worried about that simple case. Other than, i
> don't want people to think that is all that is required.
>
> For a more full example, take a look at vf610-zii-dev-rev-b.dts. The
> Freescale FEC Ethernet controller provides the base MDIO device,
> mdio1. On top of this is an MDIO mux, using a few GPIO lines to
> enable/disable 3 child MDIO busses. Each of these busses has an
> Ethernet Switch. The Ethernet switch exports up to two MDIO busses,
> and on these busses are Ethernet PHYs which are embedded inside the
> switch. The Ethernet switches are also interrupt controllers, with the
> PHYs having interrupt properties which point back to the interrupt
> controller in the switch.
>
> So i'm interested in an ACPI proposal which supports this board.
>

However interesting as an example, I'm not convinced this is what we
should aim for.

ACPI is not a replacement for DT, and it is unlikely that people would
be interested in running ACPI-only distros such as RHEL on their
Ethernet switch. DT is excellent at describing this, and there is no
need to replace it.

ACPI is about firmware abstractions: you don't need to describe every
stacked interrupt controller in minute detail to the OS if the
firmware configures it sufficiently. That way, the OS does not need to
know all these details, and vendors can update their hardware without
having to update the software as well. (Or that is the idea at least,
how that works out in practice on arm64 systems remains to be seen)

Taking the Marvell 8040 as an example: the ACPI description does not
expose the ICU interrupt controllers to the OS, but the firmware
configures them and describes their configured state as ordinary GIC
interrupts.

Also, please bear in mind that the ACPI spec is owned by the UEFI/ACPI
forum, and only members (who are all under a contract regarding
reasonable and non-discriminatory (RAND) licensing terms to IP they
own that is covered by the spec) can contribute. Individuals can
become members for free AFAIR, but that doesn't mean individuals are
typically interested in getting involved in a process that is rather
tedious and bureaucratic.

So my suggestion would be to find out to what extent the latest
version of the spec can describe non-trivial MDIO topologies, and
hopefully that includes the mvpp2 on the Marvell 8040.