[PATCH v6 05/21] net-next: stmmac: Add dwmac-sun8i

[Chen-Yu Tsai]

On Tue, Jun 27, 2017 at 5:41 PM, Maxime Ripard
<maxime.ripard at free-electrons.com> wrote:
> On Tue, Jun 27, 2017 at 10:02:45AM +0100, Andre Przywara wrote:
>> Hi,
>>
>> (CC:ing some people from that Rockchip dmwac series)
>>
>> On 27/06/17 09:21, Corentin Labbe wrote:
>> > On Tue, Jun 27, 2017 at 04:11:21PM +0800, Chen-Yu Tsai wrote:
>> >> On Tue, Jun 27, 2017 at 4:05 PM, Corentin Labbe
>> >> <clabbe.montjoie at gmail.com> wrote:
>> >>> On Mon, Jun 26, 2017 at 01:18:23AM +0100, André Przywara wrote:
>> >>>> On 31/05/17 08:18, Corentin Labbe wrote:
>> >>>>> The dwmac-sun8i is a heavy hacked version of stmmac hardware by
>> >>>>> allwinner.
>> >>>>> In fact the only common part is the descriptor management and the first
>> >>>>> register function.
>> >>>>
>> >>>> Hi,
>> >>>>
>> >>>> I know I am a bit late with this, but while adapting the U-Boot driver
>> >>>> to the new binding I was wondering about the internal PHY detection:
>> >>>>
>> >>>>
>> >>>> So here you seem to deduce the usage of the internal PHY by the PHY
>> >>>> interface specified in the DT (MII = internal, RGMII = external).
>> >>>> I think I raised this question before, but isn't it perfectly legal for
>> >>>> a board to use MII with an external PHY even on those SoCs that feature
>> >>>> an internal PHY?
>> >>>> On the first glance that does not make too much sense, but apart from
>> >>>> not being the correct binding to describe all of the SoCs features I see
>> >>>> two scenarios:
>> >>>> 1) A board vendor might choose to not use the internal PHY because it
>> >>>> has bugs, lacks features (configurability) or has other issues. For
>> >>>> instance I have heard reports that the internal PHY makes the SoC go
>> >>>> rather hot, possibly limiting the CPU frequency. By using an external
>> >>>> MII PHY (which are still cheaper than RGMII PHYs) this can be avoided.
>> >>>> 2) A PHY does not necessarily need to be directly connected to
>> >>>> magnetics. Indeed quite some boards use (RG)MII to connect to a switch
>> >>>> IC or some other network circuitry, for instance fibre connectors.
>> >>>>
>> >>>> So I was wondering if we would need an explicit:
>> >>>>       allwinner,use-internal-phy;
>> >>>> boolean DT property to signal the usage of the internal PHY?
>> >>>> Alternatively we could go with the negative version:
>> >>>>       allwinner,disable-internal-phy;
>> >>>>
>> >>>> Or what about introducing a new "allwinner,internal-mii-phy" compatible
>> >>>> string for the *PHY* node and use that?
>> >>>>
>> >>>> I just want to avoid that we introduce a binding that causes us
>> >>>> headaches later. I think we can still fix this with a followup patch
>> >>>> before the driver and its binding hit a release kernel.
>> >>>>
>> >>>> Cheers,
>> >>>> Andre.
>> >>>>
>> >>>
>> >>> I just see some patch, where "phy-mode = internal" is valid.
>> >>> I will try to find a way to use it
>> >>
>> >> Can you provide a link?
>> >
>> > https://lkml.org/lkml/2017/6/23/479
>> >
>> >>
>> >> I'm not a fan of using phy-mode for this. There's no guarantee what
>> >> mode the internal PHY uses. That's what phy-mode is for.
>>
>> I can understand Chen-Yu's concerns, but ...
>>
>> > For each soc the internal PHY mode is know and setted in emac_variant/internal_phy
>> > So its not a problem.
>>
>> that is true as well, at least for now.
>>
>> So while I agree that having a separate property to indicate the usage
>> of the internal PHY would be nice, I am bit tempted to use this easier
>> approach and piggy back on the existing phy-mode property.
>
> We're trying to fix an issue that works for now too.
>
> If we want to consider future weird cases, then we must consider all
> of them. And the phy mode changing is definitely not really far
> fetched.

I guess the issue is whether it's likely that the vendor puts 2 internal
PHYs in one SoC, and they use different modes and can be switched around.
Otherwise it's fixed for a given SoC, and we can just handle that with
the per-SoC GMAC compatible.

Maybe Florian could tell us if this was one of the intended use cases
for the "internal" phy mode.

As for Rockchip, AFAIK they have 2 MACs, one is connected to the internal
PHY, while the other is connected to the external interface, and there is
no muxing involved, unlike Allwinner's solution.

> I agree with Chen-Yu, and I really feel like the compatible solution
> you suggested would cover both your concerns, and ours.

If using a PHY compatible is the solution, we could just use the
"ethernet-phy-idAAAA.BBBB" style one, and put in the bogus ID that
Allwinner used.

Care must be taken to put this at the board level for boards using
the internal PHY, or we'd have to delete or override the property
in all other boards.

Ideally I think the internal PHY device node should _not_ be in
the SoC level .dtsi file. If we select the external interface, then
there's no connection to the internal PHY, and that device node becomes
unusable and bogus. This is something I think should be fixed regardless
of the phy-mode discussion above.

ChenYu