Re: [RFC v2 00/13] usb/mmc/power: Fix USB/LAN when TFTP booting

From: Ulf Hansson
Date: Tue May 10 2016 - 07:02:17 EST


+ Arnd

[...]

>> >> Solution
>> >> ========
>> >> This is very similar to the MMC pwrseq behavior so the idea is to:
>> >> 1. Move MMC pwrseq drivers to generic place,
>> >
>> > You can do that, but I'm going to NAK any use of pwrseq bindings outside
>> > of MMC. I think it is the wrong way to do things. The DT should describe
>>
>> Huh, I didn't know that was your view of the mmc pwrseq bindings. Why
>> didn't you NAK them before?
>
> Unfortunately, either I missed it or it was a time I couldn't spend much
> time on reviews.

Okay, I guess it's common issue among maintainers. The problem with DT
is that it gets really hard to be fixed up later. :-)

>
>> > the devices. If they happen to be "simple" then the core can walk the
>> > tree and do any setup. For example, look for "reset-gpios" and toggle
>> > that GPIO. There is no need for a special node.
>> >
>> >> 2. Extend the pwrseq-simple with regulator toggling,
>> >> 3. Add support to USB hub and port core for pwrseq,
>> >
>> > We discussed this for USB already[1] and is why we defined how to add
>> > USB child devices. The idea is not to add pwrseq to that.
>>
>> I am not familiar with the USB discussion.
>>
>> Still, let me give you some more background to the mmc pwrseq. The
>> idea from the mmc pwrseq bindings comes from the power-domain DT
>> bindings, as I thought these things were a bit related.
>> In both cases they are not directly a property of the device, but more
>> describing a HW dependency to allow the device to work.
>
> I could see this as a board level power domain. However the difference
> is we are not generally exposing internal SOC details the same way as
> board level components. Perhaps we could extend power domains to board
> level, but that is not what was done here.
>
>> One could probably use a child node instead of a phandle, but that
>> wasn't chosen back then. Of course you are the DT expert, but could
>> you perhaps tell me why a child node is better for cases like this?
>
> If there is a control path hierarchy, then we try to model that in DT
> with child nodes. In cases of SDIO and USB, there is a clear hierarchy.
> Ignoring the discovery ordering problem, we already have defined ways to
> describe GPIO connections, regulators, etc. to devices. Describing those
> things separately from the device to solve a particular issue that is
> really a kernel limitation is what I don't like.

Okay, I see.

To move forward in trying to make mmc pwrseq a generic pwrseq, could
we perhaps allow both cases?

In the mmc case, there are already deployed bindings so we need to
cope with these by using the phandle option, but for USB etc we could
force the child node option.
As long as we agree that we keep using a compatible string for the
child node as well, both options should be able to co-exist and we
should probably be able to managed them both from a common pwrseq
driver framework.

Although, I do remember from an older conversations around some of
mine submission for the mmc pwrseq code, that some people (maybe
Arnd?) wasn't keen on adding a new framework for this. Perhaps that
has changed?

Kind regards
Uffe