[lkml]   [2017]   [Dec]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 0/4] Move DP phy switch to PHY driver
Am Freitag, 1. Dezember 2017, 13:42:46 CET schrieb Doug Anderson:
> Hi,
> On Wed, Nov 29, 2017 at 6:27 PM, Chris Zhong <> wrote:
> > Hi Doug
> >
> > Thank you for mentioning this patch.
> >
> > I think the focus of the discussion is: can we put the grf control bit to
> > dts.
> >
> > The RK3399 has 2 Type-C phy, but only one DP controller, this "uphy_dp_sel"
> >
> > can help to switch these 2 phy. So I think this bit can be considered as a
> > part of
> >
> > Type-C phy, these 2 phy have different bits, just similar to other bits
> > (such as "pipe-status").
> >
> > Put them to DTS file might be a accepted practice.
> I guess the first step would be finding the person to make a decision.
> Is that Heiko? Olof? Kishon? Rob?. As I see it there are a few
> options:
> 1. Land this series as-is. This makes the new bit work just like all
> the other ones next to it. If anyone happens to try to use an old
> device tree on a new kernel they'll break. Seems rather unlikely
> given that the whole type C PHY is not really fully functional
> upstream, but technically this is a no-no from a device tree
> perspective.
> 2. Change the series to make this property optional. If it's not
> there then the code behaves like it always did. This would address
> the "compatibility" problem but likely wouldn't actually help any real
> people, and it would be extra work.
> 3. Redo the driver to deprecate all the old offsets / bits and just
> put the table in the driver, keyed off the compatible string and base
> address if the IO memory.
> I can't make this decision. It's up to those folks who would be
> landing the patch and I'd be happy with any of them. What I'm less
> happy with, however, is the indecision preventing forward progress.
> We should pick one of the above things and land it. My own personal
> bias is #1: just land the series. No real people will be hurt and
> it's just adding another property that matches the ones next to it.

I'd second that #1 . That whole type-c phy thingy never fully worked in
the past (some for the never used dp output), so personally I don't have
issues with going that route.

> From a long term perspective (AKA how I'd write the next driver like
> this) I personally lean towards to "tables in the driver, not in the
> device tree" but quite honestly I'm happy to take whatever direction
> the maintainers give.

It looks like we're in agreement here :-) . GRF stuff should not leak into
the devicetree, as it causes endless headaches later. But I guess we'll
need to live with the ones that happened so far.


 \ /
  Last update: 2017-12-01 22:58    [W:0.068 / U:0.480 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site