[RFC PATCH 1/8] drm: writeback: Refactor drm_writeback_connector structure

[Dmitry Baryshkov]

On Tue, Aug 26, 2025 at 05:48:18PM +0200, mripard at kernel.org wrote:
> On Mon, Aug 25, 2025 at 06:26:48AM +0000, Kandpal, Suraj wrote:
> > > Subject: Re: [RFC PATCH 1/8] drm: writeback: Refactor
> > > drm_writeback_connector structure
> > > 
> > > Hi,
> > > 
> > > On Sat, Aug 16, 2025 at 01:20:53AM +0300, Dmitry Baryshkov wrote:
> > > > On Thu, Aug 14, 2025 at 05:13:54PM +0100, liviu.dudau at arm.com wrote:
> > > > > Hi,
> > > > >
> > > > > On Wed, Aug 13, 2025 at 10:04:22AM +0000, Kandpal, Suraj wrote:
> > > > > > > > > };
> > > > > > > >
> > > > > > > > I still don't like that. This really doesn't belong here. If
> > > > > > > > anything, the drm_connector for writeback belongs to drm_crtc.
> > > > > > >
> > > > > > > Why? We already have generic HDMI field inside drm_connector. I
> > > > > > > am really hoping to be able to land DP parts next to it. In
> > > > > > > theory we can have a DVI- specific entry there (e.g. with the
> > > subconnector type).
> > > > > > > The idea is not to limit how the drivers subclass those structures.
> > > > > > >
> > > > > > > I don't see a good case why WB should deviate from that design.
> > > > > > >
> > > > > > > > If the issue is that some drivers need a custom drm_connector
> > > > > > > > subclass, then I'd rather turn the connector field of
> > > > > > > > drm_writeback_connector into a pointer.
> > > > > > >
> > > > > > > Having a pointer requires additional ops in order to get
> > > > > > > drm_connector from WB code and vice versa. Having
> > > > > > > drm_connector_wb inside drm_connector saves us from those ops
> > > (which don't manifest for any other kind of structure).
> > > > > > > Nor will it take any more space since union will reuse space
> > > > > > > already taken up by HDMI part.
> > > > > > >
> > > > > > > >
> > > > > >
> > > > > > Seems like this thread has died. We need to get a conclusion on the
> > > design.
> > > > > > Laurent do you have any issue with the design given Dmitry's
> > > > > > explanation as to why this Design is good for drm_writeback_connector.
> > > > >
> > > > > I'm with Laurent here. The idea for drm_connector (and a lot of drm
> > > > > structures) are to be used as base "classes" for extended
> > > > > structures. I don't know why HDMI connector ended up inside
> > > > > drm_connector as not all connectors have HDMI functionality, but that's a
> > > cleanup for another day.
> > > >
> > > > Maybe Maxime can better comment on it, but I think it was made exactly
> > > > for the purpose of not limiting the driver's design. For example, a
> > > > lot of drivers subclass drm_connector via drm_bridge_connector. If
> > > > struct drm_connector_hdmi was a wrapper around struct drm_connector,
> > > > then it would have been impossible to use HDMI helpers for bridge
> > > > drivers, while current design freely allows any driver to utilize
> > > > corresponding library code.
> > > 
> > > That's exactly why we ended up like this. With that design, we wouldn't have
> > > been able to "inherit" two connector "classes": bridge_connector is one,
> > > intel_connector another one.
> > > 
> > > See here for the rationale:
> > > https://lore.kernel.org/dri-devel/ZOTDKHxn2bOg+Xmg@phenom.ffwll.local/
> > > 
> > > I don't think the "but we'll bloat drm_connector" makes sense either.
> > > There's already a *lot* of things that aren't useful to every connector (fwnode,
> > > display_info, edid in general, scaling, vrr, etc.)
> > > 
> > > And it's not like we allocate more than a handful of them during a system's life.
> > 
> > So Are we okay with the approach mentioned here with the changes that have been proposed here like
> > Having drm_writeback_connector in union with drm_hdmi_connector
> 
> I don't think we need a union here. It artificially creates the same
> issue: we can't have two types for a connector if we do so.

Well... What kind of connector would be both HDMI and WriteBack? I think
they are mutually exclusive already.

> > Also one more thing I would like to clarify here is how everyone would
> > like the patches patches where each patch changes both the drm core
> > and all related drivers (ensures buildability but then review is tough
> > for each driver). Or patches where we have initial drm core changes
> > and then each patch does the all changes in a driver in its own
> > respective patch.
> 
> The latter should be preferred, but if you can't maintain bisectability
> that way, then it's the most important and you should fall back to the
> former.

I'd say, we should be trying our best in providing bisectability. It
really a PITA if one can not use `git bisect run`.

-- 
With best wishes
Dmitry