[PATCH 0/8] drm: writeback: clean up writeback connector initialization

[Dmitry Baryshkov]

On Mon, Aug 04, 2025 at 03:43:12PM +0100, Liviu Dudau wrote:
> Hi,
> 
> On Fri, Aug 01, 2025 at 04:51:08PM +0300, Dmitry Baryshkov wrote:
> > Drivers using drm_writeback_connector_init() / _with_encoder() don't
> > perform cleanup in a manner similar to drmm_writeback_connector_init()
> > (see drm_writeback_connector_cleanup()). Migrate all existing drivers
> > to use drmm_writeback_connector_init(), drop
> > drm_writeback_connector_init() and drm_writeback_connector::encoder
> > (it's unused afterwards).
> > 
> > This series leaves former drm_writeback_connector_init_with_encoder()
> > (renamed to drm_writeback_connector_init as a non-managed counterpart
> > for drmm_writeback_connector_init()). It is supposed to be used by
> > drivers which can not use drmm functions (like Intel). However I think
> > it would be better to drop it completely.
> 
> The intent of _init_with_encoder() was to be a special case for drivers
> that use their own specific encoder and the rest use the generic function
> that creates the virtual encoder inside the call. The API for
> _init_with_encoder() was actually introduced 4 years after the original
> patch, so that should give a hint.
> 
> drmm_writeback_connector_init() is more like _init_with_encoder() and
> I don't remember reviewing it, so I'm not sure why that was considered
> to be the better behaviour for the managed version. Now you're moving
> all the drivers to the managed version and you have to duplicate code
> in each driver to create the ENCODER_VIRTUAL encoder.

This follows e.g. the process of deprecating drm_simple_* /
drm_simple_encoder. The drivers are expected to open code empty encoder
handling on their own.

> I'm not against the changes being made in the series, I just want to
> see a better justification on why _init_with_encoder() behaviour is
> better than the previous default that you're removing.

This was triggered by the discussion of Intel writeback patchset, see
the threads for first three patches of [1]. We have an optional
non-pointer field inside drm_writeback_connector, which can be left
uninitialized (or zero-filled). I have checked and the drivers are not
actually using the embedded connector for anything after linking it to
the drm_connector. So, by removing the encoder from the
drm_writeback_connector structure we are tying a loose end.

[1] https://lore.kernel.org/dri-devel/20250725050409.2687242-1-suraj.kandpal@intel.com/

-- 
With best wishes
Dmitry