Regression on the sun4i-drm driver by cma helper change

[Noralf Trønnes]

Den 11.07.2018 17.04, skrev Daniel Vetter:
> On Wed, Jul 11, 2018 at 4:52 PM, Noralf Trønnes <noralf at tronnes.org> wrote:
>> Den 11.07.2018 16.06, skrev Maxime Ripard:
>>> CC'ing Noralf and Daniel,
>>>
>>> On Wed, Jul 11, 2018 at 09:47:53PM +0800, Icenowy Zheng wrote:
>>>> Today, during testing the Banana Pi M2 Zero HDMI patch, I found a
>>>> regression on the sun4i-drm driver, which happenes because of the cma
>>>> helper change.
>>>>
>>>> The bad commit is 894a677f4b3e6d2ab8d01bb46c1fbd5f92e4591b ("drm/cma-
>>>> helper: Use the generic fbdev emulation").
>>>>
>>>> If this commit is present, sun4i_framebuffer_init() will return
>>>> -ENODEV, because of the call of drm_client_new(). In this case sun4i-
>>>> drv will fail to probe, and all components will be unbound.
>>>>
>>>> By reverting the commit, sun4i-drv loads and HDMI output is correct.
>>>>
>>>> Could anyone investigate into this, please?
>>> Any ideas?
>>
>> I think it's because drm_fb_cma_fbdev_init() is called before
>> drm_dev_register() and that this is what returns -ENODEV:
>>
>> int drm_client_new(struct drm_device *dev, struct drm_client_dev *client,
>>             const char *name, const struct drm_client_funcs *funcs)
>> {
>> ...
>>      mutex_lock(&dev->clientlist_mutex);
>>      registered = dev->registered;
>>      if (registered)
>>          list_add(&client->list, &dev->clientlist);
>>      mutex_unlock(&dev->clientlist_mutex);
>>      if (!registered) {
>>          ret = -ENODEV;
>>          goto err_close;
>>      }
>>
>> I would have been nice to get it confirmed that indeed this is the case.
>>
>> Maybe we should remove this check temporarily until all have converted to
>> drm_fbdev_generic_setup() which will be called after drm_dev_register().
>>
>> I'll come back when I have looked at the other drivers to see in which
>> order they register fbdev and DRM.
> Yeah I don't think we can require that drivers have called
> drm_dev_register before setting up the fbdev stuff, that will break
> the world. What we might want to consider is a ->register callback
> (for e.g. framebuffer_register, but not all the stuff leading up to
> it), but for now I don't think there's a need for that.
> -Daniel

I've posted a fix, cc'ed the persons in this mail.

Noralf.