On Fri, 2021-11-12 at 17:32 +0300, Dmitry Osipenko wrote:
> 12.11.2021 13:52, Thierry Reding пишет:
> > On Tue, Nov 09, 2021 at 05:39:02PM +0300, Dmitry Osipenko wrote:
> > > 09.11.2021 17:17, Dmitry Osipenko пишет:
> > > > 09.11.2021 17:08, Dmitry Osipenko пишет:
> > > > > > +static void host1x_drm_dev_deinit(struct host1x_device *dev)
> > > > > > +{
> > > > > > + struct drm_device *drm = dev_get_drvdata(&dev->dev);
> > > > > And platform_unregister_drivers() should be moved here.
> > > > >
> > > >
> > > > Nah, that should cause deadlock. This ad-hoc is too lame.
> > >
> > > Actually, there is no problem here as I see now. The host1x driver
> > > populates DT nodes after host1x_register() [1], meaning that Host1x DRM
> > > will be always inited first.
> > >
> > > [1]
> > > https://elixir.bootlin.com/linux/v5.15/source/drivers/gpu/host1x/dev.c#L475
> > >
> > > Still I'm not a fan of the ad-hoc solution.
> >
> > Could we not fix this by making the panel "hot-pluggable"? I don't think
> > there's anything inherent to the driver that would prevent doing so. The
> > original reason for why things are as intertwined as they are now is
> > because back at the time deferred framebuffer creation didn't exist. In
> > fact I added deferred framebuffer support with Daniel's help precisely
> > to fix a similar issue for things like HDMI and DP.
>
> I don't understand what do you mean by "hot-pluggable", panel is static.
> Please either clarify more or type the patch.
>
> Keep in mind that fix should be simple and portable because stable
> kernels are wrecked.
>
> > With HDMI and DP it's slightly less critical, because a lack of mode
> > support would've created a 1024x768 framebuffer, which most HDMI/DP
> > monitors support. However, with panels we need to ensure that the exact
> > mode from the panel is used to create the framebuffer, so it can only be
> > created when the panel is available.
> >
> > But, given that deferred framebuffer creation works now (which allows
> > the framebuffer console to show up at the preferred resolution for HDMI
> > and DP), even if a monitor is attached only after the driver has probed
> > already, we should be able to make something like that work with panels
> > as well.
>
> BTW, I see now that DPAUX I2C transfer helper may access
> aux->drm_device. Hence v1 patch isn't correct anyways.
JFYI - unless I'm misunderstanding you, the aux->drm_dev accesses in the DPAUX
i2c transfer functions are just from the various drm_{dbg,err,etc.} calls,
which means that they all should be able to handle aux->drm_dev being NULL. If
you can set aux->drm_dev before i2c transfers start that's more ideal, since
otherwise you'll see the AUX device name as "(null)" in the kernel log, but
any point before device registration should work.
>
> For now I'll try to test more the ad-hoc variant with Thomas and send it
> as v2 if we won't have a better solution.
>
--
Cheers,
Lyude Paul (she/her)
Software Engineer at Red Hat
