Why drm-mipi-dsi is built-in only?

[Takashi Iwai]

On Tue, 02 Aug 2016 13:47:10 +0200,
Thierry Reding wrote:
> 
> On Tue, Aug 02, 2016 at 12:47:06PM +0200, Andrzej Hajda wrote:
> > On 08/02/2016 10:36 AM, Thierry Reding wrote:
> > > On Mon, Aug 01, 2016 at 08:04:55PM +0200, Andrzej Hajda wrote:
> > >> On 08/01/2016 03:59 PM, Jani Nikula wrote:
> > >>> Cc Andrzej, Thierry
> > >>>
> > >>> On Fri, 22 Jul 2016, Daniel Vetter <daniel at ffwll.ch> wrote:
> > >>>> On Fri, Jul 22, 2016 at 04:30:24PM +0200, Takashi Iwai wrote:
> > >>>>> Hi,
> > >>>>>
> > >>>>> is there any reason drm-mipi-dsi can't be a module?  It's fixed as a
> > >>>>> built-in since its Kconfig is bool.
> > >>>> Probably none except embedded folks eshew modules ;-) Submit patch, I'll
> > >>>> apply.
> > >>> Possibly this?
> > >>>
> > >>> postcore_initcall(mipi_dsi_bus_init);
> > >> If I remember correctly, the only reason for this is to have mipi_dsi bus
> > >> registered before mipi_dsi drivers, which usually are registered
> > >> at module initcall. But maybe bus registration can be performed at
> > >> first mipi_dsi driver registration. This way we could modularize it.
> > > I think it should work fine if this was built as a module. The purpose
> > > for having this as postcore_initcall() is simply so that the bus is
> > > fully initialized before any driver gets registered with it. If this
> > > code is built as a module, symbol dependencies will make sure that the
> > > drm_mipi_dsi.ko module will be loaded before any users.
> > 
> > If you change initcall of mipi_dsi to module and then you compile
> > it as built-in, only link order will guard correct initialization sequence.
> > As for now panels are linked after mipi-dsi, so it should be OK,
> > even if little bit hacky.
> 
> I wasn't suggesting that we turn the postcore_initcall() into
> module_init(). postcore_initcall() works just fine for modules (it's
> automatically replaced by a module_init() if the code is built as a
> module) and it will still do the right thing with regard to ordering
> when built-in.

Right, I already tested the code, just building as a module without
any other changes.  And it worked.

About the rest, what Thierry suggested:

> Two things are missing, though: 1) the drm_mipi_dsi.ko module would need
> to be reference counted so that the symbols stay around as long as there
> are any drivers (this might be covered by symbol dependencies already)

Yeah, the symbol dependency should cover it enough.

> and 2) there would need to be an exit function for the module to cleanup
> the bus.

So only this one is missing.  I'm going to cook this as well and
submit patches later.


thanks,

Takashi