Hi,

On Fri, Nov 04, 2016 at 01:05:01AM +0200, Sakari Ailus wrote:
> On Thu, Nov 03, 2016 at 11:48:43PM +0100, Sebastian Reichel wrote:
> > On Tue, Nov 01, 2016 at 12:54:08AM +0200, Sakari Ailus wrote:
> > > > > Thanks, this answered half of my questions already. ;-)
> > > > :-).
> > > > 
> > > > I'll have to go through the patches, et8ek8 driver is probably not
> > > > enough to get useful video. platform/video-bus-switch.c is needed for
> > > > camera switching, then some omap3isp patches to bind flash and
> > > > autofocus into the subdevice.
> > > > 
> > > > Then, device tree support on n900 can be added.
> > > 
> > > I briefly discussed with with Sebastian.
> > > 
> > > Do you think the elusive support for the secondary camera is worth keeping
> > > out the main camera from the DT in mainline? As long as there's a reasonable
> > > way to get it working, I'd just merge that. If someone ever gets the
> > > secondary camera working properly and nicely with the video bus switch,
> > > that's cool, we'll somehow deal with the problem then. But frankly I don't
> > > think it's very useful even if we get there: the quality is really bad.
> > 
> > If we want to keep open the option to add proper support for the
> > second camera, we could also add the bus switch and not add the
> > front camera node in DT. Then adding the front camera does not
> > require DT or userspace API changes. It would need an additional
> > DT quirk in arch/arm/mach-omap2/board-generic.c for RX51, which
> > adds the CCP2 bus settings from the camera node to the bus
> > switch node to keep isp_of_parse_node happy. That should be
> > easy to implement and not add much delay in upstreaming.
> 
> By adding the video bus switch we have a little bit more complex system as a
> whole. The V4L2 async does not currently support this. There's more here:
> 
> <URL:http://www.spinics.net/lists/linux-media/msg107262.html>

I'm not sure what part relevant for video-bus-switch is currently
not supported?

video-bus-switch registers its own async notifier and only registers
itself as subdevices to omap3isp, once its own subdevices have been
registered successfully.

> What I thought was that once we have everything that's required in
> place, we can just change what's in DT. But the software needs to
> continue to work with the old DT content.

Right, so DT is not a problem. But adding the switch would change
the media-graph, which is exposed to userspace.

> > For actually getting both cameras available with runtime-switching
> > the proper solution would probably involve moving the parsing of
> > the bus-settings to the sensor driver and providing a callback.
> > This callback can be called by omap3isp when it wants to configure
> > the phy (which is basically when it starts streaming). That seems
> > to be the only place needing the buscfg anyways.
> > 
> > Then the video-bus-switch could do something like this (pseudocode):
> > 
> > static void get_buscfg(struct *this, struct *buscfg) {
> >     if (selected_cam == 0)
> >         return this->sensor_a->get_buscfg(buscfg);
> >     else
> >         return this->sensor_b->get_buscfg(buscfg);
> > }
> > 
> > Regarding the usefulness: I noticed, that the Neo900 people also
> > plan to have the bus-switch [0]. It's still the same crappy front-cam,
> > though. Nevertheless it might be useful for testing. It has nice
> > test-image capabilities, which might be useful for regression
> > testing once everything is in place.
> > 
> > [0] http://neo900.org/stuff/block-diagrams/neo900/neo900.html
> 
> Seriously? I suppose there should be no need for that anymore, is there?
> 
> I think they wanted to save one GPIO in order to shave off 0,0001 cents from
> the manufacturing costs or something like that. And the result is...
> painful. :-I

CSI1/CCP2 is more than a single I/O pin, isn't it? Or do you
reference to the GPIO dual use to enable frontcam and switch
between the cameras? That is indeed a really ugly solution :(

-- Sebastian