On Mon October 8 2012 16:30:53 Guennadi Liakhovetski wrote:
> On Mon, 8 Oct 2012, Hans Verkuil wrote:
>
> > On Mon October 8 2012 14:23:25 Guennadi Liakhovetski wrote:
> > > Hi Hans
> > >
> > > On Fri, 5 Oct 2012, Hans Verkuil wrote:
> > >
> > > [snip]
> > >
> > > > I think the soc_camera patches should be left out for now. I suspect
> > > > that
> > > > by adding core support for async i2c handling first,
> > >
> > > Ok, let's think, what this meacs - async I2C in media / V4L2 core.
> > >
> > > The main reason for our probing order problem is the master clock,
> > > typically supplied from the camera bridge to I2C subdevices, which we
> > > only
> > > want to start when necessary, i.e. when accessing the subdevice. And the
> > > subdevice driver needs that clock running during its .probe() to be able
> > > to access and verify or configure the hardware. Our current solution is
> > > to
> > > not register I2C subdevices from the platform data, as is usual for all
> > > I2C devices, but from the bridge driver and only after it has switched on
> > > the master clock. After the subdevice driver has completed its probing we
> > > switch the clock back off until the subdevice has to be activated, e.g.
> > > for video capture.
> > >
> > > Also important - when we want to unregister the bridge driver we just
> > > also
> > > unregister the I2C device.
> > >
> > > Now, to reverse the whole thing and to allow I2C devices be registered as
> > > usual - via platform data or OF, first of all we have to teach I2C
> > > subdevice drivers to recognise the "too early" situation and request
> > > deferred probing in such a case. Then it will be reprobed after each new
> > > successful probe or unregister on the system. After the bridge driver has
> > > successfully probed the subdevice driver will be re-probed and at that
> > > time it should succeed. Now, there is a problem here too: who should
> > > switch on and off the master clock?
> > >
> > > If we do it from the bridge driver, we could install an I2C bus-notifier,
> > > _before_ the subdevice driver is probed, i.e. upon the
> > > BUS_NOTIFY_BIND_DRIVER event we could turn on the clock. If subdevice
> > > probing was successful, we can then wait for the BUS_NOTIFY_BOUND_DRIVER
> > > event to switch the clock back off. BUT - if the subdevice fails probing?
> > > How do we find out about that and turn the clock back off? There is no
> > > notification event for that... Possible solutions:
> > >
> > > 1. timer - ugly and unreliable.
> > > 2. add a "probing failed" notifier event to the device core - would this
> > > be accepted?
> > > 3. let the subdevice turn the master clock on and off for the duration of
> > > probing.
> > >
> > > My vote goes for (3). Ideally this should be done using the generic clock
> > > framework. But can we really expect all drivers and platforms to switch
> > > to
> > > it quickly enough? If not, we need a V4L2-specific callback from
> > > subdevice
> > > drivers to bridge drivers to turn the clock on and off. That's what I've
> > > done "temporarily" in this patch series for soc-camera.
> > >
> > > Suppose we decide to do the same for V4L2 centrally - add call-backs.
> > > Then
> > > we can think what else we need to add to V4L2 to support asynchronous
> > > subdevice driver probing.
> >
> > I wonder, don't we have the necessary code already? V4L2 subdev drivers can
> > have internal_ops with register/unregister ops. These are called by
> > v4l2_device_register_subdev. This happens during the bridge driver's probe.
> >
> > Suppose the subdev's probe does not actually access the i2c device, but
> > instead defers that to the register callback? The bridge driver will turn on
> > the clock before calling v4l2_device_register_subdev to ensure that the
> > register callback can access the i2c registers. The register callback will
> > do any initialization and can return an error. In case of an error the i2c
> > client is automatically unregistered as well.
>
> Yes, if v4l2_i2c_new_subdev_board() is used. This has been discussed
> several times before and always what I didn't like in this is, that I2C
> device probe() in this case succeeds without even trying to access the
> hardware. And think about DT. In this case we don't instantiate the I2C
> device, OF code does it for us. What do you do then? If you let probe()
> succeed, then you will have to somehow remember the subdevice to later
> match it against bridges...
Yes, but you need that information anyway. The bridge still needs to call
v4l2_device_register_subdev so it needs to know which subdevs are loaded.
And can't it get that from DT as well?
In my view you cannot do a proper initialization unless you have both the
bridge driver and all subdev drivers loaded and instantiated. They need one
another. So I am perfectly fine with letting the probe function do next to
nothing and postponing that until register() is called. I2C and actual probing
to check if it's the right device is a bad idea in general since you have no
idea what a hardware access to an unknown i2c device will do. There are still
some corner cases where that is needed, but I do not think that that is an
issue here.
It would simplify things a lot IMHO. Also note that the register() op will
work with any device, not just i2c. That may be a useful property as well.
> > In addition, during the register op the subdev driver can call into the
> > bridge driver since it knows the v4l2_device struct.
> >
> > This has also the advantage that subdev drivers can change to this model
> > gradually. Only drivers that need master clocks, etc. need to move any probe
> > code that actually accesses hardware to the register op. Others can remain
> > as. Nor should this change behavior of existing drivers as this happens
> > all in the V4L2 core.
> >
> > The bridge driver may still have to wait until all i2c drivers are loaded,
> > though. But that can definitely be handled centrally (i.e.: 'I need these
> > drivers, wait until all are loaded').
> >
> > > 1. We'll have to create these V4L2 clock start and stop functions, that,
> > > supplied (in case of I2C) with client address and adapter number will
> > > find
> > > the correct v4l2_device instance and call its callbacks.
> > >
> > > 2. The I2C notifier. I'm not sure, whether this one should be common. Of
> > > common tasks we have to refcount the I2C adapter and register the
> > > subdevice. Then we'd have to call the bridge driver's callback. Is it
> > > worth it doing this centrally or rather allow individual drivers to do
> > > that themselves?
> > >
> > > Also, ideally OF-compatible (I2C) drivers should run with no platform
> > > data, but soc-camera is using I2C device platform data intensively. To
> > > avoid modifying the soc-camera core and all drivers, I also trigger on
> > > the
> > > BUS_NOTIFY_BIND_DRIVER event and assign a reference to the dynamically
> > > created platform data to the I2C device. Would we also want to do this
> > > for
> > > all V4L2 bridge drivers? We could call this a "prepare" callback or
> > > something similar...
> >
> > Well, subdev drivers should either parse the OF data, or use the
> > platform_data.
> > The way soc_camera uses platform_data is one reason why it is so hard to
> > reuse subdevs for non-soc_camera drivers. All the callbacks in
> > soc_camera_link
> > should be replaced by calls to the v4l2_device notify() callback. After
> > that we
> > can see what is needed to drop struct soc_camera_link altogether as
> > platform_data.
>
> They don't have to be, they are not (or should not be) called by
> subdevices.
Then why are those callbacks in a struct that subdevs can access? I always
have a hard time with soc_camera figuring out who is using what when :-(
> > > 3. Bridge driver unregistering. Here we have to put the subdevice driver
> > > back into the deferred-probe state... Ugliness alert: I'm doing this by
> > > unregistering and re-registering the I2C device... For that I also have
> > > to
> > > create a copy of devices I2C board-info data. Lovely, ain't it? This I'd
> > > be happy to move to the V4L2 core;-)
> >
> > By just using the unregister ops this should be greatly simplified as well.
>
> Sorry, which unregister ops do you mean? internal_ops->unregistered()?
Yes.
> Yes, but only if we somehow go your way and use dummy probe() methods...
Of course.
Regards,
Hans
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html