Re: [PATCH 3/3] [media] tvp5150: Migrate to media-controller framework and add video format detection

[Mauro Carvalho Chehab <mchehab@infradead.org>]

Em 05-10-2011 20:14, Sakari Ailus escreveu:
> Hi Mauro,
>
> On Wed, Oct 05, 2011 at 06:41:27PM -0300, Mauro Carvalho Chehab wrote:
>> Em 05-10-2011 17:08, Laurent Pinchart escreveu:
> [clip]
>>> The pad-level API doesn't replace the V4L2 API, it complements it. I'm of
>>> course not advocating modifying any driver to replace V4L2 ioctls by direct
>>> subdev access. However, the pad-level API needs to be exposed to userspace, as
>>> some harware simply can't be configured through a video node only.
>>>
>>> As Hans also mentionned in his reply, the pad-level API is made of two parts:
>>> an in-kernel API made of subdev operations, and a userspace API accessed
>>> through ioctls. As the userspace API is needed for some devices, the kernel
>>> API needs to be implemented by drivers. We should phase out the non pad-level
>>> format operations in favor of pad-level operations, as the former can be
>>> implemented using the later. That has absolutely no influence on the userspace
>>> API.
>>
>> What I'm seeing is that:
>> 	- the drivers that are implementing the MC/pad API's aren't
>> compatible with generic V4L2 applications;
>
> This is currently true up to a certain degree; you'll need to configure such
> devices using media-ctl at least. Even then, such embedded systems do often
> have automatic exposure and white balance algorithms above the V4L2 (often
> proprietary implementations).
>
> To be really useful for a regular user, such algorithms would also be
> needed.
>
> The two problems are separate but they still both need to be resolved to be
> able to use general purpose V4L2 applications on such systems in a
> meaningful way.

That's true, and the MC/pad API's were added to offer support for those proprietary
plugins via libv4l.

That's fine, but the thing is that some developers seem to think that only the streaming
should be done via the V4L2 API, and all the rest via pad configs, while others have
different views. Also, I couldn't see a consense about input selection on drivers that
implement MC: they could either implement S_INPUT, create one V4L device node for each
input, or create just one devnode and let userspace (somehow) to discover the valid
inputs and set the pipelines via MC/pad.

With all those complex and different scenarios, writing a generic plugin/application
that would cover all possible alternatives with a mix of V4L/MC/pad API's would be
extremely complex, and will take years to do it right.

In other words, we need to confine the possible solutions to the ones that will
actually be supported.

>> 	- there's no way to write a generic application that works with all
>> drivers, even implementing MC/pad API's there as each driver is taking different
>> a approach on how to map things at the different API's, and pipeline configuration
>> is device-dependent;
>
> The pipeline configuration is device specific but the interfaces are not.
> Thus it's possible to write a generic plugin which configures the device.
>
> The static configuration can be kept in a text file in the first phase and
> later on hints may be added for synamic configuration on e.g. where digital
> gain, scaling and cropping should be performed and which resolutions to
> advertise in enum_framesizes.

Assuming that all drivers do the same, this would be an alternative for such
plugin.

> But we do not have such plugin yet.

As several SoC developers showed on all opportunities we've met, the complexity
for those devices is very high. The practical effect of adding a kernel driver
without the corresponding required library support is that we ended by
merging incomplete drivers and nobody, except for the manufacturers/developers
whose submitted them could actually make an userspace application to work
with. That would be ok if the drivers would be at /staging, but they aren't.
This situation should be fixed as soon as possible.

I'm seriously considering to not send the patches we have on our tree for those
drivers upstream while we don't have a solution for that.

>> 	- there's no effort to publish patches to libv4l to match the changes
>> at the kernel driver.
>
> I'd prefer concentrating all efforts towards a single, generic plugin. As
> noted before, a plugin for OMAP 3 ISP exists, but I don't think it does
> anything a generic plugin couldn't do, so it hasn't been merged.
>
> I'm working on patches to move the text-based pipeline configuration to
> libmediactl and libv4l2subdev and will post them to the list in the coming
> few days, among with a few other improvements. This is one of the first
> required steps towards such generic plugin.

That would be great, but it won't solve, as more mess is being proposed each
day. One thing is to set the pipelines. Another thing is to not support
things like S_FMT/S_STD/... ioctl's. The user knows what format he wants,
and that's what S_FMT/S_STD tells to the driver. the extra formats at the
pipeline could be handled by a policy either at the driver or at a libv4l,
although I don't think that a generic libv4l plugin would be capable of
doing it (see more about that bellow).

> Unfortunately I haven't been able to use much of my time on this; help would
> indeed be appreciated from any interested party.
>
> [clip]
>
>> I'm fine on providing raw interfaces, just like we have for some types of device
>> (like the block raw interfaces used by CD-ROM drivers) as a bonus, but this should
>> never replace an API where an application developed by a third party could work
>> with all media hardware, without needing hardware specific details.
>
> I agree.
>
> [clip]
>
>>>>>> If the application wants a different image resolution, it will use
>>>>>> S_FMT. In this case, what userspace expects is that the driver will
>>>>>> scale, if supported, or return -EINVAL otherwise.
>>>>>
>>>>> With the OMAP3 ISP, which is I believe what Javier was asking about, the
>>>>> application will set the format on the OMAP3 ISP resizer input and output
>>>>> pads to configure scaling.
>>>>
>>>> The V4L2 API doesn't tell where a function like scaler will be implemented.
>>>> So, it is fine to implement it at tvp5151 or at the omap3 resizer, when a
>>>> V4L2 call is sent.
>>>
>>> By rolling a dice ? :-)
>>
>> By using good sense. I never had a case where I had doubts about where the
>> scaling should be implemented on the drivers I've coded. For omap3/tvp5151, the
>> decision is also clear: it should be done at the bridge (omap3) resizer, as the
>> demod doesn't support scaling.
>
> Good sense in this case would require knowledge of the tv tuner in the
> camera ISP driver. One could also, and actually does, connect sensors which
> can do scaling in two steps to the same ISP. How does the ISP driver now
> know where to do scaling?

Does it actually make sense to allow scaling on both? In the case of the
generic plugin you're writing, would it do scale on both? If not, how
would it decide where to scale? Moving the problem to userspace won't solve
it, and would require userspace to take some hard decisions based on the specific
hardware designs. In practice, it means that, every time a hardware is updated
(a new sensor is added, etc), both the plugin and the driver will need to be
touched, at the same time. This also means that the userspace plugin will
be dependent on an specific kernel version, which would be a real mess to
the distributors to package the libv4l plugin.

Btw, we have some devices that support scaling on both tv decoder and bridge.
As not all tv decoders have scaling, the decision were to implement scaling
it only at the bridge. This was as good as doing it at the sensor (or even
better, as some bridges have temporal/spacial decimation filtering, with
provides a better decimation than a pure spacial filtering, and can reduce
the quantization noise).

> OMAP 3 ISP as such is a relatively simple one; there are ISPs which have two
> scalers and ability to write the image to memory also in between the two.
>
> I don't think a driver, be that tv tuner / sensor or ISP driver, has enough
> information to perform the pipeline configuration which is mandatory to
> implement the standard V4L2 ioctls for e.g. format setting and cropping. We
> actually had such an implementation of the OMAP 3 ISP driver around 2009
> (but without MC or V4L2 subdevs, still the underlying technical issues are
> the same) and it was a disaster. The patches were posted to linux-media a
> few times back then if you're interested. It was evident something had to be
> done, and now we have the Media controller and V4L2 subdev user space APIs.

As I said before, you're just transfering the problem from the kernel to
something else, without actually solving it.

> I want to reiterate that I'm very much in favour of supporting generic V4L2
> applications on these devices. The additional support that these application
> need on top of the drivers requires making policy decisions that
> applications intended for embedded devices (such as Fcam) often run on
> embedded devices need to make themselves. To support both using the same
> drivers, these policy decisions must be taken in the user space, not in the
> kernel. In my opinion libv4l is in perfect spot to fill this gap.
>
> Kind regards,
>