Re: [PATCH v4 7/7] media: open.rst: add a notice about subdev-API on vdev-centric

From: Mauro Carvalho Chehab
Date: Tue Aug 29 2017 - 05:14:48 EST


Em Tue, 29 Aug 2017 10:39:52 +0200
Hans Verkuil <hverkuil@xxxxxxxxx> escreveu:

> On 29/08/17 10:31, Ramesh Shanmugasundaram wrote:
> > Hi Hans,
> >
> >> On 28/08/17 12:30, Mauro Carvalho Chehab wrote:
> >>> Em Mon, 28 Aug 2017 12:05:06 +0200
> >>> Hans Verkuil <hverkuil@xxxxxxxxx> escreveu:
> >>>
> >>>> On 26/08/17 13:53, Mauro Carvalho Chehab wrote:
> >>>>> The documentation doesn't mention if vdev-centric hardware control
> >>>>> would have subdev API or not.
> >>>>>
> >>>>> Add a notice about that, reflecting the current status, where three
> >>>>> drivers use it, in order to support some subdev-specific controls.
> >>>>
> >>>> I posted a patch removing v4l-subdevX support for cobalt. It's only
> >>>> used within Cisco, so this is safe to do and won't break any userspace
> >> support.
> >>>
> >>> OK.
> >>>
> >>>> atmel-isc is another driver that creates subdev nodes. Like cobalt,
> >>>> this is unnecessary. There are no sensors that use private controls.
> >>>
> >>> The question is not if the driver has private controls. Private
> >>> controls can be V4L2 device node oriented.
> >>>
> >>> The real question is if userspace applications use subdevs or not in
> >>> order to set something specific to a subdev, on a pipeline where
> >>> multiple subdevs could use the same control.
> >>>
> >>> E. g. even on a simple case where the driver would have something like:
> >>>
> >>> sensor -> processing -> DMA
> >>>
> >>> both "sensor" and "processing" could provide the same control (bright,
> >>> contrast, gain, or whatever). Only by exposing such control via subdev
> >>> is possible to pinpoint what part of the hardware pipeline would be
> >>> affected when such control is changed.
> >>
> >> In theory, yes. In practice this does not happen for any of the V4L2-
> >> centric drivers. Including for the three drivers under discussion.
> >>
> >>>
> >>>> This driver is not referenced anywhere (dts or board file) in the
> >> kernel.
> >>>> It is highly unlikely anyone would use v4l-subdevX nodes when there
> >>>> is no need to do so. My suggestion is to add a kernel option for this
> >>>> driver to enable v4l-subdevX support, but set it to 'default n'.
> >>>> Perhaps with a note in the Kconfig description and a message in the
> >>>> kernel log that this will be removed in the future.
> >>>>
> >>>> The final driver is rcar_drif that uses this to set the "I2S Enable"
> >>>> private control of the max2175 driver.
> >>>>
> >>>> I remember that there was a long discussion over this control. I
> >>>> still think that there is no need to mark this private.
> >>>
> >>> The problem with I2S is that a device may have multiple places where
> >>> I2S could be used. I don't know how the rcar-drif driver uses it, but
> >>> there are several vdev-centric boards that use I2S for audio.
> >>>
> >>> On several of the devices I worked with, the I2S can be enabled, in
> >>> runtime, if the audio signal would be directed to some digital output,
> >>> or it can be disabled if the audio signal would be directed to some
> >>> analog output. Thankfully, on those devices, I2S can be indirectly
> >>> controlled via either an ALSA mixer or via VIDIOC A/V routing ioctls.
> >>> Also, there's just one I2S bus on them.
> >>>
> >>> However, on a device that have multiple I2S bus, userspace should be
> >>> able to control each of them individually, as some parts of the
> >>> pipeline may require it enabled while others may require it disabled.
> >>> So, I strongly believe that this should be a subdev control on such
> >>> hardware.
> >>>
> >>> That's said, I don't know how rcar_drif uses it. If it has just one
> >>> I2S bus and it is used only for audio, then VIDIOC A/V routing ioctls
> >>> and/or an ALSA mixer could replace it. If not, then it should be kept
> >>> as-is and the driver would need to add support for MC, in order for
> >>> applications to identify the right sub-devices that are associated
> >>> with the pipelines where I2S will be controlled.
> >>
> >> Ramesh, do applications using rcar_drif + max2175 have to manually enable
> >> the i2s? Shouldn't this be part of the device tree description instead?
> >>
> >
> > Yes, applications have to control this explicitly. It is not only enable but also disable control is used at run time and hence DT is not applicable.
> >
> > rcar_drif has two registers to write to enable rx on two data pins. It expects a sequence where the master stops output (in this max2175 i2s output - disable) - enable rcar_drif rx and then the master starts output (max2175 i2s output - enable). The application ensures this sequence today. It is one I2S bus and it is not used for audio but raw I/Q samples from max2175 tuner.
> >
> > The v4l2_subdev_tuner_ops does not have .s_stream api as in v4l2_subdev_video_ops and v4l2_subdev_audio_ops. If we plan to have one this functionality may be hidden inside it and no need for an explicit control. I too do not like a private control option.
>
> I think it would be reasonable to use the audio ops s_stream for this. We're
> streaming data after all. The audio ops most closely fits what we want to do.
>
> All this is an internal API, so can be changed in the future if needed.
>
> I like that a lot better than this weird control.

Yeah, I'm all in favor of implicitly control it.

>
> What do you think, Mauro?

Right now, we have .s_stream on both audio and video. I guess we could
just move it to v4l2_subdev_core_ops, as all subdevs may provide a
way to start streaming.

Are there any reason why we have separate s_stream callbacks
right now?

Thanks,
Mauro