Re: [PATCH v4 29/36] media: imx: mipi-csi2: enable setting and getting of frame rates

[Sakari Ailus]

Hi Russell,

On Tue, Feb 21, 2017 at 01:21:32PM +0000, Russell King - ARM Linux wrote:
> On Tue, Feb 21, 2017 at 02:37:57PM +0200, Sakari Ailus wrote:
> > Hi Russell,
> > 
> > On Tue, Feb 21, 2017 at 12:13:32AM +0000, Russell King - ARM Linux wrote:
> > > On Tue, Feb 21, 2017 at 12:04:10AM +0200, Sakari Ailus wrote:
> > > > On Wed, Feb 15, 2017 at 06:19:31PM -0800, Steve Longerbeam wrote:
> > > > > From: Russell King <rmk+kernel@xxxxxxxxxxxxxxx>
> > > > > 
> > > > > Setting and getting frame rates is part of the negotiation mechanism
> > > > > between subdevs.  The lack of support means that a frame rate at the
> > > > > sensor can't be negotiated through the subdev path.
> > > > 
> > > > Just wondering --- what do you need this for?
> > > 
> > > The v4l2 documentation contradicts the media-ctl implementation.
> > > 
> > > While v4l2 documentation says:
> > > 
> > >   These ioctls are used to get and set the frame interval at specific
> > >   subdev pads in the image pipeline. The frame interval only makes sense
> > >   for sub-devices that can control the frame period on their own. This
> > >   includes, for instance, image sensors and TV tuners. Sub-devices that
> > >   don't support frame intervals must not implement these ioctls.
> > > 
> > > However, when trying to configure the pipeline using media-ctl, eg:
> > > 
> > > media-ctl -d /dev/media1 --set-v4l2 '"imx219 pixel 0-0010":0[crop:(0,0)/3264x2464]'
> > > media-ctl -d /dev/media1 --set-v4l2 '"imx219 0-0010":1[fmt:SRGGB10/3264x2464@1/30]'
> > > media-ctl -d /dev/media1 --set-v4l2 '"imx219 0-0010":0[fmt:SRGGB8/816x616@1/30]'
> > > media-ctl -d /dev/media1 --set-v4l2 '"imx6-mipi-csi2":1[fmt:SRGGB8/816x616@1/30]'
> > > Unable to setup formats: Inappropriate ioctl for device (25)
> > > media-ctl -d /dev/media1 --set-v4l2 '"ipu1_csi0_mux":2[fmt:SRGGB8/816x616@1/30]'
> > > media-ctl -d /dev/media1 --set-v4l2 '"ipu1_csi0":2[fmt:SRGGB8/816x616@1/30]'
> > > 
> > > The problem there is that the format setting for the csi2 does not get
> > > propagated forward:
> > 
> > The CSI-2 receivers typically do not implement frame interval IOCTLs as they
> > do not control the frame interval. Some sensors or TV tuners typically do,
> > so they implement these IOCTLs.
> 
> No, TV tuners do not.  The frame rate for a TV tuner is set by the
> broadcaster, not by the tuner.  The tuner can't change that frame rate.
> The tuner may opt to "skip" fields or frames.  That's no different from
> what the CSI block in my example below is capable of doing.
> 
> Treating a tuner differently from the CSI block is inconsistent and
> completely wrong.

I agree tuners in that sense are somewhat similar, and they are not treated
differently because they are tuners (and not CSI-2 receivers). Neither can
control the frame rate of the incoming video stream.

Conceivably a tuner could implement G_FRAME_INTERVAL IOCTL, but based on a
quick glance none appears to. Neither do CSI-2 receivers. Only sensor
drivers do currently.

> 
> > There are alternative ways to specify the frame rate.
> 
> Empty statements (or hand-waving type statements) I'm afraid don't
> contribute to the discussion, because they mean nothing to me.  Please
> give an example, or flesh out what you mean.

Images are transmitted as series of lines, with each line ending in a
horizontal blanking period, and each frame ending with a similar period of
vertical blanking. The blanking configuration in the units of pixels and
lines at their pixel clock is a native unit which sensors typically use, and
some drivers expose the blanking controls directly to the user.

<URL:http://hverkuil.home.xs4all.nl/spec/uapi/v4l/extended-controls.html#image-source-control-ids>

> 
> > > $ strace media-ctl -d /dev/media1 --set-v4l2 '"imx6-mipi-csi2":1[fmt:SRGGB8/816x616@1/30]'
> > > ...
> > > open("/dev/v4l-subdev16", O_RDWR)       = 3
> > > ioctl(3, VIDIOC_SUBDEV_S_FMT, 0xbec16244) = 0
> > > ioctl(3, VIDIOC_SUBDEV_S_FRAME_INTERVAL, 0xbec162a4) = -1 ENOTTY (Inappropriate
> > > ioctl for device)
> > > fstat64(1, {st_mode=S_IFCHR|0600, st_rdev=makedev(136, 2), ...}) = 0
> > > write(1, "Unable to setup formats: Inappro"..., 61) = 61
> > > Unable to setup formats: Inappropriate ioctl for device (25)
> > > close(3)                                = 0
> > > exit_group(1)                           = ?
> > > +++ exited with 1 +++
> > > 
> > > because media-ctl exits as soon as it encouters the error while trying
> > > to set the frame rate.
> > > 
> > > This makes implementing setup of the media pipeline in shell scripts
> > > unnecessarily difficult - as you need to then know whether an entity
> > > is likely not to support the VIDIOC_SUBDEV_S_FRAME_INTERVAL call,
> > > and either avoid specifying a frame rate:
> > 
> > You should remove the frame interval setting from sub-devices that do not
> > support it.
> 
> That means we end up with horribly complex scripts.  This "solution" does
> not scale.  Therefore, it is not a "solution".

I have to disagree with that: if a piece of hardware does not offer to
control or, if a concept is not even relevant for a piece of hardware, then
a driver for that piece of hardware should not expose an interface to
control such a feature. Doing so would provide no value and at the same time
would be simply confusing for the user space.

> 
> It's fine if you want to write a script to setup the media pipeline using
> media-ctl, listing _each_ media-ctl command individually, with arguments
> specific to each step, but as I've already said, that does not scale.

Pipeline configuration as such is highly hardware specific. There are rules,
but there are details in hardware that have to be taken into account, such
as mandated cropping in certain situations. You have to simply accept that:
when it comes to camera Image Signal Processors, there are no standard
pipelines. Each ISP is different from the rest, more or less, and often more
so.

As the interface is generic, you can write generic programs that use that
interface, but you need to be able to adapt to the differences in the
functionality of the hardware.

Frankly, I think this is just needless noise stemming from a problem that's
not really difficult to solve --- if that technical problem really even
exists. But let's not debate that; I accept that dropping frames is what
you're willing to do.

> 
> I don't want to end up writing separate scripts to configure the pipeline
> for different parameters or setups.  I don't want to teach users how to
> do that either.
> 
> How are users supposed to cope with this craziness?  Are they expected to
> write their own scripts and understand this stuff?
> 
> As far as I can see, there are no applications out there at the moment that
> come close to understanding how to configure a media pipeline, so users
> have to understand how to use media-ctl to configure the pipeline manually.
> Are we really expecting users to write scripts to do this, and understand
> all these nuances?
> 
> IMHO, this is completely crazy, and hasn't been fully thought out.
> 
> > > $ strace media-ctl -d /dev/media1 --set-v4l2 '"imx6-mipi-csi2":1[fmt:SRGGB8/816x616]'
> > > ...
> > > open("/dev/v4l-subdev16", O_RDWR)       = 3
> > > ioctl(3, VIDIOC_SUBDEV_S_FMT, 0xbeb1a254) = 0
> > > open("/dev/v4l-subdev0", O_RDWR)        = 4
> > > ioctl(4, VIDIOC_SUBDEV_S_FMT, 0xbeb1a254) = 0
> > > close(4)                                = 0
> > > close(3)                                = 0
> > > exit_group(0)                           = ?
> > > +++ exited with 0 +++
> > > 
> > > or manually setting the format on the sink.
> > > 
> > > Allowing the S_FRAME_INTERVAL call seems to me to be more in keeping
> > > with the negotiation mechanism that is implemented in subdevs, and
> > > IMHO should be implemented inside the kernel as a pad operation along
> > > with the format negotiation, especially so as frame skipping is
> > > defined as scaling, in just the same way as the frame size is also
> > > scaling:
> > 
> > The origins of the S_FRAME_INTERVAL IOCTL for sub-devices are the S_PARM
> > IOCTL for video nodes. It is used to control the frame rate for more simple
> > devices that do not expose the Media controller interface. The similar
> > S_FRAME_INTERVAL was added for sub-devices as well, and it has been so far
> > used to control the frame interval for sensors (and G_FRAME_INTERVAL to
> > obtain the frame interval for TV tuners, for instance).
> > 
> > The pad argument was added there but media-ctl only supported setting the
> > frame interval on pad 0, which, coincidentally, worked well for sensor
> > devices.
> > 
> > The link validation is primarily done in order to ensure the validity of the
> > hardware configuration: streaming may not be started if the hardware
> > configuration is not valid.
> > 
> > Also, frame interval is not a static property during streaming: it may be
> > changed without the knowledge of the other sub-device drivers downstream. It
> > neither is a property of hardware receiving or processing images: if there
> > are limitations in processing pixels, then they in practice are related to
> > pixel rates or image sizes (i.e. not frame rates).
> 
> So what about the case where we have a subdev (CSI) that is capable of
> frame rate reduction, that needs to know the input frame rate and the
> desired output frame rate?  It seems to me that this has not been
> thought through...

That's because I believe you're the first one wanting to willfully throw
away perfectly good frames. :-)

If you want to do that, a simple option could be to just support
[GS]_FRAME_INTERVAL on all pads of a sub-device that can drop frames. But it
should not be include in pipeline validation, it simply does not belong
there for the reasons stated previously.

The user would be responsible for configuring the frame rates right. That
information would simply be used to configure frame dropping frequency.

I'd like to have a comment from Laurent and Hans on this.

> 
> > >        -  ``MEDIA_ENT_F_PROC_VIDEO_SCALER``
> > > 
> > >        -  Video scaler. An entity capable of video scaling must have
> > >           at least one sink pad and one source pad, and scale the
> > >           video frame(s) received on its sink pad(s) to a different
> > >           resolution output on its source pad(s). The range of
> > >           supported scaling ratios is entity-specific and can differ
> > >           between the horizontal and vertical directions (in particular
> > >           scaling can be supported in one direction only). Binning and
> > >           skipping are considered as scaling.
> > > 
> > > Although, this is vague, as it doesn't define what it means by "skipping",
> > > whether that's skipping pixels (iow, sub-sampling) or whether that's
> > > frame skipping.
> > 
> > Skipping in the context is used to refer to sub-sampling. The term is often
> > used in conjunction of sensors. The documentation could certainly be
> > clarified here.
> 
> It definitely needs to be, it's currently mis-leading.

If you're not familiar with terminology typically used with many camera
sensor, perhaps so. The documentation should indeed not assume that; I'll
submit a patch to fix that.

> 
> > > Then there's the issue where, if you have this setup:
> > > 
> > >  camera --> csi2 receiver --> csi --> capture
> > > 
> > > and the "csi" subdev can skip frames, you need to know (a) at the CSI
> > > sink pad what the frame rate is of the source (b) what the desired
> > > source pad frame rate is, so you can configure the frame skipping.
> > > So, does the csi subdev have to walk back through the media graph
> > > looking for the frame rate?  Does the capture device have to walk back
> > > through the media graph looking for some subdev to tell it what the
> > > frame rate is - the capture device certainly can't go straight to the
> > > sensor to get an answer to that question, because that bypasses the
> > > effect of the CSI frame skipping (which will lower the frame rate.)
> > > 
> > > IMHO, frame rate is just another format property, just like the
> > > resolution and data format itself, and v4l2 should be treating it no
> > > differently.
> > > 
> > > In any case, the documentation vs media-ctl create something of a very
> > > obscure situation, one that probably needs solving one way or another.
> > 
> > Before going to solutions I need to ask: what do you want to achieve?
> 
> Full and consistent support for the hardware, and a sane and consistent
> way to setup a media pipeline that is easy for everyone to understand.

That's essentially what we do have: the same interface is supported on a
large variety of different hardware devices. However, not all IOCTLs are
supported by all device drivers.

-- 
Kind regards,

Sakari Ailus
e-mail: sakari.ailus@xxxxxx	XMPP: sailus@xxxxxxxxxxxxxx