During KS/2012, there was the first Media Workshop in 2012. The discusion notes are now available.
Presentations for the topics discused there are available here.
1. V4L2 Ambiguities
The first topic is with regards to V4L2 API ambiguities that need to be addressed, in order to avoid inconsistent behavior between drivers.
1.1. Split off controls
Split off the control part from videodev2.h. Controls are almost 30% of videodev2.h (and growing). Maintaining controls would be easier if they are moved to e.g. linux/v4l2-controls.h which is included by videodev2.h
Results: OK, split them from videodev2.h to v4l2-controls.h
1.2. Driver-specific controls
Should driver-specific controls have a unique CID, or can the CIDs overlap?
Currently there are three types of controls: standard controls, controls that are specific to a chipset (e.g. cx2341x, mfc51) and driver-specific controls. The controls of the first two types have well defined and unique IDs. For driver-specific controls however there are no clear rules.
It all depends on one question: should driver-specific controls have a unique control ID as well, or can they overlap with other drivers?
If the answer is that they should be unique as well, then all driver-specific controls will have to be defined in a single header so you can be certain that there is no overlap.
New controls should not overlap.
Having all driver-specific controls in a single header file would probably be overkill.
We can instead reserve a range of CIDs for each driver, and define the range base CID only in a common header file.
Driver-specific CIDs themselves would be definedin driver-specific headers.
What should VIDIOC_STREAMON/OFF do if the stream is already started/stopped? I believe they should do nothing and just return 0. The main reason for that is it can be useful if an application can just call VIDIOC_STREAMOFF without having to check whether streaming is in progress.
Results: STREAMON/STREAMOFF should return success (0) if the stream is already started/stopped, to avoid potentially breaking existing applications.
1.4. Unsupported formats in TRY_FMT/S_FMT
What should a driver return in TRY_FMT/S_FMT if the requested format is not supported (possible behaviors include returning the currently selected format or a default format).
The spec says this: "Drivers should not return an error code unless the input is ambiguous", but it does not explain what constitutes an ambiguous input. In my opinion TRY/S_FMT should never return an error other than EINVAL (if the buffer type is unsupported) or EBUSY (for S_FMT if streaming is in progress).
Should we make a recommendation whether the currently selected format or a default format should be returned?
One proposal is to just return a default format if the requested pixelformat is unsupported. Returning the currently selected format leads to inconsistent results.
TRY_FMT/S_FMT should never return an error when the requested format is not supported. Drivers should always return a valid format, preferably a format that is as widely supported by applications as possible.
Both TRY_FMT and S_FMT should have the same behaviour. Drivers should not return different formats when getting the same input parameters.
The format returned should be a driver default format if possible (stateless behaviour) but can be stateful if needed.
The API spec should let clear that format retuns may be different when different video inputs (or outputs) are selected.
1.5. Empty VIDIOC_QUERYCAP bus_info
VIDIOC_QUERYCAP allows bus_info to be empty. Since the purpose of bus_info is to distinguish multiple identical devices this makes no sense. I propose to make the spec more strict and require that bus_info is always filled in with a unique string.
Results: bus_info should never be empty. The spec should specify the syntax of the field contents for common busses, like PCI e USB.
1.6. Deprecate V4L2_BUF_TYPE_PRIVATE
None of the kernel drivers use it, and I cannot see any good use-case for this. If some new type of buffer is needed, then that should be added instead of allowing someone to abuse this buffer type.
Results: Not used upstream, should be deprecated.
1.7. Buftype checks against node type
A driver that has both a video node and a vbi node using the same struct v4l2_ioctl_ops for both nodes will have to check in e.g. vidioc_g_fmt_vid_cap or vidioc_g_fmt_vbi_cap whether it is called from the correct node (video or vbi) and return an error if it isn't.
Should such checks be added to the V4L2 core? And if so, should we add some additional VFL types? Currently we have GRABBER (video nodes), VBI, RADIO and SUBDEV. But if we want to do proper core checks, then we would also need OUTPUT, VBI_OUT and M2M.
We could also refuse ioctls based on the VFL_TYPE: e.g. VIDIOC_ENCODER_CMD makes no sense for radio/vbi nodes. Easy to implement with the lookup table.
Results: It should first be needed to check whether this could break applications, before enforcing it at the core.
1.8. Remove experimental tag from old drivers
Remove the experimental tag from the following old drivers:
Removing this tag from these drivers might be too soon, though:
Remove the tag from the first set of old drivers.
It is too soon for VIDEO_OMAP3.
With regards to VIDEO_NOON010PC30, it was marked as experimental only because it used VIDEO_V4L2_SUBDEV_API, which in turn depends on EXPERIMENTAL. Otherwise it's pretty standard and simple sensor driver. Hence the experimental tag could probably just be removed.
1.9. Remove ioctl's experimental tag
Remove the experimental tag from the V4L2 API elements:
Video Output Overlay (OSD) Interface (from the section "video Output Overlay Interface").
VIDIOC_ENUM_FRAMESIZES and VIDIOC_ENUM_FRAMEINTERVALS ioctls.
VIDIOC_ENCODER_CMD and VIDIOC_TRY_ENCODER_CMD ioctls.
Results: Remove experimental tag from listed ioctls.
1.10. Wrong timings API use for input
What should VIDIOC_G_STD/DV_PRESET/DV_TIMINGS return if the current input or output does not support that particular timing approach? EINVAL? ENODATA?
This is relevant for the case where a driver has multiple inputs/outputs where some are SDTV (and support the STD API) and others are HDTV (and support the DV_TIMINGS API).
I propose ENODATA.
Results: EINVAL is not a good error code, ENODATA is a better possibility. Let's stick with the latest one.
1.11. New defines for DV timing API
Add these new defines:
#define V4L2_IN_CAP_TIMINGS V4L2_IN_CAP_CUSTOM_TIMINGS
#define V4L2_OUT_CAP_TIMINGS V4L2_OUT_CAP_CUSTOM_TIMINGS
Since DV_TIMINGS is now used for any HDTV timings and no longer just for custom, non-standard timings, the word "CUSTOM" is no longer appropriate.
Rename them to V4L2_IN_CAP_TIMINGS and V4L2_OUT_CAP_TIMINGS
Turn V4L2_(IN|OUT)_CAP_CUSTOM_TIMINGS into aliases for V4L2_(IN|OUT)_CAP_TIMINGS
1.12. Sequence and timestamp for output
What should video output drivers do with the sequence and timestamp fields when they return a v4l2_buffer from VIDIOC_DQBUF?
I think the spec is clear with respect to the timestamp: "The driver stores the time at which the first data byte was actually sent out in the timestamp field." For sequence the spec just says: "Set by the driver, counting the frames in the sequence."
So I think that output drivers should indeed set both sequence and timestamp.
For capture and output, the behaviour is OK, the documentation should be clarified.
For mem2mem, we need to specify the behaviour. Timestamp are not actually needed, but it is needed a way to match buffers on both sides, as the codec can reorder them, and timestamps, together with sequence numbers are used for it. An RFC should be sent to the list.
1.13. Const arg for write-only ioctls
Make the argument of write-only ioctls const in v4l2-ioctls.h. This makes it obvious to drivers that they shouldn't change the contents of the input struct since it won't make it back to userspace. It also simplifies v4l2-ioctl.c since it can rely on the fact that after the ioctl call the contents of the struct hasn't changed. Right now the struct contents is logged (if debugging is on) before the ioctl call for write-only ioctls.
1.14. How to set the timestamp?
What is the right/best way to set the timestamp? The spec says gettimeofday, but is it my understanding that ktime_get_ts is much more efficient. Some drivers are already using ktime_get_ts.
(a) all drivers must comply to the spec and use gettimeofday
(b) we change the spec and all drivers must use the more efficient ktime_get_ts
(c) we add a buffer flag V4L2_BUF_FLAG_MONOTONIC to tell userspace that a monotonic clock like ktime_get_ts is used and all drivers that use ktime_get_ts should set that flag.
If we go for (c), then we should add a recommendation to use one or the other as the preferred timestamp for new drivers.
ktime_get_ts() should be used.
Change the spec, and require all drivers to be fixed.
A new driver capability flag is needed to tell applications that the driver uses the monotonic clock.
1.15. Device specified clock
Some devices provide their own timestamp in device clock units. How should that be passed on to userspace?
Proposal: extend the struct v4l2_timecode in v4l2_buffer, either by just adding a new type, or by creating an anonymous union.
Currently there are no drivers that use v4l2_timecode.
It is questionable for v4l2_timecode to be in v4l2_buffer at all since such information will typically be carried over VBI.
Results: Reuse v4l2_timecode to pass the device clock timestamp to userspace (making it an anonymous union is an option, as it would allow different kind of timestamps later) and add a capability flag to inform applications.
If a driver supports only formats with more than one plane, should V4L2_CAP_VIDEO_CAPTURE still be defined? And if a driver also supports single-plane formats in addition to >1 plane formats, should V4L2_CAP_VIDEO_CAPTURE be compulsary?
The consensus seems to be 'No' and 'No'. The CAPs refer to whether the single and/or multiplanar API is implemented, not whether single and/or multiplanar formats are available.
Results: Flags should reflect what it is visible/usable by the userspace API; reviewing v4l-utils with that regards may be needed, due to emulation code that could be there.
VIDIOC_CROPCAP: the spec says that CROPCAP must be implemented by all capture and output devices (Section "Image Cropping, Inserting and Scaling"). In reality only a subset of the drivers support cropcap.
Should cropcap really be compulsory? Or only for drivers that can scale? And in that case, should we make a default implementation for those drivers that do not support it? (E.g.: call g_fmt and use the width/height as the default and bounds rectangles, and set the pixel aspect to 1/1)
Results: the ioctls should not be mandatory if the driver doesn't support crop or scale.
1.18. Pixel aspect ratio
Pixel aspect: currently this is only available through VIDIOC_CROPCAP. It never really belonged to VIDIOC_CROPCAP IMHO. It's just not a property of cropping/composing. It really belongs to the input/output timings (STD or DV_TIMINGS). That's where the pixel aspect ratio is determined.
While it is possible to add it to the dv_timings struct, I see no way of cleanly adding it to struct v4l2_standard (mostly because VIDIOC_ENUMSTD is now handled inside the V4L2 core and doesn't call the drivers anymore).
An alternative is to add it to struct v4l2_input/output, but I don't know if it is possible to defined a pixelaspect for inputs that are not the current input.
What I am thinking of is just to add a new ioctl for this VIDIOC_G_PIXELASPECT.
Results: TBD on the mailing list.
1.19. Tuner ownership
How to handle tuner ownership if both a video and radio node share the same tuner?
Calling S_FREQ, S_TUNER, S_MODULATOR or S_HW_FREQ_SEEK will make the filehandle the owner if possible. EBUSY is returned if someone else owns the tuner and you would need to switch the tuner mode.
Ditto for ioctls that expect a valid tuner configuration like QUERYSTD. This is likely to be driver dependent, though.
Just opening a device node should not switch ownership
But it is not clear what to do when any of these ioctls are called:
G_FREQUENCY: could just return the last set frequency for radio or TV: requires that that is remembered when switching ownership. This is what happens today, so G_FREQUENCY does not have to switch ownership.
G_TUNER: the rxsubchans, signal and afc fields all require ownership of the tuner. So in principle you would want to switch ownership when G_TUNER is called. On the other hand, that would mean that calling v4l2-ctl --all -d /dev/radio0 would change tuner ownership to radio for /dev/video0. That's rather unexpected.It is possible to just set rxsubchans, signal and afc to 0 if the device node doesn't own the tuner. I'm inclined to do that.
Should closing a device node switch ownership? E.g. if nobody has a radio device open, should the tuner switch back to TV mode automatically? I don't think it should.
How about hybrid tuners?
Results: TBD on the mailing list (or later this week).
2. V4L2 Compliance
An automated tool to test whether a driver complies with the spec was written, in order to allow being sure that drivers will match the V4L2 spec.
Known missing tests at V4L2 compliance:
Cropping: VIDIOC_CROPCAP, VIDIOC_G/S_CROP, VIDIOC_G/S_SELECTION
Framebuffer: VIDIOC_S_FBUF, OVERLAY
Encoder/Decoder: VIDIOC_(TRY_)ENCODER_CMD, VIDIOC_(TRY_)DECODER_CMD, VIDIOC_G_ENC_INDEX
Streaming: VIDIOC_QBUF, DQBUF, QUERYBUF, PREPARE_BUFS, VIDIOC_STREAMON/OFF
Subdev API, MC API (but these are possibly out-of-scope anyway)
Extra tests: it requires:
Normal and extended controls
videobuf2 (or an own video buffers implementation that is spec-compliant, videobuf1 isn"t)
Passing v4l2-compliance tests will be compulsory for new drivers under drivers/media.
The enforcement of the extra tests are ok. However, the usage of videobuf2 can't be mandatory yet for all drivers, as VB2 currently lacks support for videobuf-dvb. So, hybrid drivers may use VB1 instead, while this is not fixed on VB2.
Lots of work is required to make existing drivers compliant; as almost all fails, but merging newer drivers without passing with the tests would mean further re-work, with should really be avoided.
In order for it to work, it would also require that new tests should be added whenever a new API appears.
The V4L2 streaming profile needs to be implemented to support complex MC drivers in v4l2-compliance.
There are some pending issues related to the needs at ALSA by V4L2 drivers.
3.1. Compressed audio (ivtv, pvrusb2)
Currently, two audio drivers (ivtv and pvrusb2) currently provides pure audio mpeg TS streams, either at video or radio node, in order for the cards to work with FM radio. This is an API violation that should be fixed. Also, newer radio USB sticks might also have compressed radio support. So, it is needed a way to move the existing devices and enforce newer ones to use ALSA api for audio-only output.
During the discussions, Takashi Iwai (ALSA maintainer) said that the compressed audio format support for ALSA, including sound drivers, will be supported in 3.7. Patches to add support in driver were added on Aug, 20.
Yet, the userspace ALSA library currently doesn't support yet, as it currently on a separate library: tinycompress
Vinod Koul (firstname.lastname@example.org) is working with the userspace library. It seems that there's no technical reason why not adding it at alsalib, so it was proposed that this should be merged at alsalib, in order to allow the drivers conversions.
The headers for the kernel API are inside the git://git.kernel.org/pub/scm/linux/kernel/git/broonie/sound.git tree, at:
Userspace needs to sync audio and video. V4L2 has timestamps already. Currently, alsalib provides the PCM_STATUS: current position, system time. There is no device timestamp at kernelspace, however; the time shift is calculated by knowing the sampling time.
In opposite to pure audio devices, on audio and video devices, when capture starts, the audio can start later than video, or together with it, depending on the driver. Such time difference between audio and video could be calculated in order to adjust the initial drift, to be used as the initial timestamp for the start of the audio.
However, audio/video delay may not remain constant. For example, when changing inputs, the audio can cut off while the input change is in progress, so there need to be a way to resync the audio timestamp. Also, small time drifts may happen during audio/video stream lifetime. So, it should be possible to have a way for the driver to periodically tag the audio stream with the time reference.
In the case of compressed audio, the API has a timestamp already. So, no changes there should be needed.
Pierre-Louis Bossart is working on adding device timestamps to ALSA, just as
Laurent is doing for V4L2. If these device timestamps are both available, then
syncing between audio and video will be much easier.
3.3. Usage of media API by ALSA
There's a requirement of using the media API also by ALSA, as this is needed on some embedded devices. As there will be some discussions about that theme at LPC, on Wednesday 11:40, at Nautilus 2 room.
It was decided to not talk about that on this time slot.
There is a preliminary ALSA RFC patches with such proposal here.
And the core DMABUF documentation in the kernel is available in Documentation/dma-buf-sharing.txt.
ST-Ericsson mentioned that it is important that the Media Controller API as it is going to be used in ALSA will use the same API as is used when that API is applied to V4L2. ST-Ericsson also mentioned that they have done some work on Media Controller connector entities.
5. V4L2/DVB Issues from Userspace Perspective
Rémi Denis-Courmont gave some feedback from a userspace perspective, as seen on his presentation.
What is the best way to deal with the V4L2 API additions introduced in pretty much every kernel?
The recommendation was to just copy videodev2.h to the application.
5.2 libv4l2 release cycle
The current release version of libv4l2 is quite old compared to the master branch. Distros will typically only update when a new release is made, and it would be useful if the libv4l2 release cycle is done more often.
There are many controls, but it is not easy to determine which ones are relevant for userspace applications.
Including this information in profiles seemed the best solution.
5.4 libv4l2 doesn't support userptr.
When using the libv4l2 wrapper only the MMAP memory model is supported at the moment, and not the USERPTR model, even if the driver does support that. Why is this, and can it be fixed?
6 Adding CEC support to the kernel
HDMI CEC is used to control other HDMI devices (TV, Bluray, recorders, etc.) but also since HDMI 1.4a to control ethernet, the audio return channel and the hotplug signal.
After a presentation on what CEC does we discussed how to implement this should the CEC API be done separately through DRM or V4L2 devices, or should new /dev/cecX be created? The consensus was that that was the best approach.
Different subsystems can register themselves with the CEC device and intercept messages relevant to them. The core CEC implementation will do the standard CEC discovery part. Any messages not handled by the kernel can go to userspace.
7 Media Controller Library
Laurent presented the work he and Sakari did on the media controller library.
Sakari worked on the pipelining part: an implementation that can figure out the default pipeline that should be set up on an embedded device. The results were very promising.
Laurent worked on a Media Device Discovery Library: it enumerates the available media devices, using the Media Controller, libudev and/or sysfs. This too was very promising.
Both pieces are still work-in-progress though.
8 Android V4L2 Cam Library
Guennadi presented his Android HAL library on topof V4L2.
Currently preview, capture and still image support is working. Quite slow though due to missing support for hardware acceleration such as colorspace conversion.
Guennadi came across some missing V4L2 functionality that would be needed to do a full Android implementation. Guennadi will look further into this and provide a list of what's missing.
9 DT x V4L2
An RFC to add Device Tree bindings to the V4L2 subsystem has been discussed. Version 4 of that RFC has been developed jointly by Sylwester Nawrocki and Guennadi Liakhovetski and posted to the linux-media mailing list shortly before the conference. Both V4L2 developers and developers from other areas of the kernel with rich experience in designing device-tree bindings took part in the discussion and provided their valuable comments.
The high-level design of the V4L2 DT implementation defines DT nodes of multiple types: camera sensors, TV decoders and encoders and similar devices, typically controlled over I2C or a similar bus, should also reside on their respective (control) busses. Video bridge engines and
other data processing units, typically implemented in respective SoCs, are also each represented by a separate DT node. An important part of the media hardware representation is a description of data links between all the above parts, especially between the SoCs and external components. Such links are represented by child nodes of both receiver and transmitter DT nodes.
During the conference the general concept has been approved. Multiple improvement suggestions have been provided for clock bindings within media DT nodes, property naming, location of the bus-width property and similar properties, describing the data bus topology. An important amendment to the RFC is the addition of a "port" DT node level. Now the afore mentioned link sub-nodes should be placed under these port nodes.
As a result of the discussion, version 5 of the RFC has been designed and posted.
After further cosmetic improvements it should lead to a proper patch, adding V4L2 DT bindings to the kernel documentation and their implementation in the media drivers.