In order to help writing a media summit report, I'm sending the raw notes that were at Etherpad after the end of the meeting.
Thanks, Mauro Media summit 25-10-2018 Edinburgh https://www.spinics.net/lists/linux-media/msg141095.html Attendees Brad Love Ezequiel Garcia Gustavo Padovan Hans Verkuil Helen Koike Hidenori Yamaji Ivan Kalinin Jacopo Mondi Kieran Bingham Laurent Pinchart Mauro Chebab Maxime Ripard Michael Grzeschik Michael Ira Krufky Niklas Söderlund Patrick Lai Paul Elder Peter Griffin Ralph Clark Ricardo Ribalda Sakari Ailus Sean Young Seung-Woo Kim Stefan Klug Vinod Koul Topics CEC (Hans Verkuil) 1 pin bus, cec-gpio error injection support Tda998x (including BeagleBoard Bone support) ChromeOS CEC support DisplayPort CEC Tunneling over AUX for i915, nouveau, amdgpu Still lots of work happening around CEC utilities, the compliance test is in continuous use at cisco. URL to CEC status: https://hverkuil.home.xs4all.nl/cec-status.txt CISCO using CEC in their products, and pushing for development of CEC upstream. Automated test pipelines in place utilising cec-compliance 1.4 spec can be found quite easily online (http://d1.amobbs.com/bbs_upload782111/files_51/ourdev_716302E34B9Q.pdf) - Hans will tell us diff to 2.0 RC-Core progress report (Sean Young) Ported/removed all staging lirc drivers BPF IR decoding support. rc-core protocol decoders only support the most common protocols old lirc userspace rarely required Linux supports more IR hardware than any other operating system. Future/TODO Some driver work needed (mostly for older hardware) Build dvb without rc-core More BPF protocol decoders Support 2.4ghz remotes Needs automated testing Support with gpio and simple ir led is available Persistent storage of controls (Ricardo Ribalda - Qtec) Sensors usually come with some calibration data from the manufacturer about dead pixels etc. How to get this calibration data into the sensor. proposal: - new vid interface /dev/videoX /dev/v4l-subX - s_ext_ctrls - write array of ctrls - s_ctrl - replace specfific ctrl - dt - locate ctrl vals - (reuse v4l2 ioctls for save/restore/get value(s)) - implementation in two parts: main driver intf, storage why not set sensor vals from userspace? - user might impl wrong - if part of driver can make sure they work for specific sensor no get_min/max coz storage might not have space to save all ctrls NVMEM might be good for storage rw - Is it a filesystem (or is that more overhead) - controls validate the data in the same way as they are validated by the device sensor. Maintenance process tools - DIM (Laurent Pinchart - Ideas on Board) DRM have switched to co-maintainer model - Burden is shared - Increased commit rights to the mainline (multi-maintaner model) - no enforeced maintainer process (each their own) - multi-comitter means each has to do checks - need to standardize tools -> DIM - DIM wrapper for git kindaish - DRM inglorious maintenance tool : https://01.org/linuxgraphics/gfx-docs/maintainer-tools/dim.html - conflict resolution by commiter, not maintainer - DIM uses conflict resoultion system in git to pre-prepare the merge conflicts in a shared cache. - tool enforces quality checks drawbacks - pushing = testing = rebuilding :/ -> becomes a burden for the committers - Patches are integrated from a patchwork instance, and the commits then reference the history of the patch upstream - Adds an extra tool that must be used, and then an extra delay. - eg. takes time/difficult to ack easyish to get commit rights - Responsibility to perform correctly - spreads the review workload (creates review-economics) - trading rb tags -> lower quality :/ - "please submit patches" "also please merge conflict resolution and rebuild" - If automated testing can validate the process, it can simplify commit process for multi-commiters. - auto testing on test/build farm had this model in media for a few years - somebody merged 9k lines and lost a handful of devs - probably exception, model fine - many comitters but few maintainers - problem could be technically resolved too few ppl reviewing patches - wait time measured in months :) - multi-commiter model try to solve this problem incentive to get more maintainers - commit rights - responsilibity - get attached to code :) Automated testing (Ezequiel Garcia, Gustavo Padovan, Sakari Ailus) Testing only one config, one version, one patch application, one hardware (one test to rule them all). Sometimes testing is omitted altogether by a developer, or it has been done before making changes to a patch. A lot of problems will only be found in the additional validation steps Mauro now performs, or when the patches already have reached the master branch. We could do better. Ideal CI: patch submission -> auto test pass -> review +1, +2 -> merged! :) Isn't the core question "what level of quality standards to we want to enforce" ? The maintainance process should be modelled around the answer to that question, not the other way around. Automated testing can be part of the quality standard enforcement procedure. Three steps problem: Define the quality standards Define how to quantify them Define how to enforce them. Automated testing is part of the enforcement. Auto testing status Just v4l2- and cec- compliance Virtual device drivers (vivid, vimc, vicodec) https://linuxtv.org/wiki/index.php/Media_Open_Source_Projects:_Looking_for_Volunteers v4l2-compliance The good: * complete in terms of API testing * quick and easy to run * nice human readable machine parsable output * new drivers are required to pass the test The bad: * no stateful/statless codec support * no sdr and touch support frequently updated * only one contributor Lack of a DVB: virtual driver and dvb-compliance. Added the need of write a virtual driver and a DVB compliance tool in order to check Digital TV core to the list of things to be done at https://linuxtv.org/wiki/index.php/Media_Open_Source_Projects:_Looking_for_Volunteers. * Test format output consolidation (Test summit are working on this) Ezequiel reported that some people are complaining that v4l2-compliance is updated too frequently, but this is unavoidable according to Hans. Mauro proposed moving v4l2-compliance to the kernel source tree but Hans prefers keeping it separate as it makes it easier to develop it. v4l2-compliance has a single contributor. We should push developers submitting new APIs to implement the corresponding tests. To achieve this we first need to clean up the code base to make contributions easier. There is currently no internal API documentation, and source filies are quite big. Splitting code into separate unit tests could be useful. There are still lots of compliance tests to be developped (MC-compliance for instance). Should we have them all in a single tool or in separate tools ? Other test options kselftest Do we need a v4l2-compliance kselftest ? kunit gst-validate ktf (https://github.com/oracle/ktf, http://heim.ifi.uio.no/~knuto/ktf/) Most of our tests are low-level, do we want high-level testing ? e.g. gst-validate-1.0 --set-scenario pause_resume videotestsrc ! v4l2fwhtenc ! v4l2fwhtdec ! fakevideosink Testing of real hardware v4l2-compliance has basic support for testing that captured images don't have colours swapped Testing real hardware adds dependencies on both hardware access and possibly control of the external environment (e.g. lightning). Is this out-of-scope for KernelCI ? KernelCI - kernel CI joining LinuxFoundation recently: extend test use cases and LTS - UVC target of build integration testing (report on linux-media mailing list: RFC) - V4L testing first test beyond boot testing https://www.mail-archive.com/linux-media@vger.kernel.org/msg135787.html - MIssing components to identify driver under test - test case in v4l2-compliance output parse - Need to collaborate with kernelCI to decide test report frequency and triggering: eg. weekly report - To discuss if report regressions only or new driver/features testing - Tests could be run on any branch (but only triggered by changes: eg new commits) - Test same branch with different configurations (eg. plain V4L2, V4L2+MC, DVB with no MC etc) - Number of possible tests only depends on how many lab instances are available - 32bit compliance test utility against a 64bit kernel. - Might require simplying build of test tools (for different configurations, eg 32 vs 64 bits) - Virtual machines used to run tests are LAVA slaves (created by LAVA itself) - "Labs" instances are distributed and federated to KernelCI, each lab tests specific boards - DRM uses custom FPGA (-not sure I got this right-) to compare actual output wiht the expected one Post-presentation talk on testing Testing master branch is not useful Problems MUST be found before the patches make it to the (sub-)maintainer trees, otherwise it's too late Testing developers' branches All branches in a tree Individual developers get their code tested, even before review on list Maintainers Merging rc-1 after release works the same way Test development branch that contains the rc-1, only then push to the media tree master branch Multiple trees for folks who have a large number of branches or branches that are updated often which do not need to be tested Eveyrone: reply Ezequiel's mail on the list! Stateless codecs Support for stateless codecs will be merged for v4.20 as a staging driver being used by AllWinner how to support in userspace? bit parsing is application-specific - so there's no point in providing it should there be simple parser for compliance-testing? All main applications support libva, developed originally by Intel to provide access to their accelerators. libva backend was written for stateless cedrus decoder Mainly allwinner, but some work on Rockchip stateless codec Requirements for removal from staging is at least one other decoder and encoder using the API. API is not stable atm so should expect it to change. V4L2 Create better ioctls VIDIOC_G/S_PARM - really only used for get/set frameinterval - add VIDIOC_G/S_FRAME_INTERVAL - Just use the existing IOCTL argument memory layout, but rework the struct and rename it (for compatibility reasons the old IOCTL stays) why separate API to set ctrl? Input number VIDIOC_ENUM_FRAMEINTERVALS struct v4l2_frmivalenum - why have stepwise? - makes little sense for frame intervals - Stepwise could be removed (Qualcomm venus codec and uvc only users) - Add more types to tell the unit - add buf_type field in struct v4l2_frmivalenum and struct v4l2_frmsizeenum ns vs fracs - frac more precise - but ns is unambiguous - Facebook(OculusVR) defined a 'flick' as a common denominator of all values : https://techcrunch.com/2018/01/22/facebook-invented-a-new-time-unit-called-the-flick-and-its-truly-amazing/ https://github.com/OculusVR/Flicks Combine ENUM_FMT, ENUM_FRAMESIZES and ENUM_FRAMEINTERVALS make ioctl to combine all three that returns all possiblities/combinations wihout having to enumerate everything is it worth the effort? - might be big eg. vivid 1710 combinations - Memory requirements of the argument are not an issue as such: video buffers will require more memory anyway. - Don't change now. But revisit, if any of the following happens - Larger (or new) IOCTL struct is needed for any of the IOCTLs - If this turns out to be a performance issue somewhere struct v4l2_buffer problems: - timestamp is not 2038-safe - planes - single vs multi plane handing is really different and a mess proposal: new struct v4l2_buffer - __u64 timestamp - struct timeval could be remade - but then need to recompile - u64 simpler - also simplifies for single-multi-planar issue - modernization - support for 2038-safe struct v4l2_buffer needed in any case - struct v4l2_ext_plane planes[VIDEO_MAX_PLANES] - put planes arraywithint struct v4l2_buffer ideas: - improve plane description: offset at start of plane, padding at end of plane, fd refer to one or all planes (-1 for remaining if allowed by SoC, eg exynos4) - simplify codecs resolution change - add width/height for buffer? - per-plane pixelformat? - maybe not? (video and non-video in same buffer?) - get rid of difference between single and multiplanar - multiplanar becoming more normal - Currently the driver needs to pick which one to support. Compatibility support for the single-planar API was merged but later removed? Re-introduce it? struct v4l2_create_buffers only format.pix.sizeimage is used want to simplify - coz struct v4l2_format is kinda big - Create buffers one at a time - non-contiguous range; need to tell the index of the allocated buffer - nicely symmentrical to deleting buffers one at a time proposal: more generic stuff: - VIDIOC_CREATE_BUF - with just size per plane instead of format (if 0 then use current format's size like REQBUFS) - VIDIOC_DELETE_BUF <- currently canoot be done at all, might need it in a few years anyway - VIDIOC_DELETE_ALL_BUFS struct v4l2_format and VIDIOC_G/S/TRY_FMT v4l2_pix_format vs v4l2_pix_format-mplane - painful ...that's it? Fault Tolerant V4L2 problem: 8 cameras, one dies, they all cannot be used :/ V4L2 designed to work only when all devs working - but still need fault tolerance :/ - Media events to tell when new devices appear - A change in the graph increments the media graph version number (G_TOPOLOGY IOCTL) - Way to tell the user whether the registration is complete or not - Events are not enough: the application may not have subscribed such an event before it arrives - Media device info flag? - The property API might be usable for the purpose --- placeholders for entities that have not yet appeared? - Known enitites which are failed can be created in the Media Graph and marked disabled or 'failed' - Those entities could then have an operation to 'reprobe' - Events should be available on the media node - but the API should also be able to query state. Complex Cameras how to make generic apps work with complex hw and mc -> libcamera we gonna have more complex cameras in notebooks coz easier/cheaper to make - so we need to do this or else we're going to lose laptop webcams dev process complaints? - ppl want recognition for being involved - have intermediate multicommit tree? - ask ppl if they want commit rights rather than ask them to review - need rules for multicommit - technical, not ethical - maintainers > committers -> maintainers discuss before making decisions - commit-to-master rights vs merge rights Linuxtv.org - Hosted in a virtual machine somewhere in a German university - The long term need is to move the virtual machine elsewhere Thanks, Mauro
