Also, silly titles. Atomic has taken of for real, right now there’s 17 drivers supporting atomic modesetting merged into the DRM subsystem. And still a pile of them each release pending for review&merging. But it’s not just new drivers, there’s also been a steady stream of small improvements over the past year, I think it’s time for an update.
It seems small, but a big improvement made over the past few months is that most driver callbacks used by the helper libraries are now optional. Which means tons and tons of dummy functions and boilerplate code can be removed from drivers, leading to less clutter and easier to understand driver code. Aside: Not all drivers have been decluttered, doing that is great starter project for contributing a first few patches to the DRM subsystem. Many thanks to Boris Brezillion, Noralf Trønnes and many others for making this happen.
A long standing complaint about the DRM kernel mode setting is that it’s too complicated, especially compared to fbdev when all you have is one dumb framebuffer and nothing else. And yes, in that case there’s really no point in having distinct CRTC, plane and encoder objects, and now finally Noralf Trønnes has volunteered to write a helper library for simple display pipelines to hide all that complexity from drivers. It’s not yet merged but I’m postive it’ll land in 4.8. And it will help to make writing DRM drivers for simple hardware easy and the driver code clutter-free.
Another piece many dumb framebuffer drivers need is support for manually uploading new contents to the screen. Often on this simple panels there’s no point in doing page-flipping of entire buffers since a real render engine is nowhere to be seen. And the panels are often behind a really slow bus, making full screen uploads to expensive. Instead it’s all done by directly drawing into the frontbuffer, and then telling the driver what changed so that it can shovel the new bits over the bus to the panel. DRM has full support for this through a dirty interface and IOCTL, and legacy fbdev also has some support for this. But the fbdev emulation helpers in DRM never wired these two bits together, forcing drivers to all type their own boilerplate. Noralf has fixed this by implementing fbdev deferred I/O support for the DRM fbdev helpers.
A related improvement is generic support to disable the fbdev emulation from Archit Tajena, both through a Kconfig option and a module option. Most distributions still expect fbdev to work for the boot splash, recovery console and emergency logging. But some, like ChromeOS, are entirely legacy-free and don’t need any of this. Thus far every DRM driver had to add implement support for fbdev emulation and disabling it optionally itself. Now that’s all done in the library using dummy stub functions in the disabled case, again simplifying driver code.
Somehow most ARM-SoC display drivers start out their system suspend/resume support with a dumb register save/restore. I guess because with simple hardware that works, and regmap provides it almost for free. And then everyone learns the lessons why the atomic modeset helpers have a very strict state transition model the hard way: Display hardware gets upset extremely easily when things are done in the wrong order, or without the required delays, obeying the depencies between components and much more. Dumb register restoring does none of that. To fix this Thierry Redding implemented suspend/resume helpers for atomic drivers. Unfortunately not many drivers use this support yet, which is again a nice opportunity to get a kernel patch merged if you have the hardware for such a driver.
Another big gap in the original atomic infrastructure that’s finally getting close is generic support for nonblocking commits. The tricky part there is getting the depency tracking between commits on different display parts right, and I secretly hoped that with a few examples it would be easier to implement something that’s useful for most drivers. With 17 examples I’ve finally run out of excuse to postpone this, after more than 1 year.
But even more important than making the code prettier for atomic drivers and removing boilerplate with better helpers and libraries is in my opinion explaing it all, and making sure all drivers work the same. Over the past few months there’s been massive sphinx-based documentation toolchain - the above links are already generated using that for a peek at all the new pretty.
The flip side is testing, and on that front Collabora’s effort to convert all the kernel mode-setting tests in Tomeu Vizoso’s blog on validating changes to KMS drivers.
Finally it’s not all improvements to make it easier to write great drivers, there’s also some new feature work. Lionel Landwerlin added new atomic properties to implement color management support. And there’s work on-going to implement Z-order and blending properties, and lots more, but that’s not yet ready for merging.