[Update Jan 2011: Udlfb was promoted from the staging tree to the main kernel (in drivers/video/udlfb.c) in 2.6.38. See documentation in the kernel tree at Documentation/fb/udlfb.txt]
A 10-part udlfb patch series with several enhancements were recently accepted into the Linux kernel staging tree of 2.6.34. Greg’s maintainer patch series passing them on starts here.
This catches the official kernel staging tree with everything on git.plugable.com up to Feb, 2010.
Background
udlfb provides Linux framebuffer (fbdev) support plus a private “damage” notification ioctl used by the Roberto DeIoris’ “displaylink” X server. Udlfb automatically detects and supports all current DisplayLink USB graphics chips – 120,160,115,125,165,195. It supports 16bpp, with modes up to the maximum supported by the DisplayLink chips. Plugging any DisplayLink device into a Linux PC with this support should result in a “green screen” which means udlfb has loaded, set the monitor default mode, and drawn to the device successfully.
What udlfb does *not* do on its own is make configuring X easy – X sits on udlfb, but getting X running still involves editing configuration files like xorg.conf in different ways depending on what you’re trying to do, what distro version you have, and what primary GPU you have.
Features in Linux kernel 2.6.34 so far
- Improved performance (about 20% average improvement over a variety of “benchmarks” like x11perf, gtkperf, glxgears). Video playback, flash games, etc. are all improved
- Asynchronous urb dispatch and no mutexes held for extended periods during render
- Slightly lower CPU consumption, fewer non-localized memory accesses, slightly higher average compression
- Better handling of switching from X to fbcon VTs and back
- Driver unloads more cleanly from either bottom-up (USB removal) or top-down (shutdown). Problem reports welcome.
- More standard EDID parsing and mode handling, using fbdev’s built in mode libraries (but creates a new dependency on them being enabled in kernel)
- Performance metrics reported through sysfs (read /sys/class/graphics/fbX/*metrics* – but will move to debugfs)
- EDID reported through sysfs (read /sys/class/graphics/graphics/fbX/edid)
- Can compile as module with or without defio or sys_ dependencies
- Lots of cleanup – tested on 64-bit, closer to endian clean, checkpatch.pl clean
- Support for standard fbdev clients (via defio) – unfortunately buggy yet (see todos)
Depending on timing, some additional todo items might make it, we’ll see.
Versioning
At one point, we wanted udlfb to have its own project cycles and release versioning to help clarify things here while udlfb evolved in the early days. Got pushback on doing this from the kernel maintainers. So versioning of udlfb is purely by kernel version – “this bug is in udlfb in kernel 2.6.34rc1” or whatever. There’ll also be a newer changes that haven’t been submitted to the kernel (e.g. at http://git.plugable.com/). And it’s probably best to just version that per commit/checkin. “This bug shows up in commit 234b4e22…”. This is nice in a way – just check the stuff in somewhere let the world know what you’re doing – the only “release process” you need to coordinate with is getting the patches into the kernel itself. Or feel free also to send patches here, we’ll try to help them through the process.
Todo
- Add code to read pixel resolution limits from the device’s USB descriptors, and adhere to them. Avoids blank screen when monitors caps exceeds that of the adapter
- Add pseudo-32bpp mode to help the case of using Xinerama to extend desktop across PCIe/USB adapters. Not planning true 24/32-bpp mode yet because of complexity and perf tradeoffs
- Disable defio by default, unfortunately, because of several bugs without obvious solutions
- rendering problems with pages (lines) that no longer get updated – writes to those are no longer triggering deferred processing
- When running 2 USB adapters, dirty pages for one instance seem to affect the other, and vice versa (udlfb has no common state, so appears to be in defio or mmu handling)
- cases where we hit a kernel oops in the deferred io handler when it tries to touch the pages it’s been asked to handle.
- Move perf metrics from sysfs to debugfs and add ABI doc for what remains, per GregKH
- Fix two ignored return values that Greg’s build environment caught
- Get a safe method from DisplayLink to detect future chips that might not be compatible with current ones, and bail out of probe
There’s lot of other stuff that could be done – and any patches are very welcome. One big open question is whether fbdev has legs, or whether this effort should be converted over to KMS/DRM/DRI.