Kernel Summit + Linux Plumbers 2016

| categories: fedora

Last week was the annual kernel summit and Linux Plumbers Conference in Santa Fe, NM. Like other conferences, this involved a bunch of scheduled talks and lots of hallway track (and plenty of Mexican food).

  • Monday was the invite only kernel summit. The morning was mostly process talks. I brought up some discussion about the bugzilla, mainly that it is inconsistently used. The kernel is heavily biased towards e-mail and anything that isn't e-mail tends to get ignored/used very begrudingly and bugzilla is no exception. The power management maintainers really like the bugzilla and actually respond, most other maintainers do not. This is the problem: users who file a bug don't know whether to expect a reponse. I generally discourage the use of the bugzilla. The bugzilla will be getting an update soon so we'll see what happens after that. LWN has better summaries of all the kernel summit stuff than I can give.

  • Paul McKenney gave a talk about a formal model for memory ordering. The kernel has a document which explains how memory barriers and the like should be used in the kernel. Despite the fact that this document says it is not a specification, it gets treated like one. Several researchers have been working on actually formalizing this document to be able to give an answer whether certain sequences of code will work correctly. This doesn't have many day-to-day benefits for most kernel developers but the work is very important from a research stand point.

  • A bunch of people sat down for a meeting about what to do about the secure boot patches. All the major distros are now carrying some form of the patches so it makes sense to converge on one solution. This was a very productive discussion and there is a much better path forward for those patches.

  • There were several different MM sessions. Johannes Weiner spent a lot of time talking about process thrashing. The current OOM killer does not do a great job of detecting when processes are thrashing so processes may continue to run instead of being OOM killed. He brain stormed some better ways to detect these conditions, such as checking how frequently pages are being re-faulted. Several people (including me) brought up some issues with the SLUB allocator. The SLUB slow path is pretty slow ('glacial path' was the phrase used). SUSE still uses the SLAB allocator but is looking at work to improve SLUB for their needs.

  • There was a mini conference about FPGA/Reprogrammable Logic Devices. There's a growing interest in supporting FPGAs in the kernel as regular devices. Vendors like Altera and Xilinx have been slowly adding support. There was a lot of talk about standards which lead to a discussion about ACPI vs Devicetree for the one billionth time. FPGAs are useful across the board from the embedded world to servers so getting everyone on the same page is important. Lots of new and interesting work.

  • As part of the testing/fuzzing microconference, some developers from Google talked about their tool Syzkaller. This is a more directed fuzzing tool which tries to be slighltly more clever than random for generating system call inputs and also generate share the program that caused the crash. Part of this discussion led to a session on formalizing system calls. The goal was to have specifications generated automatically for tools and possibly man pages. There were some tentative conclusions reached so we'll see what happens in the future.

  • I gave a talk/rant about Ion at the Android microconferece. Ion has gotten enough push back that I don't think it will ever be merged in its current form. I called for a moratorium on new features to the existing architecture and to work towards a new architecture that might be more acceptable by the community. If we ignore the community aspects, Ion is an acceptable framework for Android. Keeping major frameworks out of tree is not great for the Android ecosystem so it really really really needs to be in tree and out of staging. I had some good discussions and the DRM Graphics people seem to be very willing to help shepard things along. This just leaves me doing actual work and not just complaining. The hardest task.

  • Lennart Poettering gave a talk about 'portable system services' which is basically systemd services in something like a container. The concept is needed but I remain skeptical about the idea of doing systemctl start http://my-completely-trusted-server/service.icb to run services from anywhere. This will be something to watch as it comes into Fedora.

  • kdbus was rejected as an approach sometime last year and the developers decided to start from scratch with a proposal for Bus1. They gave a very nice talk discussion the existing IPC mechanisms and some of the trade offs with each and then described their new Bus1 approach. It seemed to be reasonably well received and it wasn't just a blind port of another possibly broken IPC mechanism.

Once again, a great but very busy conference with tasks to follow up on.