July 11, 2013
Participants: John W. Linville, Tony Luck, Steven Rostedt, H. Peter Anvin, Olof Johansson.
People tagged: Jiri Kosina, Linus Torvalds, Dave Miller, Ingo Molnar, Thomas Gleixner.
John W. Linville points out the dangers of a too-tall maintainership tree, citing several examples where patches went up through several levels of maintainership, but were rejected at a higher level, requiring coordinated rework at the lower levels. John gives a couple of examples which he said were resolved quickly, but did result in pain for lower-level maintainers (he also notes that the rejections were for good reasons). John suggests that if the maintainership hierarchy grows too tall, compensating changes in the release process might be needed (for example, a multi-stage merge window), which might in turn result in higher patch latency.
Tony Luck suggests that if the examples that John cites are not isolated incidents, then top-of-tree maintainers might do well to adopt a workflow more like that of Ingo Molnar's, Thomas Gleixner's, and H. Peter Anvin's -tip tree. Tony believes that -tip tree's liberal use of topic branches reduces the pain of upstream rejections due to the fact that only the specific topic branch containing the problem need be rejected and reworked. He also speculates that heavier use of topic branches would make it easier for upstream maintainers to say “no” because there would be less pain inflicted on innocent bystanders who had the misfortune of riding the same train as the problematic patch. That said, Tony has received some negative feedback on some octopus merges, and would like some guidelines from Linus during the discussion.
H. Peter Anvin was not comfortable with the -tip workflow being called out as the Universal Grand solution, but said that he would be happy to talk at LKS about the -tip workflow and its advantages and disadvantages. Steven Rostedt said that this would be a great discussion for LKS. For his part, Olof Johansson said that he would be happy to discuss how he and Arnd Bergmann run the arm-soc tree two years on, noting that only Greg KH's driver tree and David Miller's networking tree consistently feed more patches to Linus than does the arm-soc tree. Dueling maintainers, anyone? ;-) There was a discussion on whether different subsystems needed to interact differently with Linus. H. Peter Anvin closed out the discussion by suggesting that top-level maintainers who were unsure should email Linus asking him how he would like them to operate. Peter said that lack of a reply should be interpreted as “keep doing what you have been doing”.