5-Minute Brainstorming

The final session on Monday was a brainstorming session, where presenters had five minutes to raise an issue of interest to them. Here's what came up:

• James Bottomley talked about I/O operations where use of the I/O memory management unit is optional. How, he asked, should be block layer be informed about whether the IOMMU should be used or not?

   

• Matthew Wilcox discussed a rework of how network devices work, with all function pointers to be moved into the new netdev_ops structure. This effort started as a way of eliminating duplicate implementations of the "ethtool" command set, but then expanded into a wider effort from there.

   

• Greg Kroah-Hartman suggested that, in 2.7, static device numbers should be replaced, since there is no longer any need for them. Instead, the kernel should allocate truly random device numbers, just to be sure that nobody ever hard-codes device numbers again. Greg also suggested that, in 2.7, the kernel should perform device probing in parallel, speeding up the boot process.

   

• Adam Belay stated that, in 2.7, he plans to merge resource management (i.e. control over allocations of I/O ports, I/O memory, and interrupt numbers) into the device model.

   

• Sunil Saxena talked about making better use of CPU performance counters to optimize performance-critical parts of the kernel. He also mentioned that 64-unit hyperthreaded processors may be on the horizon.

   

• Jon 'maddog' Hall discussed the apparent disconnect between developers and users at times; users think they have reported a problem, but the developers have never heard about it. Jon thinks that some effort needs to go into improving the communications channels.

   

• William Irwin mentioned that local page replacement policies do not always produce the best result. In 2.7, he plans to do some experimenting in that area.

Linus then spoke up, complaining that there has not been anywhere near enough controversy in the day's discussions. Thus, he concludes, everybody is perfectly happy with the 2.6 kernel? He asked if there were any specific worries that anybody would like express regarding the upcoming 2.6.0 release.

It was noted that asynchronous I/O still does not perform well. Linus responded that AIO has not seen much use or development; it is, essentially, unused and unmaintained. Unless more people take an interest in AIO, it will not improve much for 2.6.0.

One developer noted that a great many drivers have not been updated for the 2.5 kernel. Linus acknowledged that drivers are his main worry with this release. What is really needed here is many, many more testers who will complain when their particular hardware fails to work. Then the developers will know which drivers are actually worth the trouble of fixing. But those testers tend not to come until after the "dot-zero" release. So 2.6.0 will go out with a lot of broken drivers. Eventually (after a year or so), the ones that never get fixed can be removed.

It was asked: when will the 2.7 development series begin? The soonest possible time for the 2.7 fork is in about six months; that allows three more months for the 2.6.0 release, and three more to truly stabilize 2.6. That is a best-case scenario, however.