Kernel Log: Llano support, union filesystems
by Thorsten Leemhuis
Driver code for the graphics chip of AMD's Llano APUs has just made it into Linux 3.0, which could be given version number 3.0.0 after all. Meanwhile, there have been further discussions about the best approach for overlay, aka union, filesystems.
Late on Monday night, Linus Torvalds issued the third release candidate for Linux 3.0. With this RC, the kernel has been extended to support the graphics core of AMD's Llano processors which were introduced on Tuesday – it seems that Torvalds considered the changes small and harmless enough to integrate them into the main development branch, even though the merge window was closed more than a week ago.
After the version jump to 3.0, the developers have also made various changes to fix some problems with two-figure version numbers. In a subsequent discussion, Torvalds indicated that he might still call the next kernel "3.0.0" should the developers be unable to find some tricks to allow older programs to handle "3.0"; among the components that are struggling with such version numbers are older versions of the module-init tools (depmod and the like), mdadm, and the LVM2 and Device Mapper tools. In the long run, however, Torvalds' plan seems to be to switch to two-figure version numbers. In another discussion Torvalds has stated that developers should normally make no assumptions about the structure of the version number.
Just before the release of RC3, Rafael J. Wysocki released new regression reports. These reports state that last weekend the main development branch contained seven bugs that didn't exist in Linux 2.6.39; a further 18 bugs are known that didn't exist in 2.6.38.
Many Linux distributions use overlay, or union, filesystems for their live media so that a read-only filesystem can be overlaid with a writeable one – this makes it possible to install software when the distribution was booted from write-protected media like CD or DVD. Fedora uses the Device Mapper to make that possible; other distributions, such as Ubuntu, use overlay filesystems like aufs, which are not part of the mainline kernel. For years, various developers have been working on overlay or union functions at the filesystem level that could meet the kernel hackers' quality requirements for being integrated into the main development branch.
Developer Miklos Szeredi, who is known for his work on Fuse, has now suggested that his overlay solution be integrated into Linux 3.1. The resulting discussion gave the impression that some of the other kernel hackers would eventually also like to see an official kernel solution; however, filesystem specialist Valerie Aurora remarked that Szeredi's solution might have some problems. Aurora had been promoting an alternative union framework together with various other developers, but she largely withdrew from kernel development a few months ago.
The current discussions appear to be fizzling out – but the topic is likely to resurface again in the near future. Some background on Szeredi's approach is provided in an article on LWN.net published in September 2010. This article links to many older LWN.net articles which discuss various problems when overlaying filesystems that are maintained outside of the kernel; some of these filesystems ignore a number of known problems with overlaying.
- Sarah Sharp, who is known for her work on the USB 3.0 code, has announced a "USB mini-summit" which is to be held as part of LinuxCon in Vancouver. Her announcement has sparked a discussion revolving around some techniques for handing off USB devices to virtual machines. Among the contributors is Hans de Goede, who described a solution he has been promoting in a recent blog posting.
- The aforementioned discussion about union and overlay filesystems also included various contributions concerning userspace filesystems. In this context, Torvalds said that Fuse is suitable in certain situations, but that it is unsuitable for the root filesystem. One of the developers working on the Tuxera NTFS driver for the Linux kernel and Fuse said that the kernel driver is much faster than a userspace driver could ever be.
- Amir Goldstein has released patches that extend Ext4 to include a snapshot feature. In the resulting discussion, various developers noted that the Device Mapper's "multisnap" feature, which is still in development, offers similar functionality without requiring far-reaching changes to the Ext4 code (for example 1, 2).
- The problems in connection with the Linux (U)EFI support mentioned in the latest regular Kernel Log have triggered further discussions. In one of the postings, Torvalds is anything but positive about (U)EFI and uses strong language; however, he also describes the support of (U)EFI in Linux as fundamentally broken. Meanwhile, Matthew Garrett has discussed, on his blog, a practical effect of EFI: he said that it offers a place for storing data that can help developers find the causes for system crashes.
- Keith Packard is taking over the maintenance of the DRM/KMS driver for notebook and desktop graphics chips from Intel; in this context, he has provided and been given various tips on the use of Git in kernel development.
- In a discussion around an XFS problem that has so far not been solved in the .32 kernel, kernel veteran Theodore Tso pointed out that corrections don't appear in the stable and long-term kernels by magic, but that they require a developer to do the work; later, he added that this task is not part of the stable and long-term kernel maintainers' job.