For Monday, May 11th 2009, I'm Jon Masters with a summary of today's LKML traffic.
In today's issue: the x86 relocatable kernel, performance counters, devtmpfs, and NUMA memory affinity.
x86 relocatable kernel changes. Peter Anvin posted a followup RFC patch that aimed to address review comments posted by Eric Biederman last week. This "Protocol 2.10" allows for a more relaxed alignment requirement on the relocatable kernel image. It includes a fair amount of documentation. On the subject of x86, Yinghai Lu posted a number of patches cleaning up interrupt handling, mptable allocation, pci, acpi, and more.
Performance counters. Paul Mackerras posted a number of performance counter patches. Amongst them was a patch to fix handling of re-enabling a group leader where previously its siblings would not have been put on the PMU (Performance Monitoring Unit - a piece of hardware provided by the CPU). Paul also posted a patch that prevents the kernel from counting scheduler ticks as context switches. Finally, he posted a fix to the reset ioctl so that the full 64-bit performance count would be reset via atomic64_set, rather than just the low word of the full quantity, which had been the case due to the inadvertant use of a atomic_set call instead.
devtmpfs. A number of individuals have commented on the latest devtmpfs patches that were posted over the weekend. Listeners may recall that these patches extend tmpfs to provide a minimal in-kernel "device filesystem" in which certain devices found at boottime will have file entries made automatically in advance of udev starting. This, apparently, reduces startup time by up to several seconds, and has been shipped in at least one SuSE kernel already. The community, however, is quite concerned about the potential for this to creep into another devfs. Arjan van de Ven was amongst those offering alternatives - he repeated his suggestion for a simple text file created by the kernel providing sufficient information to udev in order to avoid it having to grovel in the bowls of the sysfs tree for several seconds, but without the kernel having to provide another filesystem. For his part, Kay provided a number of sensible justifications that will likely keep the conversation on devtmpfs going for a while to come.
NUMA memory affinity. Stefan Lankes posted a four part patch series implementing "affinity on next touch". The basic premise is that a user level process can activate this feature for a specific range of its virtual memory address space and can then expect the kernel to migrate that region of memory onto the next NUMA node that touches it. There have been suggestions of modifying the OpenMP standard to officially support this strategy.
In today's announcements: Stephen Rothwell released a linux-next tree for May 11th. Since the last tree was released on May 8th, the infiniband tree lost its build failure, the net tree lost a conflict, and the dwmw2-iommu tree gained a conflict against the x86 tree also. There remain a total of 136 source trees for the linux-next kernel tree.
The latest kernel release is still 2.6.30-rc5, which was released by Linus on Friday evening.
And finally. A big thanks to the over 6,000 people who downloaded this podcast last week. There is some kind of demand so it will continue - and in the spirit of all good Open Source projects, will evolve over time. It's a spare time effort and time consuming enough, so apologies to those who have made good suggestions if they take some time to come through. Today's addition is Transcripts. These are initially available without RSS feeds at KernelNewbies.org/KernelPodcast. Stay tuned for further updates.
That's a summary of today's LKML traffic. For further information visit kernel.org. I'm Jon Masters.