See also: last week's Kernel page.

Kernel development

The current stable kernel release remains 2.2.9, as it has been for some time. There is still no official stable kernel out there with the ICMP denial of service attack fix, unhappily (though most distributions have long since put out updated packages). Alan Cox's patches are up to 2.2.9ac4 This patch is rather smaller than previous versions - much of the more advanced stuff, which is going into 2.3, has been removed from the 2.2 patch.

A new FireWire (IEEE 1394) development effort was announced by Andreas Bombe. He actually has some working code, though it is still pretty far from prime time. Details may be found in his announcement.

Should the PCMCIA drivers be part of the standard kernel? This is not a new question, but it came up with a new force recently, due to some rather strongly worded messages from Linus. He is, evidently, not very happy with the current state of PCMCIA support (he said it "sucks"), and made some threats to start a new PCMCIA development from the beginning, much like he did with USB. David Hinds, the person who made PCMCIA happen on Linux was, not surprisingly, a little taken aback by these remarks.

Some discussion followed on the advantages of integrating PCMCIA into the standard kernel (better tracking of kernel changes, no separate package to install) versus those of keeping it separate (ability to support new cards on all kernel versions, support for non-Linux operating systems). But the real problem came out after a bit: it seems that Linus had a tremendously difficult time installing Linux on his shiny new Sony Vaio laptop. Linus has this notion that if he has a hard time with a Linux install, at least a few other users might find it a bit daunting, and he didn't like that idea.

Once the real problem was unearthed, a more solid, technical discussion followed. David Hinds outlined his understanding of the problem in this posting. The core of the solution seems not to be a full inclusion of PCMCIA into the kernel. Instead, if the kernel had enough of a minimal understanding of PCMCIA to be able to make use of devices which, at boot time, have already been initialized and set up by the system's BIOS, most of the problems would go away. A solution along these lines seems likely somewhere in the 2.3 series. (For a start, one might look at this note from Werner Almesberger, who has already implemented some of this functionality).

PCMCIA shares a problem that is coming up in a number of contexts: dealing properly with hot-pluggable devices. USB, FireWire, and the new hot-pluggable PCI bus also raise such issues. The Linux kernel still, for the most part, does not deal well with this sort of dynamic environment; among other things, see the discussions on naming issues covered in the last couple of issues of LWN.

Martin Mares has chimed in with a proposal of his own. He describes a device driver architecture that allows each driver to easily describe the devices it handles, thus making it easy for the kernel to choose the proper driver when a device is inserted. The ideas are currently in an early form; expect a more refined version later on.

Making Linux do better on web serving benchmarks. Mindcraft will be rerunning their "Linux vs NT" benchmark on June 14, see their 'open benchmark' pagefor details. There are said to be "Linux experts" involved, though it is not clear who they are at this point. Meanwhile, Juergen Schmidt published some benchmark results of his own. In his tests, Linux and Apache beat NT in many situations, but NT wins greatly in the multiple-interface case that looks most like the Mindcraft benchmark. Some problems remain in that area.

Some responses to Juergen's results include:

• The test should be run with a 2.3 kernel which has "wake one" semantics and some other tweaks. The point of the matter is, though, that these are changes in development, and still not ready for prime time.

• Many feel that Mindcraft-style benchmarks have little to do with the real world, and little to do with the reasons why people use Linux and Apache systems as web servers (and little to do with real-world web servers in general). Apache is used because it works and provides the needed functionality. These points are valid, but it is still desirable to make Apache perform better in this mode.

• How about serving static web pages directly out of the kernel? Arjan van de Ven posted an announcement of the kHTTPd kernel module. This module handles requests on the HTTP port, serves static pages (or images) directly, and passes everything else off to a user space daemon (such as Apache). The result is a striking performance increase over Apache when these sorts of requests are being serviced.

The end result seems to be that nobody expects much fun at the Mindcraft rematch, but that there is not a whole lot of concern about it either. In the longer term (six months, say), Linux and Apache will probably be a vastly faster web serving engine as several independent efforts to improve performance (including work by the Apache folks, of course) come together.

Interesting patches and updates this week included:

• Alex Buell has announced version 1.0 of his framebuffer HOWTO document.

• Hard-core performance tweakers may want to check out the performance monitoring counters patch by Mikael Pettersson, now at release 0.2.

• Version 0.1.0 ("the first really usable version") of the Linux Input Driver Suite was announced by Vojtech Pavlik. This patch reworks the keyboard and mouse drivers, simplifying things and adding a number of new features.

• Knfsd-1.3.3b by H.J. Lu. People using the 2.2 kernel NFS implementation should probably continue to track these patches as they come out.

• A new version of the power management patch by Stephen Rothwell.

• Devfs version 107 and version 99.2 (for 2.2 kernels) by Richard Gooch.

• Swsusp v6 by Gabor Kuti. This patch makes the "suspend to disk" capability often found on laptops available on all systems.

Section Editor: Jon Corbet

For other kernel news, see: Kernelnotes Kernel traffic

Next: Distributions