Home > Articles > Operating Systems, Server > Linux/UNIX/Open Source

The 2.5 Linux Kernel

  • Print
  • + Share This
Matt Sherer takes a look at the development picture for the 2.5 Linux kernel in this article, including some of the problems with the current development cycle and how we might be able to get around them.
From the author of

The 2.4 Linux kernel, released on January 6, 2001, has been waiting for a successor for a significant period of time. Developers have been looking to work on the 2.5 tree ever since 2.4 was released—and even earlier. The problem is that the 2.4 kernel made some large capability jumps over the 2.2 series and has had some significant growing pains. As a result, it has taken a significant amount of time to get known issues in the new stable tree resolved. Everyone knew that there would be a period of time when work would be solely focused on the existing 2.4 tree.

But this has gone on for months, even though Linus has said that the 2.5 series will begin "soon." Some of the works in progress were needed in the field and simply could not wait for 2.6. This has resulted in some large revisions being added to the supposedly stable tree. Reiserfs was one of the first significant additions, and the memory-management system is now suffering the same fate. While it is true that some of these projects did need to be added immediately, some of the problems could have been avoided if there was were a shorter development cycle for the kernel. With a shorter cycle, forcing a project to wait until the next development cycle wouldn't mean putting it off for 10 months (and, by extension, making it wait for 2 years until the next stable release occurs). There would also likely be more control over what changed between cycles because it doesn't make sense to add large patches that aren't related to getting the next stable kernel into the field.

Before I get berated for covering only one side of the issue, let me acknowledge the fact that, with the large developer base scattered throughout the world, working on many different scales of effort, coordination for a single release point is very difficult, to say the least. Some groups might be ready for a small release cycle because most of their changes going forward are relatively easy to implement. Others might have significant problems with this because their project could take years to re-engineer between releases—and might take just as long to properly test between revisions.

Let me reiterate my point that if the release cycle were short, putting off the inclusion of your latest work wouldn't be devastating. There would be less of a need for everyone to be coordinated in releasing all code at the same time. If your code wasn't quite ready, you could skip this release and work toward the next. If the next release was expected to be in six months, this would be a much more acceptable time frame than being forced to wait two years for the next stable release. With a shorter list of active projects in a development cycle, the problem of coordinating all these disparate projects is greatly lessened.

This mentality, if used, would take a long time to sink in. Let's take a look at what is coming for the 2.5 development tree, when it does get here. The list, as you would guess, is fairly long, but I'll try to look at only the most important points because they are the few that should be focused on first. Keeping this list short is a big step in making sure that we see 2.6.0 before 2004.

About three months after 2.4.0 was released, the Linux 2.5 Kernel Summit was held in San Jose, California. Jonathan Corbet of Linux Weekly News covered the event, and many of the items in his report are still valid items waiting for the development tree:

  • Block layer redesign—The current design for the block layer has some significant performance penalties. Fixing these would ease the migration of upper-end database services to Linux. As it stands, the kernel uses some nonoptimal methods to break down requests. In addition, some subsystems are duplicated within most large database systems (such as an I/O scheduler). Furthermore, the raw I/O interface, which allows applications to bypass a filesystem and interact with a device directly, needs to be updated to handle large operations effectively.

  • Asynchronous I/O—True asynchronous I/O has been a dream within the kernel for a long time, and it appears that a solution is coming into place. The problem is that there are many opinions on how it should work, how the interface should look, and what types of I/O should be supported. While it sounds like a fairly small addition to the kernel, it actually does imply a great amount of forethought and testing. It appears that, in 2.5, we might actually see it completed because there has been a lot of work put into preparing the approach.

  • Network driver API—In very high load situations, the networking subsystem can lose data due to interrupt overhead. The problem is this: As packet loads skyrocket, the device issues interrupts for every packet that it sees. The kernel attempts to handle each one of these packets by scheduling kernel control paths for each interrupt seen. The overhead for this becomes too much to handle, and further packet interrupts are lost, even on network cards that aren't seeing much traffic. So, for a machine that bridges several networks, large amounts of traffic on one network can impact routing even between others.

  • Kernel build and configuration management—This one hits home for more people than just kernel developers: Anyone who is doing distribution development, updating their own systems, and so on will likely have to touch the kernel build system. And most of those people who have done so will agree that it's not the cleanest or most straightforward system.

Two separate problems in the kernel build are being worked on for 2.5: configuration management and the actual build system itself. For configuration management, the current system is called CML1 and doesn't have too many supporters. Eric Raymond has rewritten this system (in Python) in what is called CML2. Included in this new version is a theorem prover, a single system for dealing with config rules, and a simpler syntax. With the new system, if the theorem prover allows the ruleset, a valid kernel image should result from the build.

The build itself is the other portion of the problem going into 2.5 development. The new kbuild system is being developed to replace the currently used one. In the new system, some of the old complicated build files are gone, along with some stages of the build process itself. The old "make dep" stage is gone because the new kbuild will resolve dependencies during the normal kernel build. Likewise, "make modules" will go away because modules will be built along with the kernel image itself. Also, the process of installing the new kernel is automated. Among other things, the new kbuild will support building kernels outside the kernel tree itself, which can simplify operations for people testing various builds from the same kernel source.

Many other components of the 2.5 discussion were presented, but, in my opinion, these are the most pressing that would need to be solved first in the event of a short 2.5 development cycle. This is not an exclusive list because there are certainly other subsystems that could handle some updating without delaying release or impacting stability of other projects. A significant one that I have skipped is the memory-management arguments, mainly because that one is a bit too close to incendiary for my taste. Besides, it is one of those problems that can't wait for a two-year development cycle, so it is already being rewritten within the current stable 2.4 tree.

Keep in mind that this list was compiled in April 2001. While these items are still valid and waiting for 2.5, some others have come up. In general, most of them are smaller requests and are oriented more toward reorganization of driver code, which is just large enough of a change that it should not generally be attempted within a stable tree. Items that are likely to be updated in particular are the frame buffer drivers, support for newer protocols (such as SCTP, among others), more filesystems, POSIX-compliant event logging, LVM updates, and much more.

As we move forward, the list of items will only grow. Opening the 2.5 tree will help in getting some of these projects further underway and get more testing, but, at the same time, it will only open the door to a large development cycle if many of the longer-term ones are accepted during the first rush. It is my opinion that a specific set of items should be selected for updating in the 2.5 tree, and that list should be expanded only on a strict case-by-case basis. Of course, this list could be significantly long, provided that many of the projects have minimal overlap. For many of the driver modifications, this is likely the case. This would help to ensure that the 2.5 cycle is short, and, in any case, other projects still can develop outside the main tree. If the 2.6 kernel can be released relatively quickly, many of the side projects will be more solidified by that point and will be more likely to be included in the 2.7 tree as being nearly completed.

In all likelihood, the chances of this model being adopted are slim—after all, the kernel developers have been attempting to reduce the development cycle for years. In the end, it is a complex system that requires a lot of code and a lot of testing, neither of which is conducive to a quick turnaround. Some have called for a specific release date for 2.6.0, to make sure that everyone knows when code needs to be ready and so that no one will attempt to send large patches at the last minute. But, as we all know, time-based deadlines are likely to fail, and the code will be ready when it is ready. The development cycle of the Linux kernel is dependent on one thing: the code itself. If the magnitude of code (read: complexity) can be reduced during the development cycles, it will yield a stable kernel in a shorter amount of time, a more stable testing environment during development, and a happier group of end users in the process.

  • + Share This
  • 🔖 Save To Your Account