Kernel 2.4.x Architectural Changes (Filesystems)
This article looks at a couple of the architectural changes in the new Linux 2.4.0 kernel. This article does not go in-depth; it is meant to introduce you to two of the many changes.
Whenever architectural changes are suggested (much less implemented) on the kernel mailing list, the ensuing debates are lively and often heated, and can appear downright unfriendly at time (depending on just how strongly the developers feel about a particular issue). This time around, one particular change that's not mandatory but is available for use, the devfs filesystem, sparked a lot of debate.
Points against devfs ranged from "Do we really need this?" (the general consensus was a very resounding "Probably") to "How should it be implemented?" A lot of suggestions and "devil's advocate" role-playing went on. And while a good number of people weren't convinced that the way it was being implemented was the way to do it, the devfs filesystem is here to stay. It could be later rewritten completely, but for now the particular implementation stands.
So why did we need this change? Well, if you look in /dev, you see lots of files. In fact, all shipping distros as of this writing include several hundred device nodes that many you may not even need. My stock system, for example, has a whole slew of SCSI device files and various proprietary CD device nodes, none of which I need. But they all take up an inode and, therefore, space on my hard disk.
With the increase in USB, IR, and other devices, the 64K maximum number of devices looks like it may soon be reached. So which devices do you ignore? I'm sure the first time you can't create a device node for a particular piece of hardware you want to use, you won't be too happy. And it's not like you could just invent a node. The nodes are currently tied via their major/minor numbers to the kernel. Your IDE hard disk isn't really known as HDA; that's just a name you use. It's known to the kernel as device major 3, minor 0. You could just as easily call HDA eggs-and-waffles, but as long as it has major 3 and minor 0, it is accessing your first IDE hard drive.
The devfs filesystem will do away with some of this and the limitations imposed by the current scheme. For example, the system I'm using with devfs has device nodes only for devices that I actually have. So, there are no SCSI nodes. Also, this filesystem exists only while the system is running, not on disk. This introduces another problem: how to create new device nodes, and how to have the system start up with other than default permissions on device nodes.
Well, the first problem is very easily solved. All devices built into the kernel or loaded as modules contain code that creates the appropriate device node. So, if you have the ide-cd driver built into the kernel, you will have a CD-ROM device node. If the ide-cd driver is built as a module, when you load that module, the device node will be created.
Does anyone see a problem here? Let's say I want to mount a CD-ROM. So, I tell the system to mount it. Currently, the mount command makes a system call to mount the CD-ROM via the device node, and if the particular driver isn't installed, the kernel module loader loads ide-cd. But with devfs, if ide-cd isn't loaded, no CD-ROM device node exists, so an error message is returned. We have a cat chasing its tail scenario. So, before we try to mount the CD-ROM, we have to know to load the ide-cd module. Okay, so the system isn't perfect yet. But it's getting there.
As for starting the system with the appropriate permissions (nondefault) on device nodes, currently that information must be saved before shutdown and then, during bootup, reapplied to the nodes. Scripts can help with this.
While devfs isn't perfect (yet), its author and maintainer, Richard Gooch, is working diligently on the problems. I've been using this new filesystem for several months (a backport also exists for 2.2.x kernels), and I've had no real problems.
But don't run out and try to change everything willy-nilly, or you'll have a problem. The reasons are that the naming scheme has changed. Also, most distributions haven't incorporated a check for devfs that will start the required user-space daemon devfsd at the appropriate time. So, you'll have to add this yourself. You'll probably also need to rebuild your kernel with support for devfs and tell it to mount devfs by default at bootup.
The first thing you'll need to do is get, compile, and install the devfsd package. This is a straightforward process. It requires no special knowledge on your part. Hopefully soon, devfs RPM and .deb packages will be available for those brave souls who are bound and determined to beat their distribution to the punch.
Changes you'll need to make include inserting code near the top of the very first boot script to run on your system so that devfsd is up and running. This must happen before any filesystems are to be mounted, or your system won't be able to find your filesystems. Error messages indicating that your system can't find any hard disk devices can cause heart palpitations, so be warned and be ready to boot back into an older kernel that doesn't include devfs support so that you can fix things.
On my particular "experimental" system, I'm running Caldera's startup scripts. The first script run out of /etc/inittab is called /etc/rc.d/rc.modules. Your distro may be very different, so look carefully at your /etc/inittab to determine which script runs first. Near the top of that script (right after all the environment and variable declarations), you'll need to call devfsd (and you'll probably need a full path to it) and tell it where to mount (normally /dev).
Read the instructions that came with devfsd to see if you need any special entries in /etc/devfsd.conf, but the defaults should work for now. Then check your fstab. You'll probably need to change it to reflect the new names, but if your devfsd.conf file has the first two MKOLDCOMPAT and RMOLDCOMPAT lines uncommented, you should be safe for your first boot. If these lines are commented out, you'll need to either uncomment them or be able to figure out in advance what the names of your devices will be under the new device-naming scheme, and use those in /etc/fstab. So, boot into the kernel with devfsd the first time with these two lines uncommented, note where the old device names are linked to the new device names, and put the new device names into fstab.
If you decide to try to go this route, good luck. You may need it. You may also be safer waiting for your distro of choice to implement this for you. Only you can be the judge of this. You'll also want to check out the documentation on devfs in the Linux kernel tree under Documentation/filesystems/devfs/*.