In Chapter 6, we installed the Udev package. Before we go into the details regarding how this works, a brief history of previous methods of handling devices is in order.
Linux systems in general traditionally use a static device creation method, whereby a great many device nodes are created under /dev (sometimes literally thousands of nodes), regardless of whether the corresponding hardware devices actually exist. This is typically done via a MAKEDEV script, which contains a number of calls to the mknod program with the relevant major and minor device numbers for every possible device that might exist in the world. Using the udev method, only those devices which are detected by the kernel get device nodes created for them. Because these device nodes will be created each time the system boots, they will be stored on a ramfs (a file system that resides entirely in memory and does not take up any disk space). Device nodes do not require much disk space, so the memory that is used in negligable.
In February 2000, a new filesystem called devfs was merged into the 2.3.46 kernel and was made available during the 2.4 series of stable kernels. Although it was present in the kernel source itself, this method of creating devices dynamically never received overwhelming support from the core kernel developers.
The main problem with the approach adopted by devfs was the way it handled device detection, creation, and naming. The latter issue, that of device node naming, was perhaps the most critical. It is generally accepted that if device names are allowed to be configurable, then the device naming policy should be up to a system administrator, not imposed on them by any particular developer(s). The devfs file system also suffers from race conditions that are inherent in its design and cannot be fixed without a substantial revision to the kernel. It has also been marked as deprecated due to a lack of recent maintenance.
With the development of the unstable 2.5 kernel tree, later released as the 2.6 series of stable kernels, a new virtual filesystem called sysfs came to be. The job of sysfs is to export a view of the system's structure to userspace processes. With this userspace visible representation, the possibility of seeing a userspace replacement for devfs became much more realistic.
The sysfs filesystem was mentioned briefly above. One may wonder how sysfs knows about the devices present on a system and what device numbers should be used. Drivers that have been compiled into the kernel directly register their objects with sysfs as they are detected by the kernel. For drivers compiled as modules, this will happen when the module is loaded. Once the sysfs filesystem is mounted (on /sys), the data which the built-in drivers registered with sysfs are available to userspace processes and to udev for device node creation.
The S10udev initscript takes care of creating these device nodes when Linux is booted. This script starts with registering /sbin/udev as a hotplug event handler. Hotplug events (discussed below) should not be generated during this stage, but udev is registered just in case they do occur. The udevstart program then walks through the /sys filesystem and creates devices under /dev that match the descriptions. For example, /sys/class/tty/vcs/dev contains the string “7:0” This string is used by udevstart to create /dev/vcs with major number 7 and minor 0. The permissions of each and every device that udevstart creates are set using files from the /etc/udev.d/permissions.d/ directory. These are numbered in a similar fashion to the LFS bootscripts. If udev cannot find a permissions file for the device it is creating, it will default permissions to 600 and ownership to root:root. The names of the nodes created under the /dev directory are configured according to the rules specified in the files within the /etc/udev/rules.d/ directory.
Once the above stage is complete, all devices that were already present and have compiled-in drivers will be available for use. What about those devices that have modular drivers?
Earlier, we mentioned the concept of a “hotplug event handler.” When a new device connection is detected by the kernel, the kernel will generate a hotplug event and look at the file /proc/sys/kernel/hotplug to find out the userspace program that handles the device's connection. The udev initscript registered udev as this handler. When these hotplug events are generated, the kernel will tell udev to check the /sys filesystem for the information pertaining to this new device and create the /dev entry for it.
This brings us to one problem that exists with udev, and likewise with devfs before it. It is commonly referred to as the “chicken and egg” problem. Most Linux distrubtions handle loading modules via entries in /etc/modules.conf. Access to a device node causes the appropriate kernel module to load. With udev, this method will not work because the device node does not exist until the module is loaded. To solve this, the S05modules bootscript was added to the lfs-bootscripts package, along with the /etc/sysconfig/modules file. By adding module names to the modules file, these modules will be loaded when the computer is starting up. This allows udev to detect the devices and create the appropriate device nodes.
Note that on slower machines or for drivers that create a lot of device nodes, the process of creating devices may take a few seconds to complete. This means that some device nodes may not be immediately accessible.
When you plug in a device, such a Universal Serial Bus (USB) MP3 player, the kernel recognizes that the device is now connected and generates a hotplug event. If the driver is already loaded (either because it was compiled into the kernel or because it was loaded via the S05modules bootscript), udev will be called upon to create the relevant device node(s) according to the sysfs data available in /sys. If the driver for the just plugged in device is available as a module but currently unloaded, then attaching the device to the system will only cause the kernel's bus driver to generate a hotplug event that notifies userspace of the new device connection and it not being attached to a driver. In effect, nothing happens and the device itself is not usable yet.
If building a system that has a lot of drivers compiled as modules rather than directly built into the kernel, using the S05modules may not be practical. The Hotplug package (see http://linux-hotplug.sourceforge.net/) can be beneficial in these cases. When the Hotplug package is installed, it will respond to the aforementioned kernel's bus driver hotplug events. The Hotplug package will load the appropriate module and make this device available by creating the device node(s) for it.
There are a few known problems when it comes to automatically creating devices nodes:
1) A kernel driver may not export its data to sysfs.
This is most common with third party drivers from outside the kernel tree. These drivers will not end up having their device nodes created. Use the /etc/sysconfig/createfiles configuration file to manually create the devices. Consult the devices.txt file inside the kernel documentation or the documentation for that driver to find the proper major/minor numbers.
2) A non-hardware device is required. This is most common with the Advanced Linux Sound Architecture (ALSA) project's Open Sound System (OSS) compatibility module. These types of devices can be handled in one of two ways:
Adding the module names to /etc/sysconfig/modules
Using an “install” line in /etc/modprobe.conf. This tells the modprobe command “when loading this module, also load this other module, at the same time.” For example:
install snd-pcm modprobe -i snd-pcm ; modprobe \ snd-pcm-oss ; true
This will cause the system to load both the snd-pcm and snd-pcm-oss modules when any request is made to load the driver snd-pcm.
Additional helpful documentation is available at the following sites:
A Userspace Implementation of devfs http://www.kroah.com/linux/talks/ols_2003_udev_paper/Reprint-Kroah-Hartman-OLS2003.pdf
The Linux Kernel Driver Model http://public.planetmirror.com/pub/lca/2003/proceedings/papers/Patrick_Mochel/Patrick_Mochel.pdf