Starting with version 1.3.51 the Linux kernel provides a watchdog driver. Not only does it take care of hardware watchdogs, but it also implements a software watchdog. It was created by Alan Cox, alan@lxorguk.ukuu.org.uk, and modularized by Angelo Haritsis, ah@doc.ic.ac.uk.

The Linux software watchdog is a timer that continuously counts down from a specified start value. Once it reaches zero a function named “watchdog_fire” is called that logs the event and initiates a hard reset. A watchdog daemon is therefore needed to refresh the timer. As long as the daemon is present the timer will never fire, but if the daemon stops working for some reason, the timer will fire and reboot the machine. However, with no additional hardware, the watchdog is not able to recover from every possible hang up. For instance, if the interrupt system itself hangs up, the software watchdog is no cure at all.

At start-up, the daemon opens the watchdog device (a character device with major number 10 and minor number 130) starting the watchdog process—an infinite loop in which the daemon alternately writes to the watchdog device to refresh the timer, then sleeps for 10 seconds. If the daemon is killed, the device file is closed, and the timer is disabled again.

Initially, I compiled the example given by Alan Cox in the source tree to be able to start, refresh and stop the software watchdog daemon. Since I always run the latest kernel and libc release and rely on the machine to be up and running for e-mail, even this simple version proved very useful for me.

As time went by, more features were added. First, WATCHDOG became a real daemon in that it automatically forks as a background process, tucking its process ID away. Then, since WATCHDOG is performing a critical job, logging to the standard syslog facility was implemented, as well as some user-friendly options:

This rudimentary version was then released as a Debian package to give more people the chance to use, and thereby, test it. The next step was to implement some enhancements, namely the ability to test the accessibility of a user-named file every time the process is awake. This ability allows WATCHDOG to catch problems with machines mounting most or all of their directory tree via NFS or similar means of remote file systems—this connection can hang despite the local machine functioning correctly. Since this hangup can cause huge problems, the WATCHDOG daemon tries to access one file on this mount every time it awakes. If the connection is down, it will not be able to refresh the timer, and the system will be rebooted. WATCHDOG can be instructed to initiate a soft-boot instead of the normal hard reset when any call returns an error message.

At this point in the development of WATCHDOG, I released the daemon as version 1.0 to all Debian archives and to tsx-11.mit.edu.

With more people using WATCHDOG some problems have been encountered and fixed, and I began working on the one problem that still worried me. From time to time the process table got so full that no more processes could be started, so I changed WATCHDOG to check the process table every time it is awake. If the table is full, the system is rebooted. Doing a hard reset in this case seemed a bad idea, so instead, WATCHDOG initiates a complete shutdown. This action also posed a problem, since the process table might stay full despite WATCHDOG's attempt to kill all processes. For example, it is not possible to start the shutdown binary when the process table is not empty. To take care of this situation WATCHDOG contains code for the complete shutdown process, including the following actions:

With this feature working correctly, version 2.0 was released on July 30, 1996 to tsx-11.mit.edu and every Debian mirror.

Of course, there are more ideas that need implementation. If anyone is in need of a check different from those described in this article, please write me at meskes@debian.org.