nluhux/vmm_clock

Experimental implementation of a kvmclock-derived clocksource for Linux guests under OpenBSD's hypervisor

vmm_clock

This is a Linux clocksource implementation based on the existing Linux kvmclock. It's specifically designed for using when running a Linux kernel as a guest under OpenBSD's vmm(4)/vmd(8) hypervisor framework.

Primary goals:

• provide a clocksource that doesn't suffer from unmanageable clock-drift as seen when using refined-jiffies with a more recent Linux LTS kernel (e.g. 5.4)
• provide a clocksource loadable as a module that does not require users build a complete kernel from source
• be platform independent, i.e. work the same on Intel and AMD hosts.

Secondary goals:

• make the code as short and tight as possible
• identify a means of testing clock drift

Known issues:

• if you try to unload the module, you will probably panic the kernel
• since kvmclock wasn't designed to be a module, it may never be possible to support module removal

Prerequisites

You'll need OpenBSD 6.8 or newer as it contains fixes I provided for race conditions in vmd(8) that cause stability problems with Linux guests. It is 100% required for this kernel module to work. If you try to use this with an older OpenBSD version, I guarantee you will have runaway vmd(8) processes or crashes.

Building & Installing

For now, follow some of the pre-reqs for virtio_vmmci and you should just be able to:

# make && make install && modprobe vmm_clock

You can load it at boot by creating /etc/modules-load.d/vmm_clock.conf with just the contents:

vmm_clock

Tested Platforms and Configs

I'm testing with the following:

• Alpine 3.12.2 and their stock v5.4 -virt kernel
• Debian Buster 10.4 and its v4.19 kernel
• Ubuntu 20.04 and its v5.4 kernel

As usual, no warranty...use at your own risk, etc.!

License

GPLv2...see LICENSE file.