VM clock goes backwards

[square]

We have been successfully running our Erlang/Elixir apps on different virtualization platforms (VMware Workstation 15 Pro, Proxmox) and we are now evaluating VirtualBox.

We are facing repetitive crashes of our apps with error OS monotonic time stepped backwards!:

[root@myserver:~]# journalctl -b -S 10:14:40 -U 10:14:50 | cat
Feb 24 10:14:44 myserver myapp[2680]: OS monotonic time stepped backwards!
Feb 24 10:14:44 myserver myapp[2680]: Previous time: 5730208658372
Feb 24 10:14:44 myserver myapp[2680]: Current time:  5730208647858
Feb 24 10:14:44 myserver systemd-coredump[4075]: Process 2680 (beam.smp) of user 399 terminated abnormally with signal 6/ABRT, processing...
Feb 24 10:14:44 myserver systemd[1]: Started Process Core Dump (PID 4075/UID 0).
Feb 24 10:14:44 myserver systemd-coredump[4076]: Removed old coredump core.beam\x2esmp.395.f5481f856fd04a718b5ca70219bc6f9d.2255.1740388064000000.zst.
Feb 24 10:14:49 myserver systemd-coredump[4076]: Process 2680 (beam.smp) of user 399 dumped core.

It seems that somehow the VirtualBox VM clock is going backwards, even with all of the following conditions combined:

• Disabled the Guest Additions Time Synchronization
• Disabled Guest Additions, for good measure
• Disabled NTP synchronization

For reference, this issue is observed with:

Host

• Windows 11 Pro 24H2
• VirtualBox 7.1.4

Guest

• NixOS 24.11.20250116.3355599 (Vicuna)
• Linux version 6.6.71

What causes the VM clock to go backwards? Is there a way to fix it?

First reported at:

• ​https://forums.virtualbox.org/viewtopic.php?t=113272
• ​https://github.com/erlang/otp/issues/9479
• ​https://elixirforum.com/t/elixir-crashes-in-virtualbox-with-monotonic-time-message/51004

[bird]

Thanks for the report. Since it's an SMP guest, I suspect it's a bad case of TSC synchronization by the guest. We'll look into ways to assist the guest in getting a zero delta between the TSC of the CPUs.

[square]

For information, I tried setting a single CPU on the VM and the issue disappeared.

Of course, proof by example is a fallacy but it went from more than 50 errors per day with 2 CPU to exactly 0 per day with a single CPU.

[bird]

I've implemented a fix that should work on more recent windows versions (19H1 and later). Test builds with revision 167943 or later should include it (not sure when I get a chance to upload).

Another way to workaround the issue is to disable Hyper-V on the host (a little complicated). That worked here.

[Klaus Espenlaub]

Fix is available on the test builds page.