"SlowServers Admin" <[email protected]> writes: > On Sat Aug 29, 2026 at 10:38 AM UTC, Dave Voutila wrote: >> Is it still running? Try vmctl pause/unpause cycle. >> >> If the pause/unpause works and this is reproduceable, see the diff >> below. Another developer reported similar issues. This diff has had some >> testing and haven't found regressions yet, but since I can't accurately >> reproduce this issue it's hard for me to say it's correct in fixing it. >> >> Having some confirmation this works (or at minimum doesn't break things) >> would give me confidence to push for this as a reliability errata. > > Hi Dave, > > vmctl pause and unpause worked! The guest is back to normal now. > > Thank you! > > I will also test Arnaud's /etc/ntpd.conf workaround: > sensor vmmci0 refid LOCL stratum 8 trusted > > Also, do you believe that this impacts both 7.8 and 7.9 hosts? And > possibly AMD only? It's interesting that I haven't seen it impact Debian > guests at all, from Intel or AMD. >
The issue is due to a flaw in the interrupt delivery to vcpus and is present in both 7.8 and 7.9. The probability of hitting that race is probably higher as vmd has changed over the past year plus. It's architecture independent and impacts both Intel and AMD. My guess is this is more likely on machines with fewer, slower cores or when heavily over-subscribed with guests. > I also have a suspicion (with way fewer datapoints) that it's only > impacted OpenBSD guests where I've been using my own ntp server, which > I found out recently is out of sync by a few minutes. My ntp configuration > has looked like this on those guests (and the hosts, themselves:) > I think that's a coincidence. So far it seems correlated to guests doing lots of disk io. > server 2602:f5ef::1:0 > sensor * > > (It's possible that isn't a great configuration to begin with.) > > Now I do have guests on Intel with the same NTP configuration as well. > >> diffstat refs/heads/master refs/heads/vmm-intr >> M sys/arch/amd64/amd64/vmm_machdep.c | 76+ 32- >> M sys/arch/amd64/include/vmmvar.h | 1+ 1- >> >> 2 files changed, 77 insertions(+), 33 deletions(-) > > Do you want me to test that patch? I might be able to put that on a 7.9 > host that doesn't have any guests yet. I can also wait for it to be > pushed out as a reliability fix and deploy it that way -- either is fine. > If you could test and report back it would be very helpful. > Thank you, > > Slow Servers > > PS: Doing the pause/unpause triggered a bug that I haven't seen before. > vmctl pause and unpause worked, but the command "hung" until I killed it > with ctrl+c. > Yes that's a separate issue I need to identify. > /var/log/daemon: > Aug 29 15:35:37 odin vmd[61999]: potato: paused vm 6 successfully > Aug 29 15:35:37 odin vmd[41111]: control_dispatch_vmd: lost control > connection: fd -1 > Aug 29 15:36:21 odin vmd[61999]: potato: unpaused vm 6 successfully. > Aug 29 15:36:21 odin vmd[41111]: control_dispatch_vmd: lost control > connection: fd -1
