[SAC] [Hosting] Hypervisor reboots on production Ganeti cluster


We’ve been having issues with our Ganeti cluster hypervisors randomly rebooting. So far we’ve had this happen four times within the last month (including this morning) at the following times (in UTC) and duration:

Event Start Time Event End Time Event Duration

2018-07-27 10:25:13 2018-07-27 10:29:29 0d 0h 4m 16s

2018-07-28 17:03:19 2018-07-28 17:06:57 0d 0h 3m 38s

2018-08-23 08:14:07 2018-08-23 08:18:00 0d 0h 3m 53s

2018-08-27 11:46:06 2018-08-27 11:48:44 0d 0h 2m 38s

The kernel log points to this [1] upstream issue for RHEL and also reported on the CentOS forum [2] (We run CentOS 7 on our servers).

[1] https://access.redhat.com/solutions/3432391
[2] https://www.centos.org/forums/viewtopic.php?t=67170

It seems the solution is “In progress” so I expect it to be included in the next released kernel in CentOS. Unfortunately the workaround is to either use an older kernel or just wait. As of right now we are going to opt to wait until the next release happens however if we see increased instability we’ll look into using an older kernel.

So far this only affects our Ganeti cluster and hasn’t been triggered in either of our OpenStack clusters (x86 & ppc64le) as of yet.

For some of our older VMs which boot using an external kernel we build, we’ve been hitting a different problem with the dracut dropping to an emergency shell if there are any fsck warnings on boot. We’ve had to manually fsck those filesystems and reboot the VMs to bring them back online. We’re hoping to move away from these external kernels soon so let us know if you’d like to do that. In the meantime, we’re going to look and see what changes we can make with the initrd so that it doesn’t drop into the emergency shell and let’s the OS do the fsck instead. We recently updated those external kernels to use an initrd which is why this has started to happen.

The RH solution page has the following upstream commit message which explains why this bug is happening in the kernel:

Because we drop cpu_base->lock around calling hrtimer::function, it is
possible for hrtimer_start() to come in between and enqueue the timer.

If hrtimer::function then returns HRTIMER_RESTART we’ll hit the BUG_ON
because HRTIMER_STATE_ENQUEUED will be set.

Since the above is a perfectly valid scenario, remove the BUG_ON and
make the enqueue_hrtimer() call conditional on the timer not being
enqueued already.

NOTE: in that concurrent scenario its entirely common for both sites
to want to modify the hrtimer, since hrtimers don’t provide
serialization themselves be sure to provide some such that the
hrtimer::function and the hrtimer_start() caller don’t both try and
fudge the expiration state at the same time.

To that effect, add a WARN when someone tries to forward an already
enqueued timer, the most common way to change the expiry of self
restarting timers. Ideally we’d put the WARN in everything modifying
the expiry but most of that is inlines and we don’t need the bloat.

If you have any questions or concerns please let us know.



Lance Albertson

Oregon State University | Open Source Lab

Hi all,

I wanted to send you an update on this situation as it seems to keep happening. Unfortunately it doesn’t seem that RHEL has released a new kernel which resolves this issue. Over the past few days I’ve been rebooting nodes to apply the L1TF vulnerability fixes which of course means these nodes are now running the kernel which has this bug. This has made the cluster more unstable unfortunately causing multiple reboots a day.

I’m currently working on deploying a new mainline kernel to mitigate this issue and hope to have it deployed by the end of today.


Lance Albertson

Oregon State University | Open Source Lab

I’ve deployed a new kernel on all of our production ganeti nodes and the systems seem to be running better now. I’ll keep an eye on it for the next few days to be sure everything is still working properly.



Lance Albertson

Oregon State University | Open Source Lab