so - for the dedicated boxes
i am hitting an issue with clocks on openbsd
Clocks as in... CPU? Or RTC? qbit
(Not that I have answers, I just help get details/triage)
RTC
What's it doing? Drifting? Lagging? And to confirm, OpenBSD is the host OS on the dedi box?
basically this: https://marc.info/?l=openbsd-bugs&m=151430928212450&w=2
the preemption_timer
i am wondering if it would be possible to get kvm-intel.preemption_timer=0 on what ever host has my dedi stuffs
What exactly is that? Some OpenBSD config() thing? A Linux string?
If you've got a dedi, you've got baremetal access, you can do whatever...
kernel boot option
Or have you been talking about a Thunder instance?
thunder
i call them dedi because : https://arpnetworks.com/dedicated
"dedicated"
"dedicated resources" but yeah, it leads to confusion IMO
So to clarify, you're asking if you could arrange for the Thunder host machine could be rebooted with a kernel parameter?
guess i am asking about the posibility of doing that
i understand that likely there are more clients.. so if it isn't an option - i will have to figure something else out
Seems like a tall ask :p but that's my opinion. And like I said, I'm just trying to distill the problem/request to help everyone out. :)
does this effect normal vps instances too?
it seems like quite a major openbsd bug
*affect :D
yes :)
qbit: What was your `sysctl kern.timecounter` anyways?
seems to be anything in kvm
brycec: acpitimer0
i think the only major change to openbsd in the recent history has been the addition of virtio random device
So not exactly the same as the mailing list guy's
mercutio: negative - i am running -current
this is a grip of changes from 6.2
was it fine prior to -current
seems to have been fine prior to the meltdown stuff
sigh
yeah so it probably will start happening on all virtual machine instances unless they fix it
starting with the next release
"FWIW, there are reports that this bug is absent from qemu-2.11.0.
hmm
i thought it was a kvm bug rather than qemu bug
https://marc.info/?l=openbsd-bugs&m=151439799628499&w=2
oh, you are probably already past that :P
well it's mostly about kvm_intel.preemption_timer being disabled apparently helping
that made me think it was kvm issue
uh, i don't think preemption_timer exists in the linux version being used
heh
it's a boot option (i put mine in grub)
it's not in /sys/module
but it's in /sys/module on newer kernel
dang
it's linux 4.4
will /sys/module always reflect avaiable moduels? or maybe loaded?
ah
not always
Not Always - Linux
but if it's not supported in there you can definitely not change it without reboot
dang
oh you can't change it while booted anyway
double dang
That's what she said!!
if you say so BryceBot
https://www.spinics.net/lists/kvm/msg161757.html
this suggests one way that may help a specific instance
qbit: this is quite easy to reproduce right?
step 1) install openbsd
snapshot?
ya
https://ftp3.usa.openbsd.org/pub/OpenBSD/snapshots/amd64/
that's all - no more steps :D
clock will drift even with ntp going
hmm install62.fs is preinstalled? :)
i can't remember seeing that before
What do you mean "is preinstalled"? It's a dd'able version of the ISO, but still just the install media (bsd.rd)
it's bsd.rd
oh
damn
with sets
qbit: Have you tested to see whether it can be reproduced from bsd.rd alone?
Or does it take the full bsd kernel config?
(I know that bsd.rd omits a lot of ACPI stuff to keep its size down, among reasons)
i don't know if bsd.rd is enough to make it happen
yeah, i doubt it would work
Also, I would've pointed mercutio to http://mirrors.arpnetworks.com/OpenBSD/snapshots/ :P
s/work/be a good test case/
<brycec> Also, I would've pointed mercutio to http://mirrors.arpnetbe a good test cases.com/OpenBSD/snapshots/ :P
:D
BryceBot: that's where i'l pulling it from :)
oh oops :)
lol qbit
Feb 20 13:03:47 spark ntpd[5226]: adjusting local clock by 18.013269s
Feb 20 13:08:07 spark ntpd[5226]: adjusting local clock by 22.688790s
damn
I agree with mercutio , it seems like quite a major openbsd bug
up_the_irons: For reference, what version of Linux/kvm/qemu are Thunder machines running?
qbit: And just to be clear, *you* have only seen this behaviour with recent snapshots? Older snapshots were fine, as were older releases?
4.4.0-109-generic #132-Ubuntu SMP Tue Jan 9 19:52:39 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
ii  qemu-kvm                             1:2.5+dfsg-5ubuntu10.16                    amd64        QEMU Full virtualization
brycec: if it was in previous snapshots - it wasn't enough ofa delay for me to notice
ha
I'm going to fire up a -current machine on a Proxmox machine, 4.4.62-1-pve + qemu 2.9.1-6~pve4 and see what happens...
i am not very observant
# date
Tue Feb 20 13:21:04 MST 2018
#
mine just finished installing
that's 6.2
i like how openbsd installs quite quick
Feb 20 13:19:22 db ntpd[2811]: adjusting local clock by 207.116095s
qbit: So you're saying 6.2 is also affected, cool (I have 6.2 VMs already in Proxmox :p)
ya
hangon my vm is using acpihept0
will that change things?
For the most part, my VMs stay just fine, but sometimes there are wild swings in drift, I'm guessing caused by load spikes in other VMs on the host. http://ix.io/N8h
By comparison, a 6.0 VM running on the same host has almost no time swing
brycec: is it using acpihept0?
my time still seems right with -current heh
maybe i need to trigger it with load?
Yes, both my 6.0 and 6.2 VM on the same Proxmox host are using acpihpet0. Only the 6.2 VM sees random drift. (Seeing it on Proxmox means it's not just an ARP issue :) And it's what I have to play with readily available.)
(i'm testing on my desktop first)
my desktop has linux 4.15, and some recent qemu
yeah - this has been confirmed on vultr too
i'm kind of curious if they've fixed it in linux or osmething
qemu 2.11.1
6.0 http://ix.io/N8s   6.2 http://ix.io/N8t   Same exact kern.timecounter values.
ahh cool
that's curious,y ou got a lot more adjustments sicne the 13th
"Something happened" on the 13th it seems.
to both of them
And re-evaluating my output, I guess 6.0 also saw some lagging afterall.
this doesn't seem extreme though
what qbit was talking about seemed more extreme
I disagree, 10s+ adjustments are extreme :p
i mean it's not good
but wasn't there like minutes of drift
like a lot bigger difference
But these 2 particular VMs of mine are real small stuff, not doing any major load (they're effectively bastion hosts). I suspect with heavier load, it exacerbates the issue.
yeah it may
it's def worse on machines with high load
i need to compile something :)
So I guess I'm back to: a) Doesn't seem to matter which OpenBSD release, even 6.0 is affected to a degree, albeit perhaps lesser. b) Even newer kernel+kvm (than ARP, but not "newest") are affected.
You guys are crazy.
likely the meltdown stuff in current snaps is gonna make it worse too
i just realised something i'm not testing with smp
is this only affecting smp hosts, or in general?
s/hosts/guests/
<mercutio> is this only affecting smp guests, or in general?
mine are all single cpu
ah ok
Ditto for me (I'm pretty sure)
i thought was on thunder
it is - but mine is proxmox -> openbsd vms
ah ok
well i did a kernel compile, and it failed to compile...
maybe -current has broken compiling 6.2
time is still right
so yeh i think either newer kernel or newer kvm has fixed it
huh
well that's cool
it seems to sync back up once load has gone down
qbit: what kernel is promox running under?
i'm thinking that the underlying thunder thing may make less difference than the proxmox layer..
mercutio: 4.4.98-6-pve
ahh so similar to ubuntu kernel
oh maybe proxmox doesn't haev a recent kernel available to try
The current stable 4.x release uses latest Ubuntu based kernel, which will be regularly updated. The first stable 4.0 release is based on 4.2 Linux kernel.
it's actually the same kernel even :)
heh