Kernel clock issue "Clocksource tsc unstable"

[Rashkae]

Hal Burgiss wrote:
> I have a wierd problem effecting two different 8.04LTS servers (fully
> updated). Their clocks stop updating. The one today said 18:05 with
> yesterdays date. And it stays that way forever. The only fix is a hard
> reset. This f*'s with anything that is time sensitive, so all logging
> stops, remote logins don't work, various networking stuff is broke,
> and so on. A PITA.
> 
> I have probably at least 6 systems on 8.04LTS, but only have this
> issue only on two of them, and these are radically different hardware.
> At first I suspected a hardware issue when I only had this issue with
> one machine. Now I have it with 2 so I can't imagine its hardware. To
> make it worse, its very sporadic. On the one system, it happens
> approximately once a month. The other system is new, has been running
> about 2 months, and this was the first occurence.
> 
> One message I did grab says:
> 
>  06:05:11 dio kernel: [1617517.563814] Clocksource tsc unstable (delta = 4686705598 ns)
>  06:05:11 dio kernel: [1617517.573798] Time: acpi_pm clocksource has been installed.
> 
> That was during a time server sync. In exactly 6 hours (at 12:05) that
> same day, the clock stopped updating, logging stopped, ssh was fubar,
> and I had to physically go to the datacenter and hit the reset button
> to fix it. 
> 
> There is no shortage to this problem on google, but most info sounds
> like workarounds or guesses. There is at least one open bug report on
> launchpad. Its labeled New from last September, which does not inspire
> confidence.
> 
> So I am curious if anyone has dealt with this, fixed it and/or
> understands the issue enough to offer a fix? These are both production
> systems, so I'd prefer to minimize guessing. One is local to where I
> am, and that is easier, but one is remote.
> 

Sorry I can't really give you a canonical answer.  tsc is an infamously
poor time keeper, but the kernel devs seem to prefer it over hpet
(Intel's High Precision Event Timer) because of performance.  (tsc is
right in the cpu, wheras hpet needs to access a chip on the mobo).

On my hard system where hpet is available, hardy defaults to that.
However, some systems do not have hpet, and kernel defaults to tsc.  In
your case, tsc is deemed unstable and switches to acpi.  Why this would
cause your clock to stop updating however, is a mystery, unless the acpi
hardware on your system is broken.

The first step to resoling this, however, will be to determine what
clocksources you have available, and then choose one on boot (as a
kernel boot parameter.)

sudo cat /sys/devices/system/clocksource/clocksource0/available_clocksource

and:

 sudo cat /sys/devices/system/clocksource/clocksource0/current_clocksource

You can specify one of the available clocksources by editing your grub
menu.lst file and appending a clocksource= option to the kernel you
boot.  For example, you can try clocksource=acpi_pm and test if acpi
works if you boot with it rather than switch to it.  If acpi_pm doesn't
work reliably on your systems, I would then try hpet or pit if those are
available.