Guest with SMP >=2 with kvm_clock has jumping clock every few seconds

I ran into the problem below running 2.6.33.2:

Problem: KVM guest running Linux 2.6.33.2 with kvm_clock as the clock source and more than 1 virtual CPU has the clock jumping forward 1 hour, 13 minutes and 17 seconds (consistently)

Hardware: AMD Athlon Dual Core 4850e @ 2511.769 MHz

I can reproduce this consistently, and it does not happen on the same machine if I launch the same KVM guest with -smp 1.

With RTC compiled into the kernel you can see that its just clock, the localtime was set to UTC:

Running the following in a loop:

cat /sys/class/rtc/rtc0/time ; date

You see:

18:07:49
Mon Apr 12 18:07:49 UTC 2010

18:08:21
Mon Apr 12 19:21:49 UTC 2010

18:08:21
Mon Apr 12 18:08:22 UTC 2010

The load average on the system is 0, the guest is doing nothing but running init. It happens regardless of whether I run one guest or over 10 guests on the same system.

I can easily reproduce this issue, so if you need someone to test patches etc, I can easily do that.

Thanks

John Buswell
http://www.carbonmountain.com

Just a follow up on this. The hypervisor had ACPI / ACPI PM clocksource support disabled, as did the guest. This problem occurred with both tsc and jiffies clock sources on the hypervisor.

With ACPI enabled and ACPI PM enabled, the clocksources on the hypervisor became hpt and acpi_pm. The guests still using kvm-clock, the clocks come up wrong. But if you set them via rtc0 (hwclock -s), they seem to maintain time ok. Only the first VM that came up exhibited an incorrect uptime.

It looks like kvm-clock overrides the setting under RTC -> Set System time from RTC on startup and resume. It might be a good idea to have the guest kernel prefer the RTC setting since its likely to correct.

Thanks

John Buswell
CEO, Carbon Mountain LLC
http://www.carbonmountain.com