Created attachment 770813 [details] Difference of normal time and guest time in seconds. Description of problem: Using virtio-net driver in windows server guests will sometimes (random delays) stop incrementing clock in guest. Stop is not incrementing long time and long time there is same time in this guest. Long time means multiple hours. Version-Release number of selected component (if applicable): virtio-win-0.1-52.iso virtio-win-0.1-59.iso qemu-kvm-1.2.2-13.fc18.x86_64 kernel-3.9.6-200.fc18.x86_64 How reproducible: daily Steps to Reproduce: 1. install windows server as guest 2. install virtio-net drivers and enable virtio-net card 3. watch time skew using some monitoring tools Actual results: see attached image from check_mk nagios plugin Expected results: running clock, at least only small time deltas Additional info: If requested, I can retest with Fedora 19.
Hi Jan, How does it work with E1000 and IDE? Thanks, Vadim.
Created attachment 770836 [details] Time skew without virtio-net network card. (In reply to Vadim Rozenfeld from comment #1) > How does it work with E1000 and IDE? Does not matter, if I use ide or virtio-disk drivers. After change of virtio-net card in libvirt to e1000 helped to workaround this problem for me. I have still time skew, but time is not stopped, just moving slowly. More info in this thread: https://lists.fedoraproject.org/pipermail/virt/2013-July/003714.html (btw, there are 2 problems in this thread. Time skew and time stopped)
Which Windows exactly? Also disable hpet for your guest.
Windows Server 2008 R2 Enterprise, SP1, 64bit All updates applied. With disabling hpet, did you mean this? <timer name='hpet' present='no'/> This is my current complete clock settings: <clock offset='variable' adjustment='2' basis='utc'> <timer name='rtc' tickpolicy='catchup'/> <timer name='pit' tickpolicy='delay'/> <timer name='hpet' present='no'/> </clock> I will tell later, how it works with disabled hpet.
HPET is disabled, clock offset is -3833.3 seconds. Looks like does not matter, if HPET is enabled or disabled.
Looks like e1000 and other fully virtualised netcards has other problem with 3.10 kernels. :-( https://lists.fedoraproject.org/pipermail/virt/2013-September/003787.html So there is no network card in current kernels for windows. :-(
Hi Jan, We believe that this is a known issue with broadcasting of clock interrupts to all Windows CPUs, where only CPU0 increments the counters. This is probably solved in kernel 3.11 (maybe 3.10), and in RHEL. Still, you have to disable HPET. Ronen.
Looks like it's fixed in current Fedora/EPEL. Tested several weeks without problems.
Great, thanks for following up Jan!