[Bug 39441] New: The reverts keep getting slower

https://bugs.winehq.org/show_bug.cgi?id=39441

Bug ID: 39441
           Summary: The reverts keep getting slower
           Product: Wine-Testbot
           Version: unspecified
          Hardware: x86
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: unknown
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

A revert that would take under 10 seconds right after creating the snapshot
would take over 7 minutes 6 months later. So every few months this would cause
the TestBot to become really sluggish and barely able to keep up with the patch
influx. Restarting libvirt, rebooting the host, restoring the VM from backup or
even transferring it to another host had no effect on the revert time.

While the revert is taking place the QEmu process fully occupies one core, no
disk I/O is performed and the VM is not running. The exact reason is not yet
known exactly but it seems to have to do with the VM's timer devices,
particularly the rtc one.

To confuse matters further not all VMs are affected: only Windows 2000, XP,
2003, 2008 and 10 suffer from this. The other post Windows Vista are immune.
Yet the guest is not active while the revert is taking place so it should not
have an impact on it.

Finally at WineConf 2015 it was discovered that the revert time of a live
snapshot is simply proportional to the snapshot's age.

This yielded a first workaround which is to refresh the live snapshots
regularly.

Further investigation showed that the common point between the impacted live
snapshots is that they all have the following clock settings
    <clock offset='localtime'>
      <timer name='rtc' tickpolicy='delay'/>
      <timer name='pit' tickpolicy='delay'/>
      <timer name='hpet' present='no'/>
    </clock>

while the unaffected VMs have
    <clock offset='localtime'/>

But switching from the former to the latter does not fix the affected VMs.

Still this lead to a better fix which has now been put in place: setting
track='guest' on the rtc timer.

Regardless, something is wrong with the way QEmu handles timers and live
snapshots so a bug was reported:
https://bugs.launchpad.net/qemu/+bug/1505041

Maybe this will shed some light on what's really happening and what the correct
timer settings are.