[libvirt] 答复: [libxl] shutdown a domain before it finishes starting

[Jim Fehlig]

On 04/01/2016 02:44 AM, Zhangbo (Oscar) wrote:
>>> Hi all:
>>>     Suppose we have a guest domain which is pvops, for example, rhel6.4.
>>>
>>>     Steps to produce the problem:
>>>     1 start the guest by virDomainCreate()
>>>     2 the API returns before the guest domain fully available, which means,
>> the disks, network interfaces and some import services are not available inside
>> the guest.
>>>     3 we call virDomainShutdown() to shutdown the guest.
>>>
>>>     Expected result:
>>>       The guest got shutdown.
>>>
>>>     The result in fact:
>>>       Because the guest is not available when we call virDomainShutdown(),
>> it couldn't  respond to our 'shutdown' xenstore request, the guest turns on
>> later, rather than shutting down.
>>
>> I don't think this is unique to a pvops guest kernel, or even a xen stack. I see
>> the same behavior with qemu. 'virsh create dom.xml && virsh shutdown dom'
>> results in the guest kernel missing the shutdown event and booting anyhow. I
>> guess SeaBIOS could still be loading when the shutdown event is issued :-). The
>> virDomainShutdownFlags documentation even states "that the guest OS may
>> ignore
>> the request". In my example, the guest OS isn't even alive yet.
>>
>>>     So , the question is:
>>>        In libxl_driver( xen-hypervisor environment), how can we tell that the
>> guest is available or not, and is it suitable to shutdown the guest at that
>> moment?
>>
>> libxl has no API to determine if a guest OS has booted. In a qemu/kvm stack, I
>> suppose qemu-ga is the preferred way to know when a guest OS has booted, or
>> is
>> far enough along to respond to shutdown events.
>>
>> One possible approach in xen, which is not supported by libvirt, would be to
>> monitor the state of a device frontend in xenstore. E.g. when
>> /local/domain/<domid>/device/vif/<vifid>/state reaches 4 (connected), you'll at
>> least know the driver in the guest is up and running.
> I've tried that way, but even the device state is not trustable, because inside the guest, it calls "add_disk" after the device state changes to 4, and  before it could respond the 'shutdown' xenstore request, which takes a while to complete.

Yeah, I thought that was a longshot. Synchronization of the front and backend
drivers doesn't necessarily mean the OS is in a position to respond to the
shutdown event. Lacking a guest agent, another option would be to wait for the
guests network stack to come alive, e.g. responds to pings or connection requests.

Regards,
Jim