[libvirt] qemu: Add timeout for monitor to avoid virsh getting stuck when monitor gets die.

[Osier Yang]

于 2011年04月07日 18:17, Jiri Denemark 写道:
> On Thu, Apr 07, 2011 at 18:04:19 +0800, Osier Yang wrote:
>> 于 2011年04月06日 16:51, Jiri Denemark 写道:
>>> This is not the right approach. Introducing a timeout into all monitor command
>>> send to qemu is a bad thing. I think the right approach is to have a simple
>>> API which would just return domain's state without talking to its monitor or
>>> doing other complicated stuff. The new API could then be used by virsh list to
>>> list all domains even though they are blocked on qemu monitor. After all,
>>> virsh is not really interested in current memory consumption of domains when
>>> listing them.
>>>
>>
>> IMHO this won't work, if you restart libvirtd after the monitor gets
>> dead, you even can't get connection to libvirtd anymore, see:
>>
>> http://www.redhat.com/archives/libvir-list/2011-March/msg01422.html
>>
>> How to reproduce:
>> 1. setup a virt guest with qemu-kvm, and start it
>> 2. stop qemu process with following:
>>      # kill -STOP  `ps aux | grep qemu | grep -v grep | awk '{print $2}'`
>> 3. run the following command:
>> 4. # service libvirtd restart
>> 5. # virsh (hangs here)
>
> Yes, but that's a separate issue, IMHO. virsh hangs at this point since
> libvirtd doesn't even get to the point when it is able to accept connections
> from client. So introducing a timeout to some specific operations to help
> solve this issue might be the right approach but doing all commands with
> timeout is certainly not what we should do. Daniel described all the issues
> very nicely in his detailed email.

Oh, I didn't mean your proposal won't work for "virsh list", that will
work well I guess. :)

I meant if we consider "virsh list can't work if the monitor dies" as
a problem, then we need to solve the even more bad problem ---
libvirtd can't accept connection anymore if it's restarted.

Regards
Osier