On 05/16/2017 12:20 PM, Daniel P. Berrange wrote:
>> @@ -678,6 +699,7 @@ qemuProcessHandleShutdown(qemuMonitorPtr
mon ATTRIBUTE_UNUSED,
>>
>> unlock:
>> virObjectUnlock(vm);
>> + qemuDomainEventQueue(driver, pre_event);
>> qemuDomainEventQueue(driver, event);
>> virObjectUnref(cfg);
>
> Nice - you send the same event as always so old clients don't break, but
> new clients can now look for the new cause.
I don't think that's right actually. IMHO, we should onyl be sending the
new event, not the original event. These are intended to indicate state
changes, and having two VIR_DOMAIN_EVENT_SHUTDOWN events sent with
different details is not really representing a state change.
WRT to compatibility, clients should always expect that 'detail' may
change or new 'detail' enum values may be added - indeed we've done
that many many times int he past. So I don't think we need to duplicate
the existing event
That may be my fault for causing a mis-understanding of
back-compatibility handling on my review of v1. In the past, when we've
had an event that returns a too-small struct, the only way to return
more information is to create a second event with the additional info,
then fire off both events at the same time (for the clients expecting
the old event semantics, and for new clients) - which really means two
separate RPC events. But here, we already have sufficient lifecycle
event parameters to return details without having to add a new RPC event
(proven by the fact that you didn't have to touch src/remote at all).
So now I'm agreeing with Daniel - the fact that we have new information
means we don't need to be back-compat to older clients: they will see
the same lifecycle event they have always seen, and not care that the
cause has changed, while new clients will see a plain cause where no
information was available from older qemu, or one of the two new causes
where qemu gave it to us.
--
Eric Blake, Principal Software Engineer
Red Hat, Inc. +1-919-301-3266
Virtualization:
qemu.org |
libvirt.org