Re: [Qemu-devel] [PATCH 0/5] support guest agent general command

[Eric Blake]

On 08/21/2012 09:26 AM, Michal Privoznik wrote:

>> Question to the qemu folks - can we enhance 'guest-info' to tell us
>> commands required to give output on success, vs. commands that are
>> expected to never answer (except on possible error), so that libvirt can
>> then make smarter decisions about whether to wait for a response for an
>> arbitrary guest agent command?

> Even if I'd gladly eliminate this --timeout I am not so sure we can do
> that. Even if qemu-ga will report which commands does reply on success.
> I guess this is the image you had in mind:
> 1) user issues command X
> 2) libvirt asks qemu-ga if X returns an success reply

2b - if no answer to guest-info in a reasonable timeout, assume guest
agent is not running, and return from API immediately

> 3a) if it does, wait for it

although a timeout is still useful here, since the guest agent can go
away at any time.

> 3b) if it doesn't just write command into agent's socket. Asynchronously.
> 4) return from API
> 
> Well, current version of qemu-ga doesn't report this kind of info yet
> (patch against qemu-ga has been submitted). So in step #3 we can't
> decide whether to go with A or B path. And we can't wait for monitor
> event like we do now for those commands not replying on success - this
> command needs to be as general as possible.

Perhaps we can be sticklers, and refuse to run arbitrary qemu-ga
commands against a guest agent that doesn't report enough information.
After all, the libvirt addition of 'virsh qemu-guest-agent-command' is
intended as a debugging aid and future development testbed, rather than
a supported libvirt interface, and when you are debugging, it is okay to
require the latest and greatest.

> 
> On the other hand, what would happen if step #2 fails, so we go with 3B
>  then? For instance, if issued command was fsfreeze-freeze, which we've
> written asynchronously, users can issue fsfreeze-status to query for status.
> 
> Have I got it right?

Yes - if the user is going to issue arbitrary guest agent commands, and
we know the command won't reply on success, we should still wait a
reasonable amount of time for an error, but we should also return
control to the caller to allow them to detect the command's success via
other means such as an event occurring.

-- 
Eric Blake   address@hidden    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature