Re: [Qemu-devel] [PATCH 0/2]: qemu-ga: make shutdown & suspend synchronous

[Michael Roth <mdroth@linux.vnet.ibm.com>]

On Mon, May 14, 2012 at 11:17:18AM -0600, Eric Blake wrote:
> On 05/14/2012 11:06 AM, Michal Privoznik wrote:
> > On 11.05.2012 21:19, Luiz Capitulino wrote:
> >> The main motivation for this series is fixing two possible race conditions
> >> in the guest-suspend-* API due to the complexity that arose from the way
> >> we handle terminated children processes today. Full details in the first
> >> patch.
> >>
> >> This series applies on top of my two other qemu-ga series submitted
> >> previously:
> >>
> >>  http://lists.gnu.org/archive/html/qemu-devel/2012-05/msg00999.html
> >>
> >>  http://lists.gnu.org/archive/html/qemu-devel/2012-05/msg01507.html
> >>
> >>  qapi-schema-guest.json   |   56 ++++++++++------
> >>  qapi/qmp-core.h          |   10 ++-
> >>  qapi/qmp-dispatch.c      |    8 ++-
> >>  qapi/qmp-registry.c      |    4 +-
> >>  qemu-ga.c                |   40 ++++++------
> >>  qga/commands-posix.c     |  162 ++++++++++++++++++----------------------------
> >>  qga/guest-agent-core.h   |    4 ++
> >>  scripts/qapi-commands.py |   14 +++-
> >>  8 files changed, 154 insertions(+), 144 deletions(-)
> >>
> > 
> > Okay, this is definitely an enhancement and fix of bogus implementation.
> > One thing that I'd like to ask is - how can user distinguish between
> > these implementations. I am asking basically from libvirt POV.
> > Because if I assume I am dealing with the previous implementation and
> > thus waiting for the {'return':{}} before I can return form an API, but
> > the GA actually uses the new implementation I will block endlessly.
> 
> I think the point was that you would block endlessly waiting for the
> {'return':{}} anyways, even with the old implementation, because there
> was no guarantee that the guest agent could issue the reply in a timely
> manner.  Therefore, the only sane implementation in libvirt is to assume
> that success will not be reported via the guest agent, and that libvirt
> must _always_ probe for the listed side effects (a change in guest
> status, qemu exiting, an event, or so forth), and optionally have a
> timeout if the associated timeout does not occur in a reasonable time.

Additionally, *all* commands carry the risk that qemu-ga can die
(killed, crashed, guest reboots, not being installed to begin with, etc) before
issuing a response, so a timeout mechanism *must* be in place if it blocks
libvirt.

Let me know if it's unclear how timeout+reset is to be implemented/used. I've
captured most of the ins/outs here but I realize it's still a bit of a
headache:

http://wiki.qemu.org/Features/QAPI/GuestAgent#QEMU_Guest_Agent_Protocol

> 
> -- 
> Eric Blake   eblake@redhat.com    +1-919-301-3266
> Libvirt virtualization library http://libvirt.org
>