On Fri, May 27, 2022 at 12:20:39PM +0200, Peter Krempa wrote:
> On Fri, May 27, 2022 at 10:47:58 +0100, Daniel P. Berrangé wrote:
> > Libvirt provides QMP passthrough APIs for the QEMU driver and these are
> > exposed in virsh. It is not especially pleasant, however, using the raw
> > QMP JSON syntax. QEMU has a tool 'qmp-shell' which can speak QMP and
> > exposes a human friendly interactive shell. It is not possible to use
> > this with libvirt managed guest, however, since only one client can
> > attach to he QMP socket at any point in time.
> >
> > The virt-qmp-proxy tool aims to solve this problem. It opens a UNIX
> > socket and listens for incoming client connections, speaking QMP on
> > the connected socket. It will forward any QMP commands received onto
> > the running libvirt QEMU guest, and forward any replies back to the
> > QMP client.
> >
> > $ virsh start demo
> > $ virt-qmp-proxy demo demo.qmp &
> > $ qmp-shell demo.qmp
> > Welcome to the QMP low-level shell!
> > Connected to QEMU 6.2.0
> >
> > (QEMU) query-kvm
> > {
> > "return": {
> > "enabled": true,
> > "present": true
> > }
> > }
> >
> > Note this tool of course has the same risks as the raw libvirt
> > QMP passthrough. It is safe to run query commands to fetch information
> > but commands which change the QEMU state risk disrupting libvirt's
> > management of QEMU, potentially resulting in data loss/corruption in
> > the worst case.
> >
> > Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
> > ---
> >
> > CC'ing QEMU since this is likely of interest to maintainers and users
> > who work with QEMU and libvirt
> >
> > Note this impl is fairly crude in that it assumes it is receiving
> > the QMP commands linewise one at a time. None the less it is good
> > enough to work with qmp-shell already, so I figured it was worth
> > exposing to the world. It also lacks support for forwarding events
> > back to the QMP client.
>
> I originally wanted to teach the qemu tools to work with libvirt
> directly similarly how 'scripts/render_block_graph.py' from the qemu
> tree already does but I guess this is also an option.
Yes, I do wonder about whether with John's new QMP python APIs,
it would be possible to plug in a livirt transport instead of
the socket transport. I've not spent enough time looking at the
Python QMP code to know if that's viable or not though.
I can look into it. It looks like render_block_graph works by actually executing a subprocess. Is there a chance of getting anything socket-like or stream-like out of libvirt to work with instead?
As long as I can get some kind of stream going it should be easily possible to just replace the fd(s) the qmp lib uses and talk to libvirt instead.
(possibly with changing some details about the handshake, yadda yadda.)
> This is an option too albeit a bit more complex to set up, but on the
> other hand a bit more universal.
The two approaches aren't mutually exclusive either. There's no
reason we can't have both options.
I've in the past thought about implementing MITM directly in the qmp library. I wrote a QMPServer() class that I did not check in upstream because it is only prototype quality, but I used it for testing. I still have plans to upstream it because it is useful for the test suite.
It has the same limitation as a lot of our Python tools have, though; it requires newlines because it doesn't have a streaming parser ...
It's been something that I want to improve upon, but since libvirt is well-behaved and sends newlines, I haven't been properly motivated to.