On Tue, Jun 23, 2026 at 12:52:53 +0100, Daniel P. Berrangé wrote:
> On Tue, Jun 23, 2026 at 12:38:10PM +0100, Daniel P. Berrangé wrote:
> > On Tue, Jun 23, 2026 at 01:03:46PM +0200, Peter Krempa wrote:
> > > On Tue, Jun 23, 2026 at 10:55:16 +0100, Daniel P. Berrangé via Devel 
> > > wrote:
> > > > Conceptually -object and object_add/object_del should be sufficient
> > > > for essentially all QEMU configuration....if only we ported all our
> > > > internal custom backends/devices/etc to QOM. That is of course a big
> > > > job which is why it hasn't happened.
> > > > 
> > > > This series started with the premise that the monitor is one of the
> > > > easier areas to convert since we have no more than three classes,
> > > > a common base, and QMP and HMP subclasses[1]. So why not give it a
> > > > go and thus unlock the ability to dynamically create/delete monitors
> > > > in QMP/HMP.
> > > > 
> > > > This series does the conversion in a great many small steps to better
> > > > understand the implications at each stage.
> > > > 
> > > > The high level outcome of this series is
> > > > 
> > > >  * HMP and QMP monitors are QOM objects, 'monitor-hmp' and
> > > >    'monitor-qmp' respectively
> > > > 
> > > >  * Both can be cold plugged and hot plugged. QMP only, can
> > > >    also be hot unplugged.
> > > > 
> > > >  * '-mon' is obsolete, deprecated and replaced by '-object',
> > > >    but -monitor, -qmp and kept  as high level syntax sugar
> > > 
> > > I was giving this a spin by implementing libvirt support for it. The
> > > basic usage seems to work well but I've encoutered a regression in
> > > behaviour which happens both when the QMP monitor is instantiated using
> > > the new syntax but also the old one
> > > ( -mon chardev=charmonitor,id=monitor,mode=control)
> > > 
> > > When libvirt wants to re-connect to a qemu running with these patches it
> > > re-issues a 'qmp_capabilities' command after connecting.
> > > 
> > > With these patches qemu does not respond to that command even when the
> > > old syntax is used. This makes libvirt stuck when re-connecting to VMs.
> > 
> > Oh that's odd, thanks for the warning. I'll investigate this.
> 
> /face-palm - a misplaced a line in the auto-delete code, and did not
> have test coverage of re-opening a monitor. Try testing libvirt with
> this additional change on top
> 
> diff --git a/monitor/qmp.c b/monitor/qmp.c
> index f6b7fe65bb..10d651ca5c 100644
> --- a/monitor/qmp.c
> +++ b/monitor/qmp.c
> @@ -613,7 +613,6 @@ static void monitor_qmp_event(void *opaque, QEMUChrEvent 
> event)
>          qobject_unref(data);
>          break;
>      case CHR_EVENT_CLOSED:
> -        mon->delete_pending = true;
>          /*
>           * Note: this is only useful when the output of the chardev
>           * backend is still open.  For example, when the backend is
> @@ -629,6 +628,7 @@ static void monitor_qmp_event(void *opaque, QEMUChrEvent 
> event)
>          case MONITOR_QMP_CLOSE_ACTION_NONE:
>              break; /* nada */
>          case MONITOR_QMP_CLOSE_ACTION_DELETE:
> +            mon->delete_pending = true;
>              /*
>               * Do NOT run in the AIO context associated with the
>               * monitor. We need to run in the default AIO context

With this, startup, reconnect and FD passing work correctly both when
instantiated via -object and via -mon and it satisfies the normal
libvirt usage.

Tested-by: Peter Krempa <[email protected]>

I've tested it with following libvirt patches:

https://lists.libvirt.org/archives/list/[email protected]/thread/PUVLG73H35JZQPYNMHSBIVUF56EKEEC2/

Reply via email to