Re: Examining device state via monitor for debugging

[Daniel P . Berrangé]

On Tue, Jun 11, 2024 at 09:13:00AM +0100, Mark Cave-Ayland wrote:
> On 11/06/2024 08:57, Daniel P. Berrangé wrote:
> 
> > On Tue, Jun 11, 2024 at 07:49:12AM +0200, Markus Armbruster wrote:
> > > Philippe Mathieu-Daudé <philmd@linaro.org> writes:
> > > 
> > > > Officialise the QMP command, use the existing
> > > > hmp_info_human_readable_text() helper.
> > > 
> > > I'm not sure "officialise" is a word :)
> > > 
> > > Taking a step back...  "info via" and its new QMP counterpart
> > > x-query-mos6522-devices dump device state.  I understand why examining
> > > device state via monitor can be useful for debugging.  However, we have
> > > more than 2000 devices in the tree.  Clearly, we don't want 2000 device
> > > state queries.  Not even 100.  Could we have more generic means instead?
> > > 
> > > We could use QOM (read-only) properties to expose device state.
> > > 
> > > If we use one QOM property per "thing", examining device state becomes
> > > quite tedious.  Also, you'd have to stop the guest to get a consistent
> > > view, and adding lots of QOM properties bloats the code.
> > > 
> > > If we use a single, object-valued property for the entire state, we get
> > > to define the objects in QAPI.  Differently tedious, and bloats the
> > > generated code.
> > > 
> > > We could use a single string-valued property.  Too much of an abuse of
> > > QOM?
> > 
> > Yeah, I'd suggest we just keep it dumb and free form, adding a
> > callback like this to the QOM base class:
> > 
> >    HumanReadableText (*debug_state)(Error **errp);
> 
> I think that's a little bit too restrictive: certainly I can see the need
> for allowing parameters for the output to be customised, and there may also
> be cases where a device may want to register more than one debug (or
> monitor) command.
> 
> Currently I quite like Manos' solution since it allows a MonitorInterface to
> be defined, and that could have separate methods for registering both "info"
> and "debug" commands. Longer term this could allow for e.g. TYPE_TCG_ACCEL
> to register "info jit" and devices can add whatever debug monitor commands
> they need.

The downside is that this wires the monitor APIs into the internals of the
devices, instead of having internals work exclusively in terms of QAPI
types.

With regards,
Daniel
-- 
|: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
|: https://libvirt.org         -o-            https://fstop138.berrange.com :|
|: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|