Re: [PATCH 08/11] mos6522: add "info via" HMP command for debugging

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Mark Cave-Ayland (mark.cave-ayland@ilande.co.uk) wrote:
> On 21/02/2022 17:11, Daniel P. Berrangé wrote:
> 
> > On Sun, Feb 20, 2022 at 05:18:33PM +0000, Mark Cave-Ayland wrote:
> > > On 08/02/2022 13:10, Daniel P. Berrangé wrote:
> > > 
> > > > On Tue, Feb 08, 2022 at 01:06:59PM +0000, Mark Cave-Ayland wrote:
> > > > > On 08/02/2022 12:49, Daniel P. Berrangé wrote:
> > > > > 
> > > > > > > I was under the impression that monitor_register_hmp_info_hrt() does all the
> > > > > > > magic here i.e. it declares the underlying QMP command with an x- prefix and
> > > > > > > effectively encapsulates the text field in a way that says "this is an
> > > > > > > unreliable text opaque for humans"?
> > > > > > 
> > > > > > The monitor_register_hmp_info_hrt only does the HMP glue side, and
> > > > > > that's only needed if you must dynamically register the HMP command.
> > > > > > For statically registered commands set '.cmd_info_hrt' directly in
> > > > > > the hml-commands-info.hx for the HMP side.
> > > > > > 
> > > > > > > If a qapi/ schema is needed could you explain what it should look like for
> > > > > > > this example and where it should go? Looking at the existing .json files I
> > > > > > > can't immediately see one which is the right place for this to live.
> > > > > > 
> > > > > > Take a look in qapi/machine.json for anyof the 'x-query-XXXX' commands
> > > > > > there. The QAPI bit is fairly simple.
> > > > > > 
> > > > > > if you want to see an illustration of what's different from a previous
> > > > > > pure HMP impl, look at:
> > > > > > 
> > > > > >      commit dd98234c059e6bdb05a52998270df6d3d990332e
> > > > > >      Author: Daniel P. Berrangé <berrange@redhat.com>
> > > > > >      Date:   Wed Sep 8 10:35:43 2021 +0100
> > > > > > 
> > > > > >        qapi: introduce x-query-roms QMP command
> > > > > 
> > > > > I see, thanks for the reference. So qapi/machine.json would be the right
> > > > > place to declare the QMP part even for a specific device?
> > > > > 
> > > > > Even this approach still wouldn't work in its current form though, since as
> > > > > mentioned in my previous email it seems that only the target CONFIG_*
> > > > > defines and not the device CONFIG_* defines are present when processing
> > > > > hmp-commands-info.hx.
> > > > 
> > > > Yeah, that's where the pain comes in.  While QAPI schema can be made
> > > > conditional on a few CONFIG_* parameters - basically those derived
> > > > from global configure time options, it is impossible for this to be
> > > > with with target specific options like the device CONFIG_* defines.
> > > > 
> > > > This is why I suggested in my othuer reply that it would need to be
> > > > done with a generic 'info dev-debug' / 'x-query-dev-debug' command
> > > > that can be registered unconditionally, and then individual devices
> > > > plug into it.
> > > 
> > > After some more experiments this afternoon I still seem to be falling
> > > through the gaps on this one. This is based upon my understanding that all
> > > new HMP commands should use a QMP HumanReadableText implementation and the
> > > new command should be restricted according to target.
> > > 
> > > Currently I am working with this change to hmp-commands-info.hx and
> > > qapi/misc-target.json:
> > 
> > [snip]
> > > i.e. qmp_marshal_output_HumanReadableText() isn't protected by the #if
> > > TARGET guards and since HumanReadableText is only used by the new
> > > qmp_x_query_via() functionality then the compiler complains and aborts the
> > > compilation.
> > > 
> > > Possibly this is an error in the QAPI generator for types hidden behind
> > > commands using "if"? Otherwise I'm not sure what is the best way to proceed,
> > > so I'd be grateful for some further pointers.
> > 
> > Yes, this is pretty much what I expect and exactly what I hit with
> > other target specific commands.
> > 
> > That's why I suggested something like a general 'x-device-debug' command
> > that is NOT conditionalized in QAPI, against which dev impls can register
> > a callback to provide detailed reporting, instead of a device type specific
> > command.
> 
> Ah so this is a known issue with this approach then. David mentioned earlier
> in the thread that he'd be okay with a HMP command if it was useful and
> restricted to the required targets, so would it be okay to add "info via"
> for now as just a (non-QMP wrapped) HMP info command if I can get that to
> work?

I still am from an HMP point of view; it sounds like the right way in
the future is to get the info devices or whatever;  I suggest you keep
it as close to a QMP implementation as possible, still with the
HumanReadableText stuff.
(Others might still be nervous about an HMP special; but I don't see
it's worth holding this trivial stuff up for it).

Dave

> 
> ATB,
> 
> Mark.
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK