Re: [PATCH] vl: Add -set options to device opts dict when using JSON syntax for -device

[Damien Hedde]

On 12/21/21 16:40, Markus Armbruster wrote:
> Paolo Bonzini <pbonzini@redhat.com> writes:
>
> > On 12/21/21 13:58, Markus Armbruster wrote:
> > > > Is this a regression?  I suspect commit 5dacda5167 "vl: Enable JSON
> > > > syntax for -device" (v6.2.0).
> > >
> > > Obviously not a regression: everything that used to work still works.
> >
> > FWIW I think -set should be deprecated.  I'm not aware of any
> > particularly useful use of it.  There are a couple in the QEMU tests
> > (in vhost-user-test and in qemu-iotests 068), but in both cases the
> > code would be easier to follow without; patches can be dusted off if
> > desired.
>
> -set has its uses, but they're kind of obscure.  When you want to use
> some canned configuration with slight modifications, then -readconfig
> canned.cfg -set ... is nicer than editing a copy of canned.cfg.  For
> what it's worth, we have a few cans in docs/config/, and we refer to at
> least one of them in the manual (in docs/system/devices/usb.rst).
>
> There are a few really good ideas in QemuOpts.  I count -readconfig,
> -writeconfig and -set among them.  Unfortunately, they have been marred
> by us not converting the whole CLI to QemuOpts as envisaged.  And now we
> never will, because our needs have long outgrown what QemuOpts can
> provide.
>
> I'd love to have unmarred, QAPI-based replacements.  However, I doubt
> maintaining backwards compatibility will be practical and worthwhile.
>
> Declare these options unstable?
I agree.

Without QemuOpts, and more precisely the ability to alter them before 
they are really handled, this kind of feature will be impossible to have.

Only way I see, is to reverse the mechanism:
+ handle set option before, it store the key,value somewhere
+ an option like -device checks that store and fetch their values.

--
Damien