Re: [Qemu-devel] [PATCH V4 4/7] qmp: Allow to change password on names block driver states.

[Kevin Wolf <kwolf@redhat.com>]

Am 10.12.2013 um 15:06 hat Luiz Capitulino geschrieben:
> On Tue, 10 Dec 2013 10:57:50 +0100
> Kevin Wolf <kwolf@redhat.com> wrote:
> 
> > Am 09.12.2013 um 17:41 hat Luiz Capitulino geschrieben:
> > > On Mon, 9 Dec 2013 17:23:09 +0100
> > > Kevin Wolf <kwolf@redhat.com> wrote:
> > > 
> > > > > > I'm leaning slightly towards the approach that Benoît took, if only for
> > > > > > the naming aspect (that is, I also thought of the idea of a bool flag,
> > > > > > but didn't suggest it because I didn't like the implications on the
> > > > > > naming).  But I can live with either approach, if anyone else has a
> > > > > > strong opinion.
> > > > > 
> > > > > Well, we can pick up any descriptive name 'treat-device-as-a-node',
> > > > > 'device-is-a-graph-node'...
> > > > 
> > > > All devices are represented by nodes, so that doesn't make sense.
> > > > If anything, 'interpret-device-name-as-node-name', which at the same
> > > > time makes it pretty clear that we're abusing a field for something it
> > > > wasn't meant for.
> > > 
> > > Having two optionals where they can't be specified at the same time
> > > and can't be left off at the same time is a clear abuse as well.
> > 
> > Is it? If you wanted to express this in the schema, we'd need to extend
> > the QAPI generator, but until now we never have. I don't think this is
> > the first time that optional fields are not completely independent, but
> > may be required/forbidden based on values of other fields. Documenting
> > it should be enough, in my opinion.
> 
> We disagree here, and what makes my objection strong is that I
> provided an alternative which I believe is less worse because it
> makes less changes to the command.

My objection to your approach is strong because Benoît already sent an
alternative which I believe is less worse because with it, arguments
actually mean what their names tell instead of having additional bools
for "oh, and I said A, but I didn't mean it, I really want B".

> > > The truth is, both proposals are bad. This makes me think that maybe
> > > we should introduce a block API 2.0 and deprecate the current one
> > > (partly or completely).
> > 
> > Nice try, but of course this is equivalent to the "new command"
> > solution.  Deprecating the old version doesn't get you rid of it, you
> > still need to support it for compatibility. And then you're back to
> > square one.
> 
> We can't get rid of anything in QMP. Deprecating means that the command
> is still available but a better replacement exists and should be used
> in new implementations.

At least one thing on which we agree.

However, suggesting to deprecate the old version would mean that we
think that referencing by device name is a bad thing that you shouldn't
be doing any more. I don't think it's true.

> > For what it's worth, I think what Benoît implemented is the outcome of
> > discussions of the Block BOF on KVM Forum that included both block layer
> > developers and API users (i.e. libvirt), after considering and
> > dismissing other options (which, of course, included separate commands).
> 
> I appreciate those discussions, but patch review and acceptance happens
> on upstream lists.

Okay. So you're going to block an interface that block layer developers
and block layer users have agreed on because it doesn't match your
aesthetic demands - and you wrote some parts of the infrastructure we're
using, so you have an inherent right to have them met?

I love productive discussions like this.

Kevin