Re: [Qemu-devel] [PATCH v3 3/6] block QMP: Drop query-block member "type" (type= in info block)

[Luiz Capitulino <lcapitulino@redhat.com>]

On Fri, 13 May 2011 16:13:55 +0200
Markus Armbruster <armbru@redhat.com> wrote:

> Luiz Capitulino <lcapitulino@redhat.com> writes:
> 
> > On Fri, 13 May 2011 10:26:38 +0200
> > Markus Armbruster <armbru@redhat.com> wrote:
> >
> >> Luiz Capitulino <lcapitulino@redhat.com> writes:
> >> 
> >> > On Thu, 12 May 2011 19:54:40 +0200
> >> > Markus Armbruster <armbru@redhat.com> wrote:
> >> >
> >> >> Luiz Capitulino <lcapitulino@redhat.com> writes:
> >> >> 
> >> >> > On Thu, 12 May 2011 19:12:56 +0200
> >> >> > Markus Armbruster <armbru@redhat.com> wrote:
> >> >> >
> >> >> >> Luiz Capitulino <lcapitulino@redhat.com> writes:
> [...]
> >> >> >> > Also, we can't just drop it from QMP. We should first note it's deprecated.
> >> >> >> 
> >> >> >> Would you accept a change to the more honest value "unknown" for the
> >> >> >> deprecation period?
> >> >> >
> >> >> > We have to avoid breaking the protocol. Changing something that has always
> >> >> > been reported as 'cdrom' to 'unknown' will likely cause as many as damages
> >> >> > as dropping the command.
> >> >> 
> >> >> I can cause damage only if somebody is using it.  Which I doubt.
> >> >
> >> > Me too and I'd agree with this patch if I was 100% sure. But it's impossible
> >> > to be sure, unless we do it by trial and error which is harmful.
> >> >
> >> >> Remember, the value is unreliable.  It's a *lie*.  We can stop lying in
> >> >> two ways: shut up (drop member "type"), or tell the truth (change the
> >> >> value to "unknown", which is a documented value of "type").
> >> >
> >> > Can we set it to 'unknown' when if=none?
> >> 
> >> Maybe.  Makes query-block mix up host and guest information again.
> >
> > It's temporary, just to respect our deprecation policy and give us time
> > to provide a viable alternative.
> >
> >> Purging guest information from block.c is the point of this series.
> >> Therefore, query-block can't be done in block.c anymore.  It needs to
> >> move to blockdev.c, where the mixed-up-by-design DriveInfo is available.
> >> It could move back when we finally clean up query-block.
> >> 
> >> Even with such compatibility gymnastics, it could still break your
> >> hypothetical client.
> >
> > Our deprecation policy is not hypothetical. There isn't case by case. Either
> > we respect it or we don't.
> 
> My point is: even your deprecation policy cannot protect your
> hypothetical client.
> 
> The only way to ensure not even the silliest hypothetical client breaks
> is not to change anything.  Which means we cannot fix query-block not to
> lie about the type, period.  Which means we have to deprecate
> query-block wholesale to fix that bug.

Why don't we force the known broken cases (like if=none) to 'unknown',
note in the doc that the field is unreliable and specify that it can be
dropped in 2 or 3 releases?

> 
> Kevin, please apply patches 1+2+6.  Feel free to drop 3-5.
> 
> [...]
>