Re: [Qemu-devel] [PATCH v6 09/10] migration: merge enforce_config_section somewhat

[Eduardo Habkost <ehabkost@redhat.com>]

On Thu, Jun 29, 2017 at 11:00:13AM +0800, Peter Xu wrote:
> On Thu, Jun 29, 2017 at 12:42:56AM +0200, Juan Quintela wrote:
> > Eduardo Habkost <ehabkost@redhat.com> wrote:
> > 
> > >
> > > So, this is a case where a user-provided config option (-machine
> > > enforce-config-section) should trigger a different default in another
> > > class (migration.send-configuration).
> > >
> > > Also, the new default triggered by -machine has a very specific
> > > priority:
> > >
> > > * AccelClass::global_props must not override "-machine enforce-config-section=on"
> > > * MachineClass::compat_props must not override
> > >   "-machine enforce-config-section=on"
> > >
> > > We must also decide in advance what should be result of:
> > > * "-machine enforce-config-section=on -global migration.send-configuration=off"
> > > * "-machine enforce-config-section=off -global migration.send-configuration=on"
> > > * "-global migration.send-configuration=off -machine enforce-config-section=off"
> > > * "-global migration.send-configuration=on -machine enforce-config-section=on"
> 
> Yes, this is considered before this patch: currently
> enforce-config-section will have the highest priority in case if
> someone used both of the old & new parameters for it (considering
> "enforce-config-section" has the word "enforce" inside, it makes some
> sense). While...
> 
> > 
> > BOOM!!!!!
> > 
> > We use old configuration or new one.
> 
> ... I agree more with Juan here, that user should not really specify
> these two parameters at the same time. If the user knows the new
> parameter, he/she should know that the new one is obsoleting the old
> one. And since even for that case this patch can handle it well (will
> take -M param), I think it's okay.

If that's the intended result, it's OK to me.  But I think the
relationship between enforce-config-section and
migration.send-configuration should be documented in qemu-options.hx.
And considering that this could break silently in future code
refactoring, an automated test would be interesting (but not critical,
as setting options contradicting each other is not a common scenario).


> 
> > 
> > >
> > > I'm not sure what we should decide about these 4 cases above, but I
> > > believe it would be safer to encode that decision at the same place we
> > > handle the priority between accel/machine/user globals:
> > > register_global_properties() at vl.c.
> > >
> > >
> > > Or maybe this extra complexity is a sign that we shouldn't try to add
> > > extra magic to make -machine affect the "migration" object properties,
> > > and keep the existing machine->enforce_config_section check in the
> > > migration code?  I'm not sure.
> > 
> > Not sure there either.  I preffer doing it in a single place, but I am
> > not the expert here.
> 
> IMHO it is not necessary to introduce such a thing in
> register_global_properties(). AFAIU this is the only place where one
> machine property can collapse with a global property? And it currently
> only happens in migration codes. Actually it is well ordered, since we
> init the migration object after register_global_properties(), so
> everthing should possibly be fine. Introducing framework-level thing
> for this may only make things more complicated imho.

True.  Considering we need to keep the "overrides everything else"
semantics of enforce-config-section, your approach is not bad.

> 
> After all we can remove all these one day when we can obsolete the
> "enforce-config-section" parameter (maybe we should add one OBSOLETE
> warning when the -M parameter is used).  Thanks,

I don't think we need a warning, but a documentation update is
important, IMO.

-- 
Eduardo