Re: [Qemu-devel] [PATCH v2 3/6] migration: move global_state.optional out

[Peter Xu <peterx@redhat.com>]

On Tue, Jun 13, 2017 at 08:16:35AM -0300, Eduardo Habkost wrote:
> On Tue, Jun 13, 2017 at 11:41:02AM +0800, Peter Xu wrote:
> > On Mon, Jun 12, 2017 at 04:18:06PM +0800, Peter Xu wrote:
> > > On Fri, Jun 09, 2017 at 10:40:10AM -0300, Eduardo Habkost wrote:
> > > > On Fri, Jun 09, 2017 at 11:48:59AM +0800, Peter Xu wrote:
> > > > > Put it into MigrationState then we can use the properties to specify
> > > > > whether to enable storing global state.
> > > > > 
> > > > > Removing global_state_set_optional() since now we can use HW_COMPAT_2_3
> > > > > for x86/power in general, and the register_compat_prop() for xen_init().
> > > > > 
> > > > > Signed-off-by: Peter Xu <peterx@redhat.com>
> [...]
> > > > > diff --git a/hw/xen/xen-common.c b/hw/xen/xen-common.c
> > > > > index 0bed577..8240d50 100644
> > > > > --- a/hw/xen/xen-common.c
> > > > > +++ b/hw/xen/xen-common.c
> > > > > @@ -138,7 +138,13 @@ static int xen_init(MachineState *ms)
> > > > >      }
> > > > >      qemu_add_vm_change_state_handler(xen_change_state_handler, NULL);
> > > > >  
> > > > > -    global_state_set_optional();
> > > > > +    /*
> > > > > +     * TODO: make sure global MigrationState has not yet been created
> > > > > +     * (otherwise the compat trick won't work). For now we are in
> > > > > +     * configure_accelerator() so we are mostly good. Better to
> > > > > +     * confirm that in the future.
> > > > > +     */
> > > > 
> > > > So, this makes accelerator initialization very sensitive to
> > > > ordering.
> > > > 
> > > > > +    register_compat_prop("migration", "store-global-state", "off");
> > > > 
> > > > It's already hard to track down machine-type stuff that is
> > > > initialized at machine->init() time but it's not introspectable,
> > > > let's not do the same mistake with accelerators.
> > > > 
> > > > Can't this be solved by a AcceleratorClass::global_props field,
> > > > so we are sure there's a single place in the code where globals
> > > > are registered?  (Currently, they are all registered by the few
> > > > lines around the machine_register_compat_props() call in main()).
> > > > 
> > > > I think I see other use cases for a new
> > > > AcceleratorClass::global_props field.  e.g.: replacing
> > > > kvm_default_props and tcg_default_props in target/i386/cpu.c.
> > > 
> > > Hmm, this sounds nice.  Then we can also get rid of the TODO above.
> > > 
> > > Thanks for your suggestion and review! I'll see whether I can writeup
> > > something as RFC for this.
> > 
> > After a second thought, I see these requirements are slightly
> > different.
> > 
> > For xen_init(), AccelClass::global_props may help since that's mainly
> > properties to be setup for TYPE_MIGRATION (let's assume it's okay that
> > MigrationState can be a QDev).
> > 
> > For kvm_default_props and tcg_default_props, AccelClass::global_props
> > may not help since they are setting up properties for TYPE_CPU, and
> > TYPE_CPU cannot really use global property features yet. :(
> 
> TYPE_CPU can use global properties, for sure.  TYPE_CPU is a
> TYPE_DEVICE subclass, and -cpu is implemented internally using
> global properties.

Indeed. :)

Thanks, 

-- 
Peter Xu