Re: [Qemu-devel] [PATCH 21/26] kvmclock: use realize for kvmclock

[Hu Tao]

On Sun, Jun 30, 2013 at 04:36:13PM +0200, Andreas Färber wrote:
> Am 25.06.2013 19:45, schrieb Eduardo Habkost:
> > On Tue, Jun 25, 2013 at 10:20:08AM +0800, Hu Tao wrote:
> > [...]
> >>> Is TYPE_SYS_BUS_DEVICE guaranteed to never override ->realize() itself?
> >>>
> >>> From DeviceClass documentation:
> >>>
> >>>  * If a type derived directly from TYPE_DEVICE implements @realize, it 
> >>> does
> >>>  * not need to implement @init and therefore does not need to store and 
> >>> call
> >>>  * #DeviceClass' default @realize callback.
> >>>  * For other types consult the documentation and implementation of the
> >>>  * respective parent types.
> >>>
> >>> The problem is that there's no documentation about ->realize() on
> >>> SysBusDeviceClass. Can we please explicitly document SysBusDeviceClass
> >>> expectations about ->realize() first, before making those changes?
> 
> If someone wants to add a paragraph to sysbus.h:SysBusDeviceClass
> documentation I would happily ack or pick that up. :)
> 
> >> IIUC, subclass's overriding ->realize should call parent's ->realize at
> >> some point. Peter Crosthwaite has a patchset to access a object's parent
> >> class at 
> >> http://lists.nongnu.org/archive/html/qemu-devel/2013-06/msg02982.html
> >>
> >> Regarding SysBusDevice::init and SysBusDevice::realize, I think it's the
> >> same as in the case of DeviceClass. If you agree I'll document it as in
> >> DeviceClass.
> > 
> > I believe it is reasonable to document that SysBusDevice subclasses
> > don't need to call the parent ->realize() method, like on DeviceClass.
> > This is exactly what this patch set does, after all, and I haven't seen
> > anybody complaining about it yet.
> 
> So the thing is that SysBusDevice's DeviceClass::init implementation,
> called by DeviceState's DeviceClass::realize implementation, just calls
> SysBusDeviceClass::init if non-NULL. When we introduce our own device's
> realizefn to replace our SysBusDeviceClass::init implementation, there
> is no point calling that effectively no-op DeviceClass::realize
> implementation.

This is true because we are in transition from DeviceClass:init to
DeviceClass:realize, by calling sub-class's DeviceClass:init in
DeviceClass's realize. But once the transition is done, and
DeviceClass's (and any intermediate devired classes') realize does
do something, we can't just ignore it in overriding realize.

Fix me if i'm wrong.

>                 And if we tried to, ...
> * ... how would we decide whether to call the parent's implementation
> first or last? It's not yes or no, it's no or when. Changing between
> either is more than just moving one line, it affects error handling and
> with knowledge about parent implementation never failing we could so far
> save us some work.

Agreed.

> * ... with the current QOM method scheme we'd go insane introducing a
> FooClass per device to save SysBusDevice's DeviceClass::realize in
> FooClass::parent_realize. Still waiting for Anthony on whether and how
> we want to change our scheme.
> 
> Long story short: If someone wants to mess with base classes they get to
> check derived classes for compatibility with their change.
> 
> Ideally qtest would help automate that to some degree.
> I would be all in favor if someone wanted to add a dummy test case per
> non-default, non-KVM device converted so that we can see that we didn't
> screw up -device instantiation too badly.
> 
> Regards,
> Andreas
> 
> -- 
> SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
> GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg