Re: [Qemu-devel] [RFC PATCH v1 0/7] QOM Super class access

[Andreas Färber]

Am 18.06.2013 12:39, schrieb Michael S. Tsirkin:
> On Tue, Jun 18, 2013 at 08:35:22PM +1000, Peter Crosthwaite wrote:
>> Hi Michael,
>>
>> On Tue, Jun 18, 2013 at 8:23 PM, Michael S. Tsirkin <address@hidden> wrote:
>>> On Tue, Jun 18, 2013 at 07:43:11PM +1000, address@hidden wrote:
>>>> From: Peter Crosthwaite <address@hidden>
>>>>
>>>>
>>>> This series enables QOM super class access and demostrates some usages.
>>>> Replaces the save->override->call via FooClass technique, to reduce
>>>> some of the boiler plate in recently fully QOMified devices.
>>>>
>>>> Applied the change to ARM CPU, MB CPU and some of Andreas's recently
>>>> QOMified i386 devices, all which have the save->override->call issue.
>>>> ARMCPU I've done a brief test on and seems to work.
>>>>
>>>> ARM CPU was particularly difficult, as it has 3 layers of heirachy,
>>>> where a non-concrete class (TYPE_ARM_CPU) need to super class itself
>>>> (to TYPE_CPU). This sees the need for super-classers to specify their
>>>> expected base class level. See patches for illustration.
>>>>
>>>> The main future work to the series is to apply the change pattern to
>>>> the reset of the tree
>>>
>>> Looks good to me overall.
>>> Some nits:
>>> - Super is an immediate parent in java and python.
>>
>> s/Super/parent might be the go. But it is designed to work like
>> py/java. Its the immediate parent of the specified level, and it is
>> analogous to the java super.
>>
>>> - One of the design points of QOM is that it let
>>>   you ignore which class is a parent and which is a child.
>>>   All casts look the same.
>>>
>>> So, why do we need the new APIs with _SUPER?
>>
>> The SUPER APIs have a nice consistent appearance with the GET_CLASS
>> and FOO_CLASS APIs and they are likely to live alongside each other in
>> QOM fns.
>>
>>> What's wrong with simple
>>>         object_class_by_name()
>>> and casting to that?
>>>
>>
>> There a performance consequence - object_class_by_name is a lookup,
>> whereas this approach is able to just walk the pointer through the
>> inheritance heirachy till it hits.
> 
> None of these uses has a chance to be at all performance
> sensitive.
> 
> 
>> Tying it to an object also brings
>> into play the possibility of a cast cache should that be needed.
> 
> Doesn't make sense to me. This is object construction,
> how many classes do you expect there to be so
> this would affect qemu startup speed noticeably?
> 1000000?

It's not just about startup, also about CPU reset for instance.
But still not that performance-sensitive.

If we want to performance-optimize it, sharing code with
object_class_by_name() or any other class-comparing function would be nice.

Andreas

> 
>> Thanks for the review.
>>
>> Regards,
>> Peter
>>
>>>
>>>>
>>>> Peter Crosthwaite (7):
>>>>   target-arm/cpu.c: delete un-needed instance/class sizes
>>>>   qom: Add super class accessor
>>>>   qdev-core: Introduce DEVICE super class cast macros
>>>>   qom/cpu: Introduce CPU super class cast macros
>>>>   target-arm: Remove ARMCPUClass
>>>>   target-microblaze: Remove MicroblazeCPUClass
>>>>   i8254: Remove [KVM]PITClass
>>>>
>>>>  hw/i386/kvm/i8254.c         | 17 ++---------------
>>>>  hw/timer/i8254.c            | 16 ++--------------
>>>>  include/hw/qdev-core.h      |  4 ++++
>>>>  include/qom/cpu.h           |  4 ++++
>>>>  include/qom/object.h        | 18 ++++++++++++++++++
>>>>  qom/object.c                | 15 +++++++++++++++
>>>>  target-arm/cpu-qom.h        | 20 --------------------
>>>>  target-arm/cpu.c            | 16 +++++-----------
>>>>  target-microblaze/cpu-qom.h | 20 --------------------
>>>>  target-microblaze/cpu.c     | 13 ++++---------
>>>>  10 files changed, 54 insertions(+), 89 deletions(-)
>>>>
>>>> --
>>>> 1.8.3.rc1.44.gb387c77.dirty
>>>


-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg