Re: [Qemu-devel] [PATCH 09/10] s390x/css: turn on channel subsystem migration

[Halil Pasic]

On 05/08/2017 08:37 PM, Dr. David Alan Gilbert wrote:
> * Halil Pasic (address@hidden) wrote:
>>
>>
>> On 05/08/2017 07:27 PM, Dr. David Alan Gilbert wrote:
>>> * Halil Pasic (address@hidden) wrote:
>>>> Turn on migration for the channel subsystem and the new scheme for
>>>> migrating virtio-ccw proxy devices (instead of letting the transport
>>>> independent child device migrate it's proxy, use the usual
>>>> DeviceClass.vmsd mechanism) for future machine versions.
>>
>> [..]
>>
>>>> @@ -1365,6 +1373,11 @@ static void virtio_ccw_device_plugged(DeviceState 
>>>> *d, Error **errp)
>>>>      sch->id.cu_model = virtio_bus_get_vdev_id(&dev->bus);
>>>>  
>>>>  
>>>> +    /* Avoid generating unknown section for legacy migration target. */
>>>> +    if (!css_migration_enabled()) {
>>>> +        DEVICE_GET_CLASS(ccw_dev)->vmsd = NULL;
>>>> +    }
>>>> +
>>>
>>> That's a very odd thing to do; can't you use a .needed at the
>>> top level of the vmstate_virtio_ccw_dev to avoid having to
>>> set it to NULL like this?
>>>
>>
>> I agree it's odd. As far as I remember I can't use .needed but
>> I will double check.
> 
> You can have top level .needed's - both vmstate_globalstate in
> migration.c and colo's colo_state use them.
> 

Works like charm. I'm really happy to get rid of this ugly hunk.  Thanks
a lot! 

I was probably confused by the fact that I want to use the same vmsd with
vmstate_save_state when the needed is false. That works, but I have probably 
blindly assumed (back then) it does not. Of course it does make sense to
ignore .needed in that function, because for a vmsd coming from a
recursive call while processing a filed then the non-presence of a field
should be indicated by field_exists.

I wonder if adding a comment at the definition site would be helpful.
Something like:


 struct VMStateDescription {
....
     void (*pre_save)(void *opaque);
+    /* Controls the existence of sections and subsections, but not fields. */
     bool (*needed)(void *opaque);
     VMStateField *fields;
     const VMStateDescription **subsections;
 };

Halil

> Dave
> 
>> Many thanks for your review!
>>
>> Halil
>>
> --
> Dr. David Alan Gilbert / address@hidden / Manchester, UK
>