Re: [Qemu-devel] [PATCH 1/3] virtio: add bus_plugged() callback to VirtioDeviceClass

[Frederic Konrad <fred.konrad@greensocs.com>]

On 04/06/2013 19:35, Andreas Färber wrote:
> Hi,
>
> Am 04.06.2013 18:22, schrieb Jesse Larrew:
>> Virtio devices are initialized prior to plugging them into a bus. However,
>> other initializations (such as host_features) don't occur until after the
>> device is plugged into the bus. If a device needs to modify it's
>> configuration based on host_features, then it needs to be notified when the
>> bus is attached and host_features is available for use.
>>
>> This patch extends struct VirtioDeviceClass to add a bus_plugged() method.
>> If implemented by a device, it will be called after the device is attached
>> to a bus.
>>
>> Signed-off-by: Jesse Larrew <jlarrew@linux.vnet.ibm.com>
> I think this is backwards...
>
> First of all, why is host_features not available before?

Hi Andreas,

The major issue here is that host_features is modified after virtio 
devices are inited.

in virtio_pci_device_plugged:
     proxy->host_features |= 0x1 << VIRTIO_F_NOTIFY_ON_EMPTY;
     proxy->host_features |= 0x1 << VIRTIO_F_BAD_FEATURE;
     proxy->host_features = virtio_bus_get_vdev_features(bus,
proxy->host_features);

and virtio_pci_device_plugged must be called after the virtio device is 
inited.

> A hook on the bus makes sense because it allows central handling for any
> devices on that bus.
> However for a device, first TypeInfo::instance_init is run, then
> qdev_set_parent_bus() connects the bus and finally DeviceClass::realize
> is run - and we want to postpone realize further in the future.
>
> So why can't this be in VirtioDevice's or VirtIONet's realize method? At
> realize time we should definitely be on the bus in this case. I.e.,

Is that possible with hotplugging virtio device on virtio-mmio?
> create vdev->config only after we know how large it needs to be rather
> than creating and later resizing it, which might fail.
>
> Regards,
> Andreas
>
Fred