Peter Maydell <peter.mayd...@linaro.org> writes:
> On 7 September 2015 at 17:40, Markus Armbruster <arm...@redhat.com> wrote:
>> Peter Maydell <peter.mayd...@linaro.org> writes:
>>
>>> Convert the pxa2xx_mmci device to be a sysbus device.
>
>>> +static Property pxa2xx_mmci_properties[] = {
>>> + /* Note: pointer property 'drive' may remain NULL, thus no need
>>> + * for dc->cannot_instantiate_with_device_add_yet = true;
>>> + * Unfortunately this can't be a DEFINE_PROP_DRIVE, because
>>> + * setting a 'drive' property results in a call to blk_attach_dev()
>>> + * attaching the BlockBackend to this device; that then means that
>>> + * the call in sd_init() to blk_attach_dev_nofail() which tries to
>>> + * attach the BlockBackend to the SD card object aborts.
>>> + */
>>> + DEFINE_PROP_PTR("drive", PXA2xxMMCIState, blk),
>>> + DEFINE_PROP_END_OF_LIST(),
>>> +};
>>
>> As far as I can tell, this problem is an artifact of our interface to
>> the common sd-card code, namely sd_init(). sd_init() was made for the
>> pre-qdev world: it creates and completely initializes the common
>> SDState.
>>
>> In qdev, we do this in three separate steps: create, set properties,
>> realize. Split up sd_init(), and the problem should go away.
>
> Yes, but I don't really want to gate QOMification of mmc
> controller devices on the more complicated problem of
> QOMifying sd.c itself, especially since we already have several
> QOMified mmc controllers...
Is serial.c QOMified? I don't think so, it's merely structured in a
QOM-friendly way: typedef SerialState, realize helper
serial_realize_core(), unrealize helper serial_exit_core(). If
SerialState had more properties, we'd also need a macro to define them.
I believe sd.c could be made like that, with not much effort.
