Re: [RFC PATCH 3/4] virtio: introudce a mdev based transport

[Jason Wang <jasowang@redhat.com>]

On 2019/9/11 下午5:36, Michael S. Tsirkin wrote:
> On Wed, Sep 11, 2019 at 10:38:39AM +0800, Jason Wang wrote:
>> On 2019/9/10 下午9:52, Michael S. Tsirkin wrote:
>>> On Tue, Sep 10, 2019 at 09:13:02PM +0800, Jason Wang wrote:
>>>> On 2019/9/10 下午6:01, Michael S. Tsirkin wrote:
>>>>>> +#ifndef _LINUX_VIRTIO_MDEV_H
>>>>>> +#define _LINUX_VIRTIO_MDEV_H
>>>>>> +
>>>>>> +#include <linux/interrupt.h>
>>>>>> +#include <linux/vringh.h>
>>>>>> +#include <uapi/linux/virtio_net.h>
>>>>>> +
>>>>>> +/*
>>>>>> + * Ioctls
>>>>>> + */
>>>>> Pls add a bit more content here. It's redundant to state these
>>>>> are ioctls. Much better to document what does each one do.
>>>> Ok.
>>>>
>>>>
>>>>>> +
>>>>>> +struct virtio_mdev_callback {
>>>>>> +	irqreturn_t (*callback)(void *);
>>>>>> +	void *private;
>>>>>> +};
>>>>>> +
>>>>>> +#define VIRTIO_MDEV 0xAF
>>>>>> +#define VIRTIO_MDEV_SET_VQ_CALLBACK _IOW(VIRTIO_MDEV, 0x00, \
>>>>>> +					 struct virtio_mdev_callback)
>>>>>> +#define VIRTIO_MDEV_SET_CONFIG_CALLBACK _IOW(VIRTIO_MDEV, 0x01, \
>>>>>> +					struct virtio_mdev_callback)
>>>>> Function pointer in an ioctl parameter? How does this ever make sense?
>>>> I admit this is hacky (casting).
>>>>
>>>>
>>>>> And can't we use a couple of registers for this, and avoid ioctls?
>>>> Yes, how about something like interrupt numbers for each virtqueue and
>>>> config?
>>> Should we just reuse VIRTIO_PCI_COMMON_Q_XXX then?
>>
>> You mean something like VIRTIO_PCI_COMMON_Q_MSIX? Then it becomes a PCI
>> transport in fact. And using either MSIX or irq number is actually another
>> layer of indirection. So I think we can just write callback function and
>> parameter through registers.
> I just realized, all these registers are just encoded so you
> can pass stuff through read/write. But it can instead be
> just a normal C function call with no messy encoding.
> So why do we want to do this encoding?


Just because it was easier to start as a POC since mdev_parent_ops is 
the only way to communicate between mdev driver and mdev device right 
now. We can invent private ops besides mdev_parent_ops, e.g a private 
pointer in mdev_parent_ops. I can try this in next version.

Thanks