[Intel-gfx] [RFC PATCH 2/2] mdev: introduce device specific ops

[Jason Wang]

On 2019/9/18 上午10:57, Tian, Kevin wrote:
>> From: Jason Wang [mailto:jasowang at redhat.com]
>> Sent: Tuesday, September 17, 2019 6:17 PM
>>
>> On 2019/9/17 下午4:09, Tian, Kevin wrote:
>>>> From: Jason Wang
>>>> Sent: Thursday, September 12, 2019 5:40 PM
>>>>
>>>> Currently, except for the crate and remove. The rest fields of
>>>> mdev_parent_ops is just designed for vfio-mdev driver and may not
>> help
>>>> for kernel mdev driver. So follow the device id support by previous
>>>> patch, this patch introduces device specific ops which points to
>>>> device specific ops (e.g vfio ops). This allows the future drivers
>>>> like virtio-mdev to implement its own device specific ops.
>>> Can you give an example about what ops might be required to support
>>> kernel mdev driver? I know you posted a link earlier, but putting a small
>>> example here can save time and avoid inconsistent understanding. Then
>>> it will help whether the proposed split makes sense or there is a
>>> possibility of redefining the callbacks to meet the both requirements.
>>> imo those callbacks fulfill some basic requirements when mediating
>>> a device...
>> I put it in the cover letter.
>>
>> The link ishttps://lkml.org/lkml/2019/9/10/135  which abuses the current
>> VFIO based mdev parent ops.
>>
>> Thanks
> So the main problem is the handling of userspace pointers vs.
> kernel space pointers. You still implement read/write/ioctl
> callbacks which is a subset of current parent_ops definition.
> In that regard is it better to introduce some helper to handle
> the pointer difference in mdev core, while still keeping the
> same set of parent ops (in whatever form suitable for both)?


Pointers is one of the issues. And read/write/ioctl is designed for 
userspace API not kernel. Technically, we can use them for kernel but it 
would not be as simple and straightforward a set of device specific 
callbacks functions. The link above is just an example, e.g we can 
simply pass the vring address through a dedicated API instead of 
mandatory an offset of a file.

Thanks

>