On 2019/11/20 下午9:41, Jason Gunthorpe wrote:
On Wed, Nov 20, 2019 at 12:07:59PM +0800, Jason Wang wrote:
1) create sub fucntion and do must to have pre configuration through devlink
2) only after sub function is created one more available instance was added
and shown through sysfs
3) user can choose to create and use that mdev instance as it did for other
type of device like vGPU
4) devlink can still use to report other stuffs
Why do we want the extra step #3? The user already indicated they want
a mdev via #1
It's about the compatibility, but if you wish, I think we can develop
devlink based lifecycle for mdev for sure.
I have the same question for the PF and VF cases, why doesn't a mdev
get created automatically when the VF is probed? Why does this need
the guid stuff?
All you said here is possible, it's a design choice for the management
interface.
The guid stuff was intended for, essentially, multi-function devices
that could be sliced up, I don't think it makes sense to use it for
single-function VF devices like the ICF driver.
It doesn't harm, and indeed we have other choice, we can do it gradually
on top.
Overall the guid thing should be optional. Drivers providing mdev
should be able to use another scheme, like devlink, to on demand
create their mdevs.
Yes, that's for sure. I'm not against to devlink for mdev/subdev, I just
say we should not make devlink the only choice for mdev/subdev.
Thanks
Jason
