Re: [RFC v4 0/3] vhost: introduce mdev based hardware backend

[Jason Wang <jasowang@redhat.com>]

On 2019/9/17 上午9:02, Tiwei Bie wrote:
> This RFC is to demonstrate below ideas,
>
> a) Build vhost-mdev on top of the same abstraction defined in
>     the virtio-mdev series [1];
>
> b) Introduce /dev/vhost-mdev to do vhost ioctls and support
>     setting mdev device as backend;
>
> Now the userspace API looks like this:
>
> - Userspace generates a compatible mdev device;
>
> - Userspace opens this mdev device with VFIO API (including
>    doing IOMMU programming for this mdev device with VFIO's
>    container/group based interface);
>
> - Userspace opens /dev/vhost-mdev and gets vhost fd;
>
> - Userspace uses vhost ioctls to setup vhost (userspace should
>    do VHOST_MDEV_SET_BACKEND ioctl with VFIO group fd and device
>    fd first before doing other vhost ioctls);
>
> Only compile test has been done for this series for now.


Have a hard thought on the architecture:

1) Create a vhost char device and pass vfio mdev device fd to it as a 
backend and translate vhost-mdev ioctl to virtio mdev transport (e.g 
read/write). DMA was done through the VFIO DMA mapping on the container 
that is attached.

We have two more choices:

2) Use vfio-mdev but do not create vhost-mdev device, instead, just 
implement vhost ioctl on vfio_device_ops, and translate them into 
virtio-mdev transport or just pass ioctl to parent.

3) Don't use vfio-mdev, create a new vhost-mdev driver, during probe 
still try to add dev to vfio group and talk to parent with device 
specific ops

So I have some questions:

1) Compared to method 2, what's the advantage of creating a new vhost 
char device? I guess it's for keep the API compatibility?

2) For method 2, is there any easy way for user/admin to distinguish e.g 
ordinary vfio-mdev for vhost from ordinary vfio-mdev?  I saw you 
introduce ops matching helper but it's not friendly to management.

3) A drawback of 1) and 2) is that it must follow vfio_device_ops that 
assumes the parameter comes from userspace, it prevents support kernel 
virtio drivers.

4) So comes the idea of method 3, since it register a new vhost-mdev 
driver, we can use device specific ops instead of VFIO ones, then we can 
have a common API between vDPA parent and vhost-mdev/virtio-mdev drivers.

What's your thoughts?

Thanks


>
> RFCv3: https://patchwork.kernel.org/patch/11117785/
>
> [1] https://lkml.org/lkml/2019/9/10/135
>
> Tiwei Bie (3):
>    vfio: support getting vfio device from device fd
>    vfio: support checking vfio driver by device ops
>    vhost: introduce mdev based hardware backend
>
>   drivers/vfio/mdev/vfio_mdev.c    |   3 +-
>   drivers/vfio/vfio.c              |  32 +++
>   drivers/vhost/Kconfig            |   9 +
>   drivers/vhost/Makefile           |   3 +
>   drivers/vhost/mdev.c             | 462 +++++++++++++++++++++++++++++++
>   drivers/vhost/vhost.c            |  39 ++-
>   drivers/vhost/vhost.h            |   6 +
>   include/linux/vfio.h             |  11 +
>   include/uapi/linux/vhost.h       |  10 +
>   include/uapi/linux/vhost_types.h |   5 +
>   10 files changed, 573 insertions(+), 7 deletions(-)
>   create mode 100644 drivers/vhost/mdev.c
>