Re: [PATCH v2] vhost: introduce mdev based hardware backend

[Tiwei Bie <tiwei.bie@intel.com>]

On Tue, Oct 29, 2019 at 06:48:27PM +0800, Jason Wang wrote:
> On 2019/10/29 下午5:57, Tiwei Bie wrote:
> > On Mon, Oct 28, 2019 at 11:50:49AM +0800, Jason Wang wrote:
> >> On 2019/10/28 上午9:58, Tiwei Bie wrote:
> >>> On Fri, Oct 25, 2019 at 08:16:26AM -0400, Michael S. Tsirkin wrote:
> >>>> On Fri, Oct 25, 2019 at 05:54:55PM +0800, Jason Wang wrote:
> >>>>> On 2019/10/24 下午6:42, Jason Wang wrote:
> >>>>>> Yes.
> >>>>>>
> >>>>>>
> >>>>>>>    And we should try to avoid
> >>>>>>> putting ctrl vq and Rx/Tx vqs in the same DMA space to prevent
> >>>>>>> guests having the chance to bypass the host (e.g. QEMU) to
> >>>>>>> setup the backend accelerator directly.
> >>>>>> That's really good point.  So when "vhost" type is created, parent
> >>>>>> should assume addr of ctrl_vq is hva.
> >>>>>>
> >>>>>> Thanks
> >>>>> This works for vhost but not virtio since there's no way for virtio kernel
> >>>>> driver to differ ctrl_vq with the rest when doing DMA map. One possible
> >>>>> solution is to provide DMA domain isolation between virtqueues. Then ctrl vq
> >>>>> can use its dedicated DMA domain for the work.
> >>> It might not be a bad idea to let the parent drivers distinguish
> >>> between virtio-mdev mdevs and vhost-mdev mdevs in ctrl-vq handling
> >>> by mdev's class id.
> >> Yes, that should work, I have something probable better, see below.
> >>
> >>
> >>>>> Anyway, this could be done in the future. We can have a version first that
> >>>>> doesn't support ctrl_vq.
> >>> +1, thanks
> >>>
> >>>>> Thanks
> >>>> Well no ctrl_vq implies either no offloads, or no XDP (since XDP needs
> >>>> to disable offloads dynamically).
> >>>>
> >>>>          if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
> >>>>              && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
> >>>>                  virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
> >>>>                  virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
> >>>>                  virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
> >>>>                  virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
> >>>>                  NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
> >>>>                  return -EOPNOTSUPP;
> >>>>          }
> >>>>
> >>>> neither is very attractive.
> >>>>
> >>>> So yes ok just for development but we do need to figure out how it will
> >>>> work down the road in production.
> >>> Totally agree.
> >>>
> >>>> So really this specific virtio net device does not support control vq,
> >>>> instead it supports a different transport specific way to send commands
> >>>> to device.
> >>>>
> >>>> Some kind of extension to the transport? Ideas?
> >> So it's basically an issue of isolating DMA domains. Maybe we can start with
> >> transport API for querying per vq DMA domain/ASID?
> >>
> >> - for vhost-mdev, userspace can query the DMA domain for each specific
> >> virtqueue. For control vq, mdev can return id for software domain, for the
> >> rest mdev will return id of VFIO domain. Then userspace know that it should
> >> use different API for preparing the virtqueue, e.g for vq other than control
> >> vq, it should use VFIO DMA API. The control vq it should use hva instead.
> >>
> >> - for virito-mdev, we can introduce per-vq DMA device, and route DMA mapping
> >> request for control vq back to mdev instead of the hardware. (We can wrap
> >> them into library or helpers to ease the development of vendor physical
> >> drivers).
> > Thanks for this proposal! I'm thinking about it these days.
> > I think it might be too complicated. I'm wondering whether we
> > can have something simpler. I will post a RFC patch to show
> > my idea today.
> 
> 
> Thanks, will check.
> 
> Btw, for virtio-mdev, the change should be very minimal, will post an
> RFC as well. For vhost-mdev, it could be just a helper to return an ID
> for DMA domain like ID_VFIO or ID_HVA.
> 
> Or a more straightforward way is to force queues like control vq to use PA.

Will check. Thanks!

> 
> 
> >
> > Thanks,
> > Tiwei
> >
>