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

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

On Wed, Oct 23, 2019 at 01:46:23PM +0800, Jason Wang wrote:
> On 2019/10/23 上午11:02, Tiwei Bie wrote:
> > On Tue, Oct 22, 2019 at 09:30:16PM +0800, Jason Wang wrote:
> > > On 2019/10/22 下午5:52, Tiwei Bie wrote:
> > > > This patch introduces a mdev based hardware vhost backend.
> > > > This backend is built on top of the same abstraction used
> > > > in virtio-mdev and provides a generic vhost interface for
> > > > userspace to accelerate the virtio devices in guest.
> > > > 
> > > > This backend is implemented as a mdev device driver on top
> > > > of the same mdev device ops used in virtio-mdev but using
> > > > a different mdev class id, and it will register the device
> > > > as a VFIO device for userspace to use. Userspace can setup
> > > > the IOMMU with the existing VFIO container/group APIs and
> > > > then get the device fd with the device name. After getting
> > > > the device fd of this device, userspace can use vhost ioctls
> > > > to setup the backend.
> > > > 
> > > > Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> > > > ---
> > > > This patch depends on below series:
> > > > https://lkml.org/lkml/2019/10/17/286
> > > > 
> > > > v1 -> v2:
> > > > - Replace _SET_STATE with _SET_STATUS (MST);
> > > > - Check status bits at each step (MST);
> > > > - Report the max ring size and max number of queues (MST);
> > > > - Add missing MODULE_DEVICE_TABLE (Jason);
> > > > - Only support the network backend w/o multiqueue for now;
> > > 
> > > Any idea on how to extend it to support devices other than net? I think we
> > > want a generic API or an API that could be made generic in the future.
> > > 
> > > Do we want to e.g having a generic vhost mdev for all kinds of devices or
> > > introducing e.g vhost-net-mdev and vhost-scsi-mdev?
> > One possible way is to do what vhost-user does. I.e. Apart from
> > the generic ring, features, ... related ioctls, we also introduce
> > device specific ioctls when we need them. As vhost-mdev just needs
> > to forward configs between parent and userspace and even won't
> > cache any info when possible,
> 
> 
> So it looks to me this is only possible if we expose e.g set_config and
> get_config to userspace.

The set_config and get_config interface isn't really everything
of device specific settings. We also have ctrlq in virtio-net.

> 
> 
> > I think it might be better to do
> > this in one generic vhost-mdev module.
> 
> 
> Looking at definitions of VhostUserRequest in qemu, it mixed generic API
> with device specific API. If we want go this ways (a generic vhost-mdev),
> more questions needs to be answered:
> 
> 1) How could userspace know which type of vhost it would use? Do we need to
> expose virtio subsystem device in for userspace this case?
> 
> 2) That generic vhost-mdev module still need to filter out unsupported
> ioctls for a specific type. E.g if it probes a net device, it should refuse
> API for other type. This in fact a vhost-mdev-net but just not modularize it
> on top of vhost-mdev.
> 
> 
> > 
> > > 
> > > > - Some minor fixes and improvements;
> > > > - Rebase on top of virtio-mdev series v4;
[...]
> > > > +
> > > > +static long vhost_mdev_get_features(struct vhost_mdev *m, u64 __user *featurep)
> > > > +{
> > > > +	if (copy_to_user(featurep, &m->features, sizeof(m->features)))
> > > > +		return -EFAULT;
> > > 
> > > As discussed in previous version do we need to filter out MQ feature here?
> > I think it's more straightforward to let the parent drivers to
> > filter out the unsupported features. Otherwise it would be tricky
> > when we want to add more features in vhost-mdev module,
> 
> 
> It's as simple as remove the feature from blacklist?

It's not really that easy. It may break the old drivers.

> 
> 
> > i.e. if
> > the parent drivers may expose unsupported features and relay on
> > vhost-mdev to filter them out, these features will be exposed
> > to userspace automatically when they are enabled in vhost-mdev
> > in the future.
> 
> 
> The issue is, it's only that vhost-mdev knows its own limitation. E.g in
> this patch, vhost-mdev only implements a subset of transport API, but parent
> doesn't know about that.
> 
> Still MQ as an example, there's no way (or no need) for parent to know that
> vhost-mdev does not support MQ.

The mdev is a MDEV_CLASS_ID_VHOST mdev device. When the parent
is being developed, it should know the currently supported features
of vhost-mdev.

> And this allows old kenrel to work with new
> parent drivers.

The new drivers should provide things like VIRTIO_MDEV_F_VERSION_1
to be compatible with the old kernels. When VIRTIO_MDEV_F_VERSION_1
is provided/negotiated, the behaviours should be consistent.

> 
> So basically we have three choices here:
> 
> 1) Implement what vhost-user did and implement a generic vhost-mdev (but may
> still have lots of device specific code). To support advanced feature which
> requires the access to config, still lots of API that needs to be added.
> 
> 2) Implement what vhost-kernel did, have a generic vhost-mdev driver and a
> vhost bus on top for match a device specific API e.g vhost-mdev-net. We
> still have device specific API but limit them only to device specific
> module. Still require new ioctls for advanced feature like MQ.
> 
> 3) Simply expose all virtio-mdev transport to userspace.

Currently, virtio-mdev transport is a set of function callbacks
defined in kernel. How to simply expose virtio-mdev transport to
userspace?


> A generic module
> without any type specific code (like virtio-mdev). No need dedicated API for
> e.g MQ. But then the API will look much different than current vhost did.
> 
> Consider the limitation of 1) I tend to choose 2 or 3. What's you opinion?
> 
>