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

[Michael S. Tsirkin]

On Fri, Sep 27, 2019 at 04:47:43PM +0800, Jason Wang wrote:
> 
> On 2019/9/27 äå12:54, Tiwei Bie wrote:
> > On Fri, Sep 27, 2019 at 11:46:06AM +0800, Jason Wang wrote:
> > > On 2019/9/26 äå12:54, Tiwei Bie wrote:
> > > > +
> > > > +static long vhost_mdev_start(struct vhost_mdev *m)
> > > > +{
> > > > +	struct mdev_device *mdev = m->mdev;
> > > > +	const struct virtio_mdev_device_ops *ops = mdev_get_dev_ops(mdev);
> > > > +	struct virtio_mdev_callback cb;
> > > > +	struct vhost_virtqueue *vq;
> > > > +	int idx;
> > > > +
> > > > +	ops->set_features(mdev, m->acked_features);
> > > > +
> > > > +	mdev_add_status(mdev, VIRTIO_CONFIG_S_FEATURES_OK);
> > > > +	if (!(mdev_get_status(mdev) & VIRTIO_CONFIG_S_FEATURES_OK))
> > > > +		goto reset;
> > > > +
> > > > +	for (idx = 0; idx < m->nvqs; idx++) {
> > > > +		vq = &m->vqs[idx];
> > > > +
> > > > +		if (!vq->desc || !vq->avail || !vq->used)
> > > > +			break;
> > > > +
> > > > +		if (ops->set_vq_state(mdev, idx, vq->last_avail_idx))
> > > > +			goto reset;
> > > If we do set_vq_state() in SET_VRING_BASE, we won't need this step here.
> > Yeah, I plan to do it in the next version.
> > 
> > > > +
> > > > +		/*
> > > > +		 * In vhost-mdev, userspace should pass ring addresses
> > > > +		 * in guest physical addresses when IOMMU is disabled or
> > > > +		 * IOVAs when IOMMU is enabled.
> > > > +		 */
> > > A question here, consider we're using noiommu mode. If guest physical
> > > address is passed here, how can a device use that?
> > > 
> > > I believe you meant "host physical address" here? And it also have the
> > > implication that the HPA should be continuous (e.g using hugetlbfs).
> > The comment is talking about the virtual IOMMU (i.e. iotlb in vhost).
> > It should be rephrased to cover the noiommu case as well. Thanks for
> > spotting this.
> > 
> > 
> > > > +
> > > > +	switch (cmd) {
> > > > +	case VHOST_MDEV_SET_STATE:
> > > > +		r = vhost_set_state(m, argp);
> > > > +		break;
> > > > +	case VHOST_GET_FEATURES:
> > > > +		r = vhost_get_features(m, argp);
> > > > +		break;
> > > > +	case VHOST_SET_FEATURES:
> > > > +		r = vhost_set_features(m, argp);
> > > > +		break;
> > > > +	case VHOST_GET_VRING_BASE:
> > > > +		r = vhost_get_vring_base(m, argp);
> > > > +		break;
> > > Does it mean the SET_VRING_BASE may only take affect after
> > > VHOST_MEV_SET_STATE?
> > Yeah, in this version, SET_VRING_BASE won't set the base to the
> > device directly. But I plan to not delay this anymore in the next
> > version to support the SET_STATUS.
> > 
> > > > +	default:
> > > > +		r = vhost_dev_ioctl(&m->dev, cmd, argp);
> > > > +		if (r == -ENOIOCTLCMD)
> > > > +			r = vhost_vring_ioctl(&m->dev, cmd, argp);
> > > > +	}
> > > > +
> > > > +	mutex_unlock(&m->mutex);
> > > > +	return r;
> > > > +}
> > > > +
> > > > +static const struct vfio_device_ops vfio_vhost_mdev_dev_ops = {
> > > > +	.name		= "vfio-vhost-mdev",
> > > > +	.open		= vhost_mdev_open,
> > > > +	.release	= vhost_mdev_release,
> > > > +	.ioctl		= vhost_mdev_unlocked_ioctl,
> > > > +};
> > > > +
> > > > +static int vhost_mdev_probe(struct device *dev)
> > > > +{
> > > > +	struct mdev_device *mdev = mdev_from_dev(dev);
> > > > +	const struct virtio_mdev_device_ops *ops = mdev_get_dev_ops(mdev);
> > > > +	struct vhost_mdev *m;
> > > > +	int nvqs, r;
> > > > +
> > > > +	m = kzalloc(sizeof(*m), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
> > > > +	if (!m)
> > > > +		return -ENOMEM;
> > > > +
> > > > +	mutex_init(&m->mutex);
> > > > +
> > > > +	nvqs = ops->get_queue_max(mdev);
> > > > +	m->nvqs = nvqs;
> > > The name could be confusing, get_queue_max() is to get the maximum number of
> > > entries for a virtqueue supported by this device.
> > OK. It might be better to rename it to something like:
> > 
> > 	get_vq_num_max()
> > 
> > which is more consistent with the set_vq_num().
> > 
> > > It looks to me that we need another API to query the maximum number of
> > > virtqueues supported by the device.
> > Yeah.
> > 
> > Thanks,
> > Tiwei
> 
> 
> One problem here:
> 
> Consider if we want to support multiqueue, how did userspace know about
> this?

There's a feature bit for this, isn't there?

> Note this information could be fetched from get_config() via a device
> specific way, do we want ioctl for accessing that area?
> 
> Thanks