Re: [PATCH vhost next 10/10] vdpa/mlx5: Add VDPA driver for supported mlx5 devices

[Eli Cohen]

On Thu, Jul 16, 2020 at 05:14:32PM +0800, Jason Wang wrote:
> 
> >+static void suspend_vqs(struct mlx5_vdpa_net *ndev)
> >+{
> >+	int i;
> >+
> >+	for (i = 0; i < MLX5_MAX_SUPPORTED_VQS; i++)
> >+		suspend_vq(ndev, &ndev->vqs[i]);
> 
> 
> In teardown_virtqueues() it has a check of mvq->num_ent, any reason
> not doing it here?
> 

Looks like checking intialized is enough. Will fix this.
 
> >+
> >+static void mlx5_vdpa_set_vq_ready(struct vdpa_device *vdev, u16 idx, bool ready)
> >+{
> >+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> >+	struct mlx5_vdpa_virtqueue *mvq = &ndev->vqs[idx];
> >+	int err;
> >+
> >+	if (!mvq->ready && ready && mvq->fw_state != MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY) {
> >+		err = modify_virtqueue(ndev, mvq, MLX5_VIRTIO_NET_Q_OBJECT_STATE_RDY);
> >+		if (err) {
> >+			mlx5_vdpa_warn(mvdev, "failed to modify virtqueue(%d)\n", err);
> >+			return;
> >+		}
> >+	}
> 
> 
> I wonder what's the reason of changing vq state on the hardware
> here. I think we can defer it to set_status().
> 

I can defer this to set status.

I just wonder if it is possible that the core vdpa driver may call this
function with ready equals false and after some time call it with ready
equals true.


> (Anyhow we don't sync vq address in set_vq_address()).
> 
> 
> >+static u64 mlx5_vdpa_get_features(struct vdpa_device *vdev)
> >+{
> >+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> >+	u16 dev_features;
> >+
> >+	dev_features = MLX5_CAP_DEV_VDPA_EMULATION(mvdev->mdev, device_features_bits_mask);
> >+	ndev->mvdev.mlx_features = mlx_to_vritio_features(dev_features);
> >+	if (MLX5_CAP_DEV_VDPA_EMULATION(mvdev->mdev, virtio_version_1_0))
> >+		ndev->mvdev.mlx_features |= BIT(VIRTIO_F_VERSION_1);
> 
> 
> This is interesting. This suggests !VIRTIO_F_VERSION_1 &&
> VIRTIO_F_IOMMU_PLATFORM is valid. But virito spec doesn't allow such
> configuration.

Will fix
> 
> So I think you need either:
> 
> 1) Fail vDPA device probe when VERSION_1 is not supported
> 2) clear IOMMU_PLATFORM if VERSION_1 is not negotiated
> 
> For 2) I guess it can't work, according to the spec, without
> IOMMU_PLATFORM, device need to use PA to access the memory
> 
> 
> >+	ndev->mvdev.mlx_features |= BIT(VIRTIO_F_ANY_LAYOUT);
I think this should be removed too

> >+	ndev->mvdev.mlx_features |= BIT(VIRTIO_F_IOMMU_PLATFORM);
> >+	if (mlx5_vdpa_max_qps(ndev->mvdev.max_vqs) > 1)
> >+		ndev->mvdev.mlx_features |= BIT(VIRTIO_NET_F_MQ);
Also this, since multqueue requires configuration vq which is not
supported in this version.

> >+
> >+	print_features(mvdev, ndev->mvdev.mlx_features, false);
> >+	return ndev->mvdev.mlx_features;
> >+}
> >+
> >+static int verify_min_features(struct mlx5_vdpa_dev *mvdev, u64 features)
> >+{
> >+	/* FIXME: qemu currently does not set all the feaures due to a bug.
> >+	 * Add checks when this is fixed.
> >+	 */
> 
> 
> I think we should add the check now then qemu can get notified. (E.g
> IOMMU_PLATFORM)

Will do.
> 
> 
> >+}
> >+
> >+#define MLX5_VDPA_MAX_VQ_ENTRIES 256
> 
> 
> Is this a hardware limitation, qemu can support up to 1K which the
> requirement of some NFV cases.
> 

Theoretically the device is limit is much higher. In the near future we
will have a device capability for this. I wanted to stay on the safe side
with this but I can change this if you think it's necessary.
> 
> >+
> >+static void mlx5_vdpa_get_config(struct vdpa_device *vdev, unsigned int offset, void *buf,
> >+				 unsigned int len)
> >+{
> >+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> >+
> >+	if (offset + len < sizeof(struct virtio_net_config))
> >+		memcpy(buf, &ndev->config + offset, len);
> 
> 
> Note that guest expect LE, what's the endian for ndev->config?
This is struct virtio_net_config from include/uapi/linux/virtio_net.h.
Looking there I see it has mixed endianess. I currently don't touch it
but if I do, I should follow endianess guidance per the struct
definition. So I don't think I should care about endianess when copying.

> 
> 
> >+}
> >+
> >+static void mlx5_vdpa_set_config(struct vdpa_device *vdev, unsigned int offset, const void *buf,
> >+				 unsigned int len)
> >+{
> >+	/* not supported */
> >+}
> >+
> >+static u32 mlx5_vdpa_get_generation(struct vdpa_device *vdev)
> >+{
> >+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >+
> >+	return mvdev->generation;
> >+}
> >+
> >+static int mlx5_vdpa_set_map(struct vdpa_device *vdev, struct vhost_iotlb *iotlb)
> >+{
> >+	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> >+	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> >+	bool change_map;
> >+	int err;
> >+
> >+	err = mlx5_vdpa_handle_set_map(mvdev, iotlb, &change_map);
> >+	if (err) {
> >+		mlx5_vdpa_warn(mvdev, "set map failed(%d)\n", err);
> >+		return err;
> >+	}
> >+
> >+	if (change_map)
> >+		return mlx5_vdpa_change_map(ndev, iotlb);
> 
> 
> Any reason for not doing this inside mlx5_handle_set_map()?
> 

All memory registration related operations are done inside mr.c in the
common code directory. But change map involves operations on other
objects managed in this file.

> >+
> >+void mlx5_vdpa_remove_dev(struct mlx5_vdpa_dev *mvdev)
> >+{
> >+	struct mlx5_vdpa_net *ndev;
> >+
> >+	mvdev->status = 0;
> 
> 
> This is probably unnecessary.
> 

Right, will remove.