Re: [PATCH RFC 2/3] vdpa/mlx5: Support different address spaces for control and data

[Eugenio Perez Martin]

On Thu, Jun 16, 2022 at 3:27 PM Eli Cohen <elic@xxxxxxxxxx> wrote:
>
> Partition virtqueues to two different address spaces: oce for control
> virtqueue which is implemented in software, and one for data virtqueus.
>
> Signed-off-by: Eli Cohen <elic@xxxxxxxxxx>
> ---
>  drivers/vdpa/mlx5/core/mlx5_vdpa.h |  11 ++++
>  drivers/vdpa/mlx5/net/mlx5_vnet.c  | 101 +++++++++++++++++++++++++----
>  2 files changed, 101 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> index 44104093163b..6af9fdbb86b7 100644
> --- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> +++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> @@ -70,6 +70,16 @@ struct mlx5_vdpa_wq_ent {
>         struct mlx5_vdpa_dev *mvdev;
>  };
>
> +enum {
> +       MLX5_VDPA_DATAVQ_GROUP,
> +       MLX5_VDPA_CVQ_GROUP,
> +       MLX5_VDPA_NUMVQ_GROUPS
> +};
> +
> +enum {
> +       MLX5_VDPA_NUM_AS = MLX5_VDPA_NUMVQ_GROUPS
> +};
> +
>  struct mlx5_vdpa_dev {
>         struct vdpa_device vdev;
>         struct mlx5_core_dev *mdev;
> @@ -85,6 +95,7 @@ struct mlx5_vdpa_dev {
>         struct mlx5_vdpa_mr mr;
>         struct mlx5_control_vq cvq;
>         struct workqueue_struct *wq;
> +       unsigned int group2asid[MLX5_VDPA_NUMVQ_GROUPS];
>  };
>
>  int mlx5_vdpa_alloc_pd(struct mlx5_vdpa_dev *dev, u32 *pdn, u16 uid);
> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> index ea4bc8a0cd25..34bd81cb697c 100644
> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> @@ -2125,9 +2125,14 @@ static u32 mlx5_vdpa_get_vq_align(struct vdpa_device *vdev)
>         return PAGE_SIZE;
>  }
>
> -static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdpa, u16 idx)
> +static u32 mlx5_vdpa_get_vq_group(struct vdpa_device *vdev, u16 idx)
>  {
> -       return 0;
> +       struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> +
> +       if (is_ctrl_vq_idx(mvdev, idx))
> +               return MLX5_VDPA_CVQ_GROUP;
> +
> +       return MLX5_VDPA_DATAVQ_GROUP;
>  }
>
>  enum { MLX5_VIRTIO_NET_F_GUEST_CSUM = 1 << 9,
> @@ -2541,6 +2546,15 @@ static void mlx5_vdpa_set_status(struct vdpa_device *vdev, u8 status)
>         up_write(&ndev->reslock);
>  }
>
> +static void init_group_to_asid_map(struct mlx5_vdpa_dev *mvdev)
> +{
> +       int i;
> +
> +       /* default mapping all groups are mapped to asid 0 */
> +       for (i = 0; i < MLX5_VDPA_NUMVQ_GROUPS; i++)
> +               mvdev->group2asid[i] = 0;
> +}
> +
>  static int mlx5_vdpa_reset(struct vdpa_device *vdev)
>  {
>         struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> @@ -2559,7 +2573,9 @@ static int mlx5_vdpa_reset(struct vdpa_device *vdev)
>         ndev->mvdev.cvq.completed_desc = 0;
>         memset(ndev->event_cbs, 0, sizeof(*ndev->event_cbs) * (mvdev->max_vqs + 1));
>         ndev->mvdev.actual_features = 0;
> +       init_group_to_asid_map(mvdev);
>         ++mvdev->generation;
> +
>         if (MLX5_CAP_GEN(mvdev->mdev, umem_uid_0)) {
>                 if (mlx5_vdpa_create_mr(mvdev, NULL))
>                         mlx5_vdpa_warn(mvdev, "create MR failed\n");
> @@ -2597,26 +2613,76 @@ static u32 mlx5_vdpa_get_generation(struct vdpa_device *vdev)
>         return mvdev->generation;
>  }
>
> -static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
> -                            struct vhost_iotlb *iotlb)
> +static u32 get_group(struct mlx5_vdpa_dev *mvdev, unsigned int asid)
> +{
> +       u32 group;
> +
> +       for (group = 0; group < MLX5_VDPA_NUMVQ_GROUPS; group++) {
> +               if (mvdev->group2asid[group] == asid)
> +                       return group;
> +       }
> +       return -EINVAL;
> +}
> +
> +static int set_map_control(struct mlx5_vdpa_dev *mvdev, struct vhost_iotlb *iotlb)
> +{
> +       u64 start = 0ULL, last = 0ULL - 1;
> +       struct vhost_iotlb_map *map;
> +       int err = 0;
> +
> +       spin_lock(&mvdev->cvq.iommu_lock);
> +       vhost_iotlb_reset(mvdev->cvq.iotlb);
> +
> +       for (map = vhost_iotlb_itree_first(iotlb, start, last); map;
> +            map = vhost_iotlb_itree_next(map, start, last)) {
> +               err = vhost_iotlb_add_range(mvdev->cvq.iotlb, map->start,
> +                                           map->last, map->addr, map->perm);
> +               if (err)
> +                       goto out;
> +       }
> +
> +out:
> +       spin_unlock(&mvdev->cvq.iommu_lock);
> +       return err;
> +}
> +
> +static int set_map_data(struct mlx5_vdpa_dev *mvdev, 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;
>
> -       down_write(&ndev->reslock);
> -
>         err = mlx5_vdpa_handle_set_map(mvdev, iotlb, &change_map);
>         if (err) {
>                 mlx5_vdpa_warn(mvdev, "set map failed(%d)\n", err);
> -               goto err;
> +               return err;
>         }
>
>         if (change_map)
>                 err = mlx5_vdpa_change_map(mvdev, iotlb);
>
> -err:
> +       return err;
> +}
> +
> +static int mlx5_vdpa_set_map(struct vdpa_device *vdev, unsigned int asid,
> +                            struct vhost_iotlb *iotlb)
> +{
> +       struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> +       struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
> +       u32 group;
> +       int err;
> +
> +       down_write(&ndev->reslock);
> +       group = get_group(mvdev, asid);
> +       switch (group) {
> +       case MLX5_VDPA_DATAVQ_GROUP:
> +               err = set_map_data(mvdev, iotlb);
> +               break;
> +       case MLX5_VDPA_CVQ_GROUP:
> +               err = set_map_control(mvdev, iotlb);
> +               break;
> +       default:
> +               err = -EINVAL;
> +       }

This shouldn't be a switch, but to check the asid assigned to the
different vqs individually.

In the current qemu version with no ASID support, all vq groups (data
and cvq) are assigned to asid 0 at the device reset. In this case,
emulated cvq also needs to receive the mappings, because guest's CVQ
commands will go from the guest's ASID directly.

Thanks!

>         up_write(&ndev->reslock);
>         return err;
>  }
> @@ -2796,6 +2862,18 @@ static int mlx5_vdpa_suspend(struct vdpa_device *vdev, bool suspend)
>         return 0;
>  }
>
> +static int mlx5_set_group_asid(struct vdpa_device *vdev, u32 group,
> +                              unsigned int asid)
> +{
> +       struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> +
> +       if (group >= MLX5_VDPA_NUMVQ_GROUPS)
> +               return -EINVAL;
> +
> +       mvdev->group2asid[group] = asid;
> +       return 0;
> +}
> +
>  static const struct vdpa_config_ops mlx5_vdpa_ops = {
>         .set_vq_address = mlx5_vdpa_set_vq_address,
>         .set_vq_num = mlx5_vdpa_set_vq_num,
> @@ -2825,6 +2903,7 @@ static const struct vdpa_config_ops mlx5_vdpa_ops = {
>         .set_config = mlx5_vdpa_set_config,
>         .get_generation = mlx5_vdpa_get_generation,
>         .set_map = mlx5_vdpa_set_map,
> +       .set_group_asid = mlx5_set_group_asid,
>         .free = mlx5_vdpa_free,
>         .suspend = mlx5_vdpa_suspend,
>  };
> @@ -3047,7 +3126,7 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
>         }
>
>         ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev->device, &mlx5_vdpa_ops,
> -                                1, 1, name, false);
> +                                MLX5_VDPA_NUMVQ_GROUPS, MLX5_VDPA_NUM_AS, name, false);
>         if (IS_ERR(ndev))
>                 return PTR_ERR(ndev);
>
> --
> 2.35.1
>