Re: [PATCH vhost 14/16] vdpa/mlx5: Enable hw support for vq descriptor mapping

[Eugenio Perez Martin]

On Thu, Oct 5, 2023 at 2:16 PM Dragos Tatulea <dtatulea@xxxxxxxxxx> wrote:
>
> On Thu, 2023-10-05 at 11:42 +0200, Eugenio Perez Martin wrote:
> > On Thu, Sep 28, 2023 at 6:50 PM Dragos Tatulea <dtatulea@xxxxxxxxxx> wrote:
> > >
> > > Vq descriptor mappings are supported in hardware by filling in an
> > > additional mkey which contains the descriptor mappings to the hw vq.
> > >
> > > A previous patch in this series added support for hw mkey (mr) creation
> > > for ASID 1.
> > >
> > > This patch fills in both the vq data and vq descriptor mkeys based on
> > > group ASID mapping.
> > >
> > > The feature is signaled to the vdpa core through the presence of the
> > > .get_vq_desc_group op.
> > >
> > > Signed-off-by: Dragos Tatulea <dtatulea@xxxxxxxxxx>
> > > ---
> > >  drivers/vdpa/mlx5/net/mlx5_vnet.c  | 26 ++++++++++++++++++++++++--
> > >  include/linux/mlx5/mlx5_ifc_vdpa.h |  7 ++++++-
> > >  2 files changed, 30 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > index 25bd2c324f5b..46441e41892c 100644
> > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > > @@ -823,6 +823,7 @@ static int create_virtqueue(struct mlx5_vdpa_net *ndev,
> > > struct mlx5_vdpa_virtque
> > >         u32 out[MLX5_ST_SZ_DW(create_virtio_net_q_out)] = {};
> > >         struct mlx5_vdpa_dev *mvdev = &ndev->mvdev;
> > >         struct mlx5_vdpa_mr *vq_mr;
> > > +       struct mlx5_vdpa_mr *vq_desc_mr;
> > >         void *obj_context;
> > >         u16 mlx_features;
> > >         void *cmd_hdr;
> > > @@ -878,6 +879,11 @@ static int create_virtqueue(struct mlx5_vdpa_net *ndev,
> > > struct mlx5_vdpa_virtque
> > >         vq_mr = mvdev->mr[mvdev->group2asid[MLX5_VDPA_DATAVQ_GROUP]];
> > >         if (vq_mr)
> > >                 MLX5_SET(virtio_q, vq_ctx, virtio_q_mkey, vq_mr->mkey);
> > > +
> > > +       vq_desc_mr = mvdev->mr[mvdev-
> > > >group2asid[MLX5_VDPA_DATAVQ_DESC_GROUP]];
> > > +       if (vq_desc_mr)
> > > +               MLX5_SET(virtio_q, vq_ctx, desc_group_mkey, vq_desc_mr-
> > > >mkey);
> > > +
> > >         MLX5_SET(virtio_q, vq_ctx, umem_1_id, mvq->umem1.id);
> > >         MLX5_SET(virtio_q, vq_ctx, umem_1_size, mvq->umem1.size);
> > >         MLX5_SET(virtio_q, vq_ctx, umem_2_id, mvq->umem2.id);
> > > @@ -2265,6 +2271,16 @@ static u32 mlx5_vdpa_get_vq_group(struct vdpa_device
> > > *vdev, u16 idx)
> > >         return MLX5_VDPA_DATAVQ_GROUP;
> > >  }
> > >
> > > +static u32 mlx5_vdpa_get_vq_desc_group(struct vdpa_device *vdev, u16 idx)
> > > +{
> > > +       struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> > > +
> > > +       if (is_ctrl_vq_idx(mvdev, idx))
> > > +               return MLX5_VDPA_CVQ_GROUP;
> > > +
> > > +       return MLX5_VDPA_DATAVQ_DESC_GROUP;
> > > +}
> > > +
> > >  static u64 mlx_to_vritio_features(u16 dev_features)
> > >  {
> > >         u64 result = 0;
> > > @@ -3139,7 +3155,7 @@ static int mlx5_set_group_asid(struct vdpa_device
> > > *vdev, u32 group,
> > >  {
> > >         struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> > >
> > > -       if (group >= MLX5_VDPA_NUMVQ_GROUPS)
> > > +       if (group >= MLX5_VDPA_NUMVQ_GROUPS || asid >= MLX5_VDPA_NUM_AS)
> >
> > Nit: the check for asid >= MLX5_VDPA_NUM_AS is redundant, as it will
> > be already checked by VHOST_VDPA_SET_GROUP_ASID handler in
> > drivers/vhost/vdpa.c:vhost_vdpa_vring_ioctl. Not a big deal.
> Ack.
>
> >
> > >                 return -EINVAL;
> > >
> > >         mvdev->group2asid[group] = asid;
> > > @@ -3160,6 +3176,7 @@ static const struct vdpa_config_ops mlx5_vdpa_ops = {
> > >         .get_vq_irq = mlx5_get_vq_irq,
> > >         .get_vq_align = mlx5_vdpa_get_vq_align,
> > >         .get_vq_group = mlx5_vdpa_get_vq_group,
> > > +       .get_vq_desc_group = mlx5_vdpa_get_vq_desc_group, /* Op disabled if
> > > not supported. */
> > >         .get_device_features = mlx5_vdpa_get_device_features,
> > >         .set_driver_features = mlx5_vdpa_set_driver_features,
> > >         .get_driver_features = mlx5_vdpa_get_driver_features,
> > > @@ -3258,6 +3275,7 @@ struct mlx5_vdpa_mgmtdev {
> > >         struct vdpa_mgmt_dev mgtdev;
> > >         struct mlx5_adev *madev;
> > >         struct mlx5_vdpa_net *ndev;
> > > +       struct vdpa_config_ops vdpa_ops;
> > >  };
> > >
> > >  static int config_func_mtu(struct mlx5_core_dev *mdev, u16 mtu)
> > > @@ -3371,7 +3389,7 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev
> > > *v_mdev, const char *name,
> > >                 max_vqs = 2;
> > >         }
> > >
> > > -       ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev-
> > > >device, &mlx5_vdpa_ops,
> > > +       ndev = vdpa_alloc_device(struct mlx5_vdpa_net, mvdev.vdev, mdev-
> > > >device, &mgtdev->vdpa_ops,
> > >                                  MLX5_VDPA_NUMVQ_GROUPS, MLX5_VDPA_NUM_AS,
> > > name, false);
> > >         if (IS_ERR(ndev))
> > >                 return PTR_ERR(ndev);
> > > @@ -3546,6 +3564,10 @@ static int mlx5v_probe(struct auxiliary_device *adev,
> > >                 MLX5_CAP_DEV_VDPA_EMULATION(mdev, max_num_virtio_queues) +
> > > 1;
> > >         mgtdev->mgtdev.supported_features = get_supported_features(mdev);
> > >         mgtdev->madev = madev;
> > > +       mgtdev->vdpa_ops = mlx5_vdpa_ops;
> > > +
> > > +       if (!MLX5_CAP_DEV_VDPA_EMULATION(mdev, desc_group_mkey_supported))
> > > +               mgtdev->vdpa_ops.get_vq_desc_group = NULL;
> >
> > I think this is better handled by splitting mlx5_vdpa_ops in two: One
> > with get_vq_desc_group and other without it. You can see an example of
> > this in the simulator, where one version supports .dma_map incremental
> > updating with .dma_map and the other supports .set_map. Otherwise,
> > this can get messy if more members opt-out or opt-in.
> >
> I implemented it this way because the upcoming resumable vq support will also
> need to selectively implement .resume if the hw capability is there. That would
> result in needing 4 different ops for all combinations. The other option would
> be to force these two ops together (.get_vq_desc_group and .resume). But I would
> prefer to not do that.
>

That's a good point. As more features are optional per device, maybe
this approach is better.

I'm not sure what Jason prefers, but I think it would be easy to
change it on top.

Thanks!

> > But I'm ok with this too, so whatever version you choose:
> >
> > Acked-by: Eugenio Pérez <eperezma@xxxxxxxxxx>
> >
> > >
> > >         err = vdpa_mgmtdev_register(&mgtdev->mgtdev);
> > >         if (err)
> > > diff --git a/include/linux/mlx5/mlx5_ifc_vdpa.h
> > > b/include/linux/mlx5/mlx5_ifc_vdpa.h
> > > index 9becdc3fa503..b86d51a855f6 100644
> > > --- a/include/linux/mlx5/mlx5_ifc_vdpa.h
> > > +++ b/include/linux/mlx5/mlx5_ifc_vdpa.h
> > > @@ -74,7 +74,11 @@ struct mlx5_ifc_virtio_q_bits {
> > >         u8    reserved_at_320[0x8];
> > >         u8    pd[0x18];
> > >
> > > -       u8    reserved_at_340[0xc0];
> > > +       u8    reserved_at_340[0x20];
> > > +
> > > +       u8    desc_group_mkey[0x20];
> > > +
> > > +       u8    reserved_at_380[0x80];
> > >  };
> > >
> > >  struct mlx5_ifc_virtio_net_q_object_bits {
> > > @@ -141,6 +145,7 @@ enum {
> > >         MLX5_VIRTQ_MODIFY_MASK_STATE                    = (u64)1 << 0,
> > >         MLX5_VIRTQ_MODIFY_MASK_DIRTY_BITMAP_PARAMS      = (u64)1 << 3,
> > >         MLX5_VIRTQ_MODIFY_MASK_DIRTY_BITMAP_DUMP_ENABLE = (u64)1 << 4,
> > > +       MLX5_VIRTQ_MODIFY_MASK_DESC_GROUP_MKEY          = (u64)1 << 14,
> > >  };
> > >
> > >  enum {
> > > --
> > > 2.41.0
> > >
> >
>