Re: [PATCH vhost 16/16] vdpa/mlx5: Update cvq iotlb mapping on ASID change
From: Dragos Tatulea
Date: Thu Oct 05 2023 - 10:09:46 EST
On Thu, 2023-10-05 at 12:41 +0200, Eugenio Perez Martin wrote:
> On Thu, Sep 28, 2023 at 6:50 PM Dragos Tatulea <dtatulea@xxxxxxxxxx> wrote:
> >
> > For the following sequence:
> > - cvq group is in ASID 0
> > - .set_map(1, cvq_iotlb)
> > - .set_group_asid(cvq_group, 1)
> >
> > ... the cvq mapping from ASID 0 will be used. This is not always correct
> > behaviour.
> >
> > This patch adds support for the above mentioned flow by saving the iotlb
> > on each .set_map and updating the cvq iotlb with it on a cvq group change.
> >
> > Signed-off-by: Dragos Tatulea <dtatulea@xxxxxxxxxx>
> > ---
> > drivers/vdpa/mlx5/core/mlx5_vdpa.h | 2 ++
> > drivers/vdpa/mlx5/core/mr.c | 26 ++++++++++++++++++++++++++
> > drivers/vdpa/mlx5/net/mlx5_vnet.c | 9 ++++++++-
> > 3 files changed, 36 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> > b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> > index ae09296f4270..db988ced5a5d 100644
> > --- a/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> > +++ b/drivers/vdpa/mlx5/core/mlx5_vdpa.h
> > @@ -32,6 +32,8 @@ struct mlx5_vdpa_mr {
> > unsigned long num_directs;
> > unsigned long num_klms;
> >
> > + struct vhost_iotlb *iotlb;
> > +
> > bool user_mr;
> > };
> >
> > diff --git a/drivers/vdpa/mlx5/core/mr.c b/drivers/vdpa/mlx5/core/mr.c
> > index a4135c16b5bf..403c08271489 100644
> > --- a/drivers/vdpa/mlx5/core/mr.c
> > +++ b/drivers/vdpa/mlx5/core/mr.c
> > @@ -499,6 +499,8 @@ static void _mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev
> > *mvdev, struct mlx5_vdpa_
> > destroy_user_mr(mvdev, mr);
> > else
> > destroy_dma_mr(mvdev, mr);
> > +
> > + vhost_iotlb_free(mr->iotlb);
> > }
> >
> > void mlx5_vdpa_destroy_mr(struct mlx5_vdpa_dev *mvdev,
> > @@ -558,6 +560,30 @@ static int _mlx5_vdpa_create_mr(struct mlx5_vdpa_dev
> > *mvdev,
> > else
> > err = create_dma_mr(mvdev, mr);
> >
> > + if (err)
> > + return err;
> > +
> > + mr->iotlb = vhost_iotlb_alloc(0, 0);
> > + if (!mr->iotlb) {
> > + err = -ENOMEM;
> > + goto err_mr;
> > + }
> > +
> > + err = dup_iotlb(mr->iotlb, iotlb);
> > + if (err)
> > + goto err_iotlb;
> > +
> > + return 0;
> > +
> > +err_iotlb:
> > + vhost_iotlb_free(mr->iotlb);
> > +
> > +err_mr:
> > + if (iotlb)
> > + destroy_user_mr(mvdev, mr);
> > + else
> > + destroy_dma_mr(mvdev, mr);
> > +
> > return err;
> > }
> >
> > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > index 46441e41892c..fc5d6b989a5a 100644
> > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> > @@ -3154,12 +3154,19 @@ static int mlx5_set_group_asid(struct vdpa_device
> > *vdev, u32 group,
> > unsigned int asid)
> > {
> > struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
> > + int err = 0;
> >
> > if (group >= MLX5_VDPA_NUMVQ_GROUPS || asid >= MLX5_VDPA_NUM_AS)
> > return -EINVAL;
> >
> > mvdev->group2asid[group] = asid;
> > - return 0;
> > +
> > + mutex_lock(&mvdev->mr_mtx);
> > + if (group == MLX5_VDPA_CVQ_GROUP && mvdev->mr[asid])
> > + err = mlx5_vdpa_update_cvq_iotlb(mvdev, mvdev->mr[asid]-
> > >iotlb, asid);
>
> Do we need to protect here in case userspace sets the same ASID twice?
> mlx5_vdpa_update_cvq_iotlb shouldn't call dup_iotlb with the same src
> and dst.
Why would src and dst be identical? Doesn't dup_iotlb create a copy of src?
That being said, there should be a check in dup_iotlb for dst == src. Not sure
if this is an error or success though...
Thanks,
Dragos
-
> Apart from that:
>
> Acked-by: Eugenio Pérez <eperezma@xxxxxxxxxx>
>
> > + mutex_unlock(&mvdev->mr_mtx);
> > +
> > + return err;
> > }
> >
> > static const struct vdpa_config_ops mlx5_vdpa_ops = {
> > --
> > 2.41.0
> >
>