| Message ID | 20200908123346.GA169007@mtl-vdi-166.wap.labs.mlnx |
|---|---|
| State | Not Applicable |
| Delegated to: | David Miller |
| Headers | show |
| Series | [v2] vdpa/mlx5: Setup driver only if VIRTIO_CONFIG_S_DRIVER_OK | expand |
----- Original Message ----- > set_map() is used by mlx5 vdpa to create a memory region based on the > address map passed by the iotlb argument. If we get successive calls, we > will destroy the current memory region and build another one based on > the new address mapping. We also need to setup the hardware resources > since they depend on the memory region. > > If these calls happen before DRIVER_OK, It means that driver VQs may > also not been setup and we may not create them yet. In this case we want > to avoid setting up the other resources and defer this till we get > DRIVER OK. > > Fixes: 1a86b377aa21 ("vdpa/mlx5: Add VDPA driver for supported mlx5 devices") > Signed-off-by: Eli Cohen <elic@nvidia.com> > --- > V1->V2: Improve changelog description > > drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 +++ > 1 file changed, 3 insertions(+) > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c > b/drivers/vdpa/mlx5/net/mlx5_vnet.c > index 9df69d5efe8c..c89cd48a0aab 100644 > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c > @@ -1645,6 +1645,9 @@ static int mlx5_vdpa_change_map(struct mlx5_vdpa_net > *ndev, struct vhost_iotlb * > if (err) > goto err_mr; > > + if (!(ndev->mvdev.status & VIRTIO_CONFIG_S_DRIVER_OK)) > + return 0; > + Is there any reason that we still need to do vq suspending and saving before? Thanks > restore_channels_info(ndev); > err = setup_driver(ndev); > if (err) > -- > 2.26.0 > >
On Tue, Sep 08, 2020 at 10:08:39PM -0400, Jason Wang wrote: > > > ----- Original Message ----- > > set_map() is used by mlx5 vdpa to create a memory region based on the > > address map passed by the iotlb argument. If we get successive calls, we > > will destroy the current memory region and build another one based on > > the new address mapping. We also need to setup the hardware resources > > since they depend on the memory region. > > > > If these calls happen before DRIVER_OK, It means that driver VQs may > > also not been setup and we may not create them yet. In this case we want > > to avoid setting up the other resources and defer this till we get > > DRIVER OK. > > > > Fixes: 1a86b377aa21 ("vdpa/mlx5: Add VDPA driver for supported mlx5 devices") > > Signed-off-by: Eli Cohen <elic@nvidia.com> > > --- > > V1->V2: Improve changelog description > > > > drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 +++ > > 1 file changed, 3 insertions(+) > > > > diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > index 9df69d5efe8c..c89cd48a0aab 100644 > > --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c > > +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c > > @@ -1645,6 +1645,9 @@ static int mlx5_vdpa_change_map(struct mlx5_vdpa_net > > *ndev, struct vhost_iotlb * > > if (err) > > goto err_mr; > > > > + if (!(ndev->mvdev.status & VIRTIO_CONFIG_S_DRIVER_OK)) > > + return 0; > > + > > Is there any reason that we still need to do vq suspending and saving before? > Though suspend_vqs() and save_channels_info() will be called, they will not actually do any work because the mvq->initialized is false. Since we don't expect so many false map updates I think it makes sense to avoid logic around the calls suspend_vqs() and save_channels_info(). > Thanks > > > restore_channels_info(ndev); > > err = setup_driver(ndev); > > if (err) > > -- > > 2.26.0 > > > > >
On 2020/9/9 下午1:30, Eli Cohen wrote: > On Tue, Sep 08, 2020 at 10:08:39PM -0400, Jason Wang wrote: >> >> ----- Original Message ----- >>> set_map() is used by mlx5 vdpa to create a memory region based on the >>> address map passed by the iotlb argument. If we get successive calls, we >>> will destroy the current memory region and build another one based on >>> the new address mapping. We also need to setup the hardware resources >>> since they depend on the memory region. >>> >>> If these calls happen before DRIVER_OK, It means that driver VQs may >>> also not been setup and we may not create them yet. In this case we want >>> to avoid setting up the other resources and defer this till we get >>> DRIVER OK. >>> >>> Fixes: 1a86b377aa21 ("vdpa/mlx5: Add VDPA driver for supported mlx5 devices") >>> Signed-off-by: Eli Cohen <elic@nvidia.com> >>> --- >>> V1->V2: Improve changelog description >>> >>> drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 +++ >>> 1 file changed, 3 insertions(+) >>> >>> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c >>> b/drivers/vdpa/mlx5/net/mlx5_vnet.c >>> index 9df69d5efe8c..c89cd48a0aab 100644 >>> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c >>> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c >>> @@ -1645,6 +1645,9 @@ static int mlx5_vdpa_change_map(struct mlx5_vdpa_net >>> *ndev, struct vhost_iotlb * >>> if (err) >>> goto err_mr; >>> >>> + if (!(ndev->mvdev.status & VIRTIO_CONFIG_S_DRIVER_OK)) >>> + return 0; >>> + >> Is there any reason that we still need to do vq suspending and saving before? >> > Though suspend_vqs() and save_channels_info() will be called, they will > not actually do any work because the mvq->initialized is false. > > Since we don't expect so many false map updates I think it makes sense > to avoid logic around the calls suspend_vqs() and save_channels_info(). I see. Acked-by: Jason Wang <jasowang@redhat.com> Thanks > >> Thanks >> >>> restore_channels_info(ndev); >>> err = setup_driver(ndev); >>> if (err) >>> -- >>> 2.26.0 >>> >>>
diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c index 9df69d5efe8c..c89cd48a0aab 100644 --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c @@ -1645,6 +1645,9 @@ static int mlx5_vdpa_change_map(struct mlx5_vdpa_net *ndev, struct vhost_iotlb * if (err) goto err_mr; + if (!(ndev->mvdev.status & VIRTIO_CONFIG_S_DRIVER_OK)) + return 0; + restore_channels_info(ndev); err = setup_driver(ndev); if (err)
set_map() is used by mlx5 vdpa to create a memory region based on the address map passed by the iotlb argument. If we get successive calls, we will destroy the current memory region and build another one based on the new address mapping. We also need to setup the hardware resources since they depend on the memory region. If these calls happen before DRIVER_OK, It means that driver VQs may also not been setup and we may not create them yet. In this case we want to avoid setting up the other resources and defer this till we get DRIVER OK. Fixes: 1a86b377aa21 ("vdpa/mlx5: Add VDPA driver for supported mlx5 devices") Signed-off-by: Eli Cohen <elic@nvidia.com> --- V1->V2: Improve changelog description drivers/vdpa/mlx5/net/mlx5_vnet.c | 3 +++ 1 file changed, 3 insertions(+)