RE: [PATCH 1/2] vdpa: mlx5: prevent cvq work from hogging CPU

[Eli Cohen <elic@nvidia.com>]

> -----Original Message-----
> From: Jason Wang <jasowang@redhat.com>
> Sent: Monday, March 21, 2022 8:04 AM
> To: mst@redhat.com; jasowang@redhat.com
> Cc: Eli Cohen <elic@nvidia.com>; virtualization@lists.linux-foundation.org; linux-kernel@vger.kernel.org
> Subject: [PATCH 1/2] vdpa: mlx5: prevent cvq work from hogging CPU
> 
> A userspace triggerable infinite loop could happen in
> mlx5_cvq_kick_handler() if userspace keeps sending a huge amount of
> cvq requests.
> 
> Fixing this by introducing a quota and re-queue the work if we're out
not requeuing the work ...
> of the budget. While at it, using a per device workqueue to avoid on
using per device work. The workqueue is already per device.
> demand memory allocation for cvq.
> 
> Fixes: 5262912ef3cfc ("vdpa/mlx5: Add support for control VQ and MAC setting")
> Signed-off-by: Jason Wang <jasowang@redhat.com>

Reviewed-by: Eli Cohen <elic@nvidia.com>

> ---
>  drivers/vdpa/mlx5/net/mlx5_vnet.c | 28 +++++++++++++++-------------
>  1 file changed, 15 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> index d0f91078600e..d5a6fb3f9c41 100644
> --- a/drivers/vdpa/mlx5/net/mlx5_vnet.c
> +++ b/drivers/vdpa/mlx5/net/mlx5_vnet.c
> @@ -163,6 +163,7 @@ struct mlx5_vdpa_net {
>  	u32 cur_num_vqs;
>  	struct notifier_block nb;
>  	struct vdpa_callback config_cb;
> +	struct mlx5_vdpa_wq_ent cvq_ent;
>  };
> 
>  static void free_resources(struct mlx5_vdpa_net *ndev);
> @@ -1600,6 +1601,8 @@ static virtio_net_ctrl_ack handle_ctrl_mq(struct mlx5_vdpa_dev *mvdev, u8 cmd)
>  	return status;
>  }
> 
> +#define MLX5_CVQ_BUDGET 16
> +
>  static void mlx5_cvq_kick_handler(struct work_struct *work)
>  {
>  	virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
> @@ -1609,17 +1612,17 @@ static void mlx5_cvq_kick_handler(struct work_struct *work)
>  	struct mlx5_control_vq *cvq;
>  	struct mlx5_vdpa_net *ndev;
>  	size_t read, write;
> -	int err;
> +	int err, n = 0;
> 
>  	wqent = container_of(work, struct mlx5_vdpa_wq_ent, work);
>  	mvdev = wqent->mvdev;
>  	ndev = to_mlx5_vdpa_ndev(mvdev);
>  	cvq = &mvdev->cvq;
>  	if (!(ndev->mvdev.actual_features & BIT_ULL(VIRTIO_NET_F_CTRL_VQ)))
> -		goto out;
> +		return;
> 
>  	if (!cvq->ready)
> -		goto out;
> +		return;
> 
>  	while (true) {
>  		err = vringh_getdesc_iotlb(&cvq->vring, &cvq->riov, &cvq->wiov, &cvq->head,
> @@ -1653,9 +1656,13 @@ static void mlx5_cvq_kick_handler(struct work_struct *work)
> 
>  		if (vringh_need_notify_iotlb(&cvq->vring))
>  			vringh_notify(&cvq->vring);
> +
> +		n++;
> +		if (n > MLX5_CVQ_BUDGET) {
> +			queue_work(mvdev->wq, &wqent->work);
> +			break;
> +		}
>  	}
> -out:
> -	kfree(wqent);
>  }
> 
>  static void mlx5_vdpa_kick_vq(struct vdpa_device *vdev, u16 idx)
> @@ -1663,7 +1670,6 @@ static void mlx5_vdpa_kick_vq(struct vdpa_device *vdev, u16 idx)
>  	struct mlx5_vdpa_dev *mvdev = to_mvdev(vdev);
>  	struct mlx5_vdpa_net *ndev = to_mlx5_vdpa_ndev(mvdev);
>  	struct mlx5_vdpa_virtqueue *mvq;
> -	struct mlx5_vdpa_wq_ent *wqent;
> 
>  	if (!is_index_valid(mvdev, idx))
>  		return;
> @@ -1672,13 +1678,7 @@ static void mlx5_vdpa_kick_vq(struct vdpa_device *vdev, u16 idx)
>  		if (!mvdev->cvq.ready)
>  			return;
> 
> -		wqent = kzalloc(sizeof(*wqent), GFP_ATOMIC);
> -		if (!wqent)
> -			return;
> -
> -		wqent->mvdev = mvdev;
> -		INIT_WORK(&wqent->work, mlx5_cvq_kick_handler);
> -		queue_work(mvdev->wq, &wqent->work);
> +		queue_work(mvdev->wq, &ndev->cvq_ent.work);
>  		return;
>  	}
> 
> @@ -2668,6 +2668,8 @@ static int mlx5_vdpa_dev_add(struct vdpa_mgmt_dev *v_mdev, const char *name,
>  	if (err)
>  		goto err_mr;
> 
> +	ndev->cvq_ent.mvdev = mvdev;
> +	INIT_WORK(&ndev->cvq_ent.work, mlx5_cvq_kick_handler);
>  	mvdev->wq = create_singlethread_workqueue("mlx5_vdpa_wq");
>  	if (!mvdev->wq) {
>  		err = -ENOMEM;
> --
> 2.18.1