RDMA/mlx5: Refactor netdev affinity code
authorMark Bloch <markb@mellanox.com>
Thu, 28 Mar 2019 13:27:39 +0000 (15:27 +0200)
committerJason Gunthorpe <jgg@mellanox.com>
Wed, 10 Apr 2019 18:05:39 +0000 (15:05 -0300)
The design of representors is such that once an IB representor is created,
the netdev of representor already exists, we can use that fact to simplify
the netdev affinity code.

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/hw/mlx5/ib_rep.c
drivers/infiniband/hw/mlx5/main.c

index d3988f6..7946cf2 100644 (file)
@@ -71,6 +71,8 @@ mlx5_ib_vport_rep_load(struct mlx5_core_dev *dev, struct mlx5_eswitch_rep *rep)
 
        ibdev->is_rep = true;
        ibdev->port[0].rep = rep;
+       ibdev->port[0].roce.netdev =
+               mlx5_ib_get_rep_netdev(dev->priv.eswitch, rep->vport);
        ibdev->mdev = dev;
        ibdev->num_ports = num_ports;
 
index 38c7156..0c3335d 100644 (file)
@@ -156,6 +156,34 @@ static int get_port_state(struct ib_device *ibdev,
        return ret;
 }
 
+static struct mlx5_roce *mlx5_get_rep_roce(struct mlx5_ib_dev *dev,
+                                          struct net_device *ndev,
+                                          u8 *port_num)
+{
+       struct mlx5_eswitch *esw = dev->mdev->priv.eswitch;
+       struct net_device *rep_ndev;
+       struct mlx5_ib_port *port;
+       int i;
+
+       for (i = 0; i < dev->num_ports; i++) {
+               port  = &dev->port[i];
+               if (!port->rep)
+                       continue;
+
+               read_lock(&port->roce.netdev_lock);
+               rep_ndev = mlx5_ib_get_rep_netdev(esw,
+                                                 port->rep->vport);
+               if (rep_ndev == ndev) {
+                       read_unlock(&port->roce.netdev_lock);
+                       *port_num = i + 1;
+                       return &port->roce;
+               }
+               read_unlock(&port->roce.netdev_lock);
+       }
+
+       return NULL;
+}
+
 static int mlx5_netdev_event(struct notifier_block *this,
                             unsigned long event, void *ptr)
 {
@@ -172,22 +200,17 @@ static int mlx5_netdev_event(struct notifier_block *this,
 
        switch (event) {
        case NETDEV_REGISTER:
+               /* Should already be registered during the load */
+               if (ibdev->is_rep)
+                       break;
                write_lock(&roce->netdev_lock);
-               if (ibdev->is_rep) {
-                       struct mlx5_eswitch *esw = ibdev->mdev->priv.eswitch;
-                       struct mlx5_eswitch_rep *rep = ibdev->port[0].rep;
-                       struct net_device *rep_ndev;
-
-                       rep_ndev = mlx5_ib_get_rep_netdev(esw, rep->vport);
-                       if (rep_ndev == ndev)
-                               roce->netdev = ndev;
-               } else if (ndev->dev.parent == &mdev->pdev->dev) {
+               if (ndev->dev.parent == &mdev->pdev->dev)
                        roce->netdev = ndev;
-               }
                write_unlock(&roce->netdev_lock);
                break;
 
        case NETDEV_UNREGISTER:
+               /* In case of reps, ib device goes away before the netdevs */
                write_lock(&roce->netdev_lock);
                if (roce->netdev == ndev)
                        roce->netdev = NULL;
@@ -205,6 +228,10 @@ static int mlx5_netdev_event(struct notifier_block *this,
                        dev_put(lag_ndev);
                }
 
+               if (ibdev->is_rep)
+                       roce = mlx5_get_rep_roce(ibdev, ndev, &port_num);
+               if (!roce)
+                       return NOTIFY_DONE;
                if ((upper == ndev || (!upper && ndev == roce->netdev))
                    && ibdev->ib_active) {
                        struct ib_event ibev = { };