IB/mlx5: Make netdev notifications multiport capable
authorDaniel Jurgens <danielj@mellanox.com>
Thu, 4 Jan 2018 15:25:35 +0000 (17:25 +0200)
committerJason Gunthorpe <jgg@mellanox.com>
Mon, 8 Jan 2018 18:42:21 +0000 (11:42 -0700)
When multiple RoCE ports are supported registration for events on
multiple netdevs is required. Refactor the event registration and
handling to support multiple ports.

Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
drivers/infiniband/hw/mlx5/main.c
drivers/infiniband/hw/mlx5/mlx5_ib.h
drivers/infiniband/hw/mlx5/qp.c
include/linux/mlx5/driver.h

index 0c24cd4..5fcb2ed 100644 (file)
@@ -113,24 +113,30 @@ static int get_port_state(struct ib_device *ibdev,
 static int mlx5_netdev_event(struct notifier_block *this,
                             unsigned long event, void *ptr)
 {
+       struct mlx5_roce *roce = container_of(this, struct mlx5_roce, nb);
        struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
-       struct mlx5_ib_dev *ibdev = container_of(this, struct mlx5_ib_dev,
-                                                roce.nb);
+       u8 port_num = roce->native_port_num;
+       struct mlx5_core_dev *mdev;
+       struct mlx5_ib_dev *ibdev;
+
+       ibdev = roce->dev;
+       mdev = ibdev->mdev;
 
        switch (event) {
        case NETDEV_REGISTER:
        case NETDEV_UNREGISTER:
-               write_lock(&ibdev->roce.netdev_lock);
-               if (ndev->dev.parent == &ibdev->mdev->pdev->dev)
-                       ibdev->roce.netdev = (event == NETDEV_UNREGISTER) ?
-                                            NULL : ndev;
-               write_unlock(&ibdev->roce.netdev_lock);
+               write_lock(&roce->netdev_lock);
+
+               if (ndev->dev.parent == &mdev->pdev->dev)
+                       roce->netdev = (event == NETDEV_UNREGISTER) ?
+                                       NULL : ndev;
+               write_unlock(&roce->netdev_lock);
                break;
 
        case NETDEV_CHANGE:
        case NETDEV_UP:
        case NETDEV_DOWN: {
-               struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(ibdev->mdev);
+               struct net_device *lag_ndev = mlx5_lag_get_roce_netdev(mdev);
                struct net_device *upper = NULL;
 
                if (lag_ndev) {
@@ -138,27 +144,28 @@ static int mlx5_netdev_event(struct notifier_block *this,
                        dev_put(lag_ndev);
                }
 
-               if ((upper == ndev || (!upper && ndev == ibdev->roce.netdev))
+               if ((upper == ndev || (!upper && ndev == roce->netdev))
                    && ibdev->ib_active) {
                        struct ib_event ibev = { };
                        enum ib_port_state port_state;
 
-                       if (get_port_state(&ibdev->ib_dev, 1, &port_state))
-                               return NOTIFY_DONE;
+                       if (get_port_state(&ibdev->ib_dev, port_num,
+                                          &port_state))
+                               goto done;
 
-                       if (ibdev->roce.last_port_state == port_state)
-                               return NOTIFY_DONE;
+                       if (roce->last_port_state == port_state)
+                               goto done;
 
-                       ibdev->roce.last_port_state = port_state;
+                       roce->last_port_state = port_state;
                        ibev.device = &ibdev->ib_dev;
                        if (port_state == IB_PORT_DOWN)
                                ibev.event = IB_EVENT_PORT_ERR;
                        else if (port_state == IB_PORT_ACTIVE)
                                ibev.event = IB_EVENT_PORT_ACTIVE;
                        else
-                               return NOTIFY_DONE;
+                               goto done;
 
-                       ibev.element.port_num = 1;
+                       ibev.element.port_num = port_num;
                        ib_dispatch_event(&ibev);
                }
                break;
@@ -167,7 +174,7 @@ static int mlx5_netdev_event(struct notifier_block *this,
        default:
                break;
        }
-
+done:
        return NOTIFY_DONE;
 }
 
@@ -183,11 +190,11 @@ static struct net_device *mlx5_ib_get_netdev(struct ib_device *device,
 
        /* Ensure ndev does not disappear before we invoke dev_hold()
         */
-       read_lock(&ibdev->roce.netdev_lock);
-       ndev = ibdev->roce.netdev;
+       read_lock(&ibdev->roce[port_num - 1].netdev_lock);
+       ndev = ibdev->roce[port_num - 1].netdev;
        if (ndev)
                dev_hold(ndev);
-       read_unlock(&ibdev->roce.netdev_lock);
+       read_unlock(&ibdev->roce[port_num - 1].netdev_lock);
 
        return ndev;
 }
@@ -3579,33 +3586,33 @@ static void mlx5_eth_lag_cleanup(struct mlx5_ib_dev *dev)
        }
 }
 
-static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev)
+static int mlx5_add_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
 {
        int err;
 
-       dev->roce.nb.notifier_call = mlx5_netdev_event;
-       err = register_netdevice_notifier(&dev->roce.nb);
+       dev->roce[port_num].nb.notifier_call = mlx5_netdev_event;
+       err = register_netdevice_notifier(&dev->roce[port_num].nb);
        if (err) {
-               dev->roce.nb.notifier_call = NULL;
+               dev->roce[port_num].nb.notifier_call = NULL;
                return err;
        }
 
        return 0;
 }
 
-static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev)
+static void mlx5_remove_netdev_notifier(struct mlx5_ib_dev *dev, u8 port_num)
 {
-       if (dev->roce.nb.notifier_call) {
-               unregister_netdevice_notifier(&dev->roce.nb);
-               dev->roce.nb.notifier_call = NULL;
+       if (dev->roce[port_num].nb.notifier_call) {
+               unregister_netdevice_notifier(&dev->roce[port_num].nb);
+               dev->roce[port_num].nb.notifier_call = NULL;
        }
 }
 
-static int mlx5_enable_eth(struct mlx5_ib_dev *dev)
+static int mlx5_enable_eth(struct mlx5_ib_dev *dev, u8 port_num)
 {
        int err;
 
-       err = mlx5_add_netdev_notifier(dev);
+       err = mlx5_add_netdev_notifier(dev, port_num);
        if (err)
                return err;
 
@@ -3626,7 +3633,7 @@ err_disable_roce:
                mlx5_nic_vport_disable_roce(dev->mdev);
 
 err_unregister_netdevice_notifier:
-       mlx5_remove_netdev_notifier(dev);
+       mlx5_remove_netdev_notifier(dev, port_num);
        return err;
 }
 
@@ -4066,7 +4073,6 @@ static int mlx5_ib_stage_init_init(struct mlx5_ib_dev *dev)
        if (!dev->port)
                return -ENOMEM;
 
-       rwlock_init(&dev->roce.netdev_lock);
        err = get_port_caps(dev);
        if (err)
                goto err_free_port;
@@ -4246,12 +4252,21 @@ static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
        struct mlx5_core_dev *mdev = dev->mdev;
        enum rdma_link_layer ll;
        int port_type_cap;
+       u8 port_num = 0;
        int err;
+       int i;
 
        port_type_cap = MLX5_CAP_GEN(mdev, port_type);
        ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 
        if (ll == IB_LINK_LAYER_ETHERNET) {
+               for (i = 0; i < dev->num_ports; i++) {
+                       rwlock_init(&dev->roce[i].netdev_lock);
+                       dev->roce[i].dev = dev;
+                       dev->roce[i].native_port_num = i + 1;
+                       dev->roce[i].last_port_state = IB_PORT_DOWN;
+               }
+
                dev->ib_dev.get_netdev  = mlx5_ib_get_netdev;
                dev->ib_dev.create_wq    = mlx5_ib_create_wq;
                dev->ib_dev.modify_wq    = mlx5_ib_modify_wq;
@@ -4264,10 +4279,9 @@ static int mlx5_ib_stage_roce_init(struct mlx5_ib_dev *dev)
                        (1ull << IB_USER_VERBS_EX_CMD_DESTROY_WQ) |
                        (1ull << IB_USER_VERBS_EX_CMD_CREATE_RWQ_IND_TBL) |
                        (1ull << IB_USER_VERBS_EX_CMD_DESTROY_RWQ_IND_TBL);
-               err = mlx5_enable_eth(dev);
+               err = mlx5_enable_eth(dev, port_num);
                if (err)
                        return err;
-               dev->roce.last_port_state = IB_PORT_DOWN;
        }
 
        return 0;
@@ -4278,13 +4292,14 @@ static void mlx5_ib_stage_roce_cleanup(struct mlx5_ib_dev *dev)
        struct mlx5_core_dev *mdev = dev->mdev;
        enum rdma_link_layer ll;
        int port_type_cap;
+       u8 port_num = 0;
 
        port_type_cap = MLX5_CAP_GEN(mdev, port_type);
        ll = mlx5_port_type_cap_to_rdma_ll(port_type_cap);
 
        if (ll == IB_LINK_LAYER_ETHERNET) {
                mlx5_disable_eth(dev);
-               mlx5_remove_netdev_notifier(dev);
+               mlx5_remove_netdev_notifier(dev, port_num);
        }
 }
 
index 57405cf..6106dde 100644 (file)
@@ -667,6 +667,8 @@ struct mlx5_roce {
        struct notifier_block   nb;
        atomic_t                next_port;
        enum ib_port_state last_port_state;
+       struct mlx5_ib_dev      *dev;
+       u8                      native_port_num;
 };
 
 struct mlx5_ib_dbg_param {
@@ -757,7 +759,7 @@ struct mlx5_ib_profile {
 struct mlx5_ib_dev {
        struct ib_device                ib_dev;
        struct mlx5_core_dev            *mdev;
-       struct mlx5_roce                roce;
+       struct mlx5_roce                roce[MLX5_MAX_PORTS];
        int                             num_ports;
        /* serialize update of capability mask
         */
index 4b08472..ae36db3 100644 (file)
@@ -2962,8 +2962,9 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
                    (ibqp->qp_type == IB_QPT_XRC_INI) ||
                    (ibqp->qp_type == IB_QPT_XRC_TGT)) {
                        if (mlx5_lag_is_active(dev->mdev)) {
+                               u8 p = mlx5_core_native_port_num(dev->mdev);
                                tx_affinity = (unsigned int)atomic_add_return(1,
-                                               &dev->roce.next_port) %
+                                               &dev->roce[p].next_port) %
                                                MLX5_MAX_PORTS + 1;
                                context->flags |= cpu_to_be32(tx_affinity << 24);
                        }
index 5b0443c..2873352 100644 (file)
@@ -1234,6 +1234,11 @@ static inline bool mlx5_rl_is_supported(struct mlx5_core_dev *dev)
        return !!(dev->priv.rl_table.max_size);
 }
 
+static inline int mlx5_core_native_port_num(struct mlx5_core_dev *dev)
+{
+       return 1;
+}
+
 enum {
        MLX5_TRIGGERED_CMD_COMP = (u64)1 << 32,
 };