IB/core: Fix oops in netdev_next_upper_dev_rcu()

When support for bonding of RoCE devices was added, there was
necessarily a link between the RoCE device and the paired netdevice that
was part of the bond.  If you remove the mlx4_en module, that paired
association is broken (the RoCE device is still present but the paired
netdevice has been released).  We need to account for this in
is_upper_ndev_bond_master_filter() and filter out those links with a
broken pairing or else we later oops in netdev_next_upper_dev_rcu().

Fixes: 408f1242d940 ("IB/core: Delete lower netdevice default GID entries in bonding scenario")
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

diff --git a/drivers/infiniband/core/roce_gid_mgmt.c b/drivers/infiniband/core/roce_gid_mgmt.c
index 25d43c8f1c2a869ffc10548f1d91da7049669b66..558de0b9895cb979745b5002a623568bb5b21f7b 100644 (file)
--- a/drivers/infiniband/core/roce_gid_mgmt.c
+++ b/drivers/infiniband/core/roce_gid_mgmt.c
@@ -267,6 +267,9 @@ is_upper_ndev_bond_master_filter(struct ib_device *ib_dev, u8 port,
        struct net_device *cookie_ndev = cookie;
        bool match = false;
 
+       if (!rdma_ndev)
+               return false;
+
        rcu_read_lock();
        if (netif_is_bond_master(cookie_ndev) &&
            rdma_is_upper_dev_rcu(rdma_ndev, cookie_ndev))