RDMA/cma: Skip device which doesn't support CM
authorParav Pandit <parav@nvidia.com>
Sun, 18 Apr 2021 13:55:52 +0000 (16:55 +0300)
committerJason Gunthorpe <jgg@nvidia.com>
Wed, 21 Apr 2021 23:27:52 +0000 (20:27 -0300)
A switchdev RDMA device do not support IB CM. When such device is added to
the RDMA CM's device list, when application invokes rdma_listen(), cma
attempts to listen to such device, however it has IB CM attribute
disabled.

Due to this, rdma_listen() call fails to listen for other non switchdev
devices as well.

A below error message can be seen.

infiniband mlx5_0: RDMA CMA: cma_listen_on_dev, error -38

A failing call flow is below.

  cma_listen_on_all()
    cma_listen_on_dev()
      _cma_attach_to_dev()
        rdma_listen() <- fails on a specific switchdev device

This is because rdma_listen() is hardwired to only work with iwarp or IB
CM compatible devices.

Hence, when a IB device doesn't support IB CM or IW CM, avoid adding such
device to the cma list so rdma_listen() can't even be called.

Link: https://lore.kernel.org/r/f9cac00d52864ea7c61295e43fb64cf4db4fdae6.1618753862.git.leonro@nvidia.com
Signed-off-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
drivers/infiniband/core/cma.c

index bb7bd02..2dc302a 100644 (file)
@@ -4900,6 +4900,17 @@ static void cma_process_remove(struct cma_device *cma_dev)
        wait_for_completion(&cma_dev->comp);
 }
 
+static bool cma_supported(struct ib_device *device)
+{
+       u32 i;
+
+       rdma_for_each_port(device, i) {
+               if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
+                       return true;
+       }
+       return false;
+}
+
 static int cma_add_one(struct ib_device *device)
 {
        struct rdma_id_private *to_destroy;
@@ -4909,6 +4920,9 @@ static int cma_add_one(struct ib_device *device)
        int ret;
        u32 i;
 
+       if (!cma_supported(device))
+               return -EOPNOTSUPP;
+
        cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
        if (!cma_dev)
                return -ENOMEM;