Re: [RFC v1 for accelerated IPoIB 05/25] IB/ipoib: Support ipoib acceleration options callbacks

On Wed, Mar 15, 2017 at 09:58:02AM -0600, Jason Gunthorpe wrote:

On Wed, Mar 15, 2017 at 08:47:51AM +0200, Leon Romanovsky wrote:

quoted

On Tue, Mar 14, 2017 at 10:00:21AM -0600, Jason Gunthorpe wrote:

quoted

On Tue, Mar 14, 2017 at 04:42:55PM +0200, Erez Shitrit wrote:

quoted

quoted

quoted

+       if (!hca->alloc_rdma_netdev)
+               dev = ipoib_create_netdev_default(hca, name,
ipoib_setup_common);
+       else
+               dev = hca->alloc_rdma_netdev(hca, port, RDMA_NETDEV_IPOIB,
+                                            name, NET_NAME_UNKNOWN,
+                                            ipoib_setup_common);
+       if (!dev) {
+               kfree(priv);
+               return NULL;
+       }

This will break ipoib on hfi1 as hfi1 will define alloc_rdma_netdev for
OPA_VNIC type. We should probably look for a dedicated return type
(-ENODEV?) to determine of the driver supports specified rdma netdev type.
Or use a ib device attribute to suggest driver support ipoib rdma netdev.

sorry, I don't understand that, we are in ipoib driver, so the type is
RDMA_NETDEV_IPOIB, if hfi wants to implement it should use the same
flag, and to use OPA_VNIC for vnic.

He means it should look like this:

 if (hca->alloc_rdma_netdev)
     dev = hca->alloc_rdma_netdev(hca, port, RDMA_NETDEV_IPOIB,
                                            name, NET_NAME_UNKNOWN,
                                            ipoib_setup_common);

 if (IS_ERR(dev) && PTR_ERR(dev) != ENOTSUP)
      goto out;

 dev = ipoib_create_netdev_default(hca, name, ipoib_setup_common);
 if (IS_ERR(dev))
      goto out;

 WARN_ON(dev == NULL);

  [...]

out:
      return PTR_ERR(dev);

And I'm confused why 'ipoib_create_netdev_default' doesn't need the
same function signature as hca->alloc_rdma_netdev

And now, I'm confused.
In your's proposal, the "dev" will be overwritten, in Erez's proposal,
"dev" will be one of two: defaults or device specific.

Well, not Erez's version allowed dev to be ERR_PTR too. More like this then

struct rdma_netdev *get_netdev(..)
{
   if (hca-alloc_rdma_netdev) {
       dev = hca-alloc_rdma_netdev(hca, port, RDMA_NETDEV_IPOIB,
                                              name, NET_NAME_UNKNOWN,
                                              ipoib_setup_common);

       if (!IS_ERR(dev) || PTR_ERR(dev) != ENOTSUP)
           return dev;
   }

   return ipoib_create_netdev_default(hca, name, ipoib_setup_common);
}

Thanks, I agree with you, the interfaces should properly handle errors
paths from day one.

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html