On Tue, Jun 27, 2017 at 12:08:38PM -0700, Vishwanathapura, Niranjana wrote: > On Tue, Jun 27, 2017 at 11:41:56AM -0600, Jason Gunthorpe wrote: > > On Tue, Jun 27, 2017 at 10:34:07AM -0700, Vishwanathapura, Niranjana wrote: > > > >diff --git a/drivers/infiniband/ulp/ipoib/ipoib_main.c b/drivers/infiniband/ulp/ipoib/ipoib_main.c > > > >index 0ddd9709e1df..91fae34bdd4f 100644 > > > >+++ b/drivers/infiniband/ulp/ipoib/ipoib_main.c > > > >@@ -2301,7 +2301,10 @@ static void ipoib_remove_one(struct ib_device *device, void *client_data) > > > > flush_workqueue(priv->wq); > > > > > > > > unregister_netdev(priv->dev); > > > >- free_netdev(priv->dev); > > > >+ if (device->free_rdma_netdev) > > > >+ device->free_rdma_netdev(priv->dev); > > > >+ else > > > >+ free_netdev(priv->dev); > > > > > > This is causing an crash in hfi1 driver. > > > If the call to alloc_rdma_netdev() has returned -EOPNOTSUPP, we shouldn't be > > > calling free_rdma_netdev() here (as device driver doesn't own that netdev). > > > I will send out a patch to hfi1 driver to guard check for the type in its > > > free_rdma_netdev() call to handle such issues. > > > > I think instead you should move free_rdma_netdev from struct ib_device > > to struct rdma_netdev to prevent this mistake in general. > > > > Thanks, that is another option. With that, we will be freeing the > rdma_netdev from a function pointer call that belongs to that rdma_netdev. > Is that ok? (Though I think it should work fine). The downside of this that the code will be non-symmetrical, for alloc_rdma_netdev vs. free_rdma_netdev. 2198 /** 2199 * rdma netdev operations 2200 * 2201 * Driver implementing alloc_rdma_netdev must return -EOPNOTSUPP if it 2202 * doesn't support the specified rdma netdev type. 2203 */ 2204 struct net_device *(*alloc_rdma_netdev)( 2205 struct ib_device *device, 2206 u8 port_num, 2207 enum rdma_netdev_t type, 2208 const char *name, 2209 unsigned char name_assign_type, 2210 void (*setup)(struct net_device *)); 2211 void (*free_rdma_netdev)(struct net_device *netdev); > > Also, I think this needs updating ib_verbs.h, hfi1, ipoib and mlx5 (under > drivers/net) together. I can make the change, if this is the direction we > are taking. > > Another option is to add a new type RDMA_NETDEV_IPOIB_DEF and use only > within ipoib driver to differentiate default case. It is less optimal option, I would prefer to see IPoIB the same for all flows. Thanks > > Thanks, > Niranjana >
Attachment:
signature.asc
Description: PGP signature
