On Tue, Apr 28, 2020 at 08:15:25PM -0300, Jason Gunthorpe wrote:
> On Sun, Apr 26, 2020 at 10:17:12AM +0300, Maor Gottlieb wrote:
> > +int rdma_lag_get_ah_roce_slave(struct ib_device *device,
> > + struct rdma_ah_attr *ah_attr,
> > + struct net_device **xmit_slave)
>
> Please do not use ** and also return int. The function should return
> net_device directly and use ERR_PTR()
>
> > +{
> > + struct net_device *master;
> > + struct net_device *slave;
> > + int err = 0;
> > +
> > + *xmit_slave = NULL;
> > + if (!(ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE &&
> > + ah_attr->grh.sgid_attr->gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP))
> > + return 0;
> > +
> > + rcu_read_lock();
> > + master = rdma_read_gid_attr_ndev_rcu(ah_attr->grh.sgid_attr);
> > + if (IS_ERR(master)) {
> > + err = PTR_ERR(master);
> > + goto unlock;
> > + }
> > + dev_hold(master);
>
> What is the point of this dev_hold? This whole thing is under
> rcu_read_lock()
>
> > +
> > + if (!netif_is_bond_master(master))
> > + goto put;
> > +
> > + slave = rdma_get_xmit_slave_udp(device, master, ah_attr);
>
> IMHO it is probably better to keep with the dev_hold and drop the RCU
> while doing rdma_build_skb so that the allocation in here doesn't have
> to be atomic. This isn't performance sensitive so the extra atomic for
> the dev_hold is better than the unnecessary GFP_ATOMIC allocation
Though if you do this be mindful that the create_ah call site is
conditionally non-sleeping, the best thing to do would be to make the
GFP_ATOMIC conditional on !RDMA_CREATE_AH_SLEEPABLE - ie pass in a gfp
flags argument.
Jason
