> From: Long Li <longli@xxxxxxxxxxxxx> > Sent: Thursday, November 21, 2024 5:34 AM > > > > > > > Actually, another alternative solution for mana_ib is always set the > > > slave device, but in the GID mgmt code we need the following patch. > > > The problem is that it may require testing/confirmation from other > > > ib providers > > as in the worst case some GIDs will not be listed. > > > > is_eth_active_slave_of_bonding_rcu() is for bonding. > > Sorry, need to bring this issue up again. > > This patch has broken user-space programs (e.g DPDK) that requires to > export a kernel device to user-mode. > > With this patch, the RDMA driver grabbed a reference from the master > device, it's impossible to move the master device to user-mode. > > I think the root cause is that the individual driver should not decide on which > (master or slave) address should be used for GID. roce_gid_mgmt.c should > handle this situation. > > I think Konstantin's suggestion makes sense, how about we do this (don't > need to define netdev_is_slave(dev)): > > --- a/drivers/infiniband/core/roce_gid_mgmt.c > +++ b/drivers/infiniband/core/roce_gid_mgmt.c > @@ -161,7 +161,7 @@ is_eth_port_of_netdev_filter(struct ib_device > *ib_dev, u32 port, > res = ((rdma_is_upper_dev_rcu(rdma_ndev, cookie) && > (is_eth_active_slave_of_bonding_rcu(rdma_ndev, real_dev) & > REQUIRED_BOND_STATES)) || > - real_dev == rdma_ndev); > + (real_dev == rdma_ndev && > + !netif_is_bond_slave(rdma_ndev))); > > rcu_read_unlock(); > return res; > > > is_eth_port_of_netdev_filter() should not return true if this netdev is a > bonded slave. In this case, only use the address of its bonded master. > Right. This change makes sense to me. I don't have a setup presently to verify it to ensure I didn't miss a corner case. Leon, Can you or others please test the regression once with the formal patch?
