RE: [PATCH rdma-next] IB/cma: Honor traffic class from lower netdevice for RoCE

[Parav Pandit <parav@mellanox.com>]

> -----Original Message-----
> From: Jason Gunthorpe <jgg@ziepe.ca>
> Sent: Tuesday, October 8, 2019 2:10 PM
> To: Leon Romanovsky <leon@kernel.org>
> Cc: Doug Ledford <dledford@redhat.com>; Parav Pandit
> <parav@mellanox.com>; RDMA mailing list <linux-rdma@vger.kernel.org>;
> Leon Romanovsky <leonro@mellanox.com>
> Subject: Re: [PATCH rdma-next] IB/cma: Honor traffic class from lower
> netdevice for RoCE
> 
> 
> On Wed, Oct 02, 2019 at 03:19:59PM +0300, Leon Romanovsky wrote:
> > From: Parav Pandit <parav@mellanox.com>
> >
> > When macvlan netdevice is used for RoCE, consider the tos->prio->tc
> > mapping as SL using its lower netdevice.
> > 1. If lower netdevice is VLAN netdevice, consider such VLAN netdevice
> > and it's parent netdevice for mapping 2. If lower netdevice is not a
> > VLAN netdevice, consider tc mapping directly from such lower netdevice
> >
> > Signed-off-by: Parav Pandit <parav@mellanox.com>
> > Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
> > drivers/infiniband/core/cma.c | 59 +++++++++++++++++++++++++++++------
> >  1 file changed, 50 insertions(+), 9 deletions(-)
> >
> > diff --git a/drivers/infiniband/core/cma.c
> > b/drivers/infiniband/core/cma.c index 0e3cf3461999..18b5ad8c7d5f
> > 100644
> > +++ b/drivers/infiniband/core/cma.c
> > @@ -2827,22 +2827,63 @@ static int cma_resolve_iw_route(struct
> rdma_id_private *id_priv)
> >  	return 0;
> >  }
> >
> > -static int iboe_tos_to_sl(struct net_device *ndev, int tos)
> > +static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
> >  {
> > -	int prio;
> >  	struct net_device *dev;
> >
> > -	prio = rt_tos2priority(tos);
> > -	dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
> > +	dev = vlan_dev_real_dev(vlan_ndev);
> >  	if (dev->num_tc)
> >  		return netdev_get_prio_tc_map(dev, prio);
> >
> > -#if IS_ENABLED(CONFIG_VLAN_8021Q)
> > +	return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
> > +		VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT; }
> > +
> > +struct iboe_prio_tc_map {
> > +	int input_prio;
> > +	int output_tc;
> > +	bool found;
> > +};
> > +
> > +static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
> > +{
> > +	struct iboe_prio_tc_map *map = data;
> > +
> > +	if (is_vlan_dev(dev))
> > +		map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
> > +	else if (dev->num_tc)
> > +		map->output_tc = netdev_get_prio_tc_map(dev, map-
> >input_prio);
> > +	else
> > +		map->output_tc = 0;
> > +	/* We are interested only in first level VLAN device, so always
> > +	 * return 1 to stop iterating over next level devices.
> > +	 */
> > +	map->found = true;
> > +	return 1;
> > +}
> > +
> > +static int iboe_tos_to_sl(struct net_device *ndev, int tos) {
> > +	struct iboe_prio_tc_map prio_tc_map = {};
> > +	int prio = rt_tos2priority(tos);
> > +
> > +	/* If VLAN device, get it directly from the VLAN netdev */
> >  	if (is_vlan_dev(ndev))
> > -		return (vlan_dev_get_egress_qos_mask(ndev, prio) &
> > -			VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
> > -#endif
> > -	return 0;
> > +		return get_vlan_ndev_tc(ndev, prio);
> > +
> > +	prio_tc_map.input_prio = prio;
> > +	netdev_walk_all_lower_dev_rcu(ndev,
> > +				      get_lower_vlan_dev_tc,
> > +				      &prio_tc_map);
> 
> Kinda looks like you have to hold rcu before calling this?
> 
Oh yes, my bad.
rcu lock unlock calls are missing around netdev_walk_all_lower_dev_rcu().
Will respin through Leon's tree.

> Jason