bonding: refine IFF_XMIT_DST_RELEASE capability

Some workloads greatly benefit of IFF_XMIT_DST_RELEASE capability
on output net device, avoiding dirtying dst refcount.

bonding currently disables IFF_XMIT_DST_RELEASE unconditionally.

If all slaves have the IFF_XMIT_DST_RELEASE bit set, then
bonding master can also have it in its priv_flags
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jay Vosburgh <fubar@us.ibm.com>
Cc: Andy Gospodarek <andy@greyhouse.net>
Cc: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/bonding/bond_main.c

@@ -1382,6 +1382,7 @@ static void bond_compute_features(struct bonding *bond)
 	netdev_features_t vlan_features = BOND_VLAN_FEATURES;
 	unsigned short max_hard_header_len = ETH_HLEN;
 	int i;
+	unsigned int flags, dst_release_flag = IFF_XMIT_DST_RELEASE;
 
 	read_lock(&bond->lock);
 
@@ -1392,6 +1393,7 @@ static void bond_compute_features(struct bonding *bond)
 		vlan_features = netdev_increment_features(vlan_features,
 			slave->dev->vlan_features, BOND_VLAN_FEATURES);
 
+		dst_release_flag &= slave->dev->priv_flags;
 		if (slave->dev->hard_header_len > max_hard_header_len)
 			max_hard_header_len = slave->dev->hard_header_len;
 	}
@@ -1400,6 +1402,9 @@ static void bond_compute_features(struct bonding *bond)
 	bond_dev->vlan_features = vlan_features;
 	bond_dev->hard_header_len = max_hard_header_len;
 
+	flags = bond_dev->priv_flags & ~IFF_XMIT_DST_RELEASE;
+	bond_dev->priv_flags = flags | dst_release_flag;
+
 	read_unlock(&bond->lock);
 
 	netdev_change_features(bond_dev);