veth: Avoid drops by oversized packets when XDP is enabled
authorToshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Fri, 3 Aug 2018 07:58:11 +0000 (16:58 +0900)
committerDaniel Borkmann <daniel@iogearbox.net>
Fri, 10 Aug 2018 14:12:20 +0000 (16:12 +0200)
Oversized packets including GSO packets can be dropped if XDP is
enabled on receiver side, so don't send such packets from peer.

Drop TSO and SCTP fragmentation features so that veth devices themselves
segment packets with XDP enabled. Also cap MTU accordingly.

v4:
- Don't auto-adjust MTU but cap max MTU.

Signed-off-by: Toshiaki Makita <makita.toshiaki@lab.ntt.co.jp>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
drivers/net/veth.c

index d3b9f10bea242a2d892102a59c5cb8f759038d2a..9edf104739dade9dc67c4048454e26377c12bd77 100644 (file)
@@ -543,6 +543,23 @@ static int veth_get_iflink(const struct net_device *dev)
        return iflink;
 }
 
+static netdev_features_t veth_fix_features(struct net_device *dev,
+                                          netdev_features_t features)
+{
+       struct veth_priv *priv = netdev_priv(dev);
+       struct net_device *peer;
+
+       peer = rtnl_dereference(priv->peer);
+       if (peer) {
+               struct veth_priv *peer_priv = netdev_priv(peer);
+
+               if (peer_priv->_xdp_prog)
+                       features &= ~NETIF_F_GSO_SOFTWARE;
+       }
+
+       return features;
+}
+
 static void veth_set_rx_headroom(struct net_device *dev, int new_hr)
 {
        struct veth_priv *peer_priv, *priv = netdev_priv(dev);
@@ -572,6 +589,7 @@ static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
        struct veth_priv *priv = netdev_priv(dev);
        struct bpf_prog *old_prog;
        struct net_device *peer;
+       unsigned int max_mtu;
        int err;
 
        old_prog = priv->_xdp_prog;
@@ -585,6 +603,15 @@ static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                        goto err;
                }
 
+               max_mtu = PAGE_SIZE - VETH_XDP_HEADROOM -
+                         peer->hard_header_len -
+                         SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+               if (peer->mtu > max_mtu) {
+                       NL_SET_ERR_MSG_MOD(extack, "Peer MTU is too large to set XDP");
+                       err = -ERANGE;
+                       goto err;
+               }
+
                if (dev->flags & IFF_UP) {
                        err = veth_enable_xdp(dev);
                        if (err) {
@@ -592,14 +619,29 @@ static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
                                goto err;
                        }
                }
+
+               if (!old_prog) {
+                       peer->hw_features &= ~NETIF_F_GSO_SOFTWARE;
+                       peer->max_mtu = max_mtu;
+               }
        }
 
        if (old_prog) {
-               if (!prog && dev->flags & IFF_UP)
-                       veth_disable_xdp(dev);
+               if (!prog) {
+                       if (dev->flags & IFF_UP)
+                               veth_disable_xdp(dev);
+
+                       if (peer) {
+                               peer->hw_features |= NETIF_F_GSO_SOFTWARE;
+                               peer->max_mtu = ETH_MAX_MTU;
+                       }
+               }
                bpf_prog_put(old_prog);
        }
 
+       if ((!!old_prog ^ !!prog) && peer)
+               netdev_update_features(peer);
+
        return 0;
 err:
        priv->_xdp_prog = old_prog;
@@ -644,6 +686,7 @@ static const struct net_device_ops veth_netdev_ops = {
        .ndo_poll_controller    = veth_poll_controller,
 #endif
        .ndo_get_iflink         = veth_get_iflink,
+       .ndo_fix_features       = veth_fix_features,
        .ndo_features_check     = passthru_features_check,
        .ndo_set_rx_headroom    = veth_set_rx_headroom,
        .ndo_bpf                = veth_xdp,