On Sat, Apr 15, 2023 at 4:06 AM Lorenzo Bianconi <lorenzo@xxxxxxxxxx> wrote:
>
> > On 4/15/23 12:10 AM, Lorenzo Bianconi wrote:
> > > > On 4/14/23 11:21 PM, Lorenzo Bianconi wrote:
> > > > > NETDEV_XDP_ACT_NDO_XMIT is not enabled by default for veth driver but it
> > > > > depends on the device configuration. Fix XDP_REDIRECT xdp-features in
> > > > > xdp_bonding selftest loading a dummy XDP program on veth2_2 device.
> > > > >
> > > > > Fixes: fccca038f300 ("veth: take into account device reconfiguration for xdp_features flag")
> > > >
> > > > Hm, does that mean we're changing^breaking existing user behavior iff after
> > > > fccca038f300 you can only make it work by loading dummy prog?
> > >
> > > nope, even before in order to enable ndo_xdp_xmit for veth you should load a dummy
> > > program on the device peer or enable gro on the device peer:
> > >
> > > https://github.com/torvalds/linux/blob/master/drivers/net/veth.c#L477
> > >
> > > we are just reflecting this behaviour in the xdp_features flag.
> >
> > Ok, I'm confused then why it passed before?
>
> ack, you are right. I guess the issue is in veth driver code. In order to
> enable NETDEV_XDP_ACT_NDO_XMIT for device "veth0", we need to check the peer
> veth1 configuration since the check in veth_xdp_xmit() is on the peer rx queue.
> Something like:
>
> diff --git a/drivers/net/veth.c b/drivers/net/veth.c
> index e1b38fbf1dd9..4b3c6647edc6 100644
> --- a/drivers/net/veth.c
> +++ b/drivers/net/veth.c
> @@ -1262,11 +1262,12 @@ static void veth_set_xdp_features(struct net_device *dev)
>
> peer = rtnl_dereference(priv->peer);
> if (peer && peer->real_num_tx_queues <= dev->real_num_rx_queues) {
> + struct veth_priv *priv_peer = netdev_priv(peer);
> xdp_features_t val = NETDEV_XDP_ACT_BASIC |
> NETDEV_XDP_ACT_REDIRECT |
> NETDEV_XDP_ACT_RX_SG;
>
> - if (priv->_xdp_prog || veth_gro_requested(dev))
> + if (priv_peer->_xdp_prog || veth_gro_requested(peer))
> val |= NETDEV_XDP_ACT_NDO_XMIT |
> NETDEV_XDP_ACT_NDO_XMIT_SG;
> xdp_set_features_flag(dev, val);
> @@ -1504,19 +1505,23 @@ static int veth_set_features(struct net_device *dev,
> {
> netdev_features_t changed = features ^ dev->features;
> struct veth_priv *priv = netdev_priv(dev);
> + struct net_device *peer;
> int err;
>
> if (!(changed & NETIF_F_GRO) || !(dev->flags & IFF_UP) || priv->_xdp_prog)
> return 0;
>
> + peer = rtnl_dereference(priv->peer);
> if (features & NETIF_F_GRO) {
> err = veth_napi_enable(dev);
> if (err)
> return err;
>
> - xdp_features_set_redirect_target(dev, true);
> + if (peer)
> + xdp_features_set_redirect_target(peer, true);
> } else {
> - xdp_features_clear_redirect_target(dev);
> + if (peer)
> + xdp_features_clear_redirect_target(peer);
> veth_napi_del(dev);
> }
> return 0;
> @@ -1598,13 +1603,13 @@ static int veth_xdp_set(struct net_device *dev, struct bpf_prog *prog,
> peer->max_mtu = max_mtu;
> }
>
> - xdp_features_set_redirect_target(dev, true);
> + xdp_features_set_redirect_target(peer, true);
> }
>
> if (old_prog) {
> if (!prog) {
> - if (!veth_gro_requested(dev))
> - xdp_features_clear_redirect_target(dev);
> + if (peer && !veth_gro_requested(dev))
> + xdp_features_clear_redirect_target(peer);
>
> if (dev->flags & IFF_UP)
> veth_disable_xdp(dev);
>
> What do you think?
Please send an official patch.
We need to fix this regression asap.
