Re: [net-next PATCH 01/15] net: Combine GENEVE and VXLAN port offload notifiers into single functions

On Mon, Jun 13, 2016 at 3:17 PM, Tom Herbert [off-list ref] wrote:

On Mon, Jun 13, 2016 at 2:51 PM, Alexander Duyck
[off-list ref] wrote:

quoted

On Mon, Jun 13, 2016 at 1:36 PM, Tom Herbert [off-list ref] wrote:

quoted

On Mon, Jun 13, 2016 at 1:24 PM, Alexander Duyck
[off-list ref] wrote:

quoted

On Mon, Jun 13, 2016 at 12:55 PM, Tom Herbert [off-list ref] wrote:

quoted

On Mon, Jun 13, 2016 at 10:47 AM, Alexander Duyck [off-list ref] wrote:

quoted

This patch merges the GENEVE and VXLAN code so that both functions pass
through a shared code path.  This way we can start the effort of using a
single function on the network device drivers to handle both of these
tunnel offload types.

Signed-off-by: Alexander Duyck <redacted>
---
 drivers/net/geneve.c     |   48 ++++-------------------------
 drivers/net/vxlan.c      |   46 ++++-----------------------
 include/net/udp_tunnel.h |   12 +++++++
 net/ipv4/udp_tunnel.c    |   77 ++++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 103 insertions(+), 80 deletions(-)

diff --git a/drivers/net/geneve.c b/drivers/net/geneve.c
index cadefe4fdaa2..f5ce41532cf4 100644
--- a/drivers/net/geneve.c
+++ b/drivers/net/geneve.c

@@ -399,19 +399,7 @@ static struct socket *geneve_create_sock(struct net *net, bool ipv6,

 static void geneve_notify_add_rx_port(struct geneve_sock *gs)
 {
-       struct net_device *dev;
-       struct sock *sk = gs->sock->sk;
-       struct net *net = sock_net(sk);
-       sa_family_t sa_family = geneve_get_sk_family(gs);
-       __be16 port = inet_sk(sk)->inet_sport;
-
-       rcu_read_lock();
-       for_each_netdev_rcu(net, dev) {
-               if (dev->netdev_ops->ndo_add_geneve_port)
-                       dev->netdev_ops->ndo_add_geneve_port(dev, sa_family,
-                                                            port);
-       }
-       rcu_read_unlock();
+       udp_tunnel_notify_add_rx_port(gs->sock, UDP_ENC_OFFLOAD_TYPE_GENEVE);
 }

 static int geneve_hlen(struct genevehdr *gh)

@@ -550,20 +538,7 @@ static struct geneve_sock *geneve_socket_create(struct net *net, __be16 port,

 static void geneve_notify_del_rx_port(struct geneve_sock *gs)
 {
-       struct net_device *dev;
-       struct sock *sk = gs->sock->sk;
-       struct net *net = sock_net(sk);
-       sa_family_t sa_family = geneve_get_sk_family(gs);
-       __be16 port = inet_sk(sk)->inet_sport;
-
-       rcu_read_lock();
-       for_each_netdev_rcu(net, dev) {
-               if (dev->netdev_ops->ndo_del_geneve_port)
-                       dev->netdev_ops->ndo_del_geneve_port(dev, sa_family,
-                                                            port);
-       }
-
-       rcu_read_unlock();
+       udp_tunnel_notify_add_rx_port(gs->sock, UDP_ENC_OFFLOAD_TYPE_GENEVE);
 }

 static void __geneve_sock_release(struct geneve_sock *gs)

@@ -1165,29 +1140,20 @@ static struct device_type geneve_type = {
        .name = "geneve",
 };

-/* Calls the ndo_add_geneve_port of the caller in order to
+/* Calls the ndo_add_udp_enc_port of the caller in order to
  * supply the listening GENEVE udp ports. Callers are expected
- * to implement the ndo_add_geneve_port.
+ * to implement the ndo_add_udp_enc_port.
  */
 static void geneve_push_rx_ports(struct net_device *dev)
 {
        struct net *net = dev_net(dev);
        struct geneve_net *gn = net_generic(net, geneve_net_id);
        struct geneve_sock *gs;
-       sa_family_t sa_family;
-       struct sock *sk;
-       __be16 port;
-
-       if (!dev->netdev_ops->ndo_add_geneve_port)
-               return;

        rcu_read_lock();
-       list_for_each_entry_rcu(gs, &gn->sock_list, list) {
-               sk = gs->sock->sk;
-               sa_family = sk->sk_family;
-               port = inet_sk(sk)->inet_sport;
-               dev->netdev_ops->ndo_add_geneve_port(dev, sa_family, port);
-       }
+       list_for_each_entry_rcu(gs, &gn->sock_list, list)
+               udp_tunnel_push_rx_port(dev, gs->sock,
+                                       UDP_ENC_OFFLOAD_TYPE_GENEVE);
        rcu_read_unlock();
 }

diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c
index f999db2f97b4..43f634282726 100644
--- a/drivers/net/vxlan.c
+++ b/drivers/net/vxlan.c

@@ -622,37 +622,13 @@ static int vxlan_gro_complete(struct sock *sk, struct sk_buff *skb, int nhoff)
 /* Notify netdevs that UDP port started listening */
 static void vxlan_notify_add_rx_port(struct vxlan_sock *vs)
 {
-       struct net_device *dev;
-       struct sock *sk = vs->sock->sk;
-       struct net *net = sock_net(sk);
-       sa_family_t sa_family = vxlan_get_sk_family(vs);
-       __be16 port = inet_sk(sk)->inet_sport;
-
-       rcu_read_lock();
-       for_each_netdev_rcu(net, dev) {
-               if (dev->netdev_ops->ndo_add_vxlan_port)
-                       dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
-                                                           port);
-       }
-       rcu_read_unlock();
+       udp_tunnel_notify_add_rx_port(vs->sock, UDP_ENC_OFFLOAD_TYPE_VXLAN);
 }

 /* Notify netdevs that UDP port is no more listening */
 static void vxlan_notify_del_rx_port(struct vxlan_sock *vs)
 {
-       struct net_device *dev;
-       struct sock *sk = vs->sock->sk;
-       struct net *net = sock_net(sk);
-       sa_family_t sa_family = vxlan_get_sk_family(vs);
-       __be16 port = inet_sk(sk)->inet_sport;
-
-       rcu_read_lock();
-       for_each_netdev_rcu(net, dev) {
-               if (dev->netdev_ops->ndo_del_vxlan_port)
-                       dev->netdev_ops->ndo_del_vxlan_port(dev, sa_family,
-                                                           port);
-       }
-       rcu_read_unlock();
+       udp_tunnel_notify_del_rx_port(vs->sock, UDP_ENC_OFFLOAD_TYPE_VXLAN);
 }

 /* Add new entry to forwarding table -- assumes lock held */

@@ -2525,30 +2501,22 @@ static struct device_type vxlan_type = {
        .name = "vxlan",
 };

-/* Calls the ndo_add_vxlan_port of the caller in order to
+/* Calls the ndo_add_udp_enc_port of the caller in order to
  * supply the listening VXLAN udp ports. Callers are expected
- * to implement the ndo_add_vxlan_port.
+ * to implement the ndo_add_udp_enc_port.
  */
 static void vxlan_push_rx_ports(struct net_device *dev)
 {
        struct vxlan_sock *vs;
        struct net *net = dev_net(dev);
        struct vxlan_net *vn = net_generic(net, vxlan_net_id);
-       sa_family_t sa_family;
-       __be16 port;
        unsigned int i;

-       if (!dev->netdev_ops->ndo_add_vxlan_port)
-               return;
-
        spin_lock(&vn->sock_lock);
        for (i = 0; i < PORT_HASH_SIZE; ++i) {
-               hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist) {
-                       port = inet_sk(vs->sock->sk)->inet_sport;
-                       sa_family = vxlan_get_sk_family(vs);
-                       dev->netdev_ops->ndo_add_vxlan_port(dev, sa_family,
-                                                           port);
-               }
+               hlist_for_each_entry_rcu(vs, &vn->sock_list[i], hlist)
+                       udp_tunnel_push_rx_port(dev, vs->sock,
+                                               UDP_ENC_OFFLOAD_TYPE_VXLAN);
        }
        spin_unlock(&vn->sock_lock);
 }

diff --git a/include/net/udp_tunnel.h b/include/net/udp_tunnel.h
index 9d14f707e534..704f931fd9ad 100644
--- a/include/net/udp_tunnel.h
+++ b/include/net/udp_tunnel.h

@@ -84,6 +84,18 @@ struct udp_tunnel_sock_cfg {
 void setup_udp_tunnel_sock(struct net *net, struct socket *sock,
                           struct udp_tunnel_sock_cfg *sock_cfg);

+/* List of offloadable UDP tunnel types */
+enum udp_enc_offloads {
+       UDP_ENC_OFFLOAD_TYPE_VXLAN,     /* RFC 7348 */
+       UDP_ENC_OFFLOAD_TYPE_GENEVE,    /* draft-ietf-nvo3-geneve */
+};
+

We've already had a lot of discussion on this. The clear outcome from
netdev was that we need to support generic offloads and move away from
protocol specific offload. Generalizing the interface to allow vendors
to unnecessarily leak out protocol specific features undermines that
effort.

Then in turn we get dirty hacks like what we have right now where
VXLAN-GPE is attempting to reuse the VXLAN offload functions or
drivers that just hard-code GENEVE ports.

Going full obstructionist on this isn't going to work.  We need to be
able to support these type of offloads because the switch vendors are
going to force the NIC vendors to do so.  We will likely never be able
to convince Cisco to implement an outer transmit checksum on their
switches.  In order to make offloads work without the outer checksum
we will need to be able to parse the frames in order to be able to
validate the inner checksum values.

NIC vendors can support checksum-complete. This works with any form of
UDP encapsulation, and IP protocol (like extension headers), and other
form of tunneling we can dream up. That was the whole point of Dave's
keynote at netdev.

Right.  That covers one tiny piece of the whole problem, but you are
holding out for hardware that may not be introduced for another 3 to 5
years.  The fact is trying to getting NIC vendors to support
checksum-complete is all well and good, but you seem to have forgotten
that NIC vendors are incredibly slow when it comes to implementing
anything.  In the meantime we will have the stuff that was already in
the pipeline coming out over the next several years.

How about the fact that we need to know that there is a tunnel there
if we want to do anything like try to parse the inner headers of a
given tunnel on Rx?  How do you propose to solve the RSS problem?

Solved by doing RSS and ECMP hash over 3-tuple of IP addresses and
IPv6 flow label (not ports). Non-zero flow labels will soon be widely
used over the Internet. IOS already is already setting them, Android
should pick up support in the next rebase, and the MS guys have
assured me that they will add support to next version of Windows. Like
a generic checksum offload, flow label works with an IP protocol,
extension header, fragmentation, UDP encapsulation, etc. With this
there is no reason for devices to parse L4 headers just to forward a
packet. HW vendors (both switches and NICs) are strongly encouraged to
support them.

Right so with all this encouragement, where are we at on getting the
switches out there to support this?  The switches are what is pushing
most of the tunnel workloads and such, and it is the switches that are
going to be the painful interfaces for us to try to offload data from.

Also have you looked into the effect on regular traffic for trying to
deal with an extra IPv4 options header?  The Intel NICs should be able
to deal with it, at least for Tx, but my concern would be the NICs
that perform parsing based offloads and such.

It seems like we might be arguing past each other as I am focused on
what we can do right now, and it seems like you are arguing the point
that we need features that probably won't be enabled in the ecosystem
for 5 to 10 years depending on the hardware release cycles.  Hopefully
we can agree to disagree and can address this again when new features
start to become available.  At that time we could probably look at
deprecating the driver API for the UDP ports assuming we have other
means of resolving this available by then.

- Alex