[ovs-dev] [PATCH v2 7/7] dpif-netlink: Probe for out-of-tree tunnels, decides used interface

[Eric Garver]

On dpif init, probe for whether tunnels are created using in-tree
(upstream linux) or out-of-tree (OVS). This is done by probing for the
existence of "ovs_geneve" via rtnetlink. This is used to determine how
to create the tunnel devices.

For out-of-tree tunnels, only try genetlink/compat.
For in-tree kernel tunnels, try rtnetlink then fallback to genetlink.

Signed-off-by: Eric Garver <e at erig.me>
---
 NEWS                    |  3 +++
 lib/dpif-netlink-rtnl.c | 38 ++++++++++++++++++++++++++++++++++++++
 lib/dpif-netlink-rtnl.h |  8 ++++++++
 lib/dpif-netlink.c      | 25 ++++++++++++++++++++++---
 4 files changed, 71 insertions(+), 3 deletions(-)

diff --git a/NEWS b/NEWS
index bd13d7b7ab19..e70ce01a752b 100644
--- a/NEWS
+++ b/NEWS
@@ -3,6 +3,9 @@ Post-v2.7.0
    - Tunnels:
      * Added support to set packet mark for tunnel endpoint using
        `egress_pkt_mark` OVSDB option.
+     * When using Linux kernel datapath tunnels may be created using rtnetlink.
+       This will allow us to take advantage of new tunnel features without
+       having to make changes to the vport modules.
    - EMC insertion probability is reduced to 1% and is configurable via
      the new 'other_config:emc-insert-inv-prob' option.
    - DPDK:
diff --git a/lib/dpif-netlink-rtnl.c b/lib/dpif-netlink-rtnl.c
index 560bf6cc7420..0e7e5245be2a 100644
--- a/lib/dpif-netlink-rtnl.c
+++ b/lib/dpif-netlink-rtnl.c
@@ -463,3 +463,41 @@ dpif_netlink_rtnl_port_destroy(const char *name, const char *type)
     }
     return 0;
 }
+
+/**
+ * Probe for whether the modules are out-of-tree (openvswitch) or in-tree
+ * (upstream kernel).
+ *
+ * We probe for "ovs_geneve" via rtnetlink. As long as this returns something
+ * other than EOPNOTSUPP we know that the module in use is the out-of-tree one.
+ * This will be used to determine which netlink interface to use when creating
+ * ports; rtnetlink or compat/genetlink.
+ *
+ * See ovs_tunnels_out_of_tree
+ */
+bool
+dpif_netlink_rtnl_probe_oot_tunnels(void)
+{
+    struct netdev *netdev = NULL;
+    bool out_of_tree = false;
+    int error;
+
+    error = netdev_open("ovs-system-probe", "geneve", &netdev);
+    if (!error) {
+        error = dpif_netlink_rtnl_geneve_create_kind(netdev, "ovs_geneve");
+        if (error != EOPNOTSUPP) {
+            if (!error) {
+                char namebuf[NETDEV_VPORT_NAME_BUFSIZE];
+                const char *dp_port;
+
+                dp_port = netdev_vport_get_dpif_port(netdev, namebuf,
+                                                     sizeof namebuf);
+                dpif_netlink_rtnl_geneve_destroy(dp_port);
+            }
+            out_of_tree = true;
+        }
+        netdev_close(netdev);
+    }
+
+    return out_of_tree;
+}
diff --git a/lib/dpif-netlink-rtnl.h b/lib/dpif-netlink-rtnl.h
index 952c0d4187e5..5c790e0bc06f 100644
--- a/lib/dpif-netlink-rtnl.h
+++ b/lib/dpif-netlink-rtnl.h
@@ -25,6 +25,8 @@
 int dpif_netlink_rtnl_port_create(struct netdev *netdev);
 int dpif_netlink_rtnl_port_destroy(const char *name, const char *type);
 
+bool dpif_netlink_rtnl_probe_oot_tunnels(void);
+
 #ifndef __linux__
 /* Dummy implementations for non Linux builds. */
 
@@ -41,6 +43,12 @@ dpif_netlink_rtnl_port_destroy(const char *name OVS_UNUSED,
     return EOPNOTSUPP;
 }
 
+static inline bool
+dpif_netlink_rtnl_probe_oot_tunnels(void)
+{
+    return true;
+}
+
 #endif
 
 #endif /* DPIF_NETLINK_RTNL_H */
diff --git a/lib/dpif-netlink.c b/lib/dpif-netlink.c
index 572e365e8f49..161616e5531b 100644
--- a/lib/dpif-netlink.c
+++ b/lib/dpif-netlink.c
@@ -211,6 +211,12 @@ static int ovs_packet_family;
  * Initialized by dpif_netlink_init(). */
 static unsigned int ovs_vport_mcgroup;
 
+/* If true, tunnel devices are created using OVS compat/genetlink.
+ * If false, tunnel devices are created with rtnetlink and using light weight
+ * tunnels. If we fail to create the tunnel the rtnetlink+LWT, then we fallback
+ * to using the compat interface. */
+static bool ovs_tunnels_out_of_tree = true;
+
 static int dpif_netlink_init(void);
 static int open_dpif(const struct dpif_netlink_dp *, struct dpif **);
 static uint32_t dpif_netlink_port_get_pid(const struct dpif *,
@@ -947,7 +953,7 @@ dpif_netlink_port_add_compat(struct dpif_netlink *dpif, struct netdev *netdev,
 
 }
 
-static int OVS_UNUSED
+static int
 dpif_netlink_rtnl_port_create_and_add(struct dpif_netlink *dpif,
                                       struct netdev *netdev,
                                       odp_port_t *port_nop)
@@ -981,10 +987,15 @@ dpif_netlink_port_add(struct dpif *dpif_, struct netdev *netdev,
                       odp_port_t *port_nop)
 {
     struct dpif_netlink *dpif = dpif_netlink_cast(dpif_);
-    int error;
+    int error = EOPNOTSUPP;
 
     fat_rwlock_wrlock(&dpif->upcall_lock);
-    error = dpif_netlink_port_add_compat(dpif, netdev, port_nop);
+    if (!ovs_tunnels_out_of_tree) {
+        error = dpif_netlink_rtnl_port_create_and_add(dpif, netdev, port_nop);
+    }
+    if (error) {
+        error = dpif_netlink_port_add_compat(dpif, netdev, port_nop);
+    }
     fat_rwlock_unlock(&dpif->upcall_lock);
 
     return error;
@@ -1019,6 +1030,12 @@ dpif_netlink_port_del__(struct dpif_netlink *dpif, odp_port_t port_no)
 
     vport_del_channels(dpif, port_no);
 
+    if (!error && !ovs_tunnels_out_of_tree) {
+        error = dpif_netlink_rtnl_port_destroy(dpif_port.name, dpif_port.type);
+        if (error == EOPNOTSUPP) {
+            error = 0;
+        }
+    }
     dpif_port_destroy(&dpif_port);
 
     return error;
@@ -2532,6 +2549,8 @@ dpif_netlink_init(void)
                                            &ovs_vport_mcgroup);
         }
 
+        ovs_tunnels_out_of_tree = dpif_netlink_rtnl_probe_oot_tunnels();
+
         ovsthread_once_done(&once);
     }
 
-- 
2.10.0