[PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Yi Yang]

In ovs compat mode, ovs won't use LWT in current kernel, this is to
make sure ovs can work on the old kernels, Linux kernel v4.7 includes
VxLAN-gpe support but many Linux distributions' kernels are odler than
v4.7, this fix will ensure that ovs can create VxLAN-gpe port correctly
on old kernels, it has been verified on Ubuntu 16.04 x86_64 with Linux
kernel 4.4.0-53-generic.

This does touch compat code, but it is necessary as Pravin commented.

Without this fix, ovs can't create VxLAN-gpe port, it is still a VxLAN
port.

vxlan_sys_4790 Link encap:Ethernet  HWaddr 72:23:60:c2:8b:8d
          inet6 addr: fe80::7023:60ff:fec2:8b8d/64 Scope:Link
          UP BROADCAST RUNNING MULTICAST  MTU:65485  Metric:1
          RX packets:0 errors:0 dropped:0 overruns:0 frame:0
          TX packets:0 errors:0 dropped:8 overruns:0 carrier:0
          collisions:0 txqueuelen:1000
          RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)

But with this fix applied, a real L3 port is created

vxlan_sys_4790 Link encap:UNSPEC  HWaddr
00-00-00-00-00-00-00-00-00-00-00-00-00-00-00-00
          UP POINTOPOINT RUNNING NOARP MULTICAST  MTU:65485  Metric:1
          RX packets:0 errors:0 dropped:0 overruns:0 frame:0
          TX packets:0 errors:0 dropped:0 overruns:0 carrier:0
          collisions:0 txqueuelen:1000
          RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)

Signed-off-by: Yi Yang <yi.y.yang-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
---
 include/uapi/linux/openvswitch.h |  1 +
 net/openvswitch/vport-vxlan.c    | 15 +++++++++++++++
 2 files changed, 16 insertions(+)

diff --git a/include/uapi/linux/openvswitch.h b/include/uapi/linux/openvswitch.h
index 375d812..b0e27b3 100644
--- a/include/uapi/linux/openvswitch.h
+++ b/include/uapi/linux/openvswitch.h
@@ -265,6 +265,7 @@ enum ovs_vport_attr {
 enum {
 	OVS_VXLAN_EXT_UNSPEC,
 	OVS_VXLAN_EXT_GBP,	/* Flag or __u32 */
+	OVS_VXLAN_EXT_GPE,	/* Flag or __u32 */
 	__OVS_VXLAN_EXT_MAX,
 };
 
diff --git a/net/openvswitch/vport-vxlan.c b/net/openvswitch/vport-vxlan.c
index 7eb955e..42e46af 100644
--- a/net/openvswitch/vport-vxlan.c
+++ b/net/openvswitch/vport-vxlan.c
@@ -52,6 +52,18 @@ static int vxlan_get_options(const struct vport *vport, struct sk_buff *skb)
 			return -EMSGSIZE;
 
 		nla_nest_end(skb, exts);
+	} else if (vxlan->flags & VXLAN_F_GPE) {
+		struct nlattr *exts;
+
+		exts = nla_nest_start(skb, OVS_TUNNEL_ATTR_EXTENSION);
+		if (!exts)
+			return -EMSGSIZE;
+
+		if (vxlan->flags & VXLAN_F_GPE &&
+		    nla_put_flag(skb, OVS_VXLAN_EXT_GPE))
+			return -EMSGSIZE;
+
+		nla_nest_end(skb, exts);
 	}
 
 	return 0;
@@ -59,6 +71,7 @@ static int vxlan_get_options(const struct vport *vport, struct sk_buff *skb)
 
 static const struct nla_policy exts_policy[OVS_VXLAN_EXT_MAX + 1] = {
 	[OVS_VXLAN_EXT_GBP]	= { .type = NLA_FLAG, },
+	[OVS_VXLAN_EXT_GPE]	= { .type = NLA_FLAG, },
 };
 
 static int vxlan_configure_exts(struct vport *vport, struct nlattr *attr,
@@ -76,6 +89,8 @@ static int vxlan_configure_exts(struct vport *vport, struct nlattr *attr,
 
 	if (exts[OVS_VXLAN_EXT_GBP])
 		conf->flags |= VXLAN_F_GBP;
+	else if (exts[OVS_VXLAN_EXT_GPE])
+		conf->flags |= VXLAN_F_GPE;
 
 	return 0;
 }
-- 
1.9.3

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Jiri Benc]

On Thu,  8 Dec 2016 16:20:10 +0800, Yi Yang wrote:
> In ovs compat mode, ovs won't use LWT in current kernel, this is to
> make sure ovs can work on the old kernels, Linux kernel v4.7 includes
> VxLAN-gpe support but many Linux distributions' kernels are odler than
> v4.7, this fix will ensure that ovs can create VxLAN-gpe port correctly
> on old kernels, it has been verified on Ubuntu 16.04 x86_64 with Linux
> kernel 4.4.0-53-generic.

NAK. We do have a way to configure this and that's rtnetlink. Open
vSwitch should use that to configure tunnels. Out of tree modules are
on their own. Upstream kernel does not accommodate out of tree modules.

 Jiri

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Yang, Yi]

On Thu, Dec 08, 2016 at 09:54:00AM +0100, Jiri Benc wrote:
> On Thu,  8 Dec 2016 16:20:10 +0800, Yi Yang wrote:
> > In ovs compat mode, ovs won't use LWT in current kernel, this is to
> > make sure ovs can work on the old kernels, Linux kernel v4.7 includes
> > VxLAN-gpe support but many Linux distributions' kernels are odler than
> > v4.7, this fix will ensure that ovs can create VxLAN-gpe port correctly
> > on old kernels, it has been verified on Ubuntu 16.04 x86_64 with Linux
> > kernel 4.4.0-53-generic.
> 
> NAK. We do have a way to configure this and that's rtnetlink. Open
> vSwitch should use that to configure tunnels. Out of tree modules are
> on their own. Upstream kernel does not accommodate out of tree modules.
>
Jiri, this has used rtnetlink to confgiure, te below is my test code in ovs.
As Pravin mentioned, in compat mode, ovs won't use current in-kernel
module which is vxlan in upstream kernel, but ovs has its own vport_vxlan
module for this, it has different behaviour from LWT in upstream kernel.
If you try this in the kernels below v4.7, you will clearly know this.

ovs will port this patch into the below files in ovs soure code
datapath/linux/compat/include/linux/openvswitch.h and datapath/vport-vxlan.c once it is accepted,this is ovs upstream process, if Linux kernel doesn't include this, ovs won't accept it.

So ovs out of tree modules need to adapt to upstream kernel, any
kernel-related changes must be accepted by Linux kernel at first. Pravin
is a dedicated person doing such work, your L3 patches merged into
net-next will be ported into ovs out of tree modules by Pravin in the
same way.

diff --git a/lib/dpif-netlink.c b/lib/dpif-netlink.c
index 0d03334..7d8a0f4 100644
--- a/lib/dpif-netlink.c
+++ b/lib/dpif-netlink.c
@@ -1006,7 +1006,11 @@ netdev_vxlan_create(struct netdev *netdev)
     nl_msg_put_string(&request, IFLA_IFNAME, name);
     nl_msg_put_u32(&request, IFLA_MTU, UINT16_MAX);
     linkinfo_off = nl_msg_start_nested(&request, IFLA_LINKINFO);
+#ifdef USE_UPSTREAM_TUNNEL
         nl_msg_put_string(&request, IFLA_INFO_KIND, "vxlan");
+#else
+        nl_msg_put_string(&request, IFLA_INFO_KIND, "ovs_vxlan");
+#endif
         infodata_off = nl_msg_start_nested(&request, IFLA_INFO_DATA);
             nl_msg_put_u8(&request, IFLA_VXLAN_LEARNING, 0);
             nl_msg_put_u8(&request, IFLA_VXLAN_COLLECT_METADATA, 1);

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Jiri Benc]

On Thu, 8 Dec 2016 18:57:51 +0800, Yang, Yi wrote:
> So ovs out of tree modules need to adapt to upstream kernel, any
> kernel-related changes must be accepted by Linux kernel at first.

I'm perfectly aware of that and I'm saying that your patch is
unacceptable for upstream kernel. This is a long standing policy of the
kernel: there's no way you can get a patch into the kernel to
accommodate an out of tree kernel module. The policy is there for good
reasons and as paradoxical as it may sound, it benefits the projects
that employ out of tree modules in the long run.

If Open vSwitch wants to carry a non-upstream patch, it's its choice
and we can have that discussion but that's not something to discuss on
netdev@vger nor propose for net-next.

 Jiri

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Yang, Yi]

On Thu, Dec 08, 2016 at 12:37:56PM +0100, Jiri Benc wrote:
> On Thu, 8 Dec 2016 18:57:51 +0800, Yang, Yi wrote:
> > So ovs out of tree modules need to adapt to upstream kernel, any
> > kernel-related changes must be accepted by Linux kernel at first.
> 
> I'm perfectly aware of that and I'm saying that your patch is
> unacceptable for upstream kernel. This is a long standing policy of the
> kernel: there's no way you can get a patch into the kernel to
> accommodate an out of tree kernel module. The policy is there for good
> reasons and as paradoxical as it may sound, it benefits the projects
> that employ out of tree modules in the long run.
> 
> If Open vSwitch wants to carry a non-upstream patch, it's its choice
> and we can have that discussion but that's not something to discuss on
> netdev@vger nor propose for net-next.
>
Jiri, according to your statement, we have to switch Linux 4.7 or above
if we want to use ovs VxLAN-gpe, Ubuntu 16.04 has had new kernel, but it
just use Linux kernel 4.4, you know Linux distributuions nerver uses the
latest stable Linux kernel because they have their own patches to
maintain, that will be a nightmare if they take the latest stable
kernel. You know RHEL also follows the same philosophy.

Current ovs master can be built on Ubuntu 14.04 which have Linux kernel
3.13, I think compatibility backward is very important, out of tree
modules are very important to ovs. If ovs installation will depend on
the latest kernel and force users to switch to new kernel, I believe it
nerver will be so popular in the industies.

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Pravin Shelar]

On Thu, Dec 8, 2016 at 12:20 AM, Yi Yang <yi.y.yang-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org> wrote:
> In ovs compat mode, ovs won't use LWT in current kernel, this is to
> make sure ovs can work on the old kernels, Linux kernel v4.7 includes
> VxLAN-gpe support but many Linux distributions' kernels are odler than
> v4.7, this fix will ensure that ovs can create VxLAN-gpe port correctly
> on old kernels, it has been verified on Ubuntu 16.04 x86_64 with Linux
> kernel 4.4.0-53-generic.
>
> This does touch compat code, but it is necessary as Pravin commented.
>
> Without this fix, ovs can't create VxLAN-gpe port, it is still a VxLAN
> port.
>
> vxlan_sys_4790 Link encap:Ethernet  HWaddr 72:23:60:c2:8b:8d
>           inet6 addr: fe80::7023:60ff:fec2:8b8d/64 Scope:Link
>           UP BROADCAST RUNNING MULTICAST  MTU:65485  Metric:1
>           RX packets:0 errors:0 dropped:0 overruns:0 frame:0
>           TX packets:0 errors:0 dropped:8 overruns:0 carrier:0
>           collisions:0 txqueuelen:1000
>           RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)
>
> But with this fix applied, a real L3 port is created
>
> vxlan_sys_4790 Link encap:UNSPEC  HWaddr
> 00-00-00-00-00-00-00-00-00-00-00-00-00-00-00-00
>           UP POINTOPOINT RUNNING NOARP MULTICAST  MTU:65485  Metric:1
>           RX packets:0 errors:0 dropped:0 overruns:0 frame:0
>           TX packets:0 errors:0 dropped:0 overruns:0 carrier:0
>           collisions:0 txqueuelen:1000
>           RX bytes:0 (0.0 B)  TX bytes:0 (0.0 B)
>
> Signed-off-by: Yi Yang <yi.y.yang-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org>
> ---
>  include/uapi/linux/openvswitch.h |  1 +
>  net/openvswitch/vport-vxlan.c    | 15 +++++++++++++++
>  2 files changed, 16 insertions(+)
>
There is no need for this patch in upstream kernel module. I am open
to having such a patch in out of tree kernel if it simplifies feature
compatibility code.

Re: [PATCH net-next] openvswitch: fix VxLAN-gpe port can't be created in ovs compat mode

[Yang, Yi]

On Thu, Dec 08, 2016 at 11:41:58AM -0800, Pravin Shelar wrote:
> On Thu, Dec 8, 2016 at 12:20 AM, Yi Yang <yi.y.yang@intel.com> wrote:
> >
> > Signed-off-by: Yi Yang <yi.y.yang@intel.com>
> > ---
> >  include/uapi/linux/openvswitch.h |  1 +
> >  net/openvswitch/vport-vxlan.c    | 15 +++++++++++++++
> >  2 files changed, 16 insertions(+)
> >
> There is no need for this patch in upstream kernel module. I am open
> to having such a patch in out of tree kernel if it simplifies feature
> compatibility code.

I'm very glad to hear this :-), the goal is to enable current ovs to create
vxlan-gpe port in compat mode without new kernel help, I'll post a patch
for ovs, thanks a lot.