vxlan regression: multiple VXLANs on a single multicast group broken in kernel 4.1

vxlan regression: multiple VXLANs on a single multicast group broken in kernel 4.1

[John Nielsen]

After updating the kernel on a KVM host, I discovered yesterday that I can no longer use more than one vxlan interface on a single multicast group/address. The interfaces can be created but only the first one may be brought up:

# ip link add vx100 type vxlan id 100 group 239.0.0.1 dev eth0
# ip link add vx200 type vxlan id 200 group 239.0.0.1 dev eth0
# ip link set up dev vx100
# ip link set up dev vx200
RTNETLINK answers: Address already in use

Using unique group addresses (one VNI per group) still works as expected. The problem is not present in kernel 4.0.5 but it is in kernel 4.1.0.

I suspect (but have not verified) that the regression was caused by this commit:
vxlan: Move socket initialization to within rtnl scope 56ef9c909b40483d2c8cb63fcbf83865f162d5ec

https://git.kernel.org/cgit/linux/kernel/git/stable/linux-stable.git/commit/drivers/net/vxlan.c?id=56ef9c909b40483d2c8cb63fcbf83865f162d5ec

I am working to verify that that is the culprit but I thought I’d see in the mean time if anyone else has encountered and/or fixed this, or can at least shed more light on the problem.

Thanks!

JN

Re: vxlan regression: multiple VXLANs on a single multicast group broken in kernel 4.1

[Marcelo Ricardo Leitner]

Em 25-08-2015 14:35, John Nielsen escreveu:
> After updating the kernel on a KVM host, I discovered yesterday that I can no longer use more than one vxlan interface on a single multicast group/address. The interfaces can be created but only the first one may be brought up:
>
> # ip link add vx100 type vxlan id 100 group 239.0.0.1 dev eth0
> # ip link add vx200 type vxlan id 200 group 239.0.0.1 dev eth0
> # ip link set up dev vx100
> # ip link set up dev vx200
> RTNETLINK answers: Address already in use
>
> Using unique group addresses (one VNI per group) still works as expected. The problem is not present in kernel 4.0.5 but it is in kernel 4.1.0.
>
> I suspect (but have not verified) that the regression was caused by this commit:
> vxlan: Move socket initialization to within rtnl scope 56ef9c909b40483d2c8cb63fcbf83865f162d5ec
>
> https://git.kernel.org/cgit/linux/kernel/git/stable/linux-stable.git/commit/drivers/net/vxlan.c?id=56ef9c909b40483d2c8cb63fcbf83865f162d5ec
>
> I am working to verify that that is the culprit but I thought I’d see in the mean time if anyone else has encountered and/or fixed this, or can at least shed more light on the problem.

Yes that commit is the culprit. Previously it was ignoring igmp_join 
errors, and now it's not but at the same time, it's trying to double 
join that group on the socket that is reused.

Will try to cook a patch here, thanks for the report!

If you need a hack for now, just ignore vxlan_igmp_join() return value 
at vxlan_open() and it should be fine.

   Marcelo

[PATCH net] vxlan: re-ignore EADDRINUSE from igmp_join

[Marcelo Ricardo Leitner]

Before 56ef9c909b40[1] it used to ignore all errors from igmp_join().
That commit enhanced that and made it error out whatever error happened
with igmp_join(), but that's not good because when using multicast
groups vxlan will try to join it multiple times if the socket is reused
and then the 2nd and further attempts will fail with EADDRINUSE.

As we don't track to which groups the socket is already subscribed, it's
okay to just ignore that error.

Fixes: 56ef9c909b40 ("vxlan: Move socket initialization to within rtnl scope")
Reported-by: John Nielsen <lists@jnielsen.net>
Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
---

John, please see how this goes for you. It worked in here.

Dave, please consider this for stable trees. At least 4.1 is affected.

Thanks.

 drivers/net/vxlan.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/drivers/net/vxlan.c b/drivers/net/vxlan.c
index 61b457b9ec00517037e4833790bea97ac53aa832..b0a2da5d4e57e602c915b65abd34a83939c4c473 100644
--- a/drivers/net/vxlan.c
+++ b/drivers/net/vxlan.c
@@ -2279,6 +2279,8 @@ static int vxlan_open(struct net_device *dev)
 
 	if (vxlan_addr_multicast(&vxlan->default_dst.remote_ip)) {
 		ret = vxlan_igmp_join(vxlan);
+		if (ret == -EADDRINUSE)
+			ret = 0;
 		if (ret) {
 			vxlan_sock_release(vs);
 			return ret;
-- 
2.4.3

Re: [PATCH net] vxlan: re-ignore EADDRINUSE from igmp_join

[David Miller]

From: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Date: Tue, 25 Aug 2015 20:22:35 -0300

> Before 56ef9c909b40[1] it used to ignore all errors from igmp_join().
> That commit enhanced that and made it error out whatever error happened
> with igmp_join(), but that's not good because when using multicast
> groups vxlan will try to join it multiple times if the socket is reused
> and then the 2nd and further attempts will fail with EADDRINUSE.
> 
> As we don't track to which groups the socket is already subscribed, it's
> okay to just ignore that error.
> 
> Fixes: 56ef9c909b40 ("vxlan: Move socket initialization to within rtnl scope")
> Reported-by: John Nielsen <lists@jnielsen.net>
> Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
> ---
> 
> John, please see how this goes for you. It worked in here.
> 
> Dave, please consider this for stable trees. At least 4.1 is affected.

Applied and queued up for -stable, thanks.

Re: [PATCH net] vxlan: re-ignore EADDRINUSE from igmp_join

[John Nielsen]

On Aug 25, 2015, at 5:22 PM, Marcelo Ricardo Leitner <marcelo.leitner@gmail.com> wrote:

> Before 56ef9c909b40[1] it used to ignore all errors from igmp_join().
> That commit enhanced that and made it error out whatever error happened
> with igmp_join(), but that's not good because when using multicast
> groups vxlan will try to join it multiple times if the socket is reused
> and then the 2nd and further attempts will fail with EADDRINUSE.
> 
> As we don't track to which groups the socket is already subscribed, it's
> okay to just ignore that error.
> 
> Fixes: 56ef9c909b40 ("vxlan: Move socket initialization to within rtnl scope")
> Reported-by: John Nielsen <lists@jnielsen.net>
> Signed-off-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
> ---
> 
> John, please see how this goes for you. It worked in here.

Confirmed working here. Thanks again!