On Wed, Oct 19, 2016 at 02:17:03PM +0200, Jiri Benc wrote:

On Tue, 18 Oct 2016 22:33:31 -0400, Jarod Wilson wrote:

quoted

--- a/drivers/net/vxlan.c
+++ b/drivers/net/vxlan.c

@@ -2367,43 +2367,31 @@ static void vxlan_set_multicast_list(struct net_device *dev)
 {
 }
 
-static int __vxlan_change_mtu(struct net_device *dev,
-			      struct net_device *lowerdev,
-			      struct vxlan_rdst *dst, int new_mtu, bool strict)
+static int vxlan_change_mtu(struct net_device *dev, int new_mtu)
 {
-	int max_mtu = IP_MAX_MTU;
-
-	if (lowerdev)
-		max_mtu = lowerdev->mtu;
+	struct vxlan_dev *vxlan = netdev_priv(dev);
+	struct vxlan_rdst *dst = &vxlan->default_dst;
+	struct net_device *lowerdev = __dev_get_by_index(vxlan->net,
+							 dst->remote_ifindex);
+	bool use_ipv6 = false;
 
 	if (dst->remote_ip.sa.sa_family == AF_INET6)
-		max_mtu -= VXLAN6_HEADROOM;
-	else
-		max_mtu -= VXLAN_HEADROOM;
-
-	if (new_mtu < 68)
-		return -EINVAL;
+		use_ipv6 = true;
 
-	if (new_mtu > max_mtu) {
-		if (strict)
+	/* We re-check this, because users *could* alter the mtu of the
+	 * lower device after we've initialized dev->max_mtu.
+	 */
+	if (lowerdev) {
+		dev->max_mtu = lowerdev->mtu -
+			       (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
+		if (new_mtu > dev->max_mtu)
 			return -EINVAL;
-
-		new_mtu = max_mtu;
 	}
 
 	dev->mtu = new_mtu;
 	return 0;
 }

Sorry for the silly question, how does the min_mtu and max_mtu stuff
works? I noticed your patches but haven't looked in depth into them.

When the ndo_change_mtu callback is defined, is the dev->min_mtu and
dev->max_mtu checked first and if the desired mtu is not within range,
ndo_change_mtu is not called?

Or does ndo_change_mtu override the checks?

The former. If the new value is outside min/max, ndo_change_mtu doesn't
get called, which is exactly the chicken and egg problem I introduced by
setting max_mtu to 1500 in ether_setup before having all drivers that call
ether_setup set a more appropriate max_mtu first. :\

In either case, the code does not look correct. In the first case,
increasing of lowerdev MTU wouldn't allow increasing of vxlan MTU
without deleting and recreating the vxlan interface. In the second
case, you're missing check against the min_mtu.

Okay, this sounds like a similar case to bridge that Sabrina pointed out.
Looks like virtual devices will need to just set no max_mtu directly (or
IP_MAX_MTU), and do dynamic checks in their ndo_change_mtu if they need to
compare against underlying devices on the fly.

...

quoted

@@ -2847,9 +2842,14 @@ static int vxlan_dev_configure(struct net *src_net, struct net_device *dev,
 	}
 
 	if (conf->mtu) {
-		err = __vxlan_change_mtu(dev, lowerdev, dst, conf->mtu, false);
-		if (err)
-			return err;
+		if (lowerdev)
+			dev->max_mtu = lowerdev->mtu;
+		dev->max_mtu -= (use_ipv6 ? VXLAN6_HEADROOM : VXLAN_HEADROOM);
+
+		dev->mtu = conf->mtu;
+
+		if (conf->mtu > dev->max_mtu)
+			dev->mtu = dev->max_mtu;
 	}

You removed the check for min_mtu but it's needed here. The conf->mtu
value comes from the user space and can be anything.

Hm. Not sure why I did that... Will put it back now...

-- 
Jarod Wilson
jarod@redhat.com