Skip to contentSkip to conversation

IPv6: Change addrconf code to deal with link changes better

PATCH In review Brian Haley 3 participants 4 messages Sep 10, 2010 – Sep 14, 2010

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 IPv6: Change addrconf code to deal with link changes better

I recently started noticing that sometimes I wasn't seeing all the "normal" IPv6 packets (DAD, MLD, Router Solicit) when I was bringing-up interfaces. I found that sometimes the link on some NICs was going up-down-up like this:

ADDRCONF(NETDEV_UP): eth3: link is not ready
e1000e: eth3 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: None ADDRCONF(NETDEV_CHANGE): eth3: link becomes ready
(DAD queued here)
e1000e: eth3 NIC Link is Down
e1000e: eth3 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: None

This is all tcpdump showed on a remote system:

IP6 fe80::21f:29ff:fe59:fac9 > ff02::2: ICMP6, router solicitation, length 16 IP6 fe80::21f:29ff:fe59:fac9 > ff02::2: ICMP6, router solicitation, length 16

The addrconf code isn't written to handle this case, and won't restart DAD. With the patch below I get this (DAD_KICK was for debugging):

ADDRCONF(NETDEV_UP): eth6: link is not ready
e1000e: eth6 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: None ADDRCONF(NETDEV_CHANGE): eth6: link becomes ready DAD_KICK(fe80::21f:29ff:fe59:faca): eth6
e1000e: eth6 NIC Link is Down
ADDRCONF(NETDEV_CHANGE): eth6: link is not ready
e1000e: eth6 NIC Link is Up 1000 Mbps Full Duplex, Flow Control: None ADDRCONF(NETDEV_CHANGE): eth6: link becomes ready DAD_KICK(fe80::21f:29ff:fe59:faca): eth6

And corresponding tcpdump:

IP6 :: > ff02::1:ff59:faca: ICMP6, neighbor solicitation, who has fe80::21f:29ff:fe59:faca, length 24 IP6 :: > ff02::16: HBH ICMP6, multicast listener report v2, 1 group record(s), length 28 IP6 fe80::21f:29ff:fe59:faca > ff02::2: ICMP6, router solicitation, length 16

I also noticed that sometimes during these blips that addrconf_dad_kick() could be called twice, transmitting two DAD packets and resetting the timer another second into the future, so I introduced a "Need DAD" state to signify that DAD has not been started yet. With both of these applied the behavior is much better.

--

Change the IPv6 addrconf code to handle the case where the NIC link state goes UP-DOWN-UP by removing all the autoconfigured addresses when it goes down, so that when it comes back up they will get added again and DAD will be triggered. Also added a "Need DAD" state so that we don't accidentally start it twice for the same address, sending a duplicate packet and delaying when the address is available for use.

include/net/if_inet6.h +1 -0
index f95ff8d..57432f1 100644
@@ -33,6 +33,7 @@
 #ifdef __KERNEL__
 
 enum {
+	INET6_IFADDR_STATE_NEEDDAD,
 	INET6_IFADDR_STATE_DAD,
 	INET6_IFADDR_STATE_POSTDAD,
 	INET6_IFADDR_STATE_UP,
net/ipv6/addrconf.c +21 -6
index 5bc893e..fabafc7 100644
@@ -2517,13 +2517,27 @@ static int addrconf_notify(struct notifier_block *this, unsigned long event,
 			if (!idev && dev->mtu >= IPV6_MIN_MTU)
 				idev = ipv6_add_dev(dev);
 
-			if (idev) {
-				idev->if_flags |= IF_READY;
+			/*
+			 * Only start DAD if the device is ready, IF_READY
+			 * will have been set by ipv6_add_dev() if so.
+			 */
+			if (idev && idev->if_flags & IF_READY)
 				run_pending = 1;
-			}
 		} else {
 			if (!addrconf_qdisc_ok(dev)) {
-				/* device is still not ready. */
+				/*
+				 * Device is still not ready.  If we've already
+				 * configured addresses, remove them now as
+				 * we'll need to start DAD all over again.
+				 */
+				printk(KERN_INFO
+					"ADDRCONF(NETDEV_CHANGE): %s: "
+					"link is not ready\n",
+					dev->name);
+
+				if (idev && idev->if_flags & IF_READY)
+					addrconf_ifdown(dev, 0);
+
 				break;
 			}
 
@@ -2736,7 +2750,7 @@ static int addrconf_ifdown(struct net_device *dev, int how)
 
 			/* Flag it for later restoration when link comes up */
 			ifa->flags |= IFA_F_TENTATIVE;
-			ifa->state = INET6_IFADDR_STATE_DAD;
+			ifa->state = INET6_IFADDR_STATE_NEEDDAD;
 
 			write_unlock_bh(&idev->lock);
 
@@ -2847,6 +2861,7 @@ static void addrconf_dad_kick(struct inet6_ifaddr *ifp)
 		rand_num = net_random() % (idev->cnf.rtr_solicit_delay ? : 1);
 
 	ifp->probes = idev->cnf.dad_transmits;
+	ifp->state = INET6_IFADDR_STATE_DAD;
 	addrconf_mod_timer(ifp, AC_DAD, rand_num);
 }
 
@@ -2992,7 +3007,7 @@ static void addrconf_dad_run(struct inet6_dev *idev)
 	list_for_each_entry(ifp, &idev->addr_list, if_list) {
 		spin_lock(&ifp->lock);
 		if (ifp->flags & IFA_F_TENTATIVE &&
-		    ifp->state == INET6_IFADDR_STATE_DAD)
+		    ifp->state == INET6_IFADDR_STATE_NEEDDAD)
 			addrconf_dad_kick(ifp);
 		spin_unlock(&ifp->lock);
 	}
#1 lore raw
Quoted 8 lines · Brian Haley

[snipped]

Change the IPv6 addrconf code to handle the case where the NIC link state goes UP-DOWN-UP by removing all the autoconfigured addresses when it goes down, so that when it comes back up they will get added again and DAD will be triggered.

Surely this is going to result in me sending you lots of rude emails everytime cursing you when my SSH connections are terminated for no real reason on a link flap? Privacy addresses are handy on servers too such as DNS servers...

When on occasion I am forced to use Windows, those darn PuTTY disconnect popups I see each time I disconnect/connect from our wireless network nearly put axe swinging mood as it is. Please, spare me experiencing this pain under Linux...

Of course I could be talking nonsense, so ignore this email and keep up the good work :)

Cheers

Signature · 3 lines
Alexander Clouter
.sigmonster says: Cole's Law:
                  	Thinly sliced cabbage.
#2 lore raw
Quoted 9 lines · Alexander Clouter
Quoted 8 lines · Brian Haley

[snipped]

Change the IPv6 addrconf code to handle the case where the NIC link state goes UP-DOWN-UP by removing all the autoconfigured addresses when it goes down, so that when it comes back up they will get added again and DAD will be triggered.

Quoted 4 lines

Surely this is going to result in me sending you lots of rude emails everytime cursing you when my SSH connections are terminated for no real reason on a link flap? Privacy addresses are handy on servers too such as DNS servers...

No, you'd be sending me some rude emails, I didn't test that particular case, so Dave, don't apply this. I guess we've lived with it long enough the way it is that dropping a few packets at initialization isn't *that* bad. I'll see if I can figure out a better way in the meantime...

-Brian

maintainer
#3 lore raw

From: Brian Haley <[email protected]>
Date: Fri, 10 Sep 2010 21:10:29 -0400

Quoted 2 lines

No, you'd be sending me some rude emails, I didn't test that particular case, so Dave, don't apply this.

Ok.