Skip to contentSkip to conversation

IPV6 address management

RFC In review Stephen Hemminger 4 participants 15 messages Jan 8, 2009 – Jan 9, 2009

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

I notice that with IPV6 when interface is brought down, all addresses are removed. This looks like intentional behaviour but means the administrator (or daemon) must reinsert addresses. Since this is not what the users expect it does lead to confusion.

Maybe the kernel should do this?

maintainer
#1 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 09:34:30 -0800

Quoted 6 lines

I notice that with IPV6 when interface is brought down, all addresses are removed. This looks like intentional behaviour but means the administrator (or daemon) must reinsert addresses. Since this is not what the users expect it does lead to confusion.

Maybe the kernel should do this?

After we've had this behavior for as long as we have, I doubt we can safely change it.

I do agree it is unintuitive, and it also disagrees with ipv4 as you already know.

There are several crucial areas where our behavior deviates between ipv4 and ipv6, including this case. And as more people start making real use of ipv6 we will see these complaints pop up over and over again from users.

#2 lore raw

What about this?

If it works (still testing), I'll submit it.

---
 Documentation/networking/ip-sysctl.txt |    8 ++++++
 include/linux/ipv6.h                   |    2 +
 net/ipv6/addrconf.c                    |   38 +++++++++++++++++++++-----------
 3 files changed, 35 insertions(+), 13 deletions(-)
Documentation/networking/ip-sysctl.txt +8 -0
index e2b2894..fa05b06 100644
@@ -937,6 +937,14 @@ dad_transmits - INTEGER
 	The amount of Duplicate Address Detection probes to send.
 	Default: 1
 	
+delete_address_ifdown - BOOLEAN
+	Delete all addresses on device down
+	Default: TRUE
+
+	When network device is disabled: 
+	  FALSE - only delete temporary addresses (similar to IPV4)
+	  TRUE  - delete all addresses
+
 forwarding - BOOLEAN
 	Configure interface-specific Host/Router behaviour.  
 
include/linux/ipv6.h +2 -0
index 641e026..b2d219a 100644
@@ -166,6 +166,7 @@ struct ipv6_devconf {
 #endif
 	__s32		disable_ipv6;
 	__s32		accept_dad;
+	__s32		delete_address_ifdown;
 	void		*sysctl;
 };
 #endif
@@ -200,6 +201,7 @@ enum {
 	DEVCONF_MC_FORWARDING,
 	DEVCONF_DISABLE_IPV6,
 	DEVCONF_ACCEPT_DAD,
+	DEVCONF_DELETE_ADDRESS_IFDOWN,
 	DEVCONF_MAX
 };
 
net/ipv6/addrconf.c +25 -13
index 2635fa2..9133c20 100644
@@ -186,6 +186,7 @@ static struct ipv6_devconf ipv6_devconf __read_mostly = {
 	.accept_source_route	= 0,	/* we do not accept RH0 by default. */
 	.disable_ipv6		= 0,
 	.accept_dad		= 1,
+	.delete_address_ifdown  = 1,
 };
 
 static struct ipv6_devconf ipv6_devconf_dflt __read_mostly = {
@@ -220,6 +221,7 @@ static struct ipv6_devconf ipv6_devconf_dflt __read_mostly = {
 	.accept_source_route	= 0,	/* we do not accept RH0 by default. */
 	.disable_ipv6		= 0,
 	.accept_dad		= 1,
+	.delete_address_ifdown  = 1,
 };
 
 /* IPv6 Wildcard Address and Loopback Address defined by RFC2553 */
@@ -2673,19 +2675,20 @@ static int addrconf_ifdown(struct net_device *dev, int how)
 		write_lock_bh(&idev->lock);
 	}
 #endif
-	while ((ifa = idev->addr_list) != NULL) {
-		idev->addr_list = ifa->if_next;
-		ifa->if_next = NULL;
-		ifa->dead = 1;
-		addrconf_del_timer(ifa);
-		write_unlock_bh(&idev->lock);
-
-		__ipv6_ifa_notify(RTM_DELADDR, ifa);
-		atomic_notifier_call_chain(&inet6addr_chain, NETDEV_DOWN, ifa);
-		in6_ifa_put(ifa);
-
-		write_lock_bh(&idev->lock);
-	}
+	if (idev->cnf.delete_address_ifdown)
+		while ((ifa = idev->addr_list) != NULL) {
+			idev->addr_list = ifa->if_next;
+			ifa->if_next = NULL;
+			ifa->dead = 1;
+			addrconf_del_timer(ifa);
+			write_unlock_bh(&idev->lock);
+
+			__ipv6_ifa_notify(RTM_DELADDR, ifa);
+			atomic_notifier_call_chain(&inet6addr_chain,
+						   NETDEV_DOWN, ifa);
+			in6_ifa_put(ifa);
+			write_lock_bh(&idev->lock);
+		}
 	write_unlock_bh(&idev->lock);
 
 	/* Step 5: Discard multicast list */
@@ -3693,6 +3696,7 @@ static inline void ipv6_store_devconf(struct ipv6_devconf *cnf,
 #endif
 	array[DEVCONF_DISABLE_IPV6] = cnf->disable_ipv6;
 	array[DEVCONF_ACCEPT_DAD] = cnf->accept_dad;
+	array[DEVCONF_DELETE_ADDRESS_IFDOWN] = cnf->delete_address_ifdown;
 }
 
 static inline size_t inet6_if_nlmsg_size(void)
@@ -4268,6 +4272,14 @@ static struct addrconf_sysctl_table
 			.proc_handler	=	&proc_dointvec,
 		},
 		{
+			.ctl_name	=	CTL_UNNUMBERED,
+			.procname	=	"delete_address_ifdown",
+			.data		=	&ipv6_devconf.delete_address_ifdown,
+			.maxlen		=	sizeof(int),
+			.mode		=	0644,
+			.proc_handler	=	&proc_dointvec,
+		},
+		{
 			.ctl_name	=	0,	/* sentinel */
 		}
 	},
Signature · 1 line
1.5.6.3
maintainer
#3 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 12:12:20 -0800

Quoted 3 lines

What about this?

If it works (still testing), I'll submit it.

So what is your plan?

Make the routing daemons depend upon the non-default behavior in order to act correctly?

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

I disagree with both plans, and with that the facility is basically useless.

We absolutely have to live with the behavior we have now, and for a long time if not forever.

#4 lore raw
Quoted 11 lines · David Miller

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 12:12:20 -0800

Quoted 3 lines

What about this?

If it works (still testing), I'll submit it.

So what is your plan?

Make the routing daemons depend upon the non-default behavior in order to act correctly?

They could check and take corrective action or tell the user to set it one way or another. Robust and complex apps already have to check all sorts of things anyway.

Quoted 9 lines

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

I disagree with both plans, and with that the facility is basically useless.

We absolutely have to live with the behavior we have now, and for a long time if not forever.

It could be argued we need to have the default behaviour stay the same, but that's no reason to keep things difficult for people/programs flexible enough to deal with change.

Thanks,
Ben

Signature · 2 lines
Ben Greear <[email protected]>
Candela Technologies Inc  http://www.candelatech.com
#5 lore raw
In reply to David Miller
Quoted 11 lines · On Thu, 08 Jan 2009 12:58:30 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 12:12:20 -0800

Quoted 3 lines

What about this?

If it works (still testing), I'll submit it.

So what is your plan?

Make the routing daemons depend upon the non-default behavior in order to act correctly?

routing daemons don't deal with addresses directly now, and getting them to do it would be painful.

Quoted 3 lines

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

No plan to ever change the default. Just ship with sysctl.conf setting.

Quoted 5 lines

I disagree with both plans, and with that the facility is basically useless.

We absolutely have to live with the behavior we have now, and for a long time if not forever.

maintainer
#6 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:44:02 -0800

Quoted 8 lines
Quoted 3 lines · On Thu, 08 Jan 2009 12:58:30 -0800 (PST) David Miller <[email protected]>…

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

No plan to ever change the default. Just ship with sysctl.conf setting.

If the distributions all ship with the sysctl changed to the non-default, our "default" is pretty meaningless wouldn't you say?

#7 lore raw
Quoted 9 lines · On Thu, 08 Jan 2009 13:51:25 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:44:02 -0800

Quoted 6 lines
Quoted 3 lines · On Thu, 08 Jan 2009 12:58:30 -0800 (PST) David Miller <[email protected]>…

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

Quoted 6 lines
Quoted 2 lines

No plan to ever change the default. Just ship with sysctl.conf setting.

If the distributions all ship with the sysctl changed to the non-default, our "default" is pretty meaningless wouldn't you say?

It seems the only logical way to undo a poor choice in the original design

maintainer
#8 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:56:14 -0800

Quoted 15 lines
Quoted 12 lines · On Thu, 08 Jan 2009 13:51:25 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:44:02 -0800

Quoted 8 lines
Quoted 3 lines · On Thu, 08 Jan 2009 12:58:30 -0800 (PST) David Miller <[email protected]>…

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

No plan to ever change the default. Just ship with sysctl.conf setting.

Quoted 6 lines
Quoted 3 lines

If the distributions all ship with the sysctl changed to the non-default, our "default" is pretty meaningless wouldn't you say?

It seems the only logical way to undo a poor choice in the original design

If the dists can do it so unilaterally, why can't we?

Everything about these proposals is a contradiction. That's why I don't like them at all.

I say we keep the behavior, we don't change or break anything, and people need to learn how to cope with it.

#9 lore raw
Quoted 25 lines · On Thu, 08 Jan 2009 13:58:19 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:56:14 -0800

Quoted 15 lines
Quoted 12 lines · On Thu, 08 Jan 2009 13:51:25 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 13:44:02 -0800

Quoted 8 lines
Quoted 3 lines · On Thu, 08 Jan 2009 12:58:30 -0800 (PST) David Miller <[email protected]>…

Or is it to gradually get people to use the non-default (via distribution sysctl settings etc.) and eventually make it the default?

No plan to ever change the default. Just ship with sysctl.conf setting.

Quoted 6 lines
Quoted 3 lines

If the distributions all ship with the sysctl changed to the non-default, our "default" is pretty meaningless wouldn't you say?

It seems the only logical way to undo a poor choice in the original design

Quoted 7 lines

If the dists can do it so unilaterally, why can't we?

Everything about these proposals is a contradiction. That's why I don't like them at all.

I say we keep the behavior, we don't change or break anything, and people need to learn how to cope with it.

Fine, it won't be the first or last vendor specific kernel patch.

maintainer
#10 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 14:01:22 -0800

Quoted 1 line

Fine, it won't be the first or last vendor specific kernel patch.

You're just supporting my argument even more.

If all the dists ship it and turn it on, then logically we should include it and have it on by default.

But you know we can't do that.

Therefore, it's detrimental for the dists to ship it too because it does break things.

I know you want to shove this in, via some avenue, but it is really so undesirable to cope with existing behavior?

#11 lore raw
Quoted 10 lines · On Thu, 08 Jan 2009 14:03:39 -0800 (PST) David Miller <[email protected]>…

From: Stephen Hemminger <[email protected]>
Date: Thu, 8 Jan 2009 14:01:22 -0800

Quoted 1 line

Fine, it won't be the first or last vendor specific kernel patch.

You're just supporting my argument even more.

If all the dists ship it and turn it on, then logically we should include it and have it on by default.

Quoted 7 lines

But you know we can't do that.

Therefore, it's detrimental for the dists to ship it too because it does break things.

I know you want to shove this in, via some avenue, but it is really so undesirable to cope with existing behavior?

Reading back in the archives, this has come up before. The real problem is that when link goes down and comes back the address needs to be reverified for DAD. Maybe instead of deleting addresses they should be moved to another list for restart.

It is a real problem for routing daemons like hostapd and zebra since the kernel address deletion looks a lot like a user address deletion. Since Vyatta has the luxury of a configuration/scripting layer, this means the state space is contained (we know what the address was), I'll just hack around it there for now until a better solution arises.

#12 lore raw

PATCH 1/1 ipv6: address preservation on link down

In reply to David Miller

When an interface goes down, IPV6 deletes all addresses unlike IPV4. This is a problem since it can break routing protocols and other applications. It looks like this was done to handle DAD, but is a big stick solution when other code is possible.

The following patch changes the behaviour to delete link local addresses but keep all configured addresses and restart DAD when interface comes back up.

See also: https://kerneltrap.org/mailarchive/linux-netdev/2008/3/25/1252064/thread

net/ipv6/addrconf.c +20 -12
@@ -2678,18 +2678,26 @@ static int addrconf_ifdown(struct net_de
 		write_lock_bh(&idev->lock);
 	}
 #endif
-	while ((ifa = idev->addr_list) != NULL) {
-		idev->addr_list = ifa->if_next;
-		ifa->if_next = NULL;
-		ifa->dead = 1;
-		addrconf_del_timer(ifa);
-		write_unlock_bh(&idev->lock);
-
-		__ipv6_ifa_notify(RTM_DELADDR, ifa);
-		atomic_notifier_call_chain(&inet6addr_chain, NETDEV_DOWN, ifa);
-		in6_ifa_put(ifa);
-
-		write_lock_bh(&idev->lock);
+	/* clear regular address list on removal */
+	bifa = &idev->addr_list;
+	while ((ifa = *bifa) != NULL) {
+		if (how || (ipv6_addr_type(&ifa->addr) & IPV6_ADDR_LINKLOCAL)) {
+			*bifa = ifa->if_next;
+			ifa->if_next = NULL;
+			ifa->dead = 1;
+			addrconf_del_timer(ifa);
+			write_unlock_bh(&idev->lock);
+
+			__ipv6_ifa_notify(RTM_DELADDR, ifa);
+			atomic_notifier_call_chain(&inet6addr_chain, NETDEV_DOWN, ifa);
+			in6_ifa_put(ifa);
+
+			write_lock_bh(&idev->lock);
+		} else {
+			/* Force duplicate address detection when link resumes */
+			ifa->flags |= IFA_F_TENTATIVE;
+			bifa = &ifa->if_next;
+		}
 	}
 	write_unlock_bh(&idev->lock);
 
#13 lore raw

Re: [PATCH] ipv6: address preservation on link down

Quoted 8 lines · you

When an interface goes down, IPV6 deletes all addresses unlike IPV4. This is a problem since it can break routing protocols and other applications. It looks like this was done to handle DAD, but is a big stick solution when other code is possible.

The following patch changes the behaviour to delete link local addresses but keep all configured addresses and restart DAD when interface comes back up.

Let alone the backward compatibility problem for a minute. I have two _other_ problems with this:

1/ If DAD should fail when the interface is brought back up, wouldn't all the EAU-64 autoconfigured address become invalid, rather than just the link-local one?

2/ In some case the link-local address is assigned by userland, no different from the other ones. At least miredo is behaving this way.

Signature · 2 lines
Rémi Denis-Courmont
Maemo Software, Nokia Devices R&D
#14 lore raw

Re: [PATCH] ipv6: address preservation on link down

Quoted 13 lines · On Fri, 9 Jan 2009 10:02:06 +0200 "Rémi Denis-Courmont"…
Quoted 8 lines · you

When an interface goes down, IPV6 deletes all addresses unlike IPV4. This is a problem since it can break routing protocols and other applications. It looks like this was done to handle DAD, but is a big stick solution when other code is possible.

The following patch changes the behaviour to delete link local addresses but keep all configured addresses and restart DAD when interface comes back up.

Let alone the backward compatibility problem for a minute. I have two _other_ problems with this:

Quoted 7 lines

1/ If DAD should fail when the interface is brought back up, wouldn't all the EAU-64 autoconfigured address become invalid, rather than just the link-local one?

2/ In some case the link-local address is assigned by userland, no different from the other ones. At least miredo is behaving this way.

1: The address DAD path is same as the initial setup. I.e. it is same as if
the address was added manually (after ifup)
2: New patch honors the permanent flag. Another alternative would be to add a new
IFA_F_DYNAMIC flag

Updated patch that combines earlier efforts..
---

By default, IPV6 deletes all addresses when an interface is deconfigured. This behaviour is different the IPV4 and confusing, so make it optional. Add a new sysctl /proc/sys/net/ipv6/conf/ethX/address_flush that controls this.

The new behaviour that happens only if address_flush is set to zero is:
* all addresses are flushed on device deletion
* linklocal addresses are flushed unless the permanent flag is set
* other addresses are saved but duplicate address detection is restarted unless nodad flag has been set.

---
 Documentation/networking/ip-sysctl.txt |    9 +++++
 include/linux/ipv6.h                   |    2 +
 net/ipv6/addrconf.c                    |   50 ++++++++++++++++++++++++++-------
 3 files changed, 51 insertions(+), 10 deletions(-)
Documentation/networking/ip-sysctl.txt +9 -0
@@ -925,6 +925,15 @@ accept_source_route - INTEGER
 
 	Default: 0
 
+address_flush - BOOLEAN
+	Flush all addresses when link goes down
+	Default: TRUE
+
+	When network device is set to admin down:
+	  TRUE  - Delete all addresses
+	  FALSE - Only addresses that are temporary or linklocal and not permanent
+	          are deleted. Other addresses restart DAD (if configured).
+
 autoconf - BOOLEAN
 	Autoconfigure addresses using Prefix Information in Router 
 	Advertisements.
include/linux/ipv6.h +2 -0
@@ -166,6 +166,7 @@ struct ipv6_devconf {
 #endif
 	__s32		disable_ipv6;
 	__s32		accept_dad;
+	__s32		address_flush;
 	void		*sysctl;
 };
 #endif
@@ -200,6 +201,7 @@ enum {
 	DEVCONF_MC_FORWARDING,
 	DEVCONF_DISABLE_IPV6,
 	DEVCONF_ACCEPT_DAD,
+	DEVCONF_ADDRESS_FLUSH,
 	DEVCONF_MAX
 };
 
net/ipv6/addrconf.c +40 -10

82 lines not on this page · raw message