Re: [PATCH] ipv6: addrconf: drop "BUG: " prefix from pr_warn()

[Ido Schimmel <idosch@nvidia.com>]

On Tue, Sep 08, 2026 at 12:31:43AM -0400, AnishMulay wrote:
> Hi Ido,
> 
> I ran into this same warning independently (different syzbot report,
> address 2001::fb rather than fc00::, extid 57f410c9a4f8d7a441d6) and
> ended up tracing it down before finding this thread.
> 
> The mechanism in my case: keep_addr_on_down is set, addrconf_ifdown()
> clears ifp->rt and deletes the route for the kept address but skips
> its notifier. On the following up, fixup_permanent_addr() allocates a
> new fib6_info and schedules addrconf_dad_work asynchronously instead
> of inserting the route itself. If a second addrconf_ifdown() for the
> same device runs before that work item gets to execute (I see this
> happen within a single "ip link set lo up" call, which drives both a
> NETDEV_UP and a NETDEV_CHANGE through addrconf_notify()), the fresh
> route gets torn down again with no notifier to re-arm anything. The
> work item then runs with ifp->rt NULL and nothing left to insert.
> 
> I have a patch that regenerates the route at that point instead of
> only warning, verified against a C reproducer translated from the
> syzbot repro (no warning, and the /128 route is present after the
> down/up cycle, across repeated runs). Since you mentioned you were
> looking into avoiding this state, I wanted to check before sending it,
> in case you already have something further along or a different angle
> on it. Happy to send what I have either way.

Does your reproducer rely on both keep_addr_on_down being set on the
loopback device and its MTU going below 1280?

The loopback device retains global addresses when this happens, unlike
any other device:

# sysctl -wq net.ipv6.conf.lo.keep_addr_on_down=1
# ip -6 address add 2001:db8:1::1/64 dev lo
# ip link set dev lo mtu 1000
# ip -6 address show dev lo
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 1000 qdisc noqueue state UNKNOWN group default qlen 1000
    inet6 2001:db8:1::1/64 scope global tentative
       valid_lft forever preferred_lft forever

# ip link add name dummy1 up type dummy
# sysctl -wq net.ipv6.conf.dummy1.keep_addr_on_down=1
# ip -6 address add 2001:db8:2::1/64 dev dummy1
# ip link set dev dummy1 mtu 1000
# ip -6 address show dev dummy1

And the local address (::1) is not restored when the MTU goes above the
minimum IPv6 MTU:

# ip link set dev lo mtu 2000
# ip -6 address show dev lo
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 2000 qdisc noqueue state UNKNOWN group default qlen 1000
    inet6 2001:db8:1::1/64 scope global tentative
       valid_lft forever preferred_lft forever

So, given that this state is quite broken and unlikely to be used by
anyone other than fuzzers, I would like to simply align the loopback
behavior with other devices and avoid keeping its addresses when the MTU
goes below the minimum:

diff --git a/net/ipv6/addrconf.c b/net/ipv6/addrconf.c
index 9d89be7e0544..2f7c872421b6 100644
--- a/net/ipv6/addrconf.c
+++ b/net/ipv6/addrconf.c
@@ -3908,9 +3908,12 @@ static int addrconf_ifdown(struct net_device *dev, bool unregister)
 	}
 
 	/* combine the user config with event to determine if permanent
-	 * addresses are to be removed from address hash table
+	 * addresses are to be removed from address hash table. IPv6 cannot
+	 * operate below IPV6_MIN_MTU, so treat that like disable_ipv6 and do
+	 * not keep addresses without their host routes.
 	 */
-	if (!unregister && !idev->cnf.disable_ipv6) {
+	if (!unregister && !idev->cnf.disable_ipv6 &&
+	    dev->mtu >= IPV6_MIN_MTU) {
 		/* aggregate the system setting and interface setting */
 		int _keep_addr = READ_ONCE(net->ipv6.devconf_all->keep_addr_on_down);

Regenerating the route in this case is more complexity for a case that
nobody is hitting other than fuzzers.