[PATCH 6.6.y] ipv4: start using dst_dev_rcu()

[PATCH 6.6.y] ipv4: start using dst_dev_rcu()

[Robert Garcia]

From: Eric Dumazet <edumazet@google.com>

[ Upstream commit 6ad8de3cefdb6ffa6708b21c567df0dbf82c43a8 ]

Change icmpv4_xrlim_allow(), ip_defrag() to prevent possible UAF.

Change ipmr_prepare_xmit(), ipmr_queue_fwd_xmit(), ip_mr_output(),
ipv4_neigh_lookup() to use lockdep enabled dst_dev_rcu().

Fixes: 4a6ce2b6f2ec ("net: introduce a new function dst_dev_put()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250828195823.3958522-9-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
[ Minor modifications made to adapt current code. ]
Signed-off-by: Robert Garcia <rob_garcia@163.com>
---
 net/ipv4/icmp.c        | 4 ++--
 net/ipv4/ip_fragment.c | 6 ++++--
 net/ipv4/ipmr.c        | 4 ++--
 net/ipv4/route.c       | 4 ++--
 4 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/net/ipv4/icmp.c b/net/ipv4/icmp.c
index 3fcf11f83d87..29a2162398e7 100644
--- a/net/ipv4/icmp.c
+++ b/net/ipv4/icmp.c
@@ -319,16 +319,16 @@ static bool icmpv4_xrlim_allow(struct net *net, struct rtable *rt,
 		return true;
 
 	/* No rate limit on loopback */
+	rcu_read_lock();
 	if (dst->dev && (dst->dev->flags&IFF_LOOPBACK))
 		goto out;
 
-	rcu_read_lock();
 	peer = inet_getpeer_v4(net->ipv4.peers, fl4->daddr,
 			       l3mdev_master_ifindex_rcu(dst->dev));
 	rc = inet_peer_xrlim_allow(peer,
 				   READ_ONCE(net->ipv4.sysctl_icmp_ratelimit));
-	rcu_read_unlock();
 out:
+	rcu_read_unlock();
 	if (!rc)
 		__ICMP_INC_STATS(net, ICMP_MIB_RATELIMITHOST);
 	else
diff --git a/net/ipv4/ip_fragment.c b/net/ipv4/ip_fragment.c
index 484edc8513e4..efc50d21d954 100644
--- a/net/ipv4/ip_fragment.c
+++ b/net/ipv4/ip_fragment.c
@@ -488,13 +488,15 @@ static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
 /* Process an incoming IP datagram fragment. */
 int ip_defrag(struct net *net, struct sk_buff *skb, u32 user)
 {
-	struct net_device *dev = skb->dev ? : skb_dst(skb)->dev;
-	int vif = l3mdev_master_ifindex_rcu(dev);
+	struct net_device *dev;
 	struct ipq *qp;
+	int vif;
 
 	__IP_INC_STATS(net, IPSTATS_MIB_REASMREQDS);
 
 	/* Lookup (or create) queue header */
+	dev = skb->dev ? : skb_dst_dev_rcu(skb);
+	vif = l3mdev_master_ifindex_rcu(dev);
 	qp = ip_find(net, ip_hdr(skb), user, vif);
 	if (qp) {
 		int ret;
diff --git a/net/ipv4/ipmr.c b/net/ipv4/ipmr.c
index af9412a507cf..948e826900fa 100644
--- a/net/ipv4/ipmr.c
+++ b/net/ipv4/ipmr.c
@@ -1905,7 +1905,7 @@ static void ipmr_queue_xmit(struct net *net, struct mr_table *mrt,
 		goto out_free;
 	}
 
-	encap += LL_RESERVED_SPACE(dev) + rt->dst.header_len;
+	encap += LL_RESERVED_SPACE(dst_dev_rcu(&rt->dst)) + rt->dst.header_len;
 
 	if (skb_cow(skb, encap)) {
 		ip_rt_put(rt);
@@ -1942,7 +1942,7 @@ static void ipmr_queue_xmit(struct net *net, struct mr_table *mrt,
 	 * result in receiving multiple packets.
 	 */
 	NF_HOOK(NFPROTO_IPV4, NF_INET_FORWARD,
-		net, NULL, skb, skb->dev, dev,
+		net, NULL, skb, skb->dev, dst_dev_rcu(&rt->dst),
 		ipmr_forward_finish);
 	return;
 
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index f134c59f839e..0ea017bcea47 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -416,11 +416,11 @@ static struct neighbour *ipv4_neigh_lookup(const struct dst_entry *dst,
 					   const void *daddr)
 {
 	const struct rtable *rt = container_of(dst, struct rtable, dst);
-	struct net_device *dev = dst->dev;
+	struct net_device *dev;
 	struct neighbour *n;
 
 	rcu_read_lock();
-
+	dev = dst_dev_rcu(dst);
 	if (likely(rt->rt_gw_family == AF_INET)) {
 		n = ip_neigh_gw4(dev, rt->rt_gw4);
 	} else if (rt->rt_gw_family == AF_INET6) {
-- 
2.34.1

Re: [PATCH 6.6.y] ipv4: start using dst_dev_rcu()

[Sasha Levin]

> [PATCH 6.6.y] ipv4: start using dst_dev_rcu()
> [ Upstream commit 6ad8de3cefdb6ffa6708b21c567df0dbf82c43a8 ]
> [ Minor modifications made to adapt current code. ]

Thanks for the backport. I think as proposed this introduces a bug: in 6.6.y
the predecessor a74fc62eec155c (which adds dst_dev_rcu()) is missing, so
skb_dst_dev_rcu() / dst_dev_rcu() ends up being called outside an RCU read-side
section here.

-- 
Thanks,
Sasha