mptcp: correct refcnt for sk in tcp_v{4,6}_rcv

When we receive an skb in tcp_vX_rcv if the sock is hold by user we copy
the sk inside the skb before putting the skb inside the backlog, the sk
will be used after from mptcp_backlog_rcv. If the sk refcnt is not done
properly the sk read from mptcp_backlog_rcv could point to another sk
because it may have been freed and reused in the mean time.

In mptcp_backlog_rcv, the call to skb_oprhan will call the destructor
that we set for the skb and decrement the sk refcnt. Note that we call
skb_oprhan after the increase of sk_refcnt. This ensures that we never
free the sk too quickly.

Adaptation of commit 952bd28 (mptcp: correct refcnt for sk in
tcp_v{4,6}_rcv) for mptcp_trunk.

Fixes: Zero-day-bug
Signed-off-by: Gregory Detal <gregory.detal@tessares.net>
Signed-off-by: Benjamin Hesmans <benjamin.hesmans@tessares.net>
Signed-off-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Signed-off-by: Christoph Paasch <cpaasch@apple.com>

include/net/mptcp.h

@@ -825,6 +825,7 @@ void mptcp_sub_close_wq(struct work_struct *work);
 void mptcp_sub_close(struct sock *sk, unsigned long delay);
 struct sock *mptcp_select_ack_sock(const struct sock *meta_sk);
 void mptcp_fallback_meta_sk(struct sock *meta_sk);
+void mptcp_prepare_for_backlog(struct sock *sk, struct sk_buff *skb);
 int mptcp_backlog_rcv(struct sock *meta_sk, struct sk_buff *skb);
 void mptcp_ack_handler(struct timer_list *t);
 bool mptcp_check_rtt(const struct tcp_sock *tp, int time);
@@ -1395,6 +1396,7 @@ static inline bool mptcp_fallback_infinite(const struct sock *sk, int flag)
 	return false;
 }
 static inline void mptcp_init_mp_opt(const struct mptcp_options_received *mopt) {}
+static inline void mptcp_prepare_for_backlog(struct sock *sk, struct sk_buff *skb) {}
 static inline bool mptcp_check_rtt(const struct tcp_sock *tp, int time)
 {
 	return false;

net/ipv4/tcp_ipv4.c

@@ -1814,7 +1814,7 @@ int tcp_v4_rcv(struct sk_buff *skb)
 
 		bh_lock_sock_nested(meta_sk);
 		if (sock_owned_by_user(meta_sk))
-			skb->sk = sk;
+			mptcp_prepare_for_backlog(sk, skb);
 	} else {
 		meta_sk = sk;
 		bh_lock_sock_nested(sk);

net/ipv6/tcp_ipv6.c

@@ -1608,7 +1608,7 @@ static int tcp_v6_rcv(struct sk_buff *skb)
 
 		bh_lock_sock_nested(meta_sk);
 		if (sock_owned_by_user(meta_sk))
-			skb->sk = sk;
+			mptcp_prepare_for_backlog(sk, skb);
 	} else {
 		meta_sk = sk;
 		bh_lock_sock_nested(sk);

net/mptcp/mptcp_ctrl.c

@@ -988,6 +988,16 @@ static void mptcp_sub_inherit_sockopts(const struct sock *meta_sk, struct sock *
 	inet_sk(sub_sk)->recverr = 0;
 }
 
+void mptcp_prepare_for_backlog(struct sock *sk, struct sk_buff *skb)
+{
+	/* In case of success (in mptcp_backlog_rcv) and error (in kfree_skb) of
+	 * sk_add_backlog, we will decrement the sk refcount.
+	 */
+	sock_hold(sk);
+	skb->sk = sk;
+	skb->destructor = sock_efree;
+}
+
 int mptcp_backlog_rcv(struct sock *meta_sk, struct sk_buff *skb)
 {
 	/* skb-sk may be NULL if we receive a packet immediatly after the
@@ -996,13 +1006,17 @@ int mptcp_backlog_rcv(struct sock *meta_sk, struct sk_buff *skb)
 	struct sock *sk = skb->sk ? skb->sk : meta_sk;
 	int ret = 0;
 
-	skb->sk = NULL;
-
 	if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt))) {
 		kfree_skb(skb);
 		return 0;
 	}
 
+	/* Decrement sk refcnt when calling the skb destructor.
+	 * Refcnt is incremented and skb destructor is set in tcp_v{4,6}_rcv via
+	 * mptcp_prepare_for_backlog() here above.
+	 */
+	skb_orphan(skb);
+
 	if (sk->sk_family == AF_INET)
 		ret = tcp_v4_do_rcv(sk, skb);
 #if IS_ENABLED(CONFIG_IPV6)