[PATCH net] udp: fix a potential panic in first_packet_length()

From: Eric Dumazet <edumazet@google.com>

first_packet_length() is called from udp_ioctl()

udp_ioctl(), as its name suggests, is used by UDP protocols,
but is also used by L2TP :(

We shall call udp_rmem_release() only for UDP variants.

Thanks to Andrey and syzkaller team for providing the report
and a nice reproducer.

Fixes: 7c13f97ffde63 ("udp: do fwd memory scheduling on dequeue")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Andrey Konovalov <redacted>
---
 net/ipv4/udp.c |    6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index 8aab7d78d25bc6eaa42dcc960cdbd5086f614cad..7c0807ee82cec6ca8c856da14fa6109dfdf27868 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c

@@ -1376,7 +1376,11 @@ static int first_packet_length(struct sock *sk)
 		kfree_skb(skb);
 	}
 	res = skb ? skb->len : -1;
-	if (total)
+	/* udp_ioctl() can be used by UDP/UDPLite, but also L2TP.
+	 * We only need to call udp_rmem_release() for UDP sockets.
+	 * L2TP does have a proper skb destructor invoked at kfree_skb() time.
+	 */
+	if (total && sk->sk_prot->memory_allocated == &udp_memory_allocated)
 		udp_rmem_release(sk, total, 1);
 	spin_unlock_bh(&rcvq->lock);
 	return res;