tcp: fix listening_get_next()

Alexey Vlasov found /proc/net/tcp could sometime loop and display
millions of sockets in LISTEN state.

In 2.6.29, when we converted TCP hash tables to RCU, we left two
sk_next() calls in listening_get_next().

We must instead use sk_nulls_next() to properly detect an end of chain.
Reported-by: NAlexey Vlasov <renton@renton.name>
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

net/ipv4/tcp_ipv4.c

@@ -2030,7 +2030,7 @@ static void *listening_get_next(struct seq_file *seq, void *cur)
 get_req:
 			req = icsk->icsk_accept_queue.listen_opt->syn_table[st->sbucket];
 		}
-		sk	  = sk_next(st->syn_wait_sk);
+		sk	  = sk_nulls_next(st->syn_wait_sk);
 		st->state = TCP_SEQ_STATE_LISTENING;
 		read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
 	} else {
@@ -2039,7 +2039,7 @@ static void *listening_get_next(struct seq_file *seq, void *cur)
 		if (reqsk_queue_len(&icsk->icsk_accept_queue))
 			goto start_req;
 		read_unlock_bh(&icsk->icsk_accept_queue.syn_wait_lock);
-		sk = sk_next(sk);
+		sk = sk_nulls_next(sk);
 	}
 get_sk:
 	sk_nulls_for_each_from(sk, node) {