tcp: fix SO_RCVLOWAT hangs with fat skbs

[ Upstream commit 24adbc1 ]

We autotune rcvbuf whenever SO_RCVLOWAT is set to account for 100%
overhead in tcp_set_rcvlowat()

This works well when skb->len/skb->truesize ratio is bigger than 0.5

But if we receive packets with small MSS, we can end up in a situation
where not enough bytes are available in the receive queue to satisfy
RCVLOWAT setting.
As our sk_rcvbuf limit is hit, we send zero windows in ACK packets,
preventing remote peer from sending more data.

Even autotuning does not help, because it only triggers at the time
user process drains the queue. If no EPOLLIN is generated, this
can not happen.

Note poll() has a similar issue, after commit
c700448 ("tcp: Respect SO_RCVLOWAT in tcp_poll().")

Fixes: 03f45c8 ("tcp: avoid extra wakeups for SO_RCVLOWAT users")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

include/net/tcp.h

@@ -1373,6 +1373,19 @@ static inline int tcp_full_space(const struct sock *sk)
 	return tcp_win_from_space(sk, sk->sk_rcvbuf);
 }
 
+/* We provision sk_rcvbuf around 200% of sk_rcvlowat.
+ * If 87.5 % (7/8) of the space has been consumed, we want to override
+ * SO_RCVLOWAT constraint, since we are receiving skbs with too small
+ * len/truesize ratio.
+ */
+static inline bool tcp_rmem_pressure(const struct sock *sk)
+{
+	int rcvbuf = READ_ONCE(sk->sk_rcvbuf);
+	int threshold = rcvbuf - (rcvbuf >> 3);
+
+	return atomic_read(&sk->sk_rmem_alloc) > threshold;
+}
+
 extern void tcp_openreq_init_rwin(struct request_sock *req,
 				  const struct sock *sk_listener,
 				  const struct dst_entry *dst);

net/ipv4/tcp.c

@@ -488,9 +488,17 @@ static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 					  int target, struct sock *sk)
 {
-	return (READ_ONCE(tp->rcv_nxt) - tp->copied_seq >= target) ||
-		(sk->sk_prot->stream_memory_read ?
-		sk->sk_prot->stream_memory_read(sk) : false);
+	int avail = READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq);
+
+	if (avail > 0) {
+		if (avail >= target)
+			return true;
+		if (tcp_rmem_pressure(sk))
+			return true;
+	}
+	if (sk->sk_prot->stream_memory_read)
+		return sk->sk_prot->stream_memory_read(sk);
+	return false;
 }
 
 /*

net/ipv4/tcp_input.c

@@ -4683,7 +4683,8 @@ void tcp_data_ready(struct sock *sk)
 	const struct tcp_sock *tp = tcp_sk(sk);
 	int avail = tp->rcv_nxt - tp->copied_seq;
 
-	if (avail < sk->sk_rcvlowat && !sock_flag(sk, SOCK_DONE))
+	if (avail < sk->sk_rcvlowat && !tcp_rmem_pressure(sk) &&
+	    !sock_flag(sk, SOCK_DONE))
 		return;
 
 	sk->sk_data_ready(sk);