tcp: avoid premature drops in tcp_add_backlog()

While testing TCP performance with latest trees,
I saw suspect SOCKET_BACKLOG drops.

tcp_add_backlog() computes its limit with :

    limit = (u32)READ_ONCE(sk->sk_rcvbuf) +
            (u32)(READ_ONCE(sk->sk_sndbuf) >> 1);
    limit += 64 * 1024;

This does not take into account that sk->sk_backlog.len
is reset only at the very end of __release_sock().

Both sk->sk_backlog.len and sk->sk_rmem_alloc could reach
sk_rcvbuf in normal conditions.

We should double sk->sk_rcvbuf contribution in the formula
to absorb bubbles in the backlog, which happen more often
for very fast flows.

This change maintains decent protection against abuses.

Fixes: c377411 ("net: sk_add_backlog() take rmem_alloc into account")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240423125620.3309458-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

net/ipv4/tcp_ipv4.c

@@ -1995,7 +1995,7 @@ int tcp_v4_early_demux(struct sk_buff *skb)
 bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
 		     enum skb_drop_reason *reason)
 {
-	u32 limit, tail_gso_size, tail_gso_segs;
+	u32 tail_gso_size, tail_gso_segs;
 	struct skb_shared_info *shinfo;
 	const struct tcphdr *th;
 	struct tcphdr *thtail;
@@ -2004,6 +2004,7 @@ bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
 	bool fragstolen;
 	u32 gso_segs;
 	u32 gso_size;
+	u64 limit;
 	int delta;
 
 	/* In case all data was pulled from skb frags (in __pskb_pull_tail()),
@@ -2099,14 +2100,22 @@ bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb,
 	__skb_push(skb, hdrlen);
 
 no_coalesce:
-	limit = (u32)READ_ONCE(sk->sk_rcvbuf) + (u32)(READ_ONCE(sk->sk_sndbuf) >> 1);
+	/* sk->sk_backlog.len is reset only at the end of __release_sock().
+	 * Both sk->sk_backlog.len and sk->sk_rmem_alloc could reach
+	 * sk_rcvbuf in normal conditions.
+	 */
+	limit = ((u64)READ_ONCE(sk->sk_rcvbuf)) << 1;
+
+	limit += ((u32)READ_ONCE(sk->sk_sndbuf)) >> 1;
 
 	/* Only socket owner can try to collapse/prune rx queues
 	 * to reduce memory overhead, so add a little headroom here.
 	 * Few sockets backlog are possibly concurrently non empty.
 	 */
 	limit += 64 * 1024;
 
+	limit = min_t(u64, limit, UINT_MAX);
+
 	if (unlikely(sk_add_backlog(sk, skb, limit))) {
 		bh_unlock_sock(sk);
 		*reason = SKB_DROP_REASON_SOCKET_BACKLOG;