Merge branch 'tcp-receive-side-improvements'

Eric Dumazet says:

====================
tcp: receive side improvements

We have set tcp_rmem[2] to 15 MB for about 8 years at Google,
but had some issues for high speed flows on very small RTT.

TCP rx autotuning has a tendency to overestimate the RTT,
thus tp->rcvq_space.space and sk->sk_rcvbuf.

This makes TCP receive queues much bigger than necessary,
to a point cpu caches are evicted before application can
copy the data, on cpus using DDIO.

This series aims to fix this.

- First patch adds tcp_rcvbuf_grow() tracepoint, which was very
  convenient to study the various issues fixed in this series.

- Seven patches fix receiver autotune issues.

- Two patches fix sender side issues.

- Final patch increases tcp_rmem[2] so that TCP speed over WAN
  can meet modern needs.

Tested on a 200Gbit NIC, average max throughput of a single flow:

Before:
 73593 Mbit.

After:
 122514 Mbit.
====================

Link: https://patch.msgid.link/20250513193919.1089692-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

7 files changed, 134 insertions, 66 deletions

diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst
index b43222ee57cf9..0f1251cce3149 100644
--- a/Documentation/networking/ip-sysctl.rst
+++ b/Documentation/networking/ip-sysctl.rst
@@ -735,7 +735,7 @@ tcp_rmem - vector of 3 INTEGERs: min, default, max
 	net.core.rmem_max.  Calling setsockopt() with SO_RCVBUF disables
 	automatic tuning of that socket's receive buffer size, in which
 	case this value is ignored.
-	Default: between 131072 and 6MB, depending on RAM size.
+	Default: between 131072 and 32MB, depending on RAM size.
 
 tcp_sack - BOOLEAN
 	Enable select acknowledgments (SACKS).
@@ -1099,7 +1099,7 @@ tcp_limit_output_bytes - INTEGER
 	limits the number of bytes on qdisc or device to reduce artificial
 	RTT/cwnd and reduce bufferbloat.
 
-	Default: 1048576 (16 * 65536)
+	Default: 4194304 (4 MB)
 
 tcp_challenge_ack_limit - INTEGER
 	Limits number of Challenge ACK sent per second, as recommended
diff --git a/include/linux/tcp.h b/include/linux/tcp.h
index a8af71623ba7c..29f59d50dc73f 100644
--- a/include/linux/tcp.h
+++ b/include/linux/tcp.h
@@ -340,7 +340,7 @@ struct tcp_sock {
 	} rcv_rtt_est;
 /* Receiver queue space */
 	struct {
-		u32	space;
+		int	space;
 		u32	seq;
 		u64	time;
 	} rcvq_space;
diff --git a/include/trace/events/tcp.h b/include/trace/events/tcp.h
index 53e878fa14d14..006c2116c8f61 100644
--- a/include/trace/events/tcp.h
+++ b/include/trace/events/tcp.h
@@ -213,6 +213,79 @@ DEFINE_EVENT(tcp_event_sk, tcp_rcv_space_adjust,
 	TP_ARGS(sk)
 );
 
+TRACE_EVENT(tcp_rcvbuf_grow,
+
+	TP_PROTO(struct sock *sk, int time),
+
+	TP_ARGS(sk, time),
+
+	TP_STRUCT__entry(
+		__field(int, time)
+		__field(__u32, rtt_us)
+		__field(__u32, copied)
+		__field(__u32, inq)
+		__field(__u32, space)
+		__field(__u32, ooo_space)
+		__field(__u32, rcvbuf)
+		__field(__u8, scaling_ratio)
+		__field(__u16, sport)
+		__field(__u16, dport)
+		__field(__u16, family)
+		__array(__u8, saddr, 4)
+		__array(__u8, daddr, 4)
+		__array(__u8, saddr_v6, 16)
+		__array(__u8, daddr_v6, 16)
+		__field(const void *, skaddr)
+		__field(__u64, sock_cookie)
+	),
+
+	TP_fast_assign(
+		struct inet_sock *inet = inet_sk(sk);
+		struct tcp_sock *tp = tcp_sk(sk);
+		__be32 *p32;
+
+		__entry->time = time;
+		__entry->rtt_us = tp->rcv_rtt_est.rtt_us >> 3;
+		__entry->copied = tp->copied_seq - tp->rcvq_space.seq;
+		__entry->inq = tp->rcv_nxt - tp->copied_seq;
+		__entry->space = tp->rcvq_space.space;
+		__entry->ooo_space = RB_EMPTY_ROOT(&tp->out_of_order_queue) ? 0 :
+				     TCP_SKB_CB(tp->ooo_last_skb)->end_seq -
+				     tp->rcv_nxt;
+
+		__entry->rcvbuf = sk->sk_rcvbuf;
+		__entry->scaling_ratio = tp->scaling_ratio;
+		__entry->sport = ntohs(inet->inet_sport);
+		__entry->dport = ntohs(inet->inet_dport);
+		__entry->family = sk->sk_family;
+
+		p32 = (__be32 *) __entry->saddr;
+		*p32 = inet->inet_saddr;
+
+		p32 = (__be32 *) __entry->daddr;
+		*p32 = inet->inet_daddr;
+
+		TP_STORE_ADDRS(__entry, inet->inet_saddr, inet->inet_daddr,
+			       sk->sk_v6_rcv_saddr, sk->sk_v6_daddr);
+
+		__entry->skaddr = sk;
+		__entry->sock_cookie = sock_gen_cookie(sk);
+	),
+
+	TP_printk("time=%u rtt_us=%u copied=%u inq=%u space=%u ooo=%u scaling_ratio=%u rcvbuf=%u "
+		  "family=%s sport=%hu dport=%hu saddr=%pI4 daddr=%pI4 "
+		  "saddrv6=%pI6c daddrv6=%pI6c skaddr=%p sock_cookie=%llx",
+		  __entry->time, __entry->rtt_us, __entry->copied,
+		  __entry->inq, __entry->space, __entry->ooo_space,
+		  __entry->scaling_ratio, __entry->rcvbuf,
+		  show_family_name(__entry->family),
+		  __entry->sport, __entry->dport,
+		  __entry->saddr, __entry->daddr,
+		  __entry->saddr_v6, __entry->daddr_v6,
+		  __entry->skaddr,
+		  __entry->sock_cookie)
+);
+
 TRACE_EVENT(tcp_retransmit_synack,
 
 	TP_PROTO(const struct sock *sk, const struct request_sock *req),
diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 0ae265d39184e..b7b6ab41b496f 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -5231,7 +5231,7 @@ void __init tcp_init(void)
 	/* Set per-socket limits to no more than 1/128 the pressure threshold */
 	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
 	max_wshare = min(4UL*1024*1024, limit);
-	max_rshare = min(6UL*1024*1024, limit);
+	max_rshare = min(32UL*1024*1024, limit);
 
 	init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
 	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index a35018e2d0ba2..8ec92dec321a9 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -664,10 +664,12 @@ EXPORT_IPV6_MOD(tcp_initialize_rcv_mss);
  */
 static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 {
-	u32 new_sample = tp->rcv_rtt_est.rtt_us;
-	long m = sample;
+	u32 new_sample, old_sample = tp->rcv_rtt_est.rtt_us;
+	long m = sample << 3;
 
-	if (new_sample != 0) {
+	if (old_sample == 0 || m < old_sample) {
+		new_sample = m;
+	} else {
 		/* If we sample in larger samples in the non-timestamp
 		 * case, we could grossly overestimate the RTT especially
 		 * with chatty applications or bulk transfer apps which
@@ -678,17 +680,12 @@ static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 		 * else with timestamps disabled convergence takes too
 		 * long.
 		 */
-		if (!win_dep) {
-			m -= (new_sample >> 3);
-			new_sample += m;
-		} else {
-			m <<= 3;
-			if (m < new_sample)
-				new_sample = m;
-		}
-	} else {
-		/* No previous measure. */
-		new_sample = m << 3;
+		if (win_dep)
+			return;
+		/* Do not use this sample if receive queue is not empty. */
+		if (tp->rcv_nxt != tp->copied_seq)
+			return;
+		new_sample = old_sample - (old_sample >> 3) + sample;
 	}
 
 	tp->rcv_rtt_est.rtt_us = new_sample;
@@ -712,7 +709,7 @@ new_measure:
 	tp->rcv_rtt_est.time = tp->tcp_mstamp;
 }
 
-static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
+static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp, u32 min_delta)
 {
 	u32 delta, delta_us;
 
@@ -722,7 +719,7 @@ static s32 tcp_rtt_tsopt_us(const struct tcp_sock *tp)
 
 	if (likely(delta < INT_MAX / (USEC_PER_SEC / TCP_TS_HZ))) {
 		if (!delta)
-			delta = 1;
+			delta = min_delta;
 		delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ);
 		return delta_us;
 	}
@@ -740,13 +737,39 @@ static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 
 	if (TCP_SKB_CB(skb)->end_seq -
 	    TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss) {
-		s32 delta = tcp_rtt_tsopt_us(tp);
+		s32 delta = tcp_rtt_tsopt_us(tp, 0);
 
-		if (delta >= 0)
+		if (delta > 0)
 			tcp_rcv_rtt_update(tp, delta, 0);
 	}
 }
 
+static void tcp_rcvbuf_grow(struct sock *sk)
+{
+	const struct net *net = sock_net(sk);
+	struct tcp_sock *tp = tcp_sk(sk);
+	int rcvwin, rcvbuf, cap;
+
+	if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) ||
+	    (sk->sk_userlocks & SOCK_RCVBUF_LOCK))
+		return;
+
+	/* slow start: allow the sender to double its rate. */
+	rcvwin = tp->rcvq_space.space << 1;
+
+	if (!RB_EMPTY_ROOT(&tp->out_of_order_queue))
+		rcvwin += TCP_SKB_CB(tp->ooo_last_skb)->end_seq - tp->rcv_nxt;
+
+	cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
+
+	rcvbuf = min_t(u32, tcp_space_from_win(sk, rcvwin), cap);
+	if (rcvbuf > sk->sk_rcvbuf) {
+		WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
+		/* Make the window clamp follow along.  */
+		WRITE_ONCE(tp->window_clamp,
+			   tcp_win_from_space(sk, rcvbuf));
+	}
+}
 /*
  * This function should be called every time data is copied to user space.
  * It calculates the appropriate TCP receive buffer space.
@@ -754,8 +777,7 @@ static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 void tcp_rcv_space_adjust(struct sock *sk)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
-	u32 copied;
-	int time;
+	int time, inq, copied;
 
 	trace_tcp_rcv_space_adjust(sk);
 
@@ -766,45 +788,18 @@ void tcp_rcv_space_adjust(struct sock *sk)
 
 	/* Number of bytes copied to user in last RTT */
 	copied = tp->copied_seq - tp->rcvq_space.seq;
+	/* Number of bytes in receive queue. */
+	inq = tp->rcv_nxt - tp->copied_seq;
+	copied -= inq;
 	if (copied <= tp->rcvq_space.space)
 		goto new_measure;
 
-	/* A bit of theory :
-	 * copied = bytes received in previous RTT, our base window
-	 * To cope with packet losses, we need a 2x factor
-	 * To cope with slow start, and sender growing its cwin by 100 %
-	 * every RTT, we need a 4x factor, because the ACK we are sending
-	 * now is for the next RTT, not the current one :
-	 * <prev RTT . ><current RTT .. ><next RTT .... >
-	 */
-
-	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
-	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
-		u64 rcvwin, grow;
-		int rcvbuf;
-
-		/* minimal window to cope with packet losses, assuming
-		 * steady state. Add some cushion because of small variations.
-		 */
-		rcvwin = ((u64)copied << 1) + 16 * tp->advmss;
+	trace_tcp_rcvbuf_grow(sk, time);
 
-		/* Accommodate for sender rate increase (eg. slow start) */
-		grow = rcvwin * (copied - tp->rcvq_space.space);
-		do_div(grow, tp->rcvq_space.space);
-		rcvwin += (grow << 1);
-
-		rcvbuf = min_t(u64, tcp_space_from_win(sk, rcvwin),
-			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
-		if (rcvbuf > sk->sk_rcvbuf) {
-			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
-
-			/* Make the window clamp follow along.  */
-			WRITE_ONCE(tp->window_clamp,
-				   tcp_win_from_space(sk, rcvbuf));
-		}
-	}
 	tp->rcvq_space.space = copied;
 
+	tcp_rcvbuf_grow(sk);
+
 new_measure:
 	tp->rcvq_space.seq = tp->copied_seq;
 	tp->rcvq_space.time = tp->tcp_mstamp;
@@ -3226,7 +3221,7 @@ static bool tcp_ack_update_rtt(struct sock *sk, const int flag,
 	 */
 	if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp &&
 	    tp->rx_opt.rcv_tsecr && flag & FLAG_ACKED)
-		seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp);
+		seq_rtt_us = ca_rtt_us = tcp_rtt_tsopt_us(tp, 1);
 
 	rs->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet (or -1) */
 	if (seq_rtt_us < 0)
@@ -5173,6 +5168,7 @@ end:
 		skb_condense(skb);
 		skb_set_owner_r(skb, sk);
 	}
+	tcp_rcvbuf_grow(sk);
 }
 
 static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb,
@@ -6873,6 +6869,9 @@ tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 		if (!tp->srtt_us)
 			tcp_synack_rtt_meas(sk, req);
 
+		if (tp->rx_opt.tstamp_ok)
+			tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
+
 		if (req) {
 			tcp_rcv_synrecv_state_fastopen(sk);
 		} else {
@@ -6898,9 +6897,6 @@ tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb)
 		tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
 		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
 
-		if (tp->rx_opt.tstamp_ok)
-			tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
-
 		if (!inet_csk(sk)->icsk_ca_ops->cong_control)
 			tcp_update_pacing_rate(sk);
 
diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c
index d5b5c32115d2e..6a14f9e6fef64 100644
--- a/net/ipv4/tcp_ipv4.c
+++ b/net/ipv4/tcp_ipv4.c
@@ -3495,8 +3495,8 @@ static int __net_init tcp_sk_init(struct net *net)
 	 * which are too large can cause TCP streams to be bursty.
 	 */
 	net->ipv4.sysctl_tcp_tso_win_divisor = 3;
-	/* Default TSQ limit of 16 TSO segments */
-	net->ipv4.sysctl_tcp_limit_output_bytes = 16 * 65536;
+	/* Default TSQ limit of 4 MB */
+	net->ipv4.sysctl_tcp_limit_output_bytes = 4 << 20;
 
 	/* rfc5961 challenge ack rate limiting, per net-ns, disabled by default. */
 	net->ipv4.sysctl_tcp_challenge_ack_limit = INT_MAX;
diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c
index 13295a59d22e6..3ac8d2d17e1ff 100644
--- a/net/ipv4/tcp_output.c
+++ b/net/ipv4/tcp_output.c
@@ -2619,9 +2619,8 @@ static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
 	limit = max_t(unsigned long,
 		      2 * skb->truesize,
 		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
-	if (sk->sk_pacing_status == SK_PACING_NONE)
-		limit = min_t(unsigned long, limit,
-			      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
+	limit = min_t(unsigned long, limit,
+		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
 	limit <<= factor;
 
 	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&