Merge branch 'tcp-backlog-processing-optims'

Eric Dumazet says:

====================
tcp: backlog processing optims

First patches are mostly preparing the ground for the last one.

Last patch of the series implements sort of ACK reduction
only for the cases a TCP receiver is under high stress,
which happens for high throughput flows.

This gives us a ~20% increase of single TCP flow (100Gbit -> 120Gbit)
====================

Link: https://lore.kernel.org/r/20230911170531.828100-1-edumazet@google.comSigned-off-by: Paolo Abeni <pabeni@redhat.com>

Documentation/networking/ip-sysctl.rst

@@ -745,6 +745,13 @@ tcp_comp_sack_nr - INTEGER
 
 	Default : 44
 
+tcp_backlog_ack_defer - BOOLEAN
+	If set, user thread processing socket backlog tries sending
+	one ACK for the whole queue. This helps to avoid potential
+	long latencies at end of a TCP socket syscall.
+
+	Default : true
+
 tcp_slow_start_after_idle - BOOLEAN
 	If set, provide RFC2861 behavior and time out the congestion
 	window after an idle period.  An idle period is defined at

include/linux/tcp.h

@@ -463,15 +463,17 @@ enum tsq_enum {
 	TCP_MTU_REDUCED_DEFERRED,  /* tcp_v{4|6}_err() could not call
 				    * tcp_v{4|6}_mtu_reduced()
 				    */
+	TCP_ACK_DEFERRED,	   /* TX pure ack is deferred */
 };
 
 enum tsq_flags {
-	TSQF_THROTTLED			= (1UL << TSQ_THROTTLED),
-	TSQF_QUEUED			= (1UL << TSQ_QUEUED),
-	TCPF_TSQ_DEFERRED		= (1UL << TCP_TSQ_DEFERRED),
-	TCPF_WRITE_TIMER_DEFERRED	= (1UL << TCP_WRITE_TIMER_DEFERRED),
-	TCPF_DELACK_TIMER_DEFERRED	= (1UL << TCP_DELACK_TIMER_DEFERRED),
-	TCPF_MTU_REDUCED_DEFERRED	= (1UL << TCP_MTU_REDUCED_DEFERRED),
+	TSQF_THROTTLED			= BIT(TSQ_THROTTLED),
+	TSQF_QUEUED			= BIT(TSQ_QUEUED),
+	TCPF_TSQ_DEFERRED		= BIT(TCP_TSQ_DEFERRED),
+	TCPF_WRITE_TIMER_DEFERRED	= BIT(TCP_WRITE_TIMER_DEFERRED),
+	TCPF_DELACK_TIMER_DEFERRED	= BIT(TCP_DELACK_TIMER_DEFERRED),
+	TCPF_MTU_REDUCED_DEFERRED	= BIT(TCP_MTU_REDUCED_DEFERRED),
+	TCPF_ACK_DEFERRED		= BIT(TCP_ACK_DEFERRED),
 };
 
 #define tcp_sk(ptr) container_of_const(ptr, struct tcp_sock, inet_conn.icsk_inet.sk)

include/net/netns/ipv4.h

@@ -132,6 +132,7 @@ struct netns_ipv4 {
 	u8 sysctl_tcp_syncookies;
 	u8 sysctl_tcp_migrate_req;
 	u8 sysctl_tcp_comp_sack_nr;
+	u8 sysctl_tcp_backlog_ack_defer;
 	int sysctl_tcp_reordering;
 	u8 sysctl_tcp_retries1;
 	u8 sysctl_tcp_retries2;

include/net/sock.h

@@ -1823,12 +1823,11 @@ static inline bool sock_owned_by_user_nocheck(const struct sock *sk)
 
 static inline void sock_release_ownership(struct sock *sk)
 {
-	if (sock_owned_by_user_nocheck(sk)) {
-		sk->sk_lock.owned = 0;
+	DEBUG_NET_WARN_ON_ONCE(!sock_owned_by_user_nocheck(sk));
+	sk->sk_lock.owned = 0;
 
-		/* The sk_lock has mutex_unlock() semantics: */
-		mutex_release(&sk->sk_lock.dep_map, _RET_IP_);
-	}
+	/* The sk_lock has mutex_unlock() semantics: */
+	mutex_release(&sk->sk_lock.dep_map, _RET_IP_);
 }
 
 /* no reclassification while locks are held */

net/core/sock.c

@@ -3001,6 +3001,9 @@ void __sk_flush_backlog(struct sock *sk)
 {
 	spin_lock_bh(&sk->sk_lock.slock);
 	__release_sock(sk);
+
+	if (sk->sk_prot->release_cb)
+		sk->sk_prot->release_cb(sk);
 	spin_unlock_bh(&sk->sk_lock.slock);
 }
 EXPORT_SYMBOL_GPL(__sk_flush_backlog);
@@ -3519,9 +3522,6 @@ void release_sock(struct sock *sk)
 	if (sk->sk_backlog.tail)
 		__release_sock(sk);
 
-	/* Warning : release_cb() might need to release sk ownership,
-	 * ie call sock_release_ownership(sk) before us.
-	 */
 	if (sk->sk_prot->release_cb)
 		sk->sk_prot->release_cb(sk);
 

net/ipv4/sysctl_net_ipv4.c

@@ -1366,6 +1366,15 @@ static struct ctl_table ipv4_net_table[] = {
 		.proc_handler	= proc_dou8vec_minmax,
 		.extra1		= SYSCTL_ZERO,
 	},
+	{
+		.procname	= "tcp_backlog_ack_defer",
+		.data		= &init_net.ipv4.sysctl_tcp_backlog_ack_defer,
+		.maxlen		= sizeof(u8),
+		.mode		= 0644,
+		.proc_handler	= proc_dou8vec_minmax,
+		.extra1		= SYSCTL_ZERO,
+		.extra2		= SYSCTL_ONE,
+	},
 	{
 		.procname       = "tcp_reflect_tos",
 		.data           = &init_net.ipv4.sysctl_tcp_reflect_tos,

net/ipv4/tcp_input.c

@@ -5553,6 +5553,14 @@ static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
 	    tcp_in_quickack_mode(sk) ||
 	    /* Protocol state mandates a one-time immediate ACK */
 	    inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOW) {
+		/* If we are running from __release_sock() in user context,
+		 * Defer the ack until tcp_release_cb().
+		 */
+		if (sock_owned_by_user_nocheck(sk) &&
+		    READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_backlog_ack_defer)) {
+			set_bit(TCP_ACK_DEFERRED, &sk->sk_tsq_flags);
+			return;
+		}
 send_now:
 		tcp_send_ack(sk);
 		return;

net/ipv4/tcp_ipv4.c

@@ -3263,6 +3263,7 @@ static int __net_init tcp_sk_init(struct net *net)
 	net->ipv4.sysctl_tcp_comp_sack_delay_ns = NSEC_PER_MSEC;
 	net->ipv4.sysctl_tcp_comp_sack_slack_ns = 100 * NSEC_PER_USEC;
 	net->ipv4.sysctl_tcp_comp_sack_nr = 44;
+	net->ipv4.sysctl_tcp_backlog_ack_defer = 1;
 	net->ipv4.sysctl_tcp_fastopen = TFO_CLIENT_ENABLE;
 	net->ipv4.sysctl_tcp_fastopen_blackhole_timeout = 0;
 	atomic_set(&net->ipv4.tfo_active_disable_times, 0);

net/ipv4/tcp_output.c

@@ -1077,7 +1077,8 @@ static void tcp_tasklet_func(struct tasklet_struct *t)
 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
 			  TCPF_WRITE_TIMER_DEFERRED |	\
 			  TCPF_DELACK_TIMER_DEFERRED |	\
-			  TCPF_MTU_REDUCED_DEFERRED)
+			  TCPF_MTU_REDUCED_DEFERRED |	\
+			  TCPF_ACK_DEFERRED)
 /**
  * tcp_release_cb - tcp release_sock() callback
  * @sk: socket
@@ -1101,16 +1102,6 @@ void tcp_release_cb(struct sock *sk)
 		tcp_tsq_write(sk);
 		__sock_put(sk);
 	}
-	/* Here begins the tricky part :
-	 * We are called from release_sock() with :
-	 * 1) BH disabled
-	 * 2) sk_lock.slock spinlock held
-	 * 3) socket owned by us (sk->sk_lock.owned == 1)
-	 *
-	 * But following code is meant to be called from BH handlers,
-	 * so we should keep BH disabled, but early release socket ownership
-	 */
-	sock_release_ownership(sk);
 
 	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 		tcp_write_timer_handler(sk);
@@ -1124,6 +1115,8 @@ void tcp_release_cb(struct sock *sk)
 		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 		__sock_put(sk);
 	}
+	if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
+		tcp_send_ack(sk);
 }
 EXPORT_SYMBOL(tcp_release_cb);