Commit 8fc8911b authored by Paolo Abeni's avatar Paolo Abeni
Browse files

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.com


Signed-off-by: default avatarPaolo Abeni <pabeni@redhat.com>
parents cd8bae85 133c4c0d

Documentation/networking/ip-sysctl.rst

+7 −0
Original line number Diff line number Diff line
@@ -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

+8 −6
Original line number Diff line number Diff line
@@ -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

+1 −0
Original line number Diff line number Diff line
@@ -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

+4 −5
Original line number Diff line number Diff line
@@ -1823,13 +1823,12 @@ 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)) {

	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_);

}

}



/* no reclassification while locks are held */

static inline bool sock_allow_reclassification(const struct sock *csk)

net/core/sock.c

+3 −3
Original line number Diff line number Diff line
@@ -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);