svn commit: r307319 - in head/sys/netinet: . tcp_stacks

[Jonathan T. Looney]

Author: jtl
Date: Fri Oct 14 14:57:43 2016
New Revision: 307319
URL: https://svnweb.freebsd.org/changeset/base/307319

Log:
  The code currently resets the keepalive timer each time a packet is
  received on a TCP session that has entered the ESTABLISHED state. This
  results in a lot of calls to reset the keepalive timer.
  
  This patch changes the behavior so we set the keepalive timer for the
  keepalive idle time (TP_KEEPIDLE). When the keepalive timer fires, it will
  first check to see if the session has been idle for TP_KEEPIDLE ticks. If
  not, it will reschedule the keepalive timer for the time the session will
  have been idle for TP_KEEPIDLE ticks.
  
  For a session with regular communication, the keepalive timer should fire
  approximately once every TP_KEEPIDLE ticks. For sessions with irregular
  communication, the keepalive timer might fire more often. But, the
  disruption from a periodic keepalive timer should be less than the regular
  cost of resetting the keepalive timer on every packet.
  
  (FWIW, this change saved approximately 1.73% of the busy CPU cycles on a
  particular test system with a heavy TCP output load. Of course, the
  actual impact is very specific to the particular hardware and workload.)
  
  Reviewed by:	gallatin, rrs
  MFC after:	2 weeks
  Sponsored by:	Netflix
  Differential Revision:	https://reviews.freebsd.org/D8243

Modified:
  head/sys/netinet/tcp_input.c
  head/sys/netinet/tcp_stacks/fastpath.c
  head/sys/netinet/tcp_timer.c

Modified: head/sys/netinet/tcp_input.c
==============================================================================
--- head/sys/netinet/tcp_input.c	Fri Oct 14 12:03:04 2016	(r307318)
+++ head/sys/netinet/tcp_input.c	Fri Oct 14 14:57:43 2016	(r307319)
@@ -1565,8 +1565,6 @@ tcp_do_segment(struct mbuf *m, struct tc
 	 * validation to ignore broken/spoofed segs.
 	 */
 	tp->t_rcvtime = ticks;
-	if (TCPS_HAVEESTABLISHED(tp->t_state))
-		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
 
 	/*
 	 * Scale up the window into a 32-bit value.

Modified: head/sys/netinet/tcp_stacks/fastpath.c
==============================================================================
--- head/sys/netinet/tcp_stacks/fastpath.c	Fri Oct 14 12:03:04 2016	(r307318)
+++ head/sys/netinet/tcp_stacks/fastpath.c	Fri Oct 14 14:57:43 2016	(r307319)
@@ -1819,8 +1819,6 @@ tcp_do_segment_fastslow(struct mbuf *m, 
 	 * validation to ignore broken/spoofed segs.
 	 */
 	tp->t_rcvtime = ticks;
-	if (TCPS_HAVEESTABLISHED(tp->t_state))
-		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
 
 	/*
 	 * Unscale the window into a 32-bit value.
@@ -2266,8 +2264,6 @@ tcp_do_segment_fastack(struct mbuf *m, s
 	 * validation to ignore broken/spoofed segs.
 	 */
 	tp->t_rcvtime = ticks;
-	if (TCPS_HAVEESTABLISHED(tp->t_state))
-		tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp));
 
 	/*
 	 * Unscale the window into a 32-bit value.

Modified: head/sys/netinet/tcp_timer.c
==============================================================================
--- head/sys/netinet/tcp_timer.c	Fri Oct 14 12:03:04 2016	(r307318)
+++ head/sys/netinet/tcp_timer.c	Fri Oct 14 14:57:43 2016	(r307319)
@@ -468,6 +468,26 @@ tcp_timer_keep(void *xtp)
 	}
 	KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0,
 		("%s: tp %p tcpcb can't be stopped here", __func__, tp));
+
+	/*
+	 * Because we don't regularly reset the keepalive callout in
+	 * the ESTABLISHED state, it may be that we don't actually need
+	 * to send a keepalive yet. If that occurs, schedule another
+	 * call for the next time the keepalive timer might expire.
+	 */
+	if (TCPS_HAVEESTABLISHED(tp->t_state)) {
+		u_int idletime;
+
+		idletime = ticks - tp->t_rcvtime;
+		if (idletime < TP_KEEPIDLE(tp)) {
+			callout_reset(&tp->t_timers->tt_keep,
+			    TP_KEEPIDLE(tp) - idletime, tcp_timer_keep, tp);
+			INP_WUNLOCK(inp);
+			CURVNET_RESTORE();
+			return;
+		}
+	}
+
 	/*
 	 * Keep-alive timer went off; send something
 	 * or drop connection if idle for too long.