diff --git a/sys/kern/subr_trap.c b/sys/kern/subr_trap.c
index 6160b3fcdf3..42ce9e7988b 100644
--- a/sys/kern/subr_trap.c
+++ b/sys/kern/subr_trap.c
@@ -76,6 +76,8 @@
 #include <sys/epoch.h>
 #endif
 
+void	(*tcp_hpts_softclock)(void);
+
 /*
  * Define the code needed before returning to user mode, for trap and
  * syscall.
@@ -127,16 +129,18 @@ userret(struct thread *td, struct trapframe *frame)
 	if (PMC_THREAD_HAS_SAMPLES(td))
 		PMC_CALL_HOOK(td, PMC_FN_THR_USERRET, NULL);
 #endif
-#ifdef TCPHPTS
 	/*
-	 * @gallatin is adament that this needs to go here, I
-	 * am not so sure. Running hpts is a lot like
-	 * a lro_flush() that happens while a user process
-	 * is running. But he may know best so I will go
-	 * with his view of accounting. :-)
+	 * Calling tcp_hpts_softclock() here allows us to avoid frequent,
+	 * expensive callouts that trash the cache and lead to a much higher
+	 * number of interrupts and context switches.  Testing on busy web
+	 * servers at Netflix has shown that this improves CPU use by 7% over
+	 * relying only on callouts to drive HPTS, and also results in idle
+	 * power savings on mostly idle servers.
+	 * This was inspired by the paper "Soft Timers: Efficient Microsecond
+	 * Software Timer Support for Network Processing"
+	 * by Mohit Aron and Peter Druschel.
 	 */
-	tcp_run_hpts();
-#endif
+	tcp_hpts_softclock();
 	/*
 	 * Let the scheduler adjust our priority etc.
 	 */
diff --git a/sys/netinet/tcp_hpts.h b/sys/netinet/tcp_hpts.h
index 8ca21daf60d..7eb1b2e08cb 100644
--- a/sys/netinet/tcp_hpts.h
+++ b/sys/netinet/tcp_hpts.h
@@ -152,7 +152,6 @@ void __tcp_set_hpts(struct tcpcb *tp, int32_t line);
 
 void tcp_set_inp_to_drop(struct inpcb *inp, uint16_t reason);
 
-extern void (*tcp_hpts_softclock)(void);
 void tcp_lro_hpts_init(void);
 
 extern int32_t tcp_min_hptsi_time;
diff --git a/sys/netinet/tcp_lro.c b/sys/netinet/tcp_lro.c
index 255e543ae21..921d28f8251 100644
--- a/sys/netinet/tcp_lro.c
+++ b/sys/netinet/tcp_lro.c
@@ -89,7 +89,6 @@ SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 
 long tcplro_stacks_wanting_mbufq;
 int	(*tcp_lro_flush_tcphpts)(struct lro_ctrl *lc, struct lro_entry *le);
-void	(*tcp_hpts_softclock)(void);
 
 counter_u64_t tcp_inp_lro_direct_queue;
 counter_u64_t tcp_inp_lro_wokeup_queue;
@@ -1262,8 +1261,7 @@ tcp_lro_flush_all(struct lro_ctrl *lc)
 done:
 	/* flush active streams */
 	tcp_lro_rx_done(lc);
-	if (tcp_hpts_softclock != NULL)
-		tcp_hpts_softclock();
+	tcp_hpts_softclock();
 	lc->lro_mbuf_count = 0;
 }
 
diff --git a/sys/sys/systm.h b/sys/sys/systm.h
index 2367a7880cf..0e90b6062b7 100644
--- a/sys/sys/systm.h
+++ b/sys/sys/systm.h
@@ -380,6 +380,12 @@ void	cpu_et_frequency(struct eventtimer *et, uint64_t newfreq);
 extern int	cpu_disable_c2_sleep;
 extern int	cpu_disable_c3_sleep;
 
+extern void	(*tcp_hpts_softclock)(void);
+#define	tcp_hpts_softclock()	do {					\
+		if (tcp_hpts_softclock != NULL)				\
+			tcp_hpts_softclock();				\
+} while (0)
+
 char	*kern_getenv(const char *name);
 void	freeenv(char *env);
 int	getenv_int(const char *name, int *data);