1 files changed, 23 insertions, 3 deletions

diff --git a/usr/src/uts/common/disp/disp.c b/usr/src/uts/common/disp/disp.c
index 0c2c0b4993..5f9c2c68a2 100644
--- a/usr/src/uts/common/disp/disp.c
+++ b/usr/src/uts/common/disp/disp.c
@@ -23,6 +23,10 @@
  * Use is subject to license terms.
  */
 
+/*
+ * Copyright (c) 2012, Joyent, Inc. All rights reserved.
+ */
+
 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
 /*	  All Rights Reserved  	*/
 
@@ -105,7 +109,7 @@ static void	cpu_resched(cpu_t *cp, pri_t tpri);
 /*
  * If this is set, only interrupt threads will cause kernel preemptions.
  * This is done by changing the value of kpreemptpri.  kpreemptpri
- * will either be the max sysclass pri + 1 or the min interrupt pri.
+ * will either be the max sysclass pri or the min interrupt pri.
  */
 int	only_intr_kpreempt;
 
@@ -252,7 +256,23 @@ dispinit(void)
 				maxglobpri = cl_maxglobpri;
 		}
 	}
-	kpreemptpri = (pri_t)v.v_maxsyspri + 1;
+
+	/*
+	 * Historically, kpreemptpri was set to v_maxsyspri + 1 -- which is
+	 * to say, maxclsyspri + 1.  However, over time, the system has used
+	 * more and more asynchronous kernel threads, with an increasing number
+	 * of these doing work on direct behalf of higher-level software (e.g.,
+	 * network processing).  This has led to potential priority inversions:
+	 * threads doing low-priority lengthy kernel work can effectively
+	 * delay kernel-level processing of higher-priority data. To minimize
+	 * such inversions, we set kpreemptpri to be v_maxsyspri; anything in
+	 * the kernel that runs at maxclsyspri will therefore induce kernel
+	 * preemption, and this priority should be used if/when an asynchronous
+	 * thread (or, as is often the case, task queue) is performing a task
+	 * on behalf of higher-level software (or any task that is otherwise
+	 * latency-sensitve).
+	 */
+	kpreemptpri = (pri_t)v.v_maxsyspri;
 	if (kpqpri == KPQPRI)
 		kpqpri = kpreemptpri;
 
@@ -2258,7 +2278,7 @@ disp_getbest(disp_t *dp)
 		 * placed earlier.
 		 */
 		if (tcp == NULL ||
-		    pri >= minclsyspri ||
+		    (pri >= minclsyspri && tp->t_procp == &p0) ||
 		    tp->t_cpu != tcp)
 			break;