Index: kern/sched_ule.c
===================================================================
RCS file: /home/ncvs/src/sys/kern/sched_ule.c,v
retrieving revision 1.223
diff -u -r1.223 sched_ule.c
--- kern/sched_ule.c	23 Jan 2008 03:10:18 -0000	1.223
+++ kern/sched_ule.c	23 Jan 2008 23:04:51 -0000
@@ -240,6 +240,7 @@
 static int pick_pri = 1;
 static int affinity;
 static int tryself = 1;
+static int oldtryself = 0;
 static int steal_htt = 1;
 static int steal_idle = 1;
 static int steal_thresh = 2;
@@ -677,6 +678,7 @@
 		 * IPI the target cpu to force it to reschedule with the new
 		 * workload.
 		 */
+		pcpu_find(TDQ_ID(low))->pc_curthread->td_owepreempt = 1;
 		ipi_selected(1 << TDQ_ID(low), IPI_PREEMPT);
 	}
 	tdq_unlock_pair(high, low);
@@ -845,7 +847,7 @@
 	if (pri > preempt_thresh)
 		return;
 sendipi:
-	ctd->td_flags |= TDF_NEEDRESCHED;
+	ctd->td_owepreempt = 1;
 	ipi_selected(1 << cpu, IPI_PREEMPT);
 }
 
@@ -1118,6 +1120,11 @@
 		    curthread->td_priority);
 		return (self);
 	}
+	if (oldtryself && pri < curthread->td_priority) {
+		CTR1(KTR_ULE, "tryself %d",
+		    curthread->td_priority);
+		return (self);
+	}
 	/*
  	 * Now search for the cpu running the lowest priority thread with
 	 * the least load.
@@ -2676,6 +2683,7 @@
 SYSCTL_INT(_kern_sched, OID_AUTO, affinity, CTLFLAG_RW, &affinity, 0,
     "Number of hz ticks to keep thread affinity for");
 SYSCTL_INT(_kern_sched, OID_AUTO, tryself, CTLFLAG_RW, &tryself, 0, "");
+SYSCTL_INT(_kern_sched, OID_AUTO, oldtryself, CTLFLAG_RW, &oldtryself, 0, "");
 SYSCTL_INT(_kern_sched, OID_AUTO, balance, CTLFLAG_RW, &rebalance, 0,
     "Enables the long-term load balancer");
 SYSCTL_INT(_kern_sched, OID_AUTO, balance_interval, CTLFLAG_RW,
Index: amd64/amd64/mp_machdep.c
===================================================================
RCS file: /home/ncvs/src/sys/amd64/amd64/mp_machdep.c,v
retrieving revision 1.287
diff -u -r1.287 mp_machdep.c
--- amd64/amd64/mp_machdep.c	2 Aug 2007 21:17:58 -0000	1.287
+++ amd64/amd64/mp_machdep.c	23 Jan 2008 23:04:51 -0000
@@ -964,9 +964,8 @@
 	if (ipi_bitmap & (1 << IPI_PREEMPT)) {
 		struct thread *running_thread = curthread;
 		thread_lock(running_thread);
-		if (running_thread->td_critnest > 1) 
-			running_thread->td_owepreempt = 1;
-		else 		
+		if (running_thread->td_critnest == 1 &&
+		    running_thread->td_owepreempt) 
 			mi_switch(SW_INVOL | SW_PREEMPT, NULL);
 		thread_unlock(running_thread);
 	}