runtime: faster entersyscall/exitsyscall

Replace cas with xadd in scheduler.
Suggested by Dmitriy in last code review.
Verified with Promela model.

When there's actual contention for the atomic word,
this avoids the looping that compare-and-swap requires.

benchmark                            old ns/op    new ns/op    delta
runtime_test.BenchmarkSyscall               32           26  -17.08%
runtime_test.BenchmarkSyscall-2            155           59  -61.81%
runtime_test.BenchmarkSyscall-3            112           52  -52.95%
runtime_test.BenchmarkSyscall-4             94           48  -48.57%
runtime_test.BenchmarkSyscallWork          871          872   +0.11%
runtime_test.BenchmarkSyscallWork-2        481          477   -0.83%
runtime_test.BenchmarkSyscallWork-3        338          335   -0.89%
runtime_test.BenchmarkSyscallWork-4        263          256   -2.66%

R=golang-dev, iant
CC=golang-dev
https://golang.org/cl/4800047

src/pkg/runtime/proc.c

@@ -773,7 +773,7 @@ runtime·gosched(void)
 void
 runtime·entersyscall(void)
 {
-	uint32 v, w;
+	uint32 v;
 
 	if(runtime·sched.predawn)
 		return;
@@ -796,24 +796,14 @@ runtime·entersyscall(void)
 	//	mcpu--
 	//	gwait not true
 	//	waitstop && mcpu <= mcpumax not true
-	// If we can do the same with a single atomic read/write,
+	// If we can do the same with a single atomic add,
 	// then we can skip the locks.
-	for(;;) {
-		v = runtime·sched.atomic;
-		if(atomic_gwaiting(v))
-			break;
-		if(atomic_waitstop(v) && atomic_mcpu(v)-1 <= atomic_mcpumax(v))
-			break;
-		w = v;
-		w += (-1<<mcpuShift);
-		if(runtime·cas(&runtime·sched.atomic, v, w))
+	v = runtime·xadd(&runtime·sched.atomic, -1<<mcpuShift);
+	if(!atomic_gwaiting(v) && (!atomic_waitstop(v) || atomic_mcpu(v) > atomic_mcpumax(v)))
 		return;
-	}
 
 	schedlock();
-
-	// atomic { mcpu--; }
-	v = runtime·xadd(&runtime·sched.atomic, (-1<<mcpuShift));
+	v = runtime·atomicload(&runtime·sched.atomic);
 	if(atomic_gwaiting(v)) {
 		matchmg();
 		v = runtime·atomicload(&runtime·sched.atomic);
@@ -837,29 +827,18 @@ runtime·entersyscall(void)
 void
 runtime·exitsyscall(void)
 {
-	uint32 v, w;
+	uint32 v;
 
 	if(runtime·sched.predawn)
 		return;
 
 	// Fast path.
-	// If we can do the mcpu-- bookkeeping and
+	// If we can do the mcpu++ bookkeeping and
 	// find that we still have mcpu <= mcpumax, then we can
 	// start executing Go code immediately, without having to
 	// schedlock/schedunlock.
-	for(;;) {
-		// If the profiler frequency needs updating,
-		// take the slow path.
-		if(m->profilehz != runtime·sched.profilehz)
-			break;
-
-		v = runtime·sched.atomic;
-		if(atomic_mcpu(v) >= atomic_mcpumax(v))
-			break;
-
-		w = v;
-		w += (1<<mcpuShift);
-		if(runtime·cas(&runtime·sched.atomic, v, w)) {
+	v = runtime·xadd(&runtime·sched.atomic, (1<<mcpuShift));
+	if(m->profilehz == runtime·sched.profilehz && atomic_mcpu(v) <= atomic_mcpumax(v)) {
 		// There's a cpu for us, so we can run.
 		g->status = Grunning;
 		// Garbage collector isn't running (since we are),
@@ -867,13 +846,9 @@ runtime·exitsyscall(void)
 		g->gcstack = nil;
 		return;
 	}
-	}
 
 	schedlock();
 
-	// atomic { mcpu++; }
-	runtime·xadd(&runtime·sched.atomic, (1<<mcpuShift));
-
 	// Tell scheduler to put g back on the run queue:
 	// mostly equivalent to g->status = Grunning,
 	// but keeps the garbage collector from thinking

src/pkg/runtime/proc.p

@@ -3,8 +3,9 @@
 // license that can be found in the LICENSE file.
 
 /*
-model for proc.c as of 2011/07/15.
-takes 4300 seconds to explore 1128130 states
+model for proc.c as of 2011/07/22.
+
+takes 4900 seconds to explore 1189070 states
 with G=3, var_gomaxprocs=1
 on a Core i7 L640 2.13 GHz Lenovo X201s.
 
@@ -329,33 +330,53 @@ inline schedule() {
 	nextgandunlock()
 }
 
+/*
+ * schedpend is > 0 if a goroutine is about to committed to
+ * entering the scheduler but has not yet done so.
+ * Just as we don't test for the undesirable conditions when a
+ * goroutine is in the scheduler, we don't test for them when
+ * a goroutine will be in the scheduler shortly.
+ * Modeling this state lets us replace mcpu cas loops with
+ * simpler mcpu atomic adds.
+ */
+byte schedpend;
+
 /*
  * entersyscall is like the C function.
  */
 inline entersyscall() {
+	bit willsched;
+
 	/*
 	 * Fast path.  Check all the conditions tested during schedlock/schedunlock
 	 * below, and if we can get through the whole thing without stopping, run it
 	 * in one atomic cas-based step.
 	 */
 	atomic {
+		atomic_mcpu--;
 		if
 		:: atomic_gwaiting ->
 			skip
-		:: atomic_waitstop && atomic_mcpu-1 <= atomic_mcpumax ->
+		:: atomic_waitstop && atomic_mcpu <= atomic_mcpumax ->
 			skip
 		:: else ->
-			atomic_mcpu--;
 			goto Lreturn_entersyscall;
-		fi
+		fi;
+		willsched = 1;
+		schedpend++;
 	}
 
 	/*
 	 * Normal path.
 	 */
 	schedlock()
-	d_step {
-		atomic_mcpu--;
+	opt_dstep {
+		if
+		:: willsched ->
+			schedpend--;
+			willsched = 0
+		:: else
+		fi
 	}
 	if
 	:: atomic_gwaiting ->
@@ -382,11 +403,11 @@ inline exitsyscall() {
 	 */
 	atomic {
 		// omitted profilehz check
+		atomic_mcpu++;
 		if
 		:: atomic_mcpu >= atomic_mcpumax ->
 			skip
 		:: else ->
-			atomic_mcpu++;
 			goto Lreturn_exitsyscall
 		fi
 	}
@@ -396,7 +417,6 @@ inline exitsyscall() {
 	 */
 	schedlock();
 	d_step {
-		atomic_mcpu++;
 		if
 		:: atomic_mcpu <= atomic_mcpumax ->
 			skip
@@ -497,10 +517,10 @@ active proctype monitor() {
 
 	do
 	// Should never have goroutines waiting with procs available.
-	:: !sched_lock && gwait > 0 && atomic_mcpu < atomic_mcpumax ->
+	:: !sched_lock && schedpend==0 && gwait > 0 && atomic_mcpu < atomic_mcpumax ->
 		assert 0
 	// Should never have gc waiting for stop if things have already stopped.
-	:: !sched_lock && atomic_waitstop && atomic_mcpu <= atomic_mcpumax ->
+	:: !sched_lock && schedpend==0 && atomic_waitstop && atomic_mcpu <= atomic_mcpumax ->
 		assert 0
 	od
 }