diff --git a/src/runtime/proc.c b/src/runtime/proc.c
index 3fe08df94d..0b509bb875 100644
--- a/src/runtime/proc.c
+++ b/src/runtime/proc.c
@@ -397,7 +397,7 @@ nextgandunlock(void)
 		throw("all goroutines are asleep - deadlock!");
 	m->nextg = nil;
 	noteclear(&m->havenextg);
-	if(sched.waitstop) {
+	if(sched.waitstop && sched.mcpu <= sched.mcpumax) {
 		sched.waitstop = 0;
 		notewakeup(&sched.stopped);
 	}
@@ -590,6 +590,10 @@ sys·entersyscall(uint64 callerpc, int64 trap)
 	sched.msyscall++;
 	if(sched.gwait != 0)
 		matchmg();
+	if(sched.waitstop && sched.mcpu <= sched.mcpumax) {
+		sched.waitstop = 0;
+		notewakeup(&sched.stopped);
+	}
 	unlock(&sched);
 	// leave SP around for gc; poison PC to make sure it's not used
 	g->sched.SP = (byte*)&callerpc;
diff --git a/src/runtime/rt1_amd64_darwin.c b/src/runtime/rt1_amd64_darwin.c
index b614756721..c476f89b58 100644
--- a/src/runtime/rt1_amd64_darwin.c
+++ b/src/runtime/rt1_amd64_darwin.c
@@ -277,19 +277,25 @@ xadd(uint32 volatile *val, int32 delta)
 void
 lock(Lock *l)
 {
+	if(m->locks < 0)
+		throw("lock count");
+	m->locks++;
+
 	// Allocate semaphore if needed.
 	if(l->sema == 0)
 		initsema(&l->sema);
 
 	if(xadd(&l->key, 1) > 1)	// someone else has it; wait
 		mach_semacquire(l->sema);
-	m->locks++;
 }
 
 void
 unlock(Lock *l)
 {
 	m->locks--;
+	if(m->locks < 0)
+		throw("lock count");
+
 	if(xadd(&l->key, -1) > 0)	// someone else is waiting
 		mach_semrelease(l->sema);
 }
diff --git a/src/runtime/rt1_amd64_linux.c b/src/runtime/rt1_amd64_linux.c
index c0c2038053..5b3e458094 100644
--- a/src/runtime/rt1_amd64_linux.c
+++ b/src/runtime/rt1_amd64_linux.c
@@ -301,13 +301,11 @@ futexwakeup(uint32 *addr)
 // 	else return 0;
 // but atomically.
 
-void
-lock(Lock *l)
+static void
+futexlock(Lock *l)
 {
 	uint32 v;
 
-	m->locks++;
-
 again:
 	v = l->key;
 	if((v&1) == 0){
@@ -346,13 +344,11 @@ again:
 	goto again;
 }
 
-void
-unlock(Lock *l)
+static void
+futexunlock(Lock *l)
 {
 	uint32 v;
 
-	m->locks--;
-
 	// Atomically get value and clear lock bit.
 again:
 	v = l->key;
@@ -366,6 +362,24 @@ again:
 		futexwakeup(&l->key);
 }
 
+void
+lock(Lock *l)
+{
+	if(m->locks < 0)
+		throw("lock count");
+	m->locks++;
+	futexlock(l);
+}
+
+void
+unlock(Lock *l)
+{
+	m->locks--;
+	if(m->locks < 0)
+		throw("lock count");
+	futexunlock(l);
+}
+
 
 // One-time notifications.
 //
@@ -383,20 +397,20 @@ void
 noteclear(Note *n)
 {
 	n->lock.key = 0;	// memset(n, 0, sizeof *n)
-	lock(&n->lock);
+	futexlock(&n->lock);
 }
 
 void
 notewakeup(Note *n)
 {
-	unlock(&n->lock);
+	futexunlock(&n->lock);
 }
 
 void
 notesleep(Note *n)
 {
-	lock(&n->lock);
-	unlock(&n->lock);	// Let other sleepers find out too.
+	futexlock(&n->lock);
+	futexunlock(&n->lock);	// Let other sleepers find out too.
 }