runtime: utilize EVFILT_USER more effectively

Re-work kqueue_event wakeup logic to use one-shot events. In an
event of waking up a wrong thread, simply re-post the event.

This saves close to 1 system call per wakeup on average, since
chances of non-blocking poller picking it up is pretty low.

src/runtime/netpoll_kqueue.go

@@ -129,10 +129,11 @@ retry:
 		ev := &events[i]
 
 		if isWakeup(ev) {
-			if delay != 0 {
+			isBlocking := delay != 0
+			processWakeupEvent(kq, isBlocking)
+			if isBlocking {
 				// netpollBreak could be picked up by a nonblocking poll.
-				// Only call drainWakeupEvent and reset the netpollWakeSig if blocking.
-				drainWakeupEvent(kq)
+				// Only reset the netpollWakeSig if blocking.
 				netpollWakeSig.Store(0)
 			}
 			continue

src/runtime/netpoll_kqueue_event.go

@@ -16,7 +16,7 @@ func addWakeupEvent(kq int32) {
 	ev := keventt{
 		ident:  kqIdent,
 		filter: _EVFILT_USER,
-		flags:  _EV_ADD,
+		flags:  _EV_ADD | _EV_CLEAR,
 	}
 	for {
 		n := kevent(kq, &ev, 1, nil, 0, nil)
@@ -38,7 +38,6 @@ func wakeNetpoll(kq int32) {
 	ev := keventt{
 		ident:  kqIdent,
 		filter: _EVFILT_USER,
-		flags:  _EV_ENABLE,
 		fflags: _NOTE_TRIGGER,
 	}
 	for {
@@ -66,13 +65,11 @@ func isWakeup(ev *keventt) bool {
 	return false
 }
 
-func drainWakeupEvent(kq int32) {
-	ev := keventt{
-		ident:  kqIdent,
-		filter: _EVFILT_USER,
-		flags:  _EV_DISABLE,
+func processWakeupEvent(kq int32, isBlocking bool) {
+	if !isBlocking {
+		// Got a wrong thread, relay
+		wakeNetpoll(kq)
 	}
-	kevent(kq, &ev, 1, nil, 0, nil)
 }
 
 func netpollIsPollDescriptor(fd uintptr) bool {

src/runtime/netpoll_kqueue_pipe.go

@@ -63,7 +63,11 @@ func isWakeup(ev *keventt) bool {
 	return false
 }
 
-func drainWakeupEvent(_ int32) {
+func processWakeupEvent(_ int32, isBlocking bool) {
+	// Only drain if blocking.
+	if !isBlocking {
+		return
+	}
 	var buf [16]byte
 	read(int32(netpollBreakRd), noescape(unsafe.Pointer(&buf[0])), int32(len(buf)))
 }