runtime: avoid futile mark worker acquisition

During the GC mark phase, one of the first behaviors of findRunnable is
to check if it should execute a GC mark worker. Mark workers often run
for many milliseconds in a row, so programs that invoke the scheduler
more frequently will see that condition trigger only a tiny fraction of
the time.

Obtaining a mark worker from the gcBgMarkWorkerPool involves a CAS on a
single memory location that's shared across the process. When GOMAXPROCS
is large, the resulting contention can waste a significant amount of CPU
time. But a sufficiently large GOMAXPROCS also means there's no need for
fractional mark workers, making it easier to check ahead of time if we
need to run a worker.

Check, without committing to a particular worker, whether we would even
want to run one.

For #68399

Change-Id: I5d8578c2101ee20a8a4156a029584356095ea118
Reviewed-on: https://go-review.googlesource.com/c/go/+/602477
Reviewed-by: Michael Pratt <[email protected]>
Auto-Submit: Rhys Hiltner <[email protected]>
Reviewed-by: Michael Knyszek <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>

src/runtime/mgcpacer.go

@@ -752,6 +752,17 @@ func (c *gcControllerState) findRunnableGCWorker(pp *p, now int64) (*g, int64) {
 		return nil, now
 	}
 
+	if c.dedicatedMarkWorkersNeeded.Load() <= 0 && c.fractionalUtilizationGoal == 0 {
+		// No current need for dedicated workers, and no need at all for
+		// fractional workers. Check before trying to acquire a worker; when
+		// GOMAXPROCS is large, that can be expensive and is often unnecessary.
+		//
+		// When a dedicated worker stops running, the gcBgMarkWorker loop notes
+		// the need for the worker before returning it to the pool. If we don't
+		// see the need now, we wouldn't have found it in the pool anyway.
+		return nil, now
+	}
+
 	// Grab a worker before we commit to running below.
 	node := (*gcBgMarkWorkerNode)(gcBgMarkWorkerPool.pop())
 	if node == nil {