runtime: make stack barrier locking more robust

The stack barrier locking functions use a simple cas lock because they
need to support trylock, but currently don't increment g.m.locks. This
is okay right now because they always run on the system stack or the
signal stack and are hence non-preemtible, but this could lead to
difficult-to-reproduce deadlocks if these conditions change in the
future.

Make these functions more robust by incrementing g.m.locks and making
them nosplit to enforce non-preemtibility.

Change-Id: I73d60a35bd2ad2d81c73aeb20dbd37665730eb1b
Reviewed-on: https://go-review.googlesource.com/17058
Run-TryBot: Austin Clements <austin@google.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Ingo Oeser <nightlyone@googlemail.com>
Reviewed-by: Russ Cox <rsc@golang.org>

src/runtime/mstkbar.go

@@ -300,16 +300,26 @@ func setNextBarrierPC(pc uintptr) {
 // This is necessary because a sigprof during barrier installation or
 // removal could observe inconsistencies between the stkbar array and
 // the stack itself and crash.
+//
+//go:nosplit
 func gcLockStackBarriers(gp *g) {
+	acquirem()
 	for !atomic.Cas(&gp.stackLock, 0, 1) {
 		osyield()
 	}
 }
 
+//go:nosplit
 func gcTryLockStackBarriers(gp *g) bool {
-	return atomic.Cas(&gp.stackLock, 0, 1)
+	mp := acquirem()
+	result := atomic.Cas(&gp.stackLock, 0, 1)
+	if !result {
+		releasem(mp)
+	}
+	return result
 }
 
 func gcUnlockStackBarriers(gp *g) {
 	atomic.Store(&gp.stackLock, 0)
+	releasem(getg().m)
 }