null check postPhase in Worker.run to resolve race condition at terminate

[mbutrovich]

ThreadBench and WorkloadState rely on WorkloadState.currentPhase being null as the indicator that the benchmark is over. For example:

benchbase/src/main/java/com/oltpbenchmark/WorkloadState.java

Lines 205 to 206 in 1c43cad

	if (this.currentPhase == null)
	// Benchmark is over---wake everyone up so they can terminate

However, Worker relies on BenchmarkState.state being DONE as its indicator that the benchmark is over. For example, this switch assumes that because state is still MEASURE that the benchmark is still running and that the result of getCurrentPhase is safe to access:

benchbase/src/main/java/com/oltpbenchmark/api/Worker.java

Lines 288 to 296 in 1c43cad

	switch (postState) {
	case MEASURE:
	// Non-serial measurement. Only measure if the state both
	// before and after was MEASURE, and the phase hasn't
	// changed, otherwise we're recording results for a query
	// that either started during the warmup phase or ended
	// after the timer went off.
	Phase postPhase = workloadState.getCurrentPhase();
	if (preState == MEASURE && postPhase.getId() == prePhase.getId()) {

The race occurs between the previously mentioned switch in Worker and in ThreadBench where this line advances to the next phase, which in the case of the benchmark ending, puts the current phase to null. If a Worker tries to read the current phase before the state is set to DONE when ThreadBench calls startCoolDown() then benchbase goes boom.

benchbase/src/main/java/com/oltpbenchmark/ThreadBench.java

Lines 222 to 229 in 1c43cad

	workState.switchToNextPhase();
	lowestRate = Integer.MAX_VALUE;
	phase = workState.getCurrentPhase();
	interruptWorkers();
	if (phase == null && !lastEntry) {
	// Last phase
	lastEntry = true;
	testState.startCoolDown();

I believe the easiest fix here is a documented null check in Worker. Maybe there's a better way to sync up the indicators of the benchmark ending between all of these classes, but this seems like the lowest risk change.

Fixes #37.

[apavlo]

LGTM