Prevent double flush due to race in SingleDirectoryDbLedgerStorage

[michaeljmarshall]

Disclaimer: I am not very familiar with this code, so please review carefully.

Motivation

We use the isFlushOngoing and the hasFlushBeenTriggered atomic booleans to prevent concurrent flushes. See:

bookkeeper/bookkeeper-server/src/main/java/org/apache/bookkeeper/bookie/storage/ldb/SingleDirectoryDbLedgerStorage.java

Lines 514 to 530 in 2986e04

	// Write cache is full, we need to trigger a flush so that it gets rotated
	// If the flush has already been triggered or flush has already switched the
	// cache, we don't need to trigger another flush
	if (!isFlushOngoing.get() && hasFlushBeenTriggered.compareAndSet(false, true)) {
	// Trigger an early flush in background
	log.info("Write cache is full, triggering flush");
	executor.execute(() -> {
	long startTime = System.nanoTime();
	try {
	flush();
	} catch (IOException e) {
	log.error("Error during flush", e);
	} finally {
	flushExecutorTime.addLatency(MathUtils.elapsedNanos(startTime), TimeUnit.NANOSECONDS);
	}
	});
	}

By my reading of the code, it seems there is a brief period of time where the calling code will think it is valid to flush, even though there is a flush in progress.

Changes

In the SingleDirectoryDbLedgerStorage#swapWriteCache method modified by this PR, I propose a change to set isFlushOngoing to true before setting hasFlushBeenTriggered to false.

Master Issue: none

[shoothzj]

LGTM

[dlg99]

closing to re-trigger CI