17.11 VS hang in EndBuild

[JanKrivanek]

Context

VS hangs during solution close - turins out to be caused by hang in BuildManager.EndBuild
There are almost 3k cases hit in wild on 17.11

Analysis

Cab: https://watsonportal.microsoft.com/CabAnalysis?CabIdentifier=https://eaus2watcab01.blob.core.windows.net/global-202409/f1d6f9d9-57c1-47af-91c9-1973b6d19519.zip
Part of: https://devdiv.visualstudio.com/DevDiv/_workitems/edit/2254960

VS hanging due to BuildManager.EndBuild stuck waiting on the _noActiveSubmissionsEvent or _noNodesActiveEvent (not possible to distinguish due to release build optimizations).

Based on stacks the CancelAllSubmissionsAsync was called.
At the same time the InProcNode is still running:

Though based on the code it seems like we do not want to kill InProcNode in VS:

msbuild/src/Build/BackEnd/BuildManager/BuildParameters.cs

Lines 596 to 604 in ab7c289

	/// <summary>
	/// Shutdown the inprocess node when the build finishes. By default this is false
	/// since visual studio needs to keep the inprocess node around after the build finishes.
	/// </summary>
	public bool ShutdownInProcNodeOnBuildFinish
	{
	get => _shutdownInProcNodeOnBuildFinish;
	set => _shutdownInProcNodeOnBuildFinish = value;
	}

However it would mean that the node count (_activeNodes.Count) wouldn't reach zero and we wouldn't be able to signal the _noNodesActiveEvent:

msbuild/src/Build/BackEnd/BuildManager/BuildManager.cs

Lines 2558 to 2587 in ab7c289

	private void CheckForActiveNodesAndCleanUpSubmissions()
	{
	Debug.Assert(Monitor.IsEntered(_syncLock));
	if (_activeNodes.Count == 0)
	{
	var submissions = new List<BuildSubmissionBase>(_buildSubmissions.Values);
	foreach (BuildSubmissionBase submission in submissions)
	{
	// The submission has not started do not add it to the results cache
	if (!submission.IsStarted)
	{
	continue;
	}
	if (!CompleteSubmissionFromCache(submission))
	{
	submission.CompleteResultsWithException(new BuildAbortedException());
	}
	// If we never received a project started event, consider logging complete anyhow, since the nodes have
	// shut down.
	submission.CompleteLogging();
	CheckSubmissionCompletenessAndRemove(submission);
	}
	_noNodesActiveEvent?.Set();
	}
	}

This feels unlikely - as it would hang always when we'd have InProcNode in VS.

Next steps in investigation

Debug through solution close in VS and find out if CheckForActiveNodesAndCleanUpSubmissions ever signals

[JanKrivanek]

On internal user hitting this reported the issue stopped occuring after enabling long paths support on his system

[JanKrivanek]

I got an internal customer dump - the blocking event is for sure _noActiveSubmissionsEvent
deeper analysis pending...

[JanKrivanek]

It turns out there is a race between CancelAllSubmissions and EndBuild - that are seemingly synchronous, but CancelAllSubmissions enqueues bit of logic for async processing.

When both methods called after each other - which is expected during solution close - they can race and deadlock