When threads have low free stack space, StackSpiller may infinite-loop through the thread pool

[kouvel]

When threads have marginally more than 128 KB of stack space, RuntimeHelpers.TryEnsureSufficientExecutionStack discounts 128 KB of stack space and only considers the remaining space, and there may not be enough even for a non-nested/non-recursive thread pool work item. In that situation, StackGuard.RunOnEmptyStackCore is triggered again from its continuation and keeps doing so since TryEnsureSufficientExecutionStack always returns false.

This appears to be the continuation triggered by RunOnEmptyStackCore, which triggers itself continually:

runtime/src/libraries/System.Linq.Expressions/src/System/Linq/Expressions/Compiler/StackSpiller.Generated.cs

Lines 16 to 28 in 16300e0

	private Result RewriteExpression(Expression node, Stack stack)
	{
	if (node == null)
	{
	return new Result(RewriteAction.None, null);
	}
	// When compiling deep trees, we run the risk of triggering a terminating StackOverflowException,
	// so we use the StackGuard utility here to probe for sufficient stack and continue the work on
	// another thread when we run out of stack space.
	if (!_guard.TryEnterOnCurrentStack())
	{
	return _guard.RunOnEmptyStack((StackSpiller @this, Expression n, Stack s) => @this.RewriteExpression(n, s), this, node, stack);

[ghost]

Tagging subscribers to this area: @cston
See info in area-owners.md if you want to be subscribed.

Issue Details

---

When threads have marginally more than 128 KB of stack space, RuntimeHelpers.TryEnsureSufficientExecutionStack discounts 128 KB of stack space and only considers the remaining space, and there may not be enough even for a non-nested/non-recursive thread pool work item. In that situation, StackGuard.RunOnEmptyStackCore is triggered again from its continuation and keeps doing so since TryEnsureSufficientExecutionStack always returns false.

This appears to be the continuation triggered by RunOnEmptyStackCore, which triggers itself continually:

runtime/src/libraries/System.Linq.Expressions/src/System/Linq/Expressions/Compiler/StackSpiller.Generated.cs

Lines 16 to 28 in 16300e0

	private Result RewriteExpression(Expression node, Stack stack)
	{
	if (node == null)
	{
	return new Result(RewriteAction.None, null);
	}
	// When compiling deep trees, we run the risk of triggering a terminating StackOverflowException,
	// so we use the StackGuard utility here to probe for sufficient stack and continue the work on
	// another thread when we run out of stack space.
	if (!_guard.TryEnterOnCurrentStack())
	{
	return _guard.RunOnEmptyStack((StackSpiller @this, Expression n, Stack s) => @this.RewriteExpression(n, s), this, node, stack);

Author:	kouvel
Assignees:	kouvel
Labels:	area-System.Linq.Expressions
Milestone:	7.0.0

[kouvel]

Repro is on an Ubuntu 20.04 VM, to run the System.Linq.Expressions tests with the following environment variable set (128 KB stack size for threads):

COMPlus_DefaultStackSize=0x20000

dotnet build /p:Configuration=Release /t:Test from src/libraries/System.Linq.Expressions/tests appears to hang for the same reason when the env var is set, and running the tests directly with the build output also hangs with some tests reported as long-running:

   System.Linq.Expressions.Tests: [Long Running Test] 'System.Linq.Expressions.Tests.LiftedSubtractNullableTests.CheckLiftedSubtractNullableULongTest', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.AsTests.CheckCustomArrayAsIEnumerableOfInterfaceTest', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.ListBindTests.NonAddableListType', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.Assign.ReferenceAssignable', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.ConvertTests.ConvertNullableDoubleToNullableEnumTest', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.MemberInitTests.Reduce', Elapsed: 00:02:11
[Long Running Test] 'System.Dynamic.Tests.BindingRestrictionsTests.TypeRestrictionTrueForMatchType', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.ConvertCheckedTests.ConvertCheckedNullableFloatToSByteTest', Elapsed: 00:02:11
[Long Running Test] 'System.Dynamic.Tests.ExpandoObjectProxyTests.KeyCollectionCorrectlyViewed', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.LiftedMultiplyCheckedNullableTests.CheckLiftedMultiplyCheckedNullableSByteTest', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.BinaryNullableAndTests.CheckNullableULongAndTest', Elapsed: 00:02:11
[Long Running Test] 'System.Linq.Expressions.Tests.LambdaMultiplyTests.LambdaMultiplyUShortTest', Elapsed: 00:02:11

I attempted something simple, just to avoid checking the StackGuard when the queued work item runs, but it doesn't seem to be enough to fix the hang, I guess it may be more involved.