Remove static modifiers from ThreadPool events in NativeRuntimeEventSource.PortableThreadPool.NativeSinks.cs

[nickyMcDonald]

Here is a fix for issue #78006 that works with win11 .NET 7.

The EventPipeEventDispatcher cycles through EventPipeInternal via QCall and sends them to be processed by the NativeRuntimeEventSource.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventPipeEventDispatcher.cs

Lines 169 to 181 in 0c612cf

	while (!m_stopDispatchTask && EventPipeInternal.GetNextEvent(m_sessionID, &instanceData))
	{
	eventsReceived = true;
	// Filter based on provider.
	if (instanceData.ProviderID == m_RuntimeProviderID)
	{
	// Dispatch the event.
	ReadOnlySpan<byte> payload = new ReadOnlySpan<byte>((void*)instanceData.Payload, (int)instanceData.PayloadLength);
	DateTime dateTimeStamp = TimeStampToDateTime(instanceData.TimeStamp);
	NativeRuntimeEventSource.Log.ProcessEvent(instanceData.EventID, instanceData.ThreadID, dateTimeStamp, instanceData.ActivityId, instanceData.ChildActivityId, payload);
	}
	}

From there they are sent to be decoded by the EventPipePayloadDecoder.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/NativeRuntimeEventSource.cs

Line 55 in a08818e

object[] decodedPayloadFields = EventPipePayloadDecoder.DecodePayload(ref m_eventData[eventID], payload);

There is where the NRE occurs because the metadata is uninitialized. Parameters is default null and attempting to get its length results in NRE. Looking back at the NativeRuntimeEventSource the NRE occurs with eventID 64 (ThreadPoolIODequeue).

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventPipePayloadDecoder.cs

Lines 16 to 19 in a08818e

	internal static object[] DecodePayload(ref EventSource.EventMetadata metadata, ReadOnlySpan<byte> payload)
	{
	ParameterInfo[] parameters = metadata.Parameters;
	object[] decodedFields = new object[parameters.Length];

In the EventSource. There are no BindingFlags for static methods. This is important because any static methods from the NativeRuntimeEventSource won’t get loaded. Because this data is used to initialize m_eventData values, any methods left out will leave uninitialized slots in m_eventData. Slot 64 being one of them.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventSource.cs

Line 3042 in 0c612cf

MethodInfo[] methods = eventSourceType.GetMethods(BindingFlags.DeclaredOnly | BindingFlags.NonPublic | BindingFlags.Public | BindingFlags.Instance);

The fix will allow for event 64 and 63 to be loaded into m_eventData before pr #66333 again.

[ghost]

Tagging subscribers to this area: @mangod9
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Here is a fix for issue #78006 that works with win11 .NET 7.

The EventPipeEventDispatcher cycles through EventPipeInternal via QCall and sends them to be processed by the NativeRuntimeEventSource.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventPipeEventDispatcher.cs

Lines 169 to 181 in 0c612cf

	while (!m_stopDispatchTask && EventPipeInternal.GetNextEvent(m_sessionID, &instanceData))
	{
	eventsReceived = true;
	// Filter based on provider.
	if (instanceData.ProviderID == m_RuntimeProviderID)
	{
	// Dispatch the event.
	ReadOnlySpan<byte> payload = new ReadOnlySpan<byte>((void*)instanceData.Payload, (int)instanceData.PayloadLength);
	DateTime dateTimeStamp = TimeStampToDateTime(instanceData.TimeStamp);
	NativeRuntimeEventSource.Log.ProcessEvent(instanceData.EventID, instanceData.ThreadID, dateTimeStamp, instanceData.ActivityId, instanceData.ChildActivityId, payload);
	}
	}

From there they are sent to be decoded by the EventPipePayloadDecoder.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/NativeRuntimeEventSource.cs

Line 55 in a08818e

object[] decodedPayloadFields = EventPipePayloadDecoder.DecodePayload(ref m_eventData[eventID], payload);

There is where the NRE occurs because the metadata is uninitialized. Parameters is default null and attempting to get its length results in NRE. Looking back at the NativeRuntimeEventSource the NRE occurs with eventID 64 (ThreadPoolIODequeue).

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventPipePayloadDecoder.cs

Lines 16 to 19 in a08818e

	internal static object[] DecodePayload(ref EventSource.EventMetadata metadata, ReadOnlySpan<byte> payload)
	{
	ParameterInfo[] parameters = metadata.Parameters;
	object[] decodedFields = new object[parameters.Length];

In the EventSource. There are no BindingFlags for static methods. This is important because any static methods from the NativeRuntimeEventSource won’t get loaded. Because this data is used to initialize m_eventData values, any methods left out will leave uninitialized slots in m_eventData. Slot 64 being one of them.

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventSource.cs

Line 3042 in 0c612cf

MethodInfo[] methods = eventSourceType.GetMethods(BindingFlags.DeclaredOnly | BindingFlags.NonPublic | BindingFlags.Public | BindingFlags.Instance);

The fix will allow for event 64 and 63 to be loaded into m_eventData before pr #66333 again.

Author:	n77y
Assignees:	-
Labels:	area-System.Threading
Milestone:	-

[jkotas]

cc @marek-safar This is regression introduced by https://github.com/dotnet/runtime/pull/66333/files#diff-f7c0dd1b698275f0f183b9517883bc82280850fc6f256e8b7405c02ce0d0c22fR152

[jkotas]

Is it possible to add a test for this?

e.g. here: https://github.com/dotnet/runtime/blob/main/src/tests/tracing/runtimeeventsource/NativeRuntimeEventSourceTest.cs

[hoyosjs]

Hmm... tracing/runtimeeventsource/nativeruntimeeventsource/nativeruntimeeventsource.sh is now timing out and a dump was collected on macOS x64

[nickyMcDonald]

Hmm... tracing/runtimeeventsource/nativeruntimeeventsource/nativeruntimeeventsource.sh is now timing out and a dump was collected on macOS x64

After adding tests to NativeRuntimeEventSourceTest.cs and returning the static modifiers to NativeRuntimeEventSource.PortableThreadPool.NativeSinks.cs the tests successfully failed. I will now remove the static modifiers and the tracing tests should now pass.

[tommcdon]

@davmason ptal