Add memory tracking telemetry support for Tracy by ccp-chargeback · Pull Request #22 · carbonengine/core

ccp-chargeback · 2026-06-03T12:23:00Z

Changes include:

Add memory tracking support in Tracy
A new TracyTestClient that mimics the behavior of the Tracy GUI / tracy-capture from connection point of view
Hardening on the internal TaskletZoneStore state
Making use of Tracy port 0.13.1#1

- Controllable via ENABLE_TELEMETRY_MEMORY_TRACKING, - Set to same value as CCP_TELEMETRY_ENABLED

Only rely on Tracy started/connected when sending memory free events

This unlocks the ability to do proper inspection of our interactions with Tracy. Nota bene: this class was created by Claude Code, and for tracy 0.11.1, so it may need some more TLC before it is fully usable. It might also be a breaking point when updating to newer tracy versions since it relies on Tracy's internal protocol.

So that tests can use it. Also provides a `TickTelemetry` wrapper around `CcpTickTelemetry` to deal with the synchronization between the test thread and the TracyTestClient thread.

Because that is how tracy works underneath the hood as well. Change made by claude code.

Interesting one! `SimpleZoneTest` crashes when entering the zone and accessing `t_taskletZoneStore`. This happens because the default value for `t_activeFiber` points at the first element in the `fiberNameStore`. Now, for whatever reason, when starting the telemetry integration set the unnamed root fiber name to trigger initialization of the corresponding zone stack, this does return the same iterator as `t_activeFiber` already contains. As a consequence, the initialization performs an early out, and does not create a relevant TaskletZoneStore. And without such a store, entering a zone clearly accesses invalid memory, causing the segmentation fault.

The internal book-keeping allows for multiple tracy zones sharing the same key (e.g. the `TaskletZoneStore` is a `stack`). However, due to the on-demand profiling book-keeping, there is an additional `set` that tracks whether a zone was manually created with a given key. Since that key is not required to be unique, it may only be removed from the manually tracked set once there are no more zones on that zone store's stack.

…nto tracy_memory_tracking

Our TracyTestClient needs access to (some of) the source files of Tracy in order to build correctly.

Likely still does not work on Windows, but at least it's not because of relying on Tracy internals.

Untested on Windows, written by claude on macOS.

…nto tracy_memory_tracking # Conflicts: # tests/CMakeLists.txt

…ations. CcpTelemetryIsConnected() has the check for: - TracyIsStarted - TracyIsConnected - checks s_profilerState == Started This means that any multiple start/stops within the same profiling session (in Tracy GUI visualizer) may fail. That's because it may try to allocate memory to the same memory address as previously done, but without registering the free event for that memory address. Because it got the free event while in a Tracy stopped state.

- Additionally sprinkle the code with fprintf statements, for debug purposes. To be removed later.

- This sets the m_shutdown boolean which controls the RecvLoop. - Also sprinkle in some debug fprintf statements (to be removed later)

- Call RequestShutdown() when StopRequested - Call ShutdownProfiler() when Stopped and safe to do so - Cleanup fiber names store and erease map, along with thread local zones stores - Make sure not to emit data for zones if telemetry is not connected. - Change indent of temp debug fprintf statements (all debug info to be removed later)

Added skeleton code for the following tests: - StartConnectStopDisconnectProfiling - StartConnectStopProfiling - StartConnectDisconnectProfiling - StartConnectDisconnectStopProfiling - ReconnectAfterStop

This maps onto TRACY_NO_CRASH_HANDLER, which should disable the internal Tracy crash handler. The no-crash-handler feature is specific to the carbonengine/vcpkg-registry port of Tracy. It replaces the default crash-handler INVERTED FEATURE in the official vcpkg registry.

Previously, the socket was only closed when `TracyTestClient::Disconnect` was called. However, it also needs to be closed when the profiler itself signals a disconnect.

# Conflicts: # CMakeLists.txt

Changes include: - Remove explicit call to tracy::ShutdownProfiler() - Revert Zone state bookkeeping changes/cleanup associated with earlier ShutdownProfiler change

…e into tracy_memory_tracking

Not sure why the tests weren't able to correctly deduce that it's a string literal without `constexpr`. The occurrence in `Throw` was a valid concern, as that just blindly trusts the input, which the compiler cannot assume as always being a string literal.

This keeps usage of the underlying profiling framework isolated.

Also: - Cleaned up unused/wrong tests - Added overload for SetUp() that keeps existing connection to TracyTestClient - Removed TODO statements

CCP-Aporia

See inline comments.

CCP-Aporia · 2026-06-03T12:29:05Z

+		SetUp(true);
+	}
+
+	void SetUp(bool doTestClientConnect)


I'm not entirely sure what the boolean parameter here is for. If there is desire to have different set up behaviour, this should be done through a specialised Test class instead.

CCP-Aporia · 2026-06-03T13:11:31Z

+	auto tracyZones = m_tracyClient.GetZones();
+	// CcpTelemetryEnterZone passes the zone name as the Tracy "function" field
+	// (via the 6-param ___tracy_alloc_srcloc), so match against both fields.
+	auto pred = [&zoneName]( const TracyTestClient::ZoneInfo& elem ) -> bool {
+		return elem.function == zoneName;
+	};
+	EXPECT_NE( tracyZones.end(), std::find_if( tracyZones.begin(), tracyZones.end(), pred ) );


Seeing that this snippet is now used in multiple places, it would benefit from being refactored into a helper function that does the assertion.

CCP-Aporia · 2026-06-03T13:12:09Z

+	tracyZones = m_tracyClient.GetZones();
+	EXPECT_EQ( tracyZones.end(), std::find_if( tracyZones.begin(), tracyZones.end(), pred ) );


This is a good candidate for a helper function, too, since it's the inverse of the snippet right above.

CCP-Aporia · 2026-06-03T13:13:08Z

+	auto tracyZones = m_tracyClient.GetZones();
+	auto pred = [&zoneName1]( const TracyTestClient::ZoneInfo& elem ) -> bool {
+		return elem.function == zoneName1;
+	};
+	EXPECT_NE( tracyZones.end(), std::find_if( tracyZones.begin(), tracyZones.end(), pred ) );


This shouldn't have been copy/pasted. Also, why is this necessary to check this as part of the restart-after-stop scenario?

CCP-Aporia · 2026-06-03T13:13:41Z

+	EXPECT_TRUE( m_tracyClient.GetZones().empty() );
+}
+
+TEST_F( CcpTelemetryTest, ReStartAfterStop )


Which specific restart-after-stop scenario is simulated here? The telemetry integration, or the client application?

This is simulating:

Press Start in ESP/Client

Press Stop in ESP/Client

Press Start again in ESP/Client

CCP-Aporia · 2026-06-03T13:22:55Z

+	CcpStopTelemetry();
+	TickTelemetry(); // This processes the "StopRequested" state.
+	TickTelemetry(); // This processes the "Stopped" state.
+	EXPECT_TRUE( m_tracyClient.IsConnected() ) << "Connection should still be true at this point because the TracyTestClient hasn't been disconnected";


Is this important to test here? Also, is this is a valid assumption? m_tracyClient.IsConnected() returns false once the TracyTestClient notices that the socket connection went away.

m_connected only becomes false on TracyTestClient::Disconnect() and at the end of the RevcLoop() after sock.Close().
This test is not disconnecting from the TracyTestClient, to this will always be true.
But you're right in this test, we don't need to check if m_tracyClient.IsConnected() == true (because it will always be true as there is no disconnect happening)

CCP-Aporia · 2026-06-03T13:26:05Z

+	TickTelemetry(); // This processes the "StopRequested" state.
+	TickTelemetry(); // This processes the "Stopped" state.
+	EXPECT_TRUE( m_tracyClient.IsConnected() ) << "Connection should still be true at this point because the TracyTestClient hasn't been disconnected";
+	EXPECT_FALSE( CcpTelemetryIsStarted() ) << "Internal profiler state should have changed: Started->StopRequested->Stopped";


The assertion as written also holds if we accidentally end up in StartRequested. And it should be unnecessary in a world where we have a helper method for the tests to perform a stop of the telemetry integration.

CCP-Aporia · 2026-06-03T13:27:34Z

+	EXPECT_FALSE( CcpTelemetryIsStarted() ) << "Internal profiler state should have changed: Started->StopRequested->Stopped";
+
+	// Simulate a new call to StartTelemetry
+	SetUp( false );


As pointed out in my comments on the SetUp( bool ) overload, this really ought to call an explicit helper method.

CCP-Aporia · 2026-06-03T13:29:07Z

+	// Simulate a new call to StartTelemetry
+	SetUp( false );
+	EXPECT_TRUE( m_tracyClient.IsConnected() ) << "Connection should still be true because the TracyTestClient hasn't never been disconnected";
+	EXPECT_TRUE( CcpTelemetryIsStarted() ) << "Internal profiler state should have changed: Started->StopRequested->Stopped";


The message seems to be copied from above, and it's missing the StartRequested phase in the flow. Also, it's an assertion that's implied by calling an appropriate helper method to started the telemetry integration.

CCP-Aporia · 2026-06-03T13:43:29Z

+	auto tracyZones2ndStart = m_tracyClient.GetZones();
+	auto pred2nd = [&zoneName2]( const TracyTestClient::ZoneInfo& elem ) -> bool {
+		return elem.function == zoneName2;
+	};
+	EXPECT_NE( tracyZones2ndStart.end(), std::find_if( tracyZones2ndStart.begin(), tracyZones2ndStart.end(), pred2nd ) );


As mentioned above, this should be a helper method instead of copy/paste.

ccp-chargeback and others added 30 commits April 17, 2026 17:39

Add basic Tracy memory tracking

f98c60d

- Controllable via ENABLE_TELEMETRY_MEMORY_TRACKING, - Set to same value as CCP_TELEMETRY_ENABLED

Fix memory tracking in v0.11.1

f99f8b7

Only rely on Tracy started/connected when sending memory free events

Control memory tracking with a switch

06e4be0

Only do TracyIsStarted/Connected when tracking memory allocations

2dd5320

Add CcpMemoryProfilingIsEnabled()

52fed36

Start and stop TracyTestClient as part of the CcpTelemetryTest fixture

c4a0837

So that tests can use it. Also provides a `TickTelemetry` wrapper around `CcpTickTelemetry` to deal with the synchronization between the test thread and the TracyTestClient thread.

Update TracyTestClient to track zones as stacks per fiber and thread

896d5cd

Because that is how tracy works underneath the hood as well. Change made by claude code.

Add a simple test for entering / leaving a telemetry zone

63fb470

Merge remote-tracking branch 'ccp-chargeback/tracy_memory_tracking' i…

4f110ed

…nto tracy_memory_tracking

Fix TracyTestClient when running on Windows

b9c5317

Our TracyTestClient needs access to (some of) the source files of Tracy in order to build correctly.

Update to tracy 0.13.1

1d06343

Rewrite TracyTestClient to no longer rely on Tracy internal sources

07b7397

Likely still does not work on Windows, but at least it's not because of relying on Tracy internals.

Add preliminary Windows support to TracyTestClient

e69d2db

Untested on Windows, written by claude on macOS.

Merge remote-tracking branch 'ccp-chargeback/tracy_memory_tracking' i…

d309309

…nto tracy_memory_tracking # Conflicts: # tests/CMakeLists.txt

Add ws2_32 to test project when linking against Windows

2c15aaf

Merge branch 'main' into tracy_memory_tracking

9af3e70

Create first draft of ReconnectAfterStop test.

da7f492

- Additionally sprinkle the code with fprintf statements, for debug purposes. To be removed later.

Handle kQueueTerminate in TracyTestClient

6021115

- This sets the m_shutdown boolean which controls the RecvLoop. - Also sprinkle in some debug fprintf statements (to be removed later)

Add some StartStopConnectDisconnect profiler tests

3675939

Added skeleton code for the following tests: - StartConnectStopDisconnectProfiling - StartConnectStopProfiling - StartConnectDisconnectProfiling - StartConnectDisconnectStopProfiling - ReconnectAfterStop

Fix disconnect logic in TracyTestClient

7072e13

Previously, the socket was only closed when `TracyTestClient::Disconnect` was called. However, it also needs to be closed when the profiler itself signals a disconnect.

Merge branch 'main' into tracy_memory_tracking

ded1346

# Conflicts: # CMakeLists.txt

Revert ShutdownProfiler() changes

8de74de

Changes include: - Remove explicit call to tracy::ShutdownProfiler() - Revert Zone state bookkeeping changes/cleanup associated with earlier ShutdownProfiler change

Remove debug printf statements in CcpTelemetry.cpp

4ac288e

Rename s_telemetryConfig back to s_config

8fbbc0b

ccp-chargeback and others added 5 commits May 27, 2026 11:45

Fix wrong size being sent to Tracy in CcpPlatformCalloc

41fdf6a

Merge branch 'tracy_memory_tracking' of github.com:ccp-chargeback/cor…

879c18d

…e into tracy_memory_tracking

Wrap Tracy memory API calls behind CcpTelemetry methods

1434367

This keeps usage of the underlying profiling framework isolated.

Add ReStartAfterStop test

b852e0f

Also: - Cleaned up unused/wrong tests - Added overload for SetUp() that keeps existing connection to TracyTestClient - Removed TODO statements

ccp-chargeback requested a review from CCP-Aporia June 3, 2026 12:23

Remove fprintf statements from tests/CcpTelemetry.cpp

4ccb200

CCP-Aporia requested changes Jun 3, 2026

View reviewed changes

ccp-chargeback added 2 commits June 3, 2026 14:01

Add calls to base Test class SetUp() and TearDown() functions.

0ed5a38

Use distinct FiberNames in TestFiberSwitching test

bbe02bb

		tracyZones = m_tracyClient.GetZones();
		EXPECT_EQ( tracyZones.end(), std::find_if( tracyZones.begin(), tracyZones.end(), pred ) );

Conversation

ccp-chargeback commented Jun 3, 2026

Uh oh!

CCP-Aporia left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants