GLFW multiwindow crash with cursor/mouse callbacks

[desiwko]

Summary

When realizing multiple windows using the GLFW backend the application will crash in Silk.NET.GLFW.Glfw.PollEvents() after exercising the windows (change focus, resize, mouse over, etc.) with a callstack similar to the following:

Process terminated. A callback was made on a garbage collected delegate of type 'Silk.NET.GLFW!Silk.NET.GLFW.GlfwCallbacks+CursorEnterCallback::Invoke'.
   at Silk.NET.GLFW.Glfw.PollEvents()
   at Silk.NET.Windowing.Internals.ViewImplementationBase.Silk.NET.Windowing.IView.DoEvents()

The crash timing is random but consistently happens. I am not able to reproduce this with the SDL backend.

Steps to reproduce

• Platform: Desktop
• Framework Version: .NET Core 6, .NET Core 7
• API: OpenGL, GLFW
• API Version: OpenGL 4.5

1. Instantiates two or more Silk windows
2. Instantiate an InputContext (window.CreateInput()) for each window
3. Run the render loop
4. Allow the mouse to traverse between windows to allow for the mouse enter message to event through
5. Observe crashing exception (note: likely will not happen immediately)

Comments

I've been able to reproduce this on two separate workstations with different hardware profiles in different code bases. I have seen other garbage collected delegate exceptions but most involve some sort of mouse input delegate attached to Glfw.PollEvent().

Minimal example:

using Silk.NET.Windowing;
using Silk.NET.Windowing.Extensions;
using Silk.NET.Input;
using Silk.NET.OpenGL;

namespace ogl;

public class OpenGL
{
    private WindowManager windowManager;
    private List<IInputContext> inputContexts = new List<IInputContext>();

    public OpenGL()
    {
        windowManager = new WindowManager();

        for (int i = 0; i < 4; i++)
            windowManager.CreateWindow(WindowOptions.Default);        
    }

    public void Start()
    {
        foreach(var window in windowManager.Windows)
        {
            window.Load += () =>
            {
                var inputContext = window.CreateInput();
                inputContexts.Add(inputContext);
            };

            window.Render += (time) =>
            {
                var context = GL.GetApi(window.GLContext);

                context.Enable(EnableCap.DepthTest);
                context.ClearColor(0f, .2f, .2f, 1f);
                context.Clear(ClearBufferMask.DepthBufferBit | ClearBufferMask.ColorBufferBit);
            };
        }

        windowManager.Run();
    }
}

[Perksey]

Wow, it's neat that they added a nice exception message.

Yep know exactly what's causing this, will investigate this on Friday.

[Perksey]

I've pushed up a PR that changes the codegen like so:

        public unsafe partial global::Silk.NET.GLFW.GlfwCallbacks.CharCallback SetCharCallback(global::Silk.NET.GLFW.WindowHandle* window, global::Silk.NET.GLFW.GlfwCallbacks.CharCallback callback)
        {
            global::System.IntPtr n4;
            ; // BEGIN RESOLVE 1 | READCOUNT: 1 | ALLOW_INLINE: True | FORCE_INLINE: False | CACHED: False
            // RESOLVED 1 INLINED
            // END RESOLVE 1
            this.PinUntilNextCall((/* VAR 1 */
-           callback), -712178223); // BEGIN RESOLVE 4 | READCOUNT: 1 | ALLOW_INLINE: False | FORCE_INLINE: False | CACHED: False
+           callback), global::System.HashCode.Combine(-712178223, (nint)window)); // BEGIN RESOLVE 4 | READCOUNT: 1 | ALLOW_INLINE: False | FORCE_INLINE: False | CACHED: False
            // BEGIN RESOLVE 0 | READCOUNT: 1 | ALLOW_INLINE: True | FORCE_INLINE: False | CACHED: False
            // RESOLVED 0 INLINED
            // END RESOLVE 0
            // BEGIN RESOLVE 2 | READCOUNT: 1 | ALLOW_INLINE: True | FORCE_INLINE: False | CACHED: False
            // ALREADY RESOLVED 1
            // RESOLVED 2 INLINED
            // END RESOLVE 2
            // RESOLVED 4 TO VARIABLE n4
            n4 = ((delegate* unmanaged[Cdecl]<global::Silk.NET.GLFW.WindowHandle*, global::System.IntPtr, global::System.IntPtr> )(CurrentVTable as GeneratedVTable).glfwSetCharCallback)((/* VAR 0 */
            window), (/* VAR 1 */
            /* VAR 2 */
            callback == null ? IntPtr.Zero : System.Runtime.InteropServices.Marshal.GetFunctionPointerForDelegate(/* VAR 1 */
            callback))); // END RESOLVE 4
            // BEGIN RESOLVE 3 | READCOUNT: 1 | ALLOW_INLINE: True | FORCE_INLINE: False | CACHED: False
            // ALREADY RESOLVED 4
            // RESOLVED 3 INLINED
            // END RESOLVE 3
            return (/* VAR 4 */
            /* VAR 3 */
            n4 == IntPtr.Zero ? null : System.Runtime.InteropServices.Marshal.GetDelegateForFunctionPointer<global::Silk.NET.GLFW.GlfwCallbacks.CharCallback>(/* VAR 4 */
            n4));
        }

This will ensure we discriminate against both the EntryPoint hashcode as well as the hashcode of whatever related discriminants (in this case the window handle) when unpinning on next call

[Perksey]

By the way we're keen to get feedback on WindowManager - it's a quite obscure part of the library and we're definitely looking to make the multi-window experience better.

[desiwko]

By the way we're keen to get feedback on WindowManager - it's a quite obscure part of the library and we're definitely looking to make the multi-window experience better.

I bet! I'll honestly say I didn't even know it existed until I wanted to investigate this issue a little bit more myself and ended up using it for the minimal example above. I will investigate using WindowManager a bit more as part of the reason multiple windows is of interest to me is in trying to get the ImGui multiple viewports working. This issue has been a blocker on that front though, so I appreciate the PR!

[Arugin]

I got the same crash. Is there any pre release version with this fix that I could get?

[Perksey]

https://dotnet.github.io/Silk.NET/docs/silk.net/experimental-feed.html