Does Direct3D11CaptureFramePool actually require a DispatcherQueue?

[jpark37]

Despite the comment saying a DispatcherQueue is required, I don't instantiate one in my application, and window capture seems to work fine. What are the consequences of not doing this?

[jpark37]

I ask because we're hearing infrequent reports of shutdown hangs, and I'm wondering if this is the reason.

[robmikh]

When the frame pool is created with Direct3D11CaptureFramePool::Create, we use the DispatcherQueue to fire the frame arrived event to the thread that created the frame pool. I could have sworn calling Create without a dispatcher queue present would cause us to blow up in an internal call. Although testing it now, this behavior seems to have regressed...

Regardless, the most important part of the DispatcherQueue requirement is that you are pumping messages on that thread. As long as you're doing that, I wouldn't expect it to be the problem. Usually when I see odd behavior on tear-down, it was because I wasn't properly initializing COM on the thread. This can cause odd life-time related bugs, and may be what you're seeing but I can't be certain.

I know that you mentioned hearing infrequent reports of this, but if you get a consistent repro or a memory dump when this happens I can take a look.

[robmikh]

Additionally, I would recommend creating a DispatcherQueue for your thread by calling CreateDispatcherQueueController. The fact that it's working without it has me a bit concerned about if the system changes in the future.

Or you can forgo the dispatcher requirement all together and use Direct3D11CaptureFramePool::CreateFreeThreaded, and then use your own mechanism to marshal the frame to a thread of your choosing.

[jpark37]

I can let a dispatcher queue live for the life of the thread. For simplicity, we do all WGC things on our render thread since we have that's where our D3D11 renderer operates. I suppose this would be advised?

void render_thread()
{
    // Add dispatcher queue here.
    queue = ...

    while (running) {
        tick sources, possibly create/destroy WGC captures
        render sources
        pump win32 messages until empty <- frame arrival callback, D3D11 texture copy here
    }

    // Let queue fall out of scope I guess?
}

We have a separate message pump that lives on the main thread that is implicit to Qt. Hopefully these aren't fighting each other, but my Win32/COM knowledge isn't that strong. The render thread pump was added recently just to support WGC.

I can find out if we have any memory dumps.

[robmikh]

That sounds reasonable to me, and having the two different pumps on two different threads should be ok. The only thing I would add is that I would add a call to RoInitialize (or in your case, winrt::init_apartmnet, since you seem to be using C++/WinRT) before interacting/activating any WinRT objects on that thread.

[jpark37]

Thanks, you reminded me I didn't set up the winrt apartment either. I also noticed we were keeping WGC objects alive across render threads (we kill and recreate the thread in some situations), so I added a fix to destroy on thread stop and recreate on thread start. Hopefully everything is robust and stable now.

[notr1ch]

I have a couple of memory dumps from the shutdown freezes we experienced, these are full heap dumps so I don't want to link them publicly. The WGC thread was stuck in a COM call to the "Capture Service" process. Debugging the service showed no stuck threads, almost like it wasn't aware of the COM call.

rpcrt4!LRPC_BASE_CCALL::DoSendReceive+0x112
rpcrt4!LRPC_CCALL::SendReceive+0x51
rpcrt4!I_RpcSendReceive+0x6f
combase!CMessageCall::CallI_RpcSendReceive+0x30 [onecore\com\combase\dcomrem\call.cxx @ 3956] 
combase!ThreadSendReceive+0xf0 [onecore\com\combase\dcomrem\channelb.cxx @ 7316] 
combase!CSyncClientCall::SwitchAptAndDispatchCall+0xa0 [onecore\com\combase\dcomrem\channelb.cxx @ 5742] 
combase!CSyncClientCall::SendReceive2+0x182 [onecore\com\combase\dcomrem\channelb.cxx @ 5377] 
combase!SyncClientCallRetryContext::SendReceiveWithRetry+0x25 [onecore\com\combase\dcomrem\callctrl.cxx @ 1617] 
combase!CSyncClientCall::SendReceiveInRetryContext+0x25 [onecore\com\combase\dcomrem\callctrl.cxx @ 567] 
combase!DefaultSendReceive+0x64 [onecore\com\combase\dcomrem\callctrl.cxx @ 525] 
combase!CSyncClientCall::SendReceive+0x128 [onecore\com\combase\dcomrem\ctxchnl.cxx @ 783] 
combase!CClientChannel::SendReceive+0x84 [onecore\com\combase\dcomrem\ctxchnl.cxx @ 653] 
combase!NdrExtpProxySendReceive+0x4e [onecore\com\combase\ndr\ndrole\proxy.cxx @ 2002] 
rpcrt4!NdrpClientCall3+0x395
combase!ObjectStublessClient+0x13b [onecore\com\combase\ndr\ndrole\amd64\stblsclt.cxx @ 369] 
combase!ObjectStubless+0x42 [onecore\com\combase\ndr\ndrole\amd64\stubless.asm @ 176] 
GraphicsCapture!winrt::Windows::Graphics::Capture::implementation::ServerCaptureSessionCore::StopCapture+0x15
GraphicsCapture!winrt::Windows::Graphics::Capture::implementation::GraphicsCaptureSession::Close+0x34
GraphicsCapture!winrt::impl::produce<winrt::Windows::Graphics::Capture::implementation::GraphicsCaptureSession,winrt::Windows::Foundation::IClosable>::Close+0x1f
libobs_winrt!winrt::impl::consume_Windows_Foundation_IClosable<winrt::Windows::Graphics::Capture::GraphicsCaptureSession>::Close+0x26 [c:\program files (x86)\windows kits\10\include\10.0.18362.0\cppwinrt\winrt\windows.foundation.h @ 17] 
libobs_winrt!winrt_capture_free+0xa3 [d:\obs2\libobs-winrt\winrt-capture.cpp @ 453] 
win_capture!wc_tick+0xb6 [d:\obs2\plugins\win-capture\window-capture.c @ 402] 
obs!obs_source_video_tick+0x1a2 [d:\obs2\libobs\obs-source.c @ 1159] 
obs!tick_sources+0x103 [d:\obs2\libobs\obs-video.c @ 71] 
obs!obs_graphics_thread+0x204 [d:\obs2\libobs\obs-video.c @ 905] 
w32_pthreads!ptw32_threadStart+0x8a [d:\obs2\deps\w32-pthreads\ptw32_threadstart.c @ 225] 
ucrtbase!thread_start<unsigned int (__cdecl*)(void *),1>+0x42
kernel32!BaseThreadInitThunk+0x14
ntdll!RtlUserThreadStart+0x21```

[robmikh]

Thanks, @notr1ch! It would be interesting to know if this still happens after initializing COM on the thread. If the proxy stub is blocked and the service is unaware of it, it might be an initialization issue.

If it persists, we can maybe talk through another channel about looking at the dump if you feel comfortable. A dump of the capture service would be interesting as well.

[robmikh]

Closing this issue. If you still see this after properly initializing COM, please let us know.

[jpark37]

We added the DispatchQueue and RoInitialize call, and we haven't seen the problem since, so we're probably good now.

I also added code to disable the yellow border recently, but I had to switch our thread from STA to MTA to get rid of the assert from blocking on the RequestAccessAsync result in-place. Do you think this is fine?