potential dangling resources in CommandListResourceStateTracker

[BlurryLight]

A minimal repoducable example is as following:

  void TestRelease() {
    nvrhi::BufferDesc desc{};
    desc.keepInitialState = true;
    desc.initialState = nvrhi::ResourceStates::Common;
    desc.byteSize = 4;

    auto CmdList = GetDevice()->createCommandList();
    CmdList->open();
    {
      auto buf = GetDevice()->createBuffer(desc);
      CmdList->setBufferState(buf,nvrhi::ResourceStates::ConstantBuffer);
    }
    CmdList->close();
  }

It's because CommandListResourceStateTracker store all resources references in raw pointers without ref counter.
When cmdlists are about to close, they try to iterate over these resources ( keepBufferInitialStates / keepTextureInitialStates) and may trigger a crash because these resources may have been freed.

potential fix

CommandListResourceStateTracker might extend resources lifetime until closed.

[BlurryLight]

there is another dangling case: CommandList::Set(Permanant)Buffer/TextureState don't record reference in TrackedCommandBuffer.referencedResources;. The resources may have been freed when pipelinebarriers haven't been executed on GPU.

[apanteleev]

Thank you for reporting the issue. I figured that storing strong references inside the ..StateTracker class is overkill because its methods are called from many CommandList functions, and those functions already keep strong refs to their resources. Also, IResource* type pointers are not passed to the state tracker at all. So, the user-facing set(Texture|Buffer|AccelStruct)State functions now store strong refs explicitly.
Please let me know if I missed anything.