Running wgpu in a secondary thread segfaults when main thread ends.

[rukai]

OS: Arch Linux
GPU: GTX 960
Driver: Nvidia proprietary driver

fn main() {
    std::thread::spawn(move || {
        triangle_example();
    });

    std::thread::sleep(std::time::Duration::from_millis(500));
}

Running the triangle example in a separate thread and then ending the main thread causes the wgpu thread to segfault.

repro: https://github.com/rukai/wgpu-rs/blob/segfault_repro/examples/repro/main.rs

[Aaron1011]

Minimized reproduction: https://github.com/Aaron1011/wgpu-segfault

The issue occurs in gfx_backend_gl, which uses libloading to dynamically load the OpenGL shared library: https://github.com/gfx-rs/gfx/blob/07043744b4b05893363066ca0e50c5e38348f20c/src/backend/gl/src/window/egl.rs#L20-L23

When the libloading::Instance gets dropped when the thread function returns, it unloads the shared library. However, it appears that a TLS destructor remains registered for the shared library, despite the code being unloaded. When libc attempts to run the registered TLS destrutors on thread exit (after the Rust code has run), we get a segfault.

See nagisa/rust_libloading#41 for more details

[kvark]

@Aaron1011 any idea on what we should be doing?

[Aaron1011]

@kvark: There's talk about changing libloading to only call dlopen when explicitly requested (instead of automatically on Drop): nagisa/rust_libloading#41 (comment)

[Aaron1011]

For now, the best solution is probably to do something similar to this workaround: nagisa/rust_libloading#41 (comment)

Specifically, gfx_backend_gl should open the graphics driver shared library with RTLD_NODELETE.

[Aaron1011]

From my testing, it looks like unloading libX11.so doesn't cause any problems - it's unloading libEGL.so that causes problems. For example, NVIDIA's libglvnd registers a pthread key destructor