window example fails on Weston and Mesa 17.3.0 (and a workaround)

[antonycourtney]

If I build glutin at HEAD and run the window example on Ubuntu (16.04) using Mesa 17.3.0 and Weston 1.11.90, it fails with:

$ ./target/debug/examples/window
thread 'main' panicked at 'eglCreateContext failed: 0x3009', src/api/egl/mod.rs:775:18

Digging in a bit and doing some printf debugging, I found this failing because the call to eglCreateContext from create_context in mod.rs is attempting to pass (3,2) as the (major,minor) version pair for ffi::egl::CONTEXT_MAJOR_VERSION and ffi::egl::CONTEXT_MINOR_VERSION. Mesa returns ffi::egl::BAD_MATCH in this case, but the error handler in create_context() in api/egl/mod.rs only handles ffi::egl::BAD_ATTRIBUTE and panics otherwise.

Tracking down where these version numbers are coming from, I found the following: First, in Context::new, we attempt to get the EGL version numbers (egl_version) via a call to egl.Initialize: https://github.com/tomaka/glutin/blob/master/src/api/egl/mod.rs#L187

Then, based on this EGL version number, we attempt to bind to different variants of OpenGL APIs, based on the specific OpenGL API requested in opengl.version: https://github.com/tomaka/glutin/blob/master/src/api/egl/mod.rs#L211

In the current example, we end up getting 1.4 back from EGL, opengl.version is GlRequest::Latest, and the call to egl.BindAPI(ffi::egl::OPENGL_API) succeeds, so (version,api) is bound to (None, Api::OpenGl). The value GLRequest::Latest bound to opengl.version appears to be the default value for self.gl_attr.version for ContextBuilder.

My straightforward workaround is to request OpenGlEs instead of OpenGL and specify (2, 0) as the version pair by changing this line of window.rs so it reads:

    let context = glutin::ContextBuilder::new().with_gl(GlRequest::Specific(Api::OpenGlEs, (2, 0)));

(Of course I also added the appropriate use calls). With this tiny change, the example works again.

But, of course, this is just a workaround. I'm new to GL programming, EGL and Rust, so I'm not entirely sure what the correct fix would be, but a couple of speculative ideas:

1. Perhaps the error handler in create_context() should handle BAD_MATCH in addition to BAD_ATTRIBUTE, so that the call to eglCreateContext returns an error instead of calling panic. I think that would let the logic in ContextPrototype::finish_impl here make a couple of more tries which would eventually call create_context with (1, 0) as the version pair.

2. Perhaps the version numbers being passed to eglCreateContext are just off, or are being passed in the wrong way? One thing I noticed during printf debugging that seems a bit suspect to me is that ffi::egl::CONTEXT_CLIENT_VERSION and ffi::egl::CONTEXT_MAJOR_VERSION have the same value.

I notice that #988 seems to fail in the same way (panic during Egl context creation), so perhaps it is the same issue and could possibly use the same workaround.

[goddessfreya]

Hey, could you post the results of glxinfo && glinfo && lspci?