NVIDIA Linux drivers don't support GLX_BIND_TO_TEXTURE_RGBA_EXT

[kmcallister]

See mailing list discussion.

[kmcallister]

Confirmed that Mesa with Intel graphics doesn't have this problem.

[larsbergstrom]

The first part (GLX_BIND_TO_TEXTURE_RGBA_EXT) is a bug on our part. We were calling glxChooseVisual, which was giving us a random visual that supported RGBA but not for textures, and then walking through the fb_configs looking for the fb_config that matched the visual and failing. Just calling glXChooseFBConfig (when supported) and getting the visual from that clears that part up.

Fix:
larsbergstrom/rust-layers@5dc342c

The second part on the mailing list, though, about getting BadAlloc on GPU rendering (and BadMatch on CPU, not mentioned there but present for RGBA) seems a bit trickier. BadAlloc indicates that we've already associated a GLX FBConfig with that X pixmap. Patch notes indicated maybe fixed in newer Nvidia drivers (Ubuntu is on 319; Nvidia is now on 331). Or we're still doing things wrong.

[larsbergstrom]

Indeed, I've confirmed that when the BadAlloc crash happens, we are attempting to create a GLX pixmap against an X pixmap that has already had a GLX pixmap created and destroyed against it. That first creation/destruction happens outside of rust-layers and I'm still tracking that part down (since there are two other places apart from rust-layers where we call glXDestroyPixmap - both skia and azure). I suspect this isn't a problem on Intel drivers because it appears to be implementation-specific whether destroying the GLX pixmap will also destroy the X pixmap or not.

For future people debugging related issues, https://github.com/apitrace/apitrace is super useful.

[mletterle]

I can't tell you how happy I am that this is being looked at....