Use EGL surfaceless platform when windowing system is not found

[dsseng]

Connections
Addresses #1551

Description
Falling back to egl::DEFAULT_DISPLAY usually results in X11 EGL platform being picked and then rejected because of unavailability on a head/GPU-less system. EGL_PLATFORM_SURFACELESS_MESA works with both radeonsi and llvmpipe/swrast when Xorg/Wayland sockets are being hidden from application. Needs to be tested in a truly GPU-less environment such as CI it is required to run in.

Testing
MESA_LOADER_DRIVER_OVERRIDE=llvmpipe RUST_BACKTRACE=1 RUST_LOG=info WAYLAND_DISPLAY= DISPLAY= WGPU_BACKEND=gl cargo run --example capture
MESA_LOADER_DRIVER_OVERRIDE=llvmpipe RUST_BACKTRACE=1 RUST_LOG=info WAYLAND_DISPLAY= DISPLAY= WGPU_BACKEND=vulkan cargo run --example capture

Tested on a desktop, but without X/Wayland accessible for test program (same conditions led to eglInitialize error). red.png is being created and has proper content.

@ maintainers please test in CI

[dsseng]

If CI tests pass even without this, something definitely needs to be enabled there. Could you please enable those tests (or instruct me to do so if it's in code) to see if they work fine in CI with surfaceless.

[cwfitzgerald]

https://github.com/gfx-rs/wgpu/blob/master/.github/workflows/ci.yml#L67 if you add gl back to that should being running all the basic tests on GL as well as vulkan.

[dsseng]

Thanks, that worked. Test failure is in skybox_etc2 unit test. Going to diagnose locally forcing llvmpipe. Only one failure, so might be some llvmpipe limit?

[cwfitzgerald]

You just need to bump the outlier count in that test: bump https://github.com/gfx-rs/wgpu/blob/master/wgpu/examples/skybox/main.rs#L502 to 110 or something, it's reporting 102 outliers.

[dsseng]

Okay, thank you!

[dsseng]

Couldn't find the exact failure 🤷

[cwfitzgerald]

I can't either, idk why it is returing -1...

[dsseng]

There're some panics, but test logs just seem messy overall (some panics and backtraces) both on CI and my system (both radeonsi/RADV and llvmpipe).

[dsseng]

Failure is related to skybox and texture compression again for some reason.

[dsseng]

test skybox_astc ... 102 outliers over max difference 90
ok
error: XDG_RUNTIME_DIR not set in the environment.
[2022-01-03T19:29:21Z ERROR wgpu_hal::gles::egl] EGL 'eglInitialize' code 0x3001: DRI2: failed to load driver
test skybox_bc1 ... TEST SKIPPED: MISSING FEATURES TEXTURE_COMPRESSION_BC
ok
error: XDG_RUNTIME_DIR not set in the environment.
[2022-01-03T19:29:21Z ERROR wgpu_hal::gles::egl] EGL 'eglInitialize' code 0x3001: DRI2: failed to load driver
test skybox_etc2 ... 102 outliers over max difference 90
ok

test result: ok. 4 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out; finished in 7.49s

I have previously increased max outliers to 105. Weird. Why does it say 4 passed while tests themselves fail and suite fails?

[kvark]

I have previously increased max outliers to 105. Weird. Why does it say 4 passed while tests themselves fail and suite fails?

the observed outlier numbers are lower, that's why it reports "ok".

[dsseng]

How do I fix that test? What's the exact cause of pipeline failure?

[kvark]

From what I see, the failure is:

error: XDG_RUNTIME_DIR not set in the environment.
[2022-01-03T19:29:15Z ERROR wgpu_hal::gles::egl] EGL 'eglInitialize' code 0x3001: DRI2: failed to load driver
thread 'main' panicked at 'Image data mismatch! Outlier count 1242592 over limit 3. Max difference 91', wgpu/examples/skybox/../../tests/common/image.rs:134:13

So we need to find out why it fails to load the driver. Maybe some packages are missing? Or XDG_RUNTIME_DIR environment needs to be set?

[dsseng]

This driver-related message have been seen earlier (on my machine with llvmpipe), but everything worked.

[dsseng]

Is there a way to extract the failing test snapshot as a build artifact to inspect it? Probably the test can be skipped since it's not directly related to EGL surface creation?

[kvark]

You can probably modify the CI in your fork of the repo and upload test results as artifacts to githubs.
I agree it may not be strictly necessary to fix it, given that GL was disabled on CI before this PR. So we can:

1. either blocklist the test on GL somehow
2. not enable GL on CI at all in this PR, leave for follow-ups

[dsseng]

Probably first option seems preferable. Precisely we should disable (return passed) the test in renderer is llvmpipe

[cwfitzgerald]

The weird thing is the test passes:

test skybox ... GOT EXPECTED TEST FAILURE: BACKEND
** <snip> **
ok

I'm not sure what is returning -1 from here.

[dsseng]

Any ideas on merging this? Some rules to blocklist those tests on llvmpipe/CI/headless?

[kvark]

Let's merge it without enabling GL testing on CI, so that the important code isn't left hanging around.

[dsseng]

Maybe skip that test only?

[cwfitzgerald]

The issue is I can't see any test in the set that is failing, it passes all the tests, then dies.

[cwfitzgerald]

Hey, sorry for letting this languish, lets nix the CI change and merge this in.

[cwfitzgerald]

Upgrading to a new test framework in #2495 pinpointed what the issue was right away. Thank you for the contribution and sorry again for the delay!

[dsseng]

No need to say sorry! Thank you for reviews and collaboration! It's great to work with your project!