[wgpu-hal] fall back from broken VK_KHR_timeline_semaphore impls.

[eddyb]

Checklist

• Run cargo clippy.
• Run RUSTFLAGS=--cfg=web_sys_unstable_apis cargo clippy --target wasm32-unknown-unknown if applicable.
• Add change to CHANGELOG.md. See simple instructions inside file.

Connections
Link to the issues addressed by this PR, or dependent PRs in other repositories

Description
I encountered this while testing an wgpu example on the Oculus Quest 2, but the important part is this:

04-14 14:43:26.051 11925 11925 W Thread-1: type=1400 audit(0.0:321): avc: denied { ioctl } for path="/dev/kgsl-3d0" dev="tmpfs" ino=1206 ioctlcmd=0x958 scontext=u:r:untrusted_app:s0:c94,c256,c512,c768 tcontext=u:object_r:gpu_device:s0 tclass=chr_file permissive=0
...
04-14 14:43:26.054 11925 11950 I AdrenoVK-0: QUALCOMM build          : d66bb48e17, I18aabff2be
04-14 14:43:26.054 11925 11950 I AdrenoVK-0: Build Date              : 11/28/22
...
04-14 14:43:26.057 11925 11950 D wgpu_hal::vulkan::ada..: Supported extensions: ["VK_KHR_swapchain", "VK_KHR_image_format_list", "VK_KHR_imageless_framebuffer", "VK_KHR_driver_properties", "VK_KHR_timeline_semaphore", "VK_EXT_image_robustness", "VK_EXT_robustness2"]
04-14 14:43:26.059 11925 11950 W wgpu_hal::vulkan: Unrecognized device error ERROR_INITIALIZATION_FAILED
04-15 10:42:47.264 11611 11635 E wgpu::backend::direct: Error in Adapter::request_device: Not enough memory left

So my understanding is roughly as follows:

• Oculus Quest 2's latest update includes an AdrenoVK version that advertises VK_KHR_timeline_semaphore
• calling VkCreateSemaphore to use the extension fails due to SELinux(?) rules not having been updated
  • AFAICT path="/dev/kgsl-3d0" ... ioctlcmd=0x958 is IOCTL_KGSL_TIMELINE_CREATE
  • searching around seems to confirm other users have ran into the same bug: https://communityforums.atmeta.com/t5/Quest-Development/Issue-with-VK-KHR-timeline-semaphore/td-p/1027690
  • absolutely no idea why that SELinux(?) audit message shows out of of order
    • (maybe that's an interaction between android_logger and other Android log buffers?)
• before this PR: VkCreateSemaphore failing w/ ERROR_INITIALIZATION_FAILED leads to wgpu::Adapter::request_device returning Err(RequestDeviceError)
• after this PR: that specific error is special-cased and the fallback (fence pool) implementation is used
  • more logging would probably be a good idea, also treating unknown errors as "out of memory" seems wrong?

Testing
The same wgpu-using demo that failed with the above, works with this change, and as expected it logs:

04-15 11:11:13.472 13889 13915 W wgpu_hal::vulkan::dev..: `VkCreateSemaphore` failed with `ERROR_INITIALIZATION_FAILED`, falling back to fence pool

[cwfitzgerald]

Could you link back to the issue explaining why this is needed?

[cwfitzgerald]

This actually isn't really a warning, it's at most an info, maybe even a debug. The user can't do anything about it, and it has absolutely no effect on the user either.

[cwfitzgerald]

Poke from triage

[cwfitzgerald]

As this MR is currently stale, I'm going to close it for triage purposes. Please feel-free to re-open there is renewed interest in landing these changes!