Using two encoders in hello-triangle example causes hang

[fintelia]

Description
Wgpu hangs when rendering a frame that involved multiple consecutive command encoders. This patch to the hello-triangle example is enough to trigger the bug.

Repro steps

$ git clone https://github.com/fintelia/wgpu && cd wgpu
$ git checkout double-encoder-hang
$ cd wgpu
$ cargo run --example hello-triangle

Expected vs observed behavior
Expected hello-triangle example to run normally. Instead it displays all black and the window cannot be closed.

Platform

• Arch Linux
• Intel integrated GPU
• wgpu commit 39e8921

[fintelia]

Using git bisect I've narrowed down this regression to being caused by 5419947:

$ git checkout 39e8921
$ git bisect start
$ git bisect bad
$ git bisect good 91437ea7
$ git bisect test.sh

test.sh

#!/bin/bash

cat >test.patch << EOF
diff --git a/wgpu/examples/hello-triangle/main.rs b/wgpu/examples/hello-triangle/main.rs
index 4d27af83..25cfef21 100644
--- a/wgpu/examples/hello-triangle/main.rs
+++ b/wgpu/examples/hello-triangle/main.rs
@@ -95,6 +95,11 @@ async fn run(event_loop: EventLoop<()>, window: Window) {
                     .get_current_frame()
                     .expect("Failed to acquire next swap chain texture")
                     .output;
+
+                *control_flow = ControlFlow::Exit;
+                let e2 = device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });
+                queue.submit(Some(e2.finish()));
+
                 let mut encoder =
                     device.create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });
                 {
EOF

cd wgpu/examples
if git apply test.patch && cargo build --example hello-triangle
then
	# run project specific test and report its status
    timeout 15 cargo run --example hello-triangle
	status=$?
else
	# tell the caller this is untestable
	status=125
fi
cd ../..

# undo the tweak to allow clean flipping to the next commit
git reset --hard

# return control
exit $status

[kvark]

Thank you for investigating!
I wonder if our use of the relay semaphore is incorrect. Vulkan validation has a new experimental option for synchronization, so that could be the next step to enable and see.

[fintelia]

I just tried running with the validation layers included with SDK version 1.2.182.0, but they produced no output. Here is a copy of the API dump. The hang seems to be happening during the call to vkQueuePresentKHR, which matches the backtrace I got when pausing in a debugger.

api_dump.txt

[kvark]

Not seeing any issue with this on Linux/NV
Edit: also no issues on Windows/Intel 550 ...

[mrk-its]

Ubuntu 20.04.2, HD Graphics 630 (KBL GT2)

I've just tried to reproduce this and I have black window not reacting on events for ~8 seconds, then triangle is displayed (and window can be closed). But it seems it is not related to 'using two encoders', I have the same behavior also on current wgpu master. When I force GL backend instead of VULCAN it works fine (triangle is displayed immediately)

[fintelia]

I'm using an Intel® HD Graphics 520 iGPU with mesa 21.2.1-1. There's no other GPU so the linked PR doesn't seem applicable.

However reading #1672, I think this might be a duplicate of that given that the hang in vkQueuePresentKHR is during an ioctl