Resize error on DX12 backend in wgpu 0.17

[mikeny07]

Description

when attempting to run wgpu examples using the DX12 backend in version 0.17 of wgpu, I encountered the following error:

[2023-07-23T20:02:29Z ERROR wgpu_hal::dx12] ResizeBuffers failed: 0x887A0001
[2023-07-23T20:02:29Z ERROR wgpu_core::device::global] surface configuration failed: window is in use
thread 'main' panicked at 'Error in Surface::configure: Validation Error

Caused by:
Invalid surface
', C:\Users****.cargo\registry\src\index.crates.io-6f17d22bba15001f\wgpu-0.17.0\src\backend\direct.rs:771:18

Repro steps

For example, in the "hello-triangle" example, simply replacing

let instance = wgpu::Instance::default();

by

 let instance = wgpu::Instance::new(wgpu::InstanceDescriptor {
        backends: wgpu::Backends::DX12,
        dx12_shader_compiler: Default::default(),
 });

this will generate the error when resizing the output window.

Expected vs observed behavior

This error only occurs in wgpu 0.17, while wgpu <= 0.16 never gives this error.

Platform

Windows 11, RTX 3060, wgpu 0.17.

[ludovitkramar]

I'm learning wgpu and I have this same error on Windows 10 with GT740.
It doesn't matter which present mode I choose, the DX12 backend produces the same error when resizing. (or just pressing maximize once.)

[2023-07-31T07:01:14Z ERROR wgpu_hal::dx12] ResizeBuffers failed: 0x887A0001
[2023-07-31T07:01:14Z ERROR wgpu_core::device::global] surface configuration failed: window is in use
thread 'main' panicked at 'Error in Surface::configure: Validation Error

Caused by:
    Invalid surface

Can confirm that 0.16 works as expected.

[teoxoy]

Traced it back to #3936.

Looking at the Remarks section of ResizeBuffers,

You can't resize a swap chain unless you release all outstanding references to its back buffers. You must release all of its direct and indirect references on the back buffers in order for ResizeBuffers to succeed.

Direct references are held by the application after it calls AddRef on a resource.

Indirect references are held by views to a resource, binding a view of the resource to a device context, a command list that used the resource, a command list that used a view to that resource, a command list that executed another command list that used the resource, and so on.

from https://learn.microsoft.com/en-us/windows/win32/api/dxgi/nf-dxgi-idxgiswapchain-resizebuffers#remarks

and us cloning the ComPtr here,

wgpu/wgpu-hal/src/dx12/mod.rs

Line 796 in 28334ac

resource: sc.resources[index].clone(),

it's plausible we still hold onto the resources when calling ResizeBuffers and not respecting its invariants.

[mikeny07]

Why this error only occurs when running the official examples in wgpu 0.17, but not in older versions such as 0.16?

"You can't resize a swap chain unless you release all outstanding references to its back buffers. You must release all of its direct and indirect references on the back buffers in order for ResizeBuffers to succeed."

Why older versions didn't have such issue? what change from older versions to 0.17 leads to this problem?

[teoxoy]

The PR I linked to above #3936.

[Capati]

Using wgpu::Device::poll before recreate swapchain didn't produce the error. This function check for resource cleanups, is this a bug or the proper way to do?

[cwfitzgerald]

It's a bug on our side, we need to do the appropriate waits internally.

[Elabajaba]

Ran into this updating bevy to wgpu 0.17. It also crashes when you minimize the window.

[DotrixXP]

Calling wgpu::Instance::poll_all before reconfiguring surface and recreating framebuffers worked for me.