Bevy segfaults on some Windows systems after Rust 1.79 update

[mockersf]

The issue was first noticed in CI: https://github.com/bevyengine/bevy/actions/runs/9505180010/job/26199467296

The same job works on Rust 1.78: https://github.com/mockersf/bevy/actions/runs/9506123706/job/26202601792

A few users reported the same error, so it's not just on virtualised hardware. It's also not on all Windows machines

It seems to be triggered when rendering. Running an example from the Bevy repository that renders something reliably causes a segfault on some systems when using DirectX12. Running directly examples from wgpu don't seem to cause the crash.

A user report that it fails on nightly-2024-04-18, and works on nightly-2024-04-17

Sorry that's not much to go on, but I don't have an affected system available.

[lqd]

A few users reported the same error, so it's not just on virtualised hardware. It's also not on all Windows machines

https://github.com/rust-lang/cargo-bisect-rustc would help locating the nightly (and validate it’s nightly-2024-04-18) and in particular the PR where this issue appeared.

[janhohenheim]

I'm a bit sleepy, but I'll try cargo-bisect-rustc now and see how far I get
Update 1: Running the following now

$env:WGPU_BACKEND='dx12'; cargo-bisect-rustc --start=2024-04-17 --end=2024-04-18 --prompt --with-src -- run --example alien_cake_addict

Update 2: Can confirm the reported range (fails on nightly-2024-04-18, and works on nightly-2024-04-17) is correct
Update 3: Let's hope it's not a rollup PR, since that won't get us the exact failing commit on a Windows-only bug:

And even further, if the regression is in a rollup PR, then it will bisect the individual PRs within the rollup. This final bisection is only available for x86_64-unknown-linux-gnu since it is using the builds made for the rustc performance tracker.

[janhohenheim]

Per another user running the tools, the offending commit is 38104f3
I'm still running the bisect to confirm this.

[lqd]

cc PR author @Mark-Simulacrum on that bisection

[janhohenheim]

My bisection is done and I can confirm the result. This is the output:

searched nightlies: from nightly-2024-04-17 to nightly-2024-04-18
regressed nightly: nightly-2024-04-18
searched commit range: 1cec373...becebb3
regressed commit: 38104f3

bisected with cargo-bisect-rustc v0.6.8

Host triple: x86_64-pc-windows-msvc
Reproduce with:

cargo bisect-rustc --end=2024-04-18 --prompt --with-src -- run --example alien_cake_addict

note by me: also requires $env:WGPU_BACKEND='dx12'; to actually get the crash

[alice-i-cecile]

The offending PR: #123936

[workingjubilee]

If that is the offending commit, then it seems very unlikely that this has not unveiled a bug inside wgpu or whatever else is busy talking to DX12.

[workingjubilee]

Can you get a backtrace in WinDbg or whatever's fashionable?

Or can a "minimalized"¹ repro case be produced?

Footnotes

1. where "minimalized" is "slightly less than the entire .exe there", I am aware it is not going to be small ↩

[Brezak]

When looking at WinDbg the issue seems to arise here. wgpu-hal was converting a normal rust str into a c_char ptr and passing that across ffi to a function expecting a null terminated string. Unluckily for us this string was a empty string "". Before #123936 that meant we'd get a pointer to somewhere we can read from. After the PR we get a 0x1 pointer that will cause a segfault when read from.

[workingjubilee]

Agh, that is the classic error that I predicted in my remarks here.

[workingjubilee]

This passes, so changing to c"" should be fine.

#[test]
fn cstrs_dont_dangle() {
    let c_str = c"";
    assert_ne!(c_str.as_ptr(), std::ptr::NonNull::dangling().as_ptr());
}

However... @Brezak Are you sure that's the issue? It seems that this string was created from a CString::new()?

https://github.com/gfx-rs/wgpu/blob/d7a35ecda06742acb4cdd5a83c5b9737f375a70b/wgpu-hal/src/dx12/device.rs#L281-L288

Does that CString actually survive the lifetime of that function call?