I am getting a stack overflow error from loading too many animations at once

[victor7543]

Bevy version

Bevy 0.14.2

[Optional] Relevant system information

using rust 1.81.0, Windows 11

AdapterInfo { name: "NVIDIA GeForce GTX 1070", vendor: 4318, device: 7041, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "561.09", backend: Vulkan }

What you did

I downloaded a model from mixamo with 50 animations and I am trying to load them with the usual method I saw in bevy's examples, but I am getting an error: thread 'IO Task Pool (0)' has overflowed its stack. Here's the part of the code that is causing the problem:

let mut graph = AnimationGraph::new();
    let animations = graph
        .add_clips(
            [
                GltfAssetLabel::Animation(0).from_asset("Paladin1.glb"),
                GltfAssetLabel::Animation(1).from_asset("Paladin1.glb"),
                ...
                GltfAssetLabel::Animation(49).from_asset("Paladin1.glb"),
            ]
            .into_iter()
            .map(|path| assets.load(path)),    // this seems to be causing the stack overflow, but only when I load too many animations 
            1.0,
            graph.root,
        )
        .collect();

    // Insert a resource with the current scene information
    let graph = graphs.add(graph);
    commands.insert_resource(Animations {
        animations,
        graph: graph.clone(),
    });

What went wrong

I am getting an error:

Running target\debug\bevy_test.exe
2024-09-17T17:43:47.913011Z INFO bevy_diagnostic::system_information_diagnostics_plugin::internal: SystemInfo { os: "Windows 11 Pro", kernel: "22631", cpu: "Intel(R) Core(TM) i5-9600KF CPU @ 3.70GHz", core_count: "6", memory: "31.9 GiB" }
2024-09-17T17:43:48.052848Z ERROR wgpu_hal::vulkan::instance: GENERAL [Loader Message (0x0)]
Failed to open dynamic library "C:\ProgramData\ReShade.\ReShade64.dll" with error 1114
2024-09-17T17:43:48.053205Z ERROR wgpu_hal::vulkan::instance: objects: (type: INSTANCE, hndl: 0x18d4bbefc00, name: ?)
2024-09-17T17:43:48.241307Z INFO bevy_render::renderer: AdapterInfo { name: "NVIDIA GeForce GTX 1070", vendor: 4318, device: 7041, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "561.09", backend: Vulkan }
2024-09-17T17:43:48.915565Z INFO bevy_winit::system: Creating new window "App" (Entity { index: 0, generation: 1 })
thread 'IO Task Pool (0)' has overflowed its stack
error: process didn't exit successfully: target\debug\bevy_test.exe (exit code: 0xc00000fd, STATUS_STACK_OVERFLOW)

If I try to load only 20 or 30 animations it works just fine, but when I try to load all 50 I get the error.

Additional information

From my own tests, it seems that the error is triggered by the call of assets.load(). I just tested by just commenting everything else in the code that was unnecessary.

It's actually really easy to reproduce this error. I just modified the animated_fox example to load the mixamo model and all it's animations instead of the fox and that's all it took. I am uploading the model and the modified animated_fox.rs file. Just place the model in the same folder as the fox and replace the example file.

code_and_model.zip

it's worth noting that I downloaded all animations with 60 fps instead of just 30.

[alice-i-cecile]

I wonder if we're not deduplicating asset dependencies properly 🤔 Regardless, we should definitely not be crashing here.

[john-sharratt]

Not sure if this is useful for your analysis but I'm also getting a stack overflow regularly in similar circumstances although my branch is a fork at the moment (I will submit pull requests later once its stable). Something I did was to use Cow with static buffers loaded via mmap to accelerate the game loading process significantly and that means many Gltf now loads heaps of things in parallel.

[john-sharratt]

I suspect it might be here... (bit of a guess though)
(looks like the async task engine is filling up the stack until it overflows, probably because it does an inline block_on)

[john-sharratt]

Yeah I believe its here...

    // TODO: use the threaded impl on wasm once wasm thread pool doesn't deadlock on it
    // See https://github.com/bevyengine/bevy/issues/1924 for more details
    // The taskpool use is also avoided when there is only one texture for performance reasons and
    // to avoid https://github.com/bevyengine/bevy/pull/2725
    // PERF: could this be a Vec instead? Are gltf texture indices dense?
    fn process_loaded_texture(
        load_context: &mut LoadContext,
        handles: &mut Vec<Handle<Image>>,
        texture: ImageOrPath,
    ) {

The processing of textures after loading the Gltf invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline), which loads another Gltf that invokes the task engine again (inline)... stack overflow!