fix(💥): detach swapchains before destroying the device (fixes Android Canvas unmount crash)

[eduardoborges]

Fixes the native crash reported in #468.

The bug

When a <Canvas> unmounts on Android, the app crashes about 2 out of 3 times with a SIGSEGV in Dawn's Vulkan backend:

F libc    : Fatal signal 11 (SIGSEGV), fault addr 0x20
  dawn::native::vulkan::FencedDeleter::DeleteWhenUnused
  dawn::native::vulkan::SwapChain::DetachFromSurfaceImpl
  dawn::native::Surface::~Surface
  rnwgpu::SurfaceInfo::detach(bool)
  Java_com_webgpu_WebGPUView_onViewDestroyed

iOS (Metal) never crashes with the same code.

Root cause

three.js's WebGPURenderer.dispose() calls device.destroy() synchronously in the React effect cleanup. On Android the native view is dropped one frame later, so SurfaceInfo::detach() releases a wgpu::Surface whose swapchain is still attached. ~Surface then runs SwapChain::DetachFromSurfaceImpl, which reads the already-destroyed device's FencedDeleter. That pointer is freed, so Dawn dereferences null at 0x20.

Calling Surface.Unconfigure() at destroy time is not enough. What touches the device is the swapchain detach inside ~Surface, so the surface has to be released (its destructor has to run) while the device is still alive.

The fix

GPUDevice::destroy() now walks the SurfaceRegistry and, for every SurfaceInfo bound to this device, unconfigures and releases the wgpu::Surface right away, before _instance.Destroy(). Releasing it runs ~Surface and SwapChain::DetachFromSurface against a live device. The native window pointer stays in _nativeSurface, so the later SurfaceInfo::detach() from the view teardown still returns the window to the platform.

Two small additions, with no change on the happy path. SurfaceInfo::unconfigureIfDevice(device) detaches and releases the surface when it belongs to device, and SurfaceRegistry::unconfigureDevice(device) calls it for every entry.

Verification

Pixel emulator, Android API 37, -gpu host (real Vulkan via virtio-gpu), minSdkVersion 26, three.js WebGPURenderer rendering a glTF in a <Canvas> inside a React Navigation screen.

Before the fix: navigate in, wait for the first frame, press back, repeat. Crashes 2 out of 3 times.

After the fix: the same loop with renderer.dispose() called immediately in cleanup, no workaround on the app side. Five clean exits out of five, same PID throughout. Instrumentation confirms the detach after teardown now sees surface=0, already released.

It also runs on iOS, where unconfigureDevice is a cheap walk that does nothing but keeps both platforms on the same teardown path.

A standalone reproduction is at https://github.com/eduardoborges/react-native-webgpu-unmount-repro.

[wcandillon]

Let me know if you can provide me with a reproducible example

[eduardoborges]

Standalone reproduction (RN 0.87.1, react-native-webgpu 0.9.0, three 0.184.0): https://github.com/eduardoborges/react-native-webgpu-unmount-repro

Mount the medal, then hide it. On an Android emulator with a real Vulkan GPU (-gpu host) the app crashes on hide without this PR, and survives every toggle with it.

[wcandillon]

That's a great find, thank you for this 🙏

[wcandillon]

@eduardoborges I've made a bunch of commits to your PR but I will revert it. I am considering a completely different approach to that problem

[eduardoborges]

@wcandillon no problem! feel at home

[wcandillon]

@eduardoborges Sorry I messed your branch and thank you for bringing this to my attention. I've made a PR against Dawn to fix it: google/dawn#84
Now I am on the verge to merge/publish a workaround for it in the meantime: #470

closing in favor of #470