[enzomanuelmangano]

Note

Written by Claude Opus 5.5 on behalf of Enzo (@enzomanuelmangano).

Fixes #581.

Problem

On Android, importDevice(Skia.getNativeDevice()) returns a device whose async work never settles (mapAsync stays pending). createInstance() looks Skia's getter up with dlsym(RTLD_DEFAULT, "rnskia_getWGPUInstance"). System.loadLibrary loads librnskia.so with RTLD_LOCAL, so the global lookup does not see it, even though librnskia.so exports the symbol. react-native-webgpu then creates its own instance and never pumps Skia's device.

Fix

On Android, when the global lookup fails, createInstance() asks for the handle of librnskia.so with dlopen("librnskia.so", RTLD_NOW | RTLD_NOLOAD) and looks the getter up on that handle.

• RTLD_NOLOAD returns the library only if it is already loaded. This never loads Skia into an app that does not use it, and it adds no build-time dependency.
• dlclose releases the reference that dlopen took. The library stays loaded, because System.loadLibrary holds it, so the function pointer stays valid.
• Both .so files are in the app's class loader namespace, so the lookup by soname resolves.
• iOS does not change. Its symptom has another cause, fixed in react-native-skia (iOS: react-native-webgpu never finds rnskia_getWGPUInstance (the linker drops it), so Skia's imported device never settles async work react-native-skia#4305, fix(📱): keep rnskia_getWGPUInstance linked on iOS so react-native-webgpu shares Skia's Dawn instance react-native-skia#4306).

The ordering requirement does not change: as with the existing dlsym, Skia must be loaded before the first GPU object is created.

Alternative considered: dl_iterate_phdr over all loaded objects. It finds the library without knowing its name, but it is more code and walks every loaded library to find one known soname.

Test plan

Reproduction app: https://github.com/enzomanuelmangano/repro-lab, demo webgpu-skia-shared-instance. Its fix/webgpu-skia-shared-instance-android branch applies this change to 0.17.0 (diff). Android API 36 emulator (arm64), release build, react-native-skia 3.7.2.

	Before	After
mapAsync on Skia's imported device	never settles (3 s timeout)	settles in 58 ms

Before	After

Video, Redmi Note 8 Pro, release build (left: 0.17.0, stays at "Running…"; right: with the fix, settles in 16 ms): https://github.com/enzomanuelmangano/repro-lab/blob/main/media/webgpu-skia-shared-instance-android.mp4

Not run: the repository's E2E suite (yarn test) and the example app. The change only adds a fallback when the existing lookup returns nullptr.

[camchis]

Confirming this on a device. I hit the same bug and patched it the same way, dlopen("librnskia.so", RTLD_NOW | RTLD_NOLOAD) then dlsym on that handle, on react-native-webgpu 0.13.0 with react-native-skia 3.3.0 (Graphite). On a Seeker (Android, Vulkan), mapAsync on buffers of a device imported with importDevice(Skia.getNativeDevice()) never settled before the change and settles normally after it. Without it WebGPU makes its own instance, and an app that mixes Skia and WebGPU waits forever on the Skia device's work. Looks good to me.

[wcandillon]

this is correct, wouldn't be cleaner for this to be in WebGPUModule.java,?

[enzomanuelmangano]

Note

Written by Claude Opus 5.5 on behalf of Enzo (@enzomanuelmangano).

Thanks a lot, and thanks @camchis for confirming on a Seeker!

Good question. I tested it on a Redmi Note 8 Pro (release build, mapAsync on a buffer of importDevice(Skia.getNativeDevice())):

Build	mapAsync
0.17.0	never settles
System.loadLibrary("rnskia") before react-native-webgpu in WebGPUModule.java, no native change	never settles
RTLD_NOLOAD lookup in C++ only (this PR)	settles in 17 ms
both	settles in 17 ms

So Java is a great place for the loading order you mentioned, and it would make this deterministic, but it can't replace the lookup: System.loadLibrary loads librnskia.so with RTLD_LOCAL, and that visibility can't be changed from Java, so dlsym(RTLD_DEFAULT, …) still misses the symbol. The only all-Java alternative I can think of is asking Skia's Java module for the instance pointer, which would couple the two packages.

If you'd like, I'm happy to add the loadLibrary("rnskia") ordering to WebGPUModule.java in this PR as well, or leave it to your follow-up. Whatever you prefer!