WebGPU Unbounded Heap Growth (Soak Test)

[thejustinwalsh]

Discovered a slow steady memory growth leak through an accidental soak test. I left the device running for several hours and came back to a low framerate.

Environment

• @nativescript/canvas 2.0.32 (master cd52d3d, 2026-05-07)
• Physical Android: Moto G 2025 (accurate oracle). Also reproducible on iOS simulator (noisier baseline).
• Renderer: three WebGPU backend (three.webgpu.js) via @nativescript/canvas-three, and the
  canvas plugin's own webgpuTriangle example.
• NativeScript runtime (V8) on both platforms.

Summary

Rendering WebGPU continuously (frameloop always) leaks native heap until the render loop stops.

Findings

Measured on a physical Moto G 2025, frameloop always, via on-device getNativeHeapAllocatedSize() and heapprofd with offline symbolization.

Three reference leaks (all native handles we held, none GC-related)

#	Leak	Fix
1	Per-frame transient wrappers (command encoder, render/compute pass, texture view, command buffer) freed only by the GC finalizer	Release at the WebGPU consumption point (end/finish/submit/present); native handle held in ArcHandle so release is exactly-once
2	present_surface read-back copy created a command encoder + command buffer every frame, never dropped	command_buffer_drop + command_encoder_drop in the read-back block
3	getCurrentTexture() registered a Texture + auto clear_view in wgpu-core's hub every frame; CanvasGPUTexture::drop was a no-op for surface textures	texture_drop (and discard-if-unpresented) in the surface branch

On-device result (foreground soak)

build	rate
baseline (released)	OOM-restarts ~115 MB; ~47 KB/frame
+ fix 1	0.31 MB/s
+ fix 2	0.18 MB/s
+ fix 3	+0.002 MB/s over 5 min — flat (= blank-app baseline)

What it was NOT

• Not GC starvation: global.gc() every 30 frames had no effect — the growth was
  reachable native memory, not collectible JS.
• Not pooling / wrapper shells: heapprofd showed the C++/V8 shells were negligible;
  the bytes were wgpu resources held in the hub.

Method

heapprofd growth analysis (per-callsite live-bytes delta, first vs last dump) + llvm-addr2line against an unstripped local build (LTO/ICF off so frames don't fold). Named the growers exactly: surface_get_current_texture (present.rs:220), device_create_command_encoder.