Wgpu: Memory issue

[lordnoriyuki]

Seeing this on latest wgpu-rs:

thread 'main' panicked at 'Memory must not be used anywhere: Memory { heap: Public(MemoryTypeId(1), <GFX9AMD_MtlBuffer: 0x7fb98b808a60>
    label = <none> 
    length = 33554432 
    cpuCacheMode = MTLCPUCacheModeDefaultCache 
    storageMode = MTLStorageModeShared 
    hazardTrackingMode = MTLHazardTrackingModeTracked 
    resourceOptions = MTLResourceCPUCacheModeDefaultCache MTLResourceStorageModeShared MTLResourceHazardTrackingModeTracked  
    purgeableState = MTLPurgeableStateNonVolatile), size: 33554432 }', /Users/ln/.cargo/git/checkouts/wgpu-53e70f8674b08dd4/549ea00/wgpu-core/src/device/alloc.rs:227:46
stack backtrace:
   0: rust_begin_unwind
             at /rustc/d9a105fdd46c926ae606777a46dd90e5b838f92f/library/std/src/panicking.rs:493:5
   1: core::panicking::panic_fmt
             at /rustc/d9a105fdd46c926ae606777a46dd90e5b838f92f/library/core/src/panicking.rs:92:14
   2: core::option::expect_none_failed
             at /rustc/d9a105fdd46c926ae606777a46dd90e5b838f92f/library/core/src/option.rs:1268:5
   3: gpu_alloc::linear::LinearAllocator<M>::dealloc
   4: gpu_alloc::allocator::GpuAllocator<M>::dealloc
   5: wgpu_core::device::life::LifetimeTracker<B>::cleanup
   6: wgpu_core::device::Device<B>::maintain
   7: wgpu_core::device::queue::<impl wgpu_core::hub::Global<G>>::queue_submit
   8: <wgpu::backend::direct::Context as wgpu::Context>::queue_submit

I'll get a trace.

[lordnoriyuki]

Closing in favor of opening issue on wgpu directly

[kvark]

@lordnoriyuki no need, we can transfer issues between repos

[kvark]

@zakarumych anything we missed, still, in the linear allocator?

[zakarumych]

This happens if Memory's clone is still around when allocator returns it to device.

With gfx backend gpu-alloc operate with 'Arc\

All blocks should be deallocated when memory object gets deallocated, but user of MemoryBlock could clone its parent 'Arc\

[lordnoriyuki]

Trace attached.

trace.zip

[lordnoriyuki]

And it was doing a lot of work - performance was really bad before it died (1000 fps in this particular scenario normally was <200 fps).

[kvark]

@zakarumych this is our wrapper - https://github.com/gfx-rs/wgpu/blob/master/wgpu-core/src/device/alloc.rs
It doesn't expose Arc<Memory> to the outside, and just a fairly boring wrapper.
Could this issue also be caused by a logic flow in gpu-alloc itself?

[zakarumych]

It can be anything.
gpu-alloc clones memory only when new block is allocated.

It is possible that gpu-alloc tries to deallocate memory before all blocks are freed.
Printing whole Allocator can reveal if this is the case.

[zakarumych]

Found the issue.
Last block that gets deallocated was holding another clone of the Arc.
Fixed in https://github.com/zakarumych/gpu-alloc/commit/d07be73f9439a37c89f5b72f2500cbf0eb4ff613

[lordnoriyuki]

Confirmed if I patch the new gpu-alloc everything is working now.