Potential memory leak in RingBuffer

[Cole Richardson]

When I run the following code, Valgrind reports that there was 300 bytes leaked from malloc:

auto rb = RingBuffer();
rb.Alloc(3, 100);

I took a look at the code in question (RingBuffer.inl) and I think the malloc'd memory pointer is being overwritten by a call to cudaMalloc?

mBuffers[n] = malloc(size);

if( CUDA_FAILED(cudaMalloc(&mBuffers[n], size)) )
{
    LogError(LOG_CUDA "RingBuffer -- failed to allocate CUDA buffer of %zu bytes\n", size);
    return false;
}

It looks like cudaFree is being called on each of the buffers but not free. If the ring buffer is only designed to allocate GPU/Mapped memory, is this malloc even needed?

[Dustin Franklin]

Ahh, thanks @SeeRich, yes this would appear to be a bug. If you remove mBuffers[n] = malloc(size); and recompile, do you get any runtime errors?

[Cole Richardson]

@dusty-nv If I comment out that line, the Valgrind report is clean.

[Cole Richardson]

Just from searching the code, it doesn't look like this is being used. RingBuffer seems to only be used in gstBufferManager.cpp and gstEncoder.cpp and both use cases specify RingBuffer::ZeroCopy which would bypass this issue. On the other hand, if a consumer of this project was using non-ZeroCopy, they would probably see a performance gain (and lower memory usage).

[Dustin Franklin]

Thanks @SeeRich!, yes agreed 👍

Patched this in d374d42 and updated the submodule in jetson-inference in dusty-nv/jetson-inference@aa91738