[BUG]: cudaGraphGetEdges() suffers use-after-free for cudaGraphEdgeData

[galv]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Silent Failure

Component

cuda.bindings

Describe the bug

Take a look here:

https://github.com/NVIDIA/cuda-python/blob/346afc808db7f7573f9b7a7c2d297868ea7e4894/cuda_bindings/cuda/bindings/runtime.pyx.in#L35898-L35901

We are creating a python cudaGraphEdgeData object via a pointer to the underlying C cudaGraphEdgeData struct.

That underlying constructor https://github.com/galv/cuda-python/blob/346afc808db7f7573f9b7a7c2d297868ea7e4894/cuda_bindings/cuda/bindings/runtime.pyx.in#L18241 just saves that pointer, rather doing a deep copy. Thus, it relies upon that underlying pointer staying alive.

Then, cudaGraphGetEdges() immediately frees the underlying pointer that the python objects reference, https://github.com/NVIDIA/cuda-python/blob/346afc808db7f7573f9b7a7c2d297868ea7e4894/cuda_bindings/cuda/bindings/runtime.pyx.in#L35901, and returns the list of python objects which are now backed by dead pointers.

Thus, the cudaGraphEdgeData returned by cudaGraphGetEdges() always suffers a use-after-free. You need to do a deep-copy.

The driver code suffers from the same issue.

Please fix this.

How to Reproduce

I don't think you need a repro to see the problem in this case.

Expected behavior

Do a deep copy of the underlying data.

Operating System

No response

nvidia-smi output

No response