Observing Host OS panic/crash due to use-after-free error related to CPU OS memory when gdr_unmap is not called before gdr_close

[realarnavgoel]

Impacted platform
All server side products, first observed on Grace-Hopper system

Impacted gdrcopy versions
2.0 and later

Impacted gdrcopy configs
Both persistent and non-persistent mode

Scenarios
If an application opens a connection to the driver (gdr_open), allocates a GPU memory via CUDA, pins and maps the allocated memory to CPU (gdr_pin_buffer, gdr_map) for read/write operations. Subsequently, if the application closes the connection, without explicitly unmapping the GPU memory, it results in a use-after-free (UAF) condition of OS memory, which can result in functional issues in unrelated areas, or even kernel panic or crash.

Known Mitigations
applications should explicitly call gdr_unmap before gdr_close.

Fixed gdrcopy version
2.4.4