Repository navigation
nvidia-uvm: propagate EGM DMA mapping failures - #1423
arpitkjaiswal wants to merge 1 commit into
Conversation
Return the recorded status from uvm_gpu_map_cpu_pages_for_egm() instead of unconditionally reporting NV_OK. This propagates DMA link and sync failures and restores the caller's NV_ERR_NOT_SUPPORTED fallback when the DMA IOVA API is unavailable. Add host regression tests that compile the actual mapping function and its caller with platform stubs in both API configurations. Verify error propagation, unlink behavior, accounting, and mapping-cache updates.
|
uvm_gpu_map_cpu_pages_for_egm() records DMA link/sync failures and NV_ERR_NOT_SUPPORTED, but always returns NV_OK. Its caller consequently records a mapping that was never established, and the unsupported-API fallback is unreachable. Return the recorded status. Existing success paths, sync-failure cleanup, output handling, and byte accounting are preserved. The caller can now propagate real errors or fall back to non-EGM access without caching an invalid mapping. Add a host regression suite that extracts and compiles the actual mapping function and uvm_cpu_chunk_map_gpu_egm() with platform stubs, with the DMA IOVA API both enabled and disabled. Validation: All 10 tests pass after the fix. Tests use GCC 13.3 and Python 3.12. Platform dependencies are stubbed; a full kernel module build and GPU hardware testing have not been performed. |
1 participant
Return the recorded status from uvm_gpu_map_cpu_pages_for_egm() instead of unconditionally reporting NV_OK. This propagates DMA link and sync failures and restores the caller's NV_ERR_NOT_SUPPORTED fallback when the DMA IOVA API is unavailable.
Add host regression tests that compile the actual mapping function and its caller with platform stubs in both API configurations. Verify error propagation, unlink behavior, accounting, and mapping-cache updates.