The correct way to implement it would be to prevent the initiation of DMA operations when the device is locked, but to let current ones continue. I have no idea if that's what happened, though.
I'm unsure how the DMA operations are structured though. A single 'operation' from the user's perspective might involve multiple operations at a lower level where locking in the middle might allow a single low-level atomic (in-progress) operation to complete, but the overall job to fail.
