GitLab

There seems to be a hardware race condition in the data channel where
we could end up receiving IRQs for more data than we have available.
Fix this by disabling IRQs when we run out of data; if the device
really does want more data, we'll underrun and flag an error
upstream.