An old bugfix was missed in driver version 5.3. I've had this patch for 2 years and I think I tested it at the time but we have moved past that version.
I discovered that mellanox/nvidia provide a smaller tarball we can download to build netmap that does not bundle large debs. This smaller tarball has the same sha256sum as the first inner tarball of the larger one.
We are starting to test driver version 5.8 but early tests show no obvious problems with TX or RX. If others want to test it please do.
We use dkms to package but never shared the "recipe".