The minic driver reads the TX timestamp from the device after sending.
Unfortunately, it only waited for the data to be enqueued to be sent.
Sometimes this lead to a race condition where the old TX stamp was read.

As a result, the PTP calculation goes wrong by as much as an entire second.
When in track_phase, this manifests as a bad phase offset sent to softpll.
The softpll takes years to reach that phase offset, so it is always busy.
Thus the clock drifts further and further away, never correcting.