(From Stelian Pop)
Upon investigation, the reason is that trying to set the performance
to 80% in longrun_determine_freqs leaves the performance to 100%.
The performance level, at least on this particular model, can be lowered
only in 33% steps. And in order to put the performance to 66%, the
code should try to set the barrier to 70%.
                                                                                                           
The following patch does even more, it tries every value from 80%
to 10% in 10% steps, until it succeeds in lowering the performance.
I'm not sure this is the best way to do it but in any case,
it works for me (and should continue to work for everybody else).