The test for pid_filter (and pgrp_filter) currently
causes a return from op_do_nmi in case a filter has
been set and the current->pid doesn't match. When it
returns from the interrupt without resetting the
perfctr that caused the NMI, the next interrupt either
doesn't come, or takes a long time to wrap back around
(?). The result is that new samples are not taken.
I've attached my patch for 0.0.7.