> > In processor_idle.c target residency is calculated by exit latency multiplied by > > a latency factor parameter (2 by default). So when these systems have the > buggy > > _CST C2 value of 400, that's where the 800 comes from for residency, not > _LPI. > > Yes, that's right, sorry for the confusion. > > My point was mostly whether or not it would be sufficient to swap the > values for the last two states or if some new values needed to be used > and the answer appears to be the latter. > > So I don't see any way to address this that would be cleaner than the > quirk mechanism as proposed. The root of the problem is that the states get skipped. So yes technically your suggestion of swapping the values of the last two states would be enough to avoid this issue as well. My thought is to do it by checking whether all the latencies are arithmetic progressions and if not then emit a notice this is a FW bug and sort at the end of acpi_processor_power_verify(pr). This would be done on all systems. Let me know if that sounds good or if you would prefer to stick to the quirk approach as proposed or something hybrid.
