On Wednesday, March 19, 2014 03:30:48 PM Srivatsa S. Bhat wrote: > On 03/19/2014 10:33 AM, Viresh Kumar wrote: > > On 18 March 2014 17:46, Srivatsa S. Bhat > > <srivatsa.b...@linux.vnet.ibm.com> wrote: > >> Agreed. As far as I understand, for ->target drivers, today we use GOV_STOP > >> to stop managing the CPU going offline. And for ->setpolicy drivers, we > >> will > >> use this new callback to achieve the same goal. > > > > So a better question would be: What's the purpose of ->stop() call for a > > policy? > > Ideally, it should remove the outgoing CPU from the policy and "stop managing > that CPU", whatever that means to the driver (for intel_pstate, it means > setting it to min P state and destroying the timer). > > > Stop managing CPUs of that policy? > > Stop managing only the particular CPU going offline. IOW, we should somehow > communicate to the ->stop() callback that we are taking CPU 'x' offline. > > If adding a ->stop() callback in the cpufreq_driver is not the best way to > achieve it, then lets think of an alternative. The way I look at it, this > new mechanism what we want, should allow ->setpolicy drivers to do what the > GOV_STOP will do for regular drivers. That is, allow it to "shutdown the > CPU from a cpufreq perspective", whatever that means to the driver. > We can think of a completely different way of achieving it, if ->stop() > is not suitable for that purpose.
I agree. That said, for the intel_pstate case ->stop() as proposed by Dirk is demonstrably sufficient and there are no other ->setpolicy drivers in sight wanting or needing anything else. So to me, (1) the new ->stop() should *only* be called for ->setpolicy drivers, because the purpose of it should be to "allow ->setpolicy drivers to do what the GOV_STOP will do for regular drivers" as you put it above, and (2) some code in the original intel_pstate's ->exit() may/should stay in there (instead of being moved to the new ->stop()), which is the only possibly remaining issue here. The whole discussion about possibly re-using ->stop() for ->target drivers goes in a totally wrong direction, because *if* ->target drivers need a new callback to be executed around where ->stop() is called for ->setpolicy drivers, *then* that has to be a *different* callback. And by the way, ->get() in fact has a different meaning for ->setpolicy drivers, so it would be good to consider logical separation of ->setpolicy and ->target drivers so that each kind has its own separate set of callbacks with no overlaps. That would make it easier to avoid breakage resulting from changes made with ->setpolicy drivers that also affect ->target drivers in unpredictable ways and the other way around. -- I speak only for myself. Rafael J. Wysocki, Intel Open Source Technology Center. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
