Messages in this thread | | | From | "Rafael J. Wysocki" <> | Date | Thu, 6 Sep 2018 11:18:55 +0200 | Subject | Re: [PATCH v6 13/14] sched/topology: Make Energy Aware Scheduling depend on schedutil |
| |
On Tue, Sep 4, 2018 at 12:59 PM Quentin Perret <quentin.perret@arm.com> wrote: > > On Monday 20 Aug 2018 at 10:44:19 (+0100), Quentin Perret wrote: > > Energy Aware Scheduling (EAS) is designed with the assumption that > > frequencies of CPUs follow their utilization value. When using a CPUFreq > > governor other than schedutil, the chances of this assumption being true > > are small, if any. When schedutil is being used, EAS' predictions are at > > least consistent with the frequency requests. Although those requests > > have no guarantees to be honored by the hardware, they should at least > > guide DVFS in the right direction and provide some hope in regards to the > > EAS model being accurate. > > > > To make sure EAS is only used in a sane configuration, create a strong > > dependency on schedutil being used. Since having sugov compiled-in does > > not provide that guarantee, extend the existing CPUFreq policy notifier > > with a new case on governor changes. That allows the scheduler to > > register a callback on this notifier to rebuild the scheduling domains > > when governors are changed, and enable/disable EAS accordingly. > > > > cc: Ingo Molnar <mingo@redhat.com> > > cc: Peter Zijlstra <peterz@infradead.org> > > Signed-off-by: Quentin Perret <quentin.perret@arm.com> > > > > --- > > This patch could probably be squashed into another one, but I kept it > > separate to ease the review. Also, it's probably optional as not having > > it will not 'break' things per se. > > > > I went for the smallest possible solution I could find, which has the > > good side of being simple, but it's definitely not the only one. > > > > Another possibility would be to hook things in sugov_start() and > > sugov_stop(), but that requires some more work. In this case, it > > wouldn't be possible to just re-build the sched_domains() from there, > > because when sugov_stop() is called, the 'governor' field of the policy > > hasn't been updated yet, so the condition (if gov == schedutil) in > > build_freq_domains() doesn't work. > > > > To workaround the issue we'll need to find a way to pass a cpumask to > > the topology code to specifically say 'sugov has been stopped on these > > CPUs'. That would mean more code to handle that, but that would also > > mean we don't have to mess around with the CPUFreq notifiers ... > > > > Not sure what's best, so all feedback is more than welcome. > > Hi, > > Does anybody have concerns with this patch ? Is it a reasonable option > to use the CPUFreq notifiers in this case ? If there is anything I can > do to ease the review please let me know.
I'm not a particular fan of notifiers to be honest and you don't need to add an extra chain just in order to be able to register a callback from a single user. That can be achieved with a single callback pointer too, but also you could just call a function exported by the scheduler directly from where in the cpufreq code it needs to be called.
> Also, is there any hope that the 12 first patches could make it in 4.20 > on their own ? Or is it already too late ?
I'm walking through them right now, albeit somewhat slowly due to various distractions, so we'll see.
Thanks, Rafael
| |