On 10-06-17, 23:21, Joel Fernandes wrote: > On Sat, Jun 10, 2017 at 2:11 AM, Joel Fernandes <joe...@google.com> wrote: > > On Fri, Jun 9, 2017 at 3:15 AM, Viresh Kumar <viresh.ku...@linaro.org> > > wrote: > >> While reducing frequency if there are no frequencies available between > >> "current" and "next" calculated frequency, then the core will never > >> select the "next" frequency. > >> > >> For example, consider the possible range of frequencies as 900 MHz, 1 > >> GHz, 1.1 GHz, and 1.2 GHz. If the current frequency is 1.1 GHz and the > >> next frequency (based on current utilization) is 1 GHz, then the > >> schedutil governor will try to set the average of these as the next > >> frequency (i.e. 1.05 GHz). > >> > >> Because we always try to find the lowest frequency greater than equal to > >> the target frequency, cpufreq_driver_resolve_freq() will end up > >> returning 1.1 GHz only. And we will not be able to reduce the frequency > >> eventually. The worst hit is the policy->min frequency as that will > >> never get selected after the frequency is increased once. > > > > But once utilization goes to 0, it will select the min frequency > > (because it selects lowest frequency >= target)? > > Never mind my comment about util 0, I see the problem you mention. > However I feel that this entire series adds complexity all to handle > the case of a false cache-miss which I think might not be that bad, > and the tradeoff with complexity/readability of the code kind of > negates the benefit. That's just my opinion about it fwiw.
Right and that's why I said in the cover letter that we may want to revert the offending commit for the time being as the solutions provided here have too much dependency on the resolve_freq() callback. -- viresh